diff --git a/FP8_Emulator/cmodel/__init__.py b/FP8_Emulator/cmodel/__init__.py new file mode 100644 index 00000000..9809e350 --- /dev/null +++ b/FP8_Emulator/cmodel/__init__.py @@ -0,0 +1 @@ +from . import simple diff --git a/FP8_Emulator/cmodel/simple.py b/FP8_Emulator/cmodel/simple.py new file mode 100644 index 00000000..608de9d4 --- /dev/null +++ b/FP8_Emulator/cmodel/simple.py @@ -0,0 +1,219 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import torch +from torch import nn +from torch.autograd import Function +import sys + +import simple_gemm_dev +import simple_conv2d_dev +# backup the original torch functions +fallback_addmm = torch.addmm +fallback_matmul = torch.matmul +fallback_mm = torch.mm + +def is_transposed(input): + if input.is_contiguous(): + return input, False + elif input.t().is_contiguous(): + return input.t(), True + else: + return input.contiguous(), False + +def addmm(input, mat1, mat2, beta=1.0, alpha=1.0, out=None): + if input.dtype == torch.float32 and mat1.dtype == torch.float32 and \ + mat1.dim() == 2 and mat2.dim() == 2 and mat1.size(1) == mat2.size(0): + if out: + output = out + else: + output = torch.zeros([mat1.size(0), mat2.size(1)]) + a_mat, a_trans = is_transposed(mat1) + b_mat, b_trans = is_transposed(mat2) + output = SimpleAddmm.apply(output, input, a_mat, b_mat, alpha, beta, a_trans, b_trans) + ret = output + else: + warnings.warn('simple.addmm does not support the input dimensions - input :{}, mat1: {}, mat2: {}, falling back to torch.addmm'.format( + input.size(), mat1.size(), mat2.size())) + ret = fallback_addmm(input, mat1, mat2, beta=beta, alpha=alpha, out=out) + return ret + +def matmul(input, other, out=None): + if input.dtype == torch.float32 and other.dtype == torch.float32 and \ + input.dim() == 2 and other.dim() == 2 and input.size(1) == other.size(0): + if out: + output = out + else: + output = torch.zeros([input.size(0), other.size(1)]) + a_mat, a_trans = is_transposed(input) + b_mat, b_trans = is_transposed(other) + output = SimpleMatmul.apply(output, a_mat, b_mat, 1.0, a_trans, b_trans) + return output + # Batch MatMul implementation + elif input.dtype == torch.float32 and other.dtype == torch.float32 and \ + input.dim() == 3 and other.dim() == 2 and input.size(2) == other.size(0): + if out: + output = out + else: + output = torch.zeros([input.size(0), input.size(1), other.size(1)]) + a_mat, a_trans = is_transposed(input) + b_mat, b_trans = is_transposed(other) + output = torch.stack(tuple([SimpleMatmul.apply(out1, a_mat1, b_mat, 1.0, a_trans, b_trans) \ + for a_mat1, out1 in zip(a_mat, output)])) + return output + else: + warnings.warn('simple.matmul does not support the input dimensions - input :{}, other: {}, falling back to torch.matmul'.format( + input.size(), other.size())) + return fallback_matmul(input, other, out=out) + +def mm(input, mat2, out=None): + if input.dtype == torch.float32 and mat2.dtype == torch.float32 and \ + input.dim() == 2 and mat2.dim() == 2 and input.size(1) == mat2.size(0): + if out: + output = out + else: + output = torch.zeros([input.size(0), mat2.size(1)]) + a_mat, a_trans = is_transposed(input) + b_mat, b_trans = is_transposed(mat2) + output = SimpleMatmul.apply(output, a_mat, b_mat, 1.0, a_trans, b_trans) + return output + else: + warnings.warn('simple.mm does not support the input dimensions - input :{}, mat2: {}, falling back to torch.mm'.format( + input.size(), mat2.size())) + return fallback_mm(input, mat2, out=out) + +def conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1): + N = input.size()[0] + C = input.size()[1] + H = input.size()[2] + W = input.size()[3] + K = weight.size()[0] + C1 = weight.size()[1] + R = weight.size()[2] + S = weight.size()[3] + + if dilation[0] > 1: + sys.exit("ERROR: simple_conv2d does not support dilated convolutions.") + if padding[0] != padding[1]: + sys.exit("ERROR: simple_conv2d does not support non-uniform padding; pad_h must be equal to pad_w.") + if groups > 1: + sys.exit("ERROR: simple_conv2d does not support grouped convolutions; set groups to 1.") + + H_out = ((H + (2*padding[0]) - dilation[0] * (R-1) -1)/stride[0]) + 1 + W_out = ((W + (2*padding[1]) - dilation[1] * (S-1) -1)/stride[1]) + 1 + output = torch.empty([N, K, int(H_out), int(W_out)]) + output = SimpleConv2dFunction.apply(output, input, weight, bias, stride, padding, dilation, groups) + return output + +class SimpleAddmm(Function): + @staticmethod + def forward(ctx, output, input, mat1, mat2, alpha, beta, a_trans, b_trans): + ctx.save_for_backward(mat1, mat2) + ctx.a_trans = a_trans + ctx.b_trans = b_trans + ctx.alpha = alpha + + simple_gemm_dev.gemm(output, mat1, mat2, alpha, a_trans, b_trans) + output += beta * input; + ctx.mark_dirty(output) + return output + + @staticmethod + def backward(ctx, grad_output): + mat1, mat2 = ctx.saved_tensors + + alpha = ctx.alpha + a_trans = ctx.a_trans + b_trans = ctx.b_trans + + grad_mat1 = torch.zeros_like(mat1) + grad_mat2 = torch.zeros_like(mat2) + grad_out, out_trans = is_transposed(grad_output) + + if a_trans: + simple_gemm_dev.gemm(grad_mat1, mat2, grad_out, alpha, b_trans, not out_trans) + else: + simple_gemm_dev.gemm(grad_mat1, grad_out, mat2, alpha, out_trans, not b_trans) + + if b_trans: + simple_gemm_dev.gemm(grad_mat2, grad_out, mat1, alpha, not out_trans, a_trans) + else: + simple_gemm_dev.gemm(grad_mat2, mat1, grad_out, alpha, not a_trans, out_trans) + + return (grad_output, grad_output, grad_mat1, grad_mat2, None, None, None, None) + +class SimpleMatmul(Function): + @staticmethod + def forward(ctx, output, mat1, mat2, alpha, a_trans, b_trans): + ctx.save_for_backward(mat1, mat2) + ctx.a_trans = a_trans + ctx.b_trans = b_trans + ctx.alpha = alpha + + simple_gemm_dev.gemm(output, mat1, mat2, alpha, a_trans, b_trans) + ctx.mark_dirty(output) + return output + + @staticmethod + def backward(ctx, grad_output): + mat1, mat2 = ctx.saved_tensors + alpha = ctx.alpha + a_trans = ctx.a_trans + b_trans = ctx.b_trans + + grad_mat1 = torch.empty_like(mat1) + grad_mat2 = torch.empty_like(mat2) + grad_out, out_trans = is_transposed(grad_output) + + if a_trans: + simple_gemm_dev.gemm(grad_mat1, mat2, grad_out, alpha, b_trans, not out_trans) + else: + simple_gemm_dev.gemm(grad_mat1, grad_out, mat2, alpha, out_trans, not b_trans) + + if b_trans: + simple_gemm_dev.gemm(grad_mat2, grad_out, mat1, alpha, not out_trans, a_trans) + else: + simple_gemm_dev.gemm(grad_mat2, mat1, grad_out, alpha, not a_trans, out_trans) + return (grad_output, grad_mat1, grad_mat2, None, None, None) + + +class SimpleConv2dFunction(Function): + @staticmethod + def forward(ctx, output, inputs, weights, bias, stride, padding, dilation, groups): + #print("### conv2d fwd called input size: ", inputs.size(), weights.size(), stride, padding, dilation, groups) + ctx.save_for_backward(inputs, weights)#, bias) + ctx.stride = stride#[0] + ctx.padding = padding#[0] + ctx.dilation = dilation#[0] + ctx.groups = groups + + if bias is None: + bias_fw = torch.zeros(output.size()[1]) + else : + bias_fw = bias + + simple_conv2d_dev.conv2d_fp(output, inputs, weights, bias_fw, stride[0], padding[0], dilation[0], groups) + ctx.mark_dirty(output) + return output + + @staticmethod + def backward(ctx, grad_output): + #inputs, weights, bias = ctx.saved_tensors + inputs, weights = ctx.saved_tensors + stride = ctx.stride + padding = ctx.padding + dilation = ctx.dilation + groups = ctx.groups + #print("### conv2d bwd called input size: ", inputs.size(), weights.size(), stride, padding, dilation, groups) + grad_inp = torch.zeros_like(inputs) + grad_wts = torch.zeros_like(weights) + + simple_conv2d_dev.conv2d_bp(grad_inp, grad_output, weights, stride[0], padding[0], dilation[0], groups) + simple_conv2d_dev.conv2d_wu(grad_wts, grad_output, inputs, stride[0], padding[0], dilation[0], groups) + return (grad_output, grad_inp, grad_wts, None, None, None, None, None) diff --git a/FP8_Emulator/cmodel/simple/simple_conv2d.cpp b/FP8_Emulator/cmodel/simple/simple_conv2d.cpp new file mode 100644 index 00000000..8498efe7 --- /dev/null +++ b/FP8_Emulator/cmodel/simple/simple_conv2d.cpp @@ -0,0 +1,127 @@ +/*----------------------------------------------------------------------------* + * Copyright (c) 2023, Intel Corporation - All rights reserved. + * This file is part of FP8-Emulation-Toolkit + * + * SPDX-License-Identifier: BSD-3-Clause + *----------------------------------------------------------------------------* + * Naveen Mellempudi (Intel Corporation) + *----------------------------------------------------------------------------*/ + +#include +#include +#include +#include +#include +#include +#include + +extern int simple_conv2d_impl_fp(float* outputs, float *inputs, float *weights, float* bias, int N, int C, int iH, int iW, + int K, int R, int S, int stride, int padding, int dilation, int groups); +extern int simple_conv2d_impl_bp(float* inputs, float *outputs, float *weights, int N, int C, int iH, int iW, + int K, int R, int S, int stride, int padding, int dilation, int groups); +extern int simple_conv2d_impl_wu(float *weights, float *outputs, float *inputs, int N, int C, int iH, int iW, + int K, int R, int S, int stride, int padding, int dilation, int groups); + +#define gettid() ((int)syscall(SYS_gettid)) + +using namespace torch::autograd::profiler; + + +double get_time() { + static bool init_done = false; + static struct timespec stp = {0,0}; + struct timespec tp; + clock_gettime(CLOCK_REALTIME, &tp); + + if(!init_done) { + init_done = true; + stp = tp; + } + double ret = (tp.tv_sec - stp.tv_sec) * 1e3 + (tp.tv_nsec - stp.tv_nsec)*1e-6; + return ret; +} + +at::Tensor simple_conv2d_fp(torch::Tensor& output, torch::Tensor input, torch::Tensor weight, torch::Tensor bias, + int stride, int padding, int dilation, int groups) +{ + RECORD_FUNCTION("simple_conv2d_fp", std::vector({input, weight, bias})); + + auto N = input.size(0); + auto C = input.size(1); + auto H = input.size(2); + auto W = input.size(3); + + auto K = weight.size(0); + //auto C1 = weight.size(1); + auto R = weight.size(2); + auto S = weight.size(3); + + float *input_ptr = input.data_ptr(); + float *weight_ptr = weight.data_ptr(); + float *output_ptr = output.data_ptr(); + float *bias_ptr = bias.data_ptr(); + + simple_conv2d_impl_fp(output_ptr, input_ptr, weight_ptr, bias_ptr, N, C, H, W, + K, R, S, stride, padding, dilation, groups); + + //thnn_conv2d_out(output, input, weight, + return output; +} + +at::Tensor simple_conv2d_bp(torch::Tensor& input, torch::Tensor output, torch::Tensor weight, + int stride, int padding, int dilation, int groups) +{ + RECORD_FUNCTION("simple_conv2d_bp", std::vector({output, weight})); + + auto N = input.size(0); + auto C = input.size(1); + auto H = input.size(2); + auto W = input.size(3); + + auto K = weight.size(0); + //auto C1 = weight.size(1); + auto R = weight.size(2); + auto S = weight.size(3); + + float *input_ptr = input.data_ptr(); + float *weight_ptr = weight.data_ptr(); + float *output_ptr = output.data_ptr(); + + simple_conv2d_impl_bp(input_ptr, output_ptr, weight_ptr, N, C, H, W, + K, R, S, stride, padding, dilation, groups); + + //thnn_conv2d_out(output, input, weight, + return input; +} + +at::Tensor simple_conv2d_wu(torch::Tensor& weight, torch::Tensor output, torch::Tensor input, + int stride, int padding, int dilation, int groups) +{ + RECORD_FUNCTION("simple_conv2d_wu", std::vector({output, input})); + + auto N = input.size(0); + auto C = input.size(1); + auto H = input.size(2); + auto W = input.size(3); + + auto K = weight.size(0); + //auto C1 = weight.size(1); + auto R = weight.size(2); + auto S = weight.size(3); + + float *input_ptr = input.data_ptr(); + float *weight_ptr = weight.data_ptr(); + float *output_ptr = output.data_ptr(); + + simple_conv2d_impl_wu(weight_ptr, output_ptr, input_ptr, N, C, H, W, + K, R, S, stride, padding, dilation, groups); + + //thnn_conv2d_out(output, input, weight, + return weight; +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("conv2d_fp", &simple_conv2d_fp, "simple conv_fp implementation"); + m.def("conv2d_bp", &simple_conv2d_bp, "simple conv_bp implementation"); + m.def("conv2d_wu", &simple_conv2d_wu, "simple conv_wu implementation"); +} diff --git a/FP8_Emulator/cmodel/simple/simple_conv2d_impl.cpp b/FP8_Emulator/cmodel/simple/simple_conv2d_impl.cpp new file mode 100644 index 00000000..b4923da1 --- /dev/null +++ b/FP8_Emulator/cmodel/simple/simple_conv2d_impl.cpp @@ -0,0 +1,977 @@ +/*----------------------------------------------------------------------------* + * Copyright (c) 2023, Intel Corporation - All rights reserved. + * This file is part of FP8-Emulation-Toolkit + * + * SPDX-License-Identifier: BSD-3-Clause + *----------------------------------------------------------------------------* + * Naveen Mellempudi (Intel Corporation) + *----------------------------------------------------------------------------*/ + +#include +#include +#include +#include +#include +#if defined(_OPENMP) + +#include +#endif +#include + +#define CHANNEL_BLOCK 64 + +extern void MMEngine_avx2_ps(int m, int n, int k, float alpha, float *A, int lda, + float *B, int ldb, float beta, float *C, int ldc); +extern void MMEngine_strideB_avx2_ps(int m, int n, int k, float alpha, float *A, int lda, + float *B, int ldb, float beta, float *C, int ldc, int strideB); + +typedef struct { + int nImg; + int ifm; + int nBIfm; + int nbIfm; + int nBOfm; + int nbOfm; + int ifhp; + int ifwp; + int ifh; + int ifw; + int ofhp; + int ofwp; + int ofh; + int ofw; + int pad_h; + int pad_w; + /* additional padding arams for feature maps -- used in WU kernel*/ + int pad_iw; + int pad_ow; + int nbofw; + int kh; + int kw; + int stride_h; + int stride_w; +} gemm_conv_t; + +INLINE void zero_buf(float* buf, long size) { + int i; +#if defined(_OPENMP) +#pragma omp parallel for private(i) +#endif + for (i = 0; i < size; ++i) { + buf[i] = 0.0f; + } +} + +INLINE void copy_buf(float* src, float* dst, long size) { + int i; +#if defined(_OPENMP) +#pragma omp parallel for private(i) +#endif + for (i = 0; i < size; ++i) { + dst[i] = src[i]; + } +} + +INLINE void init_buf(float* buf, long size, int initPos, int initOne) +{ + int i; + zero_buf(buf, size); +#if defined(_OPENMP) +#pragma omp parallel for private(i) +#endif + for (i = 0; i < size; ++i) { + buf[i] = (float)((initOne != 0) ? 1.0 : ((initPos != 0) ? drand48() : (0.05 - drand48()/10.0))); + } +} + +INLINE void set_zeropad_nchw(float* nchw, int N, int C, int H, int W, int pad_h, int pad_w) +{ + DECLARE_VLA(4, float, input, nchw, C, H, W); + int n, h, w, c; + +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(n, h, w, c) +#endif + for ( n = 0; n < N; n++ ) { + for ( c = 0; c < C; c++ ) { + for ( h = 0; h < H; h++ ) { + for ( w = 0; w < W; w++ ) { + if(h < pad_h || h >= H-pad_h || w < pad_w || w >= W-pad_w) + ACCESS_VLA(4, input, n, c, h, w, C, H, W) = 0.0; + } + } + } + } +} + +INLINE +void copy_NCHW_to_GEMM(const float* nchw, float* gemm, const int N, const int H, const int W, const int C, const int cblock) +{ + DECLARE_VLA(5, float, output, gemm, C/cblock, H, W, cblock); + DECLARE_VLA(4, const float, input, nchw, C, H, W); + int n, h, w, c1, c2; + +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(n, h, w, c1, c2) +#endif + for ( n = 0; n < N; n++ ) { + for ( c1 = 0; c1 < C/cblock; c1++ ) { + for ( h = 0; h < H; h++ ) { + for ( w = 0; w < W; w++ ) { + for ( c2 = 0; c2 < cblock; c2++ ) { + ACCESS_VLA(5, output, n, c1, h, w, c2, C/cblock, H, W, cblock) = + ACCESS_VLA(4, input, n, (c1*cblock)+c2, h, w, C, H, W); + } + } + } + } + } +} + +INLINE +void copy_pad_NCHW_to_GEMM(const float* nchw, float* gemm, const int N, const int H, const int W, const int C, const int cblock, + int pad_h, int pad_w, int pad_c) +{ + int HP, WP; + HP = H + 2*pad_h; + WP = W + 2*pad_w; + int CP = C + pad_c; + DECLARE_VLA(5, float, output, gemm, CP/cblock, HP, WP, cblock); + DECLARE_VLA(4, const float, input, nchw, C, H, W); + int n, h, w, c1, c2; + /* if channels are smaller than cblock, use channels as the cblock */ + int lcblock = cblock; + if (C < cblock) lcblock = C; + +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(n, h, w, c1, c2) +#endif + for ( n = 0; n < N; n++ ) { + for ( c1 = 0; c1 < C/lcblock; c1++ ) { + for ( h = 0; h < H; h++ ) { + for ( w = 0; w < W; w++ ) { + for ( c2 = 0; c2 < lcblock; c2++ ) { + ACCESS_VLA(5, output, n, c1, h+pad_h, w+pad_w, c2, C/cblock, HP, WP, cblock) = + ACCESS_VLA(4, input, n, (c1*lcblock)+c2, h, w, C, H, W); + } + } + } + } + } +} + +INLINE +void copy_pad_NCHW_to_GEMM_ex(const float* nchw, float* gemm, const int N, const int H, const int W, const int C, const int cblock, + int pad_h, int pad_w, int pad_wex, int pad_c) +{ + int HP, WP; + HP = H + 2*pad_h; + WP = W + 2*pad_w + pad_wex; + int CP = C + pad_c; + DECLARE_VLA(5, float, output, gemm, CP/cblock, HP, WP, cblock); + DECLARE_VLA(4, const float, input, nchw, C, H, W); + int n, h, w, c1, c2; + /* if channels are smaller than cblock, use channels as the cblock */ + int lcblock = cblock; + if (C < cblock) lcblock = C; + +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(n, h, w, c1, c2) +#endif + for ( n = 0; n < N; n++ ) { + for ( c1 = 0; c1 < C/lcblock; c1++ ) { + for ( h = 0; h < H; h++ ) { + for ( w = 0; w < W; w++ ) { + for ( c2 = 0; c2 < lcblock; c2++ ) { + ACCESS_VLA(5, output, n, c1, h+pad_h, w+pad_w, c2, C/cblock, HP, WP, cblock) = + ACCESS_VLA(4, input, n, (c1*lcblock)+c2, h, w, C, H, W); + } + } + } + } + } +} + + +INLINE +void copy_GEMM_to_NCHW(const float* gemm, float* nchw, const int N, const int H, const int W, const int C, const int cblock) +{ + DECLARE_VLA(5, const float, input, gemm, C/cblock, H, W, cblock); + DECLARE_VLA(4, float, output, nchw, C, H, W); + int n, h, w, c1, c2; +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(n, h, w, c1, c2) +#endif + for ( n = 0; n < N; n++ ) { + for ( c1 = 0; c1 < C/cblock; c1++ ) { + for ( h = 0; h < H; h++ ) { + for ( w = 0; w < W; w++ ) { + for ( c2 = 0; c2 < cblock; c2++ ) { + ACCESS_VLA(4, output, n, (c1*cblock)+c2, h, w, C, H, W) = + ACCESS_VLA(5, input, n, c1, h, w, c2, C/cblock, H, W, cblock); + } + } + } + } + } +} +INLINE +void copy_pad_GEMM_to_NCHW(const float* gemm, float* nchw, const int N, const int H, const int W, const int C, const int cblock, + int pad_h, int pad_w, int pad_c) +{ + int HP, WP; + HP = H + 2*pad_h; + WP = W + 2*pad_w; + int CP = C + pad_c; + + DECLARE_VLA(5, const float, input, gemm, CP/cblock, HP, WP, cblock); + DECLARE_VLA(4, float, output, nchw, C, H, W); + int n, h, w, c1, c2; + + /* if number of channles are smaller than cblock, use C as cblock */ + int lcblock = cblock; + if (cblock > C) lcblock = C; +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(n, h, w, c1, c2) +#endif + for ( n = 0; n < N; n++ ) { + for ( c1 = 0; c1 < C/lcblock; c1++ ) { + for ( h = 0; h < H; h++ ) { + for ( w = 0; w < W; w++ ) { + for ( c2 = 0; c2 < lcblock; c2++ ) { + ACCESS_VLA(4, output, n, (c1*lcblock)+c2, h, w, C, H, W) = + ACCESS_VLA(5, input, n, c1, h+pad_h, w+pad_w, c2, C/cblock, HP, WP, cblock); + } + } + } + } + } +} + +INLINE +void copy_KCRS_to_GEMM(const float* kcrs, float* gemm, const int R, const int S, const int C, const int K, const int cblock, const int kblock) +{ + DECLARE_VLA(6, float, output, gemm, C/cblock, R, S, cblock, kblock); + DECLARE_VLA(4, const float, input, kcrs, C, R, S); + int r, s, c1, c2, k1, k2; +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(r, s, c1, c2, k1, k2) +#endif + for ( k1 = 0; k1 < K/kblock; k1++ ) { + for ( c1 = 0; c1 < C/cblock; c1++ ) { + for ( r = 0; r < R; r++ ) { + for ( s = 0; s < S; s++ ) { + for ( c2 = 0; c2 < cblock; c2++ ) { + for ( k2 = 0; k2 < kblock; k2++ ) { + ACCESS_VLA(6, output, k1, c1, r, s, c2, k2, C/cblock, R, S, cblock, kblock) = + ACCESS_VLA(4, input, (k1*kblock)+k2, (c1*cblock)+c2, r, s, C, R, S); + } + } + } + } + } + } +} + +INLINE +void copy_pad_KCRS_to_GEMM(const float* kcrs, float* gemm, const int R, const int S, const int C, const int K, const int cblock, + const int kblock, int pad_c, int pad_k) +{ + int CP = C + pad_c; + int KP = K + pad_k; + int lcblock = cblock; + int lkblock = kblock; + if (C < cblock) lcblock = C; + if (K < kblock) lkblock = K; + + DECLARE_VLA(6, float, output, gemm, CP/cblock, R, S, cblock, kblock); + DECLARE_VLA(4, const float, input, kcrs, C, R, S); + int r, s, c1, c2, k1, k2; + +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(r, s, c1, c2, k1, k2) shared(CP, KP, lcblock, lkblock) +#endif + for ( k1 = 0; k1 < KP/kblock; k1++ ) { + for ( c1 = 0; c1 < CP/cblock; c1++ ) { + for ( r = 0; r < R; r++ ) { + for ( s = 0; s < S; s++ ) { + for ( c2 = 0; c2 < lcblock; c2++ ) { + for ( k2 = 0; k2 < lkblock; k2++ ) { + ACCESS_VLA(6, output, k1, c1, r, s, c2, k2, CP/cblock, R, S, cblock, kblock) = + ACCESS_VLA(4, input, (k1*lkblock)+k2, (c1*lcblock)+c2, r, s, C, R, S); + } + } + } + } + } + } +} + + +INLINE +void copy_GEMM_to_KCRS(const float* gemm, float* kcrs, const int R, const int S, const int C, const int K, const int cblock, const int kblock) +{ + DECLARE_VLA(6, const float, input, gemm, C/cblock, R, S, cblock, kblock); + DECLARE_VLA(4, float, output, kcrs, C, R, S); + int r, s, c1, c2, k1, k2; + +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(r, s, c1, c2, k1, k2) +#endif + for ( k1 = 0; k1 < K/kblock; k1++ ) { + for ( c1 = 0; c1 < C/cblock; c1++ ) { + for ( r = 0; r < R; r++ ) { + for ( s = 0; s < S; s++ ) { + for ( c2 = 0; c2 < cblock; c2++ ) { + for ( k2 = 0; k2 < kblock; k2++ ) { + ACCESS_VLA(4, output, (k1*kblock)+k2, (c1*cblock)+c2, r, s, C, R, S) = + ACCESS_VLA(6, input, k1, c1, r, s, c2, k2, C/cblock, R, S, cblock, kblock); + } + } + } + } + } + } +} + +INLINE +void copy_pad_GEMM_to_KCRS(const float* gemm, float* kcrs, const int R, const int S, const int C, const int K, const int cblock, + const int kblock, int pad_c, int pad_k) +{ + + int CP = C + pad_c; + int KP = K + pad_k; + int lcblock = cblock; + int lkblock = kblock; + if (C < cblock) lcblock = C; + if (K < kblock) lkblock = K; + + DECLARE_VLA(6, const float, input, gemm, CP/cblock, R, S, cblock, kblock); + DECLARE_VLA(4, float, output, kcrs, C, R, S); + int r, s, c1, c2, k1, k2; + +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(r, s, c1, c2, k1, k2) shared(CP, KP, lcblock, lkblock) +#endif + for ( k1 = 0; k1 < KP/kblock; k1++ ) { + for ( c1 = 0; c1 < CP/cblock; c1++ ) { + for ( r = 0; r < R; r++ ) { + for ( s = 0; s < S; s++ ) { + for ( c2 = 0; c2 < lcblock; c2++ ) { + for ( k2 = 0; k2 < lkblock; k2++ ) { + ACCESS_VLA(4, output, (k1*lkblock)+k2, (c1*lcblock)+c2, r, s, C, R, S) = + ACCESS_VLA(6, input, k1, c1, r, s, c2, k2, CP/cblock, R, S, cblock, kblock); + } + } + } + } + } + } +} + +INLINE +void gemm_kernel_conv_fp(gemm_conv_t* param, float* output, float* input, float* filter, const float* bias) +{ + int nImg = param->nImg; + int nBIfm = param->nBIfm; + int nbIfm = param->nbIfm; + int nBOfm = param->nBOfm; + int nbOfm = param->nbOfm; + int ifhp = param->ifhp; + int ifwp = param->ifwp; + int ofhp = param->ofhp; + int ofwp = param->ofwp; + //int ifh = param->ifh; + //int ifw = param->ifw; + int ofh = param->ofh; + int ofw = param->ofw; + //int nbofw = param->nbofw; + int pad_h = param->pad_h; + int pad_w = param->pad_w; + int kh = param->kh; + int kw = param->kw; + int stride_h = param->stride_h; + int stride_w = param->stride_w; + /* loop counters */ + int img, ofm1, ifm1, oj, oi, ij, kj, ki, ofm2; +#ifdef NAIVE_CODE + int ii, ifm2; +#endif + DECLARE_VLA(5, float, output_t, output + (pad_h * ofwp * nbOfm + pad_w * nbOfm), nBOfm, ofhp, ofwp, nbOfm); + DECLARE_VLA(5, float, input_t, input , nBIfm, ifhp, ifwp, nbIfm); + DECLARE_VLA(6, float, filter_t, filter, nBIfm, kh, kw, nbIfm, nbOfm); + + int lda = CHANNEL_BLOCK; + int ldb = stride_w*CHANNEL_BLOCK; + int ldc = CHANNEL_BLOCK; +#if 0 + /* compute new n-blocking to fit limitation of N<=16 of TMUL */ + for ( oii = 16; oii > 0; --oii ) { + if ( ofw % oii == 0 ) { + nbofw = oii; + break; + } + } +#endif +#if defined(_OPENMP) +#pragma omp parallel for collapse(2) private(img, ofm1, ofm2, oj, oi ) +#endif + /* pre-initializing with bias */ + for (img = 0; img < nImg; ++img) { + for (ofm1 = 0; ofm1 < nBOfm; ++ofm1) { + for (oj = 0; oj < ofh; ++oj) { + for (oi = 0; oi < ofw; ++oi) { + for (ofm2 = 0; ofm2 < nbOfm; ++ofm2) { + ACCESS_VLA(5, output_t, img, ofm1, oj, oi, ofm2, nBOfm, ofhp, ofwp, nbOfm) = bias[ofm1*nbOfm + ofm2]; + } + } + } + } + } + +#if defined(_OPENMP) +#ifdef NAIVE_CODE +#pragma omp parallel for collapse(4) private(img, ofm1, ofm2, ifm1, ifm2, oj, oi, ij, ii, kj, ki) +#else +#pragma omp parallel for collapse(4) private(img, ofm1, ofm2, ifm1, oj, oi, ij, kj, ki) +#endif +#endif + for (img = 0; img < nImg; ++img) { + for (ofm1 = 0; ofm1 < nBOfm; ++ofm1) { + for (ifm1 = 0; ifm1 < nBIfm; ++ifm1) { + for (oj = 0; oj < ofh; ++oj) { + ij = oj * stride_h; + for (kj = 0; kj < kh; ++kj) { + for (ki = 0; ki < kw; ++ki) { + /* let's do a 64 x ofw x 64 GEMM : M=nbOfm, N=ofw, K=nbIfm (col-major) */ +#ifdef NAIVE_CODE + for (oi = 0; oi < ofw; ++oi) { + ii = oi * stride_w; + for (ofm2 = 0; ofm2 < nbOfm; ++ofm2) { + for (ifm2 = 0; ifm2 < nbIfm; ++ifm2) { + ACCESS_VLA(5, output_t, img, ofm1, oj, oi, ofm2, nBOfm, ofhp, ofwp, nbOfm) += /* C */ + ACCESS_VLA(6, filter_t, ofm1, ifm1, kj, ki, ifm2, ofm2, nBIfm, kh, kw, nbIfm, nbOfm) /* A */ + * ACCESS_VLA(5, input_t, img, ifm1, ij + kj, ii + ki, ifm2, nBIfm, ifhp, ifwp, nbIfm); /* B */ + } + } + } +#else + MMEngine_avx2_ps (nbOfm, ofw, nbIfm, 1.0, + &ACCESS_VLA(6, filter_t, ofm1, ifm1, kj, ki, 0, 0, nBIfm, kh, kw, nbIfm, nbOfm), lda, + &ACCESS_VLA(5, input_t, img, ifm1, ij + kj, ki, 0, nBIfm, ifhp, ifwp, nbIfm), ldb, 1.0, + &ACCESS_VLA(5, output_t, img, ofm1, oj, 0, 0, nBOfm, ofhp, ofwp, nbOfm), ldc); +#endif + } + } + } + } + } + } +} + +INLINE void gemm_kernel_conv_bp(gemm_conv_t* param, float* input, float* output, float* filter, float* tr_filter) +{ + int nImg = param->nImg; + int nBIfm = param->nBIfm; + int nbIfm = param->nbIfm; + int nBOfm = param->nBOfm; + int nbOfm = param->nbOfm; + int ifhp = param->ifhp; + int ifwp = param->ifwp; + int ofhp = param->ofhp; + int ofwp = param->ofwp; + //int ifh = param->ifh; + //int ifw = param->ifw; + int ofh = param->ofh; + int ofw = param->ofw; + //int nbofw = param->nbofw; + int pad_h = param->pad_h; + int pad_w = param->pad_w; + int kh = param->kh; + int kw = param->kw; + int stride_h = param->stride_h; + int stride_w = param->stride_w; + + /* loop counters */ + int img, ofm1, ifm1, ofm2, ifm2, oj, ij, kj, ki;/*, oii;*/ +#ifdef NAIVE_CODE + int ii, oi; +#endif + DECLARE_VLA(5, float, output_t, output + (pad_h * ofwp * nbOfm + pad_w * nbOfm), nBOfm, ofhp, ofwp, nbOfm); + DECLARE_VLA(5, float, input_t, input , nBIfm, ifhp, ifwp, nbIfm); + DECLARE_VLA(6, float, filter_t, filter, nBIfm, kh, kw, nbIfm, nbOfm); + DECLARE_VLA(6, float, tr_filter_t, tr_filter, nBIfm, kh, kw, nbOfm, nbIfm); + +#if defined(_OPENMP) +# pragma omp parallel for collapse(4) private(ofm1, ifm1, kj, ki, ofm2, ifm2) +#endif + for (ofm1 = 0; ofm1 < nBOfm; ++ofm1) { + for (ifm1 = 0; ifm1 < nBIfm; ++ifm1) { + for (kj = 0; kj < kh; ++kj) { + for (ki = 0; ki < kw; ++ki) { + for (ofm2 = 0; ofm2 < nbOfm; ++ofm2) { + for (ifm2 = 0; ifm2 < nbIfm; ++ifm2) { + ACCESS_VLA(6, tr_filter_t, ofm1, ifm1, kj, ki, ofm2, ifm2, nBIfm, kh, kw, nbOfm, nbIfm) = + ACCESS_VLA( 6, filter_t, ofm1, ifm1, kj, ki, ifm2, ofm2, nBIfm, kh, kw, nbIfm, nbOfm); + } + } + } + } + } + } + + int lda = CHANNEL_BLOCK; + int ldb = CHANNEL_BLOCK; + int ldc = stride_w*CHANNEL_BLOCK; +#if 0 + /* compute new n-blocking */ + for ( oii = 16; oii > 0; --oii ) { + if ( ofw % oii == 0 ) { + nbofw = oii; + break; + } + } +#endif +#if defined(_OPENMP) +#ifdef NAIVE_CODE +# pragma omp parallel for collapse(4) private(img, ofm1, ifm1, ofm2, ifm2, oj, oi, ij, ii, kj, ki) +#else +# pragma omp parallel for collapse(4) private(img, ofm1, ifm1, ofm2, ifm2, oj, ij, kj, ki) +#endif +#endif + for (img = 0; img < nImg; ++img) { + for (ifm1 = 0; ifm1 < nBIfm; ++ifm1) { + for (ofm1 = 0; ofm1 < nBOfm; ++ofm1) { + for (oj = 0; oj < ofh; ++oj) { + ij = oj * stride_h; + for (kj = 0; kj < kh; ++kj) { + for (ki = 0; ki < kw; ++ki) { +#ifdef NAIVE_CODE + for (ifm2 = 0; ifm2 < nbIfm; ++ifm2) { + for (oi = 0; oi < ofw; ++oi) { + ii = oi * stride_w; + for (ofm2 = 0; ofm2 < nbOfm; ++ofm2) { + ACCESS_VLA(5, input_t, img, ifm1, ij + kj, ii + ki, ifm2, nBIfm, ifhp, ifwp, nbIfm) += /* C */ + ACCESS_VLA(6, tr_filter_t, ofm1, ifm1, kj, ki, ofm2, ifm2, nBIfm, kh, kw, nbOfm, nbIfm) /* A */ + * ACCESS_VLA(5, output_t, img, ofm1, oj, oi, ofm2, nBOfm, ofhp, ofwp, nbOfm); /* B */ + } + } + } +#else + MMEngine_avx2_ps ( nbIfm, ofw, nbOfm, 1.0, + &ACCESS_VLA(6, tr_filter_t, ofm1, ifm1, kj, ki, 0, 0, nBIfm, kh, kw, nbOfm, nbIfm), lda, + &ACCESS_VLA(5, output_t, img, ofm1, oj, 0, 0, nBOfm, ofhp, ofwp, nbOfm), ldb, 1.0, + &ACCESS_VLA(5, input_t, img, ifm1, ij + kj, ki, 0, nBIfm, ifhp, ifwp, nbIfm), ldc); +#endif + } + } + } + } + } + } +} + +INLINE void gemm_kernel_conv_wu(gemm_conv_t* param, float* filter, float* output, float* input, float* tr_input) +{ + int nImg = param->nImg; + int nBIfm = param->nBIfm; + int nbIfm = param->nbIfm; + int nBOfm = param->nBOfm; + int nbOfm = param->nbOfm; + int ifhp = param->ifhp; + int ifwp = param->ifwp; + int ofhp = param->ofhp; + int ofwp = param->ofwp; + int ifh = param->ifh; + int ifw = param->ifw; + int ofh = param->ofh; + int ofw = param->ofw; + //int nbofw = param->nbofw; + int pad_h = param->pad_h; + int pad_w = param->pad_w; + //int pad_iw = param->pad_iw; + int pad_ow = param->pad_ow; + int kh = param->kh; + int kw = param->kw; + int stride_h = param->stride_h; + int stride_w = param->stride_w; + + /* loop counters */ + int img, ofm1, ifm1, ifm2, oj, ij, ii, kj, ki;/*, oii;*/ +#ifdef NAIVE_CODE + int ofm2, oi; +#endif + + DECLARE_VLA(5, float, output_t, output + (pad_h * ofwp * nbOfm + pad_w * nbOfm), nBOfm, ofhp, ofwp, nbOfm); + DECLARE_VLA(5, float, input_t, (float*)input, nBIfm, ifhp, ifwp, nbIfm); + DECLARE_VLA(5, float, tr_input_t, tr_input, nBIfm, ifhp+(2*pad_h), nbIfm, ifwp+(2*pad_w)); + DECLARE_VLA(6, float, filter_t, filter, nBIfm, kh, kw, nbIfm, nbOfm); + +#if defined(_OPENMP) +# pragma omp parallel for collapse(2) private(img, ifm1, ij, ii, ifm2) +#endif + for (img = 0; img < nImg; ++img) { + for (ifm1 = 0; ifm1 < nBIfm; ++ifm1) { + for (ij = 0; ij < ifh; ++ij) { + for (ifm2 = 0; ifm2 < nbIfm; ++ifm2) { + for (ii = 0; ii < ifw; ++ii) { + ACCESS_VLA(5, tr_input_t, img, ifm1, ij+pad_h, ifm2, ii+pad_w, nBIfm, ifhp+(2*pad_h), nbIfm, ifwp+(2*pad_w)) = + ACCESS_VLA(5, input_t, img, ifm1, ij+pad_h, ii+pad_w, ifm2, nBIfm, ifhp, ifwp, nbIfm); + } + } + } + } + } + + int lda = CHANNEL_BLOCK; + int ldb = ifwp + (2*pad_w); + int ldc = CHANNEL_BLOCK; +#if 0 + int nbifm = nbIfm; + for ( oii = 16; oii > 0; --oii ) { + if ( nbIfm % oii == 0 ) { + nbifm = oii; + break; + } + } +#endif +#if defined(_OPENMP) +#ifdef NAIVE_CODE +# pragma omp parallel for collapse(2) private(img, ofm1, ifm1, ofm2, ifm2, oj, oi, ij, ii, kj, ki) +#else +# pragma omp parallel for collapse(2) private(img, ofm1, ifm1, ifm2, oj, ij, ii, kj, ki) +#endif +#endif + for (ofm1 = 0; ofm1 < nBOfm; ++ofm1) { + for (ifm1 = 0; ifm1 < nBIfm; ++ifm1) { + for (img = 0; img < nImg; ++img) { + for (oj = 0; oj < ofh; ++oj) { + ij = oj * stride_h; + for (kj = 0; kj < kh; ++kj) { + for (ki = 0; ki < kw; ++ki) { +#ifdef NAIVE_CODE + for (ifm2 = 0; ifm2 < nbIfm; ++ifm2) { + for (oi = 0; oi < ofw; ++oi) { + ii = oi * stride_w; + for (ofm2 = 0; ofm2 < nbOfm; ++ofm2) { + ACCESS_VLA( 6, filter_t, ofm1, ifm1, kj, ki, ifm2, ofm2, nBIfm, kh, kw, nbIfm, nbOfm) /* C */ += + ACCESS_VLA(5, output_t, img, ofm1, oj, oi, ofm2, nBOfm, ofhp, ofwp, nbOfm) /* A */ + * ACCESS_VLA(5, tr_input_t, img, ifm1, ij + kj, ifm2, ii + ki, nBIfm, ifhp+(2*pad_h), nbIfm, ifwp+(2*pad_w)) /* B */; + } + } + } +#else + MMEngine_strideB_avx2_ps (nbOfm, nbIfm, ofw+pad_ow, 1.0, + &ACCESS_VLA(5, output_t, img, ofm1, oj, 0, 0, nBOfm, ofhp, ofwp, nbOfm), lda, + &ACCESS_VLA(5, tr_input_t, img, ifm1, ij + kj, 0, ki, nBIfm, ifhp+(2*pad_h), nbIfm, ifwp+(2*pad_w)), ldb, 1.0, + &ACCESS_VLA(6, filter_t, ofm1, ifm1, kj, ki, 0, 0, nBIfm, kh, kw, nbIfm, nbOfm), ldc, stride_w); +#endif + } + } + } + } + } + } +} + +int simple_conv2d_impl_fp(float* outputs, float *inputs, float *weights, float* bias, int N, int C, int iH, int iW, + int K, int R, int S, int stride, int padding, int dilation, int groups) +{ + float *gemm_input, *gemm_output, *gemm_filter; + int ifhp, ifwp, ofhp, ofwp, ofh, ofw; + int stride_h, stride_w, pad_h, pad_w; + + int ifw = iW; /* input width, "W" */ + int ifh = iH; /* input height, "H" */ + int nImg = N; /* mini-batch size, "N" */ + int nIfm = C; /* number of input feature maps, "C" */ + int nOfm = K; /* number of output feature maps, "K" */ + int kh = R; /* filter height, "R" */ + int kw = S; /* filter width, "S" */ + pad_h = padding; /* padding in output */ + pad_w = padding; /* padding in output */ + + /* apply stride in both dimensions */ + stride_w = stride; + stride_h = stride; + + /* deriving some values image size */ + ofh = (ifh + 2 * pad_h - kh) / stride_h + 1; + ofw = (ifw + 2 * pad_w - kw) / stride_w + 1; + //ofh = (ifh + 2 * pad_h - dilation * (kh-1) - 1) / stride_h + 1; + //ofw = (ifw + 2 * pad_w - dilation * (kw-1) - 1) / stride_w + 1; + + ifhp = ifh + 2 * pad_h; + ifwp = ifw + 2 * pad_w; + ofhp = ofh + 2 * pad_h; + ofwp = ofw + 2 * pad_w; + + /*pad nIfm and nOfm to multiples of 64 */ + int ifm_pad = 0; + int ofm_pad = 0; + int nIfmp = nIfm; + int nOfmp = nOfm; + if (nIfm % 64 != 0){ + ifm_pad = (64-(nIfm%64)); + nIfmp += ifm_pad; + } + if (nOfm % 64 != 0) { + ofm_pad = (64-(nOfm%64)); + nOfmp += ofm_pad; + } + + gemm_conv_t gemm_param; + /* set struct for naive convolution */ + gemm_param.nImg = nImg; + gemm_param.nBIfm = nIfmp/CHANNEL_BLOCK; + gemm_param.nbIfm = CHANNEL_BLOCK; + gemm_param.nBOfm = nOfmp/CHANNEL_BLOCK; + gemm_param.nbOfm = CHANNEL_BLOCK; + gemm_param.ifhp = ifhp; + gemm_param.ifwp = ifwp; + gemm_param.ofhp = ofhp; + gemm_param.ofwp = ofwp; + gemm_param.ifh = ifh; + gemm_param.ifw = ifw; + gemm_param.ofh = ofh; + gemm_param.ofw = ofw; + gemm_param.pad_h = pad_h; + gemm_param.pad_w = pad_w; + + if ( ofw == 56 ) { + gemm_param.nbofw = 28; + } else { + gemm_param.nbofw = ofw; + } + gemm_param.kh = kh; + gemm_param.kw = kw; + gemm_param.stride_h = stride_h; + gemm_param.stride_w = stride_w; + + gemm_input = (float*)_mm_malloc( nImg*nIfmp*ifhp*ifwp*sizeof(float), 2097152); + gemm_filter = (float*)_mm_malloc( nOfmp*nIfmp*kh*kw* sizeof(float), 2097152); + gemm_output = (float*)_mm_malloc( nImg*nOfmp*ofhp*ofwp*sizeof(float), 2097152); + zero_buf(gemm_input, nImg*nIfmp*ifhp*ifwp); + if (nIfm % 64 != 0 || nOfm % 64 != 0){ + zero_buf(gemm_filter, nOfmp*nIfmp*kh*kw); + zero_buf(gemm_output, nImg*nOfmp*ofhp*ofwp); + } + + /* copy data into GEMM optimized format */ + copy_pad_NCHW_to_GEMM(inputs, gemm_input, nImg, ifh, ifw, nIfm, CHANNEL_BLOCK, pad_h, pad_w, ifm_pad); + copy_pad_KCRS_to_GEMM(weights, gemm_filter, kh, kw, nIfm, nOfm, CHANNEL_BLOCK, CHANNEL_BLOCK, ifm_pad, ofm_pad); + gemm_kernel_conv_fp(&gemm_param, gemm_output, gemm_input, gemm_filter, bias); + /* copy out data */ + copy_pad_GEMM_to_NCHW(gemm_output, outputs, nImg, ofh, ofw, nOfm, CHANNEL_BLOCK, pad_h, pad_w, ofm_pad); + + _mm_free(gemm_input); + _mm_free(gemm_filter); + _mm_free(gemm_output); + return 0; +} + +int simple_conv2d_impl_bp(float* inputs, float *outputs, float *weights, int N, int C, int iH, int iW, + int K, int R, int S, int stride, int padding, int dilation, int groups) +{ + float *gemm_input, *gemm_output, *gemm_filter, *gemm_filter_tr; + int ifhp, ifwp, ofhp, ofwp, ofh, ofw; + int stride_h, stride_w, pad_h, pad_w; + + int ifw = iW; /* input width, "W" */ + int ifh = iH; /* input height, "H" */ + int nImg = N; /* mini-batch size, "N" */ + int nIfm = C; /* number of input feature maps, "C" */ + int nOfm = K; /* number of output feature maps, "K" */ + int kh = R; /* filter height, "R" */ + int kw = S; /* filter width, "S" */ + pad_h = padding; /* padding in output */ + pad_w = padding; /* padding in output */ + + /* apply stride in both dimensions */ + stride_w = stride; + stride_h = stride; + /* deriving some values image size */ + ofh = (ifh + 2 * pad_h - kh) / stride_h + 1; + ofw = (ifw + 2 * pad_w - kw) / stride_w + 1; + //ofh = (ifh + 2 * pad_h - dilation * (kh-1) - 1) / stride_h + 1; + //ofw = (ifw + 2 * pad_w - dilation * (kw-1) - 1) / stride_w + 1; + ifhp = ifh + 2 * pad_h; + ifwp = ifw + 2 * pad_w; + ofhp = ofh + 2 * pad_h; + ofwp = ofw + 2 * pad_w; + /*pad nIfm and nOfm to multiples of 64 */ + int ifm_pad = 0; + int ofm_pad = 0; + int nIfmp = nIfm; + int nOfmp = nOfm; + if (nIfm % 64 != 0){ + ifm_pad = (64-(nIfm%64)); + nIfmp += ifm_pad; + } + if (nOfm % 64 != 0) { + ofm_pad = (64-(nOfm%64)); + nOfmp += ofm_pad; + } + + gemm_conv_t gemm_param; + /* set struct for naive convolution */ + gemm_param.nImg = nImg; + gemm_param.nBIfm = nIfmp/CHANNEL_BLOCK; + gemm_param.nbIfm = CHANNEL_BLOCK; + gemm_param.nBOfm = nOfmp/CHANNEL_BLOCK; + gemm_param.nbOfm = CHANNEL_BLOCK; + gemm_param.ifhp = ifhp; + gemm_param.ifwp = ifwp; + gemm_param.ofhp = ofhp; + gemm_param.ofwp = ofwp; + gemm_param.ifh = ifh; + gemm_param.ifw = ifw; + gemm_param.ofh = ofh; + gemm_param.ofw = ofw; + gemm_param.pad_h = pad_h; + gemm_param.pad_w = pad_w; + + if ( ofw == 56 ) { + gemm_param.nbofw = 28; + } else { + gemm_param.nbofw = ofw; + } + + gemm_param.kh = kh; + gemm_param.kw = kw; + gemm_param.stride_h = stride_h; + gemm_param.stride_w = stride_w; + gemm_input = (float*)_mm_malloc( nImg*nIfmp*ifhp*ifwp*sizeof(float), 2097152); + gemm_filter = (float*)_mm_malloc( nOfmp*nIfmp*kh*kw* sizeof(float), 2097152); + gemm_filter_tr = (float*)_mm_malloc( nOfmp*nIfmp*kh*kw* sizeof(float), 2097152); + gemm_output = (float*)_mm_malloc( nImg*nOfmp*ofhp*ofwp*sizeof(float), 2097152); + + zero_buf(gemm_input, nImg*nIfmp*ifhp*ifwp); + if (nIfm % 64 != 0 || nOfm % 64 != 0){ + zero_buf(gemm_filter, nOfmp*nIfmp*kh*kw); + zero_buf(gemm_output, nImg*nOfmp*ofhp*ofwp); + } + + /* copy data into GEMM optimized format */ + copy_pad_NCHW_to_GEMM(outputs, gemm_output, nImg, ofh, ofw, nOfm, CHANNEL_BLOCK, pad_h, pad_w, ofm_pad); + copy_pad_KCRS_to_GEMM(weights, gemm_filter, kh, kw, nIfm, nOfm, CHANNEL_BLOCK, CHANNEL_BLOCK, ifm_pad, ofm_pad); + gemm_kernel_conv_bp(&gemm_param, gemm_input, gemm_output, gemm_filter, gemm_filter_tr); + /* copy out data */ + copy_pad_GEMM_to_NCHW(gemm_input, inputs, nImg, ifh, ifw, nIfm, CHANNEL_BLOCK, pad_h, pad_w, ifm_pad); + + _mm_free(gemm_input); + _mm_free(gemm_filter); + _mm_free(gemm_filter_tr); + _mm_free(gemm_output); + return 0; +} + +int simple_conv2d_impl_wu(float *weights, float *outputs, float *inputs, int N, int C, int iH, int iW, + int K, int R, int S, int stride, int padding, int dilation, int groups) +{ + float *gemm_input, *gemm_input_tr, *gemm_output, *gemm_filter; + int ifhp, ifwp, ofhp, ofwp, ofh, ofw; + int stride_h, stride_w, pad_h, pad_w; + + int ifw = iW; /* input width, "W" */ + int ifh = iH; /* input height, "H" */ + int nImg = N; /* mini-batch size, "N" */ + int nIfm = C; /* number of input feature maps, "C" */ + int nOfm = K; /* number of output feature maps, "K" */ + int kh = R; /* filter height, "R" */ + int kw = S; /* filter width, "S" */ + pad_h = padding; /* padding in output */ + pad_w = padding; /* padding in output */ + + /* apply stride in both dimensions */ + stride_w = stride; + stride_h = stride; + + /* deriving some values image size */ + ofw = (ifw + 2 * pad_w - kw) / stride_w + 1; + ofh = (ifh + 2 * pad_h - kh) / stride_h + 1; + //ofw = (ifw + 2 * pad_w - dilation * (kw-1) - 1) / stride_w + 1; + //ofh = (ifh + 2 * pad_h - dilation * (kh-1) - 1) / stride_h + 1; + + /* padding feature map width to be a multiple of 4 to perform VNN4 operations */ + int pad_iw = 0; + int pad_ow = 0; + + if (ofw%4 != 0){ + pad_ow = 4-(ofw%4); + } + if (ifw%4 != 0){ + pad_iw = 4-(ifw%4); + } + pad_iw = stride_w * pad_ow; + + /*pad ofw and ifw to to be multiples of 4 (VNNI4) */ + ifhp = ifh + 2 * pad_h; + ifwp = ifw + 2 * pad_w + pad_iw; + ofhp = ofh + 2 * pad_h; + ofwp = ofw + 2 * pad_w + pad_ow; + + /*pad nIfm and nOfm to multiples of 64 */ + int ifm_pad = 0; + int ofm_pad = 0; + int nIfmp = nIfm; + int nOfmp = nOfm; + if (nIfm % 64 != 0){ + ifm_pad = (64-(nIfm%64)); + nIfmp += ifm_pad; + } + if (nOfm % 64 != 0) { + ofm_pad = (64-(nOfm%64)); + nOfmp += ofm_pad; + } + + gemm_conv_t gemm_param; + /* set struct for naive convolution */ + gemm_param.nImg = nImg; + gemm_param.nBIfm = nIfmp/CHANNEL_BLOCK; + gemm_param.nbIfm = CHANNEL_BLOCK; + gemm_param.nBOfm = nOfmp/CHANNEL_BLOCK; + gemm_param.nbOfm = CHANNEL_BLOCK; + gemm_param.ifhp = ifhp; + gemm_param.ifwp = ifwp; + gemm_param.ofhp = ofhp; + gemm_param.ofwp = ofwp; + gemm_param.ifh = ifh; + gemm_param.ifw = ifw; + gemm_param.ofh = ofh; + gemm_param.ofw = ofw; + gemm_param.pad_h = pad_h; + gemm_param.pad_w = pad_w; + gemm_param.pad_iw = pad_iw; + gemm_param.pad_ow = pad_ow; + + if ( ofw == 56 ) { + gemm_param.nbofw = 28; + } else { + gemm_param.nbofw = ofw; + } + + gemm_param.kh = kh; + gemm_param.kw = kw; + gemm_param.stride_h = stride_h; + gemm_param.stride_w = stride_w; + gemm_input = (float*)_mm_malloc( nImg*nIfmp*ifhp*ifwp*sizeof(float), 2097152); + gemm_input_tr = (float*)_mm_malloc( nImg*nIfmp*(ifhp+(2*pad_h))*(ifwp+(2*pad_w))*sizeof(float), 2097152); + gemm_filter = (float*)_mm_malloc( nOfmp*nIfmp*kh*kw* sizeof(float), 2097152); + gemm_output = (float*)_mm_malloc( nImg*nOfmp*ofhp*ofwp*sizeof(float), 2097152); + zero_buf(gemm_input, nImg*nIfmp*ifhp*ifwp); + zero_buf(gemm_input_tr, nImg*(ifhp+(2*pad_h))*nIfmp*(ifwp+(2*pad_w))); + zero_buf(gemm_output, nImg*nOfmp*ofhp*ofwp); + zero_buf(gemm_filter, nOfmp*nIfmp*kh*kw); + /* copy data into GEMM optimized format */ + /* compensate for the VNNI4 padding */ + copy_pad_NCHW_to_GEMM_ex(inputs, gemm_input, nImg, ifh, ifw, nIfm, CHANNEL_BLOCK, pad_h, pad_w, pad_iw, ifm_pad); + copy_pad_NCHW_to_GEMM_ex(outputs, gemm_output, nImg, ofh, ofw, nOfm, CHANNEL_BLOCK, pad_h, pad_w, pad_ow , ofm_pad); + gemm_kernel_conv_wu(&gemm_param, gemm_filter, gemm_output, gemm_input, gemm_input_tr); + copy_pad_GEMM_to_KCRS(gemm_filter, weights, kh, kw, nIfm, nOfm, CHANNEL_BLOCK, CHANNEL_BLOCK, ifm_pad, ofm_pad); + _mm_free(gemm_input); + _mm_free(gemm_input_tr); + _mm_free(gemm_filter); + _mm_free(gemm_output); + return 0; +} diff --git a/FP8_Emulator/cmodel/simple/simple_gemm.cpp b/FP8_Emulator/cmodel/simple/simple_gemm.cpp new file mode 100644 index 00000000..86e277e8 --- /dev/null +++ b/FP8_Emulator/cmodel/simple/simple_gemm.cpp @@ -0,0 +1,71 @@ +/*----------------------------------------------------------------------------* + * Copyright (c) 2023, Intel Corporation - All rights reserved. + * This file is part of FP8-Emulation-Toolkit + * + * SPDX-License-Identifier: BSD-3-Clause + *----------------------------------------------------------------------------* + * Naveen Mellempudi (Intel Corporation) + *----------------------------------------------------------------------------*/ + +#include +#include +#include +#include +#include +#include +#include +#if 10 +//extern "C" { +extern int simple_sgemm_impl( char* transa, char* transb, int m, int n, int k, + float alpha, float* a, int lda, float* b, int ldb, + float beta, float* c, int ldc ); +//} +#endif +#define gettid() ((int)syscall(SYS_gettid)) + +using namespace torch::autograd::profiler; +using namespace torch; +using namespace torch::autograd; +using at::Tensor; + +double get_time() { + static bool init_done = false; + static struct timespec stp = {0,0}; + struct timespec tp; + clock_gettime(CLOCK_REALTIME, &tp); + + if(!init_done) { + init_done = true; + stp = tp; + } + double ret = (tp.tv_sec - stp.tv_sec) * 1e3 + (tp.tv_nsec - stp.tv_nsec)*1e-6; + return ret; +} + +at::Tensor simple_gemm(torch::Tensor& C, torch::Tensor A, torch::Tensor B, float alpha, bool a_trans, bool b_trans) +{ + RECORD_FUNCTION("simple_gemm", std::vector({A, B, alpha})); + + const char *aT = a_trans ? "T" : "N"; + const char *bT = b_trans ? "T" : "N"; + + auto M = C.size(0); + auto N = C.size(1); + auto K = a_trans ? A.size(0) : A.size(1); + auto lda = A.size(1); + auto ldb = B.size(1); + auto ldc = C.size(1); + + float beta = 0.0; + + float *Aptr = A.data_ptr(); + float *Bptr = B.data_ptr(); + float *Cptr = C.data_ptr(); + + simple_sgemm_impl((char*)bT, (char*)aT, N, M, K, alpha, Bptr, ldb, Aptr, lda, beta, Cptr, ldc); + return C; +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("gemm", &simple_gemm, "TMUL GEMM Implementation"); +} diff --git a/FP8_Emulator/cmodel/simple/simple_gemm_impl.cpp b/FP8_Emulator/cmodel/simple/simple_gemm_impl.cpp new file mode 100644 index 00000000..39f96f67 --- /dev/null +++ b/FP8_Emulator/cmodel/simple/simple_gemm_impl.cpp @@ -0,0 +1,233 @@ +/*----------------------------------------------------------------------------* + * Copyright (c) 2023, Intel Corporation - All rights reserved. + * This file is part of FP8-Emulation-Toolkit + * + * SPDX-License-Identifier: BSD-3-Clause + *----------------------------------------------------------------------------* + * Naveen Mellempudi (Intel Corporation) + *----------------------------------------------------------------------------*/ + +#include +#include +#include +#include +#include +#include +#include + +#define SCRATCH_SIZE 4294967296 +int lazy_init = 1; +float* myscratch = NULL; + +extern void MMEngine_avx2_ps(int m, int n, int k, float alpha, float *A, int lda, + float *B, int ldb, float beta, float *C, int ldc); + +__extern_always_inline +void copy_matrix_and_pad(const float* in, float* out, const int LD, const int OD, int ld_pad, int od_pad) +{ + int LDP, ODP; + LDP = LD + ld_pad; + ODP = OD + od_pad; + int ld, od; + int lp, op; + + for ( op = 0; op < OD; op++ ) { + for ( lp = LD-1; lp < LDP; lp++ ) { + out[op*LDP+lp] = 0.0; + } + } + for ( op = OD-1; op < ODP; op++ ) { + for ( lp = 0; lp < LDP; lp++ ) { + out[op*LDP+lp] = 0.0; + } + } + +#if defined(_OPENMP) +#pragma omp parallel for private(ld, od) +#endif + for ( od = 0; od < OD; od++ ) { + for ( ld = 0; ld < LD; ld++ ) { + out[od*LDP+ld] = in[od*LD+ld]; + } + } +} + +__extern_always_inline +void copy_matrix_and_strip_pading(const float* in, float* out, const int LD, const int OD, int ld_pad, int od_pad) +{ + int LDP; + LDP = LD + ld_pad; + int ld, od; +#if defined(_OPENMP) +#pragma omp parallel for private(ld, od) +#endif + for ( od = 0; od < OD; od++ ) { + for ( ld = 0; ld < LD; ld++ ) { + out[od*LD+ld] = in[od*LDP+ld]; + } + } +} + +int simple_sgemm_impl( char* transa, char* transb, int m, int n, int k, + float alpha, float* a, int lda, float* b, int ldb, + float beta, float* c, int ldc ) { + float* myA = NULL; + float* myB = NULL; + float* myC = NULL; + size_t ptlda = (size_t)(lda); + size_t ptldb = (size_t)(ldb); + size_t mylda = (size_t)(lda); + size_t myldb = (size_t)(ldb); + size_t myldc = (size_t)(ldc); + int mym = (size_t)(m); + int myn = (size_t)(n); + int myk = (size_t)(k); + int m_pad = 0; + int n_pad = 0; + int k_pad = 0; + + int o,p,q,pp,oo; + + /* check for size matching our TMUL emulation */ + if ( mym % 16 != 0 ) { + m_pad = (16-(mym%16)); + mym += m_pad; + } + if ( myk % 64 != 0 ) { + k_pad = (64-(myk%64)); + myk += k_pad; + } + if ( myn % 16 != 0 ) { + n_pad = (16-(myn%16)); + myn += n_pad; + } + /* update leading dimensions with padded values */ + mylda = mym; + myldb = myk; + myldc = mym; + + /* lazy init of fp8_gemm state */ + if (lazy_init != 0) { + lazy_init = 0; + myscratch = (float*) _mm_malloc( SCRATCH_SIZE*sizeof(float), 4096 ); + } + /* check for sufficient scratch size */ + if ( (*transa == 'N') && (*transb == 'N') ) { + if ( ((ptlda*myk)+(ptldb*myn)) > SCRATCH_SIZE ) { + return -1; + } + } else if ( (*transa == 'T') && (*transb == 'N') ) { + if ( ((ptlda*mym)+(ptldb*myn)) > SCRATCH_SIZE ) { + return -2; + } + mylda = mym; + } else if ( (*transa == 'N') && (*transb == 'T') ) { + if ( ((ptlda*myk)+(ptldb*myk)) > SCRATCH_SIZE ) { + return -3; + } + myldb = myk; + } else if ( (*transa == 'T') && (*transb == 'T') ) { + if ( ((ptlda*mym)+(ptldb*myk)) > SCRATCH_SIZE ) { + return -4; + } + mylda = mym; + myldb = myk; + } else { + assert((0 && "Error : Invalid parameters")); + return -5; + } + + /* set temp A and B pointers */ + myA = myscratch; + myB = myscratch + (mylda*myk); + myC = myscratch + (mylda*myk) + (myldb*myn); + + if ( *transa == 'T' ) { + /* fill the padding with zeros */ + for ( p = 0; p < k; p++ ) { + for ( o = m-1; o < mym; o++ ) { + myA[p*mylda+o] = 0.0; + } + } + for ( p = k-1; p < myk; p++ ) { + for ( o = 0; o < mym; o++ ) { + myA[p*mylda+o] = 0.0; + } + } + + /* let's transpose data first */ +#if defined(_OPENMP) +#pragma omp parallel for private(o,p) collapse(2) +#endif + for ( p = 0; p < k; p++ ) { + for ( o = 0; o < m; o++ ) { + myA[(p*mylda)+o] = a[(o*ptlda)+p]; + } + } + } else if ( m_pad > 0 || k_pad > 0 ) { + copy_matrix_and_pad(a, myA, m, k, m_pad, k_pad); + } else { + myA = a; + } + + if ( *transb == 'T' ) { + /* fill the padding with zeros */ + for ( p = 0; p < n; p++ ) { + for ( o = k-1; o < myk; o++ ) { + myB[p*myldb+o] = 0.0; + } + } + for ( p = n-1; p < myn; p++ ) { + for ( o = 0; o < myk; o++ ) { + myB[p*myldb+o] = 0.0; + } + } + + /* let's transpose data first */ +#if defined(_OPENMP) +#pragma omp parallel for private(o,p) collapse(2) +#endif + for ( p = 0; p < n; p++ ) { + for ( o = 0; o < k; o++ ) { + myB[(p*myldb)+o] = b[(o*ptldb)+p]; + } + } + } else if ( k_pad > 0 || n_pad > 0 ) { + copy_matrix_and_pad(b, myB, k, n, k_pad, n_pad); + } else { + myB = b; + } + + if ( m_pad > 0 || n_pad > 0 ) { + copy_matrix_and_pad(c, myC, m, n, m_pad, n_pad); + } else { + myC = c; + } + /* run gemm */ +#if defined(_OPENMP) +#pragma omp parallel for private(o,p,q,pp,oo) collapse(2) +#endif + for ( o = 0; o < mym; o += 16 ) { + for ( p = 0; p < myn; p += 16 ) { + float ctmp[256]; + for ( pp = 0; pp < 16; pp++ ) { + for ( oo = 0; oo < 16; oo++ ) { + ctmp[(pp*16)+oo] = 0.0f; + } + } + for ( q = 0; q < myk; q += 64 ) { + MMEngine_avx2_ps(mym, myn, myk, alpha, &(myA[(mylda*q)+o]), mylda, + &(myB[(myldb*p)+q]), myldb, beta, ctmp, myldc); + } + for ( pp = 0; pp < 16; pp++ ) { + for ( oo = 0; oo < 16; oo++ ) { + myC[((p+pp)*myldc)+(o+oo)] += ((alpha)*ctmp[(pp*16)+oo]) + ((beta)*myC[((p+pp)*myldc)+(o+oo)]); + } + } + } + } + if ( m_pad > 0 || n_pad > 0 ) { + copy_matrix_and_strip_pading(myC, c, m, n, m_pad, n_pad); + } + return 0; +} diff --git a/FP8_Emulator/cmodel/simple/simple_mm_engine.cpp b/FP8_Emulator/cmodel/simple/simple_mm_engine.cpp new file mode 100644 index 00000000..3ba40fb6 --- /dev/null +++ b/FP8_Emulator/cmodel/simple/simple_mm_engine.cpp @@ -0,0 +1,45 @@ +/*----------------------------------------------------------------------------* + * Copyright (c) 2023, Intel Corporation - All rights reserved. + * This file is part of FP8-Emulation-Toolkit + * + * SPDX-License-Identifier: BSD-3-Clause + *----------------------------------------------------------------------------* + * Naveen Mellempudi (Intel Corporation) + *----------------------------------------------------------------------------*/ + +#include + +/* column major dat format */ +void MMEngine_avx2_ps(int m, int n, int k, float alpha, float *A, int lda, + float *B, int ldb, float beta, float *C, int ldc) +{ + for (int j = 0; j < n; j++) { + for (int i = 0; i < m; i+=8) { + __m256 cij = _mm256_loadu_ps(&C[ j*ldc + i]); + for (int l = 0; l < k; l++) { + __m256 aik = _mm256_loadu_ps(&A[ i + l * lda]); + __m256 bkj = _mm256_broadcast_ss(&B[ l + j * ldb]); + cij = _mm256_add_ps(cij, _mm256_mul_ps(aik, bkj)); + } + _mm256_storeu_ps(&C[ j * ldc + i ], cij); + } + } +} + +/* column major dat format */ +void MMEngine_strideB_avx2_ps(int m, int n, int k, float alpha, float *A, int lda, + float *B, int ldb, float beta, float *C, int ldc, int strideB) +{ + for (int j = 0; j < n; j++) { + for (int i = 0; i < m; i+=8) { + __m256 cij = _mm256_loadu_ps(&C[ j*ldc + i]); + for (int l = 0; l < k; l++) { + __m256 aik = _mm256_loadu_ps(&A[ i + l * lda]); + __m256 bkj = _mm256_broadcast_ss(&B[ l*strideB + j * ldb]); + cij = _mm256_add_ps(cij, _mm256_mul_ps(aik, bkj)); + } + _mm256_storeu_ps(&C[ j * ldc + i], cij); + } + } +} + diff --git a/FP8_Emulator/cmodel/simple/vla.h b/FP8_Emulator/cmodel/simple/vla.h new file mode 100644 index 00000000..f461878d --- /dev/null +++ b/FP8_Emulator/cmodel/simple/vla.h @@ -0,0 +1,47 @@ +/*----------------------------------------------------------------------------* + * Copyright (c) 2023, Intel Corporation - All rights reserved. + * This file is part of FP8-Emulation-Toolkit + * + * SPDX-License-Identifier: BSD-3-Clause + *----------------------------------------------------------------------------* + * Naveen Mellempudi (Intel Corporation) + *----------------------------------------------------------------------------*/ + +#ifndef VLA_H +#define VLA_H +#include +#include +#include +#include + +#define ALWAYS_INLINE __attribute__((always_inline)) +#define INLINE inline +#define RESTRICT __restrict__ +#define VLA_POSTFIX _ + +#define INDEX1_1(...) ((size_t)SELECT_HEAD(__VA_ARGS__)) +#define INDEX1_2(I0, I1, S1) (INDEX1_1(I0) * ((size_t)S1) + (size_t)I1) +#define INDEX1_3(I0, I1, I2, S1, S2) (INDEX1_2(I0, I1, S1) * ((size_t)S2) + (size_t)I2) +#define INDEX1_4(I0, I1, I2, I3, S1, S2, S3) (INDEX1_3(I0, I1, I2, S1, S2) * ((size_t)S3) + (size_t)I3) +#define INDEX1_5(I0, I1, I2, I3, I4, S1, S2, S3, S4) (INDEX1_4(I0, I1, I2, I3, S1, S2, S3) * ((size_t)S4) + (size_t)I4) +#define INDEX1_6(I0, I1, I2, I3, I4, I5, S1, S2, S3, S4, S5) (INDEX1_5(I0, I1, I2, I3, I4, S1, S2, S3, S4) * ((size_t)S5) + (size_t)I5) +#define INDEX1_7(I0, I1, I2, I3, I4, I5, I6, S1, S2, S3, S4, S5, S6) (INDEX1_6(I0, I1, I2, I3, I4, I5, S1, S2, S3, S4, S5) * ((size_t)S6) + (size_t)I6) +#define INDEX1_8(I0, I1, I2, I3, I4, I5, I6, I7, S1, S2, S3, S4, S5, S6, S7) (INDEX1_7(I0, I1, I2, I3, I4, I5, I6, S1, S2, S3, S4, S5, S6) * ((size_t)S7) + (size_t)I7) +#define INDEX1_9(I0, I1, I2, I3, I4, I5, I6, I7, I8, S1, S2, S3, S4, S5, S6, S7, S8) (INDEX1_8(I0, I1, I2, I3, I4, I5, I6, I7, S1, S2, S3, S4, S5, S6, S7) * ((size_t)S8) + (size_t)I8) +#define INDEX1_10(I0, I1, I2, I3, I4, I5, I6, I7, I8, I9, S1, S2, S3, S4, S5, S6, S7, S8, S9) (INDEX1_9(I0, I1, I2, I3, I4, I5, I6, I7, I8, S1, S2, S3, S4, S5, S6, S7, S8) * ((size_t)S9) + (size_t)I9) + +#define EXPAND(...) __VA_ARGS__ +#define CONCATENATE2(A, B) A##B +#define CONCATENATE(A, B) CONCATENATE2(A, B) +#define INDEX1(NDIMS, ...) CONCATENATE(INDEX1_, NDIMS)(__VA_ARGS__) + +#define SELECT_HEAD_AUX(A, ...) (A) +#define SELECT_HEAD(...) EXPAND(SELECT_HEAD_AUX(__VA_ARGS__, 0)) +#define SELECT_TAIL(A, ...) __VA_ARGS__ + +#define ACCESS_VLA(NDIMS, ARRAY, ...) CONCATENATE(ARRAY, VLA_POSTFIX)[INDEX1(NDIMS, __VA_ARGS__)] +#define DECLARE_VLA(NDIMS, ELEMENT_TYPE, ARRAY_VAR, ...) \ + ELEMENT_TYPE *RESTRICT CONCATENATE(ARRAY_VAR, VLA_POSTFIX) = SELECT_HEAD(__VA_ARGS__) \ + + 0 * INDEX1(NDIMS, SELECT_TAIL(__VA_ARGS__, SELECT_TAIL(__VA_ARGS__, 0))) + +#endif diff --git a/FP8_Emulator/cmodel/tests/conv_grad_test.py b/FP8_Emulator/cmodel/tests/conv_grad_test.py new file mode 100644 index 00000000..74b0c52e --- /dev/null +++ b/FP8_Emulator/cmodel/tests/conv_grad_test.py @@ -0,0 +1,58 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import torch +import time +from mpemu.cmodel import simple + +n = 64 +c = 256 +h = 28 +w = 28 +k = 512 +r = 3 +s = 3 +stride = 2 +pad = 2 + +a = torch.rand((n,c,h,w), dtype=torch.float32) +b = torch.rand((k,c,r,s), dtype=torch.float32) +bias = torch.rand((k), dtype=torch.float32) + +a64 = a.to(dtype=torch.float64, copy=True) +b64 = b.to(dtype=torch.float64, copy=True) + +a.requires_grad=True +b.requires_grad=True +a64.requires_grad=True +b64.requires_grad=True + +ref_time = time.time() +z = torch.nn.functional.conv2d(a64, b64, bias, stride=(stride,stride), padding=(pad,pad), dilation=(1,1), groups=1) +ref_time = time.time()-ref_time + +simple_time = time.time() +z2 = simple.conv2d(a, b, bias, stride=(stride,stride), padding=(pad,pad), dilation=(1,1), groups=1) +simple_time = time.time()-simple_time + +#print("Forward Time : ref_time: {}, simple_time: {} ".format(ref_time, simple_time)) +print('Forward: L2 distance output : ', torch.dist(z2.to(dtype=torch.float64, copy=True), z, 2).item()) + +ref_time = time.time() +(z[0, 0] + z[0,1]).sum().backward() +ref_time = time.time()-ref_time + +simple_time = time.time() +(z2[0, 0] + z2[0,1]).sum().backward() +simple_time = time.time()-simple_time + +#print("BackProp Time : ref_time: {}, simple_time: {}".format(ref_time, simple_time)) +torch.set_printoptions(profile="full") +print('Backward: L2 distance input_grad: ', torch.dist(a.grad.to(dtype=torch.float64, copy=True), a64.grad, 2).item()) +print('Backward: L2 distance weight_grad: ', torch.dist(b.grad.to(dtype=torch.float64, copy=True), b64.grad, 2).item()) diff --git a/FP8_Emulator/cmodel/tests/conv_test.py b/FP8_Emulator/cmodel/tests/conv_test.py new file mode 100644 index 00000000..83babee4 --- /dev/null +++ b/FP8_Emulator/cmodel/tests/conv_test.py @@ -0,0 +1,30 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import torch +from mpemu.cmodel import simple + +n = 2 +c = 64 +h = 28 +w = 28 +k = 64 +r = 3 +s = 3 + +input = torch.rand((n,c,h,w), dtype=torch.float32) +weight = torch.rand((k,c,r,s), dtype=torch.float32) +bias = torch.rand((k), dtype=torch.float32) + +output_ref = torch.nn.functional.conv2d(input, weight, bias, stride=(2,2), padding=(2,2), dilation=(1,1), groups=1) +output_simple = simple.conv2d(input, weight, bias, stride=(2,2), padding=(2,2), dilation=(1,1), groups=1) + +torch.set_printoptions(profile="full") + +print('Forward : L2 distance (simple) : ', torch.dist(output_ref, output_simple, 2).item()) diff --git a/FP8_Emulator/cmodel/tests/gemm_grad_test.py b/FP8_Emulator/cmodel/tests/gemm_grad_test.py new file mode 100644 index 00000000..4f6b2ec5 --- /dev/null +++ b/FP8_Emulator/cmodel/tests/gemm_grad_test.py @@ -0,0 +1,43 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import torch +from mpemu.cmodel import simple + +m=1024 #356320 #356305 +n=1024 #120 #256 +k=1024 #576 #128 #2020 + +# Start here +a = torch.rand((m, k), dtype=torch.float32) +b = torch.rand((n, k), dtype=torch.float32) +c = torch.zeros((m, n), dtype=torch.float32) + +a64 = a.to(dtype=torch.float64, copy=True) +b64 = b.to(dtype=torch.float64, copy=True) +c64 = c.to(dtype=torch.float64, copy=True) + +a.requires_grad=True +b.requires_grad=True +c.requires_grad=True +a64.requires_grad=True +b64.requires_grad=True +c64.requires_grad=True + +z = torch.addmm(c64, a64, b64.t()) +z2 = simple.addmm(c, a, b.t()) + +print('Forward :L2 distance output: ', torch.dist(z2.to(dtype=torch.float64, copy=True), z, 2).item()) + +(z2[0, 0] + z2[0,1]).sum().backward() +(z[0, 0] + z[0,1]).sum().backward() + +print('Backward : L2 distance a_grad: ', torch.dist(a.grad.to(dtype=torch.float64, copy=True), a64.grad, 2).item()) +print('Backward : L2 distance b_grad: ', torch.dist(b.grad.to(dtype=torch.float64, copy=True), b64.grad, 2).item()) +print('Backward : L2 distance c_grad: ', torch.dist(c.grad.to(dtype=torch.float64, copy=True), c64.grad, 2).item()) diff --git a/FP8_Emulator/cmodel/tests/gemm_irregular_test.py b/FP8_Emulator/cmodel/tests/gemm_irregular_test.py new file mode 100644 index 00000000..50541b8d --- /dev/null +++ b/FP8_Emulator/cmodel/tests/gemm_irregular_test.py @@ -0,0 +1,42 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import torch +import fuse + +m=1 #356320 #356305 +n=120 #256 +k=576 #128 #2020 + +# Start here +a = torch.rand((5, 107, 1024), dtype=torch.float32) +b = torch.rand((1024, 1024), dtype=torch.float32) +c = torch.zeros((m, n), dtype=torch.float32) + +a64 = a.to(dtype=torch.float64, copy=True) +b64 = b.to(dtype=torch.float64, copy=True) +c64 = c.to(dtype=torch.float64, copy=True) + +z = torch.matmul(a64, b64) +z3 = torch.matmul(a, b) +for a1 in a : + print("-->", a1.size()) + +#z2l = tuple([fuse.tmul.matmul(a1, b) for a1 in a]) +#z2 = torch.stack(z2l) +#z2 = torch.stack(tuple([fuse.tmul.matmul(a1, b) for a1 in a])) +z2 = fuse.tmul.matmul(a, b) + +#print (z3.size(), z2.size()) + +#print("torch :", z3.size(), z3) +#print("Ours : ", z2.size(), z2) +print('32b: L2 distance : ', torch.dist(z, z3.to(dtype=torch.float64, copy=True), 2)) +print('ours: L2 distance : ', torch.dist(z, z2.to(dtype=torch.float64, copy=True), 2)) + diff --git a/FP8_Emulator/cmodel/tests/gemm_test.py b/FP8_Emulator/cmodel/tests/gemm_test.py new file mode 100644 index 00000000..f42b8c7f --- /dev/null +++ b/FP8_Emulator/cmodel/tests/gemm_test.py @@ -0,0 +1,41 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import torch +from mpemu.cmodel import simple + +def get_grads(variables): + return [var.grad.clone() for var in variables] + +m=128 #356320 #356305 +n=120 #256 +k=576 #128 #2020 + +a = torch.rand((m, k), dtype=torch.float32) +b = torch.rand((n, k), dtype=torch.float32) +c = torch.zeros((m, n), dtype=torch.float32) + + +a64 = a.to(dtype=torch.float64, copy=True) +b64 = b.to(dtype=torch.float64, copy=True) +c64 = c.to(dtype=torch.float64, copy=True) + +z = torch.matmul(a64, b64.t()) +z2 = simple.matmul(a, b.t()) +z3 = torch.matmul(a, b.t()) + +z2gb = torch.matmul(z2, b) +z2g = simple.matmul(z2, b) +z2gwb = torch.matmul(z2.t(), a) +z2gw = simple.matmul(z2.t(), a) + +print('32b: L2 distance : ', torch.dist(z, z3.to(dtype=torch.float64, copy=True), 2)) +print('output : L2 distance : ', torch.dist(z, z2.to(dtype=torch.float64, copy=True), 2)) +print('Grad a : L2 distance : ', torch.dist(z2g, z2gb, 2)) +print('Grad b : L2 distance : ', torch.dist(z2gw, z2gwb, 2)) diff --git a/FP8_Emulator/cmodel/tests/linear_test.py b/FP8_Emulator/cmodel/tests/linear_test.py new file mode 100644 index 00000000..c5306975 --- /dev/null +++ b/FP8_Emulator/cmodel/tests/linear_test.py @@ -0,0 +1,65 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.optim as optim +from mpemu.cmodel import simple + +class Net(nn.Module): + + def __init__(self): + super(Net, self).__init__() + self.fc = nn.Linear(576, 120) + + def forward(self, x): + x = self.fc(x) + return x + + def num_flat_features(self, x): + size = x.size()[1:] # all dimensions except the batch dimension + num_features = 1 + for s in size: + num_features *= s + return num_features + +net = Net() +net1 = Net() +print(net) +input = torch.randn((1, 576), dtype=torch.float32, device="cpu") +input_new = input.to(dtype=torch.float32, copy=True) + +output = net(input) +#print("fc output:", output.size(), output) + +target = torch.randn(120, dtype=torch.float32, device="cpu") # a dummy target, for example +target = target.view(1, -1) # make it the same shape as output +criterion = nn.MSELoss() +loss = criterion(output, target) +net.zero_grad() # zeroes the gradient buffers of all parameters +#loss.backward(retain_graph=True) +loss.backward() +print("fc weight grads:", net.fc.weight.grad) + + +torch.addmm_back = torch.addmm +torch.matmul_back = torch.matmul +torch.addmm = simple.addmm +torch.matmul = simple.matmul + +output1 = net1(input_new) +#print("fc output:", output1.size(), output1) + +loss1 = criterion(output1, target) +net1.zero_grad() # zeroes the gradient buffers of all parameters +loss1.backward() +print("fc weight grads:", net1.fc.weight.grad) + +print('Linear wtgrads L2 distance : ', torch.dist(net.fc.weight.grad, net1.fc.weight.grad, 2)) diff --git a/FP8_Emulator/cmodel/tests/net.py b/FP8_Emulator/cmodel/tests/net.py new file mode 100644 index 00000000..8f2e2c74 --- /dev/null +++ b/FP8_Emulator/cmodel/tests/net.py @@ -0,0 +1,73 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.optim as optim +from mpemu.cmodel import simple + +torch_conv2d = torch.nn.functional.conv2d +torch_addmm = torch.addmm +torch.addmm = simple.addmm +torch.nn.functional.conv2d = simple.conv2d + +class Net(nn.Module): + + def __init__(self): + super(Net, self).__init__() + # 1 input image channel, 6 output channels, 3x3 square convolution + # kernel + self.conv1 = nn.Conv2d(1, 6, 3) + self.conv2 = nn.Conv2d(6, 16, 3) + # an affine operation: y = Wx + b + self.fc1 = nn.Linear(16 * 6 * 6, 120) # 6*6 from image dimension + self.fc2 = nn.Linear(120, 84) + self.fc3 = nn.Linear(84, 10) + + def forward(self, x): + # Max pooling over a (2, 2) window + x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2)) + # If the size is a square you can only specify a single number + x = F.max_pool2d(F.relu(self.conv2(x)), 2) + x = x.view(-1, self.num_flat_features(x)) + x = F.relu(self.fc1(x)) + x = F.relu(self.fc2(x)) + x = self.fc3(x) + return x + + def num_flat_features(self, x): + size = x.size()[1:] # all dimensions except the batch dimension + num_features = 1 + for s in size: + num_features *= s + return num_features + + +net = Net() +print(net) +params = list(net.parameters()) +input = torch.randn(1, 1, 32, 32, dtype=torch.float32, device="cpu") +# create your optimizer +optimizer = optim.SGD(net.parameters(), lr=0.1) +optimizer.zero_grad() # zero the gradient buffers + +output = net(input) + +print("fc3 output:", output) +target = torch.randn(10, dtype=torch.float32, device="cpu") # a dummy target, for example +target = target.view(1, -1) # make it the same shape as output +criterion = nn.MSELoss() + +loss = criterion(output, target) + +net.zero_grad() # zeroes the gradient buffers of all parameters + +loss.backward() +optimizer.step() # Does the update diff --git a/FP8_Emulator/pytquant/__init__.py b/FP8_Emulator/pytquant/__init__.py new file mode 100644 index 00000000..e04cab6a --- /dev/null +++ b/FP8_Emulator/pytquant/__init__.py @@ -0,0 +1,6 @@ +import warnings +import torch + +from . import cpp +if torch.cuda.is_available(): + from . import cuda diff --git a/FP8_Emulator/pytquant/cpp/__init__.py b/FP8_Emulator/pytquant/cpp/__init__.py new file mode 100644 index 00000000..60993b75 --- /dev/null +++ b/FP8_Emulator/pytquant/cpp/__init__.py @@ -0,0 +1 @@ +from . import fpemu as fpemu_cpp diff --git a/FP8_Emulator/pytquant/cpp/avx-fpemu.cpp b/FP8_Emulator/pytquant/cpp/avx-fpemu.cpp new file mode 100644 index 00000000..74f509ff --- /dev/null +++ b/FP8_Emulator/pytquant/cpp/avx-fpemu.cpp @@ -0,0 +1,802 @@ +#include +#include +#include +#include + +enum ROUNDING_MODES { + ROUND_RTZ = 0, + ROUND_RNE = 1, + ROUND_STOCHASTIC = 2, + ROUND_RNAZ = 3, + ROUND_RNTZ = 4, + ROUND_PINF = 5, + ROUND_NINF = 6 +}; // 枚举定义了不同的舍入模式（决定如何将一个浮点数或整数近似为另一个更接近的值的规则） + +namespace { + + typedef union half_t { + unsigned short u; + at::Half f; + } __half_t; + + typedef union ufloat32 { + unsigned u; + float f; + } __float_t; + +/* Following implementation of xoroshiro128++ PRNG is borrowed from here: + http://prng.di.unimi.it/xoshiro128plusplus.c + main page: http://prng.di.unimi.it/ +*/ + static uint32_t s1_[4] = { 1387366120, 2798441831, 888998500 , 1099633400 }; + static uint32_t s2_[4] = { 2034269327, 2125325156, 1209715489, 1931656721 }; + static uint32_t s3_[4] = { 1555452618, 650181557 , 883695203 , 627677842 }; + static uint32_t s4_[4] = { 4195248041, 2146478152, 480059239 , 1468956197 }; + static uint32_t s5_[4] = { 1252084877, 500390994 , 977516591 , 1950666000 }; + static uint32_t s6_[4] = { 3936597502, 834151069 , 1477014702, 734008143 }; + static uint32_t s7_[4] = { 1983400973, 1164103095, 2110188261, 2019272068 }; + static uint32_t s8_[4] = { 1877096364, 2833629967, 4196320416, 1774181187 }; + static uint32_t s9_[4] = { 702309618 , 4077815558, 1512057936, 1868769368 }; + static uint32_t s10_[4] = + { 510001215 , 966559856 , 776583255 , 1475621065 }; + static uint32_t s11_[4] = + { 1271806057, 1881312534, 478635452 , 814821902 }; + static uint32_t s12_[4] = + { 733990058 , 1889991804, 1108257970, 1093480892 }; + static uint32_t s13_[4] = + { 4273743809, 4167473370, 558000409 , 1594848927 }; + static uint32_t s14_[4] = + { 444870959 , 1595722866, 1064124488, 3637102547 }; + static uint32_t s15_[4] = + { 703721499 , 3896407831, 1002360059, 1427395742 }; + static uint32_t s16_[4] = + { 1295231497, 1254972431, 1423497865, 861918264 }; + +/* seed pointer array */ + static uint32_t *sptr_[16] = { s1_, s2_, s3_, s4_, s5_, s6_, s7_, s8_, s9_, + s10_, s11_, s12_, s13_, s14_, s15_, s16_ + }; + + static inline uint32_t rotl (const uint32_t x, int k) { + return (x << k) | (x >> (32 - k)); + } + + uint32_t rand_xorshft128plus_scalar (uint32_t * ps) { + const uint32_t result_plus = ps[0] + ps[3]; + const uint32_t t = ps[1] << 9; + + ps[2] ^= ps[0]; + ps[3] ^= ps[1]; + ps[1] ^= ps[2]; + ps[0] ^= ps[3]; + + ps[2] ^= t; + + ps[3] = rotl (ps[3], 11); + + return result_plus; + } + + float __double2float_rn (double inval) { + float out[4] = { 0 }; + __m128 vout = _mm_cvtpd_ps (_mm_set1_pd (inval)); + + _mm_store_ps (&out[0], vout); + return out[0]; + } + + unsigned short __float2half_rn (float inval) { + return _cvtss_sh (inval, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + } + + float __half2float (unsigned short h_val) { + return _cvtsh_ss (h_val); + } + + template < typename scalar_t > float __anyfloat2float_rn (scalar_t a_) { + float f_; + + if (std::is_same < scalar_t, double >::value) { + f_ = __double2float_rn (a_); + } else if (std::is_same < scalar_t, float >::value) { + f_ = a_; + } else if (std::is_same < scalar_t, at::Half >::value) { + f_ = __half2float ((at::Half) a_); + } + return f_; + } + + template < typename scalar_t > + void __float2anyfloat_rn (float f_, scalar_t * out) { + scalar_t a_; + + if (std::is_same < scalar_t, double >::value) { + a_ = (scalar_t) (f_); + } else if (std::is_same < scalar_t, float >::value) { + a_ = f_; + } else if (std::is_same < scalar_t, at::Half >::value) { + a_ = (at::Half) __float2half_rn (f_); + } + *out = a_; + } + + template < typename scalar_t > + unsigned short __anyfloat2half_rn (scalar_t f_) { + unsigned short h_; + + if (std::is_same < scalar_t, double >::value) { + h_ = __float2half_rn (__double2float_rn (f_)); + } else if (std::is_same < scalar_t, float >::value) { + h_ = __float2half_rn (f_); + } else if (std::is_same < scalar_t, at::Half >::value) { + unsigned short *ptrh_ = (unsigned short *) &f_; + h_ = *ptrh_; + } + return h_; + } + + template < typename scalar_t > + void __half2anyfloat (unsigned short h_, scalar_t * out, scalar_t scale=1.0) { + scalar_t f_; + + if (std::is_same < scalar_t, double >::value) { + f_ = (scalar_t) __half2float (h_); + } else if (std::is_same < scalar_t, float >::value) { + f_ = __half2float (h_); + } else if (std::is_same < scalar_t, at::Half >::value) { + f_ = *((at::Half *) & h_); + } + *out = scale*f_; + } + + template < typename scalar_t > + inline void reduce0 (scalar_t * g_data, float *g_odata, unsigned int n) { + float sum = 0.0, sumsq = 0.0; + +#pragma omp parallel for reduction(+: sum), reduction(+: sumsq) + for (unsigned int i = 0; i < n; i++) { + sum += __anyfloat2float_rn (g_data[i]); + sumsq += __anyfloat2float_rn (g_data[i]) * __anyfloat2float_rn (g_data[i]); + } + g_odata[0] = sum; + g_odata[1] = sumsq; + } + + template < typename scalar_t > + inline void absmax0 (scalar_t * g_data, float *g_odata, unsigned int n) { + float absmax = 0.0; + +#pragma omp parallel for reduction(max: absmax) + for (unsigned int i = 0; i < n; i++) { + absmax = fmaxf (absmax, fabsf (__anyfloat2float_rn (g_data[i]))); + } + g_odata[0] = absmax; + } + + static inline __m256i _mm256_rand_xorshft128plus_epi32(uint32_t *vs0, + uint32_t *vs1, + uint32_t *vs2, + uint32_t *vs3) { + const __m256i vrplus = _mm256_add_epi32(_mm256_load_si256((__m256i *)vs0), + _mm256_load_si256((__m256i *)vs3)); + const __m256i vt = + _mm256_sll_epi32(_mm256_load_si256((__m256i *)vs1), _mm_cvtsi32_si128(9)); + + _mm256_store_si256((__m256i *)vs2, + _mm256_xor_si256(_mm256_load_si256((__m256i *)vs2), + _mm256_load_si256((__m256i *)vs0))); + _mm256_store_si256((__m256i *)vs3, + _mm256_xor_si256(_mm256_load_si256((__m256i *)vs3), + _mm256_load_si256((__m256i *)vs1))); + _mm256_store_si256((__m256i *)vs1, + _mm256_xor_si256(_mm256_load_si256((__m256i *)vs1), + _mm256_load_si256((__m256i *)vs2))); + _mm256_store_si256((__m256i *)vs0, + _mm256_xor_si256(_mm256_load_si256((__m256i *)vs0), + _mm256_load_si256((__m256i *)vs3))); + _mm256_store_si256((__m256i *)vs2, + _mm256_xor_si256(_mm256_load_si256((__m256i *)vs2), vt)); + + __m256i vl = _mm256_slli_epi32(_mm256_load_si256((__m256i *)vs3), 11); + __m256i vr = _mm256_srli_epi32(_mm256_load_si256((__m256i *)vs3), 32 - 11); + + _mm256_store_si256((__m256i *)vs3, _mm256_or_si256(vl, vr)); + + return vrplus; + } + + void cvt_fp32_e5m2_rne_intrinsic (const float *__restrict__ in, float *out, + int size, float scale) { + +#pragma omp parallel for + for (int i = 0; i < size; i += 16){ + const __m256i vnaninf = _mm256_set1_epi16 (0x7c00); + const __m256i vrneadd = _mm256_set1_epi16 (0x007f); + const __m256i vfixup = _mm256_set1_epi16 (0x0001); + const __m256i vfixupmask = _mm256_set1_epi16 (0x0100); + + __m256 s_ = _mm256_set1_ps (scale); + __m256 sr_ = _mm256_set1_ps (1.0 / scale); + __m256 b = _mm256_loadu_ps (&in[i]); + __m256 a = _mm256_loadu_ps (&in[i + 8]); + + b = _mm256_mul_ps(b, s_); + a = _mm256_mul_ps(a, s_); + + __m128i ah_ = _mm256_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = _mm256_cvtps_ph(b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + const __m256i a_ = _mm256_insertf128_si256(_mm256_insertf128_si256(_mm256_setzero_si256(), bh_, 0), ah_, 1); + const __m256i maska1_ = _mm256_cmpeq_epi16(_mm256_and_si256(a_, vnaninf), vnaninf); + const __m256i maska2_ = _mm256_cmpeq_epi16(_mm256_and_si256(a_, vfixupmask), vfixupmask); + __m256i a_rne_ = _mm256_blendv_epi8(a_, _mm256_add_epi16(a_, vrneadd), _mm256_and_si256(maska1_, maska2_)); + a_rne_ = _mm256_slli_epi16(_mm256_srli_epi16(a_rne_, 8), 8); + + bh_ = _mm256_extracti128_si256(a_rne_, 0); + ah_ = _mm256_extracti128_si256(a_rne_, 1); + b = _mm256_cvtph_ps(bh_); + a = _mm256_cvtph_ps(ah_); + + _mm256_storeu_ps(&out[i], _mm256_mul_ps(b, sr_)); + _mm256_storeu_ps(&out[i + 8], _mm256_mul_ps(a, sr_)); + } + } + + void cvt_fp32_e5m2_stochastic_intrinsic(const float *__restrict__ in, float *out, + int size, float scale) { + uint32_t vs0[8] __attribute__((aligned(32))) = { + 1387366120, 279844183, 888998500, 1099633400, + 1252084877, 500390994, 977516591, 1950666000}; + uint32_t vs1[8] __attribute__((aligned(32))) = { + 2034269327, 2125325156, 1209715489, 193165672, + 187709636, 28336299, 419632041, 1774181187}; + uint32_t vs2[8] __attribute__((aligned(32))) = { + 1555452618, 650181557, 883695203, 62767784, + 127180605, 1881312534, 478635452, 814821902}; + uint32_t vs3[8] __attribute__((aligned(32))) = { + 419524804, 2146478152, 480059239, 1468956197, + 444870959, 1595722866, 1064124488, 363710254}; + +#pragma omp parallel for firstprivate(vs0, vs1, vs2, vs3) + for (int i = 0; i < size; i += 16) { + const __m256i vnaninf = _mm256_set1_epi16(0x7c00); + const __m256i vfixup = _mm256_set1_epi16(0x0001); + const __m256i vfixupmask = _mm256_set1_epi16(0x0100); + const __m256i vrneadd = _mm256_set1_epi16(0x007f); + const __m256i vdenorm = _mm256_set1_epi16(0x03ff); + const __m256i vexmant = _mm256_set1_epi16(0x7fff); + + __m256i rnd256 = _mm256_rand_xorshft128plus_epi32(vs0, vs1, vs2, vs3); + __m128i rnbits = _mm256_extractf128_si256(rnd256, 0); + + __m256 s_ = _mm256_set1_ps(scale); + __m256 sr_ = _mm256_set1_ps(1.0 / scale); + + __m256 b = _mm256_loadu_ps(&in[i]); + __m256 a = _mm256_loadu_ps(&in[i + 8]); + + b = _mm256_mul_ps(b, s_); + a = _mm256_mul_ps(a, s_); + + __m128i ah_ = _mm256_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = _mm256_cvtps_ph(b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + + const __m256i a_ = _mm256_insertf128_si256(_mm256_castsi128_si256(bh_), ah_, 1); + const __m256i maska1_ = _mm256_cmpeq_epi16(_mm256_and_si256(a_, vnaninf), vnaninf); + const __m256i maska2_ = _mm256_cmpeq_epi16(_mm256_and_si256(a_, vfixupmask), vfixupmask); + const __m256i maska4_ = _mm256_cmpgt_epi16(vdenorm, _mm256_and_si256(a_, vexmant)); + + __m256i a_sr_ = _mm256_blendv_epi8(a_, _mm256_add_epi16(a_, _mm256_cvtepu8_epi16(rnbits)), _mm256_andnot_si256(maska4_, maska1_)); + a_sr_ = _mm256_blendv_epi8(a_sr_, _mm256_add_epi16(a_sr_, vrneadd), _mm256_and_si256(maska4_, maska2_)); + a_sr_ = _mm256_slli_epi16(_mm256_srli_epi16(a_sr_, 8), 8); + + bh_ = _mm256_extracti128_si256(a_sr_, 0); + ah_ = _mm256_extracti128_si256(a_sr_, 1); + + b = _mm256_cvtph_ps(bh_); + a = _mm256_cvtph_ps(ah_); + + _mm256_storeu_ps(&out[i], _mm256_mul_ps(b, sr_)); + _mm256_storeu_ps(&out[i + 8], _mm256_mul_ps(a, sr_)); + } + } + void cvt_fp32_e5m2_scalar (const float *__restrict__ in, float *out, + int size, float scale, int rmode) { + int non_mant_bits = 5 /*exp_bits */ + 1; /* exponent + sign */ + int lshift = 10 - (8 /*mbits */ - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x00FF; + unsigned short rne_tie = 0x0180; + + float scale_reciprocal = 1.0 / scale; + + for (int gid = 0; gid < size; gid++) { + __half_t h; + float inval = scale * in[gid]; + + h.u = __anyfloat2half_rn (inval); + + unsigned short can_round = ((h.u & 0x7F00) <= 0x7B00) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + /* nearest rounding masks */ + unsigned short rnmask = (h.u & grs_bitmask); + unsigned short rnmask_tie = (h.u & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + rand_xorshft128plus_scalar (sptr_[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + h.u += can_round * is_normal * (rand & 0xFF); + /* stochastic round: denormals --> rne rounding */ + h.u += can_round * is_denorm * + (((rnmask > 0x0080) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + h.u += can_round * rne_mask * + (((rnmask > 0x0080) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + h.u += can_round * rnaz_mask * ((rnmask >= 0x0080) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + h.u += can_round * rntz_mask * ((rnmask > 0x0080) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + h.u += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0080) << lshift); + /* round to -INF, if rminf_mask is enabled */ + h.u += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0080) << lshift); + } + } + /* truncation */ + h.u = (h.u & mask_mant); + float f_; + __half2anyfloat (h.u, &f_); + out[gid] = f_ * scale_reciprocal; + } + } + void cvt_fp32_e4m3_rne_intrinsic (const float *__restrict__ in, float *out, + int size, float scale) { + const __m256i vnaninf = _mm256_set1_epi16 (0x7c00); + const __m256i vrneadd = _mm256_set1_epi16 (0x003f); + const __m256i vfixup = _mm256_set1_epi16 (0x0001); + const __m256i vfixupmask = _mm256_set1_epi16 (0x0080); + const __m256i vzero = _mm256_set1_epi16 (0x0000); + const __m256i vsign = _mm256_set1_epi16 (0x8000); + const __m256i vsatuval = _mm256_set1_epi16 (0x5F00);/* 2^8*1.110 a.k.a 448.0, largest value */ + const __m256i vflush = _mm256_set1_epi16 (0x1800);/* 2^-9, smallest denormal */ + const __m256i vxdnorm = _mm256_set1_epi16 (0x2400);/* 2^-6 smallest normal */ + + for (int i = 0; i < size; i += 16){ + __m256 s_ = _mm256_set1_ps(scale); + __m256 sr_ = _mm256_set1_ps(1.0 / scale); + __m256 b = _mm256_loadu_ps(&in[i]); + __m256 a = _mm256_loadu_ps(&in[i + 8]); + + b = _mm256_mul_ps(b, s_); + a = _mm256_mul_ps(a, s_); + + __m128i ah_ = _mm256_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = _mm256_cvtps_ph(b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i a_ = _mm256_insertf128_si256(_mm256_insertf128_si256(_mm256_setzero_si256(), bh_, 0), ah_, 1); + const __m256i maska1_ = _mm256_cmpeq_epi16(_mm256_and_si256(a_, vnaninf), vnaninf); + const __m256i maska2_ = _mm256_cmpeq_epi16(_mm256_and_si256(a_, vfixupmask), vfixupmask); + const __m256i maska3_ = _mm256_cmpgt_epi16(vsatuval, _mm256_and_si256(a_, _mm256_set1_epi16(0x7FFF))); + const __m256i maska4_ = _mm256_cmpgt_epi16(vflush, _mm256_and_si256(a_, vnaninf)); + const __m256i maska5_ = _mm256_cmpgt_epi16(vxdnorm, _mm256_and_si256(a_, vnaninf)); + + __m256i v_shft = _mm256_sub_epi16(_mm256_srli_epi16(vxdnorm, 10), _mm256_srli_epi16(_mm256_and_si256(a_, vnaninf), 10)); + __m256i a_rne_ = _mm256_blendv_epi8(a_, _mm256_add_epi16(a_rne_, vrneadd), _mm256_and_si256(maska1_, maska2_)); + a_rne_ = _mm256_slli_epi16(_mm256_srli_epi16(a_rne_, 8), 8); + bh_ = _mm256_extracti128_si256(a_rne_, 0); + ah_ = _mm256_extracti128_si256(a_rne_, 1); + b = _mm256_cvtph_ps(bh_); + a = _mm256_cvtph_ps(ah_); + _mm256_storeu_ps(&out[i], _mm256_mul_ps(b, sr_)); + _mm256_storeu_ps(&out[i + 8], _mm256_mul_ps(a, sr_)); + } + } + + void cvt_fp32_e4m3_stochastic_intrinsic (const float *__restrict__ in, + float *out, int size, float scale) { + const __m256i vnaninf = _mm256_set1_epi16(0x7c00); + const __m256i vfixup = _mm256_set1_epi16(0x0001); + const __m256i vfixupmask = _mm256_set1_epi16(0x0100); + const __m256i vrneadd = _mm256_set1_epi16(0x003f); + const __m256i vdenorm = _mm256_set1_epi16(0x03ff); + const __m256i vexmant = _mm256_set1_epi16(0x7fff); + + for (int i = 0; i < size; i += 16) { + unsigned int rndbuf[16]; + /* generate 128 random bits */ + for (int r = 0; r < 8; r++) { + rndbuf[r] = (unsigned int) rand_xorshft128plus_scalar (sptr_[r]); + } + __m128i rnbits = _mm_load_si128 ((const __m128i *) &rndbuf[0]); + + __m256 s_ = _mm256_set1_ps (scale); + __m256 sr_ = _mm256_set1_ps (1.0 / scale); + __m256 b = _mm256_loadu_ps (&in[i]); + __m256 a = _mm256_loadu_ps (&in[i + 8]); + + b = _mm256_mul_ps (b, s_); + a = _mm256_mul_ps (a, s_); + + __m128i ah_ = + _mm256_cvtps_ph(a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = + _mm256_cvtps_ph(b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + + const __m256i a_ = + _mm256_insertf128_si256(_mm256_castsi128_si256(bh_), ah_, 1); + + const __m256i maska1_ = + _mm256_cmpeq_epi16(_mm256_and_si256(a_, vnaninf), vnaninf); + const __m256i maska2_ = + _mm256_cmpeq_epi16(_mm256_and_si256(a_, vfixupmask), vfixupmask); + const __m256i maska4_ = + _mm256_cmpgt_epi16(vdenorm, _mm256_and_si256(a_, vexmant)); + + __m256i a_sr_ = _mm256_blendv_epi8( + a_, _mm256_add_epi16(a_, _mm256_cvtepu8_epi16(rnbits)), + _mm256_andnot_si256(maska4_, maska1_)); + + a_sr_ = _mm256_blendv_epi8(a_sr_, _mm256_add_epi16(a_sr_, vrneadd), + _mm256_and_si256(maska4_, maska2_)); + + a_sr_ = _mm256_slli_epi16(_mm256_srli_epi16(a_sr_, 8), 8); + + bh_ = _mm256_extracti128_si256(a_sr_, 0); + ah_ = _mm256_extracti128_si256(a_sr_, 1); + + b = _mm256_cvtph_ps(bh_); + a = _mm256_cvtph_ps(ah_); + + _mm256_storeu_ps(&out[i], _mm256_mul_ps(b, sr_)); + _mm256_storeu_ps(&out[i + 8], _mm256_mul_ps(a, sr_)); + } + } + void cvt_fp32_e4m3_scalar (const float *__restrict__ in, float *out, + int size, float scale, int rmode) { + int non_mant_bits = 4 /*exp_bits */ + 1; /* exponent + sign */ + int lshift = 10 - (8 /*mbits */ - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x007F; + unsigned short rne_tie = 0x00C0; + float scale_reciprocal = 1.0 / scale; + +#pragma omp parallel for + for (int gid = 0; gid < size; gid++) { + __half_t h; + float inval = scale * in[gid]; + + h.u = __anyfloat2half_rn (inval); + short exp_h = (short) ((h.u & 0x7C00) >> 10) - 15; + short sign_h = (h.u & 0x8000); + short mantissa_h = (h.u & 0x03FF); + + unsigned short can_round = ((h.u & 0x7FFF) < 0x5F00) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + int dshift = 0; + + if (exp_h > 8 || (can_round == 0)) { + /* Software : saturate values above to +/-448.0 to +/-448.0 */ + mantissa_h = 0x0300; + exp_h = 8; + can_round = 0; + } else if (exp_h < -9) { + /* flush values below 1-4-3 subnormal range to zero */ + exp_h = -15; + mantissa_h = 0; + } else if (exp_h < -6) { + dshift = (-6 - exp_h); + /* handle denormals */ + mantissa_h = mantissa_h >> dshift; + mantissa_h <<= dshift; + } + /* nearest rounding masks */ + unsigned short rnmask = (mantissa_h & grs_bitmask); + unsigned short rnmask_tie = (mantissa_h & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + rand_xorshft128plus_scalar (sptr_[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + mantissa_h += can_round * is_normal * (rand & 0x7F); + /* stochastic round: denormals --> rne rounding */ + mantissa_h += can_round * is_denorm * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + mantissa_h += can_round * rne_mask * (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + mantissa_h += can_round * rnaz_mask * ((rnmask >= 0x0040) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + mantissa_h += can_round * rntz_mask * ((rnmask > 0x0040) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + mantissa_h += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0040) << lshift); + /* round to -INF, if rminf_mask is enabled */ + mantissa_h += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0040) << lshift); + } + } + /* truncation */ + mantissa_h &= mask_mant; + mantissa_h += ((exp_h + 15) << 10); + mantissa_h |= sign_h; + h.u = mantissa_h; + float f; + + __half2anyfloat (h.u, &f); + out[gid] = (f * scale_reciprocal); + } + } + + template < typename scalar_t > + void E4M3_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, + int block_size, int rmode) { + float scale = in_scale; + float fmax = std::numeric_limits::max(); + if (scale > fmax) { + fprintf(stderr,"Error: Invalid scale factor : %.2e, make sure the scale is not larger than : %.2e\n", scale, fmax); + exit(1); + } + + if (block_norm == true) { + int nblocks = (size + (block_size - 1)) / block_size; + +#pragma omp parallel for + for (int b = 0; b < nblocks; b++) { + int start_index = (b * block_size); + + /* handle the last block */ + if (start_index + block_size > size) + block_size = (size - start_index); + + float maxval = 0.0; + +#pragma omp parallel for reduction (max:maxval) + for (int gid = start_index; gid < start_index + block_size; gid++) { + maxval = (maxval < fabs (in[gid])) ? fabs (in[gid]) : maxval; + } + __float_t f; + + f.f = maxval; + f.u = (f.u & 0x7F800000); + scale = 2.0 * f.f; + scale /= 8.0; + + if ((block_size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e4m3_stochastic_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } else { + cvt_fp32_e4m3_rne_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } + } else { + cvt_fp32_e4m3_scalar (&in[start_index], &out[start_index], block_size, scale, rmode); + } + } + } else { + if ((size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e4m3_stochastic_intrinsic (in, out, size, scale); + } else { + cvt_fp32_e4m3_rne_intrinsic (in, out, size, scale); + } + } else { + int vec_size = ((int) (size / 32)) * 32; + + if (vec_size > 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e4m3_stochastic_intrinsic (in, out, vec_size, scale); + } else { + cvt_fp32_e4m3_rne_intrinsic (in, out, vec_size, scale); + } + } + cvt_fp32_e4m3_scalar (&in[vec_size], &out[vec_size], size - vec_size, scale, rmode); + } + } + } + + template < typename scalar_t > + void E5M2_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, int block_size, int rmode) { + float scale = in_scale; + float fmax = std::numeric_limits::max(); + if (scale > fmax) { + fprintf(stderr,"Error: Invalid scale factor : %.2e, make sure the scale is not larger than : %.2e\n", scale, fmax); + exit(1); + } + + if (block_norm == true) { + int nblocks = (size + (block_size - 1)) / block_size; + +#pragma omp parallel for + for (int b = 0; b < nblocks; b++) { + int start_index = (b * block_size); + /* handle the last block */ + if (start_index + block_size > size) + block_size = (size - start_index); + + float maxval = 0.0; + +#pragma omp parallel for reduction (max:maxval) + for (int gid = start_index; gid < start_index + block_size; gid++) { + maxval = (maxval < fabs (in[gid])) ? fabs (in[gid]) : maxval; + } + __float_t f; + + f.f = maxval; + f.u = (f.u & 0x7F800000); + scale = 2.0 * f.f; + scale /= 16384.0; + + if ((block_size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e5m2_stochastic_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } else { + cvt_fp32_e5m2_rne_intrinsic (&in[start_index], &out[start_index], block_size, scale); + //cvt_fp32_e5m2_noinf_rne_intrinsic (&in[start_index], &out[start_index], block_size, scale); + //cvt_fp32_e5m2_flex_intrinsic(&in[start_index], &out[start_index], block_size, scale); + } + } else { + cvt_fp32_e5m2_scalar (&in[start_index], &out[start_index], block_size, scale, rmode); + } + } + } else { + if ((size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e5m2_stochastic_intrinsic (in, out, size, scale); + } else { + cvt_fp32_e5m2_rne_intrinsic (in, out, size, scale); + //cvt_fp32_e5m2_noinf_rne_intrinsic (in, out, size, scale); + } + } else { + int vec_size = ((int) (size / 32)) * 32; + + if (vec_size > 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e5m2_stochastic_intrinsic (in, out, vec_size, scale); + } else { + cvt_fp32_e5m2_rne_intrinsic (in, out, vec_size, scale); + //cvt_fp32_e5m2_noinf_rne_intrinsic (in, out, vec_size, scale); + //cvt_fp32_e5m2_flex_intrinsic(in, out, vec_size, scale); + } + } + cvt_fp32_e5m2_scalar (&in[vec_size], &out[vec_size], size - vec_size, scale, rmode); + } + } + } + + std::vector < torch::Tensor > fpemu_common_function (torch::Tensor input, + std::string mode, + int size, + bool inplace, + float scale, + bool block_norm, + int block_size) { + + torch::Tensor output; + if (!inplace) + output = torch::zeros_like (input); + + if (!mode.compare ("E4M3_STOCHASTIC")) { + E4M3_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size, ROUND_STOCHASTIC); + } else if (!mode.compare ("E4M3_RNE")) { + E4M3_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size, ROUND_RNE); + } else if (!mode.compare ("E5M2_STOCHASTIC")) { + E5M2_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size, ROUND_STOCHASTIC); + } else if (!mode.compare ("E5M2_RNE")) { + E5M2_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, scale, block_norm, block_size, ROUND_RNE); + } + + if (!inplace) { + return { + output,}; + } else { + return { + input,}; + } + } + +}//namespace + +std::vector < torch::Tensor > fpemu_forward (torch::Tensor input, + std::string mode, + int size, + bool inplace, + float scale, + bool block_norm, + int block_size) { + if (block_norm == true && block_size != size) { + if (size % block_size) { + block_norm = false; + block_size = 1; + } + } + return fpemu_common_function (input, mode, size, inplace, scale, block_norm, + block_size); +} + +std::vector < torch::Tensor > fpemu_backward (torch::Tensor grad, + std::string mode, + int size, + bool inplace, + float scale, + bool block_norm, + int block_size) { + if (block_norm == true && block_size != size) { + if (size % block_size) { + block_norm = false; + block_size = 1; + } + } + return fpemu_common_function (grad, mode, size, inplace, scale, block_norm, + block_size); +} + +PYBIND11_MODULE (TORCH_EXTENSION_NAME, m) { + m.def ("forward", &fpemu_forward, "FPEmu forward"); + m.def ("backward", &fpemu_backward, "FPEmu backward"); +} \ No newline at end of file diff --git a/FP8_Emulator/pytquant/cpp/fpemu.py b/FP8_Emulator/pytquant/cpp/fpemu.py new file mode 100644 index 00000000..9503c7c6 --- /dev/null +++ b/FP8_Emulator/pytquant/cpp/fpemu.py @@ -0,0 +1,73 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import math +from torch import nn +from torch.autograd import Function +import torch +import numpy +import fpemu_cpp + +from enum import Enum + +torch.manual_seed(42) + +""" + NONE + E5M2_RTZ + E5M2_STOCHASTIC + E5M2_RNE + E5M2_RNAZ + E5M2_RNTZ + E5M2_RPINF + E5M2_RNINF + E5M2_DAZ_STOCHASTIC + E5M2_DAZ_RNE + E5M2_DAZ_RNAZ + E5M2_DAZ_RNTZ + BFLOAT16_STOCHASTIC + BFLOAT16_RNE + FLOAT16_RNE + FLOAT16_STOCHASTIC + FLOAT16_DAZ_RNE + E4M3_RNE + E4M3_STOCHASTIC +""" + +class FPEmuOp(Function): + @staticmethod + def forward(ctx, input, mode='NONE', inplace=False, scale=1.0, blocknorm=False, blocksize=1): + if mode == 'NONE' : + ctx.mark_dirty(input) + return input + else : + if input.is_sparse : + input = input.coalesce() + size = input.values().nelement() + if inplace == True: + outputs = fpemu_cpp.forward(input._values().contiguous(), mode, size, inplace, scale, blocknorm, blocksize) + output = input + else : + outputs = fpemu_cpp.forward(input._values().contiguous(), mode, size, inplace, scale, blocknorm, blocksize) + output = torch.sparse.FloatTensor(input.indices(), outputs[0], input.size()) + else : + input = input.cpu() + size = input.nelement() + outputs = fpemu_cpp.forward(input.contiguous(), mode, size, inplace, scale, blocknorm, blocksize) + output = outputs[0] + + if inplace == True: + ctx.mark_dirty(input) + + return output + + @staticmethod + def backward(ctx, output_grad): + # straight-through estimator + return output_grad, None, None, None, None diff --git a/FP8_Emulator/pytquant/cpp/fpemu_impl.cpp b/FP8_Emulator/pytquant/cpp/fpemu_impl.cpp new file mode 100644 index 00000000..262b07cd --- /dev/null +++ b/FP8_Emulator/pytquant/cpp/fpemu_impl.cpp @@ -0,0 +1,2009 @@ +/*----------------------------------------------------------------------------* + * Copyright (c) 2023, Intel Corporation - All rights reserved. + * This file is part of FP8-Emulation-Toolkit + * + * SPDX-License-Identifier: BSD-3-Clause + *----------------------------------------------------------------------------* + * Naveen Mellempudi (Intel Corporation) + *----------------------------------------------------------------------------*/ + +#include +#include +#include +#include + +enum ROUNDING_MODES { + ROUND_RTZ = 0, + ROUND_RNE = 1, + ROUND_STOCHASTIC = 2, + ROUND_RNAZ = 3, + ROUND_RNTZ = 4, + ROUND_PINF = 5, + ROUND_NINF = 6 +}; + +namespace { + + typedef union half_t { + unsigned short u; + at::Half f; + } __half_t; + + typedef union ufloat32 { + unsigned u; + float f; + } __float_t; + +/* Following implementation of xoroshiro128++ PRNG is borrowed from here: + http://prng.di.unimi.it/xoshiro128plusplus.c + main page: http://prng.di.unimi.it/ +*/ + static uint32_t s1_[4] = { 1387366120, 2798441831, 888998500 , 1099633400 }; + static uint32_t s2_[4] = { 2034269327, 2125325156, 1209715489, 1931656721 }; + static uint32_t s3_[4] = { 1555452618, 650181557 , 883695203 , 627677842 }; + static uint32_t s4_[4] = { 4195248041, 2146478152, 480059239 , 1468956197 }; + static uint32_t s5_[4] = { 1252084877, 500390994 , 977516591 , 1950666000 }; + static uint32_t s6_[4] = { 3936597502, 834151069 , 1477014702, 734008143 }; + static uint32_t s7_[4] = { 1983400973, 1164103095, 2110188261, 2019272068 }; + static uint32_t s8_[4] = { 1877096364, 2833629967, 4196320416, 1774181187 }; + static uint32_t s9_[4] = { 702309618 , 4077815558, 1512057936, 1868769368 }; + static uint32_t s10_[4] = + { 510001215 , 966559856 , 776583255 , 1475621065 }; + static uint32_t s11_[4] = + { 1271806057, 1881312534, 478635452 , 814821902 }; + static uint32_t s12_[4] = + { 733990058 , 1889991804, 1108257970, 1093480892 }; + static uint32_t s13_[4] = + { 4273743809, 4167473370, 558000409 , 1594848927 }; + static uint32_t s14_[4] = + { 444870959 , 1595722866, 1064124488, 3637102547 }; + static uint32_t s15_[4] = + { 703721499 , 3896407831, 1002360059, 1427395742 }; + static uint32_t s16_[4] = + { 1295231497, 1254972431, 1423497865, 861918264 }; + +/* seed pointer array */ + static uint32_t *sptr_[16] = { s1_, s2_, s3_, s4_, s5_, s6_, s7_, s8_, s9_, + s10_, s11_, s12_, s13_, s14_, s15_, s16_ + }; + + static inline uint32_t rotl (const uint32_t x, int k) { + return (x << k) | (x >> (32 - k)); + } + + uint32_t rand_xorshft128plus_scalar (uint32_t * ps) { + const uint32_t result_plus = ps[0] + ps[3]; + const uint32_t t = ps[1] << 9; + + ps[2] ^= ps[0]; + ps[3] ^= ps[1]; + ps[1] ^= ps[2]; + ps[0] ^= ps[3]; + + ps[2] ^= t; + + ps[3] = rotl (ps[3], 11); + + return result_plus; + } + + inline __m512i + _mm512_rndxorshft128plus_epi32 (uint32_t * vs0, uint32_t * vs1, + uint32_t * vs2, uint32_t * vs3) { + __m512i vrplus = _mm512_add_epi32 (_mm512_load_epi32 (vs0), _mm512_load_epi32 (vs3)); + __m512i vt = _mm512_sll_epi32 (_mm512_load_epi32 (vs1), _mm_set1_epi8 (9)); + _mm512_store_epi32 (vs2, _mm512_xor_epi32 (_mm512_load_epi32 (vs2), _mm512_load_epi32 (vs0))); + _mm512_store_epi32 (vs3, _mm512_xor_epi32 (_mm512_load_epi32 (vs3), _mm512_load_epi32 (vs1))); + _mm512_store_epi32 (vs1, _mm512_xor_epi32 (_mm512_load_epi32 (vs1), _mm512_load_epi32 (vs2))); + _mm512_store_epi32 (vs0, _mm512_xor_epi32 (_mm512_load_epi32 (vs0), _mm512_load_epi32 (vs3))); + _mm512_store_epi32 (vs2, _mm512_xor_epi32 (_mm512_load_epi32 (vs2), vt)); + + __m512i vl = _mm512_sll_epi32 (_mm512_load_epi32 (vs3), _mm_set1_epi8 (11)); + __m512i vr = _mm512_sra_epi32 (_mm512_load_epi32 (vs3), _mm_set1_epi8 (21)); + _mm512_store_epi32 (vs3, _mm512_or_epi32 (vl, vr)); + return vrplus; + } + + + float __double2float_rn (double inval) { + float out[4] = { 0 }; + __m128 vout = _mm_cvtpd_ps (_mm_set1_pd (inval)); + + _mm_store_ps (&out[0], vout); + return out[0]; + } + + unsigned short __float2half_rn (float inval) { + return _cvtss_sh (inval, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + } + + float __half2float (unsigned short h_val) { + return _cvtsh_ss (h_val); + } + + template < typename scalar_t > float __anyfloat2float_rn (scalar_t a_) { + float f_; + + if (std::is_same < scalar_t, double >::value) { + f_ = __double2float_rn (a_); + } else if (std::is_same < scalar_t, float >::value) { + f_ = a_; + } else if (std::is_same < scalar_t, at::Half >::value) { + f_ = __half2float ((at::Half) a_); + } + return f_; + } + + template < typename scalar_t > + void __float2anyfloat_rn (float f_, scalar_t * out) { + scalar_t a_; + + if (std::is_same < scalar_t, double >::value) { + a_ = (scalar_t) (f_); + } else if (std::is_same < scalar_t, float >::value) { + a_ = f_; + } else if (std::is_same < scalar_t, at::Half >::value) { + a_ = (at::Half) __float2half_rn (f_); + } + *out = a_; + } + + template < typename scalar_t > + unsigned short __anyfloat2half_rn (scalar_t f_) { + unsigned short h_; + + if (std::is_same < scalar_t, double >::value) { + h_ = __float2half_rn (__double2float_rn (f_)); + } else if (std::is_same < scalar_t, float >::value) { + h_ = __float2half_rn (f_); + } else if (std::is_same < scalar_t, at::Half >::value) { + unsigned short *ptrh_ = (unsigned short *) &f_; + h_ = *ptrh_; + } + return h_; + } + + template < typename scalar_t > + void __half2anyfloat (unsigned short h_, scalar_t * out, scalar_t scale=1.0) { + scalar_t f_; + + if (std::is_same < scalar_t, double >::value) { + f_ = (scalar_t) __half2float (h_); + } else if (std::is_same < scalar_t, float >::value) { + f_ = __half2float (h_); + } else if (std::is_same < scalar_t, at::Half >::value) { + f_ = *((at::Half *) & h_); + } + *out = scale*f_; + } + + template < typename scalar_t > + inline void reduce0 (scalar_t * g_data, float *g_odata, unsigned int n) { + float sum = 0.0, sumsq = 0.0; + +#pragma omp parallel for reduction(+: sum), reduction(+: sumsq) + for (unsigned int i = 0; i < n; i++) { + sum += __anyfloat2float_rn (g_data[i]); + sumsq += __anyfloat2float_rn (g_data[i]) * __anyfloat2float_rn (g_data[i]); + } + g_odata[0] = sum; + g_odata[1] = sumsq; + } + + template < typename scalar_t > + inline void absmax0 (scalar_t * g_data, float *g_odata, unsigned int n) { + float absmax = 0.0; + +#pragma omp parallel for reduction(max: absmax) + for (unsigned int i = 0; i < n; i++) { + absmax = fmaxf (absmax, fabsf (__anyfloat2float_rn (g_data[i]))); + } + g_odata[0] = absmax; + } + + + void cvt_fp32_bf16_rne_intrinsic (const float *__restrict__ in, float *out, + int size) { +#pragma omp parallel for + for (int i = 0; i < size; i += 16) { + const __m512i vnaninf = _mm512_set1_epi32 (0x7f800000), vrneadd = _mm512_set1_epi32 (0x00007fff); + const __m512i vfixup = _mm512_set1_epi32 (0x00000001), vfixupmask = _mm512_set1_epi32 (0x00010000); + const __m512i truncmask = _mm512_set1_epi32 (0xffff0000); + __m512 a = _mm512_loadu_ps (&in[i]); + const __m512i mm512_roundbf16rne_a_ = _mm512_castps_si512 (a); + const __mmask16 mm512_roundbf16rne_mask1_ = + _mm512_cmp_epi32_mask (_mm512_and_epi32 + (mm512_roundbf16rne_a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask16 mm512_roundbf16rne_mask2_ = + _mm512_cmp_epi32_mask (_mm512_and_epi32 + (mm512_roundbf16rne_a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + __m512i roundbf16rne = + _mm512_mask_add_epi32 (mm512_roundbf16rne_a_, + mm512_roundbf16rne_mask1_, + mm512_roundbf16rne_a_, + _mm512_mask_add_epi32 (vrneadd, + mm512_roundbf16rne_mask2_, vrneadd, vfixup)); + a = _mm512_castsi512_ps (_mm512_and_epi32 (roundbf16rne, truncmask)); + + _mm512_storeu_ps (&out[i], a); + } + } + + void cvt_fp32_bf16_stochastic_intrinsic (const float *__restrict__ in, + float *out, int size) { + uint32_t vs0[16] __attribute__ ((aligned (64))) = { + 1387366120, 279844183, 888998500, 1099633400, 1252084877, 500390994, + 977516591, 1950666000, 393659750, 834151069, 1477014702, 734008143, + 1983400973, 116410309, 2110188261, 2019272068}; + uint32_t vs1[16] __attribute__ ((aligned (64))) = { + 2034269327, 2125325156, 1209715489, 193165672, 187709636, 28336299, + 419632041, 1774181187, 702309618, 407781555, 1512057936, 1868769368, + 510001215, 966559856, 776583255, 147562106}; + uint32_t vs2[16] __attribute__ ((aligned (64))) = { + 1555452618, 650181557, 883695203, 62767784, 127180605, 1881312534, + 478635452, 814821902, 733990058, 1889991804, 1108257970, 1093480892, + 427374380, 416747337, 558000409, 1594848927}; + uint32_t vs3[16] __attribute__ ((aligned (64))) = { + 419524804, 2146478152, 480059239, 1468956197, 444870959, 1595722866, + 1064124488, 363710254, 703721499, 389640783, 1002360059, 1427395742, + 1295231497, 1254972431, 1423497865, 861918264}; + +#pragma omp parallel for firstprivate (vs0, vs1, vs2, vs3) + for (int i = 0; i < size; i += 16) { + const __m512i vnaninf = _mm512_set1_epi32 (0x7f800000), vrneadd = _mm512_set1_epi32 (0x00007fff); + const __m512i vfixup = _mm512_set1_epi32 (0x00000001), vfixupmask = _mm512_set1_epi32 (0x00010000); + const __m512i truncmask = _mm512_set1_epi32 (0xffff0000); + __m512i rnd512 = _mm512_rndxorshft128plus_epi32 (vs0, vs1, vs2, vs3); + __m256i rnbits = _mm512_extracti32x8_epi32 (rnd512, 0); + + __m512 a = _mm512_loadu_ps (&in[i]); + const __m512i mm512_roundbf16sr_a_ = _mm512_castps_si512 (a); + const __mmask16 mm512_roundbf16rne_mask1_ = + _mm512_cmp_epi32_mask (_mm512_and_epi32(mm512_roundbf16sr_a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask16 mm512_roundbf16rne_mask2_ = + _mm512_cmp_epi32_mask (_mm512_and_epi32(mm512_roundbf16sr_a_, vfixupmask), vfixupmask,_MM_CMPINT_EQ); + __m512i roundbf16sr = + _mm512_mask_add_epi32 (mm512_roundbf16sr_a_, + mm512_roundbf16rne_mask1_, + mm512_roundbf16sr_a_, + _mm512_cvtepu16_epi32 (rnbits)); + roundbf16sr = + _mm512_mask_add_epi32 (roundbf16sr, mm512_roundbf16rne_mask1_, + roundbf16sr, _mm512_mask_add_epi32 (vrneadd, + mm512_roundbf16rne_mask2_, + vrneadd, + vfixup)); + a = _mm512_castsi512_ps (_mm512_and_epi32 (roundbf16sr, truncmask)); + + _mm512_storeu_ps (&out[i], a); + } + } + + void cvt_fp32_bf16_scalar (const float *in, float *out, const int size, + int rmode) { + int lshift = 16; + int rshift = lshift - 3; /* shift to preserve rounding bits */ + unsigned int mask_mant = (unsigned int) (0xFFFFFFFF << lshift); + unsigned int mask_mant_grs = (unsigned int) (0xFFFFFFFF << rshift); + + /* mask to extract G(gaurd), R (round), S (sticky) bits */ + unsigned int lsbGRS = 0xF << rshift; + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + + if (rmode == ROUND_RNE) + rne_mask = 1; + if (rmode == ROUND_STOCHASTIC) + sr_mask = 1; + + for (int gid = 0; gid < size; gid++) { + __float_t uf; + + uf.f = in[gid]; + unsigned int mant_grs = (uf.u & mask_mant_grs); + + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = + (unsigned short) + rand_xorshft128plus_scalar (sptr_[(seed_index % 16)]); + /* stochastic rounding with 16-bit random number */ + uf.u += rand; + } + /* truncation */ + uf.u &= mask_mant; + + /* round to nearest even after truncation if rne_mask is enabled */ + unsigned int rmask_tie = ((mant_grs & lsbGRS) >> rshift); + unsigned int rmask = (rmask_tie & 0x7); + + uf.u += rne_mask * (((rmask > 0x4) || (rmask_tie == 0xC)) << lshift); + + //__float2anyfloat_rn(uf.f, &out[gid]); + out[gid] = uf.f; + } + } + + template < typename scalar_t > + void BFLOAT16_Kernel (const scalar_t * in, + scalar_t * out, const int size, int rmode) { + if ((size % 16) == 0) { + if (rmode == ROUND_STOCHASTIC) + cvt_fp32_bf16_stochastic_intrinsic (in, out, size); + else + cvt_fp32_bf16_rne_intrinsic (in, out, size); + } else { + int vec_size = ((int) (size / 16)) * 16; + + if (vec_size > 0) { + if (rmode == ROUND_STOCHASTIC) + cvt_fp32_bf16_stochastic_intrinsic (in, out, vec_size); + else + cvt_fp32_bf16_rne_intrinsic (in, out, vec_size); + } + cvt_fp32_bf16_scalar (&in[vec_size], &out[vec_size], size - vec_size, rmode); + } + } + + template < typename scalar_t > + void FLOAT16_Kernel (const scalar_t * in, + scalar_t * out, + const int size, int rmode, int no_denorm) { + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + + if (rmode == ROUND_RNE) + rne_mask = 1; + if (rmode == ROUND_STOCHASTIC) + sr_mask = 1; + +#pragma omp parallel for + for (int gid = 0; gid < size; gid++) { + if (rne_mask) { + __half_t h; + + h.u = __anyfloat2half_rn (in[gid]); + unsigned short not_denorm = ((((h.u & 0x7FFF) >> 10) & 0x1F) > 0); + unsigned short is_denorm = (not_denorm == 0) ? 1 : 0; + + h.u *= !(is_denorm && no_denorm); + __half2anyfloat (h.u, &out[gid]); + } else if (sr_mask) { + unsigned int fval = ((unsigned int *) in)[gid]; + int exp_h = (int) ((fval & 0x7f800000) >> 23) - 127; + unsigned int mantissa_h = (fval & 0x7FFFFF); + unsigned int sign_h = (fval & 0x80000000); + __half_t h; + + if (exp_h == 128) { + /* handle incoming INF and NaN */ + exp_h = 0x1F; + /* handle signalling NaN */ + if (mantissa_h && ((mantissa_h & 0x400000) == 0x0)) + mantissa_h |= 0x400000; + mantissa_h |= (exp_h << 23); + mantissa_h |= (sign_h >> 3); + h.u = (unsigned short) (mantissa_h >> 13); + } else if (exp_h >= 16) { + /* saturate to INF */ + exp_h = 0x1F; + mantissa_h = 0; + mantissa_h |= (exp_h << 23); + mantissa_h |= (sign_h >> 3); + h.u = (unsigned short) (mantissa_h >> 13); + } else { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned int rand = + (unsigned int) + rand_xorshft128plus_scalar (sptr_[(seed_index % 16)]); + if (exp_h < -14) { + mantissa_h += (rand & 0x00001FFF); + /* handle denormals */ + h.u = __anyfloat2half_rn (in[gid]); + } else { + exp_h += 15; + mantissa_h |= (exp_h << 23); + mantissa_h |= (sign_h >> 3); + mantissa_h += (rand & 0x00001FFF); + h.u = (unsigned short) (mantissa_h >> 13); + } + } + __half2anyfloat (h.u, &out[gid]); + } + } + } + + + __m256i _mm256_cvt2fp16_e5m2 (__m256i a, __m256i b) { + const __m512i vnaninf = _mm512_set1_epi16 (0x7c00), vrneadd = + _mm512_set1_epi16 (0x007f); + const __m512i vfixup = _mm512_set1_epi16 (0x0001), vfixupmask = + _mm512_set1_epi16 (0x0100); + /* b: lower half, a : upper half */ + const __m512i a_ = + _mm512_inserti64x4 (_mm512_inserti64x4 (_mm512_setzero_si512 (), b, 0), a, 1); + const __mmask32 maska1_ = + _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, vnaninf), vnaninf,_MM_CMPINT_NE); + const __mmask32 maska2_ = + _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, vfixupmask), vfixupmask,_MM_CMPINT_EQ); + __m512i a_rne_ = _mm512_mask_add_epi16 (a_, maska1_, a_, + _mm512_mask_add_epi16 (vrneadd, maska2_, vrneadd,vfixup)); + return _mm512_cvtepi16_epi8 (_mm512_srli_epi16 (a_rne_, 8)); + } + + static inline __m512i _mm512_cvte5m2_fp16 (__m256i a) { + return _mm512_slli_epi16 (_mm512_cvtepi8_epi16 (a), 8); + } + + static inline void cvt_fp16_e5m2_rne_intrinsic (const short *__restrict__ in, + unsigned char *out, int size) { +#pragma omp parallel for + for (int i = 0; i < size; i += 32) { + __m256i bh_ = _mm256_lddqu_si256 ((__m256i *) & in[i]); + __m256i ah_ = _mm256_lddqu_si256 ((__m256i *) & in[i + 16]); + + _mm256_storeu_si256 ((__m256i *) & out[i], + _mm256_cvt2fp16_e5m2 (ah_, bh_)); + } + } + + __m256i _mm256_cvt2fp16_e5m2_noINF (__m256i a, __m256i b) { + const __m512i vnaninf = _mm512_set1_epi16 (0x7c00); + const __m512i vrneadd = _mm512_set1_epi16 (0x007f); + const __m512i vfixup = _mm512_set1_epi16 (0x0001); + const __m512i vfixupmask = _mm512_set1_epi16 (0x0100); + /* use a non-standard exponent offset = 16, */ + const __m512i vExp_fp16 = _mm512_set1_epi16 (0x000F); + const __m512i vExp_e5m2 = _mm512_set1_epi16 (0x0010); + const __m512i vsMant = _mm512_set1_epi16 (0x83FF); + /* Exponent Offset = 16, reclaim inf/NaN */ + const __m512i vsatuval = _mm512_set1_epi16 (0x7F00);/* 2^15*1.11 a.k.a 57344.0, largest value */ + const __m512i vinfval = _mm512_set1_epi16 (0x8000); /* -0.0 as INF */ + const __m512i a_ = _mm512_inserti64x4 (_mm512_inserti64x4 (_mm512_setzero_si512 (), b, 0), a, 1); + const __mmask32 maska1_ = _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask32 maska2_ = _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + const __mmask32 maska3_ = _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, _mm512_set1_epi16(0x7FFF)), vsatuval, _MM_CMPINT_NLE); + + __m512i vExp_ = _mm512_sub_epi16 (_mm512_srli_epi16 (_mm512_and_si512 (a_, vnaninf), 10), vExp_fp16); + vExp_ = _mm512_slli_epi16 (_mm512_add_epi16 (vExp_, vExp_e5m2), 10); + __m512i a_rne_ = _mm512_or_si512 (vExp_, _mm512_and_si512 (a_, vsMant)); + + a_rne_ = _mm512_mask_add_epi16 (a_rne_, maska1_, a_rne_, + _mm512_mask_add_epi16 (vrneadd, maska2_, vrneadd, vfixup)); + a_rne_ = _mm512_mask_mov_epi16 (a_rne_, maska3_, _mm512_or_si512(_mm512_and_si512(a_rne_, vinfval), vsatuval)); + a_rne_ = _mm512_mask_mov_epi16 (a_rne_, ~maska1_, vinfval); + return _mm512_cvtepi16_epi8 (_mm512_srli_epi16 (a_rne_, 8)); + } + + static inline __m256i _mm512_cvt2fp32_e5m2_noINF (__m512 a, __m512 b) { + __m256i ah_ = _mm512_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i bh_ = _mm512_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + return _mm256_cvt2fp16_e5m2_noINF (ah_, bh_); + } + + __m512i _mm512_cvte5m2_noinf_fp16 (__m256i a) { + const __m512i vExp_fp16 = _mm512_set1_epi16 (0x000F); + const __m512i vExp_e5m2 = _mm512_set1_epi16 (0x0010); + const __m512i vsMant = _mm512_set1_epi16 (0x83FF); + const __m512i vnaninf = _mm512_set1_epi16 (0x8000); /* -0.0 as INF */ + const __m512i vinfval = _mm512_set1_epi16 (0x7c00); + __m512i a_ = _mm512_slli_epi16 (_mm512_cvtepi8_epi16 (a), 8); + const __mmask32 mask1_ = _mm512_cmp_epi16_mask (a_, vnaninf, _MM_CMPINT_EQ); + __m512i vExp_ = _mm512_sub_epi16 (_mm512_srli_epi16 (_mm512_and_si512 (a_, vinfval), 10), vExp_e5m2); + vExp_ = _mm512_slli_epi16 (_mm512_add_epi16 (vExp_, vExp_fp16), 10); + a_ = _mm512_or_si512 (vExp_, _mm512_and_si512 (a_, vsMant)); + return _mm512_mask_mov_epi16 (a_, mask1_, vinfval); + } + + static inline void cvt_fp16_e5m2_noINF_rne_intrinsic (const short *__restrict__ in, + unsigned char *out, int size) { +#pragma omp parallel for + for (int i = 0; i < size; i += 32) { + __m256i bh_ = _mm256_lddqu_si256 ((__m256i *) & in[i]); + __m256i ah_ = _mm256_lddqu_si256 ((__m256i *) & in[i + 16]); + + _mm256_storeu_si256 ((__m256i *) & out[i], _mm256_cvt2fp16_e5m2 (ah_, bh_)); + } + } + + void cvt_fp32_e5m2_noinf_rne_intrinsic (const float *__restrict__ in, + float *out, int size, float scale) { +#pragma omp parallel for + for (int i = 0; i < size; i += 32) { + __m512 b = _mm512_loadu_ps (&in[i]); + __m512 a = _mm512_loadu_ps (&in[i + 16]); + __m256i ah_ = _mm512_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i bh_ = _mm512_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m512i a_rne_ = _mm512_cvte5m2_noinf_fp16 (_mm256_cvt2fp16_e5m2_noINF (ah_, bh_)); + bh_ = _mm512_extracti64x4_epi64 (a_rne_, 0); + ah_ = _mm512_extracti64x4_epi64 (a_rne_, 1); + b = _mm512_cvtph_ps (bh_); + a = _mm512_cvtph_ps (ah_); + _mm512_storeu_ps (&out[i], b); + _mm512_storeu_ps (&out[i + 16], a); + } + } + + static inline void cvt_fp32_e5m2_flex_intrinsic (const float *__restrict__ in, float *out, + int size, float scale) { +#pragma omp parallel for + for (int i = 0; i < size; i += 16) { + const __m512i vnaninf = _mm512_set1_epi32 (0x7f800000); + const __m512i vrneadd = _mm512_set1_epi32 (0x000fffff); + const __m512i vfixup = _mm512_set1_epi32 (0x00000001); + const __m512i vfixupmask = _mm512_set1_epi32 (0x00200000); + const __m512i vexpf32 = _mm512_set1_epi32 (0x0000007f); + const __m512i vmexp_e5m2 = _mm512_set1_epi32 (0x0000000f); + const __m512i vsign = _mm512_set1_epi32 (0x80000000); + const __m512i vmant = _mm512_set1_epi32 (0x007fffff); + const __m512i vmin_e5m2 = _mm512_set1_epi32 (0x37800000); + /*const __m512i vmax_e5m2 = _mm512_set1_epi32 (0x47600000);*/ + const __m512i vdnorm = _mm512_set1_epi32 (0x38800000); + + __m512 a = _mm512_loadu_ps (&in[i]); + const __m512i a_ = _mm512_castps_si512 (a); + const __mmask16 naninf_mask_ = _mm512_cmp_epi32_mask (_mm512_and_epi32 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask16 rneadd_mask_ = _mm512_cmp_epi32_mask (_mm512_and_epi32 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + const __mmask16 zflush_mask_ = _mm512_cmp_epi32_mask (_mm512_and_epi32 (a_, vnaninf), vmin_e5m2, _MM_CMPINT_LT); + const __mmask16 denorm_mask_ = _mm512_cmp_epi32_mask (_mm512_and_epi32 (a_, vnaninf), vdnorm, _MM_CMPINT_LT); + + __m512i a_sign = _mm512_and_epi32 (a_, vsign); + __m512i a_rne = _mm512_mask_add_epi32 (a_, naninf_mask_, a_, + _mm512_mask_add_epi32 (vrneadd, rneadd_mask_, vrneadd, vfixup)); + __m512i a_exp = _mm512_srai_epi32 (_mm512_and_epi32 (a_rne, vnaninf), 23); + a_rne = _mm512_and_epi32 (a_rne, vmant); + a_exp = _mm512_sub_epi32 (a_exp, vexpf32); + + __m512i vminexp = _mm512_sub_epi32 (_mm512_setzero_epi32 (), vmexp_e5m2); + __m512i shft_ = _mm512_sub_epi32 (vminexp, a_exp); + + /*const __mmask16 ovflow_mask_ = + _mm512_cmp_epi32_mask (a_exp, vmexp_e5m2, _MM_CMPINT_NLT);*/ + a_exp = _mm512_add_epi32 (a_exp, vmexp_e5m2); + + __m512i vrshft = _mm512_set1_epi32 (21); + + vrshft = _mm512_mask_add_epi32 (vrshft, denorm_mask_, vrshft, shft_); + __m512i vlshft = _mm512_set1_epi32 (8); + + vlshft = _mm512_mask_add_epi32 (vlshft, denorm_mask_, vlshft, shft_); + + a_rne = _mm512_sllv_epi32 (_mm512_srav_epi32 (a_rne, vrshft), vlshft); + + a_exp = _mm512_slli_epi32 (a_exp, 10); + a_rne = _mm512_or_epi32 (a_rne, a_exp); + a_rne = _mm512_or_epi32 (a_rne, _mm512_srai_epi32 (a_sign, 16)); + a_rne = _mm512_mask_set1_epi32 (a_rne, zflush_mask_, 0); + + __m256i a_rne_16 = _mm512_cvtepi32_epi16 (a_rne); + + a = _mm512_cvtph_ps (a_rne_16); + _mm512_storeu_ps (&out[i], a); + } + } + + void cvt_fp32_e5m2_rne_intrinsic (const float *__restrict__ in, float *out, + int size, float scale) { + +#pragma omp parallel for + for (int i = 0; i < size; i += 32) { + const __m512i vnaninf = _mm512_set1_epi16 (0x7c00); + const __m512i vrneadd = _mm512_set1_epi16 (0x007f); + const __m512i vfixup = _mm512_set1_epi16 (0x0001); + const __m512i vfixupmask = _mm512_set1_epi16 (0x0100); + + __m512 s_ = _mm512_set1_ps (scale); + __m512 sr_ = _mm512_set1_ps (1.0 / scale); + __m512 b = _mm512_loadu_ps (&in[i]); + __m512 a = _mm512_loadu_ps (&in[i + 16]); + + b = _mm512_mul_ps (b, s_); + a = _mm512_mul_ps (a, s_); + + __m256i ah_ = _mm512_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i bh_ = _mm512_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + const __m512i a_ = _mm512_inserti64x4 (_mm512_inserti64x4 (_mm512_setzero_si512 (), bh_, 0), ah_, 1); + const __mmask32 maska1_ = _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask32 maska2_ = _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + __m512i a_rne_ = _mm512_mask_add_epi16 (a_, maska1_, a_,_mm512_mask_add_epi16 (vrneadd, maska2_, vrneadd, vfixup)); + a_rne_ = _mm512_slli_epi16 (_mm512_srli_epi16 (a_rne_, 8), 8); + + bh_ = _mm512_extracti64x4_epi64 (a_rne_, 0); + ah_ = _mm512_extracti64x4_epi64 (a_rne_, 1); + b = _mm512_cvtph_ps (bh_); + a = _mm512_cvtph_ps (ah_); + + _mm512_storeu_ps (&out[i], _mm512_mul_ps (b, sr_)); + _mm512_storeu_ps (&out[i + 16], _mm512_mul_ps (a, sr_)); + } + } + + void cvt_fp32_e5m2_stochastic_intrinsic (const float *__restrict__ in, + float *out, int size, float scale) { + uint32_t vs0[16] __attribute__ ((aligned (64))) = { + 1387366120, 279844183, 888998500, 1099633400, 1252084877, 500390994, + 977516591, 1950666000, 393659750, 834151069, 1477014702, 734008143, + 1983400973, 116410309, 2110188261, 2019272068}; + uint32_t vs1[16] __attribute__ ((aligned (64))) = { + 2034269327, 2125325156, 1209715489, 193165672, 187709636, 28336299, + 419632041, 1774181187, 702309618, 407781555, 1512057936, 1868769368, + 510001215, 966559856, 776583255, 147562106}; + uint32_t vs2[16] __attribute__ ((aligned (64))) = { + 1555452618, 650181557, 883695203, 62767784, 127180605, 1881312534, + 478635452, 814821902, 733990058, 1889991804, 1108257970, 1093480892, + 427374380, 416747337, 558000409, 1594848927}; + uint32_t vs3[16] __attribute__ ((aligned (64))) = { + 419524804, 2146478152, 480059239, 1468956197, 444870959, 1595722866, + 1064124488, 363710254, 703721499, 389640783, 1002360059, 1427395742, + 1295231497, 1254972431, 1423497865, 861918264}; + +#pragma omp parallel for firstprivate (vs0, vs1, vs2, vs3) + for (int i = 0; i < size; i += 32) { + const __m512i vnaninf = _mm512_set1_epi16 (0x7c00); + const __m512i vfixup = _mm512_set1_epi16 (0x0001); + const __m512i vfixupmask = _mm512_set1_epi16 (0x0100); + const __m512i vrneadd = _mm512_set1_epi16 (0x007f); + const __m512i vdenorm = _mm512_set1_epi16 (0x03ff); + const __m512i vexmant = _mm512_set1_epi16 (0x7fff); + + __m512i rnd512 = _mm512_rndxorshft128plus_epi32 (vs0, vs1, vs2, vs3); + __m256i rnbits = _mm512_extracti32x8_epi32 (rnd512, 0); + + __m512 s_ = _mm512_set1_ps (scale); + __m512 sr_ = _mm512_set1_ps (1.0 / scale); + + __m512 b = _mm512_loadu_ps (&in[i]); + __m512 a = _mm512_loadu_ps (&in[i + 16]); + + b = _mm512_mul_ps (b, s_); + a = _mm512_mul_ps (a, s_); + + __m256i ah_ = _mm512_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i bh_ = _mm512_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + const __m512i a_ = _mm512_inserti64x4 (_mm512_inserti64x4 (_mm512_setzero_si512 (), bh_, 0), ah_, 1); + const __mmask32 maska1_ = _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask32 maska2_ = _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + const __mmask32 maska4_ = _mm512_cmp_epi16_mask (_mm512_and_si512 (a_, vexmant), vdenorm, _MM_CMPINT_LE); + __m512i a_sr_ = _mm512_mask_add_epi16 (a_, (maska1_ & ~maska4_), a_, _mm512_cvtepu8_epi16 (rnbits)); + a_sr_ = _mm512_mask_add_epi16 (a_sr_, maska4_, a_sr_, _mm512_mask_add_epi16 (vrneadd, maska2_, vrneadd, vfixup)); + a_sr_ = _mm512_slli_epi16 (_mm512_srli_epi16 (a_sr_, 8), 8); + + bh_ = _mm512_extracti64x4_epi64 (a_sr_, 0); + ah_ = _mm512_extracti64x4_epi64 (a_sr_, 1); + b = _mm512_cvtph_ps (bh_); + a = _mm512_cvtph_ps (ah_); + + _mm512_storeu_ps (&out[i], _mm512_mul_ps (b, sr_)); + _mm512_storeu_ps (&out[i + 16], _mm512_mul_ps (a, sr_)); + } + } + + void cvt_fp32_e5m2_scalar (const float *__restrict__ in, float *out, + int size, float scale, int rmode) { + int non_mant_bits = 5 /*exp_bits */ + 1; /* exponent + sign */ + int lshift = 10 - (8 /*mbits */ - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x00FF; + unsigned short rne_tie = 0x0180; + + float scale_reciprocal = 1.0 / scale; + + for (int gid = 0; gid < size; gid++) { + __half_t h; + float inval = scale * in[gid]; + + h.u = __anyfloat2half_rn (inval); + + unsigned short can_round = ((h.u & 0x7F00) <= 0x7B00) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + /* nearest rounding masks */ + unsigned short rnmask = (h.u & grs_bitmask); + unsigned short rnmask_tie = (h.u & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + rand_xorshft128plus_scalar (sptr_[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + h.u += can_round * is_normal * (rand & 0xFF); + /* stochastic round: denormals --> rne rounding */ + h.u += can_round * is_denorm * + (((rnmask > 0x0080) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + h.u += can_round * rne_mask * + (((rnmask > 0x0080) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + h.u += can_round * rnaz_mask * ((rnmask >= 0x0080) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + h.u += can_round * rntz_mask * ((rnmask > 0x0080) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + h.u += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0080) << lshift); + /* round to -INF, if rminf_mask is enabled */ + h.u += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0080) << lshift); + } + } + /* truncation */ + h.u = (h.u & mask_mant); + float f_; + __half2anyfloat (h.u, &f_); + out[gid] = f_ * scale_reciprocal; + } + } + + template < typename scalar_t > + void E5M2_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, int block_size, int rmode) { + float scale = in_scale; + float fmax = std::numeric_limits::max(); + if (scale > fmax) { + fprintf(stderr,"Error: Invalid scale factor : %.2e, make sure the scale is not larger than : %.2e\n", scale, fmax); + exit(1); + } + + if (block_norm == true) { + int nblocks = (size + (block_size - 1)) / block_size; + +#pragma omp parallel for + for (int b = 0; b < nblocks; b++) { + int start_index = (b * block_size); + /* handle the last block */ + if (start_index + block_size > size) + block_size = (size - start_index); + + float maxval = 0.0; + +#pragma omp parallel for reduction (max:maxval) + for (int gid = start_index; gid < start_index + block_size; gid++) { + maxval = (maxval < fabs (in[gid])) ? fabs (in[gid]) : maxval; + } + __float_t f; + + f.f = maxval; + f.u = (f.u & 0x7F800000); + scale = 2.0 * f.f; + scale /= 16384.0; + + if ((block_size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e5m2_stochastic_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } else { + cvt_fp32_e5m2_rne_intrinsic (&in[start_index], &out[start_index], block_size, scale); + //cvt_fp32_e5m2_noinf_rne_intrinsic (&in[start_index], &out[start_index], block_size, scale); + //cvt_fp32_e5m2_flex_intrinsic(&in[start_index], &out[start_index], block_size, scale); + } + } else { + cvt_fp32_e5m2_scalar (&in[start_index], &out[start_index], block_size, scale, rmode); + } + } + } else { + if ((size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e5m2_stochastic_intrinsic (in, out, size, scale); + } else { + cvt_fp32_e5m2_rne_intrinsic (in, out, size, scale); + //cvt_fp32_e5m2_noinf_rne_intrinsic (in, out, size, scale); + } + } else { + int vec_size = ((int) (size / 32)) * 32; + + if (vec_size > 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e5m2_stochastic_intrinsic (in, out, vec_size, scale); + } else { + cvt_fp32_e5m2_rne_intrinsic (in, out, vec_size, scale); + //cvt_fp32_e5m2_noinf_rne_intrinsic (in, out, vec_size, scale); + //cvt_fp32_e5m2_flex_intrinsic(in, out, vec_size, scale); + } + } + cvt_fp32_e5m2_scalar (&in[vec_size], &out[vec_size], size - vec_size, scale, rmode); + } + } + } + + template < typename scalar_t > + void E5M2_DAZ_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, const scalar_t scale, int rmode) { + int non_mant_bits = 5 /*exp_bits */ + 1; /* exponent + sign */ + int lshift = 10 - (8 /*mbits */ - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x00FF; + unsigned short rne_tie = 0x0180; + + scalar_t scale_reciprocal = 1.0/scale; + +#pragma omp parallel for + for (int gid = 0; gid < size; gid++) { + __half_t h; + float inval = in[gid] * scale; + + h.u = __anyfloat2half_rn (inval); + /* values above 57344.0, saturate them to +- Infinity */ + + unsigned short can_round = ((h.u & 0x7F00) <= 0x7B00) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + /* nearest rounding masks */ + unsigned short rnmask = (h.u & grs_bitmask); + unsigned short rnmask_tie = (h.u & rne_tie); + + if (is_naninf == 0 && is_normal) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) rand_xorshft128plus_scalar (sptr_[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + h.u += can_round * (rand & 0xFF); + } else { + /* round to nearest even, if rne_mask is enabled */ + h.u += can_round * rne_mask * + (((rnmask > 0x0080) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + h.u += can_round * rnaz_mask * ((rnmask >= 0x0080) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + h.u += can_round * rntz_mask * ((rnmask > 0x0080) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + h.u += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0080) << lshift); + /* round to -INF, if rminf_mask is enabled */ + h.u += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0080) << lshift); + } + } else if (is_denorm) { + /* Flush Denormal */ + h.u = 0; + } + /* truncation */ + h.u = (h.u & mask_mant); + __half2anyfloat (h.u, &out[gid], scale_reciprocal); + } + } + + void cvt_fp32_e4m3_rne_intrinsic (const float *__restrict__ in, float *out, + int size, float scale) { + const __m256i vnaninf = _mm256_set1_epi16 (0x7c00); + const __m256i vrneadd = _mm256_set1_epi16 (0x003f); + const __m256i vfixup = _mm256_set1_epi16 (0x0001); + const __m256i vfixupmask = _mm256_set1_epi16 (0x0080); + const __m256i vzero = _mm256_set1_epi16 (0x0000); + const __m256i vsign = _mm256_set1_epi16 (0x8000); + const __m256i vsatuval = _mm256_set1_epi16 (0x5F00);/* 2^8*1.110 a.k.a 448.0, largest value */ + const __m256i vflush = _mm256_set1_epi16 (0x1800);/* 2^-9, smallest denormal */ + const __m256i vxdnorm = _mm256_set1_epi16 (0x2400);/* 2^-6 smallest normal */ + + for (int i = 0; i < size; i += 16) { + __m256 s_ = _mm256_set1_ps (scale); + __m256 sr_ = _mm256_set1_ps (1.0 / scale); + __m256 b = _mm256_loadu_ps (&in[i]); + __m256 a = _mm256_loadu_ps (&in[i + 8]); + + b = _mm256_mul_ps (b, s_); + a = _mm256_mul_ps (a, s_); + + __m128i ah_ = _mm256_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = _mm256_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i a_ = _mm256_inserti32x4 (_mm256_inserti32x4 (_mm256_setzero_si256 (), bh_, 0), ah_, 1); + const __mmask16 maska1_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask16 maska2_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + const __mmask16 maska3_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, _mm256_set1_epi16(0x7FFF)), vsatuval, _MM_CMPINT_NLT); + const __mmask16 maska4_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vflush, _MM_CMPINT_LT); + const __mmask16 maska5_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vxdnorm, _MM_CMPINT_LT); + /* Handle denormals, shift the mantissa before rounding */ + __m256i v_shft_ = _mm256_sub_epi16 ( + _mm256_srli_epi16 (vxdnorm, 10), + _mm256_srli_epi16 (_mm256_and_si256 (a_, vnaninf), 10)); + v_shft_ = _mm256_mask_mov_epi16 (vzero, maska5_, v_shft_); + a_ = _mm256_mask_srlv_epi16 (a_, maska5_, a_, v_shft_); + a_ = _mm256_mask_sllv_epi16 (a_, maska5_, a_, v_shft_); + + __m256i a_rne_ = _mm256_mask_add_epi16 (a_, maska1_, a_, + _mm256_mask_add_epi16 (vrneadd, maska2_, vrneadd, vfixup)); + /* saturating values beyond +/-MAX */ + a_rne_ = _mm256_mask_mov_epi16 (a_rne_, maska3_, _mm256_or_si256(_mm256_and_si256(a_rne_, vsign), vsatuval)); + a_rne_ = _mm256_mask_mov_epi16 (a_rne_, maska4_, _mm256_and_si256(a_rne_, vsign)); + a_rne_ = _mm256_slli_epi16 (_mm256_srli_epi16 (a_rne_, 7), 7); + + bh_ = _mm256_extracti32x4_epi32 (a_rne_, 0); + ah_ = _mm256_extracti32x4_epi32 (a_rne_, 1); + b = _mm256_cvtph_ps (bh_); + a = _mm256_cvtph_ps (ah_); + _mm256_storeu_ps (&out[i], _mm256_mul_ps (b, sr_)); + _mm256_storeu_ps (&out[i + 8], _mm256_mul_ps (a, sr_)); + } + } + + void cvt_fp32_e4m3_stochastic_intrinsic (const float *__restrict__ in, + float *out, int size, float scale) { + const __m256i vnaninf = _mm256_set1_epi16 (0x7c00); + const __m256i vrneadd = _mm256_set1_epi16 (0x003f); + const __m256i vfixup = _mm256_set1_epi16 (0x0001); + const __m256i vfixupmask = _mm256_set1_epi16 (0x0080); + const __m256i vzero = _mm256_set1_epi16 (0x0000); + const __m256i vsign = _mm256_set1_epi16 (0x8000); + const __m256i vsatuval = _mm256_set1_epi16 (0x5F00);/* 2^8*1.110 a.k.a 448.0, largest value */ + const __m256i vflush = _mm256_set1_epi16 (0x1800);/* 2^-9, smallest denormal */ + const __m256i vxdnorm = _mm256_set1_epi16 (0x2400);/* 2^-6 smallest normal */ + + for (int i = 0; i < size; i += 16) { + unsigned int rndbuf[16]; + /* generate 128 random bits */ + for (int r = 0; r < 8; r++) { + rndbuf[r] = (unsigned int) rand_xorshft128plus_scalar (sptr_[r]); + } + __m128i rnbits = _mm_load_si128 ((const __m128i *) &rndbuf[0]); + + __m256 s_ = _mm256_set1_ps (scale); + __m256 sr_ = _mm256_set1_ps (1.0 / scale); + __m256 b = _mm256_loadu_ps (&in[i]); + __m256 a = _mm256_loadu_ps (&in[i + 8]); + + b = _mm256_mul_ps (b, s_); + a = _mm256_mul_ps (a, s_); + + __m128i ah_ = _mm256_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = _mm256_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i a_ = _mm256_inserti32x4 (_mm256_inserti32x4 (_mm256_setzero_si256 (), bh_, 0), ah_, 1); // might be avx-512 + const __mmask16 maska1_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask16 maska2_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + const __mmask16 maska3_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, _mm256_set1_epi16(0x7FFF)), vsatuval, _MM_CMPINT_NLT); + const __mmask16 maska4_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vflush, _MM_CMPINT_LT); + const __mmask16 maska5_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vxdnorm, _MM_CMPINT_LT); + /* Handle denormals, shift the mantissa before rounding */ + __m256i v_shft_ = _mm256_sub_epi16 ( + _mm256_srli_epi16 (vxdnorm, 10), + _mm256_srli_epi16 (_mm256_and_si256 (a_, vnaninf), 10)); + v_shft_ = _mm256_mask_mov_epi16 (vzero, maska5_, v_shft_); // avx-512 + a_ = _mm256_mask_srlv_epi16 (a_, maska5_, a_, v_shft_); // avx-512 + a_ = _mm256_mask_sllv_epi16 (a_, maska5_, a_, v_shft_); + + + __m256i a_sr_ = _mm256_mask_add_epi16 (a_, (maska1_ & ~maska5_), a_, + _mm256_srli_epi16 (_mm256_cvtepu8_epi16 (rnbits), 1)); + a_sr_ = _mm256_mask_add_epi16 (a_sr_, maska5_, a_sr_, + _mm256_mask_add_epi16 (vrneadd, maska2_, vrneadd, vfixup)); + + a_sr_ = _mm256_mask_mov_epi16 (a_sr_, maska3_, _mm256_or_si256(_mm256_and_si256(a_sr_, vsign), vsatuval)); + a_sr_ = _mm256_mask_mov_epi16 (a_sr_, maska4_, _mm256_and_si256(a_sr_, vsign)); + a_sr_ = _mm256_slli_epi16 (_mm256_srli_epi16 (a_sr_, 7), 7); + + bh_ = _mm256_extracti32x4_epi32 (a_sr_, 0); + ah_ = _mm256_extracti32x4_epi32 (a_sr_, 1); + b = _mm256_cvtph_ps (bh_); + a = _mm256_cvtph_ps (ah_); + _mm256_storeu_ps (&out[i], _mm256_mul_ps (b, sr_)); + _mm256_storeu_ps (&out[i + 8], _mm256_mul_ps (a, sr_)); + } + } + + void cvt_fp32_e4m3_scalar (const float *__restrict__ in, float *out, + int size, float scale, int rmode) { + int non_mant_bits = 4 /*exp_bits */ + 1; /* exponent + sign */ + int lshift = 10 - (8 /*mbits */ - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x007F; + unsigned short rne_tie = 0x00C0; + float scale_reciprocal = 1.0 / scale; + +#pragma omp parallel for + for (int gid = 0; gid < size; gid++) { + __half_t h; + float inval = scale * in[gid]; + + h.u = __anyfloat2half_rn (inval); + short exp_h = (short) ((h.u & 0x7C00) >> 10) - 15; + short sign_h = (h.u & 0x8000); + short mantissa_h = (h.u & 0x03FF); + + unsigned short can_round = ((h.u & 0x7FFF) < 0x5F00) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + int dshift = 0; + + if (exp_h > 8 || (can_round == 0)) { + /* Software : saturate values above to +/-448.0 to +/-448.0 */ + mantissa_h = 0x0300; + exp_h = 8; + can_round = 0; + } else if (exp_h < -9) { + /* flush values below 1-4-3 subnormal range to zero */ + exp_h = -15; + mantissa_h = 0; + } else if (exp_h < -6) { + dshift = (-6 - exp_h); + /* handle denormals */ + mantissa_h = mantissa_h >> dshift; + mantissa_h <<= dshift; + } + /* nearest rounding masks */ + unsigned short rnmask = (mantissa_h & grs_bitmask); + unsigned short rnmask_tie = (mantissa_h & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + rand_xorshft128plus_scalar (sptr_[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + mantissa_h += can_round * is_normal * (rand & 0x7F); + /* stochastic round: denormals --> rne rounding */ + mantissa_h += can_round * is_denorm * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + mantissa_h += can_round * rne_mask * (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + mantissa_h += can_round * rnaz_mask * ((rnmask >= 0x0040) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + mantissa_h += can_round * rntz_mask * ((rnmask > 0x0040) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + mantissa_h += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0040) << lshift); + /* round to -INF, if rminf_mask is enabled */ + mantissa_h += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0040) << lshift); + } + } + /* truncation */ + mantissa_h &= mask_mant; + mantissa_h += ((exp_h + 15) << 10); + mantissa_h |= sign_h; + h.u = mantissa_h; + float f; + + __half2anyfloat (h.u, &f); + out[gid] = (f * scale_reciprocal); + } + } + + template < typename scalar_t > + void E4M3_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, + int block_size, int rmode) { + float scale = in_scale; + float fmax = std::numeric_limits::max(); + if (scale > fmax) { + fprintf(stderr,"Error: Invalid scale factor : %.2e, make sure the scale is not larger than : %.2e\n", scale, fmax); + exit(1); + } + + if (block_norm == true) { + int nblocks = (size + (block_size - 1)) / block_size; + +#pragma omp parallel for + for (int b = 0; b < nblocks; b++) { + int start_index = (b * block_size); + + /* handle the last block */ + if (start_index + block_size > size) + block_size = (size - start_index); + + float maxval = 0.0; + +#pragma omp parallel for reduction (max:maxval) + for (int gid = start_index; gid < start_index + block_size; gid++) { + maxval = (maxval < fabs (in[gid])) ? fabs (in[gid]) : maxval; + } + __float_t f; + + f.f = maxval; + f.u = (f.u & 0x7F800000); + scale = 2.0 * f.f; + scale /= 8.0; + + if ((block_size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e4m3_stochastic_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } else { + cvt_fp32_e4m3_rne_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } + } else { + cvt_fp32_e4m3_scalar (&in[start_index], &out[start_index], block_size, scale, rmode); + } + } + } else { + if ((size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e4m3_stochastic_intrinsic (in, out, size, scale); + } else { + cvt_fp32_e4m3_rne_intrinsic (in, out, size, scale); + } + } else { + int vec_size = ((int) (size / 32)) * 32; + + if (vec_size > 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e4m3_stochastic_intrinsic (in, out, vec_size, scale); + } else { + cvt_fp32_e4m3_rne_intrinsic (in, out, vec_size, scale); + } + } + cvt_fp32_e4m3_scalar (&in[vec_size], &out[vec_size], size - vec_size, scale, rmode); + } + } + } + + void cvt_fp32_e4m3_ieee_rne_intrinsic (const float *__restrict__ in, float *out, + int size, float scale) { + const __m256i vnaninf = _mm256_set1_epi16 (0x7c00); + const __m256i vrneadd = _mm256_set1_epi16 (0x003f); + const __m256i vfixup = _mm256_set1_epi16 (0x0001); + const __m256i vfixupmask = _mm256_set1_epi16 (0x0080); + const __m256i vzero = _mm256_set1_epi16 (0x0000); + const __m256i vsign = _mm256_set1_epi16 (0x8000); + const __m256i vsatuval = _mm256_set1_epi16 (0x5B80);/* 2^7*1.111 a.k.a 240.0, largest value */ + const __m256i vflush = _mm256_set1_epi16 (0x1800);/* 2^-9, smallest denormal */ + const __m256i vxdnorm = _mm256_set1_epi16 (0x2400);/* 2^-6 smallest normal */ + + for (int i = 0; i < size; i += 16) { + __m256 s_ = _mm256_set1_ps (scale); + __m256 sr_ = _mm256_set1_ps (1.0 / scale); + __m256 b = _mm256_loadu_ps (&in[i]); + __m256 a = _mm256_loadu_ps (&in[i + 8]); + + b = _mm256_mul_ps (b, s_); + a = _mm256_mul_ps (a, s_); + + __m128i ah_ = _mm256_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = _mm256_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i a_ = _mm256_inserti32x4 (_mm256_inserti32x4 (_mm256_setzero_si256 (), bh_, 0), ah_, 1); + const __mmask16 maska1_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask16 maska2_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + const __mmask16 maska3_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, _mm256_set1_epi16(0x7FFF)), vsatuval, _MM_CMPINT_NLT); + const __mmask16 maska4_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vflush, _MM_CMPINT_LT); + const __mmask16 maska5_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vxdnorm, _MM_CMPINT_LT); + /* Handle denormals, shift the mantissa before rounding */ + __m256i v_shft_ = _mm256_sub_epi16 ( + _mm256_srli_epi16 (vxdnorm, 10), + _mm256_srli_epi16 (_mm256_and_si256 (a_, vnaninf), 10)); + v_shft_ = _mm256_mask_mov_epi16 (vzero, maska5_, v_shft_); + a_ = _mm256_mask_srlv_epi16 (a_, maska5_, a_, v_shft_); + a_ = _mm256_mask_sllv_epi16 (a_, maska5_, a_, v_shft_); + + __m256i a_rne_ = _mm256_mask_add_epi16 (a_, maska1_, a_, + _mm256_mask_add_epi16 (vrneadd, maska2_, vrneadd, vfixup)); + /* saturating values beyond +/-MAX */ + a_rne_ = _mm256_mask_mov_epi16 (a_rne_, maska3_, _mm256_or_si256(_mm256_and_si256(a_rne_, vsign), vsatuval)); + a_rne_ = _mm256_mask_mov_epi16 (a_rne_, maska4_, _mm256_and_si256(a_rne_, vsign)); + + a_rne_ = _mm256_slli_epi16 (_mm256_srli_epi16 (a_rne_, 7), 7); + + bh_ = _mm256_extracti32x4_epi32 (a_rne_, 0); + ah_ = _mm256_extracti32x4_epi32 (a_rne_, 1); + b = _mm256_cvtph_ps (bh_); + a = _mm256_cvtph_ps (ah_); + _mm256_storeu_ps (&out[i], _mm256_mul_ps (b, sr_)); + _mm256_storeu_ps (&out[i + 8], _mm256_mul_ps (a, sr_)); + } + } + + void cvt_fp32_e4m3_ieee_stochastic_intrinsic (const float *__restrict__ in, + float *out, int size, float scale) { + const __m256i vnaninf = _mm256_set1_epi16 (0x7c00); + const __m256i vrneadd = _mm256_set1_epi16 (0x003f); + const __m256i vfixup = _mm256_set1_epi16 (0x0001); + const __m256i vfixupmask = _mm256_set1_epi16 (0x0080); + const __m256i vzero = _mm256_set1_epi16 (0x0000); + const __m256i vsign = _mm256_set1_epi16 (0x8000); + const __m256i vsatuval = _mm256_set1_epi16 (0x5B80);/* 2^7*1.111 a.k.a 240.0, largest value */ + const __m256i vflush = _mm256_set1_epi16 (0x1800);/* 2^-9, smallest denormal */ + const __m256i vxdnorm = _mm256_set1_epi16 (0x2400);/* 2^-6 smallest normal */ + + for (int i = 0; i < size; i += 16) { + unsigned int rndbuf[16]; + /* generate 128 random bits */ + for (int r = 0; r < 8; r++) { + rndbuf[r] = (unsigned int) rand_xorshft128plus_scalar (sptr_[r]); + } + __m128i rnbits = _mm_load_si128 ((const __m128i *) &rndbuf[0]); + + __m256 s_ = _mm256_set1_ps (scale); + __m256 sr_ = _mm256_set1_ps (1.0 / scale); + __m256 b = _mm256_loadu_ps (&in[i]); + __m256 a = _mm256_loadu_ps (&in[i + 8]); + + b = _mm256_mul_ps (b, s_); + a = _mm256_mul_ps (a, s_); + + __m128i ah_ = _mm256_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = _mm256_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i a_ = _mm256_inserti32x4 (_mm256_inserti32x4 (_mm256_setzero_si256 (), bh_, 0), ah_, 1); + const __mmask16 maska1_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask16 maska2_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + const __mmask16 maska3_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, _mm256_set1_epi16(0x7FFF)), vsatuval, _MM_CMPINT_NLT); + const __mmask16 maska4_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vflush, _MM_CMPINT_LT); + const __mmask16 maska5_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vxdnorm, _MM_CMPINT_LT); + /* Handle denormals, shift the mantissa before rounding */ + __m256i v_shft_ = _mm256_sub_epi16 ( + _mm256_srli_epi16 (vxdnorm, 10), + _mm256_srli_epi16 (_mm256_and_si256 (a_, vnaninf), 10)); + v_shft_ = _mm256_mask_mov_epi16 (vzero, maska5_, v_shft_); + a_ = _mm256_mask_srlv_epi16 (a_, maska5_, a_, v_shft_); + a_ = _mm256_mask_sllv_epi16 (a_, maska5_, a_, v_shft_); + + + __m256i a_sr_ = _mm256_mask_add_epi16 (a_, (maska1_ & ~maska5_), a_, + _mm256_srli_epi16 (_mm256_cvtepu8_epi16 (rnbits), 1)); + a_sr_ = _mm256_mask_add_epi16 (a_sr_, maska5_, a_sr_, + _mm256_mask_add_epi16 (vrneadd, maska2_, vrneadd, vfixup)); + a_sr_ = _mm256_mask_mov_epi16 (a_sr_, maska3_, _mm256_or_si256(_mm256_and_si256(a_sr_, vsign), vsatuval)); + a_sr_ = _mm256_mask_mov_epi16 (a_sr_, maska4_, _mm256_and_si256(a_sr_, vsign)); + + a_sr_ = _mm256_slli_epi16 (_mm256_srli_epi16 (a_sr_, 7), 7); + + bh_ = _mm256_extracti32x4_epi32 (a_sr_, 0); + ah_ = _mm256_extracti32x4_epi32 (a_sr_, 1); + b = _mm256_cvtph_ps (bh_); + a = _mm256_cvtph_ps (ah_); + _mm256_storeu_ps (&out[i], _mm256_mul_ps (b, sr_)); + _mm256_storeu_ps (&out[i + 8], _mm256_mul_ps (a, sr_)); + } + } + + void cvt_fp32_e4m3_ieee_scalar (const float *__restrict__ in, float *out, + int size, float scale, int rmode) { + int non_mant_bits = 4 /*exp_bits */ + 1; /* exponent + sign */ + int lshift = 10 - (8 /*mbits */ - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x007F; + unsigned short rne_tie = 0x00C0; + float scale_reciprocal = 1.0 / scale; + +#pragma omp parallel for + for (int gid = 0; gid < size; gid++) { + __half_t h; + float inval = scale * in[gid]; + + h.u = __anyfloat2half_rn (inval); + short exp_h = (short) ((h.u & 0x7C00) >> 10) - 15; + short sign_h = (h.u & 0x8000); + short mantissa_h = (h.u & 0x03FF); + + unsigned short can_round = ((h.u & 0x7FFF) < 0x4B80) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + int dshift = 0; + + if (exp_h > 7 || (can_round == 0)) { + mantissa_h = 0x380; + exp_h = 7; + can_round = 0; + } else if (exp_h < -9) { + /* flush values below 1-4-3 subnormal range to zero */ + exp_h = -15; + mantissa_h = 0; + } else if (exp_h < -6) { + dshift = (-6 - exp_h); + /* handle denormals */ + mantissa_h = mantissa_h >> dshift; + mantissa_h <<= dshift; + } + /* nearest rounding masks */ + unsigned short rnmask = (mantissa_h & grs_bitmask); + unsigned short rnmask_tie = (mantissa_h & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + rand_xorshft128plus_scalar (sptr_[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + mantissa_h += can_round * is_normal * (rand & 0x7F); + /* stochastic round: denormals --> rne rounding */ + mantissa_h += can_round * is_denorm * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + mantissa_h += can_round * rne_mask * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + mantissa_h += can_round * rnaz_mask * ((rnmask >= 0x0040) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + mantissa_h += can_round * rntz_mask * ((rnmask > 0x0040) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + mantissa_h += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0040) << lshift); + /* round to -INF, if rminf_mask is enabled */ + mantissa_h += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0040) << lshift); + } + } + /* truncation */ + mantissa_h &= mask_mant; + mantissa_h += ((exp_h + 15) << 10); + mantissa_h |= sign_h; + h.u = mantissa_h; + float f; + + __half2anyfloat (h.u, &f); + out[gid] = (f * scale_reciprocal); + } + } + + template < typename scalar_t > + void E4M3_IEEE_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, + int block_size, int rmode) { + float scale = in_scale; + float fmax = std::numeric_limits::max(); + if (scale > fmax) { + fprintf(stderr,"Error: Invalid scale factor : %.2e, make sure the scale is not larger than : %.2e\n", scale, fmax); + exit(1); + } + + if (block_norm == true) { + int nblocks = (size + (block_size - 1)) / block_size; + +#pragma omp parallel for + for (int b = 0; b < nblocks; b++) { + int start_index = (b * block_size); + + /* handle the last block */ + if (start_index + block_size > size) + block_size = (size - start_index); + + float maxval = 0.0; + +#pragma omp parallel for reduction (max:maxval) + for (int gid = start_index; gid < start_index + block_size; gid++) { + maxval = (maxval < fabs (in[gid])) ? fabs (in[gid]) : maxval; + } + __float_t f; + + f.f = maxval; + f.u = (f.u & 0x7F800000); + scale = 2.0 * f.f; + scale /= 8.0; + + if ((block_size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e4m3_ieee_stochastic_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } else { + cvt_fp32_e4m3_ieee_rne_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } + } else { + cvt_fp32_e4m3_ieee_scalar (&in[start_index], &out[start_index], block_size, scale, rmode); + } + } + } else { + if ((size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e4m3_ieee_stochastic_intrinsic (in, out, size, scale); + } else { + cvt_fp32_e4m3_ieee_rne_intrinsic (in, out, size, scale); + } + } else { + int vec_size = ((int) (size / 32)) * 32; + if (vec_size > 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e4m3_ieee_stochastic_intrinsic (in, out, vec_size, scale); + } else { + cvt_fp32_e4m3_ieee_rne_intrinsic (in, out, vec_size, scale); + } + } + cvt_fp32_e4m3_ieee_scalar (&in[vec_size], &out[vec_size], size - vec_size, scale, rmode); + } + } + } + + void cvt_fp32_e3m4_rne_intrinsic (const float *__restrict__ in, + float *out, int size, float scale) { + + const __m256i vnaninf = _mm256_set1_epi16 (0x7c00); + const __m256i vrneadd = _mm256_set1_epi16 (0x001f); + const __m256i vfixup = _mm256_set1_epi16 (0x0001); + const __m256i vfixupmask = _mm256_set1_epi16 (0x0040); + const __m256i vzero = _mm256_set1_epi16 (0x0000); + const __m256i vsign = _mm256_set1_epi16 (0x8000); + const __m256i vsatuval = _mm256_set1_epi16 (0x4F80);/* 2^4*1.1110 a.k.a 30.0, largest value */ + const __m256i vflush = _mm256_set1_epi16 (0x2400);/* 2^-6, smallest denormal */ + const __m256i vxdnorm = _mm256_set1_epi16 (0x3400);/* 2^-2 smallest normal */ + + for (int i = 0; i < size; i += 16) { + __m256 s_ = _mm256_set1_ps (scale); + __m256 sr_ = _mm256_set1_ps (1.0 / scale); + __m256 b = _mm256_loadu_ps (&in[i]); + __m256 a = _mm256_loadu_ps (&in[i + 8]); + + b = _mm256_mul_ps (b, s_); + a = _mm256_mul_ps (a, s_); + + __m128i ah_ = _mm256_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = _mm256_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i a_ = _mm256_inserti32x4 (_mm256_inserti32x4 (_mm256_setzero_si256 (), bh_, 0), ah_, 1); + const __mmask16 maska1_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask16 maska2_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + const __mmask16 maska3_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, _mm256_set1_epi16(0x7FFF)), vsatuval, _MM_CMPINT_NLT); + const __mmask16 maska4_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vflush, _MM_CMPINT_LT); + const __mmask16 maska5_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vxdnorm, _MM_CMPINT_LT); + /* Handle denormals, shift the mantissa before rounding */ + __m256i v_shft_ = _mm256_sub_epi16 ( + _mm256_srli_epi16 (vxdnorm, 10), + _mm256_srli_epi16 (_mm256_and_si256 (a_, vnaninf), 10)); + v_shft_ = _mm256_mask_mov_epi16 (vzero, maska5_, v_shft_); + a_ = _mm256_mask_srlv_epi16 (a_, maska5_, a_, v_shft_); + a_ = _mm256_mask_sllv_epi16 (a_, maska5_, a_, v_shft_); + + __m256i a_rne_ = _mm256_mask_add_epi16 (a_, maska1_, a_, + _mm256_mask_add_epi16 (vrneadd, maska2_, vrneadd, vfixup)); + /* saturating values beyond +/-MAX */ + a_rne_ = _mm256_mask_mov_epi16 (a_rne_, maska3_, _mm256_or_si256(_mm256_and_si256(a_rne_, vsign), vsatuval)); + a_rne_ = _mm256_mask_mov_epi16 (a_rne_, maska4_, _mm256_and_si256(a_rne_, vsign)); + + a_rne_ = _mm256_slli_epi16 (_mm256_srli_epi16 (a_rne_, 6), 6); + + bh_ = _mm256_extracti32x4_epi32 (a_rne_, 0); + ah_ = _mm256_extracti32x4_epi32 (a_rne_, 1); + b = _mm256_cvtph_ps (bh_); + a = _mm256_cvtph_ps (ah_); + _mm256_storeu_ps (&out[i], _mm256_mul_ps (b, sr_)); + _mm256_storeu_ps (&out[i + 8], _mm256_mul_ps (a, sr_)); + } + } + + void cvt_fp32_e3m4_stochastic_intrinsic (const float *__restrict__ in, + float *out, int size, + float scale) { + + const __m256i vnaninf = _mm256_set1_epi16 (0x7c00); + const __m256i vrneadd = _mm256_set1_epi16 (0x001f); + const __m256i vfixup = _mm256_set1_epi16 (0x0001); + const __m256i vfixupmask = _mm256_set1_epi16 (0x0040); + const __m256i vzero = _mm256_set1_epi16 (0x0000); + const __m256i vsign = _mm256_set1_epi16 (0x8000); + const __m256i vsatuval = _mm256_set1_epi16 (0x4F80);/* 2^4*1.1110 a.k.a 30.0, largest value */ + const __m256i vflush = _mm256_set1_epi16 (0x2400);/* 2^-6, smallest denormal */ + const __m256i vxdnorm = _mm256_set1_epi16 (0x3400);/* 2^-2 smallest normal */ + + for (int i = 0; i < size; i += 16) { + unsigned int rndbuf[16]; + /* generate 128 random bits */ + for (int r = 0; r < 8; r++) { + rndbuf[r] = (unsigned int) rand_xorshft128plus_scalar (sptr_[r]); + } + __m128i rnbits = _mm_load_si128 ((const __m128i *) &rndbuf[0]); + + __m256 s_ = _mm256_set1_ps (scale); + __m256 sr_ = _mm256_set1_ps (1.0 / scale); + __m256 b = _mm256_loadu_ps (&in[i]); + __m256 a = _mm256_loadu_ps (&in[i + 8]); + + b = _mm256_mul_ps (b, s_); + a = _mm256_mul_ps (a, s_); + + __m128i ah_ = _mm256_cvtps_ph (a, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m128i bh_ = _mm256_cvtps_ph (b, (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)); + __m256i a_ = _mm256_inserti32x4 (_mm256_inserti32x4 (_mm256_setzero_si256 (), bh_, 0), ah_, 1); + const __mmask16 maska1_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vnaninf, _MM_CMPINT_NE); + const __mmask16 maska2_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vfixupmask), vfixupmask, _MM_CMPINT_EQ); + const __mmask16 maska3_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, _mm256_set1_epi16(0x7FFF)), vsatuval, _MM_CMPINT_NLT); + const __mmask16 maska4_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vflush, _MM_CMPINT_LT); + const __mmask16 maska5_ = _mm256_cmp_epi16_mask (_mm256_and_si256 (a_, vnaninf), vxdnorm, _MM_CMPINT_LT); + /* Handle denormals, shift the mantissa before rounding */ + __m256i v_shft_ = _mm256_sub_epi16 ( + _mm256_srli_epi16 (vxdnorm, 10), + _mm256_srli_epi16 (_mm256_and_si256 (a_, vnaninf), 10)); + v_shft_ = _mm256_mask_mov_epi16 (vzero, maska5_, v_shft_); + a_ = _mm256_mask_srlv_epi16 (a_, maska5_, a_, v_shft_); + a_ = _mm256_mask_sllv_epi16 (a_, maska5_, a_, v_shft_); + + __m256i a_sr_ = _mm256_mask_add_epi16 (a_, (maska1_ & ~maska5_), a_, + _mm256_srli_epi16 (_mm256_cvtepu8_epi16 (rnbits), 1)); + a_sr_ = _mm256_mask_add_epi16 (a_sr_, maska5_, a_sr_, + _mm256_mask_add_epi16 (vrneadd, maska2_, vrneadd, vfixup)); + + a_sr_ = _mm256_mask_mov_epi16 (a_sr_, maska3_, _mm256_or_si256(_mm256_and_si256(a_sr_, vsign), vsatuval)); + a_sr_ = _mm256_mask_mov_epi16 (a_sr_, maska4_, _mm256_and_si256(a_sr_, vsign)); + a_sr_ = _mm256_slli_epi16 (_mm256_srli_epi16 (a_sr_, 6), 6); + + bh_ = _mm256_extracti32x4_epi32 (a_sr_, 0); + ah_ = _mm256_extracti32x4_epi32 (a_sr_, 1); + b = _mm256_cvtph_ps (bh_); + a = _mm256_cvtph_ps (ah_); + _mm256_storeu_ps (&out[i], _mm256_mul_ps (b, sr_)); + _mm256_storeu_ps (&out[i + 8], _mm256_mul_ps (a, sr_)); + } + } + + void cvt_fp32_e3m4_scalar (const float *__restrict__ in, float *out, + int size, float scale, int rmode) { + int non_mant_bits = 3 /*exp_bits */ + 1; /* exponent + sign */ + int lshift = 10 - (8 /*mbits */ - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x003F; + unsigned short rne_tie = 0x0060; + float scale_reciprocal = 1.0 / scale; + +#pragma omp parallel for + for (int gid = 0; gid < size; gid++) { + __half_t h; + float inval = scale * in[gid]; + + h.u = __anyfloat2half_rn (inval); + short exp_h = (short) ((h.u & 0x7C00) >> 10) - 15; + short sign_h = (h.u & 0x8000); + short mantissa_h = (h.u & 0x03FF); + + unsigned short can_round = ((h.u & 0x7FFF) < 0x4F80) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + int dshift = 0; + if (exp_h > 4 || (can_round == 0)) { + /* Software : saturate values above +/-30.0 to +/-30.0 */ + mantissa_h = 0x0380; + exp_h = 4; + can_round = 0; + } else if (exp_h < -6) { + exp_h = -15; + mantissa_h = 0; + } else if (exp_h < -2) { + dshift = (-2 - exp_h); + /* handle denormals */ + mantissa_h = mantissa_h >> dshift; + mantissa_h <<= dshift; + } + /* nearest rounding masks */ + unsigned short rnmask = (mantissa_h & grs_bitmask); + unsigned short rnmask_tie = (mantissa_h & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + rand_xorshft128plus_scalar (sptr_[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + mantissa_h += can_round * is_normal * (rand & 0x3F); + /* stochastic round: denormals --> rne rounding */ + mantissa_h += can_round * is_denorm * + (((rnmask > 0x0020) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + mantissa_h += can_round * rne_mask * + (((rnmask > 0x0020) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + mantissa_h += can_round * rnaz_mask * ((rnmask >= 0x0020) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + mantissa_h += can_round * rntz_mask * ((rnmask > 0x0020) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + mantissa_h += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0020) << lshift); + /* round to -INF, if rminf_mask is enabled */ + mantissa_h += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0020) << lshift); + } + } + /* truncation */ + mantissa_h &= mask_mant; + mantissa_h += ((exp_h + 15) << 10); + mantissa_h |= sign_h; + h.u = mantissa_h; + float f; + + __half2anyfloat (h.u, &f); + out[gid] = (f * scale_reciprocal); + } + } + + template < typename scalar_t > + void E3M4_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, + int block_size, int rmode) { + float scale = in_scale; + float fmax = std::numeric_limits::max(); + if (scale > fmax) { + fprintf(stderr,"Error: Invalid scale factor : %.2e, make sure the scale is not larger than : %.2e\n", scale, fmax); + exit(1); + } + + if (block_norm == true) { + int nblocks = (size + (block_size - 1)) / block_size; + +#pragma omp parallel for + for (int b = 0; b < nblocks; b++) { + int start_index = (b * block_size); + + /* handle the last block */ + if (start_index + block_size > size) + block_size = (size - start_index); + + float maxval = 0.0; + +#pragma omp parallel for reduction (max:maxval) + for (int gid = start_index; gid < start_index + block_size; gid++) { + maxval = (maxval < fabs (in[gid])) ? fabs (in[gid]) : maxval; + } + __float_t f; + + f.f = maxval; + f.u = (f.u & 0x7F800000); + scale = 2.0 * f.f; + + if ((block_size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e3m4_stochastic_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } else { + cvt_fp32_e3m4_rne_intrinsic (&in[start_index], &out[start_index], block_size, scale); + } + } else { + cvt_fp32_e3m4_scalar (&in[start_index], &out[start_index], block_size, scale, rmode); + } + } + } else { + if ((size % 32) == 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e3m4_stochastic_intrinsic (in, out, size, scale); + } else { + cvt_fp32_e3m4_rne_intrinsic (in, out, size, scale); + } + } else { + int vec_size = ((int) (size / 32)) * 32; + if (vec_size > 0) { + if (rmode == ROUND_STOCHASTIC) { + cvt_fp32_e3m4_stochastic_intrinsic (in, out, vec_size, scale); + } else { + cvt_fp32_e3m4_rne_intrinsic (in, out, vec_size, scale); + } + } + cvt_fp32_e3m4_scalar (&in[vec_size], &out[vec_size], + size - vec_size, scale, rmode); + } + } + } + + void cvt_fp32_fp4_nearest_scalar (const float *__restrict__ in, float *out, + int size, float scale) { + + float scale_reciprocal = 1.0 / scale; + +#pragma omp parallel for + for (int gid = 0; gid < size; gid++) { + __float_t f; + float inval = scale * in[gid]; + + f.f = inval; + int exp_f = (int)(((f.u & 0x7F800000) >> 23) - 127); + int sign_f = (f.u & 0x80000000); + /* see if round up works! */ + if (exp_f < 0 && (exp_f%2)) f.f *= 1.6; + /* saturate */ + if (exp_f > 0) f.u = (sign_f | (127 << 23)); + f.u &= 0xFF800000; + /* extract the new exponent */ + exp_f = (int)(((f.u & 0x7F800000) >> 23) - 127); + /* round up did not work, round down */ + if (exp_f < 0 && (exp_f%2)) f.u = (sign_f | ((exp_f + 126) << 23)); + /* flush values smaller than 2^-12 to zero */ + //if (exp_f < -6) f.u = 0; + if (exp_f < -12) f.u = 0; + out[gid] = (f.f * scale_reciprocal); + } + } + + template < typename scalar_t > + void FP4_Nearest_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, + int block_size) { + float scale = in_scale; + float fmax = std::numeric_limits::max(); + if (scale > fmax) { + fprintf(stderr,"Error: Invalid scale factor : %.2e, make sure the scale is not larger than : %.2e\n", scale, fmax); + exit(1); + } + + if (block_norm == true) { + int nblocks = (size + (block_size - 1)) / block_size; + +#pragma omp parallel for + for (int b = 0; b < nblocks; b++) { + int start_index = (b * block_size); + + /* handle the last block */ + if (start_index + block_size > size) + block_size = (size - start_index); + + float maxval = 0.0; + +#pragma omp parallel for reduction (max:maxval) + for (int gid = start_index; gid < start_index + block_size; gid++) { + maxval = (maxval < fabs (in[gid])) ? fabs (in[gid]) : maxval; + } + /* FP4 max value is 1.0 */ + scale = 1.0/maxval; + cvt_fp32_fp4_nearest_scalar(&in[start_index], &out[start_index], block_size, scale); + } + } else { + cvt_fp32_fp4_nearest_scalar (&in[0], &out[0], size, scale); + } + } + + std::vector < torch::Tensor > fpemu_common_function (torch::Tensor input, + std::string mode, + int size, + bool inplace, + float scale, + bool block_norm, + int block_size) { + + torch::Tensor output; + if (!inplace) + output = torch::zeros_like (input); + + if (!mode.compare ("E5M2_RTZ")) { + E5M2_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, scale, block_norm, block_size, ROUND_RTZ); + } else if (!mode.compare ("E5M2_RNE")) { + E5M2_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, scale, block_norm, block_size, ROUND_RNE); + } else if (!mode.compare ("E5M2_STOCHASTIC")) { + E5M2_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, scale, block_norm, block_size, ROUND_STOCHASTIC); + } else if (!mode.compare ("E5M2_RNAZ")) { + E5M2_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, scale, block_norm, block_size, ROUND_RNAZ); + } else if (!mode.compare ("E5M2_RNTZ")) { + E5M2_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, scale, block_norm, block_size, ROUND_RNTZ); + } else if (!mode.compare ("E5M2_RPINF")) { + E5M2_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, scale, block_norm, block_size, ROUND_PINF); + } else if (!mode.compare ("E5M2_RNINF")) { + E5M2_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, scale, block_norm, block_size, ROUND_NINF); + } else if (!mode.compare ("E5M2_DAZ_RNE")) { + E5M2_DAZ_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, ROUND_RNE); + } else if (!mode.compare ("E5M2_DAZ_STOCHASTIC")) { + E5M2_DAZ_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, ROUND_STOCHASTIC); + } else if (!mode.compare ("E5M2_DAZ_RNAZ")) { + E5M2_DAZ_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, ROUND_RNAZ); + } else if (!mode.compare ("E5M2_DAZ_RNTZ")) { + E5M2_DAZ_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, ROUND_RNTZ); + } else if (!mode.compare ("FLOAT16_RNE")) { + FLOAT16_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, ROUND_RNE, 0); + } else if (!mode.compare ("FLOAT16_STOCHASTIC")) { + FLOAT16_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, ROUND_STOCHASTIC, 0); + } else if (!mode.compare ("FLOAT16_DAZ_RNE")) { + FLOAT16_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < float >(), + size, ROUND_RNE, 1); + } else if (!mode.compare ("BFLOAT16_RNE")) { + BFLOAT16_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, ROUND_RNE); + } else if (!mode.compare ("BFLOAT16_STOCHASTIC")) { + BFLOAT16_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, ROUND_STOCHASTIC); + } else if (!mode.compare ("E4M3_IEEE_RNE")) { + E4M3_IEEE_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size, ROUND_RNE); + } else if (!mode.compare ("E4M3_IEEE_STOCHASTIC")) { + E4M3_IEEE_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size, ROUND_STOCHASTIC); + } else if (!mode.compare ("E4M3_RNE")) { + E4M3_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size, ROUND_RNE); + } else if (!mode.compare ("E4M3_STOCHASTIC")) { + E4M3_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size, ROUND_STOCHASTIC); + } else if (!mode.compare ("E3M4_RNE")) { + E3M4_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size, ROUND_RNE); + } else if (!mode.compare ("E3M4_STOCHASTIC")) { + E3M4_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size, ROUND_STOCHASTIC); + } else if (!mode.compare ("FP4_NEAREST")) { + FP4_Nearest_Kernel < float >(input.data_ptr < float >(), + (inplace) ? input.data_ptr < + float >() : output.data_ptr < + float >(), size, scale, block_norm, + block_size); + } + + if (!inplace) { + return { + output,}; + } else { + return { + input,}; + } + } + +}//namespace + +std::vector < torch::Tensor > fpemu_forward (torch::Tensor input, + std::string mode, + int size, + bool inplace, + float scale, + bool block_norm, + int block_size) { + if (block_norm == true && block_size != size) { + if (size % block_size) { + block_norm = false; + block_size = 1; + } + } + return fpemu_common_function (input, mode, size, inplace, scale, block_norm, + block_size); +} + +std::vector < torch::Tensor > fpemu_backward (torch::Tensor grad, + std::string mode, + int size, + bool inplace, + float scale, + bool block_norm, + int block_size) { + if (block_norm == true && block_size != size) { + if (size % block_size) { + block_norm = false; + block_size = 1; + } + } + return fpemu_common_function (grad, mode, size, inplace, scale, block_norm, + block_size); +} + +PYBIND11_MODULE (TORCH_EXTENSION_NAME, m) { + m.def ("forward", &fpemu_forward, "FPEmu forward"); + m.def ("backward", &fpemu_backward, "FPEmu backward"); +} \ No newline at end of file diff --git a/FP8_Emulator/pytquant/cuda/__init__.py b/FP8_Emulator/pytquant/cuda/__init__.py new file mode 100644 index 00000000..893c7619 --- /dev/null +++ b/FP8_Emulator/pytquant/cuda/__init__.py @@ -0,0 +1,3 @@ +import torch +if torch.cuda.is_available(): + from . import fpemu as fpemu_cuda diff --git a/FP8_Emulator/pytquant/cuda/fpemu.py b/FP8_Emulator/pytquant/cuda/fpemu.py new file mode 100644 index 00000000..8cad717f --- /dev/null +++ b/FP8_Emulator/pytquant/cuda/fpemu.py @@ -0,0 +1,74 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +import math +from torch import nn +import torch +import numpy +import fpemu_cuda + +from enum import Enum + +torch.manual_seed(42) + +""" + NONE + E5M2_RTZ + E5M2_STOCHASTIC + E5M2_RNE + E5M2_RNAZ + E5M2_RNTZ + E5M2_RPINF + E5M2_RNINF + E5M2_DAZ_STOCHASTIC + E5M2_DAZ_RNE + E5M2_DAZ_RNAZ + E5M2_DAZ_RNTZ + BFLOAT16_STOCHASTIC + BFLOAT16_RNE + FLOAT16_RNE + FLOAT16_STOCHASTIC + FLOAT16_DAZ_RNE + E4M3_RNE + E4M3_STOCHASTIC +""" + +class FPEmuOp(torch.autograd.Function): + @staticmethod + def forward(ctx, input, mode='NONE', inplace=False, scale=1.0, blocknorm=False, blocksize=1): + if mode == 'NONE' : + ctx.mark_dirty(input) + return input + else : + if input.is_sparse : + input = input.coalesce() + size = input.values().nelement() + if inplace == True: + outputs = fpemu_cuda.forward(input._values().contiguous(), mode, size, inplace, scale, blocknorm, blocksize) + output = input + else : + outputs = fpemu_cuda.forward(input._values().contiguous(), mode, size, inplace, scale, blocknorm, blocksize) + output = torch.sparse.FloatTensor(input.indices(), outputs[0], input.size()) + else : + size = input.nelement() + outputs = fpemu_cuda.forward(input.contiguous(), mode, size, inplace, scale, blocknorm, blocksize) + output = outputs[0] + + if inplace == True: + ctx.mark_dirty(input) + return output + + @staticmethod + def backward(ctx, output_grad): + # straight-through estimator + return output_grad, None, None, None, None + + @staticmethod + def symbolic(g, input, mode='NONE', inplace=False, scale=1.0, blocknorm=False, blocksize=1): + return g.op("::FPEmuOp", input, mode, inplace, scale, blocknorm, blocksize) \ No newline at end of file diff --git a/FP8_Emulator/pytquant/cuda/fpemu_impl.cpp b/FP8_Emulator/pytquant/cuda/fpemu_impl.cpp new file mode 100644 index 00000000..deeec337 --- /dev/null +++ b/FP8_Emulator/pytquant/cuda/fpemu_impl.cpp @@ -0,0 +1,54 @@ +/*----------------------------------------------------------------------------* + * Copyright (c) 2023, Intel Corporation - All rights reserved. + * This file is part of FP8-Emulation-Toolkit + * + * SPDX-License-Identifier: BSD-3-Clause + *----------------------------------------------------------------------------* + * Naveen Mellempudi (Intel Corporation) + *----------------------------------------------------------------------------*/ + +#include +#include + +std::vector fpemu_cuda_forward( + torch::Tensor input, + std::string mode, + int size, + bool inplace, + float scale, + bool block_norm, + int block_size); + +// NOTE: AT_ASSERT has become AT_CHECK on master after 0.4. +#define CHECK_CUDA(x) AT_ASSERTM(x.device().is_cuda(), #x " must be a CUDA tensor") +#define CHECK_CONTIGUOUS(x) AT_ASSERTM(x.is_contiguous(), #x " must be contiguous") +#define CHECK_INPUT(x) CHECK_CUDA(x); CHECK_CONTIGUOUS(x) + +std::vector fpemu_forward( + torch::Tensor input, + std::string mode, + int size, + bool inplace, + float scale, + bool block_norm, + int block_size) { + CHECK_INPUT(input); + return fpemu_cuda_forward(input, mode, size, inplace, scale, block_norm, block_size); +} + +std::vector fpemu_backward( + torch::Tensor grad, + std::string mode, + int size, + bool inplace, + float scale, + bool block_norm, + int block_size) { + CHECK_INPUT(grad); + return fpemu_cuda_forward(grad, mode, size, inplace, scale, block_norm, block_size); +} + +PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) { + m.def("forward", &fpemu_forward, "FPEmu forward (CUDA)"); + m.def("backward", &fpemu_backward, "FPEmu backward (CUDA)"); +} diff --git a/FP8_Emulator/pytquant/cuda/fpemu_kernels.cu b/FP8_Emulator/pytquant/cuda/fpemu_kernels.cu new file mode 100644 index 00000000..d006be43 --- /dev/null +++ b/FP8_Emulator/pytquant/cuda/fpemu_kernels.cu @@ -0,0 +1,1381 @@ +/*----------------------------------------------------------------------------* + * Copyright (c) 2023, Intel Corporation - All rights reserved. + * This file is part of FP8-Emulation-Toolkit + * + * SPDX-License-Identifier: BSD-3-Clause + *----------------------------------------------------------------------------* + * Naveen Mellempudi (Intel Corporation) + *----------------------------------------------------------------------------*/ + +#include +#include +#include +#include +#include + +#define CUBLOCK_SIZE 256 +enum ROUNDING_MODES { ROUND_RTZ = 0, ROUND_RNE = 1, ROUND_STOCHASTIC = + 2, ROUND_RNAZ = 3, ROUND_RNTZ = 4, ROUND_PINF = 5, ROUND_NINF = 6 }; + +namespace { + + typedef union half_t { + unsigned short u; + at::Half f; + } __half_t; + + typedef union ufloat32 { + unsigned u; + float f; + } __float_t; + +/* this implementation of xoroshiro128++ PRNG is borrowed from here: + http://prng.di.unimi.it/xoshiro128plusplus.c + main page: http://prng.di.unimi.it/ +*/ + __device__ static uint32_t s1[4] = + { 1387366120, 279844183, 888998500, 1099633400 }; + __device__ static uint32_t s2[4] = + { 2034269327, 2125325156, 1209715489, 193165672 }; + __device__ static uint32_t s3[4] = + { 1555452618, 650181557, 883695203, 62767784 }; + __device__ static uint32_t s4[4] = + { 419524804, 2146478152, 480059239, 1468956197 }; + __device__ static uint32_t s5[4] = + { 1252084877, 500390994, 977516591, 1950666000 }; + __device__ static uint32_t s6[4] = + { 393659750, 834151069, 1477014702, 734008143 }; + __device__ static uint32_t s7[4] = + { 1983400973, 116410309, 2110188261, 2019272068 }; + __device__ static uint32_t s8[4] = + { 187709636, 28336299, 419632041, 1774181187 }; + __device__ static uint32_t s9[4] = + { 702309618, 407781555, 1512057936, 1868769368 }; + __device__ static uint32_t s10[4] = + { 510001215, 966559856, 776583255, 147562106 }; + __device__ static uint32_t s11[4] = + { 127180605, 1881312534, 478635452, 814821902 }; + __device__ static uint32_t s12[4] = + { 733990058, 1889991804, 1108257970, 1093480892 }; + __device__ static uint32_t s13[4] = + { 427374380, 416747337, 558000409, 1594848927 }; + __device__ static uint32_t s14[4] = + { 444870959, 1595722866, 1064124488, 363710254 }; + __device__ static uint32_t s15[4] = + { 703721499, 389640783, 1002360059, 1427395742 }; + __device__ static uint32_t s16[4] = + { 1295231497, 1254972431, 1423497865, 861918264 }; + + __device__ static uint32_t *sptr[16] = + { s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14, s15, s16 }; + + __device__ __forceinline__ uint32_t rotl_ (const uint32_t x, int k) { + return (x << k) | (x >> (32 - k)); + } + + __device__ __forceinline__ uint32_t _rand_xorshft128plus_with_seed (uint32_t + * ps) { + const uint32_t result_plus = ps[0] + ps[3]; + const uint32_t t = ps[1] << 9; + + ps[2] ^= ps[0]; + ps[3] ^= ps[1]; + ps[1] ^= ps[2]; + ps[0] ^= ps[3]; + + ps[2] ^= t; + + ps[3] = rotl_ (ps[3], 11); + + return result_plus; + } + + template < typename scalar_t > + __device__ __forceinline__ float __anyfloat2float_rn (scalar_t a_) { + float f_; + + if (std::is_same < scalar_t, double >::value) { + f_ = __double2float_rn (a_); + } else if (std::is_same < scalar_t, float >::value) { + f_ = a_; + } else if (std::is_same < scalar_t, at::Half >::value) { + f_ = __half2float ((at::Half) a_); + } + return f_; + } + + template < typename scalar_t > + __device__ __forceinline__ void __float2anyfloat_rn (float f_, + scalar_t * out) { + scalar_t a_; + + if (std::is_same < scalar_t, double >::value) { + a_ = (scalar_t) (f_); + } else if (std::is_same < scalar_t, float >::value) { + a_ = f_; + } else if (std::is_same < scalar_t, at::Half >::value) { + a_ = (at::Half) __float2half_rn (f_); + } + *out = a_; + } + + template < typename scalar_t > + __device__ __forceinline__ at::Half __anyfloat2half_rn (scalar_t f_) { + at::Half h_; + if (std::is_same < scalar_t, double >::value) { + h_ = __float2half_rn (__double2float_rn (f_)); + } else if (std::is_same < scalar_t, float >::value) { + h_ = __float2half_rn (f_); + } else if (std::is_same < scalar_t, at::Half >::value) { + h_ = (at::Half) f_; + } + return h_; + } + + template < typename scalar_t > + __device__ __forceinline__ void __half2anyfloat (at::Half h_, + scalar_t * out) { + scalar_t f_; + + if (std::is_same < scalar_t, double >::value) { + f_ = (scalar_t) __half2float ((at::Half) h_); + } else if (std::is_same < scalar_t, float >::value) { + f_ = __half2float (h_); + } else if (std::is_same < scalar_t, at::Half >::value) { + f_ = (at::Half) h_; + } + *out = f_; + } + + template < typename scalar_t > + __device__ void absmax_block (const scalar_t * in, + float *sdata, const int size) { + unsigned int tid = threadIdx.x; + unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; + + sdata[tid] = 0.0f; + + if (i < size) { + sdata[tid] = + fmaxf (fabsf (sdata[tid]), fabsf (__anyfloat2float_rn (in[i]))); + } + __syncthreads (); + + for (unsigned int s = 1; s < blockDim.x; s *= 2) { + if ((tid % (2 * s)) == 0) { + sdata[tid] = fmaxf (fabsf (sdata[tid]), fabsf (sdata[tid + s])); + } + __syncthreads (); + } + } + + __device__ static inline float atomicMaxf (float *address, float val) { + int *address_as_int = (int *) address; + int old = *address_as_int, assumed; + + while (val > __int_as_float (old)) { + assumed = old; + old = atomicCAS (address_as_int, assumed, __float_as_int (val)); + } + return __int_as_float (old); + } + + template < typename scalar_t > + __global__ void absmax0 (scalar_t * g_data, float *g_odata, + unsigned int n) { + extern __shared__ float sdata[]; + + unsigned int tid = threadIdx.x; + unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; + + sdata[tid] = 0.0f; + + if (i < n) { + sdata[tid] = + fmaxf (fabsf (sdata[tid]), fabsf (__anyfloat2float_rn (g_data[i]))); + } + __syncthreads (); + + for (unsigned int s = 1; s < blockDim.x; s *= 2) { + if ((tid % (2 * s)) == 0) { + sdata[tid] = fmaxf (fabsf (sdata[tid]), fabsf (sdata[tid + s])); + } + __syncthreads (); + } + __syncthreads (); + if (tid == 0) { + atomicMaxf (&g_odata[0], sdata[0]); + } + } + + template < typename scalar_t > + __global__ void reduce0 (scalar_t * g_data, float *g_odata, + unsigned int n) { + extern __shared__ float sdata[]; + unsigned int tid = threadIdx.x; + unsigned int i = blockIdx.x * blockDim.x + threadIdx.x; + + if (i < n) { + sdata[tid] = __anyfloat2float_rn (g_data[i]); + sdata[tid + blockDim.x] = + __anyfloat2float_rn (g_data[i]) * __anyfloat2float_rn (g_data[i]); + } else { + sdata[tid] = 0.0f; + sdata[tid + blockDim.x] = 0.0f; + } + __syncthreads (); + + for (unsigned int s = 1; s < blockDim.x; s *= 2) { + if ((tid % (2 * s)) == 0) { + sdata[tid] += sdata[tid + s]; + sdata[blockDim.x + tid] += sdata[blockDim.x + tid + s]; + } + __syncthreads (); + } + __syncthreads (); + if (tid == 0) { + atomicAdd (&g_odata[0], sdata[0]); + atomicAdd (&g_odata[1], sdata[blockDim.x]); + } + } + + + template < typename scalar_t > + __global__ void BFLOAT16_Kernel (const scalar_t * in, + scalar_t * out, + const int size, int rmode) { + int lshift = 16; + unsigned int mask_mant = (unsigned int) (0xFFFFFFFF << lshift); + unsigned int grs_bitmask = 0x0000FFFF; + unsigned int rne_tie = 0x00018000; + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + + if (rmode == ROUND_RNE) + rne_mask = 1; + if (rmode == ROUND_RNAZ) + rnaz_mask = 1; + if (rmode == ROUND_RNTZ) + rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) + sr_mask = 1; + + for (int gid = (blockIdx.x * blockDim.x) + threadIdx.x; gid < size; + gid += blockDim.x * gridDim.x) { + __float_t uf; + + uf.f = __anyfloat2float_rn (in[gid]); + unsigned int is_normal = (((uf.u & 0x7F800000) <= 0x7F000000) + && ((uf.u & 0x7F800000) >= 0x00800000)) ? 1 : 0; + unsigned int is_denorm = ((uf.u & 0x7F800000) == 0x0) ? 1 : 0; + unsigned int is_naninf = ((uf.u & 0x7F800000) == 0x7F800000) ? 1 : 0; + + /* nearest rounding masks */ + unsigned int rnmask = (uf.u & grs_bitmask); + unsigned int rnmask_tie = (uf.u & rne_tie); + + if (is_naninf == 0 && is_normal) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned int rand = _rand_xorshft128plus_with_seed (sptr[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + uf.u += (rand & 0x0000FFFF); + } else { + /* round to nearest even, if rne_mask is enabled */ + uf.u += rne_mask * + (((rnmask > 0x00008000) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + uf.u += rnaz_mask * ((rnmask >= 0x00008000) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + uf.u += rntz_mask * ((rnmask > 0x00008000) << lshift); + } + } else if (is_denorm) { + /* Flush Denormal */ + uf.u = 0; + } + /* truncation */ + uf.u = (uf.u & mask_mant); + + __float2anyfloat_rn (uf.f, &out[gid]); + } + } + + template < typename scalar_t > + __global__ void FLOAT16_Kernel (const scalar_t * in, + scalar_t * out, + const int size, + int rmode, int no_denorm) { + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + + for (int gid = (blockIdx.x * blockDim.x) + threadIdx.x; gid < size; + gid += blockDim.x * gridDim.x) { + if (std::is_same < scalar_t, at::Half >::value || rne_mask) { + __half_t h; + + at::Half hval; + hval = __anyfloat2half_rn (in[gid]); + h.f = hval; + unsigned short not_denorm = ((((h.u & 0x7FFF) >> 10) & 0x1F) > 0); + unsigned short is_denorm = (not_denorm == 0) ? 1 : 0; + + h.u *= !(is_denorm * no_denorm); + __half2anyfloat (h.f, &out[gid]); + } else if (sr_mask) { + unsigned int fval = ((unsigned int *) in)[gid]; + int exp_h = (int) ((fval & 0x7f800000) >> 23) - 127; + unsigned int mantissa_h = (fval & 0x7FFFFF); + unsigned int sign_h = (fval & 0x80000000); + __half_t h; + + if (exp_h == 128) { + /* handle incoming INF and NaN */ + exp_h = 0x1F; + /* handle signalling NaN */ + if (mantissa_h && ((mantissa_h & 0x400000) == 0x0)) + mantissa_h |= 0x400000; + mantissa_h |= (exp_h << 23); + mantissa_h |= (sign_h >> 3); + h.u = (unsigned short) (mantissa_h >> 13); + } else if (exp_h >= 16) { + /* saturate to INF */ + exp_h = 0x1F; + mantissa_h = 0; + mantissa_h |= (exp_h << 23); + mantissa_h |= (sign_h >> 3); + h.u = (unsigned short) (mantissa_h >> 13); + } else { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned int rand = (unsigned int) + _rand_xorshft128plus_with_seed (sptr[(seed_index % 16)]); + if (exp_h < -14) { + /* handle denormal values */ + h.f = __anyfloat2half_rn (in[gid]); + } else { + exp_h += 15; + mantissa_h |= (exp_h << 23); + mantissa_h |= (sign_h >> 3); + mantissa_h += (rand & 0x00001FFF); + h.u = (unsigned short) (mantissa_h >> 13); + } + } + __half2anyfloat (h.f, &out[gid]); + } + } + } + + template < typename scalar_t > + __global__ void E5M2_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, + int mbits, + int exp_bits, int rmode) { + int non_mant_bits = exp_bits + 1; /* exponent + sign */ + int lshift = 10 - (mbits - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x00FF; + unsigned short rne_tie = 0x0180; + + extern __shared__ float sdata[]; + scalar_t scale = in_scale; + + if (block_norm == true) { + absmax_block (in, sdata, size); + __float_t f; + f.f = (float) sdata[0]; + f.u = (f.u & 0x7F800000); + scale = 2.0 * f.f; + scale /= 16384.0; + } + float scale_reciprocal = 1.0 / scale; + + for (int gid = (blockIdx.x * blockDim.x) + threadIdx.x; gid < size; + gid += blockDim.x * gridDim.x) { + __half_t h; + float inval = in[gid] * scale; + __half hval = __anyfloat2half_rn (inval); + h.f = hval; + + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + unsigned short can_round = ((h.u & 0x7F00) <= 0x7B00) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + + /* nearest rounding masks */ + unsigned short rnmask = (h.u & grs_bitmask); + unsigned short rnmask_tie = (h.u & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = 0; /* use a constant seed index to reuse the same set of random numbers */ + unsigned short rand = + (unsigned short) _rand_xorshft128plus_with_seed (sptr[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + h.u += can_round * is_normal * (rand & 0xFF); + /* stochastic round: denormals --> rne rounding */ + h.u += can_round * is_denorm * + (((rnmask > 0x0080) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + h.u += can_round * rne_mask * + (((rnmask > 0x0080) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + h.u += can_round * rnaz_mask * ((rnmask >= 0x0080) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + h.u += can_round * rntz_mask * ((rnmask > 0x0080) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + h.u += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0080) << lshift); + /* round to -INF, if rminf_mask is enabled */ + h.u += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0080) << lshift); + } + } + /* truncation */ + h.u = (h.u & mask_mant); + __half2anyfloat (h.f * scale_reciprocal, &out[gid]); + } + } + + template < typename scalar_t > + __global__ void E5M2_DAZ_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t scale, + int mbits, int exp_bits, + int rmode) { + int non_mant_bits = exp_bits + 1; /* exponent + sign */ + int lshift = 10 - (mbits - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x00FF; + unsigned short rne_tie = 0x0180; + + float scale_reciprocal = 1.0 / scale; + for (int gid = (blockIdx.x * blockDim.x) + threadIdx.x; gid < size; + gid += blockDim.x * gridDim.x) { + __half_t h; + float inval = in[gid] * scale; + __half hval = __anyfloat2half_rn (inval); + h.f = hval; + /* values above 57344.0, saturate them to +- Infinity */ + unsigned short can_round = ((h.u & 0x7F00) <= 0x7B00) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + /* nearest rounding masks */ + unsigned short rnmask = (h.u & grs_bitmask); + unsigned short rnmask_tie = (h.u & rne_tie); + + if (is_naninf == 0 && is_normal) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + _rand_xorshft128plus_with_seed (sptr[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + h.u += can_round * (rand & 0xFF); + } else { + /* round to nearest even, if rne_mask is enabled */ + h.u += can_round * rne_mask * + (((rnmask > 0x0080) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + h.u += can_round * rnaz_mask * ((rnmask >= 0x0080) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + h.u += can_round * rntz_mask * ((rnmask > 0x0080) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + h.u += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0080) << lshift); + /* round to -INF, if rminf_mask is enabled */ + h.u += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0080) << lshift); + } + } else if (is_denorm) { + /* Flush Denormal */ + h.u = 0; + } + /* truncation */ + h.u = (h.u & mask_mant); + + __half2anyfloat (h.f * scale_reciprocal, &out[gid]); + } + } + + template < typename scalar_t > + __global__ void E4M3_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, + int mbits, + int exp_bits, int rmode) { + int non_mant_bits = exp_bits + 1; /* exponent + sign */ + int lshift = 10 - (mbits - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x007F; + unsigned short rne_tie = 0x00C0; + + extern __shared__ float sdata[]; + float scale = in_scale; + + if (block_norm == true) { + absmax_block (in, sdata, size); + __float_t f; + f.f = sdata[0]; + f.u = (f.u & 0x7F800000); + scale = 2 * f.f; + scale /= 8.0; + } + float scale_reciprocal = 1.0 / scale; + + for (int gid = (blockIdx.x * blockDim.x) + threadIdx.x; gid < size; + gid += blockDim.x * gridDim.x) { + __half_t h; + float inval = in[gid] * scale; + + h.f = __anyfloat2half_rn (inval); + short exp_h = (short) ((h.u & 0x7C00) >> 10) - 15; + short sign_h = (h.u & 0x8000); + short mantissa_h = (h.u & 0x03FF); + + unsigned short can_round = ((h.u & 0x7FFF) < 0x5F00) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + int dshift = 0; + + if (exp_h > 8 || (can_round == 0)) { + /* Software : saturate values above to +/-448.0 to +/-448.0 */ + mantissa_h = 0x0300; + exp_h = 8; + can_round = 0; + } else if (exp_h < -9) { + /* flush values below 1-4-3 subnormal range to zero */ + exp_h = -15; + mantissa_h = 0; + } else if (exp_h < -6) { + dshift = (-6 - exp_h); + /* handle denormals */ + mantissa_h = mantissa_h >> dshift; + mantissa_h <<= dshift; + } + /* nearest rounding masks */ + unsigned short rnmask = (mantissa_h & grs_bitmask); + unsigned short rnmask_tie = (mantissa_h & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + _rand_xorshft128plus_with_seed (sptr[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + mantissa_h += can_round * is_normal * (rand & 0x7F); + /* stochastic round: denormals --> rne rounding */ + mantissa_h += can_round * is_denorm * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + mantissa_h += can_round * rne_mask * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + mantissa_h += can_round * rnaz_mask * ((rnmask >= 0x0040) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + mantissa_h += can_round * rntz_mask * ((rnmask > 0x0040) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + mantissa_h += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0040) << lshift); + /* round to -INF, if rminf_mask is enabled */ + mantissa_h += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0040) << lshift); + } + } + /* truncation */ + mantissa_h &= mask_mant; + mantissa_h += ((exp_h + 15) << 10); + mantissa_h |= sign_h; + h.u = mantissa_h; + __half2anyfloat (h.f * scale_reciprocal, &out[gid]); + } + } + + template < typename scalar_t > + __global__ void E4M3_IEEE_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, + int mbits, int exp_bits, + int rmode) { + int non_mant_bits = exp_bits + 1; /* exponent + sign */ + int lshift = 10 - (mbits - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x007F; + unsigned short rne_tie = 0x00C0; + + extern __shared__ float sdata[]; + float scale = in_scale; + + if (block_norm == true) { + absmax_block (in, sdata, size); + __float_t f; + f.f = sdata[0]; + f.u = (f.u & 0x7F800000); + scale = 2*f.f; + scale = 2 * f.f; + scale /= 8.0; + } + float scale_reciprocal = 1.0 / scale; + + for (int gid = (blockIdx.x * blockDim.x) + threadIdx.x; gid < size; + gid += blockDim.x * gridDim.x) { + __half_t h; + float inval = in[gid] * scale; + + h.f = __anyfloat2half_rn (inval); + short exp_h = (short) ((h.u & 0x7C00) >> 10) - 15; + short sign_h = (h.u & 0x8000); + short mantissa_h = (h.u & 0x03FF); + + unsigned short can_round = ((h.u & 0x7FFF) < 0x4B80) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + int dshift = 0; + + if (exp_h > 7) { + /* Hardware : saturate +/-INF */ + mantissa_h = 0; + exp_h = 16; + is_naninf = 1; + } else if (exp_h < -9) { + exp_h = -15; + mantissa_h = 0; + } else if (exp_h < -6) { + dshift = (-6 - exp_h); + /* handle denormals */ + mantissa_h = mantissa_h >> dshift; + mantissa_h <<= dshift; + } + /* nearest rounding masks */ + unsigned short rnmask = (mantissa_h & grs_bitmask); + unsigned short rnmask_tie = (mantissa_h & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + _rand_xorshft128plus_with_seed (sptr[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + mantissa_h += can_round * is_normal * (rand & 0x7F); + /* stochastic round: denormals --> rne rounding */ + mantissa_h += can_round * is_denorm * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + mantissa_h += can_round * rne_mask * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + mantissa_h += can_round * rnaz_mask * ((rnmask >= 0x0040) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + mantissa_h += can_round * rntz_mask * ((rnmask > 0x0040) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + mantissa_h += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0040) << lshift); + /* round to -INF, if rminf_mask is enabled */ + mantissa_h += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0040) << lshift); + } + } + /* truncation */ + mantissa_h &= mask_mant; + mantissa_h += ((exp_h + 15) << 10); + mantissa_h |= sign_h; + h.u = mantissa_h; + __half2anyfloat (h.f * scale_reciprocal, &out[gid]); + } + } + + template < typename scalar_t > + __global__ void E4M3v2_Kernel (const scalar_t * + __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, int mbits, + int exp_bits, int rmode) { + int non_mant_bits = exp_bits + 1; /* exponent + sign */ + int lshift = 10 - (mbits - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x007F; + unsigned short rne_tie = 0x00C0; + + extern __shared__ float sdata[]; + float scale = in_scale; + + if (block_norm == true) { + absmax_block (in, sdata, size); + __float_t f; + f.f = sdata[0]; + f.u = (f.u & 0x7F800000); + scale = 2 * f.f; + } + float scale_reciprocal = 1.0 / scale; + + for (int gid = (blockIdx.x * blockDim.x) + threadIdx.x; gid < size; + gid += blockDim.x * gridDim.x) { + __half_t h; + float inval = in[gid] * scale; + + h.f = __anyfloat2half_rn (inval); + short exp_h = (short) ((h.u & 0x7C00) >> 10) - 15; + short sign_h = (h.u & 0x8000); + short mantissa_h = (h.u & 0x03FF); + + unsigned short can_round = ((h.u & 0x7FFF) < 0x4B80) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + int dshift = 0; + + if (exp_h > -1) { + /* handle saturation */ + mantissa_h = 0x0380; + exp_h = -1; + can_round = 0; + } + /* nearest rounding masks */ + unsigned short rnmask = (mantissa_h & grs_bitmask); + unsigned short rnmask_tie = (mantissa_h & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + _rand_xorshft128plus_with_seed (sptr[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + mantissa_h += can_round * is_normal * (rand & 0x7F); + /* stochastic round: denormals --> rne rounding */ + mantissa_h += can_round * is_denorm * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + mantissa_h += can_round * rne_mask * + (((rnmask > 0x0040) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + mantissa_h += can_round * rnaz_mask * ((rnmask >= 0x0040) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + mantissa_h += can_round * rntz_mask * ((rnmask > 0x0040) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + mantissa_h += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0040) << lshift); + /* round to -INF, if rminf_mask is enabled */ + mantissa_h += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0040) << lshift); + } + } + /* truncation */ + mantissa_h &= mask_mant; + mantissa_h <<= dshift; + mantissa_h += ((exp_h + 15) << 10); + mantissa_h |= sign_h; + h.u = mantissa_h; + __half2anyfloat (h.f * scale_reciprocal, &out[gid]); + } + } + + template < typename scalar_t > + __global__ void E3M4_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm, + int mbits, + int exp_bits, int rmode) { + int non_mant_bits = exp_bits + 1; /* exponent + sign */ + int lshift = 10 - (mbits - non_mant_bits); + + unsigned short rne_mask = 0; /* round to nearest even mask */ + unsigned short rnaz_mask = 0; /* round to nearest away from zero mask */ + unsigned short rntz_mask = 0; /* round to nearest towards zero mask */ + unsigned short sr_mask = 0; /* stochastic rounding mask */ + unsigned short rpinf_mask = 0; /* round to +INF */ + unsigned short rminf_mask = 0; /* round to -INF */ + + if (rmode == ROUND_RNE) rne_mask = 1; + if (rmode == ROUND_RNAZ) rnaz_mask = 1; + if (rmode == ROUND_RNTZ) rntz_mask = 1; + if (rmode == ROUND_STOCHASTIC) sr_mask = 1; + if (rmode == ROUND_PINF) rpinf_mask = 1; + if (rmode == ROUND_NINF) rminf_mask = 1; + + unsigned short mask_mant = (unsigned short) (0xFFFF << lshift); + unsigned short grs_bitmask = 0x003F; + unsigned short rne_tie = 0x0060; + + extern __shared__ float sdata[]; + float scale = in_scale; + + if (block_norm == true) { + absmax_block (in, sdata, size); + __float_t f; + + f.f = sdata[0]; + f.u = (f.u & 0x7F800000); + scale = 2 * f.f; + } + float scale_reciprocal = 1.0 / scale; + + for (int gid = (blockIdx.x * blockDim.x) + threadIdx.x; gid < size; + gid += blockDim.x * gridDim.x) { + __half_t h; + float inval = scale * in[gid]; + + h.f = __anyfloat2half_rn (inval); + short exp_h = (short) ((h.u & 0x7C00) >> 10) - 15; + short sign_h = (h.u & 0x8000); + short mantissa_h = (h.u & 0x03FF); + + unsigned short can_round = ((h.u & 0x7FFF) < 0x4F80) ? 1 : 0; + unsigned short is_normal = (((h.u & 0x7C00) <= 0x7800) + && ((h.u & 0x7C00) >= 0x0400)) ? 1 : 0; + unsigned short is_denorm = ((h.u & 0x7C00) == 0x0) ? 1 : 0; + unsigned short is_naninf = ((h.u & 0x7C00) == 0x7C00) ? 1 : 0; + + int dshift = 0; + + if (exp_h > 4 || (can_round == 0)) { + /* Software : saturate values above +/-30.0 to +/-30.0 */ + mantissa_h = 0x0380; + exp_h = 4; + can_round = 0; + } else if (exp_h < -6) { + exp_h = -15; + mantissa_h = 0; + } else if (exp_h < -2) { + dshift = (-2 - exp_h); + /* handle denormals */ + mantissa_h = mantissa_h >> dshift; + mantissa_h <<= dshift; + } + + /* nearest rounding masks */ + unsigned short rnmask = (mantissa_h & grs_bitmask); + unsigned short rnmask_tie = (mantissa_h & rne_tie); + + if (is_naninf == 0) { + if (sr_mask) { + /* stochastic with 16 seeds */ + int seed_index = (gid / 16); + unsigned short rand = (unsigned short) + _rand_xorshft128plus_with_seed (sptr[(seed_index % 16)]); + /* apply stochastic rounding before truncation if sr_mask is enabled */ + mantissa_h += can_round * is_normal * (rand & 0x3F); + /* stochastic round: denormals --> rne rounding */ + mantissa_h += can_round * is_denorm * + (((rnmask > 0x0020) || (rnmask_tie == rne_tie)) << lshift); + } else { + /* round to nearest even, if rne_mask is enabled */ + mantissa_h += can_round * rne_mask * + (((rnmask > 0x0020) || (rnmask_tie == rne_tie)) << lshift); + /* round to nearest away from zero, if rnaz_mask is enabled */ + mantissa_h += can_round * rnaz_mask * ((rnmask >= 0x0020) << lshift); + /* round to nearest towards zero, if rntz_mask is enabled */ + mantissa_h += can_round * rntz_mask * ((rnmask > 0x0020) << lshift); + /* round to +INF, if rpinf_mask is enabled */ + mantissa_h += can_round * rpinf_mask * (h.f > 0) * ((rnmask >= 0x0020) << lshift); + /* round to -INF, if rminf_mask is enabled */ + mantissa_h += can_round * rminf_mask * (h.f < 0) * ((rnmask >= 0x0020) << lshift); + } + } + /* truncation */ + mantissa_h &= mask_mant; + mantissa_h += ((exp_h + 15) << 10); + mantissa_h |= sign_h; + h.u = mantissa_h; + __half2anyfloat (h.f * scale_reciprocal, &out[gid]); + } + } + + template < typename scalar_t > + __global__ void FP4_Nearest_Kernel (const scalar_t * __restrict__ in, + scalar_t * __restrict__ out, + const int size, + const scalar_t in_scale, + bool block_norm) { + extern __shared__ float sdata[]; + float scale = in_scale; + + if (block_norm == true) { + absmax_block (in, sdata, size); + /* FP4 max value is 1.0 */ + scale = 1.0/sdata[0]; + } + float scale_reciprocal = 1.0 / scale; + + for (int gid = (blockIdx.x * blockDim.x) + threadIdx.x; gid < size; + gid += blockDim.x * gridDim.x) { + __float_t f; + float inval = scale * in[gid]; + + f.f = inval; + int exp_f = (int)((f.u & 0x7F800000) >> 23) - 127; + int sign_f = (f.u & 0x80000000); + /* see if round up works! */ + if (exp_f < 0 && (exp_f%2)) f.f *= 1.6; + /* saturate */ + if (exp_f > 0) f.u = (sign_f | (127 << 23)); + f.u &= 0xFF800000; + /* extract the new exponent */ + exp_f = (int)(((f.u & 0x7F800000) >> 23) - 127); + /* round up did not work, round down */ + if (exp_f < 0 && (exp_f%2)) f.u = (sign_f | ((exp_f + 126) << 23)); + /* flush values smaller than 2^-12 to zero */ + if (exp_f < -12) f.u = 0; + + out[gid] = (f.f * scale_reciprocal); + } + } + +} + +std::vector fpemu_cuda_forward( + torch::Tensor input, + std::string mode, + const int size, + bool inplace, + float scale, + bool block_norm, + int block_size) { + + float fmax = std::numeric_limits::max(); + if (scale > fmax) { + fprintf(stderr,"Error: Invalid scale factor : %.2e, make sure the scale is not larger than : %.2e\n", scale, fmax); + exit(1); + } + + torch::Tensor output; + if (!inplace ) output = torch::zeros_like(input); + + int sdata_size = 0; + int threads = CUBLOCK_SIZE; + if (block_norm == true && block_size != size) { + if (size%block_size) { + block_norm = false; + } else { + threads = block_size; + sdata_size = threads * sizeof(float); + } + } + const dim3 blocks((size + (threads-1))/threads); + + if (!mode.compare("E5M2_RTZ")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_RTZ", ([&] { + E5M2_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 5, + ROUND_RTZ); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + } else if (!mode.compare("E5M2_RNE")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_RNE", ([&] { + E5M2_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 5, + ROUND_RNE); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + } else if (!mode.compare("E5M2_STOCHASTIC")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_STOCHASTIC", ([&] { + E5M2_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 5, + ROUND_STOCHASTIC); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + } else if (!mode.compare("E5M2_RNAZ")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_RNAZ", ([&] { + E5M2_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 5, + ROUND_RNAZ); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E5M2_RNTZ")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_RNTZ", ([&] { + E5M2_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 5, + ROUND_RNTZ); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E5M2_RPINF")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_RPINF", ([&] { + E5M2_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 5, + ROUND_PINF); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E5M2_RNINF")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_RNINF", ([&] { + E5M2_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 5, + ROUND_NINF); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E5M2_DAZ_RNE")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_DAZ_RNE", ([&] { + E5M2_DAZ_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + 8, + 5, + ROUND_RNE); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E5M2_DAZ_STOCHASTIC")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_DAZ_STOCHASTIC", ([&] { + E5M2_DAZ_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + 8, + 5, + ROUND_STOCHASTIC); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E5M2_DAZ_RNAZ")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_DAZ_RNAZ", ([&] { + E5M2_DAZ_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + 8, + 5, + ROUND_RNAZ); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E5M2_DAZ_RNTZ")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E5M2_DAZ_RNTZ", ([&] { + E5M2_DAZ_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + 8, + 5, + ROUND_RNTZ); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("FLOAT16_RNE")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "FLOAT16_RNE", ([&] { + FLOAT16_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + ROUND_RNE, + 0); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("FLOAT16_STOCHASTIC")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "FLOAT16_STOCHASTIC", ([&] { + FLOAT16_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + ROUND_STOCHASTIC, + 0); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("FLOAT16_DAZ_RNE")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "FLOAT16_DAZ_RNE", ([&] { + FLOAT16_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + ROUND_RNE, + 1); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("BFLOAT16_RNE")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "BFLOAT16_RNE", ([&] { + BFLOAT16_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + ROUND_RNE); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("BFLOAT16_STOCHASTIC")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "BFLOAT16_STOCHASTIC", ([&] { + BFLOAT16_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + ROUND_STOCHASTIC); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E4M3_RNE")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E4M3_RNE", ([&] { + E4M3_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 4, + ROUND_RNE); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E4M3_STOCHASTIC")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E4M3_STOCHASTIC", ([&] { + E4M3_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 4, + ROUND_STOCHASTIC); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E4M3_IEEE_RNE")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E4M3_IEEE_RNE", ([&] { + E4M3_IEEE_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 4, + ROUND_RNE); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E4M3_IEEE_STOCHASTIC")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E4M3_IEEE_STOCHASTIC", ([&] { + E4M3_IEEE_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 4, + ROUND_STOCHASTIC); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E3M4_RNE")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E3M4_RNE", ([&] { + E3M4_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 3, + ROUND_RNE); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("E3M4_STOCHASTIC")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "E3M4_STOCHASTIC", ([&] { + E3M4_Kernel<<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm, + 8, + 3, + ROUND_STOCHASTIC); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + + } else if (!mode.compare("FP4_NEAREST")) { + AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "FP4_NEAREST", ([&] { + FP4_Nearest_Kernel <<>>( + input.data_ptr(), + (inplace) ? input.data_ptr() : output.data_ptr(), + size, + scale, + block_norm); + })); + cudaDeviceSynchronize(); + AT_CUDA_CHECK(cudaGetLastError()); + } + if (!inplace) { + return {output, }; + } else { + return {input, }; + } +} diff --git a/FP8_Emulator/pytquant/test.py b/FP8_Emulator/pytquant/test.py new file mode 100644 index 00000000..22f083c1 --- /dev/null +++ b/FP8_Emulator/pytquant/test.py @@ -0,0 +1,151 @@ +#------------------------------------------------------------------------------ +# Copyright (c) 2023, Intel Corporation - All rights reserved. +# This file is part of FP8-Emulation-Toolkit +# +# SPDX-License-Identifier: BSD-3-Clause +#------------------------------------------------------------------------------ +# Naveen Mellempudi (Intel Corporation) +#------------------------------------------------------------------------------ + +from __future__ import division +from __future__ import print_function + +import argparse +import numpy as np +import torch +import cpp.fpemu as fpemu_cpp +import mpemu +import sys + +sys.path.append('../../mpemu') +# importing +from mpemu.qutils import fpemu_device_fn + + +if torch.cuda.is_available(): + import cuda.fpemu as fpemu_cuda + +def check_equal(first, second, verbose): + if verbose: + print() + for i, (x, y) in enumerate(zip(first, second)): + x = x.cpu().detach().numpy() + y = y.cpu().detach().numpy() + + if verbose: + print("x = {}".format(x.flatten())) + print("y = {}".format(y.flatten())) + print('-' * 80) + + np.testing.assert_allclose(x, y, rtol=0.125, err_msg="Conflict : ", verbose=verbose) + +def print_tensor(first, second, verbose, sparse): + print('printing ...') + if verbose: + print() + for i, (x, y) in enumerate(zip(first, second)): + if sparse : + x = x.cpu().to_dense().detach().numpy() + y = y.cpu().to_dense().detach().numpy() + else : + x = x.cpu().detach().numpy() + y = y.cpu().detach().numpy() + if verbose: + print("x = {}".format(x.flatten())) + print("y = {}".format(y.flatten())) + print('-' * 80) + + +def zero_grad(variables): + for variable in variables: + variable.grad.zero_() + + +def get_grads(variables): + return [var.grad.clone() for var in variables] + +def check_forward(variables, with_cuda, verbose, sparse): + if with_cuda: + if not torch.cuda.is_available(): + print('CUDA is not supported on this platform ... ') + else : + #cuda_values = fpemu_cuda.FPEmuOp.apply(variables.cuda(), "E5M2_RNE") + #cuda_values = fpemu_cuda.FPEmuOp.apply(variables.cuda(), "E5M2_STOCHASTIC") + #cuda_values = fpemu_cuda.FPEmuOp.apply(variables.cuda(), "E4M3_RNE") + #cuda_values = fpemu_cuda.FPEmuOp.apply(variables.cuda(), "E4M3_STOCHASTIC") + #cuda_values = fpemu_cuda.FPEmuOp.apply(variables.cuda(), "E3M4_RNE") + #cuda_values = fpemu_cuda.FPEmuOp.apply(variables.cuda(), "E3M4_STOCHASTIC") + cuda_values = fpemu_cuda.FPEmuOp.apply(variables.cuda(), "FP4_NEAREST")#, False, 1.0, True, 4) + print('Forward: CUDA ... ', end='') + print_tensor(variables, cuda_values, verbose, sparse) + else : + #cpp_values = fpemu_cpp.FPEmuOp.apply(variables, "E5M2_RNE") + #cpp_values = fpemu_cpp.FPEmuOp.apply(variables, "E5M2_STOCHASTIC") + #cpp_values = fpemu_cpp.FPEmuOp.apply(variables, "E4M3_RNE") + #cpp_values = fpemu_cpp.FPEmuOp.apply(variables, "E4M3_STOCHASTIC") + #cpp_values = fpemu_cpp.FPEmuOp.apply(variables, "E3M4_RNE") + #cpp_values = fpemu_cpp.FPEmuOp.apply(variables, "E3M4_STOCHASTIC") + #cpp_values = fpemu_cpp.FPEmuOp.apply(variables, "FP4_NEAREST")#, False, 1.0, True, 4) + #cpp_values = fpemu_cpp.FPEmuOp.apply(variables, "INT8") + #cpp_values = fpemu_cpp.FPEmuOp.apply(variables, "INT4") + cpp_values = fpemu_device_fn(variables, "INT4", inplace=False, scale=1.0) + print('Forward: C++ ... ', end='') + print_tensor(variables, cpp_values, verbose, sparse) + + if not verbose : + print('Done.') + print('Use the option --verbose to see results') + else : + print('Done.') + + +parser = argparse.ArgumentParser() +parser.add_argument('-sp', '--sparse', action='store_true') +parser.add_argument('-c', '--cuda', action='store_true') +parser.add_argument('-v', '--verbose', action='store_true') + +options = parser.parse_args() +if options.cuda: + device = torch.device("cuda") +else: + device = torch.device("cpu") + +kwargs = {'dtype': torch.float32, + 'device': device, + 'requires_grad': True} + +''' +#FP4 values +input = torch.tensor(np.array([[ 0.0000000, + 0.000244140625, + 0.0009766, + 0.0039063, + 0.0156250, + 0.0625000, + 0.2500000, + 1.0000000 ]]), dtype=torch.float32) +''' +input = torch.tensor(np.array([[ 57344.00 , -61440.0, 65504.0, -500.0, + 448.0 , -480.0, 30.0, -31.0, + 26.0 , 15.0, -7.6505613e-00, 5.9832452e-00, + 1.5625032e-02, 3.1725775e-02, 4.3268750e-02, 6.2655000e-02, + 1.9545313e-03, 3.9045625e-03, 5.8845638e-03, 7.8089750e-03, + -6.0151856e-02, 6.9784373e-03, -1.6634936e+03, -7.6505613e-04, + 1.9545313e-03, 3.9045625e-03, 5.8845638e-03, 7.8089750e-03, + 1.5629950e-02, 1.5225650e-02, -3.1256500e-02, 9.3857500e-02, + 7.9284608e-01, 2.8815269e-01, -1.1787039e-01, 6.1035156e-05, + -1.1787039e-06, -1.0749481e-01, -1.3085605e+00, 6.5981364e-01, + -7.0325255e-02, 2.7448297e-01, 5.5694544e-01, -2.3220782e-01, + -5.9746221e-02, 15.23213444 , -0.00004323 , 1.9767435e-04, + -1.2203161e+00, 2.9099861e-01, -7.9642259e-02, 1.3200364e+00, + -1.5196867e+00, -1.2530587e+00, -2.0159689e-03, -1.9767643e+00, + 6.0834163e-04, 7.8943473e-05, 7.8247029e-04, -6.4658634e-05, + -2.3020705e-06, -1.5630834e-05, -7.4762434e-07, 2.1336775e-06]]), dtype=torch.float32) +#''' +#input = torch.randn(4, 16) + +if options.sparse : + check_forward(input.to_sparse(), options.cuda, options.verbose, options.sparse) +else : + check_forward(input, options.cuda, options.verbose, options.sparse) + diff --git a/READMEFP8.md b/READMEFP8.md new file mode 100644 index 00000000..968b97b8 --- /dev/null +++ b/READMEFP8.md @@ -0,0 +1,82 @@ +# FP8 Emulation Toolkit +## Introduction +This repository provides PyTorch tools to emulate the new `FP8` formats on top of existing floating point hardware from Intel (FP32) and NVIDIA (FP16). In addition to the two formats `E5M2` and `E4M3` defined in the joint specification from ARM-Intel-NVIDIA, the toolkit also suports a third variant named `E3M4` which follows the guidelines established for `E4M3` format. + +Following table shows the binary formats and the numeric range: + +| | E5M2 | E4M3 | E3M4 | +| -------------- | ---------------------------------------------------------------- | ---------------------------------------------------------------- | ---------------------------------------------------------------- | +| Exponent Bias | 15 | 7 | 3 | +| Infinities | S.11111.002 | N/A | N/A | +| NaNs | S.11111.{01, 10, 11}2 | S.1111.1112 | S.111.11112 | +| Zeros | S.00000.002 | S.0000.0002 | S.000.00002 | +| Max normal | S.11110.112=1.75 * 215=57344.0 | S.1111.1102=1.75 * 28=448.0 | S.111.11102=1.875 * 24=30.0 | +| Min normal | S.00001.002=2-14=6.1e-05 | S.0001.0002=2-6=1.5e-02 | S.001.00002=2-2=2.5e-01 | +| Max subnormal | S.00000.112=0.75 * 2-14=4.5e-05 | S.0000.1112=0.875 * 2-6=1.3e-02 | S.000.11112=0.9375 * 2-2=2.3e-01 | +| Min subnormal | S.00000.012=2-16=1.5e-05 | S.0000.0012=2-9=1.9e-03 | S.000.00012=2-6=1.5e-02 | + +![DataFormats](./docs/formats.png) + +## Installation + +Follow the instructions below to install FP8 Emulation Toolkit in a Python virtual environment. +Alternatively, this installation can also be performed in a docker environment. + +### Requirements +Install or upgrade the following packages on your linux machine. + +* Python >= 3.8.5 +* CUDA >= 11.1 +* gcc >= 8.4.0 + +Make sure these versions are reflected in the `$PATH` + +#### Target Hardware +* CPU >= Icelake Xeon +* GPU >= V100 + +### Create a Python virtual environment +``` +$ python3 -m ~/py-venv +$ cd ~/py-venv +$ source bin/activate +$ pip3 install --upgrade pip3 +``` +### Clone and install FP8 Emulation Toolkit +``` +$ git clone https://github.com/IntelLabs/FP8-Emulation-Toolkit.git +$ cd FP8-Emulation-Toolkit +$ pip3 install -r requirements.txt +$ python setup.py install +``` + +## Usage Examples +The emulated FP8 formats can be experimented with by integrated them into standard deep learning flows. Follow the links below for detailed instructions and code samples for exploring training and inference flows using FP8 data formats. + +* [Post-training quantization](./examples/inference) +* [Mixed precision training](./examples/training) + + +## Related Work +This implementation is based on the following research. Check out the source material for more details on the training and inference methods. + +``` +@misc{mellempudi2019mixed, + title={Mixed Precision Training With 8-bit Floating Point}, + author={Naveen Mellempudi and Sudarshan Srinivasan and Dipankar Das and Bharat Kaul}, + year={2019}, + eprint={1905.12334}, + archivePrefix={arXiv}, + primaryClass={cs.LG} +} +``` +``` +@misc{micikevicius2022fp8, + title={FP8 Formats for Deep Learning}, + author={Paulius Micikevicius and Dusan Stosic and Neil Burgess and Marius Cornea and Pradeep Dubey and Richard Grisenthwaite and Sangwon Ha and Alexander Heinecke and Patrick Judd and John Kamalu and Naveen Mellempudi and Stuart Oberman and Mohammad Shoeybi and Michael Siu and Hao Wu}, + year={2022}, + eprint={2209.05433}, + archivePrefix={arXiv}, + primaryClass={cs.LG} +} +``` diff --git a/application/imagenet_example/PTQ/ptq/ptq.py b/application/imagenet_example/PTQ/ptq/ptq.py index 39908de0..bd56b2b7 100644 --- a/application/imagenet_example/PTQ/ptq/ptq.py +++ b/application/imagenet_example/PTQ/ptq/ptq.py @@ -34,7 +34,7 @@ def get_quantize_model(model, config): extra_prepare_dict = {} if not hasattr( config, 'extra_prepare_dict') else config.extra_prepare_dict return prepare_by_platform( - model, backend_type, extra_prepare_dict) + model, backend_type, prepare_custom_config_dict=extra_prepare_dict) def deploy(model, config): diff --git a/application/imagenet_example/PTQ/ptq/ptq_main.py b/application/imagenet_example/PTQ/ptq/ptq_main.py new file mode 100644 index 00000000..3ac20967 --- /dev/null +++ b/application/imagenet_example/PTQ/ptq/ptq_main.py @@ -0,0 +1,189 @@ +import sys +import os +import time +sys.path.append(os.path.abspath('.')) +print(sys.path) + +import torchvision.models as models +import numpy as np +import time +import argparse +from data.imagenet import load_data +from models import load_model +from utils import parse_config, seed_all, evaluate +from mqbench.prepare_by_platform import prepare_by_platform, BackendType +from mqbench.advanced_ptq import ptq_reconstruction +from mqbench.convert_deploy import convert_deploy + +model_names = sorted(name for name in models.__dict__ + if name.islower() and not name.startswith("__") + and callable(models.__dict__[name])) + +parser = argparse.ArgumentParser(description='PyTorch ImageNet Training') +parser.add_argument('--data_path', metavar='DIR', + help='path to dataset', required=True) +parser.add_argument('-a', '--arch', metavar='ARCH', default='resnet18', + choices=model_names, + help='model architecture: ' + + ' | '.join(model_names) + + ' (default: resnet18)') +parser.add_argument('-b', '--batch-size', default=64, type=int, + metavar='N', + help='mini-batch size (default: 64), this is the total ' + 'batch size of all GPUs on the current node when ' + 'using Data Parallel or Distributed Data Parallel') +parser.add_argument('--backend', type=str, choices=['academic', 'tengine_u8', 'tensorrt', 'nnie', 'ppl', 'snpe', 'sophgo_tpu', 'openvino', 'tensorrt_nlp'], default='sophgo_tpu') +parser.add_argument('--cali-batch-num', default=16, type=int, + metavar='N', help='set calibration batch num (default: 16)') +parser.add_argument('--output_path', type=str, default=None) +parser.add_argument('--seed', default=None, type=int, + help='seed for initializing training. ') +parser.add_argument('--pretrained', dest='pretrained', action='store_true', + help='use pre-trained model') +parser.add_argument('--quantize_type', metavar='DIR', + help='set quantize_type', type=str, default='naive_ptq') +parser.add_argument('--deploy', action='store_true') + +BackendMap = { + 'academic': BackendType.Academic, + 'sophgo_tpu': BackendType.Sophgo_TPU, + 'nnie': BackendType.NNIE, + 'tensorrt_nlp': BackendType.Tensorrt_NLP, + 'ppl': BackendType.PPLW8A16, + 'openvino': BackendType.OPENVINO, + 'snpe': BackendType.SNPE, + 'vitis': BackendType.Vitis, + 'tengine_u8': BackendType.Tengine_u8, + 'tensorrt': BackendType.Tensorrt +} + + +def load_calibrate_data(train_loader, cali_batchsize): + cali_data = [] + for i, batch in enumerate(train_loader): + cali_data.append(batch[0]) + if i + 1 == cali_batchsize: + break + return cali_data + +def get_quantize_model(model, args): + backend_type = BackendType.Academic if not hasattr( + args, 'backend') else BackendMap[args.backend] + + if backend_type == BackendType.Academic: + extra_prepare_dict = { + "extra_qconfig_dict": { + 'w_observer': 'MinMaxObserver', + 'a_observer': 'EMAMinMaxObserver', + "w_fakequantize": "FixedFakeQuantize", + "a_fakequantize": "FixedFakeQuantize", + 'w_qscheme': { 'bit': 8, + 'symmetry': False, + 'per_channel': True, + 'pot_scale': False }, + 'a_qscheme': { 'bit': 8, + 'symmetry': False, + 'per_channel': False, + 'pot_scale': False } + } + } + elif backend_type == BackendType.Sophgo_TPU: + extra_prepare_dict = { + "extra_qconfig_dict": { + 'w_observer': 'MinMaxObserver', + 'a_observer': 'EMAMinMaxObserver',}} + else: + extra_prepare_dict = {} + return prepare_by_platform( + model, backend_type, prepare_custom_config_dict=extra_prepare_dict) + + +def deploy(model, args): + backend_type = BackendType.Academic if not hasattr( + args, 'backend') else BackendMap[args.backend] + output_path = './' if not hasattr( + args, 'output_path') else args.output_path + model_name = args.arch + deploy_to_qlinear = False if not hasattr( + args, 'deploy_to_qlinear') else args.deploy_to_qlinear + + convert_deploy(model, backend_type, { + 'input': [1, 3, 224, 224]}, output_path=output_path, model_name=model_name, deploy_to_qlinear=deploy_to_qlinear) + +def main(): + time_start = time.time() + args = parser.parse_args() + + # set output path + if args.output_path is None: + args.output_path = './' + args.output_path=os.path.join(args.output_path, args.arch) + os.system('mkdir -p {}'.format(args.output_path)) + + # set seed first + if args.seed is not None: + seed_all(args.seed) + + # create_model + if args.pretrained: + print("=> using pre-trained model '{}'".format(args.arch)) + model = models.__dict__[args.arch](pretrained=True) + else: + print("=> creating model '{}'".format(args.arch)) + model = models.__dict__[args.arch]() + model.cuda() + + # load_data + train_loader, val_loader = load_data(path=args.data_path , batch_size=args.batch_size) + + # evaluate + evaluate(val_loader, model) + + # get quantize model + model = get_quantize_model(model, args) + model.cuda() + print('>>>>>model after insert fake quantization node: ', model) + + # ptq + if args.quantize_type == 'advanced_ptq': + print('begin calibration now!') + cali_data = load_calibrate_data(train_loader, cali_batchsize=args.cali_batch_num) + from mqbench.utils.state import enable_quantization, enable_calibration_woquantization + # do activation and weight calibration seperately for quick MSE per-channel for weight one + model.eval() + import torch + with torch.no_grad(): + enable_calibration_woquantization(model, quantizer_type='act_fake_quant') + for batch in cali_data: + model(batch.cuda()) + enable_calibration_woquantization(model, quantizer_type='weight_fake_quant') + model(cali_data[0].cuda()) + print('begin advanced PTQ now!') + if hasattr(args, 'reconstruction'): + model = ptq_reconstruction( + model, cali_data, args.reconstruction) + enable_quantization(model) + evaluate(val_loader, model) + if args.deploy: + deploy(model, args) + elif args.quantize_type == 'naive_ptq': + print('begin calibration now!') + cali_data = load_calibrate_data(train_loader, cali_batchsize=args.cali_batch_num) + from mqbench.utils.state import enable_quantization, enable_calibration + model.eval() + enable_calibration(model) + for batch in cali_data: + model(batch.cuda()) + print('begin quantization now!') + enable_quantization(model) + evaluate(val_loader, model) + if args.deploy: + deploy(model, args) + else: + print("The quantize_type must in 'naive_ptq' or 'advanced_ptq',") + print("and 'advanced_ptq' need reconstruction configration.") + time_end = time.time() + print('totally time is ', time_end-time_start) + +if __name__ == '__main__': + main() \ No newline at end of file diff --git a/application/imagenet_example/main.py b/application/imagenet_example/main.py index 30349712..3301bf38 100644 --- a/application/imagenet_example/main.py +++ b/application/imagenet_example/main.py @@ -1,9 +1,18 @@ +#-- coding: gb2312 -- +import sys +import os +import time +sys.path.append(os.path.abspath('.')) +print(sys.path) + import argparse import os import random import shutil import time import warnings +import numpy as np +import copy import torch import torch.nn as nn @@ -17,7 +26,7 @@ import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models -from mqbench.convert_deploy import convert_deploy +from mqbench.convert_deploy import convert_deploy, convert_onnx from mqbench.prepare_by_platform import prepare_by_platform, BackendType from mqbench.utils.state import enable_calibration, enable_quantization, disable_all @@ -25,6 +34,7 @@ if name.islower() and not name.startswith("__") and callable(models.__dict__[name])) +cali_batch_size = 10 parser = argparse.ArgumentParser(description='PyTorch ImageNet Training') parser.add_argument('--train_data', metavar='DIR', help='path to dataset', required=True) @@ -80,14 +90,23 @@ 'multi node data parallel training') parser.add_argument('--model_path', type=str, default=None) -parser.add_argument('--backend', type=str, choices=['tengine_u8', 'tensorrt', 'nnie', 'ppl', 'snpe'], default='tensorrt') +parser.add_argument('--output_path', type=str, default=None) +parser.add_argument('--backend', type=str, choices=['tengine_u8', 'tensorrt', 'nnie', 'ppl', 'snpe', 'sophgo_tpu', 'openvino', 'tensorrt_nlp'], default='sophgo_tpu') parser.add_argument('--optim', type=str, default='sgd') parser.add_argument('--not-quant', action='store_true') parser.add_argument('--deploy', action='store_true') +parser.add_argument('--fast_test', action='store_true') +parser.add_argument('--cpu', action='store_true') +parser.add_argument('--pre_eval_and_export', action='store_true') +parser.add_argument('--deploy_batch_size', default=1, type=int, help='deploy_batch_size.') + BackendMap = {'tensorrt': BackendType.Tensorrt, + 'sophgo_tpu': BackendType.Sophgo_TPU, 'nnie': BackendType.NNIE, + 'tensorrt_nlp': BackendType.Tensorrt_NLP, 'ppl': BackendType.PPLW8A16, + 'openvino': BackendType.OPENVINO, 'snpe': BackendType.SNPE, 'vitis': BackendType.Vitis, 'tengine_u8': BackendType.Tengine_u8} @@ -95,7 +114,14 @@ best_acc1 = 0 def main(): + time_start = time.time() args = parser.parse_args() + + if args.output_path is None: + args.output_path = './' + args.output_path=os.path.join(args.output_path, args.arch) + os.system('mkdir -p {}'.format(args.output_path)) + args.quant = not args.not_quant args.backend = BackendMap[args.backend] @@ -130,6 +156,8 @@ def main(): # Simply call main_worker function main_worker(args.gpu, ngpus_per_node, args) + time_end = time.time() + print('totally time is ', time_end-time_start) def main_worker(gpu, ngpus_per_node, args): global best_acc1 @@ -155,14 +183,73 @@ def main_worker(gpu, ngpus_per_node, args): else: print("=> creating model '{}'".format(args.arch)) model = models.__dict__[args.arch]() + # print('ori module:', model) + + # model = models._utils.IntermediateLayerGetter(model,{'layer1': 'feat1', 'layer3': 'feat2'}) + # print(model(out_put)) + # for internal cluster if args.model_path: state_dict = torch.load(args.model_path) print(f'load pretrained checkpoint from: {args.model_path}') model.load_state_dict(state_dict) + + train_loader, train_sampler, val_loader, cali_loader = prepare_dataloader(args) + criterion = nn.CrossEntropyLoss().cuda(args.gpu) + if args.gpu is not None: + model = model.cuda(args.gpu) + else: + model = model.cpu() + + if args.pre_eval_and_export: + print('ԭʼonnxģ;') + validate(val_loader, model.eval(), criterion, args) #δִmodel.cuda()ᱨ + + kwargs = { + 'input_shape_dict': {'data': [args.deploy_batch_size, 3, 224, 224]}, + 'output_path': args.output_path, + 'model_name': args.arch, + 'dummy_input': None, + 'onnx_model_path': os.path.join(args.output_path, '{}_ori.onnx'.format(args.arch)), + } + module_tmp = copy.deepcopy(model) + module_tmp = module_tmp.cpu() + convert_onnx(module_tmp.eval(), **kwargs) + del module_tmp + model = model.train() #prepareǰһҪtrainģʽ + # quantize model if args.quant: - model = prepare_by_platform(model, args.backend) + prepare_custom_config_dict= { + # 'extra_qconfig_dict':{'w_fakequantize':'PACTFakeQuantize'}, + # 'work_mode':'int4_and_int8_mix', + + # 'work_mode':'all_int4_qat', #int4_and_int8_mix + # 'extra_qconfig_dict': { + # 'w_qscheme': { + # 'bit': 4, # custom bitwidth for weight, + # 'symmetry': True, # custom whether quant is symmetric for weight, + # 'per_channel': True, # custom whether quant is per-channel or per-tensor for weight, + # 'pot_scale': False, # custom whether scale is power of two for weight. + # }, + # 'a_qscheme': { + # 'bit': 4, # custom bitwidth for activation, + # 'symmetry': True, # custom whether quant is symmetric for activation, + # 'per_channel': False, # custom whether quant is per-channel or per-tensor for activation, + # 'pot_scale': False, # custom whether scale is power of two for activation. + # } + # } + } + model = prepare_by_platform(model, args.backend, input_shape_dict = {'data': [args.deploy_batch_size, 3, 224, 224]}) + print('>>>>>prepared module:', model) + + if args.fast_test: + convert_deploy(model.eval(), args.backend, input_shape_dict= + {'data': [args.deploy_batch_size, 3, 224, 224]}, + model_name='{}_mqmoble'.format(args.arch), + # work_mode ='int4_and_int8_mix', + output_path=args.output_path) + if not torch.cuda.is_available(): print('using CPU, this will be slow') elif args.distributed: @@ -192,10 +279,13 @@ def main_worker(gpu, ngpus_per_node, args): model.features = torch.nn.DataParallel(model.features) model.cuda() else: - model = torch.nn.DataParallel(model).cuda() + if args.cpu: + model = model.cpu() + else: + # model = torch.nn.DataParallel(model).cuda() #ᵼgpuѵģ޷resumecpu + model = model.cuda() # define loss function (criterion) and optimizer - criterion = nn.CrossEntropyLoss().cuda(args.gpu) if args.optim == 'sgd': optimizer = torch.optim.SGD(model.parameters(), args.lr, momentum=args.momentum, @@ -206,8 +296,15 @@ def main_worker(gpu, ngpus_per_node, args): weight_decay=args.weight_decay, amsgrad=False) - # prepare dataset - train_loader, train_sampler, val_loader, cali_loader = prepare_dataloader(args) + if args.quant: + enable_calibration(model) + calibrate(cali_loader, model, args) + + if args.quant: + enable_quantization(model) + + cudnn.benchmark = True + # cudnn.deterministic = True # # optionally resume from a checkpoint if args.resume: @@ -218,6 +315,8 @@ def main_worker(gpu, ngpus_per_node, args): else: # Map model to be loaded to specified single gpu. loc = 'cuda:{}'.format(args.gpu) + if args.cpu: + loc = 'cpu' checkpoint = torch.load(args.resume, map_location=loc) args.start_epoch = checkpoint['epoch'] best_acc1 = checkpoint['best_acc1'] @@ -226,6 +325,8 @@ def main_worker(gpu, ngpus_per_node, args): best_acc1 = best_acc1.to(args.gpu) state_dict = checkpoint['state_dict'] + # if args.cpu: + # model = torch.nn.DataParallel(model).cpu() model_dict = model.state_dict() if 'module.' in list(state_dict.keys())[0] and 'module.' not in list(model_dict.keys())[0]: for k in list(state_dict.keys()): @@ -237,26 +338,22 @@ def main_worker(gpu, ngpus_per_node, args): .format(args.resume, checkpoint['epoch'], best_acc1)) else: print("=> no checkpoint found at '{}'".format(args.resume)) - elif args.quant: - enable_calibration(model) - calibrate(cali_loader, model, args) + exit(1) - cudnn.benchmark = True - - if args.quant: - enable_quantization(model) - - if args.quant and args.deploy: - convert_deploy(model.eval(), args.backend, input_shape_dict={'data': [10, 3, 224, 224]}) - return - - if args.evaluate: - if args.quant: + if args.evaluate: + print('resumeģ;') from mqbench.convert_deploy import convert_merge_bn - convert_merge_bn(model.eval()) - validate(val_loader, model, criterion, args) - return - + module_tmp2 = copy.deepcopy(model) + convert_merge_bn(module_tmp2.eval()) + validate(val_loader, module_tmp2, criterion, args) + del module_tmp2 + gen_test_ref_data(cali_loader, model, args) + convert_deploy(model.eval(), args.backend, input_shape_dict={'data': [args.deploy_batch_size, 3, 224, 224]}, + model_name='{}_mqmoble'.format(args.arch), output_path=args.output_path) + exit(0) + + if args.fast_test: + args.epochs = 1 for epoch in range(args.start_epoch, args.epochs): if args.distributed: train_sampler.set_epoch(epoch) @@ -266,21 +363,48 @@ def main_worker(gpu, ngpus_per_node, args): train(train_loader, model, criterion, optimizer, epoch, args) # evaluate on validation set + if epoch == args.epochs - 1: + print('qatѵĴڵeval:') + else: + print(f'epoch{epoch}ѵeval:') acc1 = validate(val_loader, model, criterion, args) - # remember best acc@1 and save checkpoint - is_best = acc1 > best_acc1 - best_acc1 = max(acc1, best_acc1) + # # remember best acc@1 and save checkpoint + # is_best = acc1 > best_acc1 + # best_acc1 = max(acc1, best_acc1) + + # if not args.multiprocessing_distributed or (args.multiprocessing_distributed + # and args.rank % ngpus_per_node == 0): + # save_checkpoint({ + # 'epoch': epoch + 1, + # 'arch': args.arch, + # 'state_dict': model.state_dict(), + # 'best_acc1': best_acc1, + # 'optimizer' : optimizer.state_dict(), + # }, is_best, filename=os.path.join(args.output_path, 'checkpoint.pth.tar')) + + print('disable_allԾ:') + disable_all(model) + validate(val_loader, model, criterion, args) + enable_quantization(model) + + gen_test_ref_data(cali_loader, model, args) + convert_deploy(model.eval(), args.backend, input_shape_dict= + {'data': [args.deploy_batch_size, 3, 224, 224]}, + model_name='{}_mqmoble'.format(args.arch), + # work_mode ='int4_and_int8_mix', + output_path=args.output_path) + + '''model_path = os.path.join(args.output_path, '{}.pt'.format('{}_mqmoble'.format(args.arch))) + model_pt = torch.load(model_path) + if args.gpu is not None: + model_pt = model_pt.cuda(args.gpu) + else: + model_pt = model_pt.cpu() + print('load fused bn ptԾ:') + validate(val_loader, model_pt, criterion, args) - if not args.multiprocessing_distributed or (args.multiprocessing_distributed - and args.rank % ngpus_per_node == 0): - save_checkpoint({ - 'epoch': epoch + 1, - 'arch': args.arch, - 'state_dict': model.state_dict(), - 'best_acc1': best_acc1, - 'optimizer' : optimizer.state_dict(), - }, is_best) + validate_onnx(criterion, args)''' def prepare_dataloader(args): traindir = os.path.join(args.train_data, 'train') @@ -306,8 +430,7 @@ def prepare_dataloader(args): train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None), num_workers=args.workers, pin_memory=True, sampler=train_sampler) - cali_batch_size = 10 - cali_batch = 10 + cali_batch = 20 cali_dataset = torch.utils.data.Subset(train_dataset, indices=torch.arange(cali_batch_size * cali_batch)) cali_loader = torch.utils.data.DataLoader(cali_dataset, batch_size=cali_batch_size, shuffle=False, num_workers=args.workers, pin_memory=True) @@ -324,6 +447,24 @@ def prepare_dataloader(args): return train_loader, train_sampler, val_loader, cali_loader +def prepare_dataloader_batch(args, batch_size): + valdir = os.path.join(args.val_data, 'val') + normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], + std=[0.229, 0.224, 0.225]) + + val_loader = torch.utils.data.DataLoader( + datasets.ImageFolder(valdir, transforms.Compose([ + transforms.Resize(256), + transforms.CenterCrop(224), + transforms.ToTensor(), + normalize, + ])), + batch_size=batch_size, shuffle=False, + num_workers=args.workers, pin_memory=True) + + return val_loader + + def calibrate(cali_loader, model, args): model.eval() print("Start calibration ...") @@ -337,6 +478,96 @@ def calibrate(cali_loader, model, args): print("End calibration.") return +def get_node_name_by_module_name(qname, model): + nodes = list(model.graph.nodes) + modules = dict(model.named_modules()) + for node in nodes: + if node.target in modules and qname == node.target: + return node.name + +def get_node_input_by_module_name(qname, model): + nodes = list(model.graph.nodes) + modules = dict(model.named_modules()) + post_str = '_post_act_fake_quantizer' + input_str = '_input_act_fake_quantizer' + scale_name = None + for node in nodes: + if node.target in modules and qname == node.target: + print(f'{qname} input:', node.args[0].name) + if post_str in node.args[0].name: + scale_name = node.args[0].name + break + elif input_str in node.args[0].name: + node2 = node.args[0] + print(f'{node.args[0].name}.input:', node2.args[0].name) + if post_str in node2.args[0].name: + scale_name = node2.args[0].name + break + elif 'x' == node2.args[0].name: + return 'data' + break + if scale_name is not None: + return scale_name[:len(scale_name)-len(post_str)] + else: + return '' + +layer_names = [] +features_out_hook = {} +i = 0 +def hook(module, fea_in, fea_out): + global i + if i >= len(layer_names): + return None + name = layer_names[i] + i += 1 + global features_out_hook + features_out_hook[name] = fea_out.cpu().numpy() + return None + +def gen_test_ref_data(cali_loader, model, args): + return + model.eval() + global layer_names + hook_handles = [] + input_data = {} + # exclude_module = ['fake_quantize', 'observer', 'torch.fx', 'batchnorm', 'torch.nn.modules.module.Module'] + for name, child in model.named_modules(): + # if not any([i in str(type(child)) for i in exclude_module]): + if name.endswith('_act_fake_quantizer'): + # if '_dup' in name: + # name = name[:-5] + # layer_names.append(name.replace('.','_')) + # output = get_node_name_by_module_name(name, model) + # input = get_node_input_by_module_name(name, model) + # print(f'name:{name}, output:{output}, input:{input}') + # if input != '': + layer_names.append(name) + print(f"add hook on {name} for {str(type(child))}") + hd = child.register_forward_hook(hook=hook) + hook_handles.append(hd) + print('layer_names:', layer_names) + if args.cpu: + model = model.cpu() + with torch.no_grad(): + for i, (images, target) in enumerate(cali_loader): + if args.gpu is not None: + images = images.cuda(args.gpu, non_blocking=True) + else: + images = images.cpu() + output = model(images) + print("gen_test_ref_data ==> ", i+1) + if i == 0: + input_data['data'] = images.cpu().numpy() + np.savez(os.path.join(args.output_path, 'input_data.npz'), **input_data) + global features_out_hook + features_out_hook['data'] = images.cpu().numpy() + np.savez(os.path.join(args.output_path, 'layer_outputs.npz'), **features_out_hook) + for hd in hook_handles: + hd.remove() + break + print("End gen_test_ref_data.") + return + def train(train_loader, model, criterion, optimizer, epoch, args): batch_time = AverageMeter('Time', ':6.3f') data_time = AverageMeter('Data', ':6.3f') @@ -356,9 +587,8 @@ def train(train_loader, model, criterion, optimizer, epoch, args): # measure data loading time data_time.update(time.time() - end) - if args.gpu is not None: + if args.gpu is not None and torch.cuda.is_available(): images = images.cuda(args.gpu, non_blocking=True) - if torch.cuda.is_available(): target = target.cuda(args.gpu, non_blocking=True) # compute output @@ -383,6 +613,15 @@ def train(train_loader, model, criterion, optimizer, epoch, args): if i % args.print_freq == 0: progress.display(i) + # # ѵ̲Ƿ쳣 + # for param in model.named_parameters(): + # sum = torch.isnan(param[1]).sum() + # if sum > 0: + # print(param[0], 'has Nan', param[1].shape, 'sum:', sum) + + if args.fast_test: + if i % 64 == 0: + break def validate(val_loader, model, criterion, args): batch_time = AverageMeter('Time', ':6.3f') @@ -396,14 +635,21 @@ def validate(val_loader, model, criterion, args): # switch to evaluate mode model.eval() + if args.cpu: + model = model.cpu() + with torch.no_grad(): end = time.time() for i, (images, target) in enumerate(val_loader): - if args.gpu is not None: - images = images.cuda(args.gpu, non_blocking=True) - if torch.cuda.is_available(): - target = target.cuda(args.gpu, non_blocking=True) + if not args.cpu: + if args.gpu is not None: + images = images.cuda(args.gpu, non_blocking=True) + if torch.cuda.is_available(): + target = target.cuda(args.gpu, non_blocking=True) + else: + images = images.cpu() + target = target.cpu() # compute output output = model(images) @@ -421,7 +667,72 @@ def validate(val_loader, model, criterion, args): if i % args.print_freq == 0: progress.display(i) + + if args.fast_test: + if i % 100 == 0: + break + + # TODO: this should also be done with the ProgressMeter + print(' * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}' + .format(top1=top1, top5=top5)) + + return top1.avg + + +def validate_onnx(criterion, args): + import onnxruntime as rt + val_loader = prepare_dataloader_batch(args, args.deploy_batch_size) + model_path = os.path.join(args.output_path, '{}_deploy_model.onnx'.format('{}_mqmoble'.format(args.arch))) + sess = rt.InferenceSession(model_path, providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']) + input_name = sess.get_inputs()[0].name + + batch_time = AverageMeter('Time', ':6.3f') + losses = AverageMeter('Loss', ':.4e') + top1 = AverageMeter('Acc@1', ':6.2f') + top5 = AverageMeter('Acc@5', ':6.2f') + progress = ProgressMeter( + len(val_loader), + [batch_time, losses, top1, top5], + prefix='Test: ') + + + with torch.no_grad(): + end = time.time() + for i, (images, target) in enumerate(val_loader): + if not args.cpu: + if args.gpu is not None: + images = images.cuda(args.gpu, non_blocking=True) + if torch.cuda.is_available(): + target = target.cuda(args.gpu, non_blocking=True) + else: + images = images.cpu() + target = target.cpu() + + # compute output + # output = model(images) + output = sess.run(None, {input_name:images.clone().detach().cpu().numpy()}) + output = torch.from_numpy(output[0]).cuda(args.gpu, non_blocking=True) + loss = criterion(output, target) + + # measure accuracy and record loss + acc1, acc5 = accuracy(output, target, topk=(1, 5)) + losses.update(loss.item(), images.size(0)) + top1.update(acc1[0], images.size(0)) + top5.update(acc5[0], images.size(0)) + + # measure elapsed time + batch_time.update(time.time() - end) + end = time.time() + + if i % args.print_freq == 0: + progress.display(i) + + # if args.fast_test: + # if i % 100 == 0: + # break + # TODO: this should also be done with the ProgressMeter + print('deploy_model.onnxдonnxruntimeԾ:') print(' * Acc@1 {top1.avg:.3f} Acc@5 {top5.avg:.3f}' .format(top1=top1, top5=top5)) @@ -431,7 +742,7 @@ def validate(val_loader, model, criterion, args): def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'): torch.save(state, filename) if is_best: - shutil.copyfile(filename, 'model_best.pth.tar') + shutil.copyfile(filename, filename+'_best') class AverageMeter(object): diff --git a/application/imagenet_example/vscode_launch_file.txt b/application/imagenet_example/vscode_launch_file.txt new file mode 100644 index 00000000..ca45a684 --- /dev/null +++ b/application/imagenet_example/vscode_launch_file.txt @@ -0,0 +1,43 @@ +{ + "version": "0.2.0", + "configurations": [ + { + "name": "Python: main.py", + "type": "python", + "request": "launch", + // "preLaunchTask": "clear", //??? + //"program": "${file}", + "program": "imagenet_example/main.py", + // "program": "application/imagenet_example/PTQ/ptq/ptq.py", + "console": "integratedTerminal", + "justMyCode": false, + // "env": {"CUDA_VISIBLE_DEVICES":"0,1,2,3"}, + "args": [ + // "--arch=shufflenet_v2_x0_5", "--batch-size=320", + // "--arch=mobilenet_v2", "--batch-size=64", + // "--arch=resnet18", "--batch-size=128", + // "--arch=vgg11_bn", "--batch-size=32", + // "--arch=resnet50", "--batch-size=32", + // "--arch=squeezenet1_1", "--batch-size=128", + "--arch=mobilenet_v3_small", "--batch-size=128", + "--epochs=1","--lr=1e-4", + "--gpu=0", + "--pretrained", + "--evaluate", + // "--resume=tmp_path1/squeezenet1_1/checkpoint.pth.tar_best", + "--output_path=tmp_test", + // "--output_path=/workspace/tmp_path_1008", + "--fast_test", + "--backend=sophgo_tpu", + "--optim=sgd", + // "--backend=tensorrt_nlp", + // "--deploy", + // "--cpu", + // "--pre_eval_and_export", + "--train_data=/data/imagenet/for_train_val/", + "--val_data=/data/imagenet/for_train_val/" + ] + // "args": ["--config=application/imagenet_example/PTQ/configs/qdrop/r18_2_4.yaml"] + } + ] +} \ No newline at end of file diff --git a/application/nlp_example/config-gptq.yaml b/application/nlp_example/config-gptq.yaml new file mode 100644 index 00000000..ede98a08 --- /dev/null +++ b/application/nlp_example/config-gptq.yaml @@ -0,0 +1,145 @@ +quant: + a_qconfig: + quantizer: FixedFakeQuantize + observer: MinMaxObserver + bit: 8 + symmetric: False + per_channel: False + w_qconfig: + quantizer: GPTQFakeQuantize + observer: MinMaxObserver + bit: 8 + symmetric: True + per_channel: False + calibrate: 128 + pot_scale: False + backend: academic + +# quant: +# a_qconfig: +# quantizer: FixedFakeQuantize +# observer: MinMaxObserver +# bit: 8 +# symmetric: False +# per_channel: False +# w_qconfig: +# quantizer: FixedFakeQuantize +# observer: MinMaxObserver +# bit: 8 +# symmetric: True +# per_channel: False +# calibrate: 64 +# pot_scale: False +# backend: academic + + +# quant: +# a_qconfig: +# quantizer: GPTQFakeQuantize +# observer: MinMaxObserver +# bit: 8 +# symmetric: False +# per_channel: False +# w_qconfig: +# quantizer: GPTQFakeQuantize +# observer: MinMaxObserver +# bit: 8 +# symmetric: True +# per_channel: True +# calibrate: 64 +# pot_scale: False +# backend: academic + +# quant: +# a_qconfig: +# quantizer: FixedFakeQuantize +# observer: EMAMinMaxObserver +# bit: 8 +# symmetric: False +# per_channel: False +# w_qconfig: +# quantizer: FixedFakeQuantize +# observer: MinMaxObserver +# bit: 8 +# symmetric: True +# per_channel: True +# calibrate: 64 +# pot_scale: False +# backend: academic + +# quant: +# a_qconfig: +# quantizer: LearnableFakeQuantize +# observer: LSQObserver +# bit: 8 +# symmetric: False +# per_channel: False +# w_qconfig: +# quantizer: LearnableFakeQuantize +# observer: LSQObserver +# bit: 8 +# symmetric: True +# per_channel: True +# calibrate: 64 +# pot_scale: False +# backend: academic + +data: + task_name: mrpc + dataset_name: mrpc + dataset_config_name: null + max_seq_length: 128 + overwrite_cache: False # Overwrite the cached preprocessed datasets or not. + pad_to_max_length: True # Whether to pad all samples to 'max_seq_length' + # If False, will pad the samples dynamically when batching to the maximum length in the batch." + max_train_samples: null + max_eval_samples: null + max_predict_samples: null + train_file: null + validation_file: null + test_file: null + +model: + type: bert + model_name_or_path: Intel/bert-base-uncased-mrpc + config_name: null # pretrained config name or path if not the same as model_name + tokenizer_name: null + cache_dir: ./cache_dir + use_fast_tokenizer: True # whether to use one of the fast tokenizer (backed by the tokenizers library) or not + model_revision: main # The specific model version to use (can be a branch name, tag name or commit id). + use_auth_token: Fasle # will use the token generated when running `transformers-cli login` (necessary to use this script " + # with private models)" + +train: + seed: 42 + output_dir: output_dir_gptq + overwrite_output_dir: False # use this to continue training if output_dir points to a checkpoint directory + do_train: False + do_eval: True + do_predict: False + evaluation_strategy: "no" #The evaluation strategy to use. "no"; "steps"; "epoch" + eval_steps: null # Run an evaluation every X steps. + per_device_train_batch_size: 4 # Batch size per GPU/TPU core/CPU for training. + per_device_eval_batch_size: 4 # Batch size per GPU/TPU core/CPU for evaluation + +progress: + log_level: debug # Logger log level to use on the main node. Possible choices are the log levels as strings: 'debug', 'info', 'warning', 'error' and 'critical', plus a 'passive' level which doesn't set anything and lets the application set the level. Defaults to 'passive'. + log_level_replica: passive # Logger log level to use on replica nodes. + logging_dir: null # Tensorboard log dir. + logging_strategy: steps # The logging strategy to use. "no"; "steps"; "epoch"; + logging_steps: 500 # Log every X updates steps. + + save_strategy: "no" # The checkpoint save strategy to use. "no"; "steps"; "epoch"; + save_steps: 500 # Save checkpoint every X updates steps. + save_total_limit: null # Limit the total amount of checkpoints. + # Deletes the older checkpoints in the output_dir. Default is unlimited checkpoints + save_on_each_node: False #When doing multi-node distributed training, whether to save models and checkpoints on each node, or only on the main one + + no_cuda: False # Do not use CUDA even when it is available + run_name: null # An optional descriptor for the run. Notably used for wandb logging. + disable_tqdm: null # Whether or not to disable the tqdm progress bars. use False or True + + load_best_model_at_end: False #Whether or not to load the best model found during training at the end of training. + metric_for_best_model: null # The metric to use to compare two different models." + greater_is_better: null # Whether the `metric_for_best_model` should be maximized or not. + diff --git a/application/nlp_example/config.yaml b/application/nlp_example/config.yaml index 00f64222..9cddb1da 100644 --- a/application/nlp_example/config.yaml +++ b/application/nlp_example/config.yaml @@ -6,11 +6,12 @@ quant: symmetric: False per_channel: False w_qconfig: - quantizer: FixedFakeQuantize + quantizer: E4M3FakeQuantize observer: MinMaxObserver bit: 8 symmetric: True - per_channel: True + per_channel: False # per-channel需要更长的训练时间 + scaling_method: no_scaling # 在这个地方选取一种scale获取的手段（max/mean/no_scaling，具体的获取流程可以阅读E4/E5的fake quant） calibrate: 64 pot_scale: False backend: academic @@ -31,13 +32,13 @@ data: model: type: bert - model_name_or_path: final_pretrain_models/bert-base-uncased-mrpc + model_name_or_path: Intel/bert-base-uncased-mrpc # 需要将此处的模型更改为要测试的网络模型 config_name: null # pretrained config name or path if not the same as model_name tokenizer_name: null cache_dir: ./cache_dir use_fast_tokenizer: True # whether to use one of the fast tokenizer (backed by the tokenizers library) or not model_revision: main # The specific model version to use (can be a branch name, tag name or commit id). - use_auth_token: Fasle # will use the token generated when running `transformers-cli login` (necessary to use this script " + use_auth_token: False # will use the token generated when running `transformers-cli login` (necessary to use this script " # with private models)" train: @@ -65,7 +66,7 @@ progress: # Deletes the older checkpoints in the output_dir. Default is unlimited checkpoints save_on_each_node: False #When doing multi-node distributed training, whether to save models and checkpoints on each node, or only on the main one - no_cuda: False # Do not use CUDA even when it is available + no_cuda: False # Do not use CUDA even when it is available (CUDA版本的使用开关) run_name: null # An optional descriptor for the run. Notably used for wandb logging. disable_tqdm: null # Whether or not to disable the tqdm progress bars. use False or True diff --git a/application/nlp_example/glue_utils.py b/application/nlp_example/glue_utils.py index 3580b2f4..3f34d40c 100644 --- a/application/nlp_example/glue_utils.py +++ b/application/nlp_example/glue_utils.py @@ -45,6 +45,9 @@ def parse_config(config_file): def make_huggingface_training_args(config_train, config_progress): training_args = TrainingArguments( seed=config_train.seed, + label_names = [ + 'labels' + ], output_dir=config_train.output_dir, overwrite_output_dir=config_train.overwrite_output_dir, do_train=config_train.do_train, @@ -61,7 +64,8 @@ def make_huggingface_training_args(config_train, config_progress): run_name=config_progress.run_name, disable_tqdm=config_progress.disable_tqdm, metric_for_best_model=config_progress.metric_for_best_model, - greater_is_better=config_progress.greater_is_better + greater_is_better=config_progress.greater_is_better, + label_names = ['labels'] ) config_progress.log_level = training_args.get_process_log_level() diff --git a/application/nlp_example/gpt2qat.py b/application/nlp_example/gpt2qat.py new file mode 100644 index 00000000..8610a005 --- /dev/null +++ b/application/nlp_example/gpt2qat.py @@ -0,0 +1,534 @@ +import torch +import torch.nn as nn +import numpy as np +import random +import inspect +import argparse +import unittest +import transformers +from cleantext import clean +from nltk.corpus import stopwords +from nltk.tokenize import word_tokenize +import pandas as pd +import numpy as np +import random +import datetime +import time +import copy +from torch.utils.data import Dataset, DataLoader, random_split, RandomSampler, SequentialSampler +torch.manual_seed(42) +from transformers import GPT2LMHeadModel, GPT2Tokenizer, GPT2Config, GPT2LMHeadModel +from transformers import GPT2PreTrainedModel, GPT2Tokenizer +from transformers import AdamW, get_linear_schedule_with_warmup +from tqdm import tqdm, trange +from itertools import chain +from tqdm.auto import tqdm +from transformers import AdamW, get_scheduler +from transformers import AutoModel +from transformers import AutoTokenizer +from transformers import default_data_collator +from transformers.onnx.features import FeaturesManager +from datasets import load_dataset,load_metric +import torch.optim as optim +from mqbench.convert_deploy import convert_deploy, convert_onnx +from mqbench.prepare_by_platform import prepare_by_platform, BackendType +from mqbench.utils.state import enable_calibration, enable_quantization, disable_all +from transformers import logging +import torch.onnx +import logging +import os +import collections +import torch.nn.functional as F +import csv +from torch.nn import CrossEntropyLoss +from torch.nn.parallel import DataParallel +from transformers import AutoModelForQuestionAnswering, TrainingArguments, Trainer +from transformers import GPT2LMHeadModel, GPT2Tokenizer +from transformers import GPT2Tokenizer, GPT2LMHeadModel, AdamW, get_linear_schedule_with_warmup +from transformers import BertTokenizer, BertModel +from transformers.utils.fx import HFTracer + +parser = argparse.ArgumentParser(description='MQBench gpt2 Training') + +parser.add_argument('--epochs', default=3, type=int, metavar='N', + help='number of total epochs to run') +parser.add_argument('--b', '--batch-size', default=1, type=int, + metavar='N', + help='mini-batch size (default: 16), this is the total ') +parser.add_argument('--lr', '--learning-rate', default=2e-4, type=float, + metavar='LR', help='initial learning rate', dest='lr') +parser.add_argument('--wd', '--weight-decay', default=1e-2, type=float, + metavar='W', help='weight decay (default: 1e-4)', + dest='wd') +parser.add_argument('--wbit', default=4, type=int, + metavar='wbit', help='weight bit') +parser.add_argument('--abit', default=8, type=int, + metavar='abit', help='active bit') +parser.add_argument('--wob', default='LSQObserver', type=str, + metavar='wob', help='weight observer') +parser.add_argument('--aob', default='EMAQuantileObserver', type=str, + metavar='aob', help='active observer') +parser.add_argument('--wfq', default='LearnableFakeQuantize', type=str, + metavar='wfq', help='weight fakequantize') +parser.add_argument('--afq', default='LearnableFakeQuantize', type=str, + metavar='afq', help='active fakequantize') +#clean data +def cleaning(text,punct): + cleaned_text = clean(text, + fix_unicode=False, # fix various unicode errors + to_ascii=False, # transliterate to closest ASCII representation + lower=True, # lowercase text + no_line_breaks=False, # fully strip line breaks as opposed to only normalizing them + no_urls=False, # replace all URLs with a special token + no_emails=False, # replace all email addresses with a special token + no_phone_numbers=False, # replace all phone numbers with a special token + no_numbers=False, # replace all numbers with a special token + no_digits=False, # replace all digits with a special token + no_currency_symbols=False, # replace all currency symbols with a special token + no_punct=punct, # remove punctuations + lang="en" # set to 'de' for German special handling + ) + + tokens = word_tokenize(cleaned_text) + filtered_sentence = [w for w in tokens if not w in stopwords.words('english')] + cleaned_text_0 = ' '.join(filtered_sentence) + return cleaned_text_0 +#train +def train(model,epochs,optimizer,scheduler,train_dataloader,validation_dataloader): + + total_t0 = time.time() + training_stats = [] + model = model.to(device) + for epoch_i in range(0, epochs): + print("") + print('======== Epoch {:} / {:} ========'.format(epoch_i + 1, epochs)) + print('Training...') + t0 = time.time() + total_train_loss = 0 + model.train() + for step, batch in enumerate(train_dataloader): + b_input_ids = batch[0].to(device) + b_labels = batch[0].to(device) + b_masks = batch[1].to(device) + model.zero_grad() + outputs = model(b_input_ids, + attention_mask = b_masks, + labels=b_labels + ) + loss = outputs[0] + batch_loss = loss.item() + total_train_loss += batch_loss + loss.backward() + optimizer.step() + scheduler.step() + # Calculate the average loss over all of the batches. + avg_train_loss = total_train_loss / len(train_dataloader) + # Measure how long this epoch took. + training_time = format_time(time.time() - t0) + print("") + print(" Average training loss: {0:.2f}".format(avg_train_loss)) + print(" Training epoch took: {:}".format(training_time)) + print("") + print("Running Validation...") + + t0 = time.time() + model.eval() + total_eval_loss = 0 + nb_eval_steps = 0 + + # Evaluate data for one epoch + for batch in validation_dataloader: + b_input_ids = batch[0].to(device) + b_labels = batch[0].to(device) + b_masks = batch[1].to(device) + with torch.no_grad(): + outputs = model(b_input_ids, + # token_type_ids=None, + attention_mask = b_masks, + labels=b_labels) + loss = outputs[0] + batch_loss = loss.item() + total_eval_loss += batch_loss + avg_val_loss = total_eval_loss / len(validation_dataloader) + validation_time = format_time(time.time() - t0) + print(" Validation Loss: {0:.2f}".format(avg_val_loss)) + print(" Validation took: {:}".format(validation_time)) + # Record all statistics from this epoch. + training_stats.append( + { + 'epoch': epoch_i + 1, + 'Training Loss': avg_train_loss, + 'Valid. Loss': avg_val_loss, + 'Training Time': training_time, + 'Validation Time': validation_time + } + ) + print("") + print("Training complete!") + print("Total training took {:} (h:mm:ss)".format(format_time(time.time()-total_t0))) + return model,training_stats +#quant train +def train1(model,epochs,optimizer,scheduler,train_dataloader,validation_dataloader): + + total_t0 = time.time() + training_stats = [] + model = model.to(device) + for epoch_i in range(0, epochs): + print("") + print('======== Epoch {:} / {:} ========'.format(epoch_i + 1, epochs)) + print('Training...') + t0 = time.time() + total_train_loss = 0 + model.train() + for step, batch in enumerate(train_dataloader): + b_input_ids = batch[0].to(device) + b_labels = batch[0].to(device) + b_masks = batch[1].to(device) + model.zero_grad() + outputs = model(b_input_ids, + attention_mask = b_masks, + labels=b_labels + ) + loss = outputs[0] + batch_loss = loss.item() + total_train_loss += batch_loss + loss.backward() + optimizer.step() + scheduler.step() + # Calculate the average loss over all of the batches. + avg_train_loss = total_train_loss / len(train_dataloader) + # Measure how long this epoch took. + training_time = format_time(time.time() - t0) + print("") + print(" Average training loss: {0:.2f}".format(avg_train_loss)) + print(" Training epoch took: {:}".format(training_time)) + print("") + print("Running Validation...") + + t0 = time.time() + model.eval() + total_eval_loss = 0 + nb_eval_steps = 0 + + # Evaluate data for one epoch + for batch in validation_dataloader: + b_input_ids = batch[0].to(device) + b_labels = batch[0].to(device) + b_masks = batch[1].to(device) + with torch.no_grad(): + outputs = model(b_input_ids, + # token_type_ids=None, + attention_mask = b_masks, + labels=b_labels) + loss = outputs[0] + batch_loss = loss.item() + total_eval_loss += batch_loss + avg_val_loss = total_eval_loss / len(validation_dataloader) + validation_time = format_time(time.time() - t0) + print(" Validation Loss: {0:.2f}".format(avg_val_loss)) + print(" Validation took: {:}".format(validation_time)) + # Record all statistics from this epoch. + training_stats.append( + { + 'epoch': epoch_i + 1, + 'Training Loss': avg_train_loss, + 'Valid. Loss': avg_val_loss, + 'Training Time': training_time, + 'Validation Time': validation_time + } + ) + print("") + print("Training complete!") + print("Total training took {:} (h:mm:ss)".format(format_time(time.time()-total_t0))) + return model,training_stats +def cal_ppl_bygpt2(model,test_dataloader): + total_ppl=0 + model.eval() + with torch.no_grad(): + for step, batch in enumerate(test_dataloader): + b_input_ids = batch[0].to(device) + b_masks = batch[1].to(device) + b_labels=b_input_ids + outputs = model(b_input_ids, + attention_mask = b_masks, + labels=b_labels) + bs, sl = b_input_ids.size() + logits = outputs[1] + # Shift so that tokens < n predict n + shift_logits = logits[:, :-1, :].contiguous() + shift_labels = b_input_ids[:, 1:].contiguous() + shift_attentions = b_masks[:, 1:].contiguous() + # Flatten the tokens + loss_fct = CrossEntropyLoss(ignore_index=0, reduction="none") + loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)).detach().reshape(bs, -1) + meanloss = loss.sum(1) / shift_attentions.sum(1) + ppl = torch.exp(meanloss).cpu() + ppl=ppl.numpy().tolist() + total_ppl+=ppl[0] + avg_ppl=total_ppl/len(test_dataloader) + return avg_ppl +def cal_ppl_bygpt22(model,test_dataloader): + total_ppl=0 + model.eval() + with torch.no_grad(): + for step, batch in enumerate(test_dataloader): + b_input_ids = batch[0].to(device) + b_masks = batch[1].to(device) + b_labels=b_input_ids + outputs = model(b_input_ids, + attention_mask = b_masks, + labels=b_labels) + bs, sl = b_input_ids.size() + logits = outputs[1] + # Shift so that tokens < n predict n + shift_logits = logits[:, :-1, :].contiguous() + shift_labels = b_input_ids[:, 1:].contiguous() + shift_attentions = b_masks[:, 1:].contiguous() + # Flatten the tokens + loss_fct = CrossEntropyLoss(ignore_index=0, reduction="none") + loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)).detach().reshape(bs, -1) + meanloss = loss.sum(1) / shift_attentions.sum(1) + ppl = torch.exp(meanloss).cpu() + ppl=ppl.numpy().tolist() + total_ppl+=ppl[0] + avg_ppl=total_ppl/len(test_dataloader) + return avg_ppl +def calibrate(cali_loader, model): + model.eval() + print("Start calibration ...") + print("Calibrate data number = ", len(cali_loader)) + with torch.no_grad(): + for step, batch in enumerate(cali_loader): + b_input_ids = batch[0].to(device) + b_masks = batch[1].to(device) + b_labels=b_input_ids + outputs = model(b_input_ids, + attention_mask = b_masks, + labels=b_labels) + print("Calibration ==> ", step+1) + print("End calibration.") + return +class GPT2Dataset(Dataset): + + def __init__(self, txt_list, tokenizer, gpt2_type="gpt2", max_length=1024): + + self.tokenizer = tokenizer + self.input_ids = [] + self.attn_masks = [] + + for txt in txt_list: + encodings_dict = tokenizer( '<|startoftext|>'+ txt + '<|endoftext|>', truncation=True, max_length=max_length,padding="max_length") #, padding="max_length" + self.input_ids.append(torch.tensor(encodings_dict['input_ids'])) + self.attn_masks.append(torch.tensor(encodings_dict['attention_mask'])) + + def __len__(self): + return len(self.input_ids) + + def __getitem__(self, idx): + return self.input_ids[idx], self.attn_masks[idx] + +###################################################################################################################### + +args = parser.parse_args() +#load parameters +batch_size =args.b +epochs = args.epochs +learning_rate = args.lr +warmup_steps = 1e2 +epsilon = 1e-8 +#load data +squad = load_dataset("squad", split="train[:10000]") +#processing data +que = [] +con = [] +ans = [] +for i in squad: + que.append(cleaning(i['question'],True)) + con.append(cleaning(i['context'],True)) + ans.append(cleaning(i['answers']['text'],True)) +main_data = pd.DataFrame() +main_data['Question'] = que +main_data['Context'] = con +main_data['Answer'] = ans +main_data['train_data'] = main_data['Question']+''+main_data['Context']+''+main_data['Answer'] +main_data = main_data.sample(frac=1,random_state=32).reset_index(drop=True) +train_data = main_data[0:9000].reset_index(drop=True) +cali_data=train_data['train_data'][0:100].reset_index(drop=True) +test_data = main_data[9000:10000].reset_index(drop=True) +validation = train_data['train_data'][8500:9000] +validation1 = validation.apply( lambda x: x.split('')[0]+'') +validation1.reset_index(drop=True,inplace=True) +#save data +train_data.to_csv('training_data.csv',index=False) +test_data.to_csv('testing_data.csv',index=False) +validation1.to_csv('validation_data.csv',index=False) +#load GPT tokenizer +tokenizer = GPT2Tokenizer.from_pretrained('gpt2',bos_token='<|startoftext|>', eos_token='<|endoftext|>', pad_token='<|pad|>') #gpt2-medium +#Building GPT dataset +traindataset = GPT2Dataset(train_data['train_data'][0:8500], tokenizer) +validdataset = GPT2Dataset(validation1, tokenizer) +cali_loader=GPT2Dataset(cali_data, tokenizer) +testdataset=GPT2Dataset(test_data['train_data'], tokenizer) +#Generate Text Collection +test_set = pd.DataFrame() +test_set['train_data']=test_data['train_data'][:500] +test_set['True_end_train_data1'] = test_set['train_data'].str.split().str[-20:].apply(' '.join) +test_set['train_data1'] = test_set['train_data'].str.split().str[:-20].apply(' '.join) +#dataloader +train_dataloader = DataLoader( + traindataset, # The training samples. + sampler = RandomSampler(traindataset), # Select batches randomly + batch_size = batch_size # Trains with this batch size. + ) + +# For validation the order doesn't matter, so we'll just read them sequentially. +validation_dataloader = DataLoader( + validdataset, # The validation samples. + sampler = SequentialSampler(validdataset), # Pull out batches sequentially. + batch_size = batch_size # Evaluate with this batch size. + ) +cali_loader = DataLoader( + cali_loader, + sampler = SequentialSampler(cali_loader), + batch_size = 2 + ) +test_dataloader = DataLoader( + testdataset, + sampler = RandomSampler(testdataset), + batch_size = batch_size + ) +#load model +configuration = GPT2Config.from_pretrained('gpt2',resid_pdrop = 0.3 , output_hidden_states=False) +model = GPT2LMHeadModel.from_pretrained("gpt2", config=configuration) +model.resize_token_embeddings(len(tokenizer)) +device = 'cuda' if torch.cuda.is_available() else 'cpu' +model=model.to(device) +seed_val = 32 +random.seed(seed_val) +np.random.seed(seed_val) +torch.manual_seed(seed_val) +torch.cuda.manual_seed_all(seed_val) +#quantize +sig = inspect.signature(model.forward) +input_names =['input_ids','attention_mask'] +concrete_args = {p.name: p.default for p in sig.parameters.values() if p.name not in input_names} +extra_qconfig_dict={ + 'w_observer': args.wob,#'MinMaxObserver', + 'a_observer': args.aob,#'EMAMinMaxObserver', + 'w_fakequantize':args.wfq, #'FixedFakeQuantize', + 'a_fakequantize':args.afq, # 'LearnableFakeQuantize', + 'w_qscheme': { + 'bit':args.wbit, + 'symmetry':True, + 'per_channel':False, + 'pot_scale': False + }, + 'a_qscheme': { + 'bit':args.abit, + 'symmetry': True, + 'per_channel': False, + 'pot_scale': False + } + } +preserve_attr={'': ['config']} +prepare_custom_config_dict = { + 'concrete_args': concrete_args, + 'preserve_attr': preserve_attr, + #'work_mode':'all_int4_qat', + 'extra_qconfig_dict':extra_qconfig_dict} +#Insert quantization node +model_prepared= prepare_by_platform(model, BackendType.Academic_NLP,[], prepare_custom_config_dict, custom_tracer=HFTracer()) + +#Post processing +class Quantizegpt2(GPT2PreTrainedModel): + """ + 用于建模类似SQuAD这样的问答数据集 + """ + def __init__(self,config): + super(Quantizegpt2, self).__init__(config) + self.gpt2 = model_prepared + + def forward(self, input_ids,attention_mask,labels=None): + + gpt2_output= self.gpt2(input_ids=input_ids,attention_mask=attention_mask) + lm_logits = gpt2_output['logits'] + loss = None + if labels is not None: + shift_logits = lm_logits[..., :-1, :].contiguous() + shift_labels = labels[..., 1:].contiguous() + loss_fct =nn.CrossEntropyLoss() + loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) + return loss,lm_logits +configuration = GPT2Config.from_pretrained('gpt2',resid_pdrop = 0.3 , output_hidden_states=False) +model_prepared1=Quantizegpt2(config=configuration) + +#train parameters +optimizer = AdamW(model_prepared1.parameters(), + lr = learning_rate, + eps = epsilon + ) +total_steps = len(train_dataloader) * epochs +scheduler = get_linear_schedule_with_warmup(optimizer, + num_warmup_steps = warmup_steps, + num_training_steps = total_steps) +def format_time(elapsed): + return str(datetime.timedelta(seconds=int(round((elapsed))))) + +#Original model training +model_prepared11=copy.deepcopy(model_prepared1) +optimizer1 = AdamW(model_prepared11.parameters(), + lr = learning_rate, + eps = epsilon + ) +scheduler1 = get_linear_schedule_with_warmup(optimizer1, + num_warmup_steps = warmup_steps, + num_training_steps = total_steps) +disable_all(model_prepared11) +device = 'cuda' if torch.cuda.is_available() else 'cpu' +#model_prepared1=DataParallel(model_prepared1) +model_prepared11=model_prepared11.train() +model_prepared2,training_stats1=train(model_prepared11,epochs,optimizer1,scheduler1,train_dataloader,validation_dataloader) + +# Display floats with two decimal places. +pd.set_option('precision', 2) +# Create a DataFrame from our training statistics. +df_stats1 = pd.DataFrame(data=training_stats1) +# Use the 'epoch' as the row index. +df_stats1 = df_stats1.set_index('epoch') +# Display the table. +print(df_stats1) + +#Original model PPL +avg_ppl1=cal_ppl_bygpt2(model_prepared2,test_dataloader) +print("原始模型PPL:{}".format(avg_ppl1)) + +#calibration +device = 'cuda' if torch.cuda.is_available() else 'cpu' +enable_calibration(model_prepared1) +model_prepared1=model_prepared1.to(device) +calibrate(cali_loader, model_prepared1) + +#quantize model train +enable_quantization(model_prepared1) +model_prepared1=model_prepared1.train() +model_prepared3,training_stats2=train1(model_prepared1,epochs,optimizer,scheduler,train_dataloader,validation_dataloader) + +# Display floats with two decimal places. +pd.set_option('precision', 2) +# Create a DataFrame from our training statistics. +df_stats2 = pd.DataFrame(data=training_stats2) +# Use the 'epoch' as the row index. +df_stats2 = df_stats2.set_index('epoch') +# Display the table. +print(df_stats2) + +#quantize model PPL +avg_ppl2=cal_ppl_bygpt22(model_prepared3,test_dataloader) +print("量化模型PPL:{}".format(avg_ppl2)) + + + + + + + diff --git a/application/nlp_example/ptq-gptq.py b/application/nlp_example/ptq-gptq.py new file mode 100644 index 00000000..4b70e8a2 --- /dev/null +++ b/application/nlp_example/ptq-gptq.py @@ -0,0 +1,354 @@ +import sys +import torch +import inspect +import logging +import datasets +import argparse +import transformers +import q_model +import glue_utils +import numpy as np +from transformers import ( + DataCollatorWithPadding, + EvalPrediction, + Trainer, + PretrainedConfig, + TrainingArguments, + default_data_collator, +) +from transformers.utils.fx import HFTracer +from transformers.onnx.features import FeaturesManager +from itertools import chain +from mqbench.prepare_by_platform import prepare_by_platform, BackendType +from mqbench.convert_deploy import convert_deploy, convert_onnx +from mqbench.utils.state import enable_quantization, enable_calibration_woquantization, enable_calibration +import re +from mqbench.fake_quantize import global_var +import copy + +backends = { + 'academic': BackendType.Academic_NLP, + 'tensorrt': BackendType.Tensorrt_NLP, +} + +device = torch.device('cuda') + +logger = logging.getLogger("transformer") + +def set_logger(config_progress): + + logging.basicConfig( + format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", + datefmt="%m/%d/%Y %H:%M:%S", + handlers=[logging.StreamHandler(sys.stdout)], + ) + log_level = config_progress.log_level + logger.setLevel(log_level) + datasets.utils.logging.set_verbosity(log_level) + transformers.utils.logging.set_verbosity(log_level) + transformers.utils.logging.enable_default_handler() + transformers.utils.logging.enable_explicit_format() + +def evaluate(trainer, eval_datasets, num_samples=-1): + logger.info("*** Evaluate ***") + if isinstance(eval_datasets, tuple): + for i in range(len(eval_datasets)): + if num_samples != -1: + metrics = trainer.evaluate(eval_dataset=eval_datasets[i].shuffle().select(range(num_samples))) + else: + metrics = trainer.evaluate(eval_dataset=eval_datasets[i]) + trainer.log_metrics("eval", metrics) + trainer.save_metrics("eval", metrics) + else: + if num_samples != -1: + metrics = trainer.evaluate(eval_dataset=eval_datasets.shuffle().select(range(num_samples))) + else: + metrics = trainer.evaluate(eval_dataset=eval_datasets) + trainer.log_metrics("eval", metrics) + trainer.save_metrics("eval", metrics) + + +def quantize_model(model, config_quant): + if not hasattr(config_quant, 'backend'): + config_quant.backend = 'academic' + sig = inspect.signature(model.forward) + input_names = ['input_ids', 'attention_mask', 'token_type_ids'] + concrete_args = {p.name: p.default for p in sig.parameters.values() if p.name not in input_names} + prepare_custom_config_dict = { + 'concrete_args': concrete_args, + 'preserve_attr': {'': ['config', 'num_labels']}, + 'extra_qconfig_dict':{ + 'w_observer': config_quant.w_qconfig.observer, + 'a_observer': config_quant.a_qconfig.observer, + 'w_fakequantize': config_quant.w_qconfig.quantizer, + 'a_fakequantize': config_quant.a_qconfig.quantizer, + 'w_qscheme': { + 'bit': config_quant.w_qconfig.bit, + 'symmetry': config_quant.w_qconfig.symmetric, + 'per_channel': config_quant.w_qconfig.per_channel, + 'pot_scale': config_quant.pot_scale + }, + 'a_qscheme': { + 'bit': config_quant.a_qconfig.bit, + 'symmetry': config_quant.a_qconfig.symmetric, + 'per_channel': config_quant.a_qconfig.per_channel, + 'pot_scale': config_quant.pot_scale + } + } + } + backend = backends[config_quant.backend] + model = prepare_by_platform(model, backend, prepare_custom_config_dict, custom_tracer=HFTracer()) + return model + +inp_out_hooks = [] + +def insert_model_info(model, valid_layers=(torch.nn.Conv2d, torch.nn.Linear, transformers.Conv1D)): + print('>'*6, 'Start Insert Info') + for name, module in model.named_modules(): + try: + items = module._modules.items() + assert(len(items)) + except: + # print(name) + if('.weight_fake_quant' in name): + layer_name = name.split('.weight_fake_quant')[0] + layer_names = layer_name.split('.') + upper_layer = model + for l in layer_names: + upper_layer = getattr(upper_layer, l) + + if isinstance(upper_layer, valid_layers): + print('>'*8, 'Insert', layer_name, type(upper_layer)) + def get_inp_out(layer_name): + def tmp(_, inp, out): + # print(inp[0].data, out) + global_var.set_value(layer_name+'.weight_fake_quant.inp', inp[0].data) + global_var.set_value(layer_name+'.weight_fake_quant.out', out.data) + return tmp + inp_out_hooks.append(upper_layer.register_forward_hook(get_inp_out(layer_name))) + layer_module = copy.deepcopy(upper_layer) + layer_module.weight_fake_quant = torch.nn.Sequential() + layer_module.requires_grad = False + setattr(upper_layer.weight_fake_quant, 'layer_module', layer_module) + setattr(upper_layer.weight_fake_quant, 'layer_type', type(layer_module)) + setattr(upper_layer.weight_fake_quant, 'layer_name', layer_name+'.weight_fake_quant') + setattr(upper_layer.weight_fake_quant, 'is_gptq_valid', True) + setattr(upper_layer.weight_fake_quant, 'is_gptq_done', False) + else: + setattr(upper_layer.weight_fake_quant, 'is_gptq_valid', False) + print('>'*6, 'End Insert Info') + +def remove_model_info(model, valid_layers=(torch.nn.Conv2d, torch.nn.Linear, transformers.Conv1D)): + print('>'*6, 'Start Remove Info') + print('>'*4, 'Remove hooks') + for hook in inp_out_hooks: + hook.remove() + print('>'*4, 'Set GPTQ Done') + layer_modules = [] + for name, module in model.named_modules(): + try: + items = module._modules.items() + assert(len(items)) + except: + # print(name) + if('.weight_fake_quant' in name): + layer_name = name.split('.weight_fake_quant')[0] + layer_names = layer_name.split('.') + upper_layer = model + for l in layer_names: + upper_layer = getattr(upper_layer, l) + + if isinstance(upper_layer, valid_layers): + if (getattr(upper_layer.weight_fake_quant, 'is_gptq_valid') == True): + setattr(upper_layer.weight_fake_quant, 'is_gptq_done', True) + if hasattr(upper_layer.weight_fake_quant, 'layer_module'): + layer_modules.append(upper_layer.weight_fake_quant) + + # print('>'*8, 'Remove', layer_name, type(upper_layer)) + # inp_out_hooks[name].remove() + # else: + # setattr(upper_layer.weight_fake_quant, 'is_gptq_valid', False) + for lm in layer_modules: + if hasattr(lm, 'layer_module'): + del lm.layer_module + print('>'*6, 'End Remove Info') + +def main(config_path): + global_var._init() + config = glue_utils.parse_config(config_path) + glue_utils.set_seed(config.train.seed) + training_args = glue_utils.make_huggingface_training_args(config.train, config.progress) + set_logger(config.progress) + raw_datasets, num_labels, label_list = glue_utils.load_dataset_labels(config.data) + tokenizer, model = glue_utils.load_model(config.model, config.data, num_labels) + i = 0 + for sts in model.state_dict(): + print(sts, model.state_dict()[sts]) + i = i + 1 + if (i>10): + break + label_to_id = None + if ( + model.config.label2id != PretrainedConfig(num_labels=num_labels).label2id + and config.data.task_name is not None + and config.data.task_name != 'stsb' + ): + # Some have all caps in their config, some don't. + label_name_to_id = {k.lower(): v for k, v in model.config.label2id.items()} + if list(sorted(label_name_to_id.keys())) == list(sorted(label_list)): + label_to_id = {i: int(label_name_to_id[label_list[i]]) for i in range(num_labels)} + else: + logger.warning( + "Your model seems to have been trained with labels, but they don't match the dataset: ", + f"model labels: {list(sorted(label_name_to_id.keys()))}, dataset labels: {list(sorted(label_list))}." + "\nIgnoring the model labels as a result.", + ) + + if label_to_id is not None: + model.config.label2id = label_to_id + model.config.id2label = {id: label for label, id in model.config.label2id.items()} + elif config.data.task_name is not None and config.data.task_name != 'stsb': + model.config.label2id = {l: i for i, l in enumerate(label_list)} + model.config.id2label = {id: label for label, id in model.config.label2id.items()} + config.data.max_seq_length = min(config.data.max_seq_length, tokenizer.model_max_length) + raw_datasets = glue_utils.preprocess_dataset(config.data, training_args, raw_datasets, label_to_id, tokenizer) + + if config.data.task_name == 'mnli': + eval_datasets = ( + raw_datasets['validation_matched'], raw_datasets['validation_mismatched'] + ) + else: + eval_datasets = raw_datasets['validation'] + metric = datasets.load_metric("glue", config.data.task_name) + + if hasattr(config, 'quant'): + model = quantize_model(model, config.quant) + print("MQBench Model:") + print(model) + + def compute_metrics(p: EvalPrediction): + preds = p.predictions[0] if isinstance(p.predictions, tuple) else p.predictions + preds = np.squeeze(preds) if config.data.task_name=='stsb' else np.argmax(preds, axis=1) + + result = metric.compute(predictions=preds, references=p.label_ids) + if len(result) > 1: + result["combined_score"] = np.mean(list(result.values())).item() + return result + + # Data collator will default to DataCollatorWithPadding, so we change it if we already did the padding. + if config.data.pad_to_max_length: + data_collator = default_data_collator + elif training_args.fp16: + data_collator = DataCollatorWithPadding(tokenizer, pad_to_multiple_of=8) + else: + data_collator = None + + trainer = Trainer( + model=model, + args=training_args, + eval_dataset=eval_datasets if training_args.do_eval else None, + compute_metrics=compute_metrics, + tokenizer=tokenizer, + data_collator=data_collator, + ) + model_with_label = q_model.Quant_Bert(model) + trainer.model = model_with_label + dicts = model.state_dict() + print(len(dicts)) + for layer in dicts: + print(layer, '\t', dicts[layer].shape) + if hasattr(config, 'quant'): + # calibrate the model + # calibrate_datasets = raw_datasets['train'].shuffle().select(range(config.quant.calibrate)) + # enable_calibration_woquantization(trainer.model, quantizer_type='act_fake_quant') + # eval_dataloader = trainer.get_eval_dataloader(calibrate_datasets) + # for step, inputs in enumerate(eval_dataloader): + # print(step, inputs) + # break + # inp = (inputs['input_ids'], inputs['token_type_ids'], inputs['attention_mask']) + # # inp.to(torch.device('cpu')) + # trainer.model = trainer.model.to(torch.device('cpu')) + # torch.onnx.export(trainer.model, inp, "bert_woquantization.onnx", do_constant_folding=False) + # trainer.model = trainer.model.to(torch.device('cuda')) + # evaluate(trainer, calibrate_datasets) + # enable_calibration_woquantization(trainer.model, quantizer_type='weight_fake_quant') + # for layer in model.named_modules(): + # print("============ layer 1 ===============") + # print(layer[1]) + # break + # evaluate(trainer, calibrate_datasets.select(range(2))) + + # calibrate_datasets = raw_datasets['train'].shuffle().select(range(config.quant.calibrate)) + # enable_calibration_woquantization(trainer.model, quantizer_type='act_fake_quant') + # evaluate(trainer, calibrate_datasets) + # enable_calibration_woquantization(trainer.model, quantizer_type='weight_fake_quant') + # evaluate(trainer, calibrate_datasets.select(range(2))) + + calibrate_datasets = raw_datasets['train'].shuffle().select(range(config.quant.calibrate)) + insert_model_info(trainer.model) + # export model + torch.save({'net': trainer.model.state_dict()}, 'torch_model_before.pth') + eval_dataloader = trainer.get_eval_dataloader(calibrate_datasets) + step, inputs = next(enumerate(eval_dataloader), 'end') + # inputs = inputs.to(device) + input_ids = inputs['input_ids'].to(device) + token_type_ids = inputs['token_type_ids'].to(device) + attention_mask = inputs['attention_mask'].to(device) + # inp = (inputs['input_ids'], inputs['token_type_ids'], inputs['attention_mask']) + enable_calibration(trainer.model) + evaluate(trainer, calibrate_datasets) + # trainer.model(input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask) + enable_quantization(trainer.model) + evaluate(trainer, calibrate_datasets.select(range(1))) + # trainer.model(input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask) + print("GPTQ End.") + # evaluate(trainer, calibrate_datasets.select(range(2))) + remove_model_info(trainer.model) + torch.save({'net': trainer.model.state_dict()}, 'torch_model_after.pth') + + if training_args.do_eval: + if hasattr(config, 'quant'): + enable_quantization(trainer.model) + evaluate(trainer, eval_datasets) + + model_kind, model_onnx_config = FeaturesManager.check_supported_model_or_raise(model, feature='default') + onnx_config = model_onnx_config(model.config) + export_inputs = {} + # device = torch.device('cpu') + export_inputs['input_ids'] = torch.tensor(eval_datasets[0]['input_ids']).unsqueeze(0).to(torch.device('cpu')) + export_inputs['token_type_ids'] = torch.tensor(eval_datasets[0]['token_type_ids']).unsqueeze(0).to(torch.device('cpu')) + export_inputs['attention_mask'] = torch.tensor(eval_datasets[0]['attention_mask']).unsqueeze(0).to(torch.device('cpu')) + + model = model.to(torch.device('cpu')) + for tensor in model.state_dict(): + if (model.state_dict()[tensor].device.type != 'cpu'): + print('*'*10, 'not same device', tensor) + + # kwargs = { + # 'input_shape_dict': {'input_ids': [1, 128], 'token_type_ids': [1, 128], 'attention_mask': [1, 128]}, + # 'output_path': './', + # 'model_name': 'mqbench_model_gptq', + # 'dummy_input': export_inputs, + # 'onnx_model_path': './mqbench_model_gptq.onnx', + # } + # convert_onnx(model, **kwargs) + + convert_deploy(model, + backends[config.quant.backend], + dummy_input=(export_inputs,), + model_name='mqbench_model_gptq', + input_names=list(onnx_config.inputs.keys()), + output_names=list(onnx_config.outputs.keys()), + dynamic_axes={name: axes for name, axes in chain(onnx_config.inputs.items(), onnx_config.outputs.items())} + ) + print("Done.") + + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser(description='configuration', + formatter_class=argparse.ArgumentDefaultsHelpFormatter) + parser.add_argument('--config', default='/home/zhang/Projects/quantization/MQBench/application/nlp_example/config-gptq.yaml', type=str) + args = parser.parse_args() + main(args.config) diff --git a/application/nlp_example/ptq.py b/application/nlp_example/ptq.py index 0ac959da..57174cee 100644 --- a/application/nlp_example/ptq.py +++ b/application/nlp_example/ptq.py @@ -26,6 +26,7 @@ backends = { 'academic': BackendType.Academic_NLP, 'tensorrt': BackendType.Tensorrt_NLP, + 'sophgo_tpu': BackendType.Sophgo_TPU } logger = logging.getLogger("transformer") @@ -91,8 +92,9 @@ def quantize_model(model, config_quant): } } } + backend = backends[config_quant.backend] - model = prepare_by_platform(model, backend, prepare_custom_config_dict, custom_tracer=HFTracer()) + model = prepare_by_platform(model, backend, prepare_custom_config_dict=prepare_custom_config_dict, custom_tracer=HFTracer()) return model @@ -136,7 +138,7 @@ def main(config_path): else: eval_datasets = raw_datasets['validation'] metric = datasets.load_metric("glue", config.data.task_name) - + if hasattr(config, 'quant'): model = quantize_model(model, config.quant) @@ -177,19 +179,18 @@ def compute_metrics(p: EvalPrediction): if training_args.do_eval: if hasattr(config, 'quant'): enable_quantization(trainer.model) - evaluate(trainer, eval_datasets) + evaluate(trainer, eval_datasets) #此步骤进行了fake quant操作，注释以后即可跳过fake quant的计算操作 model_kind, model_onnx_config = FeaturesManager.check_supported_model_or_raise(model, feature='default') onnx_config = model_onnx_config(model.config) export_inputs = {} - export_inputs['input_ids'] = torch.tensor(eval_datasets[0]['input_ids']).unsqueeze(0).cuda() - export_inputs['token_type_ids'] = torch.tensor(eval_datasets[0]['token_type_ids']).unsqueeze(0).cuda() - export_inputs['attention_mask'] = torch.tensor(eval_datasets[0]['attention_mask']).unsqueeze(0).cuda() - + export_inputs['input_ids'] = torch.tensor(eval_datasets[0]['input_ids']).unsqueeze(0)#.cuda() + export_inputs['token_type_ids'] = torch.tensor(eval_datasets[0]['token_type_ids']).unsqueeze(0)#.cuda() + export_inputs['attention_mask'] = torch.tensor(eval_datasets[0]['attention_mask']).unsqueeze(0)#.cuda() convert_deploy(model, backends[config.quant.backend], dummy_input=(export_inputs,), - model_name='mqbench_model', + model_name='bert-base-uncased', input_names=list(onnx_config.inputs.keys()), output_names=list(onnx_config.outputs.keys()), dynamic_axes={name: axes for name, axes in chain(onnx_config.inputs.items(), onnx_config.outputs.items())} diff --git a/application/nlp_example/qat_bertbase_classication.py b/application/nlp_example/qat_bertbase_classication.py new file mode 100644 index 00000000..706f1626 --- /dev/null +++ b/application/nlp_example/qat_bertbase_classication.py @@ -0,0 +1,258 @@ +import torch +import torch.nn as nn +import inspect +import unittest +import argparse +import copy +from itertools import chain +from torch.utils.data import DataLoader +from tqdm.auto import tqdm +from transformers import AdamW, get_scheduler +from transformers import AutoModel +from transformers import AutoTokenizer +from transformers import BertTokenizer, BertModel +from transformers.utils.fx import HFTracer +from transformers.onnx.features import FeaturesManager +from datasets import load_dataset +import torch.optim as optim +from mqbench.convert_deploy import convert_deploy, convert_onnx +from mqbench.prepare_by_platform import prepare_by_platform, BackendType +from mqbench.utils.state import enable_calibration, enable_quantization, disable_all +from transformers import logging +import matplotlib.pyplot as plt +import torch.onnx + +parser = argparse.ArgumentParser(description='MQBench bertbase Training') + +parser.add_argument('--epochs', default=1, type=int, metavar='N', + help='number of total epochs to run') +parser.add_argument('--b', '--batch-size', default=4, type=int, + metavar='N', + help='mini-batch size (default: 16), this is the total ') +parser.add_argument('--lr', '--learning-rate', default=1e-5, type=float, + metavar='LR', help='initial learning rate', dest='lr') +parser.add_argument('--wd', '--weight-decay', default=1e-2, type=float, + metavar='W', help='weight decay (default: 1e-4)', + dest='wd') +parser.add_argument('--wbit', default=4, type=int, + metavar='wbit', help='weight bit') +parser.add_argument('--abit', default=8, type=int, + metavar='abit', help='active bit') +parser.add_argument('--wob', default='LSQObserver', type=str, + metavar='wob', help='weight observer') +parser.add_argument('--aob', default='EMAQuantileObserver', type=str, + metavar='aob', help='active observer') +parser.add_argument('--wfq', default='AdaRoundFakeQuantize', type=str, + metavar='wfq', help='weight fakequantize') +parser.add_argument('--afq', default='LearnableFakeQuantize', type=str, + metavar='afq', help='active fakequantize') + +class Dataset(torch.utils.data.Dataset): + def __init__(self, data_type): + self.data = self.load_data(data_type) + + def load_data(self, data_type): + tmp_dataset = load_dataset(path='laugustyniak/abusive-clauses-pl', split = data_type) + Data = {} + for idx, line in enumerate(tmp_dataset): + sample = line + Data[idx] = sample + return Data + + def __len__(self): + return len(self.data) + + def __getitem__(self, idx): + return self.data[idx] +def collote_fn(batch_samples): + batch_text= [] + batch_label = [] + for sample in batch_samples: + batch_text.append(sample['text']) + batch_label.append(int(sample['label'])) + X = tokenizer( + batch_text, + padding=True, + truncation=True, + return_tensors="pt" + ) + y = torch.tensor(batch_label) + return X, y +def train_loop(dataloader, model, loss_fn, optimizer, lr_scheduler, epoch, total_loss): + progress_bar = tqdm(range(len(dataloader))) + progress_bar.set_description(f'loss: {0:>7f}') + finish_batch_num = (epoch-1)*len(dataloader) + + model.train() + for batch, (X, y) in enumerate(dataloader, start=1): + X, y = X.to(device), y.to(device) + pred = model(X) + loss = loss_fn(pred, y) + + optimizer.zero_grad() + loss.backward() + optimizer.step() + lr_scheduler.step() + + total_loss += loss.item() + progress_bar.set_description(f'loss: {total_loss/(finish_batch_num + batch):>7f}') + progress_bar.update(1) + #losses.append(loss.item()) + #print(f"Epoch [{epoch+1}/{epochs}], Loss: {loss.item()}") + return total_loss #losses + +def test_loop(dataloader, model, mode='Test'): + assert mode in ['Valid', 'Test'] + size = len(dataloader.dataset) + correct = 0 + + model.eval() + with torch.no_grad(): + for X, y in dataloader: + X, y = X.to(device), y.to(device) + pred = model(X) + correct += (pred.argmax(1) == y).type(torch.float).sum().item() + + correct /= size + print(f"{mode} Accuracy: {(100*correct):>0.1f}%\n") + return correct +def calibrate(cali_loader, model): + model.eval() + print("Start calibration ...") + print("Calibrate data number = ", len(cali_loader.dataset)) + with torch.no_grad(): + for i, (X, y) in enumerate(cali_loader): + X, y = X.to(device), y.to(device) + pred = model(X) + print("Calibration ==> ", i+1) + print("End calibration.") + return + +################################################################################################################## + +args = parser.parse_args() +#load data +train_data = Dataset('train') +test_data = Dataset('test') + +#load parameters +checkpoint = "bert-base-uncased" +tokenizer = AutoTokenizer.from_pretrained(checkpoint) +device = 'cuda' if torch.cuda.is_available() else 'cpu' +print(f'Using {device} device') +learning_rate = args.lr +epoch_num = args.epochs + +#dataloader +train_dataloader = DataLoader(train_data, batch_size=args.b, shuffle=True, collate_fn=collote_fn) +test_dataloader = DataLoader(test_data, batch_size=args.b, shuffle=True, collate_fn=collote_fn) + +#quantize +model1=AutoModel.from_pretrained(checkpoint) +#量化模型参数准备 +sig = inspect.signature(model1.forward) +input_names =['input_ids','token_type_ids','attention_mask'] +concrete_args = {p.name: p.default for p in sig.parameters.values() if p.name not in input_names} +extra_qconfig_dict={ + 'w_observer': args.wob,#'MinMaxObserver', + 'a_observer': args.aob,#'EMAMinMaxObserver', + 'w_fakequantize':args.wfq, #'FixedFakeQuantize', + 'a_fakequantize':args.afq, # 'LearnableFakeQuantize', + 'w_qscheme': { + 'bit':args.wbit, + 'symmetry':True, + 'per_channel': False, + 'pot_scale': False + }, + 'a_qscheme': { + 'bit':args.abit, + 'symmetry': True, + 'per_channel': False, + 'pot_scale': False + } + } +preserve_attr={'': ['config']} +prepare_custom_config_dict = { + 'concrete_args': concrete_args, + 'preserve_attr': preserve_attr, + #'work_mode':'all_int4_qat', + 'extra_qconfig_dict':extra_qconfig_dict} +#插入量化节点 +model_prepared= prepare_by_platform(model1, BackendType.Academic_NLP,[], prepare_custom_config_dict, custom_tracer=HFTracer()) +#后处理 +class NeuralNetwork2(nn.Module): + def __init__(self): + super(NeuralNetwork2, self).__init__() + self.bert_encoder = model_prepared + self.classifier = nn.Linear(768, 2) + + def forward(self, x): + bert_output = self.bert_encoder(**x) + cls_vectors = bert_output['last_hidden_state'][:, 0] + logits = self.classifier(cls_vectors) + return logits +model_prepared1 = NeuralNetwork2().to(device) +#校准 +cali =[] +for i in range(20): + text=train_data[i] + cali.append(text) +cali_loader = DataLoader(cali, batch_size=args.b, shuffle=True, collate_fn=collote_fn) +enable_calibration(model_prepared1) +model_prepared1=model_prepared1.to(device) +calibrate(cali_loader, model_prepared1) +#原始模型精度 +model_prepared11=copy.deepcopy(model_prepared1) +disable_all(model_prepared11) +model_prepared11=model_prepared11.train() +optimizer = AdamW(model_prepared11.parameters(), lr=learning_rate) +lr_scheduler = get_scheduler( + "linear", + optimizer=optimizer, + num_warmup_steps=0, + num_training_steps=epoch_num*len(train_dataloader), +) +loss_fn = nn.CrossEntropyLoss() +total_loss = 0 +best_acc = 0 +for t in range(epoch_num): + print(f"Epoch {t+1}/{epoch_num}\n-------------------------------") + total_loss = train_loop(train_dataloader, model_prepared11, loss_fn, optimizer, lr_scheduler, t+1, total_loss) + Test_acc = test_loop(test_dataloader,model_prepared11, mode='Test') + if Test_acc > best_acc: + best_acc = Test_acc + print('saving new weights...\n') +print("Done!") + +#量化模型精度 +enable_quantization(model_prepared1) +model_prepared1=model_prepared1.train() +total_loss = 0 +best_acc = 0 +optimizer1 = AdamW(model_prepared1.parameters(), lr=learning_rate) +lr_scheduler1 = get_scheduler( + "linear", + optimizer=optimizer1, + num_warmup_steps=0, + num_training_steps=epoch_num*len(train_dataloader), +) +for t in range(epoch_num): + print(f"Epoch {t+1}/{epoch_num}\n-------------------------------") + total_loss = train_loop(train_dataloader, model_prepared1, loss_fn, optimizer1, lr_scheduler1, t+1, total_loss) + Test_acc = test_loop(test_dataloader,model_prepared1, mode='Test') + if Test_acc > best_acc: + best_acc = Test_acc + print('saving new weights...\n') +print("Done!") + +#量化模型部署 +train_dataloader1 = DataLoader(train_data, batch_size=1, shuffle=True, collate_fn=collote_fn) +batch_X1, batch_y1 = next(iter(train_dataloader1)) +model_prepared.eval() +model_kind, model_onnx_config = FeaturesManager.check_supported_model_or_raise(model_prepared, feature='default') +onnx_config = model_onnx_config(model_prepared.config) +convert_deploy(model_prepared, + BackendType.Academic_NLP, + dummy_input=(dict(batch_X1),), + model_name='bert-base-uncased-mqbench' + ) \ No newline at end of file diff --git a/application/nlp_example/qat_bertbase_questionanswer.py b/application/nlp_example/qat_bertbase_questionanswer.py new file mode 100644 index 00000000..06adfc61 --- /dev/null +++ b/application/nlp_example/qat_bertbase_questionanswer.py @@ -0,0 +1,479 @@ +#导入所需的库 +import argparse +import transformers +import torch +import torch.nn as nn +import inspect +import unittest +import copy +from itertools import chain +from torch.utils.data import DataLoader +from tqdm.auto import tqdm +import numpy as np +from transformers import AdamW, get_scheduler +from transformers import AutoModel +from transformers import AutoTokenizer +from transformers import default_data_collator +from transformers.onnx.features import FeaturesManager +from datasets import load_dataset,load_metric +import torch.optim as optim +from mqbench.convert_deploy import convert_deploy, convert_onnx +from mqbench.prepare_by_platform import prepare_by_platform, BackendType +from mqbench.utils.state import enable_calibration, enable_quantization, disable_all +from transformers import logging +import matplotlib.pyplot as plt +import torch.onnx +import pandas as pd +import json +import logging +import os +import collections +import six +from transformers import DistilBertConfig +from transformers import AutoModelForQuestionAnswering, TrainingArguments, Trainer +from transformers import BertTokenizer, BertModel +from transformers.utils.fx import HFTracer +from transformers import Trainer, TrainingArguments, PreTrainedModel + +parser = argparse.ArgumentParser(description='MQBench bertbase Training') + +parser.add_argument('--epochs', default=1, type=int, metavar='N', + help='number of total epochs to run') +parser.add_argument('--b', '--batch-size', default=16, type=int, + metavar='N', + help='mini-batch size (default: 16), this is the total ') +parser.add_argument('--lr', '--learning-rate', default=2e-5, type=float, + metavar='LR', help='initial learning rate', dest='lr') +parser.add_argument('--wd', '--weight-decay', default=1e-2, type=float, + metavar='W', help='weight decay (default: 1e-4)', + dest='wd') +parser.add_argument('--wbit', default=4, type=int, + metavar='wbit', help='weight bit') +parser.add_argument('--abit', default=8, type=int, + metavar='abit', help='active bit') +parser.add_argument('--wob', default='LSQObserver', type=str, + metavar='wob', help='weight observer') +parser.add_argument('--aob', default='EMAQuantileObserver', type=str, + metavar='aob', help='active observer') +parser.add_argument('--wfq', default='AdaRoundFakeQuantize', type=str, + metavar='wfq', help='weight fakequantize') +parser.add_argument('--afq', default='LearnableFakeQuantize', type=str, + metavar='afq', help='active fakequantize') +parser.add_argument('--backend', type=str, choices=['Academic_NLP', 'Tensorrt_NLP'], default='Academic_NLP') + + +#前处理数据 +def prepare_train_features(examples): + # Some of the questions have lots of whitespace on the left, which is not useful and will make the + # truncation of the context fail (the tokenized question will take a lots of space). So we remove that + # left whitespace + examples["question"] = [q.lstrip() for q in examples["question"]] + + # Tokenize our examples with truncation and padding, but keep the overflows using a stride. This results + # in one example possible giving several features when a context is long, each of those features having a + # context that overlaps a bit the context of the previous feature. + tokenized_examples = tokenizer( + examples["question" if pad_on_right else "context"], + examples["context" if pad_on_right else "question"], + truncation="only_second" if pad_on_right else "only_first", + max_length=max_length, + stride=doc_stride, + return_overflowing_tokens=True, + return_offsets_mapping=True, + padding="max_length", + ) + + # Since one example might give us several features if it has a long context, we need a map from a feature to + # its corresponding example. This key gives us just that. + sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") + # The offset mappings will give us a map from token to character position in the original context. This will + # help us compute the start_positions and end_positions. + offset_mapping = tokenized_examples.pop("offset_mapping") + + # Let's label those examples! + tokenized_examples["start_positions"] = [] + tokenized_examples["end_positions"] = [] + + for i, offsets in enumerate(offset_mapping): + # We will label impossible answers with the index of the CLS token. + input_ids = tokenized_examples["input_ids"][i] + cls_index = input_ids.index(tokenizer.cls_token_id) + + # Grab the sequence corresponding to that example (to know what is the context and what is the question). + sequence_ids = tokenized_examples.sequence_ids(i) + + # One example can give several spans, this is the index of the example containing this span of text. + sample_index = sample_mapping[i] + answers = examples["answers"][sample_index] + # If no answers are given, set the cls_index as answer. + if len(answers["answer_start"]) == 0: + tokenized_examples["start_positions"].append(cls_index) + tokenized_examples["end_positions"].append(cls_index) + else: + # Start/end character index of the answer in the text. + start_char = answers["answer_start"][0] + end_char = start_char + len(answers["text"][0]) + + # Start token index of the current span in the text. + token_start_index = 0 + while sequence_ids[token_start_index] != (1 if pad_on_right else 0): + token_start_index += 1 + + # End token index of the current span in the text. + token_end_index = len(input_ids) - 1 + while sequence_ids[token_end_index] != (1 if pad_on_right else 0): + token_end_index -= 1 + + # Detect if the answer is out of the span (in which case this feature is labeled with the CLS index). + if not (offsets[token_start_index][0] <= start_char and offsets[token_end_index][1] >= end_char): + tokenized_examples["start_positions"].append(cls_index) + tokenized_examples["end_positions"].append(cls_index) + else: + # Otherwise move the token_start_index and token_end_index to the two ends of the answer. + # Note: we could go after the last offset if the answer is the last word (edge case). + while token_start_index < len(offsets) and offsets[token_start_index][0] <= start_char: + token_start_index += 1 + tokenized_examples["start_positions"].append(token_start_index - 1) + while offsets[token_end_index][1] >= end_char: + token_end_index -= 1 + tokenized_examples["end_positions"].append(token_end_index + 1) + + return tokenized_examples + +def prepare_validation_features(examples): + # Some of the questions have lots of whitespace on the left, which is not useful and will make the + # truncation of the context fail (the tokenized question will take a lots of space). So we remove that + # left whitespace + examples["question"] = [q.lstrip() for q in examples["question"]] + + # Tokenize our examples with truncation and maybe padding, but keep the overflows using a stride. This results + # in one example possible giving several features when a context is long, each of those features having a + # context that overlaps a bit the context of the previous feature. + tokenized_examples = tokenizer( + examples["question" if pad_on_right else "context"], + examples["context" if pad_on_right else "question"], + truncation="only_second" if pad_on_right else "only_first", + max_length=max_length, + stride=doc_stride, + return_overflowing_tokens=True, + return_offsets_mapping=True, + padding="max_length", + ) + + # Since one example might give us several features if it has a long context, we need a map from a feature to + # its corresponding example. This key gives us just that. + sample_mapping = tokenized_examples.pop("overflow_to_sample_mapping") + + # We keep the example_id that gave us this feature and we will store the offset mappings. + tokenized_examples["example_id"] = [] + + for i in range(len(tokenized_examples["input_ids"])): + # Grab the sequence corresponding to that example (to know what is the context and what is the question). + sequence_ids = tokenized_examples.sequence_ids(i) + context_index = 1 if pad_on_right else 0 + + # One example can give several spans, this is the index of the example containing this span of text. + sample_index = sample_mapping[i] + tokenized_examples["example_id"].append(examples["id"][sample_index]) + + # Set to None the offset_mapping that are not part of the context so it's easy to determine if a token + # position is part of the context or not. + tokenized_examples["offset_mapping"][i] = [ + (o if sequence_ids[k] == context_index else None) + for k, o in enumerate(tokenized_examples["offset_mapping"][i]) + ] + + return tokenized_examples +def postprocess_qa_predictions(examples, features, raw_predictions, n_best_size = 20, max_answer_length = 30): + all_start_logits, all_end_logits = raw_predictions + # Build a map example to its corresponding features. + example_id_to_index = {k: i for i, k in enumerate(examples["id"])} + features_per_example = collections.defaultdict(list) + for i, feature in enumerate(features): + features_per_example[example_id_to_index[feature["example_id"]]].append(i) + + # The dictionaries we have to fill. + predictions = collections.OrderedDict() + + # Logging. + print(f"Post-processing {len(examples)} example predictions split into {len(features)} features.") + + # Let's loop over all the examples! + for example_index, example in enumerate(tqdm(examples)): + # Those are the indices of the features associated to the current example. + feature_indices = features_per_example[example_index] + + min_null_score = None # Only used if squad_v2 is True. + valid_answers = [] + + context = example["context"] + # Looping through all the features associated to the current example. + for feature_index in feature_indices: + # We grab the predictions of the model for this feature. + start_logits = all_start_logits[feature_index] + end_logits = all_end_logits[feature_index] + # This is what will allow us to map some the positions in our logits to span of texts in the original + # context. + offset_mapping = features[feature_index]["offset_mapping"] + + # Update minimum null prediction. + cls_index = features[feature_index]["input_ids"].index(tokenizer.cls_token_id) + feature_null_score = start_logits[cls_index] + end_logits[cls_index] + if min_null_score is None or min_null_score < feature_null_score: + min_null_score = feature_null_score + + # Go through all possibilities for the `n_best_size` greater start and end logits. + start_indexes = np.argsort(start_logits)[-1 : -n_best_size - 1 : -1].tolist() + end_indexes = np.argsort(end_logits)[-1 : -n_best_size - 1 : -1].tolist() + for start_index in start_indexes: + for end_index in end_indexes: + # Don't consider out-of-scope answers, either because the indices are out of bounds or correspond + # to part of the input_ids that are not in the context. + if ( + start_index >= len(offset_mapping) + or end_index >= len(offset_mapping) + or offset_mapping[start_index] is None + or offset_mapping[end_index] is None + ): + continue + # Don't consider answers with a length that is either < 0 or > max_answer_length. + if end_index < start_index or end_index - start_index + 1 > max_answer_length: + continue + + start_char = offset_mapping[start_index][0] + end_char = offset_mapping[end_index][1] + valid_answers.append( + { + "score": start_logits[start_index] + end_logits[end_index], + "text": context[start_char: end_char] + } + ) + + if len(valid_answers) > 0: + best_answer = sorted(valid_answers, key=lambda x: x["score"], reverse=True)[0] + else: + # In the very rare edge case we have not a single non-null prediction, we create a fake prediction to avoid + # failure. + best_answer = {"text": "", "score": 0.0} + + # Let's pick our final answer: the best one or the null answer (only for squad_v2) + if not squad_v2: + predictions[example["id"]] = best_answer["text"] + else: + answer = best_answer["text"] if best_answer["score"] > min_null_score else "" + predictions[example["id"]] = answer + + return predictions +def calibrate(cali_loader, model): + model.eval() + print("Start calibration ...") + print("Calibrate data number = ", len(cali_loader)) + with torch.no_grad(): + for i in range(len(cali_loader)): + X= next(iter(cali_loader)) + batch_input =X['input_ids'].to(device) + batch_seg = X['attention_mask'].to(device) + start_logits, end_logits = model(input_ids=batch_input, + attention_mask=batch_seg) + print("Calibration ==> ", i+1) + print("End calibration.") + return + +def prec(datasets,trainer): + validation_features1 = datasets["validation"].map( + prepare_validation_features, + batched=True, + remove_columns=datasets["validation"].column_names + ) + raw_predictions1 = trainer.predict(validation_features1) + + validation_features1.set_format(type=validation_features1.format["type"], columns=list(validation_features1.features.keys())) + examples1 = datasets["validation"] + features1 = validation_features1 + example_id_to_index1 = {k: i for i, k in enumerate(examples1["id"])} + features_per_example1 = collections.defaultdict(list) + for i, feature in enumerate(features1): + features_per_example1[example_id_to_index1[feature["example_id"]]].append(i) + + final_predictions1 = postprocess_qa_predictions(datasets["validation"], validation_features1, raw_predictions1.predictions) + metric = load_metric("squad_v2" if squad_v2 else "squad") + if squad_v2: + formatted_predictions = [{"id": k, "prediction_text": v, "no_answer_probability": 0.0} for k, v in final_predictions1.items()] + else: + formatted_predictions = [{"id": k, "prediction_text": v} for k, v in final_predictions1.items()] + references = [{"id": ex["id"], "answers": ex["answers"]} for ex in datasets["validation"]] + result=metric.compute(predictions=formatted_predictions, references=references) + print(result) + return +################################################################################################################### + +#输入参数 +args = parser.parse_args() +squad_v2 = False +model_checkpoint = "distilbert-base-uncased" +batch_size = args.b + +#导入数据 +datasets = load_dataset("squad_v2" if squad_v2 else "squad") +#快速分词 +tokenizer = AutoTokenizer.from_pretrained(model_checkpoint) +assert isinstance(tokenizer, transformers.PreTrainedTokenizerFast) +#预处理参数导入 +max_length = 384 # The maximum length of a feature (question and context) +doc_stride = 128 # The authorized overlap between two part of the context when splitting it is needed. +pad_on_right = tokenizer.padding_side == "right" +n_best_size = 20 +#对训练数据进行处理 +tokenized_datasets = datasets.map(prepare_train_features, batched=True, remove_columns=datasets["train"].column_names) +#训练参数导入 +model = AutoModelForQuestionAnswering.from_pretrained(model_checkpoint) +model_name = model_checkpoint.split("/")[-1] +args1 = TrainingArguments( + f"{model_name}-finetuned-squad", + evaluation_strategy = "epoch", + learning_rate=args.lr, + per_device_train_batch_size=batch_size, + per_device_eval_batch_size=batch_size, + num_train_epochs=1, + weight_decay=args.wd +) +data_collator = default_data_collator + +############################################################################################################### + +#量化模型参数准备 +sig = inspect.signature(model.forward) +input_names =['input_ids','token_type_ids','attention_mask'] +#input_names =['input_ids','attention_mask'] +concrete_args = {p.name: p.default for p in sig.parameters.values() if p.name not in input_names} +extra_qconfig_dict={ + 'w_observer': args.wob,#'MinMaxObserver', + 'a_observer': args.aob,#'EMAMinMaxObserver', + 'w_fakequantize':args.wfq, #'FixedFakeQuantize', + 'a_fakequantize':args.afq, # 'LearnableFakeQuantize', + 'w_qscheme': { + 'bit':args.wbit, + 'symmetry':True, + 'per_channel':False, + 'pot_scale': False + }, + 'a_qscheme': { + 'bit':args.abit, + 'symmetry': True, + 'per_channel': False, + 'pot_scale': False + } + } +preserve_attr={'': ['config']} +prepare_custom_config_dict = { + 'concrete_args': concrete_args, + 'preserve_attr': preserve_attr, + #'work_mode':'all_int4_qat', + 'extra_qconfig_dict':extra_qconfig_dict} +#插入量化节点 +model_prepared= prepare_by_platform(model, BackendType.Academic_NLP,[], prepare_custom_config_dict, custom_tracer=HFTracer()) + +#校准 +device = 'cuda' if torch.cuda.is_available() else 'cpu' +cali =[] +for i in range(64): + text=tokenized_datasets["train"][i] + cali.append(text) +cali_loader = DataLoader(cali, batch_size=16, shuffle=True, collate_fn= default_data_collator) +enable_calibration(model_prepared) +model_prepared=model_prepared.to(device) +calibrate(cali_loader, model_prepared) + +#模型后处理 +enable_quantization(model_prepared) +model_prepared.train() +class BertForQuestionAnswering(PreTrainedModel): + """ + 用于建模类似SQuAD这样的问答数据集 + """ + def __init__(self,config): + super(BertForQuestionAnswering, self).__init__(config) + self.bert = model_prepared + + def forward(self, input_ids, + attention_mask=None, + start_positions=None, + end_positions=None): + bert_output= self.bert( + input_ids=input_ids, + attention_mask=attention_mask) + start_logits=bert_output['start_logits'] + end_logits=bert_output['end_logits'] + start_logits = start_logits.squeeze(-1) + end_logits = end_logits.squeeze(-1) + + if start_positions is not None and end_positions is not None: + # 由于部分情况下start/end 位置会超过输入的长度 + # （例如输入序列的可能大于512，并且正确的开始或者结束符就在512之后） + # 那么此时就要进行特殊处理 + ignored_index = start_logits.size(1) # 取输入序列的长度 + start_positions.clamp_(0, ignored_index) + # 如果正确起始位置start_positions中，存在输入样本的开始位置大于输入长度， + # 那么直接取输入序列的长度作为开始位置 + end_positions.clamp_(0, ignored_index) + + loss_fct = nn.CrossEntropyLoss(ignore_index=ignored_index) + # 这里指定ignored_index其实就是为了忽略掉超过输入序列长度的（起始结束）位置 + # 在预测时所带来的损失，因为这些位置并不能算是模型预测错误的（只能看做是没有预测）， + # 同时如果不加ignore_index的话，那么可能会影响模型在正常情况下的语义理解能力 + start_loss = loss_fct(start_logits, start_positions) + end_loss = loss_fct(end_logits, end_positions) + return (start_loss + end_loss) / 2, start_logits, end_logits + else: + return start_logits, end_logits +config1 = DistilBertConfig.from_pretrained('distilbert-base-uncased') +# 创建自定义配置对象 +model_prepared2=BertForQuestionAnswering(config1) +# 原始模型训练 +model_prepared22=copy.deepcopy(model_prepared2) +disable_all(model_prepared22) +model_prepared22=model_prepared22.train() +trainer1 = Trainer( + model_prepared22, + args1, + train_dataset=tokenized_datasets["train"], + eval_dataset=tokenized_datasets["validation"], + data_collator=data_collator, + tokenizer=tokenizer, +) +trainer1.train() +print("原始模型精度：") +prec(datasets,trainer1) +print("**************************************************") +# 量化模型训练 +enable_quantization(model_prepared2) +model_prepared2.train() +trainer2 = Trainer( + model_prepared2, + args1, + train_dataset=tokenized_datasets["train"], + eval_dataset=tokenized_datasets["validation"], + data_collator=data_collator, + tokenizer=tokenizer, +) +trainer2.train() +print("量化模型精度：") +prec(datasets,trainer2) +print("**************************************************") + +#模型部署 +keys_to_copy = ['input_ids', 'attention_mask'] +copied_cali=[] +for i in range(len(cali)): + text= {key: cali[i][key] for key in keys_to_copy} + copied_cali.append(text) +cali_loader1 = DataLoader(copied_cali, batch_size=1, shuffle=True, collate_fn= default_data_collator) +X=next(iter(cali_loader1)) +model_prepared.eval() +model_kind, model_onnx_config = FeaturesManager.check_supported_model_or_raise(model_prepared, feature='default') +onnx_config = model_onnx_config(model_prepared.config) +convert_deploy(model_prepared, + BackendType.Academic_NLP, + dummy_input=((dict(X)),), + model_name='bert-base-uncased-mqbench-squad' + ) diff --git a/application/yolov5_example/.dockerignore b/application/yolov5_example/.dockerignore new file mode 100644 index 00000000..3b669254 --- /dev/null +++ b/application/yolov5_example/.dockerignore @@ -0,0 +1,222 @@ +# Repo-specific DockerIgnore ------------------------------------------------------------------------------------------- +.git +.cache +.idea +runs +output +coco +storage.googleapis.com + +data/samples/* +**/results*.csv +*.jpg + +# Neural Network weights ----------------------------------------------------------------------------------------------- +**/*.pt +**/*.pth +**/*.onnx +**/*.engine +**/*.mlmodel +**/*.torchscript +**/*.torchscript.pt +**/*.tflite +**/*.h5 +**/*.pb +*_saved_model/ +*_web_model/ +*_openvino_model/ + +# Below Copied From .gitignore ----------------------------------------------------------------------------------------- +# Below Copied From .gitignore ----------------------------------------------------------------------------------------- + + +# GitHub Python GitIgnore ---------------------------------------------------------------------------------------------- +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +env/ +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +*.egg-info/ +wandb/ +.installed.cfg +*.egg + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +.hypothesis/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# pyenv +.python-version + +# celery beat schedule file +celerybeat-schedule + +# SageMath parsed files +*.sage.py + +# dotenv +.env + +# virtualenv +.venv* +venv*/ +ENV*/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ + + +# https://github.com/github/gitignore/blob/master/Global/macOS.gitignore ----------------------------------------------- + +# General +.DS_Store +.AppleDouble +.LSOverride + +# Icon must end with two \r +Icon +Icon? + +# Thumbnails +._* + +# Files that might appear in the root of a volume +.DocumentRevisions-V100 +.fseventsd +.Spotlight-V100 +.TemporaryItems +.Trashes +.VolumeIcon.icns +.com.apple.timemachine.donotpresent + +# Directories potentially created on remote AFP share +.AppleDB +.AppleDesktop +Network Trash Folder +Temporary Items +.apdisk + + +# https://github.com/github/gitignore/blob/master/Global/JetBrains.gitignore +# Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio and WebStorm +# Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839 + +# User-specific stuff: +.idea/* +.idea/**/workspace.xml +.idea/**/tasks.xml +.idea/dictionaries +.html # Bokeh Plots +.pg # TensorFlow Frozen Graphs +.avi # videos + +# Sensitive or high-churn files: +.idea/**/dataSources/ +.idea/**/dataSources.ids +.idea/**/dataSources.local.xml +.idea/**/sqlDataSources.xml +.idea/**/dynamic.xml +.idea/**/uiDesigner.xml + +# Gradle: +.idea/**/gradle.xml +.idea/**/libraries + +# CMake +cmake-build-debug/ +cmake-build-release/ + +# Mongo Explorer plugin: +.idea/**/mongoSettings.xml + +## File-based project format: +*.iws + +## Plugin-specific files: + +# IntelliJ +out/ + +# mpeltonen/sbt-idea plugin +.idea_modules/ + +# JIRA plugin +atlassian-ide-plugin.xml + +# Cursive Clojure plugin +.idea/replstate.xml + +# Crashlytics plugin (for Android Studio and IntelliJ) +com_crashlytics_export_strings.xml +crashlytics.properties +crashlytics-build.properties +fabric.properties diff --git a/application/yolov5_example/.gitattributes b/application/yolov5_example/.gitattributes new file mode 100644 index 00000000..dad4239e --- /dev/null +++ b/application/yolov5_example/.gitattributes @@ -0,0 +1,2 @@ +# this drop notebooks from GitHub language stats +*.ipynb linguist-vendored diff --git a/application/yolov5_example/.gitignore b/application/yolov5_example/.gitignore new file mode 100644 index 00000000..69a00843 --- /dev/null +++ b/application/yolov5_example/.gitignore @@ -0,0 +1,256 @@ +# Repo-specific GitIgnore ---------------------------------------------------------------------------------------------- +*.jpg +*.jpeg +*.png +*.bmp +*.tif +*.tiff +*.heic +*.JPG +*.JPEG +*.PNG +*.BMP +*.TIF +*.TIFF +*.HEIC +*.mp4 +*.mov +*.MOV +*.avi +*.data +*.json +*.cfg +!setup.cfg +!cfg/yolov3*.cfg + +storage.googleapis.com +runs/* +data/* +data/images/* +!data/*.yaml +!data/hyps +!data/scripts +!data/images +!data/images/zidane.jpg +!data/images/bus.jpg +!data/*.sh + +results*.csv + +# Datasets ------------------------------------------------------------------------------------------------------------- +coco/ +coco128/ +VOC/ + +# MATLAB GitIgnore ----------------------------------------------------------------------------------------------------- +*.m~ +*.mat +!targets*.mat + +# Neural Network weights ----------------------------------------------------------------------------------------------- +*.weights +*.pt +*.pb +*.onnx +*.engine +*.mlmodel +*.torchscript +*.tflite +*.h5 +*_saved_model/ +*_web_model/ +*_openvino_model/ +darknet53.conv.74 +yolov3-tiny.conv.15 + +# GitHub Python GitIgnore ---------------------------------------------------------------------------------------------- +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +env/ +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +*.egg-info/ +/wandb/ +.installed.cfg +*.egg + + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +.hypothesis/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# pyenv +.python-version + +# celery beat schedule file +celerybeat-schedule + +# SageMath parsed files +*.sage.py + +# dotenv +.env + +# virtualenv +.venv* +venv*/ +ENV*/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ + + +# https://github.com/github/gitignore/blob/master/Global/macOS.gitignore ----------------------------------------------- + +# General +.DS_Store +.AppleDouble +.LSOverride + +# Icon must end with two \r +Icon +Icon? + +# Thumbnails +._* + +# Files that might appear in the root of a volume +.DocumentRevisions-V100 +.fseventsd +.Spotlight-V100 +.TemporaryItems +.Trashes +.VolumeIcon.icns +.com.apple.timemachine.donotpresent + +# Directories potentially created on remote AFP share +.AppleDB +.AppleDesktop +Network Trash Folder +Temporary Items +.apdisk + + +# https://github.com/github/gitignore/blob/master/Global/JetBrains.gitignore +# Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio and WebStorm +# Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839 + +# User-specific stuff: +.idea/* +.idea/**/workspace.xml +.idea/**/tasks.xml +.idea/dictionaries +.html # Bokeh Plots +.pg # TensorFlow Frozen Graphs +.avi # videos + +# Sensitive or high-churn files: +.idea/**/dataSources/ +.idea/**/dataSources.ids +.idea/**/dataSources.local.xml +.idea/**/sqlDataSources.xml +.idea/**/dynamic.xml +.idea/**/uiDesigner.xml + +# Gradle: +.idea/**/gradle.xml +.idea/**/libraries + +# CMake +cmake-build-debug/ +cmake-build-release/ + +# Mongo Explorer plugin: +.idea/**/mongoSettings.xml + +## File-based project format: +*.iws + +## Plugin-specific files: + +# IntelliJ +out/ + +# mpeltonen/sbt-idea plugin +.idea_modules/ + +# JIRA plugin +atlassian-ide-plugin.xml + +# Cursive Clojure plugin +.idea/replstate.xml + +# Crashlytics plugin (for Android Studio and IntelliJ) +com_crashlytics_export_strings.xml +crashlytics.properties +crashlytics-build.properties +fabric.properties diff --git a/application/yolov5_example/.pre-commit-config.yaml b/application/yolov5_example/.pre-commit-config.yaml new file mode 100644 index 00000000..43aca019 --- /dev/null +++ b/application/yolov5_example/.pre-commit-config.yaml @@ -0,0 +1,64 @@ +# Define hooks for code formations +# Will be applied on any updated commit files if a user has installed and linked commit hook + +default_language_version: + python: python3.8 + +# Define bot property if installed via https://github.com/marketplace/pre-commit-ci +ci: + autofix_prs: true + autoupdate_commit_msg: '[pre-commit.ci] pre-commit suggestions' + autoupdate_schedule: monthly + # submodules: true + +repos: + - repo: https://github.com/pre-commit/pre-commit-hooks + rev: v4.3.0 + hooks: + # - id: end-of-file-fixer + - id: trailing-whitespace + - id: check-case-conflict + - id: check-yaml + - id: check-toml + - id: pretty-format-json + - id: check-docstring-first + + - repo: https://github.com/asottile/pyupgrade + rev: v2.37.3 + hooks: + - id: pyupgrade + name: Upgrade code + args: [ --py37-plus ] + + - repo: https://github.com/PyCQA/isort + rev: 5.10.1 + hooks: + - id: isort + name: Sort imports + + - repo: https://github.com/pre-commit/mirrors-yapf + rev: v0.32.0 + hooks: + - id: yapf + name: YAPF formatting + + - repo: https://github.com/executablebooks/mdformat + rev: 0.7.14 + hooks: + - id: mdformat + name: MD formatting + additional_dependencies: + - mdformat-gfm + - mdformat-black + exclude: "README.md|README_cn.md" + + - repo: https://github.com/asottile/yesqa + rev: v1.3.0 + hooks: + - id: yesqa + + - repo: https://github.com/PyCQA/flake8 + rev: 5.0.2 + hooks: + - id: flake8 + name: PEP8 diff --git a/application/yolov5_example/CONTRIBUTING.md b/application/yolov5_example/CONTRIBUTING.md new file mode 100644 index 00000000..13b9b73b --- /dev/null +++ b/application/yolov5_example/CONTRIBUTING.md @@ -0,0 +1,98 @@ +## Contributing to YOLOv5 🚀 + +We love your input! We want to make contributing to YOLOv5 as easy and transparent as possible, whether it's: + +- Reporting a bug +- Discussing the current state of the code +- Submitting a fix +- Proposing a new feature +- Becoming a maintainer + +YOLOv5 works so well due to our combined community effort, and for every small improvement you contribute you will be +helping push the frontiers of what's possible in AI 😃! + +## Submitting a Pull Request (PR) 🛠️ + +Submitting a PR is easy! This example shows how to submit a PR for updating `requirements.txt` in 4 steps: + +### 1. Select File to Update + +Select `requirements.txt` to update by clicking on it in GitHub. + +

PR_step1

+ +### 2. Click 'Edit this file' + +Button is in top-right corner. + +

PR_step2

+ +### 3. Make Changes + +Change `matplotlib` version from `3.2.2` to `3.3`. + +

PR_step3

+ +### 4. Preview Changes and Submit PR + +Click on the **Preview changes** tab to verify your updates. At the bottom of the screen select 'Create a **new branch** +for this commit', assign your branch a descriptive name such as `fix/matplotlib_version` and click the green **Propose +changes** button. All done, your PR is now submitted to YOLOv5 for review and approval 😃! + +

PR_step4

+ +### PR recommendations + +To allow your work to be integrated as seamlessly as possible, we advise you to: + +- ✅ Verify your PR is **up-to-date with upstream/master.** If your PR is behind upstream/master an + automatic [GitHub Actions](https://github.com/ultralytics/yolov5/blob/master/.github/workflows/rebase.yml) merge may + be attempted by writing /rebase in a new comment, or by running the following code, replacing 'feature' with the name + of your local branch: + +```bash +git remote add upstream https://github.com/ultralytics/yolov5.git +git fetch upstream +# git checkout feature # <--- replace 'feature' with local branch name +git merge upstream/master +git push -u origin -f +``` + +- ✅ Verify all Continuous Integration (CI) **checks are passing**. +- ✅ Reduce changes to the absolute **minimum** required for your bug fix or feature addition. _"It is not daily increase + but daily decrease, hack away the unessential. The closer to the source, the less wastage there is."_ — Bruce Lee + +## Submitting a Bug Report 🐛 + +If you spot a problem with YOLOv5 please submit a Bug Report! + +For us to start investigating a possible problem we need to be able to reproduce it ourselves first. We've created a few +short guidelines below to help users provide what we need in order to get started. + +When asking a question, people will be better able to provide help if you provide **code** that they can easily +understand and use to **reproduce** the problem. This is referred to by community members as creating +a [minimum reproducible example](https://stackoverflow.com/help/minimal-reproducible-example). Your code that reproduces +the problem should be: + +- ✅ **Minimal** – Use as little code as possible that still produces the same problem +- ✅ **Complete** – Provide **all** parts someone else needs to reproduce your problem in the question itself +- ✅ **Reproducible** – Test the code you're about to provide to make sure it reproduces the problem + +In addition to the above requirements, for [Ultralytics](https://ultralytics.com/) to provide assistance your code +should be: + +- ✅ **Current** – Verify that your code is up-to-date with current + GitHub [master](https://github.com/ultralytics/yolov5/tree/master), and if necessary `git pull` or `git clone` a new + copy to ensure your problem has not already been resolved by previous commits. +- ✅ **Unmodified** – Your problem must be reproducible without any modifications to the codebase in this + repository. [Ultralytics](https://ultralytics.com/) does not provide support for custom code ⚠️. + +If you believe your problem meets all of the above criteria, please close this issue and raise a new one using the 🐛 +**Bug Report** [template](https://github.com/ultralytics/yolov5/issues/new/choose) and providing +a [minimum reproducible example](https://stackoverflow.com/help/minimal-reproducible-example) to help us better +understand and diagnose your problem. + +## License + +By contributing, you agree that your contributions will be licensed under +the [GPL-3.0 license](https://choosealicense.com/licenses/gpl-3.0/) diff --git a/application/yolov5_example/LICENSE b/application/yolov5_example/LICENSE new file mode 100644 index 00000000..92b370f0 --- /dev/null +++ b/application/yolov5_example/LICENSE @@ -0,0 +1,674 @@ +GNU GENERAL PUBLIC LICENSE + Version 3, 29 June 2007 + + Copyright (C) 2007 Free Software Foundation, Inc. + Everyone is permitted to copy and distribute verbatim copies + of this license document, but changing it is not allowed. + + Preamble + + The GNU General Public License is a free, copyleft license for +software and other kinds of works. + + The licenses for most software and other practical works are designed +to take away your freedom to share and change the works. By contrast, +the GNU General Public License is intended to guarantee your freedom to +share and change all versions of a program--to make sure it remains free +software for all its users. We, the Free Software Foundation, use the +GNU General Public License for most of our software; it applies also to +any other work released this way by its authors. You can apply it to +your programs, too. + + When we speak of free software, we are referring to freedom, not +price. Our General Public Licenses are designed to make sure that you +have the freedom to distribute copies of free software (and charge for +them if you wish), that you receive source code or can get it if you +want it, that you can change the software or use pieces of it in new +free programs, and that you know you can do these things. + + To protect your rights, we need to prevent others from denying you +these rights or asking you to surrender the rights. Therefore, you have +certain responsibilities if you distribute copies of the software, or if +you modify it: responsibilities to respect the freedom of others. + + For example, if you distribute copies of such a program, whether +gratis or for a fee, you must pass on to the recipients the same +freedoms that you received. You must make sure that they, too, receive +or can get the source code. And you must show them these terms so they +know their rights. + + Developers that use the GNU GPL protect your rights with two steps: +(1) assert copyright on the software, and (2) offer you this License +giving you legal permission to copy, distribute and/or modify it. + + For the developers' and authors' protection, the GPL clearly explains +that there is no warranty for this free software. For both users' and +authors' sake, the GPL requires that modified versions be marked as +changed, so that their problems will not be attributed erroneously to +authors of previous versions. + + Some devices are designed to deny users access to install or run +modified versions of the software inside them, although the manufacturer +can do so. This is fundamentally incompatible with the aim of +protecting users' freedom to change the software. The systematic +pattern of such abuse occurs in the area of products for individuals to +use, which is precisely where it is most unacceptable. Therefore, we +have designed this version of the GPL to prohibit the practice for those +products. If such problems arise substantially in other domains, we +stand ready to extend this provision to those domains in future versions +of the GPL, as needed to protect the freedom of users. + + Finally, every program is threatened constantly by software patents. +States should not allow patents to restrict development and use of +software on general-purpose computers, but in those that do, we wish to +avoid the special danger that patents applied to a free program could +make it effectively proprietary. To prevent this, the GPL assures that +patents cannot be used to render the program non-free. + + The precise terms and conditions for copying, distribution and +modification follow. + + TERMS AND CONDITIONS + + 0. Definitions. + + "This License" refers to version 3 of the GNU General Public License. + + "Copyright" also means copyright-like laws that apply to other kinds of +works, such as semiconductor masks. + + "The Program" refers to any copyrightable work licensed under this +License. Each licensee is addressed as "you". "Licensees" and +"recipients" may be individuals or organizations. + + To "modify" a work means to copy from or adapt all or part of the work +in a fashion requiring copyright permission, other than the making of an +exact copy. The resulting work is called a "modified version" of the +earlier work or a work "based on" the earlier work. + + A "covered work" means either the unmodified Program or a work based +on the Program. + + To "propagate" a work means to do anything with it that, without +permission, would make you directly or secondarily liable for +infringement under applicable copyright law, except executing it on a +computer or modifying a private copy. Propagation includes copying, +distribution (with or without modification), making available to the +public, and in some countries other activities as well. + + To "convey" a work means any kind of propagation that enables other +parties to make or receive copies. Mere interaction with a user through +a computer network, with no transfer of a copy, is not conveying. + + An interactive user interface displays "Appropriate Legal Notices" +to the extent that it includes a convenient and prominently visible +feature that (1) displays an appropriate copyright notice, and (2) +tells the user that there is no warranty for the work (except to the +extent that warranties are provided), that licensees may convey the +work under this License, and how to view a copy of this License. If +the interface presents a list of user commands or options, such as a +menu, a prominent item in the list meets this criterion. + + 1. Source Code. + + The "source code" for a work means the preferred form of the work +for making modifications to it. "Object code" means any non-source +form of a work. + + A "Standard Interface" means an interface that either is an official +standard defined by a recognized standards body, or, in the case of +interfaces specified for a particular programming language, one that +is widely used among developers working in that language. + + The "System Libraries" of an executable work include anything, other +than the work as a whole, that (a) is included in the normal form of +packaging a Major Component, but which is not part of that Major +Component, and (b) serves only to enable use of the work with that +Major Component, or to implement a Standard Interface for which an +implementation is available to the public in source code form. A +"Major Component", in this context, means a major essential component +(kernel, window system, and so on) of the specific operating system +(if any) on which the executable work runs, or a compiler used to +produce the work, or an object code interpreter used to run it. + + The "Corresponding Source" for a work in object code form means all +the source code needed to generate, install, and (for an executable +work) run the object code and to modify the work, including scripts to +control those activities. However, it does not include the work's +System Libraries, or general-purpose tools or generally available free +programs which are used unmodified in performing those activities but +which are not part of the work. For example, Corresponding Source +includes interface definition files associated with source files for +the work, and the source code for shared libraries and dynamically +linked subprograms that the work is specifically designed to require, +such as by intimate data communication or control flow between those +subprograms and other parts of the work. + + The Corresponding Source need not include anything that users +can regenerate automatically from other parts of the Corresponding +Source. + + The Corresponding Source for a work in source code form is that +same work. + + 2. Basic Permissions. + + All rights granted under this License are granted for the term of +copyright on the Program, and are irrevocable provided the stated +conditions are met. This License explicitly affirms your unlimited +permission to run the unmodified Program. The output from running a +covered work is covered by this License only if the output, given its +content, constitutes a covered work. This License acknowledges your +rights of fair use or other equivalent, as provided by copyright law. + + You may make, run and propagate covered works that you do not +convey, without conditions so long as your license otherwise remains +in force. You may convey covered works to others for the sole purpose +of having them make modifications exclusively for you, or provide you +with facilities for running those works, provided that you comply with +the terms of this License in conveying all material for which you do +not control copyright. Those thus making or running the covered works +for you must do so exclusively on your behalf, under your direction +and control, on terms that prohibit them from making any copies of +your copyrighted material outside their relationship with you. + + Conveying under any other circumstances is permitted solely under +the conditions stated below. Sublicensing is not allowed; section 10 +makes it unnecessary. + + 3. Protecting Users' Legal Rights From Anti-Circumvention Law. + + No covered work shall be deemed part of an effective technological +measure under any applicable law fulfilling obligations under article +11 of the WIPO copyright treaty adopted on 20 December 1996, or +similar laws prohibiting or restricting circumvention of such +measures. + + When you convey a covered work, you waive any legal power to forbid +circumvention of technological measures to the extent such circumvention +is effected by exercising rights under this License with respect to +the covered work, and you disclaim any intention to limit operation or +modification of the work as a means of enforcing, against the work's +users, your or third parties' legal rights to forbid circumvention of +technological measures. + + 4. Conveying Verbatim Copies. + + You may convey verbatim copies of the Program's source code as you +receive it, in any medium, provided that you conspicuously and +appropriately publish on each copy an appropriate copyright notice; +keep intact all notices stating that this License and any +non-permissive terms added in accord with section 7 apply to the code; +keep intact all notices of the absence of any warranty; and give all +recipients a copy of this License along with the Program. + + You may charge any price or no price for each copy that you convey, +and you may offer support or warranty protection for a fee. + + 5. Conveying Modified Source Versions. + + You may convey a work based on the Program, or the modifications to +produce it from the Program, in the form of source code under the +terms of section 4, provided that you also meet all of these conditions: + + a) The work must carry prominent notices stating that you modified + it, and giving a relevant date. + + b) The work must carry prominent notices stating that it is + released under this License and any conditions added under section + 7. This requirement modifies the requirement in section 4 to + "keep intact all notices". + + c) You must license the entire work, as a whole, under this + License to anyone who comes into possession of a copy. This + License will therefore apply, along with any applicable section 7 + additional terms, to the whole of the work, and all its parts, + regardless of how they are packaged. This License gives no + permission to license the work in any other way, but it does not + invalidate such permission if you have separately received it. + + d) If the work has interactive user interfaces, each must display + Appropriate Legal Notices; however, if the Program has interactive + interfaces that do not display Appropriate Legal Notices, your + work need not make them do so. + + A compilation of a covered work with other separate and independent +works, which are not by their nature extensions of the covered work, +and which are not combined with it such as to form a larger program, +in or on a volume of a storage or distribution medium, is called an +"aggregate" if the compilation and its resulting copyright are not +used to limit the access or legal rights of the compilation's users +beyond what the individual works permit. Inclusion of a covered work +in an aggregate does not cause this License to apply to the other +parts of the aggregate. + + 6. Conveying Non-Source Forms. + + You may convey a covered work in object code form under the terms +of sections 4 and 5, provided that you also convey the +machine-readable Corresponding Source under the terms of this License, +in one of these ways: + + a) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by the + Corresponding Source fixed on a durable physical medium + customarily used for software interchange. + + b) Convey the object code in, or embodied in, a physical product + (including a physical distribution medium), accompanied by a + written offer, valid for at least three years and valid for as + long as you offer spare parts or customer support for that product + model, to give anyone who possesses the object code either (1) a + copy of the Corresponding Source for all the software in the + product that is covered by this License, on a durable physical + medium customarily used for software interchange, for a price no + more than your reasonable cost of physically performing this + conveying of source, or (2) access to copy the + Corresponding Source from a network server at no charge. + + c) Convey individual copies of the object code with a copy of the + written offer to provide the Corresponding Source. This + alternative is allowed only occasionally and noncommercially, and + only if you received the object code with such an offer, in accord + with subsection 6b. + + d) Convey the object code by offering access from a designated + place (gratis or for a charge), and offer equivalent access to the + Corresponding Source in the same way through the same place at no + further charge. You need not require recipients to copy the + Corresponding Source along with the object code. If the place to + copy the object code is a network server, the Corresponding Source + may be on a different server (operated by you or a third party) + that supports equivalent copying facilities, provided you maintain + clear directions next to the object code saying where to find the + Corresponding Source. Regardless of what server hosts the + Corresponding Source, you remain obligated to ensure that it is + available for as long as needed to satisfy these requirements. + + e) Convey the object code using peer-to-peer transmission, provided + you inform other peers where the object code and Corresponding + Source of the work are being offered to the general public at no + charge under subsection 6d. + + A separable portion of the object code, whose source code is excluded +from the Corresponding Source as a System Library, need not be +included in conveying the object code work. + + A "User Product" is either (1) a "consumer product", which means any +tangible personal property which is normally used for personal, family, +or household purposes, or (2) anything designed or sold for incorporation +into a dwelling. In determining whether a product is a consumer product, +doubtful cases shall be resolved in favor of coverage. For a particular +product received by a particular user, "normally used" refers to a +typical or common use of that class of product, regardless of the status +of the particular user or of the way in which the particular user +actually uses, or expects or is expected to use, the product. A product +is a consumer product regardless of whether the product has substantial +commercial, industrial or non-consumer uses, unless such uses represent +the only significant mode of use of the product. + + "Installation Information" for a User Product means any methods, +procedures, authorization keys, or other information required to install +and execute modified versions of a covered work in that User Product from +a modified version of its Corresponding Source. The information must +suffice to ensure that the continued functioning of the modified object +code is in no case prevented or interfered with solely because +modification has been made. + + If you convey an object code work under this section in, or with, or +specifically for use in, a User Product, and the conveying occurs as +part of a transaction in which the right of possession and use of the +User Product is transferred to the recipient in perpetuity or for a +fixed term (regardless of how the transaction is characterized), the +Corresponding Source conveyed under this section must be accompanied +by the Installation Information. But this requirement does not apply +if neither you nor any third party retains the ability to install +modified object code on the User Product (for example, the work has +been installed in ROM). + + The requirement to provide Installation Information does not include a +requirement to continue to provide support service, warranty, or updates +for a work that has been modified or installed by the recipient, or for +the User Product in which it has been modified or installed. Access to a +network may be denied when the modification itself materially and +adversely affects the operation of the network or violates the rules and +protocols for communication across the network. + + Corresponding Source conveyed, and Installation Information provided, +in accord with this section must be in a format that is publicly +documented (and with an implementation available to the public in +source code form), and must require no special password or key for +unpacking, reading or copying. + + 7. Additional Terms. + + "Additional permissions" are terms that supplement the terms of this +License by making exceptions from one or more of its conditions. +Additional permissions that are applicable to the entire Program shall +be treated as though they were included in this License, to the extent +that they are valid under applicable law. If additional permissions +apply only to part of the Program, that part may be used separately +under those permissions, but the entire Program remains governed by +this License without regard to the additional permissions. + + When you convey a copy of a covered work, you may at your option +remove any additional permissions from that copy, or from any part of +it. (Additional permissions may be written to require their own +removal in certain cases when you modify the work.) You may place +additional permissions on material, added by you to a covered work, +for which you have or can give appropriate copyright permission. + + Notwithstanding any other provision of this License, for material you +add to a covered work, you may (if authorized by the copyright holders of +that material) supplement the terms of this License with terms: + + a) Disclaiming warranty or limiting liability differently from the + terms of sections 15 and 16 of this License; or + + b) Requiring preservation of specified reasonable legal notices or + author attributions in that material or in the Appropriate Legal + Notices displayed by works containing it; or + + c) Prohibiting misrepresentation of the origin of that material, or + requiring that modified versions of such material be marked in + reasonable ways as different from the original version; or + + d) Limiting the use for publicity purposes of names of licensors or + authors of the material; or + + e) Declining to grant rights under trademark law for use of some + trade names, trademarks, or service marks; or + + f) Requiring indemnification of licensors and authors of that + material by anyone who conveys the material (or modified versions of + it) with contractual assumptions of liability to the recipient, for + any liability that these contractual assumptions directly impose on + those licensors and authors. + + All other non-permissive additional terms are considered "further +restrictions" within the meaning of section 10. If the Program as you +received it, or any part of it, contains a notice stating that it is +governed by this License along with a term that is a further +restriction, you may remove that term. If a license document contains +a further restriction but permits relicensing or conveying under this +License, you may add to a covered work material governed by the terms +of that license document, provided that the further restriction does +not survive such relicensing or conveying. + + If you add terms to a covered work in accord with this section, you +must place, in the relevant source files, a statement of the +additional terms that apply to those files, or a notice indicating +where to find the applicable terms. + + Additional terms, permissive or non-permissive, may be stated in the +form of a separately written license, or stated as exceptions; +the above requirements apply either way. + + 8. Termination. + + You may not propagate or modify a covered work except as expressly +provided under this License. Any attempt otherwise to propagate or +modify it is void, and will automatically terminate your rights under +this License (including any patent licenses granted under the third +paragraph of section 11). + + However, if you cease all violation of this License, then your +license from a particular copyright holder is reinstated (a) +provisionally, unless and until the copyright holder explicitly and +finally terminates your license, and (b) permanently, if the copyright +holder fails to notify you of the violation by some reasonable means +prior to 60 days after the cessation. + + Moreover, your license from a particular copyright holder is +reinstated permanently if the copyright holder notifies you of the +violation by some reasonable means, this is the first time you have +received notice of violation of this License (for any work) from that +copyright holder, and you cure the violation prior to 30 days after +your receipt of the notice. + + Termination of your rights under this section does not terminate the +licenses of parties who have received copies or rights from you under +this License. If your rights have been terminated and not permanently +reinstated, you do not qualify to receive new licenses for the same +material under section 10. + + 9. Acceptance Not Required for Having Copies. + + You are not required to accept this License in order to receive or +run a copy of the Program. Ancillary propagation of a covered work +occurring solely as a consequence of using peer-to-peer transmission +to receive a copy likewise does not require acceptance. However, +nothing other than this License grants you permission to propagate or +modify any covered work. These actions infringe copyright if you do +not accept this License. Therefore, by modifying or propagating a +covered work, you indicate your acceptance of this License to do so. + + 10. Automatic Licensing of Downstream Recipients. + + Each time you convey a covered work, the recipient automatically +receives a license from the original licensors, to run, modify and +propagate that work, subject to this License. You are not responsible +for enforcing compliance by third parties with this License. + + An "entity transaction" is a transaction transferring control of an +organization, or substantially all assets of one, or subdividing an +organization, or merging organizations. If propagation of a covered +work results from an entity transaction, each party to that +transaction who receives a copy of the work also receives whatever +licenses to the work the party's predecessor in interest had or could +give under the previous paragraph, plus a right to possession of the +Corresponding Source of the work from the predecessor in interest, if +the predecessor has it or can get it with reasonable efforts. + + You may not impose any further restrictions on the exercise of the +rights granted or affirmed under this License. For example, you may +not impose a license fee, royalty, or other charge for exercise of +rights granted under this License, and you may not initiate litigation +(including a cross-claim or counterclaim in a lawsuit) alleging that +any patent claim is infringed by making, using, selling, offering for +sale, or importing the Program or any portion of it. + + 11. Patents. + + A "contributor" is a copyright holder who authorizes use under this +License of the Program or a work on which the Program is based. The +work thus licensed is called the contributor's "contributor version". + + A contributor's "essential patent claims" are all patent claims +owned or controlled by the contributor, whether already acquired or +hereafter acquired, that would be infringed by some manner, permitted +by this License, of making, using, or selling its contributor version, +but do not include claims that would be infringed only as a +consequence of further modification of the contributor version. For +purposes of this definition, "control" includes the right to grant +patent sublicenses in a manner consistent with the requirements of +this License. + + Each contributor grants you a non-exclusive, worldwide, royalty-free +patent license under the contributor's essential patent claims, to +make, use, sell, offer for sale, import and otherwise run, modify and +propagate the contents of its contributor version. + + In the following three paragraphs, a "patent license" is any express +agreement or commitment, however denominated, not to enforce a patent +(such as an express permission to practice a patent or covenant not to +sue for patent infringement). To "grant" such a patent license to a +party means to make such an agreement or commitment not to enforce a +patent against the party. + + If you convey a covered work, knowingly relying on a patent license, +and the Corresponding Source of the work is not available for anyone +to copy, free of charge and under the terms of this License, through a +publicly available network server or other readily accessible means, +then you must either (1) cause the Corresponding Source to be so +available, or (2) arrange to deprive yourself of the benefit of the +patent license for this particular work, or (3) arrange, in a manner +consistent with the requirements of this License, to extend the patent +license to downstream recipients. "Knowingly relying" means you have +actual knowledge that, but for the patent license, your conveying the +covered work in a country, or your recipient's use of the covered work +in a country, would infringe one or more identifiable patents in that +country that you have reason to believe are valid. + + If, pursuant to or in connection with a single transaction or +arrangement, you convey, or propagate by procuring conveyance of, a +covered work, and grant a patent license to some of the parties +receiving the covered work authorizing them to use, propagate, modify +or convey a specific copy of the covered work, then the patent license +you grant is automatically extended to all recipients of the covered +work and works based on it. + + A patent license is "discriminatory" if it does not include within +the scope of its coverage, prohibits the exercise of, or is +conditioned on the non-exercise of one or more of the rights that are +specifically granted under this License. You may not convey a covered +work if you are a party to an arrangement with a third party that is +in the business of distributing software, under which you make payment +to the third party based on the extent of your activity of conveying +the work, and under which the third party grants, to any of the +parties who would receive the covered work from you, a discriminatory +patent license (a) in connection with copies of the covered work +conveyed by you (or copies made from those copies), or (b) primarily +for and in connection with specific products or compilations that +contain the covered work, unless you entered into that arrangement, +or that patent license was granted, prior to 28 March 2007. + + Nothing in this License shall be construed as excluding or limiting +any implied license or other defenses to infringement that may +otherwise be available to you under applicable patent law. + + 12. No Surrender of Others' Freedom. + + If conditions are imposed on you (whether by court order, agreement or +otherwise) that contradict the conditions of this License, they do not +excuse you from the conditions of this License. If you cannot convey a +covered work so as to satisfy simultaneously your obligations under this +License and any other pertinent obligations, then as a consequence you may +not convey it at all. For example, if you agree to terms that obligate you +to collect a royalty for further conveying from those to whom you convey +the Program, the only way you could satisfy both those terms and this +License would be to refrain entirely from conveying the Program. + + 13. Use with the GNU Affero General Public License. + + Notwithstanding any other provision of this License, you have +permission to link or combine any covered work with a work licensed +under version 3 of the GNU Affero General Public License into a single +combined work, and to convey the resulting work. The terms of this +License will continue to apply to the part which is the covered work, +but the special requirements of the GNU Affero General Public License, +section 13, concerning interaction through a network will apply to the +combination as such. + + 14. Revised Versions of this License. + + The Free Software Foundation may publish revised and/or new versions of +the GNU General Public License from time to time. Such new versions will +be similar in spirit to the present version, but may differ in detail to +address new problems or concerns. + + Each version is given a distinguishing version number. If the +Program specifies that a certain numbered version of the GNU General +Public License "or any later version" applies to it, you have the +option of following the terms and conditions either of that numbered +version or of any later version published by the Free Software +Foundation. If the Program does not specify a version number of the +GNU General Public License, you may choose any version ever published +by the Free Software Foundation. + + If the Program specifies that a proxy can decide which future +versions of the GNU General Public License can be used, that proxy's +public statement of acceptance of a version permanently authorizes you +to choose that version for the Program. + + Later license versions may give you additional or different +permissions. However, no additional obligations are imposed on any +author or copyright holder as a result of your choosing to follow a +later version. + + 15. Disclaimer of Warranty. + + THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY +APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT +HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY +OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, +THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR +PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM +IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF +ALL NECESSARY SERVICING, REPAIR OR CORRECTION. + + 16. Limitation of Liability. + + IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING +WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS +THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY +GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE +USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF +DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD +PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), +EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF +SUCH DAMAGES. + + 17. Interpretation of Sections 15 and 16. + + If the disclaimer of warranty and limitation of liability provided +above cannot be given local legal effect according to their terms, +reviewing courts shall apply local law that most closely approximates +an absolute waiver of all civil liability in connection with the +Program, unless a warranty or assumption of liability accompanies a +copy of the Program in return for a fee. + + END OF TERMS AND CONDITIONS + + How to Apply These Terms to Your New Programs + + If you develop a new program, and you want it to be of the greatest +possible use to the public, the best way to achieve this is to make it +free software which everyone can redistribute and change under these terms. + + To do so, attach the following notices to the program. It is safest +to attach them to the start of each source file to most effectively +state the exclusion of warranty; and each file should have at least +the "copyright" line and a pointer to where the full notice is found. + + + Copyright (C) + + This program is free software: you can redistribute it and/or modify + it under the terms of the GNU General Public License as published by + the Free Software Foundation, either version 3 of the License, or + (at your option) any later version. + + This program is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + GNU General Public License for more details. + + You should have received a copy of the GNU General Public License + along with this program. If not, see . + +Also add information on how to contact you by electronic and paper mail. + + If the program does terminal interaction, make it output a short +notice like this when it starts in an interactive mode: + + Copyright (C) + This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. + This is free software, and you are welcome to redistribute it + under certain conditions; type `show c' for details. + +The hypothetical commands `show w' and `show c' should show the appropriate +parts of the General Public License. Of course, your program's commands +might be different; for a GUI interface, you would use an "about box". + + You should also get your employer (if you work as a programmer) or school, +if any, to sign a "copyright disclaimer" for the program, if necessary. +For more information on this, and how to apply and follow the GNU GPL, see +. + + The GNU General Public License does not permit incorporating your program +into proprietary programs. If your program is a subroutine library, you +may consider it more useful to permit linking proprietary applications with +the library. If this is what you want to do, use the GNU Lesser General +Public License instead of this License. But first, please read +. diff --git a/application/yolov5_example/README.md b/application/yolov5_example/README.md new file mode 100644 index 00000000..b368d1d6 --- /dev/null +++ b/application/yolov5_example/README.md @@ -0,0 +1,363 @@ +
+

+ + +

+ +English | [简体中文](.github/README_cn.md) +
+
+ CI CPU testing + YOLOv5 Citation + Docker Pulls +
+ Open In Colab + Open In Kaggle + Join Forum +
+ +
+

+YOLOv5 🚀 is a family of object detection architectures and models pretrained on the COCO dataset, and represents Ultralytics + open-source research into future vision AI methods, incorporating lessons learned and best practices evolved over thousands of hours of research and development. +

+ +
+ + + + + + + + + + + + + + + + + + + + +
+ + + +
+ +##
Documentation
+ +See the [YOLOv5 Docs](https://docs.ultralytics.com) for full documentation on training, testing and deployment. + +##
Quick Start Examples
+ +
+Install + +Clone repo and install [requirements.txt](https://github.com/ultralytics/yolov5/blob/master/requirements.txt) in a +[**Python>=3.7.0**](https://www.python.org/) environment, including +[**PyTorch>=1.7**](https://pytorch.org/get-started/locally/). + +```bash +git clone https://github.com/ultralytics/yolov5 # clone +cd yolov5 +pip install -r requirements.txt # install +``` + +
+ +
+Inference + +YOLOv5 [PyTorch Hub](https://github.com/ultralytics/yolov5/issues/36) inference. [Models](https://github.com/ultralytics/yolov5/tree/master/models) download automatically from the latest +YOLOv5 [release](https://github.com/ultralytics/yolov5/releases). + +```python +import torch + +# Model +model = torch.hub.load('ultralytics/yolov5', 'yolov5s') # or yolov5n - yolov5x6, custom + +# Images +img = 'https://ultralytics.com/images/zidane.jpg' # or file, Path, PIL, OpenCV, numpy, list + +# Inference +results = model(img) + +# Results +results.print() # or .show(), .save(), .crop(), .pandas(), etc. +``` + +
+ +
+Inference with detect.py + +`detect.py` runs inference on a variety of sources, downloading [models](https://github.com/ultralytics/yolov5/tree/master/models) automatically from +the latest YOLOv5 [release](https://github.com/ultralytics/yolov5/releases) and saving results to `runs/detect`. + +```bash +python detect.py --source 0 # webcam + img.jpg # image + vid.mp4 # video + path/ # directory + 'path/*.jpg' # glob + 'https://youtu.be/Zgi9g1ksQHc' # YouTube + 'rtsp://example.com/media.mp4' # RTSP, RTMP, HTTP stream +``` + +
+ +
+Training + +The commands below reproduce YOLOv5 [COCO](https://github.com/ultralytics/yolov5/blob/master/data/scripts/get_coco.sh) +results. [Models](https://github.com/ultralytics/yolov5/tree/master/models) +and [datasets](https://github.com/ultralytics/yolov5/tree/master/data) download automatically from the latest +YOLOv5 [release](https://github.com/ultralytics/yolov5/releases). Training times for YOLOv5n/s/m/l/x are +1/2/4/6/8 days on a V100 GPU ([Multi-GPU](https://github.com/ultralytics/yolov5/issues/475) times faster). Use the +largest `--batch-size` possible, or pass `--batch-size -1` for +YOLOv5 [AutoBatch](https://github.com/ultralytics/yolov5/pull/5092). Batch sizes shown for V100-16GB. + +```bash +python train.py --data coco.yaml --cfg yolov5n.yaml --weights '' --batch-size 128 + yolov5s 64 + yolov5m 40 + yolov5l 24 + yolov5x 16 +``` + + + +
+ +
+Tutorials + +- [Train Custom Data](https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data)  🚀 RECOMMENDED +- [Tips for Best Training Results](https://github.com/ultralytics/yolov5/wiki/Tips-for-Best-Training-Results)  ☘️ + RECOMMENDED +- [Multi-GPU Training](https://github.com/ultralytics/yolov5/issues/475) +- [PyTorch Hub](https://github.com/ultralytics/yolov5/issues/36) 🌟 NEW +- [TFLite, ONNX, CoreML, TensorRT Export](https://github.com/ultralytics/yolov5/issues/251) 🚀 +- [Test-Time Augmentation (TTA)](https://github.com/ultralytics/yolov5/issues/303) +- [Model Ensembling](https://github.com/ultralytics/yolov5/issues/318) +- [Model Pruning/Sparsity](https://github.com/ultralytics/yolov5/issues/304) +- [Hyperparameter Evolution](https://github.com/ultralytics/yolov5/issues/607) +- [Transfer Learning with Frozen Layers](https://github.com/ultralytics/yolov5/issues/1314) +- [Architecture Summary](https://github.com/ultralytics/yolov5/issues/6998) 🌟 NEW +- [Weights & Biases Logging](https://github.com/ultralytics/yolov5/issues/1289) +- [Roboflow for Datasets, Labeling, and Active Learning](https://github.com/ultralytics/yolov5/issues/4975)  🌟 NEW +- [ClearML Logging](https://github.com/ultralytics/yolov5/tree/master/utils/loggers/clearml) 🌟 NEW +- [Deci Platform](https://github.com/ultralytics/yolov5/wiki/Deci-Platform) 🌟 NEW + +
+ +##
Environments
+ +Get started in seconds with our verified environments. Click each icon below for details. + +
+ + + + + + + + + + + + + + +
+ +##
Integrations
+ +
+ + + + + + + + + + + +
+ +|Deci ⭐ NEW|ClearML ⭐ NEW|Roboflow|Weights & Biases +|:-:|:-:|:-:|:-:| +|Automatically compile and quantize YOLOv5 for better inference performance in one click at [Deci](https://bit.ly/yolov5-deci-platform)|Automatically track, visualize and even remotely train YOLOv5 using [ClearML](https://cutt.ly/yolov5-readme-clearml) (open-source!)|Label and export your custom datasets directly to YOLOv5 for training with [Roboflow](https://roboflow.com/?ref=ultralytics) |Automatically track and visualize all your YOLOv5 training runs in the cloud with [Weights & Biases](https://wandb.ai/site?utm_campaign=repo_yolo_readme) + + +##
Why YOLOv5
+ +

+
+ YOLOv5-P5 640 Figure (click to expand) + +

+
+
+ Figure Notes (click to expand) + +- **COCO AP val** denotes mAP@0.5:0.95 metric measured on the 5000-image [COCO val2017](http://cocodataset.org) dataset over various inference sizes from 256 to 1536. +- **GPU Speed** measures average inference time per image on [COCO val2017](http://cocodataset.org) dataset using a [AWS p3.2xlarge](https://aws.amazon.com/ec2/instance-types/p3/) V100 instance at batch-size 32. +- **EfficientDet** data from [google/automl](https://github.com/google/automl) at batch size 8. +- **Reproduce** by `python val.py --task study --data coco.yaml --iou 0.7 --weights yolov5n6.pt yolov5s6.pt yolov5m6.pt yolov5l6.pt yolov5x6.pt` + +
+ +### Pretrained Checkpoints + +| Model | size
(pixels) | mAPval
0.5:0.95 | mAPval
0.5 | Speed
CPU b1
(ms) | Speed
V100 b1
(ms) | Speed
V100 b32
(ms) | params
(M) | FLOPs
@640 (B) | +|------------------------------------------------------------------------------------------------------|-----------------------|-------------------------|--------------------|------------------------------|-------------------------------|--------------------------------|--------------------|------------------------| +| [YOLOv5n](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5n.pt) | 640 | 28.0 | 45.7 | **45** | **6.3** | **0.6** | **1.9** | **4.5** | +| [YOLOv5s](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5s.pt) | 640 | 37.4 | 56.8 | 98 | 6.4 | 0.9 | 7.2 | 16.5 | +| [YOLOv5m](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5m.pt) | 640 | 45.4 | 64.1 | 224 | 8.2 | 1.7 | 21.2 | 49.0 | +| [YOLOv5l](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5l.pt) | 640 | 49.0 | 67.3 | 430 | 10.1 | 2.7 | 46.5 | 109.1 | +| [YOLOv5x](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5x.pt) | 640 | 50.7 | 68.9 | 766 | 12.1 | 4.8 | 86.7 | 205.7 | +| | | | | | | | | | +| [YOLOv5n6](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5n6.pt) | 1280 | 36.0 | 54.4 | 153 | 8.1 | 2.1 | 3.2 | 4.6 | +| [YOLOv5s6](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5s6.pt) | 1280 | 44.8 | 63.7 | 385 | 8.2 | 3.6 | 12.6 | 16.8 | +| [YOLOv5m6](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5m6.pt) | 1280 | 51.3 | 69.3 | 887 | 11.1 | 6.8 | 35.7 | 50.0 | +| [YOLOv5l6](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5l6.pt) | 1280 | 53.7 | 71.3 | 1784 | 15.8 | 10.5 | 76.8 | 111.4 | +| [YOLOv5x6](https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5x6.pt)
+ [TTA][TTA] | 1280
1536 | 55.0
**55.8** | 72.7
**72.7** | 3136
- | 26.2
- | 19.4
- | 140.7
- | 209.8
- | + +
+ Table Notes (click to expand) + +- All checkpoints are trained to 300 epochs with default settings. Nano and Small models use [hyp.scratch-low.yaml](https://github.com/ultralytics/yolov5/blob/master/data/hyps/hyp.scratch-low.yaml) hyps, all others use [hyp.scratch-high.yaml](https://github.com/ultralytics/yolov5/blob/master/data/hyps/hyp.scratch-high.yaml). +- **mAPval** values are for single-model single-scale on [COCO val2017](http://cocodataset.org) dataset.
Reproduce by `python val.py --data coco.yaml --img 640 --conf 0.001 --iou 0.65` +- **Speed** averaged over COCO val images using a [AWS p3.2xlarge](https://aws.amazon.com/ec2/instance-types/p3/) instance. NMS times (~1 ms/img) not included.
Reproduce by `python val.py --data coco.yaml --img 640 --task speed --batch 1` +- **TTA** [Test Time Augmentation](https://github.com/ultralytics/yolov5/issues/303) includes reflection and scale augmentations.
Reproduce by `python val.py --data coco.yaml --img 1536 --iou 0.7 --augment` + +
+ +##
Classification ⭐ NEW
+ +YOLOv5 [release v6.2](https://github.com/ultralytics/yolov5/releases) brings support for classification model training, validation, prediction and export! We've made training classifier models super simple. Click below to get started. + +
+ Classification Checkpoints (click to expand) + +
+ +We trained YOLOv5-cls classification models on ImageNet for 90 epochs using a 4xA100 instance, and we trained ResNet and EfficientNet models alongside with the same default training settings to compare. We exported all models to ONNX FP32 for CPU speed tests and to TensorRT FP16 for GPU speed tests. We ran all speed tests on Google [Colab Pro](https://colab.research.google.com/signup) for easy reproducibility. + +| Model | size
(pixels) | acc
top1 | acc
top5 | Training
90 epochs
4xA100 (hours) | Speed
ONNX CPU
(ms) | Speed
TensorRT V100
(ms) | params
(M) | FLOPs
@224 (B) | +|----------------------------------------------------------------------------------------------------|-----------------------|------------------|------------------|----------------------------------------------|--------------------------------|-------------------------------------|--------------------|------------------------| +| [YOLOv5n-cls](https://github.com/ultralytics/yolov5/releases/download/v6.2/yolov5n-cls.pt) | 224 | 64.6 | 85.4 | 7:59 | **3.3** | **0.5** | **2.5** | **0.5** | +| [YOLOv5s-cls](https://github.com/ultralytics/yolov5/releases/download/v6.2/yolov5s-cls.pt) | 224 | 71.5 | 90.2 | 8:09 | 6.6 | 0.6 | 5.4 | 1.4 | +| [YOLOv5m-cls](https://github.com/ultralytics/yolov5/releases/download/v6.2/yolov5m-cls.pt) | 224 | 75.9 | 92.9 | 10:06 | 15.5 | 0.9 | 12.9 | 3.9 | +| [YOLOv5l-cls](https://github.com/ultralytics/yolov5/releases/download/v6.2/yolov5l-cls.pt) | 224 | 78.0 | 94.0 | 11:56 | 26.9 | 1.4 | 26.5 | 8.5 | +| [YOLOv5x-cls](https://github.com/ultralytics/yolov5/releases/download/v6.2/yolov5x-cls.pt) | 224 | **79.0** | **94.4** | 15:04 | 54.3 | 1.8 | 48.1 | 15.9 | +| | +| [ResNet18](https://github.com/ultralytics/yolov5/releases/download/v6.2/resnet18.pt) | 224 | 70.3 | 89.5 | **6:47** | 11.2 | 0.5 | 11.7 | 3.7 | +| [ResNet34](https://github.com/ultralytics/yolov5/releases/download/v6.2/resnet34.pt) | 224 | 73.9 | 91.8 | 8:33 | 20.6 | 0.9 | 21.8 | 7.4 | +| [ResNet50](https://github.com/ultralytics/yolov5/releases/download/v6.2/resnet50.pt) | 224 | 76.8 | 93.4 | 11:10 | 23.4 | 1.0 | 25.6 | 8.5 | +| [ResNet101](https://github.com/ultralytics/yolov5/releases/download/v6.2/resnet101.pt) | 224 | 78.5 | 94.3 | 17:10 | 42.1 | 1.9 | 44.5 | 15.9 | +| | +| [EfficientNet_b0](https://github.com/ultralytics/yolov5/releases/download/v6.2/efficientnet_b0.pt) | 224 | 75.1 | 92.4 | 13:03 | 12.5 | 1.3 | 5.3 | 1.0 | +| [EfficientNet_b1](https://github.com/ultralytics/yolov5/releases/download/v6.2/efficientnet_b1.pt) | 224 | 76.4 | 93.2 | 17:04 | 14.9 | 1.6 | 7.8 | 1.5 | +| [EfficientNet_b2](https://github.com/ultralytics/yolov5/releases/download/v6.2/efficientnet_b2.pt) | 224 | 76.6 | 93.4 | 17:10 | 15.9 | 1.6 | 9.1 | 1.7 | +| [EfficientNet_b3](https://github.com/ultralytics/yolov5/releases/download/v6.2/efficientnet_b3.pt) | 224 | 77.7 | 94.0 | 19:19 | 18.9 | 1.9 | 12.2 | 2.4 | + +
+ Table Notes (click to expand) + +- All checkpoints are trained to 90 epochs with SGD optimizer with lr0=0.001 at image size 224 and all default settings. Runs logged to https://wandb.ai/glenn-jocher/YOLOv5-Classifier-v6-2. +- **Accuracy** values are for single-model single-scale on [ImageNet-1k](https://www.image-net.org/index.php) dataset.
Reproduce by `python classify/val.py --data ../datasets/imagenet --img 224` +- **Speed** averaged over 100 inference images using a Google [Colab Pro](https://colab.research.google.com/signup) V100 High-RAM instance.
Reproduce by `python classify/val.py --data ../datasets/imagenet --img 224 --batch 1` +- **Export** to ONNX at FP32 and TensorRT at FP16 done with `export.py`.
Reproduce by `python export.py --weights yolov5s-cls.pt --include engine onnx --imgsz 224` +
+
+ +
+ Classification Usage Examples (click to expand) + +### Train +YOLOv5 classification training supports auto-download of MNIST, Fashion-MNIST, CIFAR10, CIFAR100, Imagenette, Imagewoof, and ImageNet datasets with the `--data` argument. To start training on MNIST for example use `--data mnist`. + +```bash +# Single-GPU +python classify/train.py --model yolov5s-cls.pt --data cifar100 --epochs 5 --img 224 --batch 128 + +# Multi-GPU DDP +python -m torch.distributed.run --nproc_per_node 4 --master_port 1 classify/train.py --model yolov5s-cls.pt --data imagenet --epochs 5 --img 224 --device 0,1,2,3 +``` + +### Val +Validate accuracy on a pretrained model. To validate YOLOv5s-cls accuracy on ImageNet. +```bash +bash data/scripts/get_imagenet.sh --val # download ImageNet val split (6.3G, 50000 images) +python classify/val.py --weights yolov5s-cls.pt --data ../datasets/imagenet --img 224 +``` + +### Predict +Run a classification prediction on an image. +```bash +python classify/predict.py --weights yolov5s-cls.pt --data data/images/bus.jpg +``` +```python +model = torch.hub.load('ultralytics/yolov5', 'custom', 'yolov5s-cls.pt') # load from PyTorch Hub +``` + +### Export +Export a group of trained YOLOv5-cls, ResNet and EfficientNet models to ONNX and TensorRT. +```bash +python export.py --weights yolov5s-cls.pt resnet50.pt efficientnet_b0.pt --include onnx engine --img 224 +``` +
+ + +##
Contribute
+ +We love your input! We want to make contributing to YOLOv5 as easy and transparent as possible. Please see our [Contributing Guide](CONTRIBUTING.md) to get started, and fill out the [YOLOv5 Survey](https://ultralytics.com/survey?utm_source=github&utm_medium=social&utm_campaign=Survey) to send us feedback on your experiences. Thank you to all our contributors! + + + + +##
Contact
+ +For YOLOv5 bugs and feature requests please visit [GitHub Issues](https://github.com/ultralytics/yolov5/issues). For business inquiries or +professional support requests please visit [https://ultralytics.com/contact](https://ultralytics.com/contact). + +
+
+ + + + + + + + + + + + + + + + + + + + +
+ +[assets]: https://github.com/ultralytics/yolov5/releases +[tta]: https://github.com/ultralytics/yolov5/issues/303 diff --git a/application/yolov5_example/classify/predict.py b/application/yolov5_example/classify/predict.py new file mode 100644 index 00000000..419830d4 --- /dev/null +++ b/application/yolov5_example/classify/predict.py @@ -0,0 +1,109 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Run classification inference on images + +Usage: + $ python classify/predict.py --weights yolov5s-cls.pt --source im.jpg +""" + +import argparse +import os +import sys +from pathlib import Path + +import cv2 +import torch.nn.functional as F + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from classify.train import imshow_cls +from models.common import DetectMultiBackend +from utils.augmentations import classify_transforms +from utils.general import LOGGER, check_requirements, colorstr, increment_path, print_args +from utils.torch_utils import select_device, smart_inference_mode, time_sync + + +@smart_inference_mode() +def run( + weights=ROOT / 'yolov5s-cls.pt', # model.pt path(s) + source=ROOT / 'data/images/bus.jpg', # file/dir/URL/glob, 0 for webcam + imgsz=224, # inference size + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + half=False, # use FP16 half-precision inference + dnn=False, # use OpenCV DNN for ONNX inference + show=True, + project=ROOT / 'runs/predict-cls', # save to project/name + name='exp', # save to project/name + exist_ok=False, # existing project/name ok, do not increment +): + file = str(source) + seen, dt = 1, [0.0, 0.0, 0.0] + device = select_device(device) + + # Directories + save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run + save_dir.mkdir(parents=True, exist_ok=True) # make dir + + # Transforms + transforms = classify_transforms(imgsz) + + # Load model + model = DetectMultiBackend(weights, device=device, dnn=dnn, fp16=half) + model.warmup(imgsz=(1, 3, imgsz, imgsz)) # warmup + + # Image + t1 = time_sync() + im = cv2.cvtColor(cv2.imread(file), cv2.COLOR_BGR2RGB) + im = transforms(im).unsqueeze(0).to(device) + im = im.half() if model.fp16 else im.float() + t2 = time_sync() + dt[0] += t2 - t1 + + # Inference + results = model(im) + t3 = time_sync() + dt[1] += t3 - t2 + + p = F.softmax(results, dim=1) # probabilities + i = p.argsort(1, descending=True)[:, :5].squeeze() # top 5 indices + dt[2] += time_sync() - t3 + LOGGER.info(f"image 1/1 {file}: {imgsz}x{imgsz} {', '.join(f'{model.names[j]} {p[0, j]:.2f}' for j in i)}") + + # Print results + t = tuple(x / seen * 1E3 for x in dt) # speeds per image + shape = (1, 3, imgsz, imgsz) + LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms post-process per image at shape {shape}' % t) + if show: + imshow_cls(im, f=save_dir / Path(file).name, verbose=True) + LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}") + return p + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s-cls.pt', help='model path(s)') + parser.add_argument('--source', type=str, default=ROOT / 'data/images/bus.jpg', help='file') + parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=224, help='train, val image size (pixels)') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference') + parser.add_argument('--project', default=ROOT / 'runs/predict-cls', help='save to project/name') + parser.add_argument('--name', default='exp', help='save to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + opt = parser.parse_args() + print_args(vars(opt)) + return opt + + +def main(opt): + check_requirements(exclude=('tensorboard', 'thop')) + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/application/yolov5_example/classify/train.py b/application/yolov5_example/classify/train.py new file mode 100644 index 00000000..f2b46556 --- /dev/null +++ b/application/yolov5_example/classify/train.py @@ -0,0 +1,325 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Train a YOLOv5 classifier model on a classification dataset +Datasets: --data mnist, fashion-mnist, cifar10, cifar100, imagenette, imagewoof, imagenet, or 'path/to/custom/dataset' + +Usage: + $ python classify/train.py --model yolov5s-cls.pt --data cifar100 --epochs 5 --img 128 + $ python -m torch.distributed.run --nproc_per_node 4 --master_port 1 classify/train.py --model yolov5s-cls.pt --data imagenet --epochs 5 --img 224 --device 0,1,2,3 +""" + +import argparse +import os +import subprocess +import sys +import time +from copy import deepcopy +from datetime import datetime +from pathlib import Path + +import torch +import torch.distributed as dist +import torch.hub as hub +import torch.optim.lr_scheduler as lr_scheduler +import torchvision +from torch.cuda import amp +from tqdm import tqdm + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from classify import val as validate +from models.experimental import attempt_load +from models.yolo import ClassificationModel, DetectionModel +from utils.dataloaders import create_classification_dataloader +from utils.general import (DATASETS_DIR, LOGGER, WorkingDirectory, check_git_status, check_requirements, colorstr, + download, increment_path, init_seeds, print_args, yaml_save) +from utils.loggers import GenericLogger +from utils.plots import imshow_cls +from utils.torch_utils import (ModelEMA, model_info, reshape_classifier_output, select_device, smart_DDP, + smart_optimizer, smartCrossEntropyLoss, torch_distributed_zero_first) + +LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1)) # https://pytorch.org/docs/stable/elastic/run.html +RANK = int(os.getenv('RANK', -1)) +WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1)) + + +def train(opt, device): + init_seeds(opt.seed + 1 + RANK, deterministic=True) + save_dir, data, bs, epochs, nw, imgsz, pretrained = \ + opt.save_dir, Path(opt.data), opt.batch_size, opt.epochs, min(os.cpu_count() - 1, opt.workers), \ + opt.imgsz, str(opt.pretrained).lower() == 'true' + cuda = device.type != 'cpu' + + # Directories + wdir = save_dir / 'weights' + wdir.mkdir(parents=True, exist_ok=True) # make dir + last, best = wdir / 'last.pt', wdir / 'best.pt' + + # Save run settings + yaml_save(save_dir / 'opt.yaml', vars(opt)) + + # Logger + logger = GenericLogger(opt=opt, console_logger=LOGGER) if RANK in {-1, 0} else None + + # Download Dataset + with torch_distributed_zero_first(LOCAL_RANK), WorkingDirectory(ROOT): + data_dir = data if data.is_dir() else (DATASETS_DIR / data) + if not data_dir.is_dir(): + LOGGER.info(f'\nDataset not found ⚠️, missing path {data_dir}, attempting download...') + t = time.time() + if str(data) == 'imagenet': + subprocess.run(f"bash {ROOT / 'data/scripts/get_imagenet.sh'}", shell=True, check=True) + else: + url = f'https://github.com/ultralytics/yolov5/releases/download/v1.0/{data}.zip' + download(url, dir=data_dir.parent) + s = f"Dataset download success ✅ ({time.time() - t:.1f}s), saved to {colorstr('bold', data_dir)}\n" + LOGGER.info(s) + + # Dataloaders + nc = len([x for x in (data_dir / 'train').glob('*') if x.is_dir()]) # number of classes + trainloader = create_classification_dataloader(path=data_dir / 'train', + imgsz=imgsz, + batch_size=bs // WORLD_SIZE, + augment=True, + cache=opt.cache, + rank=LOCAL_RANK, + workers=nw) + + test_dir = data_dir / 'test' if (data_dir / 'test').exists() else data_dir / 'val' # data/test or data/val + if RANK in {-1, 0}: + testloader = create_classification_dataloader(path=test_dir, + imgsz=imgsz, + batch_size=bs // WORLD_SIZE * 2, + augment=False, + cache=opt.cache, + rank=-1, + workers=nw) + + # Model + with torch_distributed_zero_first(LOCAL_RANK), WorkingDirectory(ROOT): + if Path(opt.model).is_file() or opt.model.endswith('.pt'): + model = attempt_load(opt.model, device='cpu', fuse=False) + elif opt.model in torchvision.models.__dict__: # TorchVision models i.e. resnet50, efficientnet_b0 + model = torchvision.models.__dict__[opt.model](weights='IMAGENET1K_V1' if pretrained else None) + else: + m = hub.list('ultralytics/yolov5') # + hub.list('pytorch/vision') # models + raise ModuleNotFoundError(f'--model {opt.model} not found. Available models are: \n' + '\n'.join(m)) + if isinstance(model, DetectionModel): + LOGGER.warning("WARNING: pass YOLOv5 classifier model with '-cls' suffix, i.e. '--model yolov5s-cls.pt'") + model = ClassificationModel(model=model, nc=nc, cutoff=opt.cutoff or 10) # convert to classification model + reshape_classifier_output(model, nc) # update class count + for p in model.parameters(): + p.requires_grad = True # for training + for m in model.modules(): + if not pretrained and hasattr(m, 'reset_parameters'): + m.reset_parameters() + if isinstance(m, torch.nn.Dropout) and opt.dropout is not None: + m.p = opt.dropout # set dropout + model = model.to(device) + names = trainloader.dataset.classes # class names + model.names = names # attach class names + + # Info + if RANK in {-1, 0}: + model_info(model) + if opt.verbose: + LOGGER.info(model) + images, labels = next(iter(trainloader)) + file = imshow_cls(images[:25], labels[:25], names=names, f=save_dir / 'train_images.jpg') + logger.log_images(file, name='Train Examples') + logger.log_graph(model, imgsz) # log model + + # Optimizer + optimizer = smart_optimizer(model, opt.optimizer, opt.lr0, momentum=0.9, decay=5e-5) + + # Scheduler + lrf = 0.01 # final lr (fraction of lr0) + # lf = lambda x: ((1 + math.cos(x * math.pi / epochs)) / 2) * (1 - lrf) + lrf # cosine + lf = lambda x: (1 - x / epochs) * (1 - lrf) + lrf # linear + scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) + # scheduler = lr_scheduler.OneCycleLR(optimizer, max_lr=lr0, total_steps=epochs, pct_start=0.1, + # final_div_factor=1 / 25 / lrf) + + # EMA + ema = ModelEMA(model) if RANK in {-1, 0} else None + + # DDP mode + if cuda and RANK != -1: + model = smart_DDP(model) + + # Train + t0 = time.time() + criterion = smartCrossEntropyLoss(label_smoothing=opt.label_smoothing) # loss function + best_fitness = 0.0 + scaler = amp.GradScaler(enabled=cuda) + val = test_dir.stem # 'val' or 'test' + LOGGER.info(f'Image sizes {imgsz} train, {imgsz} test\n' + f'Using {nw * WORLD_SIZE} dataloader workers\n' + f"Logging results to {colorstr('bold', save_dir)}\n" + f'Starting {opt.model} training on {data} dataset with {nc} classes for {epochs} epochs...\n\n' + f"{'Epoch':>10}{'GPU_mem':>10}{'train_loss':>12}{f'{val}_loss':>12}{'top1_acc':>12}{'top5_acc':>12}") + for epoch in range(epochs): # loop over the dataset multiple times + tloss, vloss, fitness = 0.0, 0.0, 0.0 # train loss, val loss, fitness + model.train() + if RANK != -1: + trainloader.sampler.set_epoch(epoch) + pbar = enumerate(trainloader) + if RANK in {-1, 0}: + pbar = tqdm(enumerate(trainloader), total=len(trainloader), bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') + for i, (images, labels) in pbar: # progress bar + images, labels = images.to(device, non_blocking=True), labels.to(device) + + # Forward + with amp.autocast(enabled=cuda): # stability issues when enabled + loss = criterion(model(images), labels) + + # Backward + scaler.scale(loss).backward() + + # Optimize + scaler.unscale_(optimizer) # unscale gradients + torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0) # clip gradients + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + if ema: + ema.update(model) + + if RANK in {-1, 0}: + # Print + tloss = (tloss * i + loss.item()) / (i + 1) # update mean losses + mem = '%.3gG' % (torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0) # (GB) + pbar.desc = f"{f'{epoch + 1}/{epochs}':>10}{mem:>10}{tloss:>12.3g}" + ' ' * 36 + + # Test + if i == len(pbar) - 1: # last batch + top1, top5, vloss = validate.run(model=ema.ema, + dataloader=testloader, + criterion=criterion, + pbar=pbar) # test accuracy, loss + fitness = top1 # define fitness as top1 accuracy + + # Scheduler + scheduler.step() + + # Log metrics + if RANK in {-1, 0}: + # Best fitness + if fitness > best_fitness: + best_fitness = fitness + + # Log + metrics = { + "train/loss": tloss, + f"{val}/loss": vloss, + "metrics/accuracy_top1": top1, + "metrics/accuracy_top5": top5, + "lr/0": optimizer.param_groups[0]['lr']} # learning rate + logger.log_metrics(metrics, epoch) + + # Save model + final_epoch = epoch + 1 == epochs + if (not opt.nosave) or final_epoch: + ckpt = { + 'epoch': epoch, + 'best_fitness': best_fitness, + 'model': deepcopy(ema.ema).half(), # deepcopy(de_parallel(model)).half(), + 'ema': None, # deepcopy(ema.ema).half(), + 'updates': ema.updates, + 'optimizer': None, # optimizer.state_dict(), + 'opt': vars(opt), + 'date': datetime.now().isoformat()} + + # Save last, best and delete + torch.save(ckpt, last) + if best_fitness == fitness: + torch.save(ckpt, best) + del ckpt + + # Train complete + if RANK in {-1, 0} and final_epoch: + LOGGER.info(f'\nTraining complete ({(time.time() - t0) / 3600:.3f} hours)' + f"\nResults saved to {colorstr('bold', save_dir)}" + f"\nPredict: python classify/predict.py --weights {best} --source im.jpg" + f"\nValidate: python classify/val.py --weights {best} --data {data_dir}" + f"\nExport: python export.py --weights {best} --include onnx" + f"\nPyTorch Hub: model = torch.hub.load('ultralytics/yolov5', 'custom', '{best}')" + f"\nVisualize: https://netron.app\n") + + # Plot examples + images, labels = (x[:25] for x in next(iter(testloader))) # first 25 images and labels + pred = torch.max(ema.ema((images.half() if cuda else images.float()).to(device)), 1)[1] + file = imshow_cls(images, labels, pred, names, verbose=False, f=save_dir / 'test_images.jpg') + + # Log results + meta = {"epochs": epochs, "top1_acc": best_fitness, "date": datetime.now().isoformat()} + logger.log_images(file, name='Test Examples (true-predicted)', epoch=epoch) + logger.log_model(best, epochs, metadata=meta) + + +def parse_opt(known=False): + parser = argparse.ArgumentParser() + parser.add_argument('--model', type=str, default='yolov5s-cls.pt', help='initial weights path') + parser.add_argument('--data', type=str, default='mnist', help='cifar10, cifar100, mnist, imagenet, etc.') + parser.add_argument('--epochs', type=int, default=10) + parser.add_argument('--batch-size', type=int, default=64, help='total batch size for all GPUs') + parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=128, help='train, val image size (pixels)') + parser.add_argument('--nosave', action='store_true', help='only save final checkpoint') + parser.add_argument('--cache', type=str, nargs='?', const='ram', help='--cache images in "ram" (default) or "disk"') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)') + parser.add_argument('--project', default=ROOT / 'runs/train-cls', help='save to project/name') + parser.add_argument('--name', default='exp', help='save to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--pretrained', nargs='?', const=True, default=True, help='start from i.e. --pretrained False') + parser.add_argument('--optimizer', choices=['SGD', 'Adam', 'AdamW', 'RMSProp'], default='Adam', help='optimizer') + parser.add_argument('--lr0', type=float, default=0.001, help='initial learning rate') + parser.add_argument('--label-smoothing', type=float, default=0.1, help='Label smoothing epsilon') + parser.add_argument('--cutoff', type=int, default=None, help='Model layer cutoff index for Classify() head') + parser.add_argument('--dropout', type=float, default=None, help='Dropout (fraction)') + parser.add_argument('--verbose', action='store_true', help='Verbose mode') + parser.add_argument('--seed', type=int, default=0, help='Global training seed') + parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify') + return parser.parse_known_args()[0] if known else parser.parse_args() + + +def main(opt): + # Checks + if RANK in {-1, 0}: + print_args(vars(opt)) + check_git_status() + check_requirements() + + # DDP mode + device = select_device(opt.device, batch_size=opt.batch_size) + if LOCAL_RANK != -1: + assert opt.batch_size != -1, 'AutoBatch is coming soon for classification, please pass a valid --batch-size' + assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE' + assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command' + torch.cuda.set_device(LOCAL_RANK) + device = torch.device('cuda', LOCAL_RANK) + dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo") + + # Parameters + opt.save_dir = increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok) # increment run + + # Train + train(opt, device) + + +def run(**kwargs): + # Usage: from yolov5 import classify; classify.train.run(data=mnist, imgsz=320, model='yolov5m') + opt = parse_opt(True) + for k, v in kwargs.items(): + setattr(opt, k, v) + main(opt) + return opt + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/application/yolov5_example/classify/val.py b/application/yolov5_example/classify/val.py new file mode 100644 index 00000000..0930ba8c --- /dev/null +++ b/application/yolov5_example/classify/val.py @@ -0,0 +1,158 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Validate a classification model on a dataset + +Usage: + $ python classify/val.py --weights yolov5s-cls.pt --data ../datasets/imagenet +""" + +import argparse +import os +import sys +from pathlib import Path + +import torch +from tqdm import tqdm + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from models.common import DetectMultiBackend +from utils.dataloaders import create_classification_dataloader +from utils.general import LOGGER, check_img_size, check_requirements, colorstr, increment_path, print_args +from utils.torch_utils import select_device, smart_inference_mode, time_sync + + +@smart_inference_mode() +def run( + data=ROOT / '../datasets/mnist', # dataset dir + weights=ROOT / 'yolov5s-cls.pt', # model.pt path(s) + batch_size=128, # batch size + imgsz=224, # inference size (pixels) + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + workers=8, # max dataloader workers (per RANK in DDP mode) + verbose=False, # verbose output + project=ROOT / 'runs/val-cls', # save to project/name + name='exp', # save to project/name + exist_ok=False, # existing project/name ok, do not increment + half=True, # use FP16 half-precision inference + dnn=False, # use OpenCV DNN for ONNX inference + model=None, + dataloader=None, + criterion=None, + pbar=None, +): + # Initialize/load model and set device + training = model is not None + if training: # called by train.py + device, pt, jit, engine = next(model.parameters()).device, True, False, False # get model device, PyTorch model + half &= device.type != 'cpu' # half precision only supported on CUDA + model.half() if half else model.float() + else: # called directly + device = select_device(device, batch_size=batch_size) + + # Directories + save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run + save_dir.mkdir(parents=True, exist_ok=True) # make dir + + # Load model + model = DetectMultiBackend(weights, device=device, dnn=dnn, fp16=half) + stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine + imgsz = check_img_size(imgsz, s=stride) # check image size + half = model.fp16 # FP16 supported on limited backends with CUDA + if engine: + batch_size = model.batch_size + else: + device = model.device + if not (pt or jit): + batch_size = 1 # export.py models default to batch-size 1 + LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models') + + # Dataloader + data = Path(data) + test_dir = data / 'test' if (data / 'test').exists() else data / 'val' # data/test or data/val + dataloader = create_classification_dataloader(path=test_dir, + imgsz=imgsz, + batch_size=batch_size, + augment=False, + rank=-1, + workers=workers) + + model.eval() + pred, targets, loss, dt = [], [], 0, [0.0, 0.0, 0.0] + n = len(dataloader) # number of batches + action = 'validating' if dataloader.dataset.root.stem == 'val' else 'testing' + desc = f"{pbar.desc[:-36]}{action:>36}" if pbar else f"{action}" + bar = tqdm(dataloader, desc, n, not training, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}', position=0) + with torch.cuda.amp.autocast(enabled=device.type != 'cpu'): + for images, labels in bar: + t1 = time_sync() + images, labels = images.to(device, non_blocking=True), labels.to(device) + t2 = time_sync() + dt[0] += t2 - t1 + + y = model(images) + t3 = time_sync() + dt[1] += t3 - t2 + + pred.append(y.argsort(1, descending=True)[:, :5]) + targets.append(labels) + if criterion: + loss += criterion(y, labels) + dt[2] += time_sync() - t3 + + loss /= n + pred, targets = torch.cat(pred), torch.cat(targets) + correct = (targets[:, None] == pred).float() + acc = torch.stack((correct[:, 0], correct.max(1).values), dim=1) # (top1, top5) accuracy + top1, top5 = acc.mean(0).tolist() + + if pbar: + pbar.desc = f"{pbar.desc[:-36]}{loss:>12.3g}{top1:>12.3g}{top5:>12.3g}" + if verbose: # all classes + LOGGER.info(f"{'Class':>24}{'Images':>12}{'top1_acc':>12}{'top5_acc':>12}") + LOGGER.info(f"{'all':>24}{targets.shape[0]:>12}{top1:>12.3g}{top5:>12.3g}") + for i, c in enumerate(model.names): + aci = acc[targets == i] + top1i, top5i = aci.mean(0).tolist() + LOGGER.info(f"{c:>24}{aci.shape[0]:>12}{top1i:>12.3g}{top5i:>12.3g}") + + # Print results + t = tuple(x / len(dataloader.dataset.samples) * 1E3 for x in dt) # speeds per image + shape = (1, 3, imgsz, imgsz) + LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms post-process per image at shape {shape}' % t) + LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}") + + return top1, top5, loss + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--data', type=str, default=ROOT / '../datasets/mnist', help='dataset path') + parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s-cls.pt', help='model.pt path(s)') + parser.add_argument('--batch-size', type=int, default=128, help='batch size') + parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=224, help='inference size (pixels)') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)') + parser.add_argument('--verbose', nargs='?', const=True, default=True, help='verbose output') + parser.add_argument('--project', default=ROOT / 'runs/val-cls', help='save to project/name') + parser.add_argument('--name', default='exp', help='save to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference') + opt = parser.parse_args() + print_args(vars(opt)) + return opt + + +def main(opt): + check_requirements(exclude=('tensorboard', 'thop')) + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/application/yolov5_example/data/Argoverse.yaml b/application/yolov5_example/data/Argoverse.yaml new file mode 100644 index 00000000..9d21296e --- /dev/null +++ b/application/yolov5_example/data/Argoverse.yaml @@ -0,0 +1,67 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Argoverse-HD dataset (ring-front-center camera) http://www.cs.cmu.edu/~mengtial/proj/streaming/ by Argo AI +# Example usage: python train.py --data Argoverse.yaml +# parent +# ├── yolov5 +# └── datasets +# └── Argoverse ← downloads here (31.3 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/Argoverse # dataset root dir +train: Argoverse-1.1/images/train/ # train images (relative to 'path') 39384 images +val: Argoverse-1.1/images/val/ # val images (relative to 'path') 15062 images +test: Argoverse-1.1/images/test/ # test images (optional) https://eval.ai/web/challenges/challenge-page/800/overview + +# Classes +nc: 8 # number of classes +names: ['person', 'bicycle', 'car', 'motorcycle', 'bus', 'truck', 'traffic_light', 'stop_sign'] # class names + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + import json + + from tqdm import tqdm + from utils.general import download, Path + + + def argoverse2yolo(set): + labels = {} + a = json.load(open(set, "rb")) + for annot in tqdm(a['annotations'], desc=f"Converting {set} to YOLOv5 format..."): + img_id = annot['image_id'] + img_name = a['images'][img_id]['name'] + img_label_name = f'{img_name[:-3]}txt' + + cls = annot['category_id'] # instance class id + x_center, y_center, width, height = annot['bbox'] + x_center = (x_center + width / 2) / 1920.0 # offset and scale + y_center = (y_center + height / 2) / 1200.0 # offset and scale + width /= 1920.0 # scale + height /= 1200.0 # scale + + img_dir = set.parents[2] / 'Argoverse-1.1' / 'labels' / a['seq_dirs'][a['images'][annot['image_id']]['sid']] + if not img_dir.exists(): + img_dir.mkdir(parents=True, exist_ok=True) + + k = str(img_dir / img_label_name) + if k not in labels: + labels[k] = [] + labels[k].append(f"{cls} {x_center} {y_center} {width} {height}\n") + + for k in labels: + with open(k, "w") as f: + f.writelines(labels[k]) + + + # Download + dir = Path('../datasets/Argoverse') # dataset root dir + urls = ['https://argoverse-hd.s3.us-east-2.amazonaws.com/Argoverse-HD-Full.zip'] + download(urls, dir=dir, delete=False) + + # Convert + annotations_dir = 'Argoverse-HD/annotations/' + (dir / 'Argoverse-1.1' / 'tracking').rename(dir / 'Argoverse-1.1' / 'images') # rename 'tracking' to 'images' + for d in "train.json", "val.json": + argoverse2yolo(dir / annotations_dir / d) # convert VisDrone annotations to YOLO labels diff --git a/application/yolov5_example/data/GlobalWheat2020.yaml b/application/yolov5_example/data/GlobalWheat2020.yaml new file mode 100644 index 00000000..4c43693f --- /dev/null +++ b/application/yolov5_example/data/GlobalWheat2020.yaml @@ -0,0 +1,54 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Global Wheat 2020 dataset http://www.global-wheat.com/ by University of Saskatchewan +# Example usage: python train.py --data GlobalWheat2020.yaml +# parent +# ├── yolov5 +# └── datasets +# └── GlobalWheat2020 ← downloads here (7.0 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/GlobalWheat2020 # dataset root dir +train: # train images (relative to 'path') 3422 images + - images/arvalis_1 + - images/arvalis_2 + - images/arvalis_3 + - images/ethz_1 + - images/rres_1 + - images/inrae_1 + - images/usask_1 +val: # val images (relative to 'path') 748 images (WARNING: train set contains ethz_1) + - images/ethz_1 +test: # test images (optional) 1276 images + - images/utokyo_1 + - images/utokyo_2 + - images/nau_1 + - images/uq_1 + +# Classes +nc: 1 # number of classes +names: ['wheat_head'] # class names + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + from utils.general import download, Path + + + # Download + dir = Path(yaml['path']) # dataset root dir + urls = ['https://zenodo.org/record/4298502/files/global-wheat-codalab-official.zip', + 'https://github.com/ultralytics/yolov5/releases/download/v1.0/GlobalWheat2020_labels.zip'] + download(urls, dir=dir) + + # Make Directories + for p in 'annotations', 'images', 'labels': + (dir / p).mkdir(parents=True, exist_ok=True) + + # Move + for p in 'arvalis_1', 'arvalis_2', 'arvalis_3', 'ethz_1', 'rres_1', 'inrae_1', 'usask_1', \ + 'utokyo_1', 'utokyo_2', 'nau_1', 'uq_1': + (dir / p).rename(dir / 'images' / p) # move to /images + f = (dir / p).with_suffix('.json') # json file + if f.exists(): + f.rename((dir / 'annotations' / p).with_suffix('.json')) # move to /annotations diff --git a/application/yolov5_example/data/ImageNet.yaml b/application/yolov5_example/data/ImageNet.yaml new file mode 100644 index 00000000..9f89b426 --- /dev/null +++ b/application/yolov5_example/data/ImageNet.yaml @@ -0,0 +1,156 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# ImageNet-1k dataset https://www.image-net.org/index.php by Stanford University +# Simplified class names from https://github.com/anishathalye/imagenet-simple-labels +# Example usage: python classify/train.py --data imagenet +# parent +# ├── yolov5 +# └── datasets +# └── imagenet ← downloads here (144 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/imagenet # dataset root dir +train: train # train images (relative to 'path') 1281167 images +val: val # val images (relative to 'path') 50000 images +test: # test images (optional) + +# Classes +nc: 1000 # number of classes +names: ['tench', 'goldfish', 'great white shark', 'tiger shark', 'hammerhead shark', 'electric ray', 'stingray', 'cock', + 'hen', 'ostrich', 'brambling', 'goldfinch', 'house finch', 'junco', 'indigo bunting', 'American robin', + 'bulbul', 'jay', 'magpie', 'chickadee', 'American dipper', 'kite', 'bald eagle', 'vulture', 'great grey owl', + 'fire salamander', 'smooth newt', 'newt', 'spotted salamander', 'axolotl', 'American bullfrog', 'tree frog', + 'tailed frog', 'loggerhead sea turtle', 'leatherback sea turtle', 'mud turtle', 'terrapin', 'box turtle', + 'banded gecko', 'green iguana', 'Carolina anole', 'desert grassland whiptail lizard', 'agama', + 'frilled-necked lizard', 'alligator lizard', 'Gila monster', 'European green lizard', 'chameleon', + 'Komodo dragon', 'Nile crocodile', 'American alligator', 'triceratops', 'worm snake', 'ring-necked snake', + 'eastern hog-nosed snake', 'smooth green snake', 'kingsnake', 'garter snake', 'water snake', 'vine snake', + 'night snake', 'boa constrictor', 'African rock python', 'Indian cobra', 'green mamba', 'sea snake', + 'Saharan horned viper', 'eastern diamondback rattlesnake', 'sidewinder', 'trilobite', 'harvestman', 'scorpion', + 'yellow garden spider', 'barn spider', 'European garden spider', 'southern black widow', 'tarantula', + 'wolf spider', 'tick', 'centipede', 'black grouse', 'ptarmigan', 'ruffed grouse', 'prairie grouse', 'peacock', + 'quail', 'partridge', 'grey parrot', 'macaw', 'sulphur-crested cockatoo', 'lorikeet', 'coucal', 'bee eater', + 'hornbill', 'hummingbird', 'jacamar', 'toucan', 'duck', 'red-breasted merganser', 'goose', 'black swan', + 'tusker', 'echidna', 'platypus', 'wallaby', 'koala', 'wombat', 'jellyfish', 'sea anemone', 'brain coral', + 'flatworm', 'nematode', 'conch', 'snail', 'slug', 'sea slug', 'chiton', 'chambered nautilus', 'Dungeness crab', + 'rock crab', 'fiddler crab', 'red king crab', 'American lobster', 'spiny lobster', 'crayfish', 'hermit crab', + 'isopod', 'white stork', 'black stork', 'spoonbill', 'flamingo', 'little blue heron', 'great egret', 'bittern', + 'crane (bird)', 'limpkin', 'common gallinule', 'American coot', 'bustard', 'ruddy turnstone', 'dunlin', + 'common redshank', 'dowitcher', 'oystercatcher', 'pelican', 'king penguin', 'albatross', 'grey whale', + 'killer whale', 'dugong', 'sea lion', 'Chihuahua', 'Japanese Chin', 'Maltese', 'Pekingese', 'Shih Tzu', + 'King Charles Spaniel', 'Papillon', 'toy terrier', 'Rhodesian Ridgeback', 'Afghan Hound', 'Basset Hound', + 'Beagle', 'Bloodhound', 'Bluetick Coonhound', 'Black and Tan Coonhound', 'Treeing Walker Coonhound', + 'English foxhound', 'Redbone Coonhound', 'borzoi', 'Irish Wolfhound', 'Italian Greyhound', 'Whippet', + 'Ibizan Hound', 'Norwegian Elkhound', 'Otterhound', 'Saluki', 'Scottish Deerhound', 'Weimaraner', + 'Staffordshire Bull Terrier', 'American Staffordshire Terrier', 'Bedlington Terrier', 'Border Terrier', + 'Kerry Blue Terrier', 'Irish Terrier', 'Norfolk Terrier', 'Norwich Terrier', 'Yorkshire Terrier', + 'Wire Fox Terrier', 'Lakeland Terrier', 'Sealyham Terrier', 'Airedale Terrier', 'Cairn Terrier', + 'Australian Terrier', 'Dandie Dinmont Terrier', 'Boston Terrier', 'Miniature Schnauzer', 'Giant Schnauzer', + 'Standard Schnauzer', 'Scottish Terrier', 'Tibetan Terrier', 'Australian Silky Terrier', + 'Soft-coated Wheaten Terrier', 'West Highland White Terrier', 'Lhasa Apso', 'Flat-Coated Retriever', + 'Curly-coated Retriever', 'Golden Retriever', 'Labrador Retriever', 'Chesapeake Bay Retriever', + 'German Shorthaired Pointer', 'Vizsla', 'English Setter', 'Irish Setter', 'Gordon Setter', 'Brittany', + 'Clumber Spaniel', 'English Springer Spaniel', 'Welsh Springer Spaniel', 'Cocker Spaniels', 'Sussex Spaniel', + 'Irish Water Spaniel', 'Kuvasz', 'Schipperke', 'Groenendael', 'Malinois', 'Briard', 'Australian Kelpie', + 'Komondor', 'Old English Sheepdog', 'Shetland Sheepdog', 'collie', 'Border Collie', 'Bouvier des Flandres', + 'Rottweiler', 'German Shepherd Dog', 'Dobermann', 'Miniature Pinscher', 'Greater Swiss Mountain Dog', + 'Bernese Mountain Dog', 'Appenzeller Sennenhund', 'Entlebucher Sennenhund', 'Boxer', 'Bullmastiff', + 'Tibetan Mastiff', 'French Bulldog', 'Great Dane', 'St. Bernard', 'husky', 'Alaskan Malamute', 'Siberian Husky', + 'Dalmatian', 'Affenpinscher', 'Basenji', 'pug', 'Leonberger', 'Newfoundland', 'Pyrenean Mountain Dog', + 'Samoyed', 'Pomeranian', 'Chow Chow', 'Keeshond', 'Griffon Bruxellois', 'Pembroke Welsh Corgi', + 'Cardigan Welsh Corgi', 'Toy Poodle', 'Miniature Poodle', 'Standard Poodle', 'Mexican hairless dog', + 'grey wolf', 'Alaskan tundra wolf', 'red wolf', 'coyote', 'dingo', 'dhole', 'African wild dog', 'hyena', + 'red fox', 'kit fox', 'Arctic fox', 'grey fox', 'tabby cat', 'tiger cat', 'Persian cat', 'Siamese cat', + 'Egyptian Mau', 'cougar', 'lynx', 'leopard', 'snow leopard', 'jaguar', 'lion', 'tiger', 'cheetah', 'brown bear', + 'American black bear', 'polar bear', 'sloth bear', 'mongoose', 'meerkat', 'tiger beetle', 'ladybug', + 'ground beetle', 'longhorn beetle', 'leaf beetle', 'dung beetle', 'rhinoceros beetle', 'weevil', 'fly', 'bee', + 'ant', 'grasshopper', 'cricket', 'stick insect', 'cockroach', 'mantis', 'cicada', 'leafhopper', 'lacewing', + 'dragonfly', 'damselfly', 'red admiral', 'ringlet', 'monarch butterfly', 'small white', 'sulphur butterfly', + 'gossamer-winged butterfly', 'starfish', 'sea urchin', 'sea cucumber', 'cottontail rabbit', 'hare', + 'Angora rabbit', 'hamster', 'porcupine', 'fox squirrel', 'marmot', 'beaver', 'guinea pig', 'common sorrel', + 'zebra', 'pig', 'wild boar', 'warthog', 'hippopotamus', 'ox', 'water buffalo', 'bison', 'ram', 'bighorn sheep', + 'Alpine ibex', 'hartebeest', 'impala', 'gazelle', 'dromedary', 'llama', 'weasel', 'mink', 'European polecat', + 'black-footed ferret', 'otter', 'skunk', 'badger', 'armadillo', 'three-toed sloth', 'orangutan', 'gorilla', + 'chimpanzee', 'gibbon', 'siamang', 'guenon', 'patas monkey', 'baboon', 'macaque', 'langur', + 'black-and-white colobus', 'proboscis monkey', 'marmoset', 'white-headed capuchin', 'howler monkey', 'titi', + "Geoffroy's spider monkey", 'common squirrel monkey', 'ring-tailed lemur', 'indri', 'Asian elephant', + 'African bush elephant', 'red panda', 'giant panda', 'snoek', 'eel', 'coho salmon', 'rock beauty', 'clownfish', + 'sturgeon', 'garfish', 'lionfish', 'pufferfish', 'abacus', 'abaya', 'academic gown', 'accordion', + 'acoustic guitar', 'aircraft carrier', 'airliner', 'airship', 'altar', 'ambulance', 'amphibious vehicle', + 'analog clock', 'apiary', 'apron', 'waste container', 'assault rifle', 'backpack', 'bakery', 'balance beam', + 'balloon', 'ballpoint pen', 'Band-Aid', 'banjo', 'baluster', 'barbell', 'barber chair', 'barbershop', 'barn', + 'barometer', 'barrel', 'wheelbarrow', 'baseball', 'basketball', 'bassinet', 'bassoon', 'swimming cap', + 'bath towel', 'bathtub', 'station wagon', 'lighthouse', 'beaker', 'military cap', 'beer bottle', 'beer glass', + 'bell-cot', 'bib', 'tandem bicycle', 'bikini', 'ring binder', 'binoculars', 'birdhouse', 'boathouse', + 'bobsleigh', 'bolo tie', 'poke bonnet', 'bookcase', 'bookstore', 'bottle cap', 'bow', 'bow tie', 'brass', 'bra', + 'breakwater', 'breastplate', 'broom', 'bucket', 'buckle', 'bulletproof vest', 'high-speed train', + 'butcher shop', 'taxicab', 'cauldron', 'candle', 'cannon', 'canoe', 'can opener', 'cardigan', 'car mirror', + 'carousel', 'tool kit', 'carton', 'car wheel', 'automated teller machine', 'cassette', 'cassette player', + 'castle', 'catamaran', 'CD player', 'cello', 'mobile phone', 'chain', 'chain-link fence', 'chain mail', + 'chainsaw', 'chest', 'chiffonier', 'chime', 'china cabinet', 'Christmas stocking', 'church', 'movie theater', + 'cleaver', 'cliff dwelling', 'cloak', 'clogs', 'cocktail shaker', 'coffee mug', 'coffeemaker', 'coil', + 'combination lock', 'computer keyboard', 'confectionery store', 'container ship', 'convertible', 'corkscrew', + 'cornet', 'cowboy boot', 'cowboy hat', 'cradle', 'crane (machine)', 'crash helmet', 'crate', 'infant bed', + 'Crock Pot', 'croquet ball', 'crutch', 'cuirass', 'dam', 'desk', 'desktop computer', 'rotary dial telephone', + 'diaper', 'digital clock', 'digital watch', 'dining table', 'dishcloth', 'dishwasher', 'disc brake', 'dock', + 'dog sled', 'dome', 'doormat', 'drilling rig', 'drum', 'drumstick', 'dumbbell', 'Dutch oven', 'electric fan', + 'electric guitar', 'electric locomotive', 'entertainment center', 'envelope', 'espresso machine', 'face powder', + 'feather boa', 'filing cabinet', 'fireboat', 'fire engine', 'fire screen sheet', 'flagpole', 'flute', + 'folding chair', 'football helmet', 'forklift', 'fountain', 'fountain pen', 'four-poster bed', 'freight car', + 'French horn', 'frying pan', 'fur coat', 'garbage truck', 'gas mask', 'gas pump', 'goblet', 'go-kart', + 'golf ball', 'golf cart', 'gondola', 'gong', 'gown', 'grand piano', 'greenhouse', 'grille', 'grocery store', + 'guillotine', 'barrette', 'hair spray', 'half-track', 'hammer', 'hamper', 'hair dryer', 'hand-held computer', + 'handkerchief', 'hard disk drive', 'harmonica', 'harp', 'harvester', 'hatchet', 'holster', 'home theater', + 'honeycomb', 'hook', 'hoop skirt', 'horizontal bar', 'horse-drawn vehicle', 'hourglass', 'iPod', 'clothes iron', + "jack-o'-lantern", 'jeans', 'jeep', 'T-shirt', 'jigsaw puzzle', 'pulled rickshaw', 'joystick', 'kimono', + 'knee pad', 'knot', 'lab coat', 'ladle', 'lampshade', 'laptop computer', 'lawn mower', 'lens cap', + 'paper knife', 'library', 'lifeboat', 'lighter', 'limousine', 'ocean liner', 'lipstick', 'slip-on shoe', + 'lotion', 'speaker', 'loupe', 'sawmill', 'magnetic compass', 'mail bag', 'mailbox', 'tights', 'tank suit', + 'manhole cover', 'maraca', 'marimba', 'mask', 'match', 'maypole', 'maze', 'measuring cup', 'medicine chest', + 'megalith', 'microphone', 'microwave oven', 'military uniform', 'milk can', 'minibus', 'miniskirt', 'minivan', + 'missile', 'mitten', 'mixing bowl', 'mobile home', 'Model T', 'modem', 'monastery', 'monitor', 'moped', + 'mortar', 'square academic cap', 'mosque', 'mosquito net', 'scooter', 'mountain bike', 'tent', 'computer mouse', + 'mousetrap', 'moving van', 'muzzle', 'nail', 'neck brace', 'necklace', 'nipple', 'notebook computer', 'obelisk', + 'oboe', 'ocarina', 'odometer', 'oil filter', 'organ', 'oscilloscope', 'overskirt', 'bullock cart', + 'oxygen mask', 'packet', 'paddle', 'paddle wheel', 'padlock', 'paintbrush', 'pajamas', 'palace', 'pan flute', + 'paper towel', 'parachute', 'parallel bars', 'park bench', 'parking meter', 'passenger car', 'patio', + 'payphone', 'pedestal', 'pencil case', 'pencil sharpener', 'perfume', 'Petri dish', 'photocopier', 'plectrum', + 'Pickelhaube', 'picket fence', 'pickup truck', 'pier', 'piggy bank', 'pill bottle', 'pillow', 'ping-pong ball', + 'pinwheel', 'pirate ship', 'pitcher', 'hand plane', 'planetarium', 'plastic bag', 'plate rack', 'plow', + 'plunger', 'Polaroid camera', 'pole', 'police van', 'poncho', 'billiard table', 'soda bottle', 'pot', + "potter's wheel", 'power drill', 'prayer rug', 'printer', 'prison', 'projectile', 'projector', 'hockey puck', + 'punching bag', 'purse', 'quill', 'quilt', 'race car', 'racket', 'radiator', 'radio', 'radio telescope', + 'rain barrel', 'recreational vehicle', 'reel', 'reflex camera', 'refrigerator', 'remote control', 'restaurant', + 'revolver', 'rifle', 'rocking chair', 'rotisserie', 'eraser', 'rugby ball', 'ruler', 'running shoe', 'safe', + 'safety pin', 'salt shaker', 'sandal', 'sarong', 'saxophone', 'scabbard', 'weighing scale', 'school bus', + 'schooner', 'scoreboard', 'CRT screen', 'screw', 'screwdriver', 'seat belt', 'sewing machine', 'shield', + 'shoe store', 'shoji', 'shopping basket', 'shopping cart', 'shovel', 'shower cap', 'shower curtain', 'ski', + 'ski mask', 'sleeping bag', 'slide rule', 'sliding door', 'slot machine', 'snorkel', 'snowmobile', 'snowplow', + 'soap dispenser', 'soccer ball', 'sock', 'solar thermal collector', 'sombrero', 'soup bowl', 'space bar', + 'space heater', 'space shuttle', 'spatula', 'motorboat', 'spider web', 'spindle', 'sports car', 'spotlight', + 'stage', 'steam locomotive', 'through arch bridge', 'steel drum', 'stethoscope', 'scarf', 'stone wall', + 'stopwatch', 'stove', 'strainer', 'tram', 'stretcher', 'couch', 'stupa', 'submarine', 'suit', 'sundial', + 'sunglass', 'sunglasses', 'sunscreen', 'suspension bridge', 'mop', 'sweatshirt', 'swimsuit', 'swing', 'switch', + 'syringe', 'table lamp', 'tank', 'tape player', 'teapot', 'teddy bear', 'television', 'tennis ball', + 'thatched roof', 'front curtain', 'thimble', 'threshing machine', 'throne', 'tile roof', 'toaster', + 'tobacco shop', 'toilet seat', 'torch', 'totem pole', 'tow truck', 'toy store', 'tractor', 'semi-trailer truck', + 'tray', 'trench coat', 'tricycle', 'trimaran', 'tripod', 'triumphal arch', 'trolleybus', 'trombone', 'tub', + 'turnstile', 'typewriter keyboard', 'umbrella', 'unicycle', 'upright piano', 'vacuum cleaner', 'vase', 'vault', + 'velvet', 'vending machine', 'vestment', 'viaduct', 'violin', 'volleyball', 'waffle iron', 'wall clock', + 'wallet', 'wardrobe', 'military aircraft', 'sink', 'washing machine', 'water bottle', 'water jug', + 'water tower', 'whiskey jug', 'whistle', 'wig', 'window screen', 'window shade', 'Windsor tie', 'wine bottle', + 'wing', 'wok', 'wooden spoon', 'wool', 'split-rail fence', 'shipwreck', 'yawl', 'yurt', 'website', 'comic book', + 'crossword', 'traffic sign', 'traffic light', 'dust jacket', 'menu', 'plate', 'guacamole', 'consomme', + 'hot pot', 'trifle', 'ice cream', 'ice pop', 'baguette', 'bagel', 'pretzel', 'cheeseburger', 'hot dog', + 'mashed potato', 'cabbage', 'broccoli', 'cauliflower', 'zucchini', 'spaghetti squash', 'acorn squash', + 'butternut squash', 'cucumber', 'artichoke', 'bell pepper', 'cardoon', 'mushroom', 'Granny Smith', 'strawberry', + 'orange', 'lemon', 'fig', 'pineapple', 'banana', 'jackfruit', 'custard apple', 'pomegranate', 'hay', + 'carbonara', 'chocolate syrup', 'dough', 'meatloaf', 'pizza', 'pot pie', 'burrito', 'red wine', 'espresso', + 'cup', 'eggnog', 'alp', 'bubble', 'cliff', 'coral reef', 'geyser', 'lakeshore', 'promontory', 'shoal', + 'seashore', 'valley', 'volcano', 'baseball player', 'bridegroom', 'scuba diver', 'rapeseed', 'daisy', + "yellow lady's slipper", 'corn', 'acorn', 'rose hip', 'horse chestnut seed', 'coral fungus', 'agaric', + 'gyromitra', 'stinkhorn mushroom', 'earth star', 'hen-of-the-woods', 'bolete', 'ear', + 'toilet paper'] # class names + +# Download script/URL (optional) +download: data/scripts/get_imagenet.sh diff --git a/application/yolov5_example/data/Objects365.yaml b/application/yolov5_example/data/Objects365.yaml new file mode 100644 index 00000000..4cc94753 --- /dev/null +++ b/application/yolov5_example/data/Objects365.yaml @@ -0,0 +1,114 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Objects365 dataset https://www.objects365.org/ by Megvii +# Example usage: python train.py --data Objects365.yaml +# parent +# ├── yolov5 +# └── datasets +# └── Objects365 ← downloads here (712 GB = 367G data + 345G zips) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/Objects365 # dataset root dir +train: images/train # train images (relative to 'path') 1742289 images +val: images/val # val images (relative to 'path') 80000 images +test: # test images (optional) + +# Classes +nc: 365 # number of classes +names: ['Person', 'Sneakers', 'Chair', 'Other Shoes', 'Hat', 'Car', 'Lamp', 'Glasses', 'Bottle', 'Desk', 'Cup', + 'Street Lights', 'Cabinet/shelf', 'Handbag/Satchel', 'Bracelet', 'Plate', 'Picture/Frame', 'Helmet', 'Book', + 'Gloves', 'Storage box', 'Boat', 'Leather Shoes', 'Flower', 'Bench', 'Potted Plant', 'Bowl/Basin', 'Flag', + 'Pillow', 'Boots', 'Vase', 'Microphone', 'Necklace', 'Ring', 'SUV', 'Wine Glass', 'Belt', 'Monitor/TV', + 'Backpack', 'Umbrella', 'Traffic Light', 'Speaker', 'Watch', 'Tie', 'Trash bin Can', 'Slippers', 'Bicycle', + 'Stool', 'Barrel/bucket', 'Van', 'Couch', 'Sandals', 'Basket', 'Drum', 'Pen/Pencil', 'Bus', 'Wild Bird', + 'High Heels', 'Motorcycle', 'Guitar', 'Carpet', 'Cell Phone', 'Bread', 'Camera', 'Canned', 'Truck', + 'Traffic cone', 'Cymbal', 'Lifesaver', 'Towel', 'Stuffed Toy', 'Candle', 'Sailboat', 'Laptop', 'Awning', + 'Bed', 'Faucet', 'Tent', 'Horse', 'Mirror', 'Power outlet', 'Sink', 'Apple', 'Air Conditioner', 'Knife', + 'Hockey Stick', 'Paddle', 'Pickup Truck', 'Fork', 'Traffic Sign', 'Balloon', 'Tripod', 'Dog', 'Spoon', 'Clock', + 'Pot', 'Cow', 'Cake', 'Dinning Table', 'Sheep', 'Hanger', 'Blackboard/Whiteboard', 'Napkin', 'Other Fish', + 'Orange/Tangerine', 'Toiletry', 'Keyboard', 'Tomato', 'Lantern', 'Machinery Vehicle', 'Fan', + 'Green Vegetables', 'Banana', 'Baseball Glove', 'Airplane', 'Mouse', 'Train', 'Pumpkin', 'Soccer', 'Skiboard', + 'Luggage', 'Nightstand', 'Tea pot', 'Telephone', 'Trolley', 'Head Phone', 'Sports Car', 'Stop Sign', + 'Dessert', 'Scooter', 'Stroller', 'Crane', 'Remote', 'Refrigerator', 'Oven', 'Lemon', 'Duck', 'Baseball Bat', + 'Surveillance Camera', 'Cat', 'Jug', 'Broccoli', 'Piano', 'Pizza', 'Elephant', 'Skateboard', 'Surfboard', + 'Gun', 'Skating and Skiing shoes', 'Gas stove', 'Donut', 'Bow Tie', 'Carrot', 'Toilet', 'Kite', 'Strawberry', + 'Other Balls', 'Shovel', 'Pepper', 'Computer Box', 'Toilet Paper', 'Cleaning Products', 'Chopsticks', + 'Microwave', 'Pigeon', 'Baseball', 'Cutting/chopping Board', 'Coffee Table', 'Side Table', 'Scissors', + 'Marker', 'Pie', 'Ladder', 'Snowboard', 'Cookies', 'Radiator', 'Fire Hydrant', 'Basketball', 'Zebra', 'Grape', + 'Giraffe', 'Potato', 'Sausage', 'Tricycle', 'Violin', 'Egg', 'Fire Extinguisher', 'Candy', 'Fire Truck', + 'Billiards', 'Converter', 'Bathtub', 'Wheelchair', 'Golf Club', 'Briefcase', 'Cucumber', 'Cigar/Cigarette', + 'Paint Brush', 'Pear', 'Heavy Truck', 'Hamburger', 'Extractor', 'Extension Cord', 'Tong', 'Tennis Racket', + 'Folder', 'American Football', 'earphone', 'Mask', 'Kettle', 'Tennis', 'Ship', 'Swing', 'Coffee Machine', + 'Slide', 'Carriage', 'Onion', 'Green beans', 'Projector', 'Frisbee', 'Washing Machine/Drying Machine', + 'Chicken', 'Printer', 'Watermelon', 'Saxophone', 'Tissue', 'Toothbrush', 'Ice cream', 'Hot-air balloon', + 'Cello', 'French Fries', 'Scale', 'Trophy', 'Cabbage', 'Hot dog', 'Blender', 'Peach', 'Rice', 'Wallet/Purse', + 'Volleyball', 'Deer', 'Goose', 'Tape', 'Tablet', 'Cosmetics', 'Trumpet', 'Pineapple', 'Golf Ball', + 'Ambulance', 'Parking meter', 'Mango', 'Key', 'Hurdle', 'Fishing Rod', 'Medal', 'Flute', 'Brush', 'Penguin', + 'Megaphone', 'Corn', 'Lettuce', 'Garlic', 'Swan', 'Helicopter', 'Green Onion', 'Sandwich', 'Nuts', + 'Speed Limit Sign', 'Induction Cooker', 'Broom', 'Trombone', 'Plum', 'Rickshaw', 'Goldfish', 'Kiwi fruit', + 'Router/modem', 'Poker Card', 'Toaster', 'Shrimp', 'Sushi', 'Cheese', 'Notepaper', 'Cherry', 'Pliers', 'CD', + 'Pasta', 'Hammer', 'Cue', 'Avocado', 'Hamimelon', 'Flask', 'Mushroom', 'Screwdriver', 'Soap', 'Recorder', + 'Bear', 'Eggplant', 'Board Eraser', 'Coconut', 'Tape Measure/Ruler', 'Pig', 'Showerhead', 'Globe', 'Chips', + 'Steak', 'Crosswalk Sign', 'Stapler', 'Camel', 'Formula 1', 'Pomegranate', 'Dishwasher', 'Crab', + 'Hoverboard', 'Meat ball', 'Rice Cooker', 'Tuba', 'Calculator', 'Papaya', 'Antelope', 'Parrot', 'Seal', + 'Butterfly', 'Dumbbell', 'Donkey', 'Lion', 'Urinal', 'Dolphin', 'Electric Drill', 'Hair Dryer', 'Egg tart', + 'Jellyfish', 'Treadmill', 'Lighter', 'Grapefruit', 'Game board', 'Mop', 'Radish', 'Baozi', 'Target', 'French', + 'Spring Rolls', 'Monkey', 'Rabbit', 'Pencil Case', 'Yak', 'Red Cabbage', 'Binoculars', 'Asparagus', 'Barbell', + 'Scallop', 'Noddles', 'Comb', 'Dumpling', 'Oyster', 'Table Tennis paddle', 'Cosmetics Brush/Eyeliner Pencil', + 'Chainsaw', 'Eraser', 'Lobster', 'Durian', 'Okra', 'Lipstick', 'Cosmetics Mirror', 'Curling', 'Table Tennis'] + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + from tqdm import tqdm + + from utils.general import Path, check_requirements, download, np, xyxy2xywhn + + check_requirements(('pycocotools>=2.0',)) + from pycocotools.coco import COCO + + # Make Directories + dir = Path(yaml['path']) # dataset root dir + for p in 'images', 'labels': + (dir / p).mkdir(parents=True, exist_ok=True) + for q in 'train', 'val': + (dir / p / q).mkdir(parents=True, exist_ok=True) + + # Train, Val Splits + for split, patches in [('train', 50 + 1), ('val', 43 + 1)]: + print(f"Processing {split} in {patches} patches ...") + images, labels = dir / 'images' / split, dir / 'labels' / split + + # Download + url = f"https://dorc.ks3-cn-beijing.ksyun.com/data-set/2020Objects365%E6%95%B0%E6%8D%AE%E9%9B%86/{split}/" + if split == 'train': + download([f'{url}zhiyuan_objv2_{split}.tar.gz'], dir=dir, delete=False) # annotations json + download([f'{url}patch{i}.tar.gz' for i in range(patches)], dir=images, curl=True, delete=False, threads=8) + elif split == 'val': + download([f'{url}zhiyuan_objv2_{split}.json'], dir=dir, delete=False) # annotations json + download([f'{url}images/v1/patch{i}.tar.gz' for i in range(15 + 1)], dir=images, curl=True, delete=False, threads=8) + download([f'{url}images/v2/patch{i}.tar.gz' for i in range(16, patches)], dir=images, curl=True, delete=False, threads=8) + + # Move + for f in tqdm(images.rglob('*.jpg'), desc=f'Moving {split} images'): + f.rename(images / f.name) # move to /images/{split} + + # Labels + coco = COCO(dir / f'zhiyuan_objv2_{split}.json') + names = [x["name"] for x in coco.loadCats(coco.getCatIds())] + for cid, cat in enumerate(names): + catIds = coco.getCatIds(catNms=[cat]) + imgIds = coco.getImgIds(catIds=catIds) + for im in tqdm(coco.loadImgs(imgIds), desc=f'Class {cid + 1}/{len(names)} {cat}'): + width, height = im["width"], im["height"] + path = Path(im["file_name"]) # image filename + try: + with open(labels / path.with_suffix('.txt').name, 'a') as file: + annIds = coco.getAnnIds(imgIds=im["id"], catIds=catIds, iscrowd=None) + for a in coco.loadAnns(annIds): + x, y, w, h = a['bbox'] # bounding box in xywh (xy top-left corner) + xyxy = np.array([x, y, x + w, y + h])[None] # pixels(1,4) + x, y, w, h = xyxy2xywhn(xyxy, w=width, h=height, clip=True)[0] # normalized and clipped + file.write(f"{cid} {x:.5f} {y:.5f} {w:.5f} {h:.5f}\n") + except Exception as e: + print(e) diff --git a/application/yolov5_example/data/SKU-110K.yaml b/application/yolov5_example/data/SKU-110K.yaml new file mode 100644 index 00000000..2acf34d1 --- /dev/null +++ b/application/yolov5_example/data/SKU-110K.yaml @@ -0,0 +1,53 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# SKU-110K retail items dataset https://github.com/eg4000/SKU110K_CVPR19 by Trax Retail +# Example usage: python train.py --data SKU-110K.yaml +# parent +# ├── yolov5 +# └── datasets +# └── SKU-110K ← downloads here (13.6 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/SKU-110K # dataset root dir +train: train.txt # train images (relative to 'path') 8219 images +val: val.txt # val images (relative to 'path') 588 images +test: test.txt # test images (optional) 2936 images + +# Classes +nc: 1 # number of classes +names: ['object'] # class names + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + import shutil + from tqdm import tqdm + from utils.general import np, pd, Path, download, xyxy2xywh + + + # Download + dir = Path(yaml['path']) # dataset root dir + parent = Path(dir.parent) # download dir + urls = ['http://trax-geometry.s3.amazonaws.com/cvpr_challenge/SKU110K_fixed.tar.gz'] + download(urls, dir=parent, delete=False) + + # Rename directories + if dir.exists(): + shutil.rmtree(dir) + (parent / 'SKU110K_fixed').rename(dir) # rename dir + (dir / 'labels').mkdir(parents=True, exist_ok=True) # create labels dir + + # Convert labels + names = 'image', 'x1', 'y1', 'x2', 'y2', 'class', 'image_width', 'image_height' # column names + for d in 'annotations_train.csv', 'annotations_val.csv', 'annotations_test.csv': + x = pd.read_csv(dir / 'annotations' / d, names=names).values # annotations + images, unique_images = x[:, 0], np.unique(x[:, 0]) + with open((dir / d).with_suffix('.txt').__str__().replace('annotations_', ''), 'w') as f: + f.writelines(f'./images/{s}\n' for s in unique_images) + for im in tqdm(unique_images, desc=f'Converting {dir / d}'): + cls = 0 # single-class dataset + with open((dir / 'labels' / im).with_suffix('.txt'), 'a') as f: + for r in x[images == im]: + w, h = r[6], r[7] # image width, height + xywh = xyxy2xywh(np.array([[r[1] / w, r[2] / h, r[3] / w, r[4] / h]]))[0] # instance + f.write(f"{cls} {xywh[0]:.5f} {xywh[1]:.5f} {xywh[2]:.5f} {xywh[3]:.5f}\n") # write label diff --git a/application/yolov5_example/data/VOC.yaml b/application/yolov5_example/data/VOC.yaml new file mode 100644 index 00000000..636ddc42 --- /dev/null +++ b/application/yolov5_example/data/VOC.yaml @@ -0,0 +1,81 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# PASCAL VOC dataset http://host.robots.ox.ac.uk/pascal/VOC by University of Oxford +# Example usage: python train.py --data VOC.yaml +# parent +# ├── yolov5 +# └── datasets +# └── VOC ← downloads here (2.8 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/VOC +train: # train images (relative to 'path') 16551 images + - images/train2012 + - images/train2007 + - images/val2012 + - images/val2007 +val: # val images (relative to 'path') 4952 images + - images/test2007 +test: # test images (optional) + - images/test2007 + +# Classes +nc: 20 # number of classes +names: ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', + 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'] # class names + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + import xml.etree.ElementTree as ET + + from tqdm import tqdm + from utils.general import download, Path + + + def convert_label(path, lb_path, year, image_id): + def convert_box(size, box): + dw, dh = 1. / size[0], 1. / size[1] + x, y, w, h = (box[0] + box[1]) / 2.0 - 1, (box[2] + box[3]) / 2.0 - 1, box[1] - box[0], box[3] - box[2] + return x * dw, y * dh, w * dw, h * dh + + in_file = open(path / f'VOC{year}/Annotations/{image_id}.xml') + out_file = open(lb_path, 'w') + tree = ET.parse(in_file) + root = tree.getroot() + size = root.find('size') + w = int(size.find('width').text) + h = int(size.find('height').text) + + for obj in root.iter('object'): + cls = obj.find('name').text + if cls in yaml['names'] and not int(obj.find('difficult').text) == 1: + xmlbox = obj.find('bndbox') + bb = convert_box((w, h), [float(xmlbox.find(x).text) for x in ('xmin', 'xmax', 'ymin', 'ymax')]) + cls_id = yaml['names'].index(cls) # class id + out_file.write(" ".join([str(a) for a in (cls_id, *bb)]) + '\n') + + + # Download + dir = Path(yaml['path']) # dataset root dir + url = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/' + urls = [f'{url}VOCtrainval_06-Nov-2007.zip', # 446MB, 5012 images + f'{url}VOCtest_06-Nov-2007.zip', # 438MB, 4953 images + f'{url}VOCtrainval_11-May-2012.zip'] # 1.95GB, 17126 images + download(urls, dir=dir / 'images', delete=False, curl=True, threads=3) + + # Convert + path = dir / 'images/VOCdevkit' + for year, image_set in ('2012', 'train'), ('2012', 'val'), ('2007', 'train'), ('2007', 'val'), ('2007', 'test'): + imgs_path = dir / 'images' / f'{image_set}{year}' + lbs_path = dir / 'labels' / f'{image_set}{year}' + imgs_path.mkdir(exist_ok=True, parents=True) + lbs_path.mkdir(exist_ok=True, parents=True) + + with open(path / f'VOC{year}/ImageSets/Main/{image_set}.txt') as f: + image_ids = f.read().strip().split() + for id in tqdm(image_ids, desc=f'{image_set}{year}'): + f = path / f'VOC{year}/JPEGImages/{id}.jpg' # old img path + lb_path = (lbs_path / f.name).with_suffix('.txt') # new label path + f.rename(imgs_path / f.name) # move image + convert_label(path, lb_path, year, id) # convert labels to YOLO format diff --git a/application/yolov5_example/data/VisDrone.yaml b/application/yolov5_example/data/VisDrone.yaml new file mode 100644 index 00000000..10337b46 --- /dev/null +++ b/application/yolov5_example/data/VisDrone.yaml @@ -0,0 +1,61 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# VisDrone2019-DET dataset https://github.com/VisDrone/VisDrone-Dataset by Tianjin University +# Example usage: python train.py --data VisDrone.yaml +# parent +# ├── yolov5 +# └── datasets +# └── VisDrone ← downloads here (2.3 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/VisDrone # dataset root dir +train: VisDrone2019-DET-train/images # train images (relative to 'path') 6471 images +val: VisDrone2019-DET-val/images # val images (relative to 'path') 548 images +test: VisDrone2019-DET-test-dev/images # test images (optional) 1610 images + +# Classes +nc: 10 # number of classes +names: ['pedestrian', 'people', 'bicycle', 'car', 'van', 'truck', 'tricycle', 'awning-tricycle', 'bus', 'motor'] + + +# Download script/URL (optional) --------------------------------------------------------------------------------------- +download: | + from utils.general import download, os, Path + + def visdrone2yolo(dir): + from PIL import Image + from tqdm import tqdm + + def convert_box(size, box): + # Convert VisDrone box to YOLO xywh box + dw = 1. / size[0] + dh = 1. / size[1] + return (box[0] + box[2] / 2) * dw, (box[1] + box[3] / 2) * dh, box[2] * dw, box[3] * dh + + (dir / 'labels').mkdir(parents=True, exist_ok=True) # make labels directory + pbar = tqdm((dir / 'annotations').glob('*.txt'), desc=f'Converting {dir}') + for f in pbar: + img_size = Image.open((dir / 'images' / f.name).with_suffix('.jpg')).size + lines = [] + with open(f, 'r') as file: # read annotation.txt + for row in [x.split(',') for x in file.read().strip().splitlines()]: + if row[4] == '0': # VisDrone 'ignored regions' class 0 + continue + cls = int(row[5]) - 1 + box = convert_box(img_size, tuple(map(int, row[:4]))) + lines.append(f"{cls} {' '.join(f'{x:.6f}' for x in box)}\n") + with open(str(f).replace(os.sep + 'annotations' + os.sep, os.sep + 'labels' + os.sep), 'w') as fl: + fl.writelines(lines) # write label.txt + + + # Download + dir = Path(yaml['path']) # dataset root dir + urls = ['https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-train.zip', + 'https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-val.zip', + 'https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-test-dev.zip', + 'https://github.com/ultralytics/yolov5/releases/download/v1.0/VisDrone2019-DET-test-challenge.zip'] + download(urls, dir=dir, curl=True, threads=4) + + # Convert + for d in 'VisDrone2019-DET-train', 'VisDrone2019-DET-val', 'VisDrone2019-DET-test-dev': + visdrone2yolo(dir / d) # convert VisDrone annotations to YOLO labels diff --git a/application/yolov5_example/data/coco.yaml b/application/yolov5_example/data/coco.yaml new file mode 100644 index 00000000..0c0c4ada --- /dev/null +++ b/application/yolov5_example/data/coco.yaml @@ -0,0 +1,45 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# COCO 2017 dataset http://cocodataset.org by Microsoft +# Example usage: python train.py --data coco.yaml +# parent +# ├── yolov5 +# └── datasets +# └── coco ← downloads here (20.1 GB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/coco # dataset root dir +train: train2017.txt # train images (relative to 'path') 118287 images +val: val2017.txt # val images (relative to 'path') 5000 images +test: test-dev2017.txt # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794 + +# Classes +nc: 80 # number of classes +names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', + 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', + 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', + 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', + 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', + 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', + 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', + 'hair drier', 'toothbrush'] # class names + + +# Download script/URL (optional) +download: | + from utils.general import download, Path + + + # Download labels + segments = False # segment or box labels + dir = Path(yaml['path']) # dataset root dir + url = 'https://github.com/ultralytics/yolov5/releases/download/v1.0/' + urls = [url + ('coco2017labels-segments.zip' if segments else 'coco2017labels.zip')] # labels + download(urls, dir=dir.parent) + + # Download data + urls = ['http://images.cocodataset.org/zips/train2017.zip', # 19G, 118k images + 'http://images.cocodataset.org/zips/val2017.zip', # 1G, 5k images + 'http://images.cocodataset.org/zips/test2017.zip'] # 7G, 41k images (optional) + download(urls, dir=dir / 'images', threads=3) diff --git a/application/yolov5_example/data/coco128.yaml b/application/yolov5_example/data/coco128.yaml new file mode 100644 index 00000000..2517d207 --- /dev/null +++ b/application/yolov5_example/data/coco128.yaml @@ -0,0 +1,30 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# COCO128 dataset https://www.kaggle.com/ultralytics/coco128 (first 128 images from COCO train2017) by Ultralytics +# Example usage: python train.py --data coco128.yaml +# parent +# ├── yolov5 +# └── datasets +# └── coco128 ← downloads here (7 MB) + + +# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..] +path: ../datasets/coco128 # dataset root dir +train: images/train2017 # train images (relative to 'path') 128 images +val: images/train2017 # val images (relative to 'path') 128 images +test: # test images (optional) + +# Classes +nc: 80 # number of classes +names: ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', + 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', + 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', + 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', + 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', + 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', + 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', + 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', + 'hair drier', 'toothbrush'] # class names + + +# Download script/URL (optional) +download: https://ultralytics.com/assets/coco128.zip diff --git a/application/yolov5_example/data/hyps/hyp.Objects365.yaml b/application/yolov5_example/data/hyps/hyp.Objects365.yaml new file mode 100644 index 00000000..74971740 --- /dev/null +++ b/application/yolov5_example/data/hyps/hyp.Objects365.yaml @@ -0,0 +1,34 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for Objects365 training +# python train.py --weights yolov5m.pt --data Objects365.yaml --evolve +# See Hyperparameter Evolution tutorial for details https://github.com/ultralytics/yolov5#tutorials + +lr0: 0.00258 +lrf: 0.17 +momentum: 0.779 +weight_decay: 0.00058 +warmup_epochs: 1.33 +warmup_momentum: 0.86 +warmup_bias_lr: 0.0711 +box: 0.0539 +cls: 0.299 +cls_pw: 0.825 +obj: 0.632 +obj_pw: 1.0 +iou_t: 0.2 +anchor_t: 3.44 +anchors: 3.2 +fl_gamma: 0.0 +hsv_h: 0.0188 +hsv_s: 0.704 +hsv_v: 0.36 +degrees: 0.0 +translate: 0.0902 +scale: 0.491 +shear: 0.0 +perspective: 0.0 +flipud: 0.0 +fliplr: 0.5 +mosaic: 1.0 +mixup: 0.0 +copy_paste: 0.0 diff --git a/application/yolov5_example/data/hyps/hyp.VOC.yaml b/application/yolov5_example/data/hyps/hyp.VOC.yaml new file mode 100644 index 00000000..0aa4e7d9 --- /dev/null +++ b/application/yolov5_example/data/hyps/hyp.VOC.yaml @@ -0,0 +1,40 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for VOC training +# python train.py --batch 128 --weights yolov5m6.pt --data VOC.yaml --epochs 50 --img 512 --hyp hyp.scratch-med.yaml --evolve +# See Hyperparameter Evolution tutorial for details https://github.com/ultralytics/yolov5#tutorials + +# YOLOv5 Hyperparameter Evolution Results +# Best generation: 467 +# Last generation: 996 +# metrics/precision, metrics/recall, metrics/mAP_0.5, metrics/mAP_0.5:0.95, val/box_loss, val/obj_loss, val/cls_loss +# 0.87729, 0.85125, 0.91286, 0.72664, 0.0076739, 0.0042529, 0.0013865 + +lr0: 0.00334 +lrf: 0.15135 +momentum: 0.74832 +weight_decay: 0.00025 +warmup_epochs: 3.3835 +warmup_momentum: 0.59462 +warmup_bias_lr: 0.18657 +box: 0.02 +cls: 0.21638 +cls_pw: 0.5 +obj: 0.51728 +obj_pw: 0.67198 +iou_t: 0.2 +anchor_t: 3.3744 +fl_gamma: 0.0 +hsv_h: 0.01041 +hsv_s: 0.54703 +hsv_v: 0.27739 +degrees: 0.0 +translate: 0.04591 +scale: 0.75544 +shear: 0.0 +perspective: 0.0 +flipud: 0.0 +fliplr: 0.5 +mosaic: 0.85834 +mixup: 0.04266 +copy_paste: 0.0 +anchors: 3.412 diff --git a/application/yolov5_example/data/hyps/hyp.scratch-high.yaml b/application/yolov5_example/data/hyps/hyp.scratch-high.yaml new file mode 100644 index 00000000..123cc840 --- /dev/null +++ b/application/yolov5_example/data/hyps/hyp.scratch-high.yaml @@ -0,0 +1,34 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for high-augmentation COCO training from scratch +# python train.py --batch 32 --cfg yolov5m6.yaml --weights '' --data coco.yaml --img 1280 --epochs 300 +# See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials + +lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3) +lrf: 0.1 # final OneCycleLR learning rate (lr0 * lrf) +momentum: 0.937 # SGD momentum/Adam beta1 +weight_decay: 0.0005 # optimizer weight decay 5e-4 +warmup_epochs: 3.0 # warmup epochs (fractions ok) +warmup_momentum: 0.8 # warmup initial momentum +warmup_bias_lr: 0.1 # warmup initial bias lr +box: 0.05 # box loss gain +cls: 0.3 # cls loss gain +cls_pw: 1.0 # cls BCELoss positive_weight +obj: 0.7 # obj loss gain (scale with pixels) +obj_pw: 1.0 # obj BCELoss positive_weight +iou_t: 0.20 # IoU training threshold +anchor_t: 4.0 # anchor-multiple threshold +# anchors: 3 # anchors per output layer (0 to ignore) +fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) +hsv_h: 0.015 # image HSV-Hue augmentation (fraction) +hsv_s: 0.7 # image HSV-Saturation augmentation (fraction) +hsv_v: 0.4 # image HSV-Value augmentation (fraction) +degrees: 0.0 # image rotation (+/- deg) +translate: 0.1 # image translation (+/- fraction) +scale: 0.9 # image scale (+/- gain) +shear: 0.0 # image shear (+/- deg) +perspective: 0.0 # image perspective (+/- fraction), range 0-0.001 +flipud: 0.0 # image flip up-down (probability) +fliplr: 0.5 # image flip left-right (probability) +mosaic: 1.0 # image mosaic (probability) +mixup: 0.1 # image mixup (probability) +copy_paste: 0.1 # segment copy-paste (probability) diff --git a/application/yolov5_example/data/hyps/hyp.scratch-low.yaml b/application/yolov5_example/data/hyps/hyp.scratch-low.yaml new file mode 100644 index 00000000..b77bf312 --- /dev/null +++ b/application/yolov5_example/data/hyps/hyp.scratch-low.yaml @@ -0,0 +1,34 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for low-augmentation COCO training from scratch +# python train.py --batch 64 --cfg yolov5n6.yaml --weights '' --data coco.yaml --img 640 --epochs 300 --linear +# See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials + +lr0: 0.001 # initial learning rate (SGD=1E-2, Adam=1E-3) +lrf: 0.01 # final OneCycleLR learning rate (lr0 * lrf) +momentum: 0.937 # SGD momentum/Adam beta1 +weight_decay: 0.0005 # optimizer weight decay 5e-4 +warmup_epochs: 3.0 # warmup epochs (fractions ok) +warmup_momentum: 0.8 # warmup initial momentum +warmup_bias_lr: 0.1 # warmup initial bias lr +box: 0.05 # box loss gain +cls: 0.5 # cls loss gain +cls_pw: 1.0 # cls BCELoss positive_weight +obj: 1.0 # obj loss gain (scale with pixels) +obj_pw: 1.0 # obj BCELoss positive_weight +iou_t: 0.20 # IoU training threshold +anchor_t: 4.0 # anchor-multiple threshold +# anchors: 3 # anchors per output layer (0 to ignore) +fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) +hsv_h: 0.015 # image HSV-Hue augmentation (fraction) +hsv_s: 0.7 # image HSV-Saturation augmentation (fraction) +hsv_v: 0.4 # image HSV-Value augmentation (fraction) +degrees: 0.0 # image rotation (+/- deg) +translate: 0.1 # image translation (+/- fraction) +scale: 0.5 # image scale (+/- gain) +shear: 0.0 # image shear (+/- deg) +perspective: 0.0 # image perspective (+/- fraction), range 0-0.001 +flipud: 0.0 # image flip up-down (probability) +fliplr: 0.5 # image flip left-right (probability) +mosaic: 1.0 # image mosaic (probability) +mixup: 0.0 # image mixup (probability) +copy_paste: 0.0 # segment copy-paste (probability) diff --git a/application/yolov5_example/data/hyps/hyp.scratch-med.yaml b/application/yolov5_example/data/hyps/hyp.scratch-med.yaml new file mode 100644 index 00000000..d6867d75 --- /dev/null +++ b/application/yolov5_example/data/hyps/hyp.scratch-med.yaml @@ -0,0 +1,34 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Hyperparameters for medium-augmentation COCO training from scratch +# python train.py --batch 32 --cfg yolov5m6.yaml --weights '' --data coco.yaml --img 1280 --epochs 300 +# See tutorials for hyperparameter evolution https://github.com/ultralytics/yolov5#tutorials + +lr0: 0.01 # initial learning rate (SGD=1E-2, Adam=1E-3) +lrf: 0.1 # final OneCycleLR learning rate (lr0 * lrf) +momentum: 0.937 # SGD momentum/Adam beta1 +weight_decay: 0.0005 # optimizer weight decay 5e-4 +warmup_epochs: 3.0 # warmup epochs (fractions ok) +warmup_momentum: 0.8 # warmup initial momentum +warmup_bias_lr: 0.1 # warmup initial bias lr +box: 0.05 # box loss gain +cls: 0.3 # cls loss gain +cls_pw: 1.0 # cls BCELoss positive_weight +obj: 0.7 # obj loss gain (scale with pixels) +obj_pw: 1.0 # obj BCELoss positive_weight +iou_t: 0.20 # IoU training threshold +anchor_t: 4.0 # anchor-multiple threshold +# anchors: 3 # anchors per output layer (0 to ignore) +fl_gamma: 0.0 # focal loss gamma (efficientDet default gamma=1.5) +hsv_h: 0.015 # image HSV-Hue augmentation (fraction) +hsv_s: 0.7 # image HSV-Saturation augmentation (fraction) +hsv_v: 0.4 # image HSV-Value augmentation (fraction) +degrees: 0.0 # image rotation (+/- deg) +translate: 0.1 # image translation (+/- fraction) +scale: 0.9 # image scale (+/- gain) +shear: 0.0 # image shear (+/- deg) +perspective: 0.0 # image perspective (+/- fraction), range 0-0.001 +flipud: 0.0 # image flip up-down (probability) +fliplr: 0.5 # image flip left-right (probability) +mosaic: 1.0 # image mosaic (probability) +mixup: 0.1 # image mixup (probability) +copy_paste: 0.0 # segment copy-paste (probability) diff --git a/application/yolov5_example/data/images/bus.jpg b/application/yolov5_example/data/images/bus.jpg new file mode 100644 index 00000000..b43e3111 Binary files /dev/null and b/application/yolov5_example/data/images/bus.jpg differ diff --git a/application/yolov5_example/data/images/zidane.jpg b/application/yolov5_example/data/images/zidane.jpg new file mode 100644 index 00000000..92d72ea1 Binary files /dev/null and b/application/yolov5_example/data/images/zidane.jpg differ diff --git a/application/yolov5_example/data/scripts/download_weights.sh b/application/yolov5_example/data/scripts/download_weights.sh new file mode 100644 index 00000000..a4f3becf --- /dev/null +++ b/application/yolov5_example/data/scripts/download_weights.sh @@ -0,0 +1,21 @@ +#!/bin/bash +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Download latest models from https://github.com/ultralytics/yolov5/releases +# Example usage: bash data/scripts/download_weights.sh +# parent +# └── yolov5 +# ├── yolov5s.pt ← downloads here +# ├── yolov5m.pt +# └── ... + +python - <= cls >= 0, f'incorrect class index {cls}' + + # Write YOLO label + if id not in shapes: + shapes[id] = Image.open(file).size + box = xyxy2xywhn(box[None].astype(np.float), w=shapes[id][0], h=shapes[id][1], clip=True) + with open((labels / id).with_suffix('.txt'), 'a') as f: + f.write(f"{cls} {' '.join(f'{x:.6f}' for x in box[0])}\n") # write label.txt + except Exception as e: + print(f'WARNING: skipping one label for {file}: {e}') + + + # Download manually from https://challenge.xviewdataset.org + dir = Path(yaml['path']) # dataset root dir + # urls = ['https://d307kc0mrhucc3.cloudfront.net/train_labels.zip', # train labels + # 'https://d307kc0mrhucc3.cloudfront.net/train_images.zip', # 15G, 847 train images + # 'https://d307kc0mrhucc3.cloudfront.net/val_images.zip'] # 5G, 282 val images (no labels) + # download(urls, dir=dir, delete=False) + + # Convert labels + convert_labels(dir / 'xView_train.geojson') + + # Move images + images = Path(dir / 'images') + images.mkdir(parents=True, exist_ok=True) + Path(dir / 'train_images').rename(dir / 'images' / 'train') + Path(dir / 'val_images').rename(dir / 'images' / 'val') + + # Split + autosplit(dir / 'images' / 'train') diff --git a/application/yolov5_example/detect.py b/application/yolov5_example/detect.py new file mode 100644 index 00000000..c699a749 --- /dev/null +++ b/application/yolov5_example/detect.py @@ -0,0 +1,257 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Run inference on images, videos, directories, streams, etc. + +Usage - sources: + $ python path/to/detect.py --weights yolov5s.pt --source 0 # webcam + img.jpg # image + vid.mp4 # video + path/ # directory + path/*.jpg # glob + 'https://youtu.be/Zgi9g1ksQHc' # YouTube + 'rtsp://example.com/media.mp4' # RTSP, RTMP, HTTP stream + +Usage - formats: + $ python path/to/detect.py --weights yolov5s.pt # PyTorch + yolov5s.torchscript # TorchScript + yolov5s.onnx # ONNX Runtime or OpenCV DNN with --dnn + yolov5s.xml # OpenVINO + yolov5s.engine # TensorRT + yolov5s.mlmodel # CoreML (macOS-only) + yolov5s_saved_model # TensorFlow SavedModel + yolov5s.pb # TensorFlow GraphDef + yolov5s.tflite # TensorFlow Lite + yolov5s_edgetpu.tflite # TensorFlow Edge TPU +""" + +import argparse +import os +import platform +import sys +from pathlib import Path + +import torch +import torch.backends.cudnn as cudnn + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from models.common import DetectMultiBackend +from utils.dataloaders import IMG_FORMATS, VID_FORMATS, LoadImages, LoadStreams +from utils.general import (LOGGER, check_file, check_img_size, check_imshow, check_requirements, colorstr, cv2, + increment_path, non_max_suppression, print_args, scale_coords, strip_optimizer, xyxy2xywh) +from utils.plots import Annotator, colors, save_one_box +from utils.torch_utils import select_device, smart_inference_mode, time_sync + + +@smart_inference_mode() +def run( + weights=ROOT / 'yolov5s.pt', # model.pt path(s) + source=ROOT / 'data/images', # file/dir/URL/glob, 0 for webcam + data=ROOT / 'data/coco128.yaml', # dataset.yaml path + imgsz=(640, 640), # inference size (height, width) + conf_thres=0.25, # confidence threshold + iou_thres=0.45, # NMS IOU threshold + max_det=1000, # maximum detections per image + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + view_img=False, # show results + save_txt=False, # save results to *.txt + save_conf=False, # save confidences in --save-txt labels + save_crop=False, # save cropped prediction boxes + nosave=False, # do not save images/videos + classes=None, # filter by class: --class 0, or --class 0 2 3 + agnostic_nms=False, # class-agnostic NMS + augment=False, # augmented inference + visualize=False, # visualize features + update=False, # update all models + project=ROOT / 'runs/detect', # save results to project/name + name='exp', # save results to project/name + exist_ok=False, # existing project/name ok, do not increment + line_thickness=3, # bounding box thickness (pixels) + hide_labels=False, # hide labels + hide_conf=False, # hide confidences + half=False, # use FP16 half-precision inference + dnn=False, # use OpenCV DNN for ONNX inference +): + source = str(source) + save_img = not nosave and not source.endswith('.txt') # save inference images + is_file = Path(source).suffix[1:] in (IMG_FORMATS + VID_FORMATS) + is_url = source.lower().startswith(('rtsp://', 'rtmp://', 'http://', 'https://')) + webcam = source.isnumeric() or source.endswith('.txt') or (is_url and not is_file) + if is_url and is_file: + source = check_file(source) # download + + # Directories + save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run + (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir + + # Load model + device = select_device(device) + model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half) + stride, names, pt = model.stride, model.names, model.pt + imgsz = check_img_size(imgsz, s=stride) # check image size + + # Dataloader + if webcam: + view_img = check_imshow() + cudnn.benchmark = True # set True to speed up constant image size inference + dataset = LoadStreams(source, img_size=imgsz, stride=stride, auto=pt) + bs = len(dataset) # batch_size + else: + dataset = LoadImages(source, img_size=imgsz, stride=stride, auto=pt) + bs = 1 # batch_size + vid_path, vid_writer = [None] * bs, [None] * bs + + # Run inference + model.warmup(imgsz=(1 if pt else bs, 3, *imgsz)) # warmup + seen, windows, dt = 0, [], [0.0, 0.0, 0.0] + for path, im, im0s, vid_cap, s in dataset: + t1 = time_sync() + im = torch.from_numpy(im).to(device) + im = im.half() if model.fp16 else im.float() # uint8 to fp16/32 + im /= 255 # 0 - 255 to 0.0 - 1.0 + if len(im.shape) == 3: + im = im[None] # expand for batch dim + t2 = time_sync() + dt[0] += t2 - t1 + + # Inference + visualize = increment_path(save_dir / Path(path).stem, mkdir=True) if visualize else False + pred = model(im, augment=augment, visualize=visualize) + t3 = time_sync() + dt[1] += t3 - t2 + + # NMS + pred = non_max_suppression(pred, conf_thres, iou_thres, classes, agnostic_nms, max_det=max_det) + dt[2] += time_sync() - t3 + + # Second-stage classifier (optional) + # pred = utils.general.apply_classifier(pred, classifier_model, im, im0s) + + # Process predictions + for i, det in enumerate(pred): # per image + seen += 1 + if webcam: # batch_size >= 1 + p, im0, frame = path[i], im0s[i].copy(), dataset.count + s += f'{i}: ' + else: + p, im0, frame = path, im0s.copy(), getattr(dataset, 'frame', 0) + + p = Path(p) # to Path + save_path = str(save_dir / p.name) # im.jpg + txt_path = str(save_dir / 'labels' / p.stem) + ('' if dataset.mode == 'image' else f'_{frame}') # im.txt + s += '%gx%g ' % im.shape[2:] # print string + gn = torch.tensor(im0.shape)[[1, 0, 1, 0]] # normalization gain whwh + imc = im0.copy() if save_crop else im0 # for save_crop + annotator = Annotator(im0, line_width=line_thickness, example=str(names)) + if len(det): + # Rescale boxes from img_size to im0 size + det[:, :4] = scale_coords(im.shape[2:], det[:, :4], im0.shape).round() + + # Print results + for c in det[:, -1].unique(): + n = (det[:, -1] == c).sum() # detections per class + s += f"{n} {names[int(c)]}{'s' * (n > 1)}, " # add to string + + # Write results + for *xyxy, conf, cls in reversed(det): + if save_txt: # Write to file + xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh + line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format + with open(f'{txt_path}.txt', 'a') as f: + f.write(('%g ' * len(line)).rstrip() % line + '\n') + + if save_img or save_crop or view_img: # Add bbox to image + c = int(cls) # integer class + label = None if hide_labels else (names[c] if hide_conf else f'{names[c]} {conf:.2f}') + annotator.box_label(xyxy, label, color=colors(c, True)) + if save_crop: + save_one_box(xyxy, imc, file=save_dir / 'crops' / names[c] / f'{p.stem}.jpg', BGR=True) + + # Stream results + im0 = annotator.result() + if view_img: + if platform.system() == 'Linux' and p not in windows: + windows.append(p) + cv2.namedWindow(str(p), cv2.WINDOW_NORMAL | cv2.WINDOW_KEEPRATIO) # allow window resize (Linux) + cv2.resizeWindow(str(p), im0.shape[1], im0.shape[0]) + cv2.imshow(str(p), im0) + cv2.waitKey(1) # 1 millisecond + + # Save results (image with detections) + if save_img: + if dataset.mode == 'image': + cv2.imwrite(save_path, im0) + else: # 'video' or 'stream' + if vid_path[i] != save_path: # new video + vid_path[i] = save_path + if isinstance(vid_writer[i], cv2.VideoWriter): + vid_writer[i].release() # release previous video writer + if vid_cap: # video + fps = vid_cap.get(cv2.CAP_PROP_FPS) + w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) + h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) + else: # stream + fps, w, h = 30, im0.shape[1], im0.shape[0] + save_path = str(Path(save_path).with_suffix('.mp4')) # force *.mp4 suffix on results videos + vid_writer[i] = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (w, h)) + vid_writer[i].write(im0) + + # Print time (inference-only) + LOGGER.info(f'{s}Done. ({t3 - t2:.3f}s)') + + # Print results + t = tuple(x / seen * 1E3 for x in dt) # speeds per image + LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {(1, 3, *imgsz)}' % t) + if save_txt or save_img: + s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' + LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}") + if update: + strip_optimizer(weights[0]) # update model (to fix SourceChangeWarning) + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s.pt', help='model path(s)') + parser.add_argument('--source', type=str, default=ROOT / 'data/images', help='file/dir/URL/glob, 0 for webcam') + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='(optional) dataset.yaml path') + parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w') + parser.add_argument('--conf-thres', type=float, default=0.25, help='confidence threshold') + parser.add_argument('--iou-thres', type=float, default=0.45, help='NMS IoU threshold') + parser.add_argument('--max-det', type=int, default=1000, help='maximum detections per image') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--view-img', action='store_true', help='show results') + parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') + parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') + parser.add_argument('--save-crop', action='store_true', help='save cropped prediction boxes') + parser.add_argument('--nosave', action='store_true', help='do not save images/videos') + parser.add_argument('--classes', nargs='+', type=int, help='filter by class: --classes 0, or --classes 0 2 3') + parser.add_argument('--agnostic-nms', action='store_true', help='class-agnostic NMS') + parser.add_argument('--augment', action='store_true', help='augmented inference') + parser.add_argument('--visualize', action='store_true', help='visualize features') + parser.add_argument('--update', action='store_true', help='update all models') + parser.add_argument('--project', default=ROOT / 'runs/detect', help='save results to project/name') + parser.add_argument('--name', default='exp', help='save results to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--line-thickness', default=3, type=int, help='bounding box thickness (pixels)') + parser.add_argument('--hide-labels', default=False, action='store_true', help='hide labels') + parser.add_argument('--hide-conf', default=False, action='store_true', help='hide confidences') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference') + opt = parser.parse_args() + opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1 # expand + print_args(vars(opt)) + return opt + + +def main(opt): + check_requirements(exclude=('tensorboard', 'thop')) + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/application/yolov5_example/export.py b/application/yolov5_example/export.py new file mode 100644 index 00000000..595039b2 --- /dev/null +++ b/application/yolov5_example/export.py @@ -0,0 +1,616 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Export a YOLOv5 PyTorch model to other formats. TensorFlow exports authored by https://github.com/zldrobit + +Format | `export.py --include` | Model +--- | --- | --- +PyTorch | - | yolov5s.pt +TorchScript | `torchscript` | yolov5s.torchscript +ONNX | `onnx` | yolov5s.onnx +OpenVINO | `openvino` | yolov5s_openvino_model/ +TensorRT | `engine` | yolov5s.engine +CoreML | `coreml` | yolov5s.mlmodel +TensorFlow SavedModel | `saved_model` | yolov5s_saved_model/ +TensorFlow GraphDef | `pb` | yolov5s.pb +TensorFlow Lite | `tflite` | yolov5s.tflite +TensorFlow Edge TPU | `edgetpu` | yolov5s_edgetpu.tflite +TensorFlow.js | `tfjs` | yolov5s_web_model/ + +Requirements: + $ pip install -r requirements.txt coremltools onnx onnx-simplifier onnxruntime openvino-dev tensorflow-cpu # CPU + $ pip install -r requirements.txt coremltools onnx onnx-simplifier onnxruntime-gpu openvino-dev tensorflow # GPU + +Usage: + $ python path/to/export.py --weights yolov5s.pt --include torchscript onnx openvino engine coreml tflite ... + +Inference: + $ python path/to/detect.py --weights yolov5s.pt # PyTorch + yolov5s.torchscript # TorchScript + yolov5s.onnx # ONNX Runtime or OpenCV DNN with --dnn + yolov5s.xml # OpenVINO + yolov5s.engine # TensorRT + yolov5s.mlmodel # CoreML (macOS-only) + yolov5s_saved_model # TensorFlow SavedModel + yolov5s.pb # TensorFlow GraphDef + yolov5s.tflite # TensorFlow Lite + yolov5s_edgetpu.tflite # TensorFlow Edge TPU + +TensorFlow.js: + $ cd .. && git clone https://github.com/zldrobit/tfjs-yolov5-example.git && cd tfjs-yolov5-example + $ npm install + $ ln -s ../../yolov5/yolov5s_web_model public/yolov5s_web_model + $ npm start +""" + +import argparse +import json +import os +import platform +import subprocess +import sys +import time +import warnings +from pathlib import Path + +import pandas as pd +import torch +import yaml +from torch.utils.mobile_optimizer import optimize_for_mobile + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +if platform.system() != 'Windows': + ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from models.experimental import attempt_load +from models.yolo import Detect +from utils.dataloaders import LoadImages +from utils.general import (LOGGER, check_dataset, check_img_size, check_requirements, check_version, check_yaml, + colorstr, file_size, print_args, url2file) +from utils.torch_utils import select_device, smart_inference_mode + + +def export_formats(): + # YOLOv5 export formats + x = [ + ['PyTorch', '-', '.pt', True, True], + ['TorchScript', 'torchscript', '.torchscript', True, True], + ['ONNX', 'onnx', '.onnx', True, True], + ['OpenVINO', 'openvino', '_openvino_model', True, False], + ['TensorRT', 'engine', '.engine', False, True], + ['CoreML', 'coreml', '.mlmodel', True, False], + ['TensorFlow SavedModel', 'saved_model', '_saved_model', True, True], + ['TensorFlow GraphDef', 'pb', '.pb', True, True], + ['TensorFlow Lite', 'tflite', '.tflite', True, False], + ['TensorFlow Edge TPU', 'edgetpu', '_edgetpu.tflite', False, False], + ['TensorFlow.js', 'tfjs', '_web_model', False, False],] + return pd.DataFrame(x, columns=['Format', 'Argument', 'Suffix', 'CPU', 'GPU']) + + +def export_torchscript(model, im, file, optimize, prefix=colorstr('TorchScript:')): + # YOLOv5 TorchScript model export + try: + LOGGER.info(f'\n{prefix} starting export with torch {torch.__version__}...') + f = file.with_suffix('.torchscript') + + ts = torch.jit.trace(model, im, strict=False) + d = {"shape": im.shape, "stride": int(max(model.stride)), "names": model.names} + extra_files = {'config.txt': json.dumps(d)} # torch._C.ExtraFilesMap() + if optimize: # https://pytorch.org/tutorials/recipes/mobile_interpreter.html + optimize_for_mobile(ts)._save_for_lite_interpreter(str(f), _extra_files=extra_files) + else: + ts.save(str(f), _extra_files=extra_files) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'{prefix} export failure: {e}') + + +def export_onnx(model, im, file, opset, train, dynamic, simplify, prefix=colorstr('ONNX:')): + # YOLOv5 ONNX export + try: + check_requirements(('onnx',)) + import onnx + + LOGGER.info(f'\n{prefix} starting export with onnx {onnx.__version__}...') + f = file.with_suffix('.onnx') + + torch.onnx.export( + model.cpu() if dynamic else model, # --dynamic only compatible with cpu + im.cpu() if dynamic else im, + f, + verbose=False, + opset_version=opset, + training=torch.onnx.TrainingMode.TRAINING if train else torch.onnx.TrainingMode.EVAL, + do_constant_folding=not train, + input_names=['images'], + output_names=['output'], + dynamic_axes={ + 'images': { + 0: 'batch', + 2: 'height', + 3: 'width'}, # shape(1,3,640,640) + 'output': { + 0: 'batch', + 1: 'anchors'} # shape(1,25200,85) + } if dynamic else None) + + # Checks + model_onnx = onnx.load(f) # load onnx model + onnx.checker.check_model(model_onnx) # check onnx model + + # Metadata + d = {'stride': int(max(model.stride)), 'names': model.names} + for k, v in d.items(): + meta = model_onnx.metadata_props.add() + meta.key, meta.value = k, str(v) + onnx.save(model_onnx, f) + + # Simplify + if simplify: + try: + cuda = torch.cuda.is_available() + check_requirements(('onnxruntime-gpu' if cuda else 'onnxruntime', 'onnx-simplifier>=0.4.1')) + import onnxsim + + LOGGER.info(f'{prefix} simplifying with onnx-simplifier {onnxsim.__version__}...') + model_onnx, check = onnxsim.simplify(model_onnx) + assert check, 'assert check failed' + onnx.save(model_onnx, f) + except Exception as e: + LOGGER.info(f'{prefix} simplifier failure: {e}') + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'{prefix} export failure: {e}') + + +def export_openvino(model, file, half, prefix=colorstr('OpenVINO:')): + # YOLOv5 OpenVINO export + try: + check_requirements(('openvino-dev',)) # requires openvino-dev: https://pypi.org/project/openvino-dev/ + import openvino.inference_engine as ie + + LOGGER.info(f'\n{prefix} starting export with openvino {ie.__version__}...') + f = str(file).replace('.pt', f'_openvino_model{os.sep}') + + cmd = f"mo --input_model {file.with_suffix('.onnx')} --output_dir {f} --data_type {'FP16' if half else 'FP32'}" + subprocess.check_output(cmd.split()) # export + with open(Path(f) / file.with_suffix('.yaml').name, 'w') as g: + yaml.dump({'stride': int(max(model.stride)), 'names': model.names}, g) # add metadata.yaml + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_coreml(model, im, file, int8, half, prefix=colorstr('CoreML:')): + # YOLOv5 CoreML export + try: + check_requirements(('coremltools',)) + import coremltools as ct + + LOGGER.info(f'\n{prefix} starting export with coremltools {ct.__version__}...') + f = file.with_suffix('.mlmodel') + + ts = torch.jit.trace(model, im, strict=False) # TorchScript model + ct_model = ct.convert(ts, inputs=[ct.ImageType('image', shape=im.shape, scale=1 / 255, bias=[0, 0, 0])]) + bits, mode = (8, 'kmeans_lut') if int8 else (16, 'linear') if half else (32, None) + if bits < 32: + if platform.system() == 'Darwin': # quantization only supported on macOS + with warnings.catch_warnings(): + warnings.filterwarnings("ignore", category=DeprecationWarning) # suppress numpy==1.20 float warning + ct_model = ct.models.neural_network.quantization_utils.quantize_weights(ct_model, bits, mode) + else: + print(f'{prefix} quantization only supported on macOS, skipping...') + ct_model.save(f) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return ct_model, f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + return None, None + + +def export_engine(model, im, file, train, half, dynamic, simplify, workspace=4, verbose=False): + # YOLOv5 TensorRT export https://developer.nvidia.com/tensorrt + prefix = colorstr('TensorRT:') + try: + assert im.device.type != 'cpu', 'export running on CPU but must be on GPU, i.e. `python export.py --device 0`' + try: + import tensorrt as trt + except Exception: + if platform.system() == 'Linux': + check_requirements(('nvidia-tensorrt',), cmds=('-U --index-url https://pypi.ngc.nvidia.com',)) + import tensorrt as trt + + if trt.__version__[0] == '7': # TensorRT 7 handling https://github.com/ultralytics/yolov5/issues/6012 + grid = model.model[-1].anchor_grid + model.model[-1].anchor_grid = [a[..., :1, :1, :] for a in grid] + export_onnx(model, im, file, 12, train, dynamic, simplify) # opset 12 + model.model[-1].anchor_grid = grid + else: # TensorRT >= 8 + check_version(trt.__version__, '8.0.0', hard=True) # require tensorrt>=8.0.0 + export_onnx(model, im, file, 13, train, dynamic, simplify) # opset 13 + onnx = file.with_suffix('.onnx') + + LOGGER.info(f'\n{prefix} starting export with TensorRT {trt.__version__}...') + assert onnx.exists(), f'failed to export ONNX file: {onnx}' + f = file.with_suffix('.engine') # TensorRT engine file + logger = trt.Logger(trt.Logger.INFO) + if verbose: + logger.min_severity = trt.Logger.Severity.VERBOSE + + builder = trt.Builder(logger) + config = builder.create_builder_config() + config.max_workspace_size = workspace * 1 << 30 + # config.set_memory_pool_limit(trt.MemoryPoolType.WORKSPACE, workspace << 30) # fix TRT 8.4 deprecation notice + + flag = (1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)) + network = builder.create_network(flag) + parser = trt.OnnxParser(network, logger) + if not parser.parse_from_file(str(onnx)): + raise RuntimeError(f'failed to load ONNX file: {onnx}') + + inputs = [network.get_input(i) for i in range(network.num_inputs)] + outputs = [network.get_output(i) for i in range(network.num_outputs)] + LOGGER.info(f'{prefix} Network Description:') + for inp in inputs: + LOGGER.info(f'{prefix}\tinput "{inp.name}" with shape {inp.shape} and dtype {inp.dtype}') + for out in outputs: + LOGGER.info(f'{prefix}\toutput "{out.name}" with shape {out.shape} and dtype {out.dtype}') + + if dynamic: + if im.shape[0] <= 1: + LOGGER.warning(f"{prefix}WARNING: --dynamic model requires maximum --batch-size argument") + profile = builder.create_optimization_profile() + for inp in inputs: + profile.set_shape(inp.name, (1, *im.shape[1:]), (max(1, im.shape[0] // 2), *im.shape[1:]), im.shape) + config.add_optimization_profile(profile) + + LOGGER.info(f'{prefix} building FP{16 if builder.platform_has_fast_fp16 and half else 32} engine in {f}') + if builder.platform_has_fast_fp16 and half: + config.set_flag(trt.BuilderFlag.FP16) + with builder.build_engine(network, config) as engine, open(f, 'wb') as t: + t.write(engine.serialize()) + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_saved_model(model, + im, + file, + dynamic, + tf_nms=False, + agnostic_nms=False, + topk_per_class=100, + topk_all=100, + iou_thres=0.45, + conf_thres=0.25, + keras=False, + prefix=colorstr('TensorFlow SavedModel:')): + # YOLOv5 TensorFlow SavedModel export + try: + import tensorflow as tf + from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2 + + from models.tf import TFDetect, TFModel + + LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...') + f = str(file).replace('.pt', '_saved_model') + batch_size, ch, *imgsz = list(im.shape) # BCHW + + tf_model = TFModel(cfg=model.yaml, model=model, nc=model.nc, imgsz=imgsz) + im = tf.zeros((batch_size, *imgsz, ch)) # BHWC order for TensorFlow + _ = tf_model.predict(im, tf_nms, agnostic_nms, topk_per_class, topk_all, iou_thres, conf_thres) + inputs = tf.keras.Input(shape=(*imgsz, ch), batch_size=None if dynamic else batch_size) + outputs = tf_model.predict(inputs, tf_nms, agnostic_nms, topk_per_class, topk_all, iou_thres, conf_thres) + keras_model = tf.keras.Model(inputs=inputs, outputs=outputs) + keras_model.trainable = False + keras_model.summary() + if keras: + keras_model.save(f, save_format='tf') + else: + spec = tf.TensorSpec(keras_model.inputs[0].shape, keras_model.inputs[0].dtype) + m = tf.function(lambda x: keras_model(x)) # full model + m = m.get_concrete_function(spec) + frozen_func = convert_variables_to_constants_v2(m) + tfm = tf.Module() + tfm.__call__ = tf.function(lambda x: frozen_func(x)[:4] if tf_nms else frozen_func(x)[0], [spec]) + tfm.__call__(im) + tf.saved_model.save(tfm, + f, + options=tf.saved_model.SaveOptions(experimental_custom_gradients=False) + if check_version(tf.__version__, '2.6') else tf.saved_model.SaveOptions()) + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return keras_model, f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + return None, None + + +def export_pb(keras_model, file, prefix=colorstr('TensorFlow GraphDef:')): + # YOLOv5 TensorFlow GraphDef *.pb export https://github.com/leimao/Frozen_Graph_TensorFlow + try: + import tensorflow as tf + from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2 + + LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...') + f = file.with_suffix('.pb') + + m = tf.function(lambda x: keras_model(x)) # full model + m = m.get_concrete_function(tf.TensorSpec(keras_model.inputs[0].shape, keras_model.inputs[0].dtype)) + frozen_func = convert_variables_to_constants_v2(m) + frozen_func.graph.as_graph_def() + tf.io.write_graph(graph_or_graph_def=frozen_func.graph, logdir=str(f.parent), name=f.name, as_text=False) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_tflite(keras_model, im, file, int8, data, nms, agnostic_nms, prefix=colorstr('TensorFlow Lite:')): + # YOLOv5 TensorFlow Lite export + try: + import tensorflow as tf + + LOGGER.info(f'\n{prefix} starting export with tensorflow {tf.__version__}...') + batch_size, ch, *imgsz = list(im.shape) # BCHW + f = str(file).replace('.pt', '-fp16.tflite') + + converter = tf.lite.TFLiteConverter.from_keras_model(keras_model) + converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS] + converter.target_spec.supported_types = [tf.float16] + converter.optimizations = [tf.lite.Optimize.DEFAULT] + if int8: + from models.tf import representative_dataset_gen + dataset = LoadImages(check_dataset(check_yaml(data))['train'], img_size=imgsz, auto=False) + converter.representative_dataset = lambda: representative_dataset_gen(dataset, ncalib=100) + converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8] + converter.target_spec.supported_types = [] + converter.inference_input_type = tf.uint8 # or tf.int8 + converter.inference_output_type = tf.uint8 # or tf.int8 + converter.experimental_new_quantizer = True + f = str(file).replace('.pt', '-int8.tflite') + if nms or agnostic_nms: + converter.target_spec.supported_ops.append(tf.lite.OpsSet.SELECT_TF_OPS) + + tflite_model = converter.convert() + open(f, "wb").write(tflite_model) + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_edgetpu(file, prefix=colorstr('Edge TPU:')): + # YOLOv5 Edge TPU export https://coral.ai/docs/edgetpu/models-intro/ + try: + cmd = 'edgetpu_compiler --version' + help_url = 'https://coral.ai/docs/edgetpu/compiler/' + assert platform.system() == 'Linux', f'export only supported on Linux. See {help_url}' + if subprocess.run(f'{cmd} >/dev/null', shell=True).returncode != 0: + LOGGER.info(f'\n{prefix} export requires Edge TPU compiler. Attempting install from {help_url}') + sudo = subprocess.run('sudo --version >/dev/null', shell=True).returncode == 0 # sudo installed on system + for c in ( + 'curl https://packages.cloud.google.com/apt/doc/apt-key.gpg | sudo apt-key add -', + 'echo "deb https://packages.cloud.google.com/apt coral-edgetpu-stable main" | sudo tee /etc/apt/sources.list.d/coral-edgetpu.list', + 'sudo apt-get update', 'sudo apt-get install edgetpu-compiler'): + subprocess.run(c if sudo else c.replace('sudo ', ''), shell=True, check=True) + ver = subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().split()[-1] + + LOGGER.info(f'\n{prefix} starting export with Edge TPU compiler {ver}...') + f = str(file).replace('.pt', '-int8_edgetpu.tflite') # Edge TPU model + f_tfl = str(file).replace('.pt', '-int8.tflite') # TFLite model + + cmd = f"edgetpu_compiler -s -d -k 10 --out_dir {file.parent} {f_tfl}" + subprocess.run(cmd.split(), check=True) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +def export_tfjs(file, prefix=colorstr('TensorFlow.js:')): + # YOLOv5 TensorFlow.js export + try: + check_requirements(('tensorflowjs',)) + import re + + import tensorflowjs as tfjs + + LOGGER.info(f'\n{prefix} starting export with tensorflowjs {tfjs.__version__}...') + f = str(file).replace('.pt', '_web_model') # js dir + f_pb = file.with_suffix('.pb') # *.pb path + f_json = f'{f}/model.json' # *.json path + + cmd = f'tensorflowjs_converter --input_format=tf_frozen_model ' \ + f'--output_node_names=Identity,Identity_1,Identity_2,Identity_3 {f_pb} {f}' + subprocess.run(cmd.split()) + + with open(f_json) as j: + json = j.read() + with open(f_json, 'w') as j: # sort JSON Identity_* in ascending order + subst = re.sub( + r'{"outputs": {"Identity.?.?": {"name": "Identity.?.?"}, ' + r'"Identity.?.?": {"name": "Identity.?.?"}, ' + r'"Identity.?.?": {"name": "Identity.?.?"}, ' + r'"Identity.?.?": {"name": "Identity.?.?"}}}', r'{"outputs": {"Identity": {"name": "Identity"}, ' + r'"Identity_1": {"name": "Identity_1"}, ' + r'"Identity_2": {"name": "Identity_2"}, ' + r'"Identity_3": {"name": "Identity_3"}}}', json) + j.write(subst) + + LOGGER.info(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)') + return f + except Exception as e: + LOGGER.info(f'\n{prefix} export failure: {e}') + + +@smart_inference_mode() +def run( + data=ROOT / 'data/coco128.yaml', # 'dataset.yaml path' + weights=ROOT / 'yolov5s.pt', # weights path + imgsz=(640, 640), # image (height, width) + batch_size=1, # batch size + device='cpu', # cuda device, i.e. 0 or 0,1,2,3 or cpu + include=('torchscript', 'onnx'), # include formats + half=False, # FP16 half-precision export + inplace=False, # set YOLOv5 Detect() inplace=True + train=False, # model.train() mode + keras=False, # use Keras + optimize=False, # TorchScript: optimize for mobile + int8=False, # CoreML/TF INT8 quantization + dynamic=False, # ONNX/TF/TensorRT: dynamic axes + simplify=False, # ONNX: simplify model + opset=12, # ONNX: opset version + verbose=False, # TensorRT: verbose log + workspace=4, # TensorRT: workspace size (GB) + nms=False, # TF: add NMS to model + agnostic_nms=False, # TF: add agnostic NMS to model + topk_per_class=100, # TF.js NMS: topk per class to keep + topk_all=100, # TF.js NMS: topk for all classes to keep + iou_thres=0.45, # TF.js NMS: IoU threshold + conf_thres=0.25, # TF.js NMS: confidence threshold +): + t = time.time() + include = [x.lower() for x in include] # to lowercase + fmts = tuple(export_formats()['Argument'][1:]) # --include arguments + flags = [x in include for x in fmts] + assert sum(flags) == len(include), f'ERROR: Invalid --include {include}, valid --include arguments are {fmts}' + jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs = flags # export booleans + file = Path(url2file(weights) if str(weights).startswith(('http:/', 'https:/')) else weights) # PyTorch weights + + # Load PyTorch model + device = select_device(device) + if half: + assert device.type != 'cpu' or coreml, '--half only compatible with GPU export, i.e. use --device 0' + assert not dynamic, '--half not compatible with --dynamic, i.e. use either --half or --dynamic but not both' + model = attempt_load(weights, device=device, inplace=True, fuse=True) # load FP32 model + + # Checks + imgsz *= 2 if len(imgsz) == 1 else 1 # expand + if optimize: + assert device.type == 'cpu', '--optimize not compatible with cuda devices, i.e. use --device cpu' + + # Input + gs = int(max(model.stride)) # grid size (max stride) + imgsz = [check_img_size(x, gs) for x in imgsz] # verify img_size are gs-multiples + im = torch.zeros(batch_size, 3, *imgsz).to(device) # image size(1,3,320,192) BCHW iDetection + + # Update model + model.train() if train else model.eval() # training mode = no Detect() layer grid construction + for k, m in model.named_modules(): + if isinstance(m, Detect): + m.inplace = inplace + m.onnx_dynamic = dynamic + m.export = True + + for _ in range(2): + y = model(im) # dry runs + if half and not coreml: + im, model = im.half(), model.half() # to FP16 + shape = tuple((y[0] if isinstance(y, tuple) else y).shape) # model output shape + LOGGER.info(f"\n{colorstr('PyTorch:')} starting from {file} with output shape {shape} ({file_size(file):.1f} MB)") + + # Exports + f = [''] * 10 # exported filenames + warnings.filterwarnings(action='ignore', category=torch.jit.TracerWarning) # suppress TracerWarning + if jit: + f[0] = export_torchscript(model, im, file, optimize) + if engine: # TensorRT required before ONNX + f[1] = export_engine(model, im, file, train, half, dynamic, simplify, workspace, verbose) + if onnx or xml: # OpenVINO requires ONNX + f[2] = export_onnx(model, im, file, opset, train, dynamic, simplify) + if xml: # OpenVINO + f[3] = export_openvino(model, file, half) + if coreml: + _, f[4] = export_coreml(model, im, file, int8, half) + + # TensorFlow Exports + if any((saved_model, pb, tflite, edgetpu, tfjs)): + if int8 or edgetpu: # TFLite --int8 bug https://github.com/ultralytics/yolov5/issues/5707 + check_requirements(('flatbuffers==1.12',)) # required before `import tensorflow` + assert not tflite or not tfjs, 'TFLite and TF.js models must be exported separately, please pass only one type.' + model, f[5] = export_saved_model(model.cpu(), + im, + file, + dynamic, + tf_nms=nms or agnostic_nms or tfjs, + agnostic_nms=agnostic_nms or tfjs, + topk_per_class=topk_per_class, + topk_all=topk_all, + iou_thres=iou_thres, + conf_thres=conf_thres, + keras=keras) + if pb or tfjs: # pb prerequisite to tfjs + f[6] = export_pb(model, file) + if tflite or edgetpu: + f[7] = export_tflite(model, im, file, int8=int8 or edgetpu, data=data, nms=nms, agnostic_nms=agnostic_nms) + if edgetpu: + f[8] = export_edgetpu(file) + if tfjs: + f[9] = export_tfjs(file) + + # Finish + f = [str(x) for x in f if x] # filter out '' and None + if any(f): + h = '--half' if half else '' # --half FP16 inference arg + LOGGER.info(f'\nExport complete ({time.time() - t:.2f}s)' + f"\nResults saved to {colorstr('bold', file.parent.resolve())}" + f"\nDetect: python detect.py --weights {f[-1]} {h}" + f"\nValidate: python val.py --weights {f[-1]} {h}" + f"\nPyTorch Hub: model = torch.hub.load('ultralytics/yolov5', 'custom', '{f[-1]}')" + f"\nVisualize: https://netron.app") + return f # return list of exported files/dirs + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path') + parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s.pt', help='model.pt path(s)') + parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640, 640], help='image (h, w)') + parser.add_argument('--batch-size', type=int, default=1, help='batch size') + parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--half', action='store_true', help='FP16 half-precision export') + parser.add_argument('--inplace', action='store_true', help='set YOLOv5 Detect() inplace=True') + parser.add_argument('--train', action='store_true', help='model.train() mode') + parser.add_argument('--keras', action='store_true', help='TF: use Keras') + parser.add_argument('--optimize', action='store_true', help='TorchScript: optimize for mobile') + parser.add_argument('--int8', action='store_true', help='CoreML/TF INT8 quantization') + parser.add_argument('--dynamic', action='store_true', help='ONNX/TF/TensorRT: dynamic axes') + parser.add_argument('--simplify', action='store_true', help='ONNX: simplify model') + parser.add_argument('--opset', type=int, default=12, help='ONNX: opset version') + parser.add_argument('--verbose', action='store_true', help='TensorRT: verbose log') + parser.add_argument('--workspace', type=int, default=4, help='TensorRT: workspace size (GB)') + parser.add_argument('--nms', action='store_true', help='TF: add NMS to model') + parser.add_argument('--agnostic-nms', action='store_true', help='TF: add agnostic NMS to model') + parser.add_argument('--topk-per-class', type=int, default=100, help='TF.js NMS: topk per class to keep') + parser.add_argument('--topk-all', type=int, default=100, help='TF.js NMS: topk for all classes to keep') + parser.add_argument('--iou-thres', type=float, default=0.45, help='TF.js NMS: IoU threshold') + parser.add_argument('--conf-thres', type=float, default=0.25, help='TF.js NMS: confidence threshold') + parser.add_argument('--include', + nargs='+', + default=['torchscript', 'onnx'], + help='torchscript, onnx, openvino, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs') + opt = parser.parse_args() + print_args(vars(opt)) + return opt + + +def main(opt): + for opt.weights in (opt.weights if isinstance(opt.weights, list) else [opt.weights]): + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/application/yolov5_example/hubconf.py b/application/yolov5_example/hubconf.py new file mode 100644 index 00000000..011eaa57 --- /dev/null +++ b/application/yolov5_example/hubconf.py @@ -0,0 +1,160 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +PyTorch Hub models https://pytorch.org/hub/ultralytics_yolov5/ + +Usage: + import torch + model = torch.hub.load('ultralytics/yolov5', 'yolov5s') + model = torch.hub.load('ultralytics/yolov5:master', 'custom', 'path/to/yolov5s.onnx') # file from branch +""" + +import torch + + +def _create(name, pretrained=True, channels=3, classes=80, autoshape=True, verbose=True, device=None): + """Creates or loads a YOLOv5 model + + Arguments: + name (str): model name 'yolov5s' or path 'path/to/best.pt' + pretrained (bool): load pretrained weights into the model + channels (int): number of input channels + classes (int): number of model classes + autoshape (bool): apply YOLOv5 .autoshape() wrapper to model + verbose (bool): print all information to screen + device (str, torch.device, None): device to use for model parameters + + Returns: + YOLOv5 model + """ + from pathlib import Path + + from models.common import AutoShape, DetectMultiBackend + from models.experimental import attempt_load + from models.yolo import Model + from utils.downloads import attempt_download + from utils.general import LOGGER, check_requirements, intersect_dicts, logging + from utils.torch_utils import select_device + + if not verbose: + LOGGER.setLevel(logging.WARNING) + check_requirements(exclude=('tensorboard', 'thop', 'opencv-python')) + name = Path(name) + path = name.with_suffix('.pt') if name.suffix == '' and not name.is_dir() else name # checkpoint path + try: + device = select_device(device) + if pretrained and channels == 3 and classes == 80: + try: + model = DetectMultiBackend(path, device=device, fuse=autoshape) # detection model + if autoshape: + model = AutoShape(model) # for file/URI/PIL/cv2/np inputs and NMS + except Exception: + model = attempt_load(path, device=device, fuse=False) # arbitrary model + else: + cfg = list((Path(__file__).parent / 'models').rglob(f'{path.stem}.yaml'))[0] # model.yaml path + model = Model(cfg, channels, classes) # create model + if pretrained: + ckpt = torch.load(attempt_download(path), map_location=device) # load + csd = ckpt['model'].float().state_dict() # checkpoint state_dict as FP32 + csd = intersect_dicts(csd, model.state_dict(), exclude=['anchors']) # intersect + model.load_state_dict(csd, strict=False) # load + if len(ckpt['model'].names) == classes: + model.names = ckpt['model'].names # set class names attribute + if not verbose: + LOGGER.setLevel(logging.INFO) # reset to default + return model.to(device) + + except Exception as e: + help_url = 'https://github.com/ultralytics/yolov5/issues/36' + s = f'{e}. Cache may be out of date, try `force_reload=True` or see {help_url} for help.' + raise Exception(s) from e + + +def custom(path='path/to/model.pt', autoshape=True, _verbose=True, device=None): + # YOLOv5 custom or local model + return _create(path, autoshape=autoshape, verbose=_verbose, device=device) + + +def yolov5n(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-nano model https://github.com/ultralytics/yolov5 + return _create('yolov5n', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5s(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-small model https://github.com/ultralytics/yolov5 + return _create('yolov5s', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5m(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-medium model https://github.com/ultralytics/yolov5 + return _create('yolov5m', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5l(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-large model https://github.com/ultralytics/yolov5 + return _create('yolov5l', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5x(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-xlarge model https://github.com/ultralytics/yolov5 + return _create('yolov5x', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5n6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-nano-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5n6', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5s6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-small-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5s6', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5m6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-medium-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5m6', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5l6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-large-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5l6', pretrained, channels, classes, autoshape, _verbose, device) + + +def yolov5x6(pretrained=True, channels=3, classes=80, autoshape=True, _verbose=True, device=None): + # YOLOv5-xlarge-P6 model https://github.com/ultralytics/yolov5 + return _create('yolov5x6', pretrained, channels, classes, autoshape, _verbose, device) + + +if __name__ == '__main__': + import argparse + from pathlib import Path + + import numpy as np + from PIL import Image + + from utils.general import cv2, print_args + + # Argparser + parser = argparse.ArgumentParser() + parser.add_argument('--model', type=str, default='yolov5s', help='model name') + opt = parser.parse_args() + print_args(vars(opt)) + + # Model + model = _create(name=opt.model, pretrained=True, channels=3, classes=80, autoshape=True, verbose=True) + # model = custom(path='path/to/model.pt') # custom + + # Images + imgs = [ + 'data/images/zidane.jpg', # filename + Path('data/images/zidane.jpg'), # Path + 'https://ultralytics.com/images/zidane.jpg', # URI + cv2.imread('data/images/bus.jpg')[:, :, ::-1], # OpenCV + Image.open('data/images/bus.jpg'), # PIL + np.zeros((320, 640, 3))] # numpy + + # Inference + results = model(imgs, size=320) # batched inference + + # Results + results.print() + results.save() diff --git a/application/yolov5_example/models/__init__.py b/application/yolov5_example/models/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/application/yolov5_example/models/common.py b/application/yolov5_example/models/common.py new file mode 100644 index 00000000..17e40e60 --- /dev/null +++ b/application/yolov5_example/models/common.py @@ -0,0 +1,771 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Common modules +""" + +import json +import math +import platform +import warnings +from collections import OrderedDict, namedtuple +from copy import copy +from pathlib import Path + +import cv2 +import numpy as np +import pandas as pd +import requests +import torch +import torch.nn as nn +from PIL import Image +from torch.cuda import amp + +from utils.dataloaders import exif_transpose, letterbox +from utils.general import (LOGGER, ROOT, check_requirements, check_suffix, check_version, colorstr, increment_path, + make_divisible, non_max_suppression, scale_coords, xywh2xyxy, xyxy2xywh, yaml_load) +from utils.plots import Annotator, colors, save_one_box +from utils.torch_utils import copy_attr, smart_inference_mode, time_sync + + +def autopad(k, p=None): # kernel, padding + # Pad to 'same' + if p is None: + p = k // 2 if isinstance(k, int) else [x // 2 for x in k] # auto-pad + return p + + +class Conv(nn.Module): + # Standard convolution + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups + super().__init__() + self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False) + self.bn = nn.BatchNorm2d(c2) + self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity()) + + def forward(self, x): + return self.act(self.bn(self.conv(x))) + + def forward_fuse(self, x): + return self.act(self.conv(x)) + + +class DWConv(Conv): + # Depth-wise convolution class + def __init__(self, c1, c2, k=1, s=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups + super().__init__(c1, c2, k, s, g=math.gcd(c1, c2), act=act) + + +class DWConvTranspose2d(nn.ConvTranspose2d): + # Depth-wise transpose convolution class + def __init__(self, c1, c2, k=1, s=1, p1=0, p2=0): # ch_in, ch_out, kernel, stride, padding, padding_out + super().__init__(c1, c2, k, s, p1, p2, groups=math.gcd(c1, c2)) + + +class TransformerLayer(nn.Module): + # Transformer layer https://arxiv.org/abs/2010.11929 (LayerNorm layers removed for better performance) + def __init__(self, c, num_heads): + super().__init__() + self.q = nn.Linear(c, c, bias=False) + self.k = nn.Linear(c, c, bias=False) + self.v = nn.Linear(c, c, bias=False) + self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads) + self.fc1 = nn.Linear(c, c, bias=False) + self.fc2 = nn.Linear(c, c, bias=False) + + def forward(self, x): + x = self.ma(self.q(x), self.k(x), self.v(x))[0] + x + x = self.fc2(self.fc1(x)) + x + return x + + +class TransformerBlock(nn.Module): + # Vision Transformer https://arxiv.org/abs/2010.11929 + def __init__(self, c1, c2, num_heads, num_layers): + super().__init__() + self.conv = None + if c1 != c2: + self.conv = Conv(c1, c2) + self.linear = nn.Linear(c2, c2) # learnable position embedding + self.tr = nn.Sequential(*(TransformerLayer(c2, num_heads) for _ in range(num_layers))) + self.c2 = c2 + + def forward(self, x): + if self.conv is not None: + x = self.conv(x) + b, _, w, h = x.shape + p = x.flatten(2).permute(2, 0, 1) + return self.tr(p + self.linear(p)).permute(1, 2, 0).reshape(b, self.c2, w, h) + + +class Bottleneck(nn.Module): + # Standard bottleneck + def __init__(self, c1, c2, shortcut=True, g=1, e=0.5): # ch_in, ch_out, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c_, c2, 3, 1, g=g) + self.add = shortcut and c1 == c2 + + def forward(self, x): + return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) + + +class BottleneckCSP(nn.Module): + # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False) + self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False) + self.cv4 = Conv(2 * c_, c2, 1, 1) + self.bn = nn.BatchNorm2d(2 * c_) # applied to cat(cv2, cv3) + self.act = nn.SiLU() + self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n))) + + def forward(self, x): + y1 = self.cv3(self.m(self.cv1(x))) + y2 = self.cv2(x) + return self.cv4(self.act(self.bn(torch.cat((y1, y2), 1)))) + + +class CrossConv(nn.Module): + # Cross Convolution Downsample + def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False): + # ch_in, ch_out, kernel, stride, groups, expansion, shortcut + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, (1, k), (1, s)) + self.cv2 = Conv(c_, c2, (k, 1), (s, 1), g=g) + self.add = shortcut and c1 == c2 + + def forward(self, x): + return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x)) + + +class C3(nn.Module): + # CSP Bottleneck with 3 convolutions + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c1, c_, 1, 1) + self.cv3 = Conv(2 * c_, c2, 1) # optional act=FReLU(c2) + self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n))) + + def forward(self, x): + return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1)) + + +class C3x(C3): + # C3 module with cross-convolutions + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): + super().__init__(c1, c2, n, shortcut, g, e) + c_ = int(c2 * e) + self.m = nn.Sequential(*(CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n))) + + +class C3TR(C3): + # C3 module with TransformerBlock() + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): + super().__init__(c1, c2, n, shortcut, g, e) + c_ = int(c2 * e) + self.m = TransformerBlock(c_, c_, 4, n) + + +class C3SPP(C3): + # C3 module with SPP() + def __init__(self, c1, c2, k=(5, 9, 13), n=1, shortcut=True, g=1, e=0.5): + super().__init__(c1, c2, n, shortcut, g, e) + c_ = int(c2 * e) + self.m = SPP(c_, c_, k) + + +class C3Ghost(C3): + # C3 module with GhostBottleneck() + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5): + super().__init__(c1, c2, n, shortcut, g, e) + c_ = int(c2 * e) # hidden channels + self.m = nn.Sequential(*(GhostBottleneck(c_, c_) for _ in range(n))) + + +class SPP(nn.Module): + # Spatial Pyramid Pooling (SPP) layer https://arxiv.org/abs/1406.4729 + def __init__(self, c1, c2, k=(5, 9, 13)): + super().__init__() + c_ = c1 // 2 # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1) + self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k]) + + def forward(self, x): + x = self.cv1(x) + with warnings.catch_warnings(): + warnings.simplefilter('ignore') # suppress torch 1.9.0 max_pool2d() warning + return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1)) + + +class SPPF(nn.Module): + # Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher + def __init__(self, c1, c2, k=5): # equivalent to SPP(k=(5, 9, 13)) + super().__init__() + c_ = c1 // 2 # hidden channels + self.cv1 = Conv(c1, c_, 1, 1) + self.cv2 = Conv(c_ * 4, c2, 1, 1) + self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2) + + def forward(self, x): + x = self.cv1(x) + with warnings.catch_warnings(): + warnings.simplefilter('ignore') # suppress torch 1.9.0 max_pool2d() warning + y1 = self.m(x) + y2 = self.m(y1) + return self.cv2(torch.cat((x, y1, y2, self.m(y2)), 1)) + + +class Focus(nn.Module): + # Focus wh information into c-space + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True): # ch_in, ch_out, kernel, stride, padding, groups + super().__init__() + self.conv = Conv(c1 * 4, c2, k, s, p, g, act) + # self.contract = Contract(gain=2) + + def forward(self, x): # x(b,c,w,h) -> y(b,4c,w/2,h/2) + return self.conv(torch.cat((x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]), 1)) + # return self.conv(self.contract(x)) + + +class GhostConv(nn.Module): + # Ghost Convolution https://github.com/huawei-noah/ghostnet + def __init__(self, c1, c2, k=1, s=1, g=1, act=True): # ch_in, ch_out, kernel, stride, groups + super().__init__() + c_ = c2 // 2 # hidden channels + self.cv1 = Conv(c1, c_, k, s, None, g, act) + self.cv2 = Conv(c_, c_, 5, 1, None, c_, act) + + def forward(self, x): + y = self.cv1(x) + return torch.cat((y, self.cv2(y)), 1) + + +class GhostBottleneck(nn.Module): + # Ghost Bottleneck https://github.com/huawei-noah/ghostnet + def __init__(self, c1, c2, k=3, s=1): # ch_in, ch_out, kernel, stride + super().__init__() + c_ = c2 // 2 + self.conv = nn.Sequential( + GhostConv(c1, c_, 1, 1), # pw + DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(), # dw + GhostConv(c_, c2, 1, 1, act=False)) # pw-linear + self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False), Conv(c1, c2, 1, 1, + act=False)) if s == 2 else nn.Identity() + + def forward(self, x): + return self.conv(x) + self.shortcut(x) + + +class Contract(nn.Module): + # Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40) + def __init__(self, gain=2): + super().__init__() + self.gain = gain + + def forward(self, x): + b, c, h, w = x.size() # assert (h / s == 0) and (W / s == 0), 'Indivisible gain' + s = self.gain + x = x.view(b, c, h // s, s, w // s, s) # x(1,64,40,2,40,2) + x = x.permute(0, 3, 5, 1, 2, 4).contiguous() # x(1,2,2,64,40,40) + return x.view(b, c * s * s, h // s, w // s) # x(1,256,40,40) + + +class Expand(nn.Module): + # Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160) + def __init__(self, gain=2): + super().__init__() + self.gain = gain + + def forward(self, x): + b, c, h, w = x.size() # assert C / s ** 2 == 0, 'Indivisible gain' + s = self.gain + x = x.view(b, s, s, c // s ** 2, h, w) # x(1,2,2,16,80,80) + x = x.permute(0, 3, 4, 1, 5, 2).contiguous() # x(1,16,80,2,80,2) + return x.view(b, c // s ** 2, h * s, w * s) # x(1,16,160,160) + + +class Concat(nn.Module): + # Concatenate a list of tensors along dimension + def __init__(self, dimension=1): + super().__init__() + self.d = dimension + + def forward(self, x): + return torch.cat(x, self.d) + + +class DetectMultiBackend(nn.Module): + # YOLOv5 MultiBackend class for python inference on various backends + def __init__(self, weights='yolov5s.pt', device=torch.device('cpu'), dnn=False, data=None, fp16=False, fuse=True): + # Usage: + # PyTorch: weights = *.pt + # TorchScript: *.torchscript + # ONNX Runtime: *.onnx + # ONNX OpenCV DNN: *.onnx with --dnn + # OpenVINO: *.xml + # CoreML: *.mlmodel + # TensorRT: *.engine + # TensorFlow SavedModel: *_saved_model + # TensorFlow GraphDef: *.pb + # TensorFlow Lite: *.tflite + # TensorFlow Edge TPU: *_edgetpu.tflite + from models.experimental import attempt_download, attempt_load # scoped to avoid circular import + + super().__init__() + w = str(weights[0] if isinstance(weights, list) else weights) + pt, jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs = self._model_type(w) # get backend + w = attempt_download(w) # download if not local + fp16 &= pt or jit or onnx or engine # FP16 + stride = 32 # default stride + + if pt: # PyTorch + model = attempt_load(weights if isinstance(weights, list) else w, device=device, inplace=True, fuse=fuse) + stride = max(int(model.stride.max()), 32) # model stride + names = model.module.names if hasattr(model, 'module') else model.names # get class names + model.half() if fp16 else model.float() + self.model = model # explicitly assign for to(), cpu(), cuda(), half() + elif jit: # TorchScript + LOGGER.info(f'Loading {w} for TorchScript inference...') + extra_files = {'config.txt': ''} # model metadata + model = torch.jit.load(w, _extra_files=extra_files) + model.half() if fp16 else model.float() + if extra_files['config.txt']: + d = json.loads(extra_files['config.txt']) # extra_files dict + stride, names = int(d['stride']), d['names'] + elif dnn: # ONNX OpenCV DNN + LOGGER.info(f'Loading {w} for ONNX OpenCV DNN inference...') + check_requirements(('opencv-python>=4.5.4',)) + net = cv2.dnn.readNetFromONNX(w) + elif onnx: # ONNX Runtime + LOGGER.info(f'Loading {w} for ONNX Runtime inference...') + cuda = torch.cuda.is_available() and device.type != 'cpu' + check_requirements(('onnx', 'onnxruntime-gpu' if cuda else 'onnxruntime')) + import onnxruntime + providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if cuda else ['CPUExecutionProvider'] + session = onnxruntime.InferenceSession(w, providers=providers) + meta = session.get_modelmeta().custom_metadata_map # metadata + if 'stride' in meta: + stride, names = int(meta['stride']), eval(meta['names']) + elif xml: # OpenVINO + LOGGER.info(f'Loading {w} for OpenVINO inference...') + check_requirements(('openvino',)) # requires openvino-dev: https://pypi.org/project/openvino-dev/ + from openvino.runtime import Core, Layout, get_batch + ie = Core() + if not Path(w).is_file(): # if not *.xml + w = next(Path(w).glob('*.xml')) # get *.xml file from *_openvino_model dir + network = ie.read_model(model=w, weights=Path(w).with_suffix('.bin')) + if network.get_parameters()[0].get_layout().empty: + network.get_parameters()[0].set_layout(Layout("NCHW")) + batch_dim = get_batch(network) + if batch_dim.is_static: + batch_size = batch_dim.get_length() + executable_network = ie.compile_model(network, device_name="CPU") # device_name="MYRIAD" for Intel NCS2 + output_layer = next(iter(executable_network.outputs)) + meta = Path(w).with_suffix('.yaml') + if meta.exists(): + stride, names = self._load_metadata(meta) # load metadata + elif engine: # TensorRT + LOGGER.info(f'Loading {w} for TensorRT inference...') + import tensorrt as trt # https://developer.nvidia.com/nvidia-tensorrt-download + check_version(trt.__version__, '7.0.0', hard=True) # require tensorrt>=7.0.0 + if device.type == 'cpu': + device = torch.device('cuda:0') + Binding = namedtuple('Binding', ('name', 'dtype', 'shape', 'data', 'ptr')) + logger = trt.Logger(trt.Logger.INFO) + with open(w, 'rb') as f, trt.Runtime(logger) as runtime: + model = runtime.deserialize_cuda_engine(f.read()) + context = model.create_execution_context() + bindings = OrderedDict() + fp16 = False # default updated below + dynamic = False + for index in range(model.num_bindings): + name = model.get_binding_name(index) + dtype = trt.nptype(model.get_binding_dtype(index)) + if model.binding_is_input(index): + if -1 in tuple(model.get_binding_shape(index)): # dynamic + dynamic = True + context.set_binding_shape(index, tuple(model.get_profile_shape(0, index)[2])) + if dtype == np.float16: + fp16 = True + shape = tuple(context.get_binding_shape(index)) + im = torch.from_numpy(np.empty(shape, dtype=dtype)).to(device) + bindings[name] = Binding(name, dtype, shape, im, int(im.data_ptr())) + binding_addrs = OrderedDict((n, d.ptr) for n, d in bindings.items()) + batch_size = bindings['images'].shape[0] # if dynamic, this is instead max batch size + elif coreml: # CoreML + LOGGER.info(f'Loading {w} for CoreML inference...') + import coremltools as ct + model = ct.models.MLModel(w) + else: # TensorFlow (SavedModel, GraphDef, Lite, Edge TPU) + if saved_model: # SavedModel + LOGGER.info(f'Loading {w} for TensorFlow SavedModel inference...') + import tensorflow as tf + keras = False # assume TF1 saved_model + model = tf.keras.models.load_model(w) if keras else tf.saved_model.load(w) + elif pb: # GraphDef https://www.tensorflow.org/guide/migrate#a_graphpb_or_graphpbtxt + LOGGER.info(f'Loading {w} for TensorFlow GraphDef inference...') + import tensorflow as tf + + def wrap_frozen_graph(gd, inputs, outputs): + x = tf.compat.v1.wrap_function(lambda: tf.compat.v1.import_graph_def(gd, name=""), []) # wrapped + ge = x.graph.as_graph_element + return x.prune(tf.nest.map_structure(ge, inputs), tf.nest.map_structure(ge, outputs)) + + gd = tf.Graph().as_graph_def() # graph_def + with open(w, 'rb') as f: + gd.ParseFromString(f.read()) + frozen_func = wrap_frozen_graph(gd, inputs="x:0", outputs="Identity:0") + elif tflite or edgetpu: # https://www.tensorflow.org/lite/guide/python#install_tensorflow_lite_for_python + try: # https://coral.ai/docs/edgetpu/tflite-python/#update-existing-tf-lite-code-for-the-edge-tpu + from tflite_runtime.interpreter import Interpreter, load_delegate + except ImportError: + import tensorflow as tf + Interpreter, load_delegate = tf.lite.Interpreter, tf.lite.experimental.load_delegate, + if edgetpu: # Edge TPU https://coral.ai/software/#edgetpu-runtime + LOGGER.info(f'Loading {w} for TensorFlow Lite Edge TPU inference...') + delegate = { + 'Linux': 'libedgetpu.so.1', + 'Darwin': 'libedgetpu.1.dylib', + 'Windows': 'edgetpu.dll'}[platform.system()] + interpreter = Interpreter(model_path=w, experimental_delegates=[load_delegate(delegate)]) + else: # Lite + LOGGER.info(f'Loading {w} for TensorFlow Lite inference...') + interpreter = Interpreter(model_path=w) # load TFLite model + interpreter.allocate_tensors() # allocate + input_details = interpreter.get_input_details() # inputs + output_details = interpreter.get_output_details() # outputs + elif tfjs: + raise NotImplementedError('ERROR: YOLOv5 TF.js inference is not supported') + else: + raise NotImplementedError(f'ERROR: {w} is not a supported format') + + # class names + if 'names' not in locals(): + names = yaml_load(data)['names'] if data else [f'class{i}' for i in range(999)] + if names[0] == 'n01440764' and len(names) == 1000: # ImageNet + names = yaml_load(ROOT / 'data/ImageNet.yaml')['names'] # human-readable names + + self.__dict__.update(locals()) # assign all variables to self + + def forward(self, im, augment=False, visualize=False, val=False): + # YOLOv5 MultiBackend inference + b, ch, h, w = im.shape # batch, channel, height, width + if self.fp16 and im.dtype != torch.float16: + im = im.half() # to FP16 + + if self.pt: # PyTorch + y = self.model(im, augment=augment, visualize=visualize) if augment or visualize else self.model(im) + if isinstance(y, tuple): + y = y[0] + elif self.jit: # TorchScript + y = self.model(im)[0] + elif self.dnn: # ONNX OpenCV DNN + im = im.cpu().numpy() # torch to numpy + self.net.setInput(im) + y = self.net.forward() + elif self.onnx: # ONNX Runtime + im = im.cpu().numpy() # torch to numpy + y = self.session.run([self.session.get_outputs()[0].name], {self.session.get_inputs()[0].name: im})[0] + elif self.xml: # OpenVINO + im = im.cpu().numpy() # FP32 + y = self.executable_network([im])[self.output_layer] + elif self.engine: # TensorRT + if self.dynamic and im.shape != self.bindings['images'].shape: + i_in, i_out = (self.model.get_binding_index(x) for x in ('images', 'output')) + self.context.set_binding_shape(i_in, im.shape) # reshape if dynamic + self.bindings['images'] = self.bindings['images']._replace(shape=im.shape) + self.bindings['output'].data.resize_(tuple(self.context.get_binding_shape(i_out))) + s = self.bindings['images'].shape + assert im.shape == s, f"input size {im.shape} {'>' if self.dynamic else 'not equal to'} max model size {s}" + self.binding_addrs['images'] = int(im.data_ptr()) + self.context.execute_v2(list(self.binding_addrs.values())) + y = self.bindings['output'].data + elif self.coreml: # CoreML + im = im.permute(0, 2, 3, 1).cpu().numpy() # torch BCHW to numpy BHWC shape(1,320,192,3) + im = Image.fromarray((im[0] * 255).astype('uint8')) + # im = im.resize((192, 320), Image.ANTIALIAS) + y = self.model.predict({'image': im}) # coordinates are xywh normalized + if 'confidence' in y: + box = xywh2xyxy(y['coordinates'] * [[w, h, w, h]]) # xyxy pixels + conf, cls = y['confidence'].max(1), y['confidence'].argmax(1).astype(np.float) + y = np.concatenate((box, conf.reshape(-1, 1), cls.reshape(-1, 1)), 1) + else: + k = 'var_' + str(sorted(int(k.replace('var_', '')) for k in y)[-1]) # output key + y = y[k] # output + else: # TensorFlow (SavedModel, GraphDef, Lite, Edge TPU) + im = im.permute(0, 2, 3, 1).cpu().numpy() # torch BCHW to numpy BHWC shape(1,320,192,3) + if self.saved_model: # SavedModel + y = (self.model(im, training=False) if self.keras else self.model(im)).numpy() + elif self.pb: # GraphDef + y = self.frozen_func(x=self.tf.constant(im)).numpy() + else: # Lite or Edge TPU + input, output = self.input_details[0], self.output_details[0] + int8 = input['dtype'] == np.uint8 # is TFLite quantized uint8 model + if int8: + scale, zero_point = input['quantization'] + im = (im / scale + zero_point).astype(np.uint8) # de-scale + self.interpreter.set_tensor(input['index'], im) + self.interpreter.invoke() + y = self.interpreter.get_tensor(output['index']) + if int8: + scale, zero_point = output['quantization'] + y = (y.astype(np.float32) - zero_point) * scale # re-scale + y[..., :4] *= [w, h, w, h] # xywh normalized to pixels + + if isinstance(y, np.ndarray): + y = torch.tensor(y, device=self.device) + return (y, []) if val else y + + def warmup(self, imgsz=(1, 3, 640, 640)): + # Warmup model by running inference once + warmup_types = self.pt, self.jit, self.onnx, self.engine, self.saved_model, self.pb + if any(warmup_types) and self.device.type != 'cpu': + im = torch.zeros(*imgsz, dtype=torch.half if self.fp16 else torch.float, device=self.device) # input + for _ in range(2 if self.jit else 1): # + self.forward(im) # warmup + + @staticmethod + def _model_type(p='path/to/model.pt'): + # Return model type from model path, i.e. path='path/to/model.onnx' -> type=onnx + from export import export_formats + suffixes = list(export_formats().Suffix) + ['.xml'] # export suffixes + check_suffix(p, suffixes) # checks + p = Path(p).name # eliminate trailing separators + pt, jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, xml2 = (s in p for s in suffixes) + xml |= xml2 # *_openvino_model or *.xml + tflite &= not edgetpu # *.tflite + return pt, jit, onnx, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs + + @staticmethod + def _load_metadata(f='path/to/meta.yaml'): + # Load metadata from meta.yaml if it exists + d = yaml_load(f) + return d['stride'], d['names'] # assign stride, names + + +class AutoShape(nn.Module): + # YOLOv5 input-robust model wrapper for passing cv2/np/PIL/torch inputs. Includes preprocessing, inference and NMS + conf = 0.25 # NMS confidence threshold + iou = 0.45 # NMS IoU threshold + agnostic = False # NMS class-agnostic + multi_label = False # NMS multiple labels per box + classes = None # (optional list) filter by class, i.e. = [0, 15, 16] for COCO persons, cats and dogs + max_det = 1000 # maximum number of detections per image + amp = False # Automatic Mixed Precision (AMP) inference + + def __init__(self, model, verbose=True): + super().__init__() + if verbose: + LOGGER.info('Adding AutoShape... ') + copy_attr(self, model, include=('yaml', 'nc', 'hyp', 'names', 'stride', 'abc'), exclude=()) # copy attributes + self.dmb = isinstance(model, DetectMultiBackend) # DetectMultiBackend() instance + self.pt = not self.dmb or model.pt # PyTorch model + self.model = model.eval() + if self.pt: + m = self.model.model.model[-1] if self.dmb else self.model.model[-1] # Detect() + m.inplace = False # Detect.inplace=False for safe multithread inference + + def _apply(self, fn): + # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers + self = super()._apply(fn) + if self.pt: + m = self.model.model.model[-1] if self.dmb else self.model.model[-1] # Detect() + m.stride = fn(m.stride) + m.grid = list(map(fn, m.grid)) + if isinstance(m.anchor_grid, list): + m.anchor_grid = list(map(fn, m.anchor_grid)) + return self + + @smart_inference_mode() + def forward(self, imgs, size=640, augment=False, profile=False): + # Inference from various sources. For height=640, width=1280, RGB images example inputs are: + # file: imgs = 'data/images/zidane.jpg' # str or PosixPath + # URI: = 'https://ultralytics.com/images/zidane.jpg' + # OpenCV: = cv2.imread('image.jpg')[:,:,::-1] # HWC BGR to RGB x(640,1280,3) + # PIL: = Image.open('image.jpg') or ImageGrab.grab() # HWC x(640,1280,3) + # numpy: = np.zeros((640,1280,3)) # HWC + # torch: = torch.zeros(16,3,320,640) # BCHW (scaled to size=640, 0-1 values) + # multiple: = [Image.open('image1.jpg'), Image.open('image2.jpg'), ...] # list of images + + t = [time_sync()] + p = next(self.model.parameters()) if self.pt else torch.zeros(1, device=self.model.device) # for device, type + autocast = self.amp and (p.device.type != 'cpu') # Automatic Mixed Precision (AMP) inference + if isinstance(imgs, torch.Tensor): # torch + with amp.autocast(autocast): + return self.model(imgs.to(p.device).type_as(p), augment, profile) # inference + + # Pre-process + n, imgs = (len(imgs), list(imgs)) if isinstance(imgs, (list, tuple)) else (1, [imgs]) # number, list of images + shape0, shape1, files = [], [], [] # image and inference shapes, filenames + for i, im in enumerate(imgs): + f = f'image{i}' # filename + if isinstance(im, (str, Path)): # filename or uri + im, f = Image.open(requests.get(im, stream=True).raw if str(im).startswith('http') else im), im + im = np.asarray(exif_transpose(im)) + elif isinstance(im, Image.Image): # PIL Image + im, f = np.asarray(exif_transpose(im)), getattr(im, 'filename', f) or f + files.append(Path(f).with_suffix('.jpg').name) + if im.shape[0] < 5: # image in CHW + im = im.transpose((1, 2, 0)) # reverse dataloader .transpose(2, 0, 1) + im = im[..., :3] if im.ndim == 3 else np.tile(im[..., None], 3) # enforce 3ch input + s = im.shape[:2] # HWC + shape0.append(s) # image shape + g = (size / max(s)) # gain + shape1.append([y * g for y in s]) + imgs[i] = im if im.data.contiguous else np.ascontiguousarray(im) # update + shape1 = [make_divisible(x, self.stride) if self.pt else size for x in np.array(shape1).max(0)] # inf shape + x = [letterbox(im, shape1, auto=False)[0] for im in imgs] # pad + x = np.ascontiguousarray(np.array(x).transpose((0, 3, 1, 2))) # stack and BHWC to BCHW + x = torch.from_numpy(x).to(p.device).type_as(p) / 255 # uint8 to fp16/32 + t.append(time_sync()) + + with amp.autocast(autocast): + # Inference + y = self.model(x, augment, profile) # forward + t.append(time_sync()) + + # Post-process + y = non_max_suppression(y if self.dmb else y[0], + self.conf, + self.iou, + self.classes, + self.agnostic, + self.multi_label, + max_det=self.max_det) # NMS + for i in range(n): + scale_coords(shape1, y[i][:, :4], shape0[i]) + + t.append(time_sync()) + return Detections(imgs, y, files, t, self.names, x.shape) + + +class Detections: + # YOLOv5 detections class for inference results + def __init__(self, imgs, pred, files, times=(0, 0, 0, 0), names=None, shape=None): + super().__init__() + d = pred[0].device # device + gn = [torch.tensor([*(im.shape[i] for i in [1, 0, 1, 0]), 1, 1], device=d) for im in imgs] # normalizations + self.imgs = imgs # list of images as numpy arrays + self.pred = pred # list of tensors pred[0] = (xyxy, conf, cls) + self.names = names # class names + self.files = files # image filenames + self.times = times # profiling times + self.xyxy = pred # xyxy pixels + self.xywh = [xyxy2xywh(x) for x in pred] # xywh pixels + self.xyxyn = [x / g for x, g in zip(self.xyxy, gn)] # xyxy normalized + self.xywhn = [x / g for x, g in zip(self.xywh, gn)] # xywh normalized + self.n = len(self.pred) # number of images (batch size) + self.t = tuple((times[i + 1] - times[i]) * 1000 / self.n for i in range(3)) # timestamps (ms) + self.s = shape # inference BCHW shape + + def display(self, pprint=False, show=False, save=False, crop=False, render=False, labels=True, save_dir=Path('')): + crops = [] + for i, (im, pred) in enumerate(zip(self.imgs, self.pred)): + s = f'image {i + 1}/{len(self.pred)}: {im.shape[0]}x{im.shape[1]} ' # string + if pred.shape[0]: + for c in pred[:, -1].unique(): + n = (pred[:, -1] == c).sum() # detections per class + s += f"{n} {self.names[int(c)]}{'s' * (n > 1)}, " # add to string + if show or save or render or crop: + annotator = Annotator(im, example=str(self.names)) + for *box, conf, cls in reversed(pred): # xyxy, confidence, class + label = f'{self.names[int(cls)]} {conf:.2f}' + if crop: + file = save_dir / 'crops' / self.names[int(cls)] / self.files[i] if save else None + crops.append({ + 'box': box, + 'conf': conf, + 'cls': cls, + 'label': label, + 'im': save_one_box(box, im, file=file, save=save)}) + else: # all others + annotator.box_label(box, label if labels else '', color=colors(cls)) + im = annotator.im + else: + s += '(no detections)' + + im = Image.fromarray(im.astype(np.uint8)) if isinstance(im, np.ndarray) else im # from np + if pprint: + print(s.rstrip(', ')) + if show: + im.show(self.files[i]) # show + if save: + f = self.files[i] + im.save(save_dir / f) # save + if i == self.n - 1: + LOGGER.info(f"Saved {self.n} image{'s' * (self.n > 1)} to {colorstr('bold', save_dir)}") + if render: + self.imgs[i] = np.asarray(im) + if crop: + if save: + LOGGER.info(f'Saved results to {save_dir}\n') + return crops + + def print(self): + self.display(pprint=True) # print results + print(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {tuple(self.s)}' % self.t) + + def show(self, labels=True): + self.display(show=True, labels=labels) # show results + + def save(self, labels=True, save_dir='runs/detect/exp'): + save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True) # increment save_dir + self.display(save=True, labels=labels, save_dir=save_dir) # save results + + def crop(self, save=True, save_dir='runs/detect/exp'): + save_dir = increment_path(save_dir, exist_ok=save_dir != 'runs/detect/exp', mkdir=True) if save else None + return self.display(crop=True, save=save, save_dir=save_dir) # crop results + + def render(self, labels=True): + self.display(render=True, labels=labels) # render results + return self.imgs + + def pandas(self): + # return detections as pandas DataFrames, i.e. print(results.pandas().xyxy[0]) + new = copy(self) # return copy + ca = 'xmin', 'ymin', 'xmax', 'ymax', 'confidence', 'class', 'name' # xyxy columns + cb = 'xcenter', 'ycenter', 'width', 'height', 'confidence', 'class', 'name' # xywh columns + for k, c in zip(['xyxy', 'xyxyn', 'xywh', 'xywhn'], [ca, ca, cb, cb]): + a = [[x[:5] + [int(x[5]), self.names[int(x[5])]] for x in x.tolist()] for x in getattr(self, k)] # update + setattr(new, k, [pd.DataFrame(x, columns=c) for x in a]) + return new + + def tolist(self): + # return a list of Detections objects, i.e. 'for result in results.tolist():' + r = range(self.n) # iterable + x = [Detections([self.imgs[i]], [self.pred[i]], [self.files[i]], self.times, self.names, self.s) for i in r] + # for d in x: + # for k in ['imgs', 'pred', 'xyxy', 'xyxyn', 'xywh', 'xywhn']: + # setattr(d, k, getattr(d, k)[0]) # pop out of list + return x + + def __len__(self): + return self.n # override len(results) + + def __str__(self): + self.print() # override print(results) + return '' + + +class Classify(nn.Module): + # Classification head, i.e. x(b,c1,20,20) to x(b,c2) + def __init__(self, c1, c2, k=1, s=1, p=None, g=1): # ch_in, ch_out, kernel, stride, padding, groups + super().__init__() + c_ = 1280 # efficientnet_b0 size + self.conv = Conv(c1, c_, k, s, autopad(k, p), g) + self.pool = nn.AdaptiveAvgPool2d(1) # to x(b,c_,1,1) + self.drop = nn.Dropout(p=0.0, inplace=True) + self.linear = nn.Linear(c_, c2) # to x(b,c2) + + def forward(self, x): + if isinstance(x, list): + x = torch.cat(x, 1) + return self.linear(self.drop(self.pool(self.conv(x)).flatten(1))) diff --git a/application/yolov5_example/models/experimental.py b/application/yolov5_example/models/experimental.py new file mode 100644 index 00000000..cb32d01b --- /dev/null +++ b/application/yolov5_example/models/experimental.py @@ -0,0 +1,107 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Experimental modules +""" +import math + +import numpy as np +import torch +import torch.nn as nn + +from models.common import Conv +from utils.downloads import attempt_download + + +class Sum(nn.Module): + # Weighted sum of 2 or more layers https://arxiv.org/abs/1911.09070 + def __init__(self, n, weight=False): # n: number of inputs + super().__init__() + self.weight = weight # apply weights boolean + self.iter = range(n - 1) # iter object + if weight: + self.w = nn.Parameter(-torch.arange(1.0, n) / 2, requires_grad=True) # layer weights + + def forward(self, x): + y = x[0] # no weight + if self.weight: + w = torch.sigmoid(self.w) * 2 + for i in self.iter: + y = y + x[i + 1] * w[i] + else: + for i in self.iter: + y = y + x[i + 1] + return y + + +class MixConv2d(nn.Module): + # Mixed Depth-wise Conv https://arxiv.org/abs/1907.09595 + def __init__(self, c1, c2, k=(1, 3), s=1, equal_ch=True): # ch_in, ch_out, kernel, stride, ch_strategy + super().__init__() + n = len(k) # number of convolutions + if equal_ch: # equal c_ per group + i = torch.linspace(0, n - 1E-6, c2).floor() # c2 indices + c_ = [(i == g).sum() for g in range(n)] # intermediate channels + else: # equal weight.numel() per group + b = [c2] + [0] * n + a = np.eye(n + 1, n, k=-1) + a -= np.roll(a, 1, axis=1) + a *= np.array(k) ** 2 + a[0] = 1 + c_ = np.linalg.lstsq(a, b, rcond=None)[0].round() # solve for equal weight indices, ax = b + + self.m = nn.ModuleList([ + nn.Conv2d(c1, int(c_), k, s, k // 2, groups=math.gcd(c1, int(c_)), bias=False) for k, c_ in zip(k, c_)]) + self.bn = nn.BatchNorm2d(c2) + self.act = nn.SiLU() + + def forward(self, x): + return self.act(self.bn(torch.cat([m(x) for m in self.m], 1))) + + +class Ensemble(nn.ModuleList): + # Ensemble of models + def __init__(self): + super().__init__() + + def forward(self, x, augment=False, profile=False, visualize=False): + y = [module(x, augment, profile, visualize)[0] for module in self] + # y = torch.stack(y).max(0)[0] # max ensemble + # y = torch.stack(y).mean(0) # mean ensemble + y = torch.cat(y, 1) # nms ensemble + return y, None # inference, train output + + +def attempt_load(weights, device=None, inplace=True, fuse=True): + # Loads an ensemble of models weights=[a,b,c] or a single model weights=[a] or weights=a + from models.yolo import Detect, Model + + model = Ensemble() + for w in weights if isinstance(weights, list) else [weights]: + ckpt = torch.load(attempt_download(w), map_location='cpu') # load + ckpt = (ckpt.get('ema') or ckpt['model']).to(device).float() # FP32 model + if not hasattr(ckpt, 'stride'): + ckpt.stride = torch.tensor([32.]) # compatibility update for ResNet etc. + model.append(ckpt.fuse().eval() if fuse and hasattr(ckpt, 'fuse') else ckpt.eval()) # model in eval mode + + # Compatibility updates + for m in model.modules(): + t = type(m) + if t in (nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model): + m.inplace = inplace # torch 1.7.0 compatibility + if t is Detect and not isinstance(m.anchor_grid, list): + delattr(m, 'anchor_grid') + setattr(m, 'anchor_grid', [torch.zeros(1)] * m.nl) + elif t is nn.Upsample and not hasattr(m, 'recompute_scale_factor'): + m.recompute_scale_factor = None # torch 1.11.0 compatibility + + # Return model + if len(model) == 1: + return model[-1] + + # Return detection ensemble + print(f'Ensemble created with {weights}\n') + for k in 'names', 'nc', 'yaml': + setattr(model, k, getattr(model[0], k)) + model.stride = model[torch.argmax(torch.tensor([m.stride.max() for m in model])).int()].stride # max stride + assert all(model[0].nc == m.nc for m in model), f'Models have different class counts: {[m.nc for m in model]}' + return model diff --git a/application/yolov5_example/models/hub/anchors.yaml b/application/yolov5_example/models/hub/anchors.yaml new file mode 100644 index 00000000..e4d7beb0 --- /dev/null +++ b/application/yolov5_example/models/hub/anchors.yaml @@ -0,0 +1,59 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Default anchors for COCO data + + +# P5 ------------------------------------------------------------------------------------------------------------------- +# P5-640: +anchors_p5_640: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + + +# P6 ------------------------------------------------------------------------------------------------------------------- +# P6-640: thr=0.25: 0.9964 BPR, 5.54 anchors past thr, n=12, img_size=640, metric_all=0.281/0.716-mean/best, past_thr=0.469-mean: 9,11, 21,19, 17,41, 43,32, 39,70, 86,64, 65,131, 134,130, 120,265, 282,180, 247,354, 512,387 +anchors_p6_640: + - [9,11, 21,19, 17,41] # P3/8 + - [43,32, 39,70, 86,64] # P4/16 + - [65,131, 134,130, 120,265] # P5/32 + - [282,180, 247,354, 512,387] # P6/64 + +# P6-1280: thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1280, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 19,27, 44,40, 38,94, 96,68, 86,152, 180,137, 140,301, 303,264, 238,542, 436,615, 739,380, 925,792 +anchors_p6_1280: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# P6-1920: thr=0.25: 0.9950 BPR, 5.55 anchors past thr, n=12, img_size=1920, metric_all=0.281/0.714-mean/best, past_thr=0.468-mean: 28,41, 67,59, 57,141, 144,103, 129,227, 270,205, 209,452, 455,396, 358,812, 653,922, 1109,570, 1387,1187 +anchors_p6_1920: + - [28,41, 67,59, 57,141] # P3/8 + - [144,103, 129,227, 270,205] # P4/16 + - [209,452, 455,396, 358,812] # P5/32 + - [653,922, 1109,570, 1387,1187] # P6/64 + + +# P7 ------------------------------------------------------------------------------------------------------------------- +# P7-640: thr=0.25: 0.9962 BPR, 6.76 anchors past thr, n=15, img_size=640, metric_all=0.275/0.733-mean/best, past_thr=0.466-mean: 11,11, 13,30, 29,20, 30,46, 61,38, 39,92, 78,80, 146,66, 79,163, 149,150, 321,143, 157,303, 257,402, 359,290, 524,372 +anchors_p7_640: + - [11,11, 13,30, 29,20] # P3/8 + - [30,46, 61,38, 39,92] # P4/16 + - [78,80, 146,66, 79,163] # P5/32 + - [149,150, 321,143, 157,303] # P6/64 + - [257,402, 359,290, 524,372] # P7/128 + +# P7-1280: thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1280, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 19,22, 54,36, 32,77, 70,83, 138,71, 75,173, 165,159, 148,334, 375,151, 334,317, 251,626, 499,474, 750,326, 534,814, 1079,818 +anchors_p7_1280: + - [19,22, 54,36, 32,77] # P3/8 + - [70,83, 138,71, 75,173] # P4/16 + - [165,159, 148,334, 375,151] # P5/32 + - [334,317, 251,626, 499,474] # P6/64 + - [750,326, 534,814, 1079,818] # P7/128 + +# P7-1920: thr=0.25: 0.9968 BPR, 6.71 anchors past thr, n=15, img_size=1920, metric_all=0.273/0.732-mean/best, past_thr=0.463-mean: 29,34, 81,55, 47,115, 105,124, 207,107, 113,259, 247,238, 222,500, 563,227, 501,476, 376,939, 749,711, 1126,489, 801,1222, 1618,1227 +anchors_p7_1920: + - [29,34, 81,55, 47,115] # P3/8 + - [105,124, 207,107, 113,259] # P4/16 + - [247,238, 222,500, 563,227] # P5/32 + - [501,476, 376,939, 749,711] # P6/64 + - [1126,489, 801,1222, 1618,1227] # P7/128 diff --git a/application/yolov5_example/models/hub/yolov3-spp.yaml b/application/yolov5_example/models/hub/yolov3-spp.yaml new file mode 100644 index 00000000..c6698215 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov3-spp.yaml @@ -0,0 +1,51 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# darknet53 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [32, 3, 1]], # 0 + [-1, 1, Conv, [64, 3, 2]], # 1-P1/2 + [-1, 1, Bottleneck, [64]], + [-1, 1, Conv, [128, 3, 2]], # 3-P2/4 + [-1, 2, Bottleneck, [128]], + [-1, 1, Conv, [256, 3, 2]], # 5-P3/8 + [-1, 8, Bottleneck, [256]], + [-1, 1, Conv, [512, 3, 2]], # 7-P4/16 + [-1, 8, Bottleneck, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32 + [-1, 4, Bottleneck, [1024]], # 10 + ] + +# YOLOv3-SPP head +head: + [[-1, 1, Bottleneck, [1024, False]], + [-1, 1, SPP, [512, [5, 9, 13]]], + [-1, 1, Conv, [1024, 3, 1]], + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, Conv, [1024, 3, 1]], # 15 (P5/32-large) + + [-2, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, Conv, [512, 3, 1]], # 22 (P4/16-medium) + + [-2, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P3 + [-1, 1, Bottleneck, [256, False]], + [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small) + + [[27, 22, 15], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/hub/yolov3-tiny.yaml b/application/yolov5_example/models/hub/yolov3-tiny.yaml new file mode 100644 index 00000000..b28b4431 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov3-tiny.yaml @@ -0,0 +1,41 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,14, 23,27, 37,58] # P4/16 + - [81,82, 135,169, 344,319] # P5/32 + +# YOLOv3-tiny backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [16, 3, 1]], # 0 + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 1-P1/2 + [-1, 1, Conv, [32, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 3-P2/4 + [-1, 1, Conv, [64, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 5-P3/8 + [-1, 1, Conv, [128, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 7-P4/16 + [-1, 1, Conv, [256, 3, 1]], + [-1, 1, nn.MaxPool2d, [2, 2, 0]], # 9-P5/32 + [-1, 1, Conv, [512, 3, 1]], + [-1, 1, nn.ZeroPad2d, [[0, 1, 0, 1]]], # 11 + [-1, 1, nn.MaxPool2d, [2, 1, 0]], # 12 + ] + +# YOLOv3-tiny head +head: + [[-1, 1, Conv, [1024, 3, 1]], + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, Conv, [512, 3, 1]], # 15 (P5/32-large) + + [-2, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Conv, [256, 3, 1]], # 19 (P4/16-medium) + + [[19, 15], 1, Detect, [nc, anchors]], # Detect(P4, P5) + ] diff --git a/application/yolov5_example/models/hub/yolov3.yaml b/application/yolov5_example/models/hub/yolov3.yaml new file mode 100644 index 00000000..d1ef9129 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov3.yaml @@ -0,0 +1,51 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# darknet53 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [32, 3, 1]], # 0 + [-1, 1, Conv, [64, 3, 2]], # 1-P1/2 + [-1, 1, Bottleneck, [64]], + [-1, 1, Conv, [128, 3, 2]], # 3-P2/4 + [-1, 2, Bottleneck, [128]], + [-1, 1, Conv, [256, 3, 2]], # 5-P3/8 + [-1, 8, Bottleneck, [256]], + [-1, 1, Conv, [512, 3, 2]], # 7-P4/16 + [-1, 8, Bottleneck, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P5/32 + [-1, 4, Bottleneck, [1024]], # 10 + ] + +# YOLOv3 head +head: + [[-1, 1, Bottleneck, [1024, False]], + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, Conv, [1024, 3, 1]], + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, Conv, [1024, 3, 1]], # 15 (P5/32-large) + + [-2, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Bottleneck, [512, False]], + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, Conv, [512, 3, 1]], # 22 (P4/16-medium) + + [-2, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P3 + [-1, 1, Bottleneck, [256, False]], + [-1, 2, Bottleneck, [256, False]], # 27 (P3/8-small) + + [[27, 22, 15], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/hub/yolov5-bifpn.yaml b/application/yolov5_example/models/hub/yolov5-bifpn.yaml new file mode 100644 index 00000000..504815f5 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5-bifpn.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 BiFPN head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14, 6], 1, Concat, [1]], # cat P4 <--- BiFPN change + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/hub/yolov5-fpn.yaml b/application/yolov5_example/models/hub/yolov5-fpn.yaml new file mode 100644 index 00000000..a23e9c6f --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5-fpn.yaml @@ -0,0 +1,42 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 FPN head +head: + [[-1, 3, C3, [1024, False]], # 10 (P5/32-large) + + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 1, Conv, [512, 1, 1]], + [-1, 3, C3, [512, False]], # 14 (P4/16-medium) + + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 1, Conv, [256, 1, 1]], + [-1, 3, C3, [256, False]], # 18 (P3/8-small) + + [[18, 14, 10], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/hub/yolov5-p2.yaml b/application/yolov5_example/models/hub/yolov5-p2.yaml new file mode 100644 index 00000000..554117dd --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5-p2.yaml @@ -0,0 +1,54 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: 3 # AutoAnchor evolves 3 anchors per P output layer + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head with (P2, P3, P4, P5) outputs +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [128, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 2], 1, Concat, [1]], # cat backbone P2 + [-1, 1, C3, [128, False]], # 21 (P2/4-xsmall) + + [-1, 1, Conv, [128, 3, 2]], + [[-1, 18], 1, Concat, [1]], # cat head P3 + [-1, 3, C3, [256, False]], # 24 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 27 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 30 (P5/32-large) + + [[21, 24, 27, 30], 1, Detect, [nc, anchors]], # Detect(P2, P3, P4, P5) + ] diff --git a/application/yolov5_example/models/hub/yolov5-p34.yaml b/application/yolov5_example/models/hub/yolov5-p34.yaml new file mode 100644 index 00000000..dbf0f850 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5-p34.yaml @@ -0,0 +1,41 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: 3 # AutoAnchor evolves 3 anchors per P output layer + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [ [ -1, 1, Conv, [ 64, 6, 2, 2 ] ], # 0-P1/2 + [ -1, 1, Conv, [ 128, 3, 2 ] ], # 1-P2/4 + [ -1, 3, C3, [ 128 ] ], + [ -1, 1, Conv, [ 256, 3, 2 ] ], # 3-P3/8 + [ -1, 6, C3, [ 256 ] ], + [ -1, 1, Conv, [ 512, 3, 2 ] ], # 5-P4/16 + [ -1, 9, C3, [ 512 ] ], + [ -1, 1, Conv, [ 1024, 3, 2 ] ], # 7-P5/32 + [ -1, 3, C3, [ 1024 ] ], + [ -1, 1, SPPF, [ 1024, 5 ] ], # 9 + ] + +# YOLOv5 v6.0 head with (P3, P4) outputs +head: + [ [ -1, 1, Conv, [ 512, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 6 ], 1, Concat, [ 1 ] ], # cat backbone P4 + [ -1, 3, C3, [ 512, False ] ], # 13 + + [ -1, 1, Conv, [ 256, 1, 1 ] ], + [ -1, 1, nn.Upsample, [ None, 2, 'nearest' ] ], + [ [ -1, 4 ], 1, Concat, [ 1 ] ], # cat backbone P3 + [ -1, 3, C3, [ 256, False ] ], # 17 (P3/8-small) + + [ -1, 1, Conv, [ 256, 3, 2 ] ], + [ [ -1, 14 ], 1, Concat, [ 1 ] ], # cat head P4 + [ -1, 3, C3, [ 512, False ] ], # 20 (P4/16-medium) + + [ [ 17, 20 ], 1, Detect, [ nc, anchors ] ], # Detect(P3, P4) + ] diff --git a/application/yolov5_example/models/hub/yolov5-p6.yaml b/application/yolov5_example/models/hub/yolov5-p6.yaml new file mode 100644 index 00000000..a17202f2 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5-p6.yaml @@ -0,0 +1,56 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: 3 # AutoAnchor evolves 3 anchors per P output layer + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head with (P3, P4, P5, P6) outputs +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/application/yolov5_example/models/hub/yolov5-p7.yaml b/application/yolov5_example/models/hub/yolov5-p7.yaml new file mode 100644 index 00000000..edd7d13a --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5-p7.yaml @@ -0,0 +1,67 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: 3 # AutoAnchor evolves 3 anchors per P output layer + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, Conv, [1280, 3, 2]], # 11-P7/128 + [-1, 3, C3, [1280]], + [-1, 1, SPPF, [1280, 5]], # 13 + ] + +# YOLOv5 v6.0 head with (P3, P4, P5, P6, P7) outputs +head: + [[-1, 1, Conv, [1024, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 10], 1, Concat, [1]], # cat backbone P6 + [-1, 3, C3, [1024, False]], # 17 + + [-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 21 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 25 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 29 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 26], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 32 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 22], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 35 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 18], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 38 (P6/64-xlarge) + + [-1, 1, Conv, [1024, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P7 + [-1, 3, C3, [1280, False]], # 41 (P7/128-xxlarge) + + [[29, 32, 35, 38, 41], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6, P7) + ] diff --git a/application/yolov5_example/models/hub/yolov5-panet.yaml b/application/yolov5_example/models/hub/yolov5-panet.yaml new file mode 100644 index 00000000..ccfbf900 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5-panet.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 PANet head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/hub/yolov5l6.yaml b/application/yolov5_example/models/hub/yolov5l6.yaml new file mode 100644 index 00000000..632c2cb6 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5l6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/application/yolov5_example/models/hub/yolov5m6.yaml b/application/yolov5_example/models/hub/yolov5m6.yaml new file mode 100644 index 00000000..ecc53fd6 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5m6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.67 # model depth multiple +width_multiple: 0.75 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/application/yolov5_example/models/hub/yolov5n6.yaml b/application/yolov5_example/models/hub/yolov5n6.yaml new file mode 100644 index 00000000..0c0c71d3 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5n6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.25 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/application/yolov5_example/models/hub/yolov5s-ghost.yaml b/application/yolov5_example/models/hub/yolov5s-ghost.yaml new file mode 100644 index 00000000..ff9519c3 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5s-ghost.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, GhostConv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3Ghost, [128]], + [-1, 1, GhostConv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3Ghost, [256]], + [-1, 1, GhostConv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3Ghost, [512]], + [-1, 1, GhostConv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3Ghost, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, GhostConv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3Ghost, [512, False]], # 13 + + [-1, 1, GhostConv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3Ghost, [256, False]], # 17 (P3/8-small) + + [-1, 1, GhostConv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3Ghost, [512, False]], # 20 (P4/16-medium) + + [-1, 1, GhostConv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3Ghost, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/hub/yolov5s-transformer.yaml b/application/yolov5_example/models/hub/yolov5s-transformer.yaml new file mode 100644 index 00000000..100d7c44 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5s-transformer.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3TR, [1024]], # 9 <--- C3TR() Transformer module + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/hub/yolov5s6.yaml b/application/yolov5_example/models/hub/yolov5s6.yaml new file mode 100644 index 00000000..a28fb559 --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5s6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/application/yolov5_example/models/hub/yolov5x6.yaml b/application/yolov5_example/models/hub/yolov5x6.yaml new file mode 100644 index 00000000..ba795c4a --- /dev/null +++ b/application/yolov5_example/models/hub/yolov5x6.yaml @@ -0,0 +1,60 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.33 # model depth multiple +width_multiple: 1.25 # layer channel multiple +anchors: + - [19,27, 44,40, 38,94] # P3/8 + - [96,68, 86,152, 180,137] # P4/16 + - [140,301, 303,264, 238,542] # P5/32 + - [436,615, 739,380, 925,792] # P6/64 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [768, 3, 2]], # 7-P5/32 + [-1, 3, C3, [768]], + [-1, 1, Conv, [1024, 3, 2]], # 9-P6/64 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 11 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [768, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 8], 1, Concat, [1]], # cat backbone P5 + [-1, 3, C3, [768, False]], # 15 + + [-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 19 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 23 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 20], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 26 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 16], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [768, False]], # 29 (P5/32-large) + + [-1, 1, Conv, [768, 3, 2]], + [[-1, 12], 1, Concat, [1]], # cat head P6 + [-1, 3, C3, [1024, False]], # 32 (P6/64-xlarge) + + [[23, 26, 29, 32], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5, P6) + ] diff --git a/application/yolov5_example/models/tf.py b/application/yolov5_example/models/tf.py new file mode 100644 index 00000000..b0d98cc2 --- /dev/null +++ b/application/yolov5_example/models/tf.py @@ -0,0 +1,574 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +TensorFlow, Keras and TFLite versions of YOLOv5 +Authored by https://github.com/zldrobit in PR https://github.com/ultralytics/yolov5/pull/1127 + +Usage: + $ python models/tf.py --weights yolov5s.pt + +Export: + $ python path/to/export.py --weights yolov5s.pt --include saved_model pb tflite tfjs +""" + +import argparse +import sys +from copy import deepcopy +from pathlib import Path + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +# ROOT = ROOT.relative_to(Path.cwd()) # relative + +import numpy as np +import tensorflow as tf +import torch +import torch.nn as nn +from tensorflow import keras + +from models.common import (C3, SPP, SPPF, Bottleneck, BottleneckCSP, C3x, Concat, Conv, CrossConv, DWConv, + DWConvTranspose2d, Focus, autopad) +from models.experimental import MixConv2d, attempt_load +from models.yolo import Detect +from utils.activations import SiLU +from utils.general import LOGGER, make_divisible, print_args + + +class TFBN(keras.layers.Layer): + # TensorFlow BatchNormalization wrapper + def __init__(self, w=None): + super().__init__() + self.bn = keras.layers.BatchNormalization( + beta_initializer=keras.initializers.Constant(w.bias.numpy()), + gamma_initializer=keras.initializers.Constant(w.weight.numpy()), + moving_mean_initializer=keras.initializers.Constant(w.running_mean.numpy()), + moving_variance_initializer=keras.initializers.Constant(w.running_var.numpy()), + epsilon=w.eps) + + def call(self, inputs): + return self.bn(inputs) + + +class TFPad(keras.layers.Layer): + # Pad inputs in spatial dimensions 1 and 2 + def __init__(self, pad): + super().__init__() + if isinstance(pad, int): + self.pad = tf.constant([[0, 0], [pad, pad], [pad, pad], [0, 0]]) + else: # tuple/list + self.pad = tf.constant([[0, 0], [pad[0], pad[0]], [pad[1], pad[1]], [0, 0]]) + + def call(self, inputs): + return tf.pad(inputs, self.pad, mode='constant', constant_values=0) + + +class TFConv(keras.layers.Layer): + # Standard convolution + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True, w=None): + # ch_in, ch_out, weights, kernel, stride, padding, groups + super().__init__() + assert g == 1, "TF v2.2 Conv2D does not support 'groups' argument" + # TensorFlow convolution padding is inconsistent with PyTorch (e.g. k=3 s=2 'SAME' padding) + # see https://stackoverflow.com/questions/52975843/comparing-conv2d-with-padding-between-tensorflow-and-pytorch + conv = keras.layers.Conv2D( + filters=c2, + kernel_size=k, + strides=s, + padding='SAME' if s == 1 else 'VALID', + use_bias=not hasattr(w, 'bn'), + kernel_initializer=keras.initializers.Constant(w.conv.weight.permute(2, 3, 1, 0).numpy()), + bias_initializer='zeros' if hasattr(w, 'bn') else keras.initializers.Constant(w.conv.bias.numpy())) + self.conv = conv if s == 1 else keras.Sequential([TFPad(autopad(k, p)), conv]) + self.bn = TFBN(w.bn) if hasattr(w, 'bn') else tf.identity + self.act = activations(w.act) if act else tf.identity + + def call(self, inputs): + return self.act(self.bn(self.conv(inputs))) + + +class TFDWConv(keras.layers.Layer): + # Depthwise convolution + def __init__(self, c1, c2, k=1, s=1, p=None, act=True, w=None): + # ch_in, ch_out, weights, kernel, stride, padding, groups + super().__init__() + assert c2 % c1 == 0, f'TFDWConv() output={c2} must be a multiple of input={c1} channels' + conv = keras.layers.DepthwiseConv2D( + kernel_size=k, + depth_multiplier=c2 // c1, + strides=s, + padding='SAME' if s == 1 else 'VALID', + use_bias=not hasattr(w, 'bn'), + depthwise_initializer=keras.initializers.Constant(w.conv.weight.permute(2, 3, 1, 0).numpy()), + bias_initializer='zeros' if hasattr(w, 'bn') else keras.initializers.Constant(w.conv.bias.numpy())) + self.conv = conv if s == 1 else keras.Sequential([TFPad(autopad(k, p)), conv]) + self.bn = TFBN(w.bn) if hasattr(w, 'bn') else tf.identity + self.act = activations(w.act) if act else tf.identity + + def call(self, inputs): + return self.act(self.bn(self.conv(inputs))) + + +class TFDWConvTranspose2d(keras.layers.Layer): + # Depthwise ConvTranspose2d + def __init__(self, c1, c2, k=1, s=1, p1=0, p2=0, w=None): + # ch_in, ch_out, weights, kernel, stride, padding, groups + super().__init__() + assert c1 == c2, f'TFDWConv() output={c2} must be equal to input={c1} channels' + assert k == 4 and p1 == 1, 'TFDWConv() only valid for k=4 and p1=1' + weight, bias = w.weight.permute(2, 3, 1, 0).numpy(), w.bias.numpy() + self.c1 = c1 + self.conv = [ + keras.layers.Conv2DTranspose(filters=1, + kernel_size=k, + strides=s, + padding='VALID', + output_padding=p2, + use_bias=True, + kernel_initializer=keras.initializers.Constant(weight[..., i:i + 1]), + bias_initializer=keras.initializers.Constant(bias[i])) for i in range(c1)] + + def call(self, inputs): + return tf.concat([m(x) for m, x in zip(self.conv, tf.split(inputs, self.c1, 3))], 3)[:, 1:-1, 1:-1] + + +class TFFocus(keras.layers.Layer): + # Focus wh information into c-space + def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True, w=None): + # ch_in, ch_out, kernel, stride, padding, groups + super().__init__() + self.conv = TFConv(c1 * 4, c2, k, s, p, g, act, w.conv) + + def call(self, inputs): # x(b,w,h,c) -> y(b,w/2,h/2,4c) + # inputs = inputs / 255 # normalize 0-255 to 0-1 + inputs = [inputs[:, ::2, ::2, :], inputs[:, 1::2, ::2, :], inputs[:, ::2, 1::2, :], inputs[:, 1::2, 1::2, :]] + return self.conv(tf.concat(inputs, 3)) + + +class TFBottleneck(keras.layers.Layer): + # Standard bottleneck + def __init__(self, c1, c2, shortcut=True, g=1, e=0.5, w=None): # ch_in, ch_out, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c_, c2, 3, 1, g=g, w=w.cv2) + self.add = shortcut and c1 == c2 + + def call(self, inputs): + return inputs + self.cv2(self.cv1(inputs)) if self.add else self.cv2(self.cv1(inputs)) + + +class TFCrossConv(keras.layers.Layer): + # Cross Convolution + def __init__(self, c1, c2, k=3, s=1, g=1, e=1.0, shortcut=False, w=None): + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, (1, k), (1, s), w=w.cv1) + self.cv2 = TFConv(c_, c2, (k, 1), (s, 1), g=g, w=w.cv2) + self.add = shortcut and c1 == c2 + + def call(self, inputs): + return inputs + self.cv2(self.cv1(inputs)) if self.add else self.cv2(self.cv1(inputs)) + + +class TFConv2d(keras.layers.Layer): + # Substitution for PyTorch nn.Conv2D + def __init__(self, c1, c2, k, s=1, g=1, bias=True, w=None): + super().__init__() + assert g == 1, "TF v2.2 Conv2D does not support 'groups' argument" + self.conv = keras.layers.Conv2D(filters=c2, + kernel_size=k, + strides=s, + padding='VALID', + use_bias=bias, + kernel_initializer=keras.initializers.Constant( + w.weight.permute(2, 3, 1, 0).numpy()), + bias_initializer=keras.initializers.Constant(w.bias.numpy()) if bias else None) + + def call(self, inputs): + return self.conv(inputs) + + +class TFBottleneckCSP(keras.layers.Layer): + # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, w=None): + # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv2d(c1, c_, 1, 1, bias=False, w=w.cv2) + self.cv3 = TFConv2d(c_, c_, 1, 1, bias=False, w=w.cv3) + self.cv4 = TFConv(2 * c_, c2, 1, 1, w=w.cv4) + self.bn = TFBN(w.bn) + self.act = lambda x: keras.activations.swish(x) + self.m = keras.Sequential([TFBottleneck(c_, c_, shortcut, g, e=1.0, w=w.m[j]) for j in range(n)]) + + def call(self, inputs): + y1 = self.cv3(self.m(self.cv1(inputs))) + y2 = self.cv2(inputs) + return self.cv4(self.act(self.bn(tf.concat((y1, y2), axis=3)))) + + +class TFC3(keras.layers.Layer): + # CSP Bottleneck with 3 convolutions + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, w=None): + # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c1, c_, 1, 1, w=w.cv2) + self.cv3 = TFConv(2 * c_, c2, 1, 1, w=w.cv3) + self.m = keras.Sequential([TFBottleneck(c_, c_, shortcut, g, e=1.0, w=w.m[j]) for j in range(n)]) + + def call(self, inputs): + return self.cv3(tf.concat((self.m(self.cv1(inputs)), self.cv2(inputs)), axis=3)) + + +class TFC3x(keras.layers.Layer): + # 3 module with cross-convolutions + def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, w=None): + # ch_in, ch_out, number, shortcut, groups, expansion + super().__init__() + c_ = int(c2 * e) # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c1, c_, 1, 1, w=w.cv2) + self.cv3 = TFConv(2 * c_, c2, 1, 1, w=w.cv3) + self.m = keras.Sequential([ + TFCrossConv(c_, c_, k=3, s=1, g=g, e=1.0, shortcut=shortcut, w=w.m[j]) for j in range(n)]) + + def call(self, inputs): + return self.cv3(tf.concat((self.m(self.cv1(inputs)), self.cv2(inputs)), axis=3)) + + +class TFSPP(keras.layers.Layer): + # Spatial pyramid pooling layer used in YOLOv3-SPP + def __init__(self, c1, c2, k=(5, 9, 13), w=None): + super().__init__() + c_ = c1 // 2 # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c_ * (len(k) + 1), c2, 1, 1, w=w.cv2) + self.m = [keras.layers.MaxPool2D(pool_size=x, strides=1, padding='SAME') for x in k] + + def call(self, inputs): + x = self.cv1(inputs) + return self.cv2(tf.concat([x] + [m(x) for m in self.m], 3)) + + +class TFSPPF(keras.layers.Layer): + # Spatial pyramid pooling-Fast layer + def __init__(self, c1, c2, k=5, w=None): + super().__init__() + c_ = c1 // 2 # hidden channels + self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1) + self.cv2 = TFConv(c_ * 4, c2, 1, 1, w=w.cv2) + self.m = keras.layers.MaxPool2D(pool_size=k, strides=1, padding='SAME') + + def call(self, inputs): + x = self.cv1(inputs) + y1 = self.m(x) + y2 = self.m(y1) + return self.cv2(tf.concat([x, y1, y2, self.m(y2)], 3)) + + +class TFDetect(keras.layers.Layer): + # TF YOLOv5 Detect layer + def __init__(self, nc=80, anchors=(), ch=(), imgsz=(640, 640), w=None): # detection layer + super().__init__() + self.stride = tf.convert_to_tensor(w.stride.numpy(), dtype=tf.float32) + self.nc = nc # number of classes + self.no = nc + 5 # number of outputs per anchor + self.nl = len(anchors) # number of detection layers + self.na = len(anchors[0]) // 2 # number of anchors + self.grid = [tf.zeros(1)] * self.nl # init grid + self.anchors = tf.convert_to_tensor(w.anchors.numpy(), dtype=tf.float32) + self.anchor_grid = tf.reshape(self.anchors * tf.reshape(self.stride, [self.nl, 1, 1]), [self.nl, 1, -1, 1, 2]) + self.m = [TFConv2d(x, self.no * self.na, 1, w=w.m[i]) for i, x in enumerate(ch)] + self.training = False # set to False after building model + self.imgsz = imgsz + for i in range(self.nl): + ny, nx = self.imgsz[0] // self.stride[i], self.imgsz[1] // self.stride[i] + self.grid[i] = self._make_grid(nx, ny) + + def call(self, inputs): + z = [] # inference output + x = [] + for i in range(self.nl): + x.append(self.m[i](inputs[i])) + # x(bs,20,20,255) to x(bs,3,20,20,85) + ny, nx = self.imgsz[0] // self.stride[i], self.imgsz[1] // self.stride[i] + x[i] = tf.reshape(x[i], [-1, ny * nx, self.na, self.no]) + + if not self.training: # inference + y = tf.sigmoid(x[i]) + grid = tf.transpose(self.grid[i], [0, 2, 1, 3]) - 0.5 + anchor_grid = tf.transpose(self.anchor_grid[i], [0, 2, 1, 3]) * 4 + xy = (y[..., 0:2] * 2 + grid) * self.stride[i] # xy + wh = y[..., 2:4] ** 2 * anchor_grid + # Normalize xywh to 0-1 to reduce calibration error + xy /= tf.constant([[self.imgsz[1], self.imgsz[0]]], dtype=tf.float32) + wh /= tf.constant([[self.imgsz[1], self.imgsz[0]]], dtype=tf.float32) + y = tf.concat([xy, wh, y[..., 4:]], -1) + z.append(tf.reshape(y, [-1, self.na * ny * nx, self.no])) + + return tf.transpose(x, [0, 2, 1, 3]) if self.training else (tf.concat(z, 1), x) + + @staticmethod + def _make_grid(nx=20, ny=20): + # yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)]) + # return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float() + xv, yv = tf.meshgrid(tf.range(nx), tf.range(ny)) + return tf.cast(tf.reshape(tf.stack([xv, yv], 2), [1, 1, ny * nx, 2]), dtype=tf.float32) + + +class TFUpsample(keras.layers.Layer): + # TF version of torch.nn.Upsample() + def __init__(self, size, scale_factor, mode, w=None): # warning: all arguments needed including 'w' + super().__init__() + assert scale_factor == 2, "scale_factor must be 2" + self.upsample = lambda x: tf.image.resize(x, (x.shape[1] * 2, x.shape[2] * 2), method=mode) + # self.upsample = keras.layers.UpSampling2D(size=scale_factor, interpolation=mode) + # with default arguments: align_corners=False, half_pixel_centers=False + # self.upsample = lambda x: tf.raw_ops.ResizeNearestNeighbor(images=x, + # size=(x.shape[1] * 2, x.shape[2] * 2)) + + def call(self, inputs): + return self.upsample(inputs) + + +class TFConcat(keras.layers.Layer): + # TF version of torch.concat() + def __init__(self, dimension=1, w=None): + super().__init__() + assert dimension == 1, "convert only NCHW to NHWC concat" + self.d = 3 + + def call(self, inputs): + return tf.concat(inputs, self.d) + + +def parse_model(d, ch, model, imgsz): # model_dict, input_channels(3) + LOGGER.info(f"\n{'':>3}{'from':>18}{'n':>3}{'params':>10} {'module':<40}{'arguments':<30}") + anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'] + na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors # number of anchors + no = na * (nc + 5) # number of outputs = anchors * (classes + 5) + + layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out + for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']): # from, number, module, args + m_str = m + m = eval(m) if isinstance(m, str) else m # eval strings + for j, a in enumerate(args): + try: + args[j] = eval(a) if isinstance(a, str) else a # eval strings + except NameError: + pass + + n = max(round(n * gd), 1) if n > 1 else n # depth gain + if m in [ + nn.Conv2d, Conv, DWConv, DWConvTranspose2d, Bottleneck, SPP, SPPF, MixConv2d, Focus, CrossConv, + BottleneckCSP, C3, C3x]: + c1, c2 = ch[f], args[0] + c2 = make_divisible(c2 * gw, 8) if c2 != no else c2 + + args = [c1, c2, *args[1:]] + if m in [BottleneckCSP, C3, C3x]: + args.insert(2, n) + n = 1 + elif m is nn.BatchNorm2d: + args = [ch[f]] + elif m is Concat: + c2 = sum(ch[-1 if x == -1 else x + 1] for x in f) + elif m is Detect: + args.append([ch[x + 1] for x in f]) + if isinstance(args[1], int): # number of anchors + args[1] = [list(range(args[1] * 2))] * len(f) + args.append(imgsz) + else: + c2 = ch[f] + + tf_m = eval('TF' + m_str.replace('nn.', '')) + m_ = keras.Sequential([tf_m(*args, w=model.model[i][j]) for j in range(n)]) if n > 1 \ + else tf_m(*args, w=model.model[i]) # module + + torch_m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args) # module + t = str(m)[8:-2].replace('__main__.', '') # module type + np = sum(x.numel() for x in torch_m_.parameters()) # number params + m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params + LOGGER.info(f'{i:>3}{str(f):>18}{str(n):>3}{np:>10} {t:<40}{str(args):<30}') # print + save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist + layers.append(m_) + ch.append(c2) + return keras.Sequential(layers), sorted(save) + + +class TFModel: + # TF YOLOv5 model + def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, model=None, imgsz=(640, 640)): # model, channels, classes + super().__init__() + if isinstance(cfg, dict): + self.yaml = cfg # model dict + else: # is *.yaml + import yaml # for torch hub + self.yaml_file = Path(cfg).name + with open(cfg) as f: + self.yaml = yaml.load(f, Loader=yaml.FullLoader) # model dict + + # Define model + if nc and nc != self.yaml['nc']: + LOGGER.info(f"Overriding {cfg} nc={self.yaml['nc']} with nc={nc}") + self.yaml['nc'] = nc # override yaml value + self.model, self.savelist = parse_model(deepcopy(self.yaml), ch=[ch], model=model, imgsz=imgsz) + + def predict(self, + inputs, + tf_nms=False, + agnostic_nms=False, + topk_per_class=100, + topk_all=100, + iou_thres=0.45, + conf_thres=0.25): + y = [] # outputs + x = inputs + for m in self.model.layers: + if m.f != -1: # if not from previous layer + x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers + + x = m(x) # run + y.append(x if m.i in self.savelist else None) # save output + + # Add TensorFlow NMS + if tf_nms: + boxes = self._xywh2xyxy(x[0][..., :4]) + probs = x[0][:, :, 4:5] + classes = x[0][:, :, 5:] + scores = probs * classes + if agnostic_nms: + nms = AgnosticNMS()((boxes, classes, scores), topk_all, iou_thres, conf_thres) + else: + boxes = tf.expand_dims(boxes, 2) + nms = tf.image.combined_non_max_suppression(boxes, + scores, + topk_per_class, + topk_all, + iou_thres, + conf_thres, + clip_boxes=False) + return nms, x[1] + return x[0] # output only first tensor [1,6300,85] = [xywh, conf, class0, class1, ...] + # x = x[0][0] # [x(1,6300,85), ...] to x(6300,85) + # xywh = x[..., :4] # x(6300,4) boxes + # conf = x[..., 4:5] # x(6300,1) confidences + # cls = tf.reshape(tf.cast(tf.argmax(x[..., 5:], axis=1), tf.float32), (-1, 1)) # x(6300,1) classes + # return tf.concat([conf, cls, xywh], 1) + + @staticmethod + def _xywh2xyxy(xywh): + # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right + x, y, w, h = tf.split(xywh, num_or_size_splits=4, axis=-1) + return tf.concat([x - w / 2, y - h / 2, x + w / 2, y + h / 2], axis=-1) + + +class AgnosticNMS(keras.layers.Layer): + # TF Agnostic NMS + def call(self, input, topk_all, iou_thres, conf_thres): + # wrap map_fn to avoid TypeSpec related error https://stackoverflow.com/a/65809989/3036450 + return tf.map_fn(lambda x: self._nms(x, topk_all, iou_thres, conf_thres), + input, + fn_output_signature=(tf.float32, tf.float32, tf.float32, tf.int32), + name='agnostic_nms') + + @staticmethod + def _nms(x, topk_all=100, iou_thres=0.45, conf_thres=0.25): # agnostic NMS + boxes, classes, scores = x + class_inds = tf.cast(tf.argmax(classes, axis=-1), tf.float32) + scores_inp = tf.reduce_max(scores, -1) + selected_inds = tf.image.non_max_suppression(boxes, + scores_inp, + max_output_size=topk_all, + iou_threshold=iou_thres, + score_threshold=conf_thres) + selected_boxes = tf.gather(boxes, selected_inds) + padded_boxes = tf.pad(selected_boxes, + paddings=[[0, topk_all - tf.shape(selected_boxes)[0]], [0, 0]], + mode="CONSTANT", + constant_values=0.0) + selected_scores = tf.gather(scores_inp, selected_inds) + padded_scores = tf.pad(selected_scores, + paddings=[[0, topk_all - tf.shape(selected_boxes)[0]]], + mode="CONSTANT", + constant_values=-1.0) + selected_classes = tf.gather(class_inds, selected_inds) + padded_classes = tf.pad(selected_classes, + paddings=[[0, topk_all - tf.shape(selected_boxes)[0]]], + mode="CONSTANT", + constant_values=-1.0) + valid_detections = tf.shape(selected_inds)[0] + return padded_boxes, padded_scores, padded_classes, valid_detections + + +def activations(act=nn.SiLU): + # Returns TF activation from input PyTorch activation + if isinstance(act, nn.LeakyReLU): + return lambda x: keras.activations.relu(x, alpha=0.1) + elif isinstance(act, nn.Hardswish): + return lambda x: x * tf.nn.relu6(x + 3) * 0.166666667 + elif isinstance(act, (nn.SiLU, SiLU)): + return lambda x: keras.activations.swish(x) + else: + raise Exception(f'no matching TensorFlow activation found for PyTorch activation {act}') + + +def representative_dataset_gen(dataset, ncalib=100): + # Representative dataset generator for use with converter.representative_dataset, returns a generator of np arrays + for n, (path, img, im0s, vid_cap, string) in enumerate(dataset): + im = np.transpose(img, [1, 2, 0]) + im = np.expand_dims(im, axis=0).astype(np.float32) + im /= 255 + yield [im] + if n >= ncalib: + break + + +def run( + weights=ROOT / 'yolov5s.pt', # weights path + imgsz=(640, 640), # inference size h,w + batch_size=1, # batch size + dynamic=False, # dynamic batch size +): + # PyTorch model + im = torch.zeros((batch_size, 3, *imgsz)) # BCHW image + model = attempt_load(weights, device=torch.device('cpu'), inplace=True, fuse=False) + _ = model(im) # inference + model.info() + + # TensorFlow model + im = tf.zeros((batch_size, *imgsz, 3)) # BHWC image + tf_model = TFModel(cfg=model.yaml, model=model, nc=model.nc, imgsz=imgsz) + _ = tf_model.predict(im) # inference + + # Keras model + im = keras.Input(shape=(*imgsz, 3), batch_size=None if dynamic else batch_size) + keras_model = keras.Model(inputs=im, outputs=tf_model.predict(im)) + keras_model.summary() + + LOGGER.info('PyTorch, TensorFlow and Keras models successfully verified.\nUse export.py for TF model export.') + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='weights path') + parser.add_argument('--imgsz', '--img', '--img-size', nargs='+', type=int, default=[640], help='inference size h,w') + parser.add_argument('--batch-size', type=int, default=1, help='batch size') + parser.add_argument('--dynamic', action='store_true', help='dynamic batch size') + opt = parser.parse_args() + opt.imgsz *= 2 if len(opt.imgsz) == 1 else 1 # expand + print_args(vars(opt)) + return opt + + +def main(opt): + run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/application/yolov5_example/models/yolo.py b/application/yolov5_example/models/yolo.py new file mode 100644 index 00000000..e7a9fa1a --- /dev/null +++ b/application/yolov5_example/models/yolo.py @@ -0,0 +1,360 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +YOLO-specific modules + +Usage: + $ python path/to/models/yolo.py --cfg yolov5s.yaml +""" + +import argparse +import contextlib +import os +import platform +import sys +from copy import deepcopy +from pathlib import Path + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +if platform.system() != 'Windows': + ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from models.common import * +from models.experimental import * +from utils.autoanchor import check_anchor_order +from utils.general import LOGGER, check_version, check_yaml, make_divisible, print_args +from utils.plots import feature_visualization +from utils.torch_utils import (fuse_conv_and_bn, initialize_weights, model_info, profile, scale_img, select_device, + time_sync) + +try: + import thop # for FLOPs computation +except ImportError: + thop = None + + +class Detect(nn.Module): + stride = None # strides computed during build + onnx_dynamic = False # ONNX export parameter + export = False # export mode + + def __init__(self, nc=80, anchors=(), ch=(), inplace=True): # detection layer + super().__init__() + self.nc = nc # number of classes + self.no = nc + 5 # number of outputs per anchor + self.nl = len(anchors) # number of detection layers + self.na = len(anchors[0]) // 2 # number of anchors + self.grid = [torch.zeros(1)] * self.nl # init grid + self.anchor_grid = [torch.zeros(1)] * self.nl # init anchor grid + self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2)) # shape(nl,na,2) + self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch) # output conv + self.inplace = inplace # use inplace ops (e.g. slice assignment) + + def forward(self, x): + z = [] # inference output + for i in range(self.nl): + x[i] = self.m[i](x[i]) # conv + bs, _, ny, nx = x[i].shape # x(bs,255,20,20) to x(bs,3,20,20,85) + x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous() + + if not self.training: # inference + if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]: + self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i) + + y = x[i].sigmoid() + if self.inplace: + y[..., 0:2] = (y[..., 0:2] * 2 + self.grid[i]) * self.stride[i] # xy + y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i] # wh + else: # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953 + xy, wh, conf = y.split((2, 2, self.nc + 1), 4) # y.tensor_split((2, 4, 5), 4) # torch 1.8.0 + xy = (xy * 2 + self.grid[i]) * self.stride[i] # xy + wh = (wh * 2) ** 2 * self.anchor_grid[i] # wh + y = torch.cat((xy, wh, conf), 4) + z.append(y.view(bs, -1, self.no)) + + return x if self.training else (torch.cat(z, 1),) if self.export else (torch.cat(z, 1), x) + + def _make_grid(self, nx=20, ny=20, i=0, torch_1_10=check_version(torch.__version__, '1.10.0')): + d = self.anchors[i].device + t = self.anchors[i].dtype + shape = 1, self.na, ny, nx, 2 # grid shape + y, x = torch.arange(ny, device=d, dtype=t), torch.arange(nx, device=d, dtype=t) + if torch_1_10: # torch>=1.10.0 meshgrid workaround for torch>=0.7 compatibility + yv, xv = torch.meshgrid(y, x, indexing='ij') + else: + yv, xv = torch.meshgrid(y, x) + grid = torch.stack((xv, yv), 2).expand(shape) - 0.5 # add grid offset, i.e. y = 2.0 * x - 0.5 + anchor_grid = (self.anchors[i] * self.stride[i]).view((1, self.na, 1, 1, 2)).expand(shape) + return grid, anchor_grid + + +class BaseModel(nn.Module): + # YOLOv5 base model + def forward(self, x, profile=False, visualize=False): + return self._forward_once(x, profile, visualize) # single-scale inference, train + + def _forward_once(self, x, profile=False, visualize=False): + y, dt = [], [] # outputs + for m in self.model: + if m.f != -1: # if not from previous layer + x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layers + # if profile: + # self._profile_one_layer(m, x, dt) + x = m(x) # run + y.append(x if m.i in self.save else None) # save output + # if visualize: + # feature_visualization(x, m.type, m.i, save_dir=visualize) + return x + + def _profile_one_layer(self, m, x, dt): + c = m == self.model[-1] # is final layer, copy input as inplace fix + o = thop.profile(m, inputs=(x.copy() if c else x,), verbose=False)[0] / 1E9 * 2 if thop else 0 # FLOPs + t = time_sync() + for _ in range(10): + m(x.copy() if c else x) + dt.append((time_sync() - t) * 100) + if m == self.model[0]: + LOGGER.info(f"{'time (ms)':>10s} {'GFLOPs':>10s} {'params':>10s} module") + LOGGER.info(f'{dt[-1]:10.2f} {o:10.2f} {m.np:10.0f} {m.type}') + if c: + LOGGER.info(f"{sum(dt):10.2f} {'-':>10s} {'-':>10s} Total") + + def fuse(self): # fuse model Conv2d() + BatchNorm2d() layers + LOGGER.info('Fusing layers... ') + for m in self.model.modules(): + if isinstance(m, (Conv, DWConv)) and hasattr(m, 'bn'): + m.conv = fuse_conv_and_bn(m.conv, m.bn) # update conv + delattr(m, 'bn') # remove batchnorm + m.forward = m.forward_fuse # update forward + self.info() + return self + + def info(self, verbose=False, img_size=640): # print model information + model_info(self, verbose, img_size) + + def _apply(self, fn): + # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers + self = super()._apply(fn) + m = self.model[-1] # Detect() + if isinstance(m, Detect): + m.stride = fn(m.stride) + m.grid = list(map(fn, m.grid)) + if isinstance(m.anchor_grid, list): + m.anchor_grid = list(map(fn, m.anchor_grid)) + return self + + +class DetectionModel(BaseModel): + # YOLOv5 detection model + def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, anchors=None): # model, input channels, number of classes + super().__init__() + if isinstance(cfg, dict): + self.yaml = cfg # model dict + else: # is *.yaml + import yaml # for torch hub + self.yaml_file = Path(cfg).name + with open(cfg, encoding='ascii', errors='ignore') as f: + self.yaml = yaml.safe_load(f) # model dict + + # Define model + ch = self.yaml['ch'] = self.yaml.get('ch', ch) # input channels + if nc and nc != self.yaml['nc']: + LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}") + self.yaml['nc'] = nc # override yaml value + if anchors: + LOGGER.info(f'Overriding model.yaml anchors with anchors={anchors}') + self.yaml['anchors'] = round(anchors) # override yaml value + self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch]) # model, savelist + self.names = [str(i) for i in range(self.yaml['nc'])] # default names + self.inplace = self.yaml.get('inplace', True) + + # Build strides, anchors + m = self.model[-1] # Detect() + if isinstance(m, Detect): + s = 256 # 2x min stride + m.inplace = self.inplace + m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))]) # forward + check_anchor_order(m) # must be in pixel-space (not grid-space) + m.anchors /= m.stride.view(-1, 1, 1) + self.stride = m.stride + self._initialize_biases() # only run once + + # Init weights, biases + initialize_weights(self) + self.info() + LOGGER.info('') + + def forward(self, x, augment=False, profile=False, visualize=False): + # if augment: + # return self._forward_augment(x) # augmented inference, None + return self._forward_once(x, profile, visualize) # single-scale inference, train + + def _forward_augment(self, x): + img_size = x.shape[-2:] # height, width + s = [1, 0.83, 0.67] # scales + f = [None, 3, None] # flips (2-ud, 3-lr) + y = [] # outputs + for si, fi in zip(s, f): + xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max())) + yi = self._forward_once(xi)[0] # forward + # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1]) # save + yi = self._descale_pred(yi, fi, si, img_size) + y.append(yi) + y = self._clip_augmented(y) # clip augmented tails + return torch.cat(y, 1), None # augmented inference, train + + def _descale_pred(self, p, flips, scale, img_size): + # de-scale predictions following augmented inference (inverse operation) + if self.inplace: + p[..., :4] /= scale # de-scale + if flips == 2: + p[..., 1] = img_size[0] - p[..., 1] # de-flip ud + elif flips == 3: + p[..., 0] = img_size[1] - p[..., 0] # de-flip lr + else: + x, y, wh = p[..., 0:1] / scale, p[..., 1:2] / scale, p[..., 2:4] / scale # de-scale + if flips == 2: + y = img_size[0] - y # de-flip ud + elif flips == 3: + x = img_size[1] - x # de-flip lr + p = torch.cat((x, y, wh, p[..., 4:]), -1) + return p + + def _clip_augmented(self, y): + # Clip YOLOv5 augmented inference tails + nl = self.model[-1].nl # number of detection layers (P3-P5) + g = sum(4 ** x for x in range(nl)) # grid points + e = 1 # exclude layer count + i = (y[0].shape[1] // g) * sum(4 ** x for x in range(e)) # indices + y[0] = y[0][:, :-i] # large + i = (y[-1].shape[1] // g) * sum(4 ** (nl - 1 - x) for x in range(e)) # indices + y[-1] = y[-1][:, i:] # small + return y + + def _initialize_biases(self, cf=None): # initialize biases into Detect(), cf is class frequency + # https://arxiv.org/abs/1708.02002 section 3.3 + # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1. + m = self.model[-1] # Detect() module + for mi, s in zip(m.m, m.stride): # from + b = mi.bias.view(m.na, -1).detach() # conv.bias(255) to (3,85) + b[:, 4] += math.log(8 / (640 / s) ** 2) # obj (8 objects per 640 image) + b[:, 5:] += math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum()) # cls + mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True) + + +Model = DetectionModel # retain YOLOv5 'Model' class for backwards compatibility + + +class ClassificationModel(BaseModel): + # YOLOv5 classification model + def __init__(self, cfg=None, model=None, nc=1000, cutoff=10): # yaml, model, number of classes, cutoff index + super().__init__() + self._from_detection_model(model, nc, cutoff) if model is not None else self._from_yaml(cfg) + + def _from_detection_model(self, model, nc=1000, cutoff=10): + # Create a YOLOv5 classification model from a YOLOv5 detection model + if isinstance(model, DetectMultiBackend): + model = model.model # unwrap DetectMultiBackend + model.model = model.model[:cutoff] # backbone + m = model.model[-1] # last layer + ch = m.conv.in_channels if hasattr(m, 'conv') else m.cv1.conv.in_channels # ch into module + c = Classify(ch, nc) # Classify() + c.i, c.f, c.type = m.i, m.f, 'models.common.Classify' # index, from, type + model.model[-1] = c # replace + self.model = model.model + self.stride = model.stride + self.save = [] + self.nc = nc + + def _from_yaml(self, cfg): + # Create a YOLOv5 classification model from a *.yaml file + self.model = None + + +def parse_model(d, ch): # model_dict, input_channels(3) + LOGGER.info(f"\n{'':>3}{'from':>18}{'n':>3}{'params':>10} {'module':<40}{'arguments':<30}") + anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple'] + na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors # number of anchors + no = na * (nc + 5) # number of outputs = anchors * (classes + 5) + + layers, save, c2 = [], [], ch[-1] # layers, savelist, ch out + for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']): # from, number, module, args + m = eval(m) if isinstance(m, str) else m # eval strings + for j, a in enumerate(args): + with contextlib.suppress(NameError): + args[j] = eval(a) if isinstance(a, str) else a # eval strings + + n = n_ = max(round(n * gd), 1) if n > 1 else n # depth gain + if m in (Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, MixConv2d, Focus, CrossConv, + BottleneckCSP, C3, C3TR, C3SPP, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x): + c1, c2 = ch[f], args[0] + if c2 != no: # if not output + c2 = make_divisible(c2 * gw, 8) + + args = [c1, c2, *args[1:]] + if m in [BottleneckCSP, C3, C3TR, C3Ghost, C3x]: + args.insert(2, n) # number of repeats + n = 1 + elif m is nn.BatchNorm2d: + args = [ch[f]] + elif m is Concat: + c2 = sum(ch[x] for x in f) + elif m is Detect: + args.append([ch[x] for x in f]) + if isinstance(args[1], int): # number of anchors + args[1] = [list(range(args[1] * 2))] * len(f) + elif m is Contract: + c2 = ch[f] * args[0] ** 2 + elif m is Expand: + c2 = ch[f] // args[0] ** 2 + else: + c2 = ch[f] + + m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args) # module + t = str(m)[8:-2].replace('__main__.', '') # module type + np = sum(x.numel() for x in m_.parameters()) # number params + m_.i, m_.f, m_.type, m_.np = i, f, t, np # attach index, 'from' index, type, number params + LOGGER.info(f'{i:>3}{str(f):>18}{n_:>3}{np:10.0f} {t:<40}{str(args):<30}') # print + save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelist + layers.append(m_) + if i == 0: + ch = [] + ch.append(c2) + return nn.Sequential(*layers), sorted(save) + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml') + parser.add_argument('--batch-size', type=int, default=1, help='total batch size for all GPUs') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--profile', action='store_true', help='profile model speed') + parser.add_argument('--line-profile', action='store_true', help='profile model speed layer by layer') + parser.add_argument('--test', action='store_true', help='test all yolo*.yaml') + opt = parser.parse_args() + opt.cfg = check_yaml(opt.cfg) # check YAML + print_args(vars(opt)) + device = select_device(opt.device) + + # Create model + im = torch.rand(opt.batch_size, 3, 640, 640).to(device) + model = Model(opt.cfg).to(device) + + # Options + if opt.line_profile: # profile layer by layer + model(im, profile=True) + + elif opt.profile: # profile forward-backward + results = profile(input=im, ops=[model], n=3) + + elif opt.test: # test all models + for cfg in Path(ROOT / 'models').rglob('yolo*.yaml'): + try: + _ = Model(cfg) + except Exception as e: + print(f'Error in {cfg}: {e}') + + else: # report fused model summary + model.fuse() diff --git a/application/yolov5_example/models/yolov5l.yaml b/application/yolov5_example/models/yolov5l.yaml new file mode 100644 index 00000000..ce8a5de4 --- /dev/null +++ b/application/yolov5_example/models/yolov5l.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.0 # model depth multiple +width_multiple: 1.0 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/yolov5m.yaml b/application/yolov5_example/models/yolov5m.yaml new file mode 100644 index 00000000..ad13ab37 --- /dev/null +++ b/application/yolov5_example/models/yolov5m.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.67 # model depth multiple +width_multiple: 0.75 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/yolov5n.yaml b/application/yolov5_example/models/yolov5n.yaml new file mode 100644 index 00000000..8a28a40d --- /dev/null +++ b/application/yolov5_example/models/yolov5n.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.25 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/yolov5s.yaml b/application/yolov5_example/models/yolov5s.yaml new file mode 100644 index 00000000..f35beabb --- /dev/null +++ b/application/yolov5_example/models/yolov5s.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 0.33 # model depth multiple +width_multiple: 0.50 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/models/yolov5x.yaml b/application/yolov5_example/models/yolov5x.yaml new file mode 100644 index 00000000..f617a027 --- /dev/null +++ b/application/yolov5_example/models/yolov5x.yaml @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license + +# Parameters +nc: 80 # number of classes +depth_multiple: 1.33 # model depth multiple +width_multiple: 1.25 # layer channel multiple +anchors: + - [10,13, 16,30, 33,23] # P3/8 + - [30,61, 62,45, 59,119] # P4/16 + - [116,90, 156,198, 373,326] # P5/32 + +# YOLOv5 v6.0 backbone +backbone: + # [from, number, module, args] + [[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2 + [-1, 1, Conv, [128, 3, 2]], # 1-P2/4 + [-1, 3, C3, [128]], + [-1, 1, Conv, [256, 3, 2]], # 3-P3/8 + [-1, 6, C3, [256]], + [-1, 1, Conv, [512, 3, 2]], # 5-P4/16 + [-1, 9, C3, [512]], + [-1, 1, Conv, [1024, 3, 2]], # 7-P5/32 + [-1, 3, C3, [1024]], + [-1, 1, SPPF, [1024, 5]], # 9 + ] + +# YOLOv5 v6.0 head +head: + [[-1, 1, Conv, [512, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 6], 1, Concat, [1]], # cat backbone P4 + [-1, 3, C3, [512, False]], # 13 + + [-1, 1, Conv, [256, 1, 1]], + [-1, 1, nn.Upsample, [None, 2, 'nearest']], + [[-1, 4], 1, Concat, [1]], # cat backbone P3 + [-1, 3, C3, [256, False]], # 17 (P3/8-small) + + [-1, 1, Conv, [256, 3, 2]], + [[-1, 14], 1, Concat, [1]], # cat head P4 + [-1, 3, C3, [512, False]], # 20 (P4/16-medium) + + [-1, 1, Conv, [512, 3, 2]], + [[-1, 10], 1, Concat, [1]], # cat head P5 + [-1, 3, C3, [1024, False]], # 23 (P5/32-large) + + [[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5) + ] diff --git a/application/yolov5_example/mqbench.code-workspace b/application/yolov5_example/mqbench.code-workspace new file mode 100644 index 00000000..7e768222 --- /dev/null +++ b/application/yolov5_example/mqbench.code-workspace @@ -0,0 +1,11 @@ +{ + "folders": [ + { + "path": "../../opt/conda/lib/python3.7/site-packages/MQBench-0.0.6-py3.7.egg/mqbench" + }, + { + "path": "." + } + ], + "settings": {} +} \ No newline at end of file diff --git a/application/yolov5_example/requirements.txt b/application/yolov5_example/requirements.txt new file mode 100644 index 00000000..10620566 --- /dev/null +++ b/application/yolov5_example/requirements.txt @@ -0,0 +1,43 @@ +# YOLOv5 requirements +# Usage: pip install -r requirements.txt + +# Base ---------------------------------------- +matplotlib>=3.2.2 +numpy>=1.18.5 +opencv-python>=4.1.1 +Pillow>=7.1.2 +PyYAML>=5.3.1 +requests>=2.23.0 +scipy>=1.4.1 +torch>=1.7.0 +torchvision>=0.8.1 +tqdm>=4.64.0 +protobuf<=3.20.1 # https://github.com/ultralytics/yolov5/issues/8012 + +# Logging ------------------------------------- +tensorboard>=2.4.1 +# wandb +# clearml + +# Plotting ------------------------------------ +pandas>=1.1.4 +seaborn>=0.11.0 + +# Export -------------------------------------- +# coremltools>=5.2 # CoreML export +# onnx>=1.9.0 # ONNX export +# onnx-simplifier>=0.4.1 # ONNX simplifier +# nvidia-pyindex # TensorRT export +# nvidia-tensorrt # TensorRT export +# scikit-learn==0.19.2 # CoreML quantization +# tensorflow>=2.4.1 # TFLite export (or tensorflow-cpu, tensorflow-aarch64) +# tensorflowjs>=3.9.0 # TF.js export +# openvino-dev # OpenVINO export + +# Extras -------------------------------------- +ipython # interactive notebook +psutil # system utilization +thop>=0.1.1 # FLOPs computation +# albumentations>=1.0.3 +# pycocotools>=2.0 # COCO mAP +# roboflow diff --git a/application/yolov5_example/setup.cfg b/application/yolov5_example/setup.cfg new file mode 100644 index 00000000..020a7574 --- /dev/null +++ b/application/yolov5_example/setup.cfg @@ -0,0 +1,59 @@ +# Project-wide configuration file, can be used for package metadata and other toll configurations +# Example usage: global configuration for PEP8 (via flake8) setting or default pytest arguments +# Local usage: pip install pre-commit, pre-commit run --all-files + +[metadata] +license_file = LICENSE +description_file = README.md + + +[tool:pytest] +norecursedirs = + .git + dist + build +addopts = + --doctest-modules + --durations=25 + --color=yes + + +[flake8] +max-line-length = 120 +exclude = .tox,*.egg,build,temp +select = E,W,F +doctests = True +verbose = 2 +# https://pep8.readthedocs.io/en/latest/intro.html#error-codes +format = pylint +# see: https://www.flake8rules.com/ +ignore = + E731 # Do not assign a lambda expression, use a def + F405 # name may be undefined, or defined from star imports: module + E402 # module level import not at top of file + F401 # module imported but unused + W504 # line break after binary operator + E127 # continuation line over-indented for visual indent + W504 # line break after binary operator + E231 # missing whitespace after ‘,’, ‘;’, or ‘:’ + E501 # line too long + F403 # ‘from module import *’ used; unable to detect undefined names + + +[isort] +# https://pycqa.github.io/isort/docs/configuration/options.html +line_length = 120 +# see: https://pycqa.github.io/isort/docs/configuration/multi_line_output_modes.html +multi_line_output = 0 + + +[yapf] +based_on_style = pep8 +spaces_before_comment = 2 +COLUMN_LIMIT = 120 +COALESCE_BRACKETS = True +SPACES_AROUND_POWER_OPERATOR = True +SPACE_BETWEEN_ENDING_COMMA_AND_CLOSING_BRACKET = False +SPLIT_BEFORE_CLOSING_BRACKET = False +SPLIT_BEFORE_FIRST_ARGUMENT = False +# EACH_DICT_ENTRY_ON_SEPARATE_LINE = False diff --git a/application/yolov5_example/train.py b/application/yolov5_example/train.py new file mode 100644 index 00000000..73405e54 --- /dev/null +++ b/application/yolov5_example/train.py @@ -0,0 +1,728 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Train a YOLOv5 model on a custom dataset. + +Models and datasets download automatically from the latest YOLOv5 release. +Models: https://github.com/ultralytics/yolov5/tree/master/models +Datasets: https://github.com/ultralytics/yolov5/tree/master/data +Tutorial: https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data + +Usage: + $ python path/to/train.py --data coco128.yaml --weights yolov5s.pt --img 640 # from pretrained (RECOMMENDED) + $ python path/to/train.py --data coco128.yaml --weights '' --cfg yolov5s.yaml --img 640 # from scratch +""" + +import argparse +from ast import arg +import math +import os +import random +import sys +import time +from copy import deepcopy +from datetime import datetime +from pathlib import Path + +import numpy as np +import torch +import torch.distributed as dist +import torch.nn as nn +import yaml +from torch.optim import lr_scheduler +from tqdm import tqdm +from mqbench.convert_deploy import convert_deploy, convert_onnx +from mqbench.prepare_by_platform import prepare_by_platform, BackendType +from mqbench.utils.state import enable_calibration, enable_quantization, disable_all + + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +import val # for end-of-epoch mAP +from models.experimental import attempt_load +from models.yolo import Model +from utils.autoanchor import check_anchors +from utils.autobatch import check_train_batch_size +from utils.callbacks import Callbacks +from utils.dataloaders import create_dataloader +from utils.downloads import attempt_download, is_url +from utils.general import (LOGGER, check_amp, check_dataset, check_file, check_git_status, check_img_size, + check_requirements, check_suffix, check_yaml, colorstr, get_latest_run, increment_path, + init_seeds, intersect_dicts, labels_to_class_weights, labels_to_image_weights, methods, + one_cycle, print_args, print_mutation, strip_optimizer, yaml_save) +from utils.loggers import Loggers +from utils.loggers.wandb.wandb_utils import check_wandb_resume +from utils.loss import ComputeLoss +from utils.metrics import fitness +from utils.plots import plot_evolve, plot_labels +from utils.torch_utils import (EarlyStopping, ModelEMA, de_parallel, select_device, smart_DDP, smart_optimizer, + smart_resume, torch_distributed_zero_first) + +LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1)) # https://pytorch.org/docs/stable/elastic/run.html +RANK = int(os.getenv('RANK', -1)) +WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1)) + + +def train(hyp, opt, device, callbacks): # hyp is path/to/hyp.yaml or hyp dictionary + save_dir, epochs, batch_size, weights, single_cls, evolve, data, cfg, resume, noval, nosave, workers, freeze = \ + Path(opt.save_dir), opt.epochs, opt.batch_size, opt.weights, opt.single_cls, opt.evolve, opt.data, opt.cfg, \ + opt.resume, opt.noval, opt.nosave, opt.workers, opt.freeze + callbacks.run('on_pretrain_routine_start') + + # Directories + w = save_dir / 'weights' # weights dir + (w.parent if evolve else w).mkdir(parents=True, exist_ok=True) # make dir + last, best = w / 'last.pt', w / 'best.pt' + + # Hyperparameters + if isinstance(hyp, str): + with open(hyp, errors='ignore') as f: + hyp = yaml.safe_load(f) # load hyps dict + LOGGER.info(colorstr('hyperparameters: ') + ', '.join(f'{k}={v}' for k, v in hyp.items())) + opt.hyp = hyp.copy() # for saving hyps to checkpoints + + # Save run settings + if not evolve: + yaml_save(save_dir / 'hyp.yaml', hyp) + yaml_save(save_dir / 'opt.yaml', vars(opt)) + + # Loggers + data_dict = None + if RANK in {-1, 0}: + loggers = Loggers(save_dir, weights, opt, hyp, LOGGER) # loggers instance + if loggers.clearml: + data_dict = loggers.clearml.data_dict # None if no ClearML dataset or filled in by ClearML + if loggers.wandb: + data_dict = loggers.wandb.data_dict + if resume: + weights, epochs, hyp, batch_size = opt.weights, opt.epochs, opt.hyp, opt.batch_size + + # Register actions + for k in methods(loggers): + callbacks.register_action(k, callback=getattr(loggers, k)) + + # Config + plots = not evolve and not opt.noplots # create plots + cuda = device.type != 'cpu' + init_seeds(opt.seed + 1 + RANK, deterministic=True) + with torch_distributed_zero_first(LOCAL_RANK): + data_dict = data_dict or check_dataset(data) # check if None + train_path, val_path = data_dict['train'], data_dict['val'] + nc = 1 if single_cls else int(data_dict['nc']) # number of classes + names = ['item'] if single_cls and len(data_dict['names']) != 1 else data_dict['names'] # class names + assert len(names) == nc, f'{len(names)} names found for nc={nc} dataset in {data}' # check + is_coco = isinstance(val_path, str) and val_path.endswith('coco/val2017.txt') # COCO dataset + + # Model + check_suffix(weights, '.pt') # check weights + pretrained = weights.endswith('.pt') + if pretrained: + with torch_distributed_zero_first(LOCAL_RANK): + weights = attempt_download(weights) # download if not found locally + ckpt = torch.load(weights, map_location='cpu') # load checkpoint to CPU to avoid CUDA memory leak + model = Model(cfg or ckpt['model'].yaml, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create + exclude = ['anchor'] if (cfg or hyp.get('anchors')) and not resume else [] # exclude keys + csd = ckpt['model'].float().state_dict() # checkpoint state_dict as FP32 + csd = intersect_dicts(csd, model.state_dict(), exclude=exclude) # intersect + model.load_state_dict(csd, strict=False) # load + LOGGER.info(f'Transferred {len(csd)}/{len(model.state_dict())} items from {weights}') # report + else: + model = Model(cfg, ch=3, nc=nc, anchors=hyp.get('anchors')).to(device) # create + amp = check_amp(model) # check AMP + amp = False + + # Freeze + freeze = [f'model.{x}.' for x in (freeze if len(freeze) > 1 else range(freeze[0]))] # layers to freeze + for k, v in model.named_parameters(): + v.requires_grad = True # train all layers + # v.register_hook(lambda x: torch.nan_to_num(x)) # NaN to 0 (commented for erratic training results) + if any(x in k for x in freeze): + LOGGER.info(f'freezing {k}') + v.requires_grad = False + + # Image size + gs = max(int(model.stride.max()), 32) # grid size (max stride) + imgsz = check_img_size(opt.imgsz, gs, floor=gs * 2) # verify imgsz is gs-multiple + + # Batch size + if RANK == -1 and batch_size == -1: # single-GPU only, estimate best batch size + batch_size = check_train_batch_size(model, imgsz, amp) + loggers.on_params_update({"batch_size": batch_size}) + + # Optimizer + nbs = 64 # nominal batch size + accumulate = max(round(nbs / batch_size), 1) # accumulate loss before optimizing + hyp['weight_decay'] *= batch_size * accumulate / nbs # scale weight_decay + optimizer = smart_optimizer(model, opt.optimizer, hyp['lr0'], hyp['momentum'], hyp['weight_decay']) + + print('wxc1 lr0:', hyp['lr0']) + # Scheduler + if opt.cos_lr: + lf = one_cycle(1, hyp['lrf'], epochs) # cosine 1->hyp['lrf'] + else: + lf = lambda x: (1 - x / epochs) * (1.0 - hyp['lrf']) + hyp['lrf'] # linear + scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf) # plot_lr_scheduler(optimizer, scheduler, epochs) + + # EMA + ema = ModelEMA(model) if RANK in {-1, 0} else None + + # Resume + best_fitness, start_epoch = 0.0, 0 + if pretrained: + if resume: + best_fitness, start_epoch, epochs = smart_resume(ckpt, optimizer, ema, weights, epochs, resume) + del ckpt, csd + + # DP mode + if cuda and RANK == -1 and torch.cuda.device_count() > 1: + LOGGER.warning('WARNING: DP not recommended, use torch.distributed.run for best DDP Multi-GPU results.\n' + 'See Multi-GPU Tutorial at https://github.com/ultralytics/yolov5/issues/475 to get started.') + model = torch.nn.DataParallel(model) + + # SyncBatchNorm + if opt.sync_bn and cuda and RANK != -1: + model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model).to(device) + LOGGER.info('Using SyncBatchNorm()') + + # Trainloader + train_loader, dataset = create_dataloader(train_path, + imgsz, + batch_size // WORLD_SIZE, + gs, + single_cls, + hyp=hyp, + augment=True, + cache=None if opt.cache == 'val' else opt.cache, + rect=opt.rect, + rank=LOCAL_RANK, + workers=workers, + image_weights=opt.image_weights, + quad=opt.quad, + prefix=colorstr('train: '), + shuffle=True) + labels = np.concatenate(dataset.labels, 0) + mlc = int(labels[:, 0].max()) # max label class + assert mlc < nc, f'Label class {mlc} exceeds nc={nc} in {data}. Possible class labels are 0-{nc - 1}' + + # Process 0 + if RANK in {-1, 0}: + val_loader = create_dataloader(val_path, + imgsz, + batch_size // WORLD_SIZE * 2, + gs, + single_cls, + hyp=hyp, + cache=None if noval else opt.cache, + rect=True, + rank=-1, + workers=workers * 2, + pad=0.5, + prefix=colorstr('val: '))[0] + + if not resume: + if plots: + plot_labels(labels, names, save_dir) + + # Anchors + if not opt.noautoanchor: + check_anchors(dataset, model=model, thr=hyp['anchor_t'], imgsz=imgsz) + model.half().float() # pre-reduce anchor precision + + callbacks.run('on_pretrain_routine_end') + + # DDP mode + if cuda and RANK != -1: + model = smart_DDP(model) + + # Model attributes + nl = de_parallel(model).model[-1].nl # number of detection layers (to scale hyps) + hyp['box'] *= 3 / nl # scale to layers + hyp['cls'] *= nc / 80 * 3 / nl # scale to classes and layers + hyp['obj'] *= (imgsz / 640) ** 2 * 3 / nl # scale to image size and layers + hyp['label_smoothing'] = opt.label_smoothing + model.nc = nc # attach number of classes to model + model.hyp = hyp # attach hyperparameters to model + model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc # attach class weights + model.names = names + + # Start training + t0 = time.time() + nb = len(train_loader) # number of batches + nw = max(round(hyp['warmup_epochs'] * nb), 100) # number of warmup iterations, max(3 epochs, 100 iterations) + # nw = min(nw, (epochs - start_epoch) / 2 * nb) # limit warmup to < 1/2 of training + last_opt_step = -1 + maps = np.zeros(nc) # mAP per class + results = (0, 0, 0, 0, 0, 0, 0) # P, R, mAP@.5, mAP@.5-.95, val_loss(box, obj, cls) + scheduler.last_epoch = start_epoch - 1 # do not move + scaler = torch.cuda.amp.GradScaler(enabled=amp) + stopper, stop = EarlyStopping(patience=opt.patience), False + compute_loss = ComputeLoss(model) # init loss class + callbacks.run('on_train_start') + LOGGER.info(f'Image sizes {imgsz} train, {imgsz} val\n' + f'Using {train_loader.num_workers * WORLD_SIZE} dataloader workers\n' + f"Logging results to {colorstr('bold', save_dir)}\n" + f'Starting training for {epochs} epochs...') + + model_name = opt.cfg.split('/')[-1].split('.')[0] + output_dir = os.path.join(opt.output_path, model_name) + os.system('rm -rf {};mkdir -p {}'.format(output_dir, output_dir)) + if opt.pre_eval_and_export: + import copy + print('原始onnx模型精度') + results, maps, _ = val.run(data_dict, + batch_size=batch_size // WORLD_SIZE * 2, + imgsz=imgsz, + half=amp, + model=ema.ema, + single_cls=single_cls, + dataloader=val_loader, + save_dir=save_dir, + plots=False, + callbacks=callbacks, + compute_loss=compute_loss) + kwargs = { + 'input_shape_dict': {'data': [1, 3, opt.imgsz, opt.imgsz]}, + 'output_path': output_dir, + 'model_name': model_name, + 'dummy_input': None, + 'onnx_model_path': os.path.join(output_dir, '{}_ori.onnx'.format(model_name)), + } + module_tmp = copy.deepcopy(model) + module_tmp = module_tmp.cpu() + convert_onnx(module_tmp.eval(), **kwargs) + del module_tmp + model = model.train() #prepare前一定要是train模式！！ + # exit(0) + + backend = BackendType.Sophgo_TPU + if opt.quantize: + prepare_custom_config_dict= { + # 'work_mode':'int4_and_int8_mix', + # 'extra_qconfig_dict':{'w_fakequantize':'PACTFakeQuantize'} + # 'concrete_args':{'augment':False, 'profile':False, 'visualize':False} + } + + # print('named_modules:', dict(model.named_modules())['']) + model.train() + model = model.to(device) + model = prepare_by_platform(model, backend, prepare_custom_config_dict) + # print('prepared module:', model) + enable_calibration(model) + calibration_flag = True + model = model.to(device) + + for epoch in range(start_epoch, epochs): # epoch ------------------------------------------------------------------ + callbacks.run('on_train_epoch_start') + model.train() + + # Update image weights (optional, single-GPU only) + if opt.image_weights: + cw = model.class_weights.cpu().numpy() * (1 - maps) ** 2 / nc # class weights + iw = labels_to_image_weights(dataset.labels, nc=nc, class_weights=cw) # image weights + dataset.indices = random.choices(range(dataset.n), weights=iw, k=dataset.n) # rand weighted idx + + # Update mosaic border (optional) + # b = int(random.uniform(0.25 * imgsz, 0.75 * imgsz + gs) // gs * gs) + # dataset.mosaic_border = [b - imgsz, -b] # height, width borders + + mloss = torch.zeros(3, device=device) # mean losses + if RANK != -1: + train_loader.sampler.set_epoch(epoch) + pbar = enumerate(train_loader) + LOGGER.info(('\n' + '%10s' * 7) % ('Epoch', 'gpu_mem', 'box', 'obj', 'cls', 'labels', 'img_size')) + if RANK in {-1, 0}: + pbar = tqdm(pbar, total=nb, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar + sample_size = nb//1000 + print('sample_size:', sample_size) + optimizer.zero_grad() + for i, (imgs, targets, paths, _) in pbar: # batch ------------------------------------------------------------- + if opt.fast_test and i % sample_size != 0: + continue + callbacks.run('on_train_batch_start') + ni = i + nb * epoch # number integrated batches (since train start) + imgs = imgs.to(device, non_blocking=True).float() / 255 # uint8 to float32, 0-255 to 0.0-1.0 + + # Warmup + if ni <= nw: + xi = [0, nw] # x interp + # compute_loss.gr = np.interp(ni, xi, [0.0, 1.0]) # iou loss ratio (obj_loss = 1.0 or iou) + accumulate = max(1, np.interp(ni, xi, [1, nbs / batch_size]).round()) + for j, x in enumerate(optimizer.param_groups): + # bias lr falls from 0.1 to lr0, all other lrs rise from 0.0 to lr0 + x['lr'] = np.interp(ni, xi, [hyp['warmup_bias_lr'] if j == 0 else 0.0, x['initial_lr'] * lf(epoch)]) + if 'momentum' in x: + x['momentum'] = np.interp(ni, xi, [hyp['warmup_momentum'], hyp['momentum']]) + + # Multi-scale + if opt.multi_scale: + sz = random.randrange(imgsz * 0.5, imgsz * 1.5 + gs) // gs * gs # size + sf = sz / max(imgs.shape[2:]) # scale factor + if sf != 1: + ns = [math.ceil(x * sf / gs) * gs for x in imgs.shape[2:]] # new shape (stretched to gs-multiple) + imgs = nn.functional.interpolate(imgs, size=ns, mode='bilinear', align_corners=False) + + # Forward + with torch.cuda.amp.autocast(amp): + pred = model(imgs) # forward + loss, loss_items = compute_loss(pred, targets.to(device)) # loss scaled by batch_size + # print(i, 'loss:', loss) + if RANK != -1: + loss *= WORLD_SIZE # gradient averaged between devices in DDP mode + if opt.quad: + loss *= 4. + + if opt.quantize: + if calibration_flag: + if i >= 50: + calibration_flag = False + model.zero_grad() + enable_quantization(model) + print('close calibration') + else: + print('calibration iter{}'.format(i)) + continue + + # Backward + scaler.scale(loss).backward() + + # Optimize - https://pytorch.org/docs/master/notes/amp_examples.html + if ni - last_opt_step >= accumulate: + scaler.unscale_(optimizer) # unscale gradients + torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=10.0) # clip gradients + scaler.step(optimizer) # optimizer.step + scaler.update() + optimizer.zero_grad() + if ema: + ema.update(model) + last_opt_step = ni + + # Log + if RANK in {-1, 0}: + mloss = (mloss * i + loss_items) / (i + 1) # update mean losses + mem = f'{torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0:.3g}G' # (GB) + pbar.set_description(('%10s' * 2 + '%10.4g' * 5) % + (f'{epoch}/{epochs - 1}', mem, *mloss, targets.shape[0], imgs.shape[-1])) + callbacks.run('on_train_batch_end', ni, model, imgs, targets, paths, plots) + if callbacks.stop_training: + return + # end batch ------------------------------------------------------------------------------------------------ + + # Scheduler + lr = [x['lr'] for x in optimizer.param_groups] # for loggers + scheduler.step() + + if RANK in {-1, 0}: + # mAP + callbacks.run('on_train_epoch_end', epoch=epoch) + ema.update_attr(model, include=['yaml', 'nc', 'hyp', 'names', 'stride', 'class_weights']) + final_epoch = (epoch + 1 == epochs) or stopper.possible_stop + if not noval or final_epoch: # Calculate mAP + results, maps, _ = val.run(data_dict, + batch_size=batch_size // WORLD_SIZE * 2, + imgsz=imgsz, + half=amp, + model=ema.ema, + single_cls=single_cls, + dataloader=val_loader, + save_dir=save_dir, + plots=False, + callbacks=callbacks, + compute_loss=compute_loss) + + # Update best mAP + fi = fitness(np.array(results).reshape(1, -1)) # weighted combination of [P, R, mAP@.5, mAP@.5-.95] + stop = stopper(epoch=epoch, fitness=fi) # early stop check + if fi > best_fitness: + best_fitness = fi + log_vals = list(mloss) + list(results) + lr + callbacks.run('on_fit_epoch_end', log_vals, epoch, best_fitness, fi) + + # Save model + # if (not nosave) or (final_epoch and not evolve): # if save + # ckpt = { + # 'epoch': epoch, + # 'best_fitness': best_fitness, + # 'model': deepcopy(de_parallel(model)).half(), + # 'ema': deepcopy(ema.ema).half(), + # 'updates': ema.updates, + # 'optimizer': optimizer.state_dict(), + # 'wandb_id': loggers.wandb.wandb_run.id if loggers.wandb else None, + # 'opt': vars(opt), + # 'date': datetime.now().isoformat()} + + # # Save last, best and delete + # torch.save(ckpt, last) #wangxuechuan 23.04.12, Can't pickle local object 'add_module_to_qconfig_obs_ctr..get_factory_kwargs_based_on_module_device' + # if best_fitness == fi: + # torch.save(ckpt, best) + # if opt.save_period > 0 and epoch % opt.save_period == 0: + # torch.save(ckpt, w / f'epoch{epoch}.pt') + # del ckpt + # callbacks.run('on_model_save', last, epoch, final_epoch, best_fitness, fi) + + # EarlyStopping + if RANK != -1: # if DDP training + broadcast_list = [stop if RANK == 0 else None] + dist.broadcast_object_list(broadcast_list, 0) # broadcast 'stop' to all ranks + if RANK != 0: + stop = broadcast_list[0] + if stop: + break # must break all DDP ranks + + if opt.quantize: + print(f'epoch{epoch} convert_deploy') + model_name = opt.cfg.split('/')[-1].split('.')[0] + output_dir = os.path.join(opt.output_path, model_name) + output_dir = os.path.join(output_dir, str(epoch)) + output_dir = os.path.join(output_dir, model_name) + os.system('mkdir -p {}'.format(output_dir)) + model2 = deepcopy(model) + convert_deploy(model2.eval(), backend, input_shape_dict={'data': [1, 3, opt.imgsz, opt.imgsz]}, + model_name='{}_mqmoble'.format(model_name), output_path=output_dir) + del model2 + + # end epoch ---------------------------------------------------------------------------------------------------- + # end training ----------------------------------------------------------------------------------------------------- + + if not opt.fast_test and RANK in {-1, 0}: + LOGGER.info(f'\n{epoch - start_epoch + 1} epochs completed in {(time.time() - t0) / 3600:.3f} hours.') + for f in last, best: + if f.exists(): + strip_optimizer(f) # strip optimizers + if f is best: + LOGGER.info(f'\nValidating {f}...') + results, _, _ = val.run( + data_dict, + batch_size=batch_size // WORLD_SIZE * 2, + imgsz=imgsz, + model=attempt_load(f, device).half(), + iou_thres=0.65 if is_coco else 0.60, # best pycocotools results at 0.65 + single_cls=single_cls, + dataloader=val_loader, + save_dir=save_dir, + save_json=is_coco, + verbose=True, + plots=plots, + callbacks=callbacks, + compute_loss=compute_loss) # val best model with plots + if is_coco: + callbacks.run('on_fit_epoch_end', list(mloss) + list(results) + lr, epoch, best_fitness, fi) + + callbacks.run('on_train_end', last, best, plots, epoch, results) + + if opt.quantize: + model_name = opt.cfg.split('/')[-1].split('.')[0] + output_dir = os.path.join(opt.output_path, model_name) + os.system('mkdir -p {}'.format(output_dir)) + convert_deploy(model.eval(), backend, input_shape_dict={'data': [1, 3, opt.imgsz, opt.imgsz]}, + model_name='{}_mqmoble'.format(model_name), output_path=output_dir) + + torch.cuda.empty_cache() + return results + + +def parse_opt(known=False): + parser = argparse.ArgumentParser() + parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='initial weights path') + parser.add_argument('--cfg', type=str, default='', help='model.yaml path') + parser.add_argument('--output_path', type=str, default='./', help='output path') + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path') + parser.add_argument('--hyp', type=str, default=ROOT / 'data/hyps/hyp.scratch-low.yaml', help='hyperparameters path') + parser.add_argument('--epochs', type=int, default=300) + parser.add_argument('--batch-size', type=int, default=16, help='total batch size for all GPUs, -1 for autobatch') + parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='train, val image size (pixels)') + parser.add_argument('--rect', action='store_true', help='rectangular training') + parser.add_argument('--resume', nargs='?', const=True, default=False, help='resume most recent training') + parser.add_argument('--nosave', action='store_true', help='only save final checkpoint') + parser.add_argument('--noval', action='store_true', help='only validate final epoch') + parser.add_argument('--noautoanchor', action='store_true', help='disable AutoAnchor') + parser.add_argument('--noplots', action='store_true', help='save no plot files') + parser.add_argument('--evolve', type=int, nargs='?', const=300, help='evolve hyperparameters for x generations') + parser.add_argument('--bucket', type=str, default='', help='gsutil bucket') + parser.add_argument('--cache', type=str, nargs='?', const='ram', help='--cache images in "ram" (default) or "disk"') + parser.add_argument('--image-weights', action='store_true', help='use weighted image selection for training') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--multi-scale', action='store_true', help='vary img-size +/- 50%%') + parser.add_argument('--single-cls', action='store_true', help='train multi-class data as single-class') + parser.add_argument('--optimizer', type=str, choices=['SGD', 'Adam', 'AdamW'], default='SGD', help='optimizer') + parser.add_argument('--sync-bn', action='store_true', help='use SyncBatchNorm, only available in DDP mode') + parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)') + parser.add_argument('--project', default=ROOT / 'runs/train', help='save to project/name') + parser.add_argument('--name', default='exp', help='save to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--quad', action='store_true', help='quad dataloader') + parser.add_argument('--cos-lr', action='store_true', help='cosine LR scheduler') + parser.add_argument('--label-smoothing', type=float, default=0.0, help='Label smoothing epsilon') + parser.add_argument('--patience', type=int, default=100, help='EarlyStopping patience (epochs without improvement)') + parser.add_argument('--freeze', nargs='+', type=int, default=[0], help='Freeze layers: backbone=10, first3=0 1 2') + parser.add_argument('--save-period', type=int, default=-1, help='Save checkpoint every x epochs (disabled if < 1)') + parser.add_argument('--seed', type=int, default=0, help='Global training seed') + parser.add_argument('--local_rank', type=int, default=-1, help='Automatic DDP Multi-GPU argument, do not modify') + + # Weights & Biases arguments + parser.add_argument('--entity', default=None, help='W&B: Entity') + parser.add_argument('--upload_dataset', nargs='?', const=True, default=False, help='W&B: Upload data, "val" option') + parser.add_argument('--bbox_interval', type=int, default=-1, help='W&B: Set bounding-box image logging interval') + parser.add_argument('--artifact_alias', type=str, default='latest', help='W&B: Version of dataset artifact to use') + parser.add_argument('--quantize', action='store_true', help='quantize') + parser.add_argument('--pre_eval_and_export', action='store_true', help='pre_eval_and_export') + parser.add_argument('--fast_test', action='store_true', help='fast_test') + + return parser.parse_known_args()[0] if known else parser.parse_args() + + +def main(opt, callbacks=Callbacks()): + # Checks + if RANK in {-1, 0}: + print_args(vars(opt)) + check_git_status() + check_requirements() + + # Resume + if opt.resume and not (check_wandb_resume(opt) or opt.evolve): # resume from specified or most recent last.pt + last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run()) + opt_yaml = last.parent.parent / 'opt.yaml' # train options yaml + opt_data = opt.data # original dataset + if opt_yaml.is_file(): + with open(opt_yaml, errors='ignore') as f: + d = yaml.safe_load(f) + else: + d = torch.load(last, map_location='cpu')['opt'] + opt = argparse.Namespace(**d) # replace + opt.cfg, opt.weights, opt.resume = '', str(last), True # reinstate + if is_url(opt_data): + opt.data = check_file(opt_data) # avoid HUB resume auth timeout + else: + opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = \ + check_file(opt.data), check_yaml(opt.cfg), check_yaml(opt.hyp), str(opt.weights), str(opt.project) # checks + assert len(opt.cfg) or len(opt.weights), 'either --cfg or --weights must be specified' + if opt.evolve: + if opt.project == str(ROOT / 'runs/train'): # if default project name, rename to runs/evolve + opt.project = str(ROOT / 'runs/evolve') + opt.exist_ok, opt.resume = opt.resume, False # pass resume to exist_ok and disable resume + if opt.name == 'cfg': + opt.name = Path(opt.cfg).stem # use model.yaml as name + opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok)) + + # DDP mode + device = select_device(opt.device, batch_size=opt.batch_size) + if LOCAL_RANK != -1: + msg = 'is not compatible with YOLOv5 Multi-GPU DDP training' + assert not opt.image_weights, f'--image-weights {msg}' + assert not opt.evolve, f'--evolve {msg}' + assert opt.batch_size != -1, f'AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size' + assert opt.batch_size % WORLD_SIZE == 0, f'--batch-size {opt.batch_size} must be multiple of WORLD_SIZE' + assert torch.cuda.device_count() > LOCAL_RANK, 'insufficient CUDA devices for DDP command' + torch.cuda.set_device(LOCAL_RANK) + device = torch.device('cuda', LOCAL_RANK) + dist.init_process_group(backend="nccl" if dist.is_nccl_available() else "gloo") + + # Train + if not opt.evolve: + train(opt.hyp, opt, device, callbacks) + + # Evolve hyperparameters (optional) + else: + # Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit) + meta = { + 'lr0': (1, 1e-5, 1e-1), # initial learning rate (SGD=1E-2, Adam=1E-3) + 'lrf': (1, 0.01, 1.0), # final OneCycleLR learning rate (lr0 * lrf) + 'momentum': (0.3, 0.6, 0.98), # SGD momentum/Adam beta1 + 'weight_decay': (1, 0.0, 0.001), # optimizer weight decay + 'warmup_epochs': (1, 0.0, 5.0), # warmup epochs (fractions ok) + 'warmup_momentum': (1, 0.0, 0.95), # warmup initial momentum + 'warmup_bias_lr': (1, 0.0, 0.2), # warmup initial bias lr + 'box': (1, 0.02, 0.2), # box loss gain + 'cls': (1, 0.2, 4.0), # cls loss gain + 'cls_pw': (1, 0.5, 2.0), # cls BCELoss positive_weight + 'obj': (1, 0.2, 4.0), # obj loss gain (scale with pixels) + 'obj_pw': (1, 0.5, 2.0), # obj BCELoss positive_weight + 'iou_t': (0, 0.1, 0.7), # IoU training threshold + 'anchor_t': (1, 2.0, 8.0), # anchor-multiple threshold + 'anchors': (2, 2.0, 10.0), # anchors per output grid (0 to ignore) + 'fl_gamma': (0, 0.0, 2.0), # focal loss gamma (efficientDet default gamma=1.5) + 'hsv_h': (1, 0.0, 0.1), # image HSV-Hue augmentation (fraction) + 'hsv_s': (1, 0.0, 0.9), # image HSV-Saturation augmentation (fraction) + 'hsv_v': (1, 0.0, 0.9), # image HSV-Value augmentation (fraction) + 'degrees': (1, 0.0, 45.0), # image rotation (+/- deg) + 'translate': (1, 0.0, 0.9), # image translation (+/- fraction) + 'scale': (1, 0.0, 0.9), # image scale (+/- gain) + 'shear': (1, 0.0, 10.0), # image shear (+/- deg) + 'perspective': (0, 0.0, 0.001), # image perspective (+/- fraction), range 0-0.001 + 'flipud': (1, 0.0, 1.0), # image flip up-down (probability) + 'fliplr': (0, 0.0, 1.0), # image flip left-right (probability) + 'mosaic': (1, 0.0, 1.0), # image mixup (probability) + 'mixup': (1, 0.0, 1.0), # image mixup (probability) + 'copy_paste': (1, 0.0, 1.0)} # segment copy-paste (probability) + + with open(opt.hyp, errors='ignore') as f: + hyp = yaml.safe_load(f) # load hyps dict + if 'anchors' not in hyp: # anchors commented in hyp.yaml + hyp['anchors'] = 3 + if opt.noautoanchor: + del hyp['anchors'], meta['anchors'] + opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir) # only val/save final epoch + # ei = [isinstance(x, (int, float)) for x in hyp.values()] # evolvable indices + evolve_yaml, evolve_csv = save_dir / 'hyp_evolve.yaml', save_dir / 'evolve.csv' + if opt.bucket: + os.system(f'gsutil cp gs://{opt.bucket}/evolve.csv {evolve_csv}') # download evolve.csv if exists + + for _ in range(opt.evolve): # generations to evolve + if evolve_csv.exists(): # if evolve.csv exists: select best hyps and mutate + # Select parent(s) + parent = 'single' # parent selection method: 'single' or 'weighted' + x = np.loadtxt(evolve_csv, ndmin=2, delimiter=',', skiprows=1) + n = min(5, len(x)) # number of previous results to consider + x = x[np.argsort(-fitness(x))][:n] # top n mutations + w = fitness(x) - fitness(x).min() + 1E-6 # weights (sum > 0) + if parent == 'single' or len(x) == 1: + # x = x[random.randint(0, n - 1)] # random selection + x = x[random.choices(range(n), weights=w)[0]] # weighted selection + elif parent == 'weighted': + x = (x * w.reshape(n, 1)).sum(0) / w.sum() # weighted combination + + # Mutate + mp, s = 0.8, 0.2 # mutation probability, sigma + npr = np.random + npr.seed(int(time.time())) + g = np.array([meta[k][0] for k in hyp.keys()]) # gains 0-1 + ng = len(meta) + v = np.ones(ng) + while all(v == 1): # mutate until a change occurs (prevent duplicates) + v = (g * (npr.random(ng) < mp) * npr.randn(ng) * npr.random() * s + 1).clip(0.3, 3.0) + for i, k in enumerate(hyp.keys()): # plt.hist(v.ravel(), 300) + hyp[k] = float(x[i + 7] * v[i]) # mutate + + # Constrain to limits + for k, v in meta.items(): + hyp[k] = max(hyp[k], v[1]) # lower limit + hyp[k] = min(hyp[k], v[2]) # upper limit + hyp[k] = round(hyp[k], 5) # significant digits + + # Train mutation + results = train(hyp.copy(), opt, device, callbacks) + callbacks = Callbacks() + # Write mutation results + print_mutation(results, hyp.copy(), save_dir, opt.bucket) + + # Plot results + plot_evolve(evolve_csv) + LOGGER.info(f'Hyperparameter evolution finished {opt.evolve} generations\n' + f"Results saved to {colorstr('bold', save_dir)}\n" + f'Usage example: $ python train.py --hyp {evolve_yaml}') + + +def run(**kwargs): + # Usage: import train; train.run(data='coco128.yaml', imgsz=320, weights='yolov5m.pt') + opt = parse_opt(True) + for k, v in kwargs.items(): + setattr(opt, k, v) + main(opt) + return opt + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/application/yolov5_example/tutorial.ipynb b/application/yolov5_example/tutorial.ipynb new file mode 100644 index 00000000..9fa338b1 --- /dev/null +++ b/application/yolov5_example/tutorial.ipynb @@ -0,0 +1,1151 @@ +{ + "cells": [ + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "view-in-github" + }, + "source": [ + "\"Open" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "t6MPjfT5NrKQ" + }, + "source": [ + "\n", + "\n", + "\n", + "This is the **official YOLOv5 🚀 notebook** by **Ultralytics**, and is freely available for redistribution under the [GPL-3.0 license](https://choosealicense.com/licenses/gpl-3.0/). \n", + "For more information please visit https://github.com/ultralytics/yolov5 and https://ultralytics.com. Thank you!" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "7mGmQbAO5pQb" + }, + "source": [ + "# Setup\n", + "\n", + "Clone repo, install dependencies and check PyTorch and GPU." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/" + }, + "id": "wbvMlHd_QwMG", + "outputId": "185d0979-edcd-4860-e6fb-b8a27dbf5096" + }, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "YOLOv5 🚀 v6.1-370-g20f1b7e Python-3.7.13 torch-1.12.0+cu113 CUDA:0 (Tesla V100-SXM2-16GB, 16160MiB)\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Setup complete ✅ (8 CPUs, 51.0 GB RAM, 37.4/166.8 GB disk)\n" + ] + } + ], + "source": [ + "!git clone https://github.com/ultralytics/yolov5 # clone\n", + "%cd yolov5\n", + "%pip install -qr requirements.txt # install\n", + "\n", + "import torch\n", + "import utils\n", + "display = utils.notebook_init() # checks" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "4JnkELT0cIJg" + }, + "source": [ + "# 1. Inference\n", + "\n", + "`detect.py` runs YOLOv5 inference on a variety of sources, downloading models automatically from the [latest YOLOv5 release](https://github.com/ultralytics/yolov5/releases), and saving results to `runs/detect`. Example inference sources are:\n", + "\n", + "```shell\n", + "python detect.py --source 0 # webcam\n", + " img.jpg # image \n", + " vid.mp4 # video\n", + " path/ # directory\n", + " 'path/*.jpg' # glob\n", + " 'https://youtu.be/Zgi9g1ksQHc' # YouTube\n", + " 'rtsp://example.com/media.mp4' # RTSP, RTMP, HTTP stream\n", + "```" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/" + }, + "id": "zR9ZbuQCH7FX", + "outputId": "4b13989f-32a4-4ef0-b403-06ff3aac255c" + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\u001b[34m\u001b[1mdetect: \u001b[0mweights=['yolov5s.pt'], source=data/images, data=data/coco128.yaml, imgsz=[640, 640], conf_thres=0.25, iou_thres=0.45, max_det=1000, device=, view_img=False, save_txt=False, save_conf=False, save_crop=False, nosave=False, classes=None, agnostic_nms=False, augment=False, visualize=False, update=False, project=runs/detect, name=exp, exist_ok=False, line_thickness=3, hide_labels=False, hide_conf=False, half=False, dnn=False\n", + "YOLOv5 🚀 v6.1-370-g20f1b7e Python-3.7.13 torch-1.12.0+cu113 CUDA:0 (Tesla V100-SXM2-16GB, 16160MiB)\n", + "\n", + "Downloading https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5s.pt to yolov5s.pt...\n", + "100% 14.1M/14.1M [00:00<00:00, 53.9MB/s]\n", + "\n", + "Fusing layers... \n", + "YOLOv5s summary: 213 layers, 7225885 parameters, 0 gradients\n", + "image 1/2 /content/yolov5/data/images/bus.jpg: 640x480 4 persons, 1 bus, Done. (0.016s)\n", + "image 2/2 /content/yolov5/data/images/zidane.jpg: 384x640 2 persons, 2 ties, Done. (0.021s)\n", + "Speed: 0.6ms pre-process, 18.6ms inference, 25.0ms NMS per image at shape (1, 3, 640, 640)\n", + "Results saved to \u001b[1mruns/detect/exp\u001b[0m\n" + ] + } + ], + "source": [ + "!python detect.py --weights yolov5s.pt --img 640 --conf 0.25 --source data/images\n", + "#display.Image(filename='runs/detect/exp/zidane.jpg', width=600)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "hkAzDWJ7cWTr" + }, + "source": [ + "        \n", + "" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "0eq1SMWl6Sfn" + }, + "source": [ + "# 2. Validate\n", + "Validate a model's accuracy on [COCO](https://cocodataset.org/#home) val or test-dev datasets. Models are downloaded automatically from the [latest YOLOv5 release](https://github.com/ultralytics/yolov5/releases). To show results by class use the `--verbose` flag. Note that `pycocotools` metrics may be ~1% better than the equivalent repo metrics, as is visible below, due to slight differences in mAP computation." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "eyTZYGgRjnMc" + }, + "source": [ + "## COCO val\n", + "Download [COCO val 2017](https://github.com/ultralytics/yolov5/blob/74b34872fdf41941cddcf243951cdb090fbac17b/data/coco.yaml#L14) dataset (1GB - 5000 images), and test model accuracy." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/", + "height": 49, + "referenced_widgets": [ + "c31d2039ccf74c22b67841f4877d1186", + "d4bba1727c714d94ad58a72bffa07c4c", + "9aeff9f1780b45f892422fdc96e56913", + "bf55a7c71d074d3fa88b10b997820825", + "d8b66044e2fb4f5b916696834d880c81", + "102e1deda239436fa72751c58202fa0f", + "4fd4431ced6c42368e18424912b877e4", + "cdd709c4f40941bea1b2053523c9fac8", + "a1ef2d8de2b741c78ca5d938e2ddbcdf", + "0dbce99bb6184238842cbec0587d564a", + "91ff5f93f2a24c5790ab29e347965946" + ] + }, + "id": "WQPtK1QYVaD_", + "outputId": "a9004b06-37a6-41ed-a1f2-ac956f3963b3" + }, + "outputs": [ + { + "data": { + "application/vnd.jupyter.widget-view+json": { + "model_id": "c31d2039ccf74c22b67841f4877d1186", + "version_major": 2, + "version_minor": 0 + }, + "text/plain": [ + " 0%| | 0.00/780M [00:00

\n", + "Close the active learning loop by sampling images from your inference conditions with the `roboflow` pip package\n", + "

\n", + "\n", + "Train a YOLOv5s model on the [COCO128](https://www.kaggle.com/ultralytics/coco128) dataset with `--data coco128.yaml`, starting from pretrained `--weights yolov5s.pt`, or from randomly initialized `--weights '' --cfg yolov5s.yaml`.\n", + "\n", + "- **Pretrained [Models](https://github.com/ultralytics/yolov5/tree/master/models)** are downloaded\n", + "automatically from the [latest YOLOv5 release](https://github.com/ultralytics/yolov5/releases)\n", + "- **[Datasets](https://github.com/ultralytics/yolov5/tree/master/data)** available for autodownload include: [COCO](https://github.com/ultralytics/yolov5/blob/master/data/coco.yaml), [COCO128](https://github.com/ultralytics/yolov5/blob/master/data/coco128.yaml), [VOC](https://github.com/ultralytics/yolov5/blob/master/data/VOC.yaml), [Argoverse](https://github.com/ultralytics/yolov5/blob/master/data/Argoverse.yaml), [VisDrone](https://github.com/ultralytics/yolov5/blob/master/data/VisDrone.yaml), [GlobalWheat](https://github.com/ultralytics/yolov5/blob/master/data/GlobalWheat2020.yaml), [xView](https://github.com/ultralytics/yolov5/blob/master/data/xView.yaml), [Objects365](https://github.com/ultralytics/yolov5/blob/master/data/Objects365.yaml), [SKU-110K](https://github.com/ultralytics/yolov5/blob/master/data/SKU-110K.yaml).\n", + "- **Training Results** are saved to `runs/train/` with incrementing run directories, i.e. `runs/train/exp2`, `runs/train/exp3` etc.\n", + "

\n", + "\n", + "A **Mosaic Dataloader** is used for training which combines 4 images into 1 mosaic.\n", + "\n", + "## Train on Custom Data with Roboflow 🌟 NEW\n", + "\n", + "[Roboflow](https://roboflow.com/?ref=ultralytics) enables you to easily **organize, label, and prepare** a high quality dataset with your own custom data. Roboflow also makes it easy to establish an active learning pipeline, collaborate with your team on dataset improvement, and integrate directly into your model building workflow with the `roboflow` pip package.\n", + "\n", + "- Custom Training Example: [https://blog.roboflow.com/how-to-train-yolov5-on-a-custom-dataset/](https://blog.roboflow.com/how-to-train-yolov5-on-a-custom-dataset/?ref=ultralytics)\n", + "- Custom Training Notebook: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/roboflow-ai/yolov5-custom-training-tutorial/blob/main/yolov5-custom-training.ipynb)\n", + "
\n", + "\n", + "

Label images lightning fast (including with model-assisted labeling)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "bOy5KI2ncnWd" + }, + "outputs": [], + "source": [ + "# Tensorboard (optional)\n", + "%load_ext tensorboard\n", + "%tensorboard --logdir runs/train" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "DQhI6vvaRWjR" + }, + "outputs": [], + "source": [ + "# ClearML (optional)\n", + "%pip install -q clearml\n", + "!clearml-init" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "2fLAV42oNb7M" + }, + "outputs": [], + "source": [ + "# Weights & Biases (optional)\n", + "%pip install -q wandb\n", + "import wandb\n", + "wandb.login()" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "colab": { + "base_uri": "https://localhost:8080/" + }, + "id": "1NcFxRcFdJ_O", + "outputId": "bce1b4bd-1a14-4c07-aebd-6c11e91ad24b" + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\u001b[34m\u001b[1mtrain: \u001b[0mweights=yolov5s.pt, cfg=, data=coco128.yaml, hyp=data/hyps/hyp.scratch-low.yaml, epochs=3, batch_size=16, imgsz=640, rect=False, resume=False, nosave=False, noval=False, noautoanchor=False, noplots=False, evolve=None, bucket=, cache=ram, image_weights=False, device=, multi_scale=False, single_cls=False, optimizer=SGD, sync_bn=False, workers=8, project=runs/train, name=exp, exist_ok=False, quad=False, cos_lr=False, label_smoothing=0.0, patience=100, freeze=[0], save_period=-1, seed=0, local_rank=-1, entity=None, upload_dataset=False, bbox_interval=-1, artifact_alias=latest\n", + "\u001b[34m\u001b[1mgithub: \u001b[0mup to date with https://github.com/ultralytics/yolov5 ✅\n", + "YOLOv5 🚀 v6.1-370-g20f1b7e Python-3.7.13 torch-1.12.0+cu113 CUDA:0 (Tesla V100-SXM2-16GB, 16160MiB)\n", + "\n", + "\u001b[34m\u001b[1mhyperparameters: \u001b[0mlr0=0.01, lrf=0.01, momentum=0.937, weight_decay=0.0005, warmup_epochs=3.0, warmup_momentum=0.8, warmup_bias_lr=0.1, box=0.05, cls=0.5, cls_pw=1.0, obj=1.0, obj_pw=1.0, iou_t=0.2, anchor_t=4.0, fl_gamma=0.0, hsv_h=0.015, hsv_s=0.7, hsv_v=0.4, degrees=0.0, translate=0.1, scale=0.5, shear=0.0, perspective=0.0, flipud=0.0, fliplr=0.5, mosaic=1.0, mixup=0.0, copy_paste=0.0\n", + "\u001b[34m\u001b[1mWeights & Biases: \u001b[0mrun 'pip install wandb' to automatically track and visualize YOLOv5 🚀 runs in Weights & Biases\n", + "\u001b[34m\u001b[1mClearML: \u001b[0mrun 'pip install clearml' to automatically track, visualize and remotely train YOLOv5 🚀 runs in ClearML\n", + "\u001b[34m\u001b[1mTensorBoard: \u001b[0mStart with 'tensorboard --logdir runs/train', view at http://localhost:6006/\n", + "\n", + "Dataset not found ⚠️, missing paths ['/content/datasets/coco128/images/train2017']\n", + "Downloading https://ultralytics.com/assets/coco128.zip to coco128.zip...\n", + "100% 6.66M/6.66M [00:00<00:00, 75.2MB/s]\n", + "Dataset download success ✅ (0.7s), saved to \u001b[1m/content/datasets\u001b[0m\n", + "\n", + " from n params module arguments \n", + " 0 -1 1 3520 models.common.Conv [3, 32, 6, 2, 2] \n", + " 1 -1 1 18560 models.common.Conv [32, 64, 3, 2] \n", + " 2 -1 1 18816 models.common.C3 [64, 64, 1] \n", + " 3 -1 1 73984 models.common.Conv [64, 128, 3, 2] \n", + " 4 -1 2 115712 models.common.C3 [128, 128, 2] \n", + " 5 -1 1 295424 models.common.Conv [128, 256, 3, 2] \n", + " 6 -1 3 625152 models.common.C3 [256, 256, 3] \n", + " 7 -1 1 1180672 models.common.Conv [256, 512, 3, 2] \n", + " 8 -1 1 1182720 models.common.C3 [512, 512, 1] \n", + " 9 -1 1 656896 models.common.SPPF [512, 512, 5] \n", + " 10 -1 1 131584 models.common.Conv [512, 256, 1, 1] \n", + " 11 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] \n", + " 12 [-1, 6] 1 0 models.common.Concat [1] \n", + " 13 -1 1 361984 models.common.C3 [512, 256, 1, False] \n", + " 14 -1 1 33024 models.common.Conv [256, 128, 1, 1] \n", + " 15 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest'] \n", + " 16 [-1, 4] 1 0 models.common.Concat [1] \n", + " 17 -1 1 90880 models.common.C3 [256, 128, 1, False] \n", + " 18 -1 1 147712 models.common.Conv [128, 128, 3, 2] \n", + " 19 [-1, 14] 1 0 models.common.Concat [1] \n", + " 20 -1 1 296448 models.common.C3 [256, 256, 1, False] \n", + " 21 -1 1 590336 models.common.Conv [256, 256, 3, 2] \n", + " 22 [-1, 10] 1 0 models.common.Concat [1] \n", + " 23 -1 1 1182720 models.common.C3 [512, 512, 1, False] \n", + " 24 [17, 20, 23] 1 229245 models.yolo.Detect [80, [[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]], [128, 256, 512]]\n", + "Model summary: 270 layers, 7235389 parameters, 7235389 gradients, 16.6 GFLOPs\n", + "\n", + "Transferred 349/349 items from yolov5s.pt\n", + "\u001b[34m\u001b[1mAMP: \u001b[0mchecks passed ✅\n", + "\u001b[34m\u001b[1moptimizer:\u001b[0m SGD(lr=0.01) with parameter groups 57 weight(decay=0.0), 60 weight(decay=0.0005), 60 bias\n", + "\u001b[34m\u001b[1malbumentations: \u001b[0mBlur(always_apply=False, p=0.01, blur_limit=(3, 7)), MedianBlur(always_apply=False, p=0.01, blur_limit=(3, 7)), ToGray(always_apply=False, p=0.01), CLAHE(always_apply=False, p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8))\n", + "\u001b[34m\u001b[1mtrain: \u001b[0mScanning '/content/datasets/coco128/labels/train2017' images and labels...128 found, 0 missing, 2 empty, 0 corrupt: 100% 128/128 [00:00<00:00, 7926.40it/s]\n", + "\u001b[34m\u001b[1mtrain: \u001b[0mNew cache created: /content/datasets/coco128/labels/train2017.cache\n", + "\u001b[34m\u001b[1mtrain: \u001b[0mCaching images (0.1GB ram): 100% 128/128 [00:00<00:00, 975.81it/s]\n", + "\u001b[34m\u001b[1mval: \u001b[0mScanning '/content/datasets/coco128/labels/train2017.cache' images and labels... 128 found, 0 missing, 2 empty, 0 corrupt: 100% 128/128 [00:00\n", + "\"ClearML" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "DLI1JmHU7B0l" + }, + "source": [ + "## Weights & Biases Logging\n", + "\n", + "[Weights & Biases](https://wandb.ai/site?utm_campaign=repo_yolo_notebook) (W&B) is integrated with YOLOv5 for real-time visualization and cloud logging of training runs. This allows for better run comparison and introspection, as well improved visibility and collaboration for teams. To enable W&B `pip install wandb`, and then train normally (you will be guided through setup on first use). \n", + "\n", + "During training you will see live updates at [https://wandb.ai/home](https://wandb.ai/home?utm_campaign=repo_yolo_notebook), and you can create and share detailed [Reports](https://wandb.ai/glenn-jocher/yolov5_tutorial/reports/YOLOv5-COCO128-Tutorial-Results--VmlldzozMDI5OTY) of your results. For more information see the [YOLOv5 Weights & Biases Tutorial](https://github.com/ultralytics/yolov5/issues/1289). \n", + "\n", + "\n", + "\"Weights" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "-WPvRbS5Swl6" + }, + "source": [ + "## Local Logging\n", + "\n", + "Training results are automatically logged with [Tensorboard](https://www.tensorflow.org/tensorboard) and [CSV](https://github.com/ultralytics/yolov5/pull/4148) loggers to `runs/train`, with a new experiment directory created for each new training as `runs/train/exp2`, `runs/train/exp3`, etc.\n", + "\n", + "This directory contains train and val statistics, mosaics, labels, predictions and augmentated mosaics, as well as metrics and charts including precision-recall (PR) curves and confusion matrices. \n", + "\n", + "\"Local\n" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "Zelyeqbyt3GD" + }, + "source": [ + "# Environments\n", + "\n", + "YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including [CUDA](https://developer.nvidia.com/cuda)/[CUDNN](https://developer.nvidia.com/cudnn), [Python](https://www.python.org/) and [PyTorch](https://pytorch.org/) preinstalled):\n", + "\n", + "- **Google Colab and Kaggle** notebooks with free GPU: \"Open \"Open\n", + "- **Google Cloud** Deep Learning VM. See [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart)\n", + "- **Amazon** Deep Learning AMI. See [AWS Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/AWS-Quickstart)\n", + "- **Docker Image**. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) \"Docker\n" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "6Qu7Iesl0p54" + }, + "source": [ + "# Status\n", + "\n", + "![CI CPU testing](https://github.com/ultralytics/yolov5/workflows/CI%20CPU%20testing/badge.svg)\n", + "\n", + "If this badge is green, all [YOLOv5 GitHub Actions](https://github.com/ultralytics/yolov5/actions) Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training ([train.py](https://github.com/ultralytics/yolov5/blob/master/train.py)), testing ([val.py](https://github.com/ultralytics/yolov5/blob/master/val.py)), inference ([detect.py](https://github.com/ultralytics/yolov5/blob/master/detect.py)) and export ([export.py](https://github.com/ultralytics/yolov5/blob/master/export.py)) on macOS, Windows, and Ubuntu every 24 hours and on every commit.\n" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "id": "IEijrePND_2I" + }, + "source": [ + "# Appendix\n", + "\n", + "Additional content below for PyTorch Hub, CI, reproducing results, profiling speeds, VOC training, classification training and TensorRT example." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "GMusP4OAxFu6" + }, + "outputs": [], + "source": [ + "import torch\n", + "\n", + "# PyTorch Hub Model\n", + "model = torch.hub.load('ultralytics/yolov5', 'yolov5s') # or yolov5n - yolov5x6, custom\n", + "\n", + "# Images\n", + "img = 'https://ultralytics.com/images/zidane.jpg' # or file, Path, PIL, OpenCV, numpy, list\n", + "\n", + "# Inference\n", + "results = model(img)\n", + "\n", + "# Results\n", + "results.print() # or .show(), .save(), .crop(), .pandas(), etc." + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "FGH0ZjkGjejy" + }, + "outputs": [], + "source": [ + "# YOLOv5 CI\n", + "%%shell\n", + "rm -rf runs # remove runs/\n", + "m=yolov5n # official weights\n", + "b=runs/train/exp/weights/best # best.pt checkpoint\n", + "python train.py --imgsz 64 --batch 32 --weights $m.pt --cfg $m.yaml --epochs 1 --device 0 # train\n", + "for d in 0 cpu; do # devices\n", + " for w in $m $b; do # weights\n", + " python val.py --imgsz 64 --batch 32 --weights $w.pt --device $d # val\n", + " python detect.py --imgsz 64 --weights $w.pt --device $d # detect\n", + " done\n", + "done\n", + "python hubconf.py --model $m # hub\n", + "python models/tf.py --weights $m.pt # build TF model\n", + "python models/yolo.py --cfg $m.yaml # build PyTorch model\n", + "python export.py --weights $m.pt --img 64 --include torchscript # export" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "mcKoSIK2WSzj" + }, + "outputs": [], + "source": [ + "# Reproduce\n", + "for x in (f'yolov5{x}' for x in 'nsmlx'):\n", + " !python val.py --weights {x}.pt --data coco.yaml --img 640 --task speed # speed\n", + " !python val.py --weights {x}.pt --data coco.yaml --img 640 --conf 0.001 --iou 0.65 # mAP" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "gogI-kwi3Tye" + }, + "outputs": [], + "source": [ + "# Profile\n", + "from utils.torch_utils import profile\n", + "\n", + "m1 = lambda x: x * torch.sigmoid(x)\n", + "m2 = torch.nn.SiLU()\n", + "results = profile(input=torch.randn(16, 3, 640, 640), ops=[m1, m2], n=100)" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "BSgFCAcMbk1R" + }, + "outputs": [], + "source": [ + "# VOC\n", + "for b, m in zip([64, 64, 64, 32, 16], [f'yolov5{x}' for x in 'nsmlx']): # batch, model\n", + " !python train.py --batch {b} --weights {m}.pt --data VOC.yaml --epochs 50 --img 512 --hyp hyp.VOC.yaml --project VOC --name {m} --cache" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "UWGH7H6yakVl" + }, + "outputs": [], + "source": [ + "# Classification\n", + "for m in [*(f'yolov5{x}.pt' for x in 'nsmlx'), 'resnet50.pt', 'efficientnet_b0.pt']:\n", + " for d in 'mnist', 'fashion-mnist', 'cifar10', 'cifar100', 'imagenette160', 'imagenette320', 'imagenette', 'imagewoof160', 'imagewoof320', 'imagewoof':\n", + " !python classify/train.py --model {m} --data {d} --epochs 10 --project YOLOv5-cls --name {m}-{d}" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "metadata": { + "id": "VTRwsvA9u7ln" + }, + "outputs": [], + "source": [ + "# TensorRT \n", + "!pip install -U nvidia-tensorrt --index-url https://pypi.ngc.nvidia.com # install\n", + "!python export.py --weights yolov5s.pt --include engine --imgsz 640 --device 0 # export\n", + "!python detect.py --weights yolov5s.engine --imgsz 640 --device 0 # inference" + ] + } + ], + "metadata": { + "accelerator": "GPU", + "colab": { + "collapsed_sections": [], + "include_colab_link": true, + "machine_shape": "hm", + "name": "YOLOv5 Tutorial", + "provenance": [], + "toc_visible": true + }, + "kernelspec": { + "display_name": "Python 3.6.9 64-bit", + "language": "python", + "name": "python3" + }, + "language_info": { + "name": "python", + "version": "3.6.9" + }, + "vscode": { + "interpreter": { + "hash": "31f2aee4e71d21fbe5cf8b01ff0e069b9275f58929596ceb00d14d90e3e16cd6" + } + }, + "widgets": { + "application/vnd.jupyter.widget-state+json": { + "0dbce99bb6184238842cbec0587d564a": { + "model_module": "@jupyter-widgets/base", + "model_module_version": "1.2.0", + "model_name": "LayoutModel", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } + }, + "102e1deda239436fa72751c58202fa0f": { + "model_module": "@jupyter-widgets/base", + "model_module_version": "1.2.0", + "model_name": "LayoutModel", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } + }, + "4fd4431ced6c42368e18424912b877e4": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "DescriptionStyleModel", + "state": { + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "DescriptionStyleModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "StyleView", + "description_width": "" + } + }, + "91ff5f93f2a24c5790ab29e347965946": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "DescriptionStyleModel", + "state": { + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "DescriptionStyleModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "StyleView", + "description_width": "" + } + }, + "9aeff9f1780b45f892422fdc96e56913": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "FloatProgressModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "FloatProgressModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "ProgressView", + "bar_style": "success", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_cdd709c4f40941bea1b2053523c9fac8", + "max": 818322941, + "min": 0, + "orientation": "horizontal", + "style": "IPY_MODEL_a1ef2d8de2b741c78ca5d938e2ddbcdf", + "value": 818322941 + } + }, + "a1ef2d8de2b741c78ca5d938e2ddbcdf": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "ProgressStyleModel", + "state": { + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "ProgressStyleModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "StyleView", + "bar_color": null, + "description_width": "" + } + }, + "bf55a7c71d074d3fa88b10b997820825": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "HTMLModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HTMLModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HTMLView", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_0dbce99bb6184238842cbec0587d564a", + "placeholder": "​", + "style": "IPY_MODEL_91ff5f93f2a24c5790ab29e347965946", + "value": " 780M/780M [01:10<00:00, 10.5MB/s]" + } + }, + "c31d2039ccf74c22b67841f4877d1186": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "HBoxModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HBoxModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HBoxView", + "box_style": "", + "children": [ + "IPY_MODEL_d4bba1727c714d94ad58a72bffa07c4c", + "IPY_MODEL_9aeff9f1780b45f892422fdc96e56913", + "IPY_MODEL_bf55a7c71d074d3fa88b10b997820825" + ], + "layout": "IPY_MODEL_d8b66044e2fb4f5b916696834d880c81" + } + }, + "cdd709c4f40941bea1b2053523c9fac8": { + "model_module": "@jupyter-widgets/base", + "model_module_version": "1.2.0", + "model_name": "LayoutModel", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } + }, + "d4bba1727c714d94ad58a72bffa07c4c": { + "model_module": "@jupyter-widgets/controls", + "model_module_version": "1.5.0", + "model_name": "HTMLModel", + "state": { + "_dom_classes": [], + "_model_module": "@jupyter-widgets/controls", + "_model_module_version": "1.5.0", + "_model_name": "HTMLModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/controls", + "_view_module_version": "1.5.0", + "_view_name": "HTMLView", + "description": "", + "description_tooltip": null, + "layout": "IPY_MODEL_102e1deda239436fa72751c58202fa0f", + "placeholder": "​", + "style": "IPY_MODEL_4fd4431ced6c42368e18424912b877e4", + "value": "100%" + } + }, + "d8b66044e2fb4f5b916696834d880c81": { + "model_module": "@jupyter-widgets/base", + "model_module_version": "1.2.0", + "model_name": "LayoutModel", + "state": { + "_model_module": "@jupyter-widgets/base", + "_model_module_version": "1.2.0", + "_model_name": "LayoutModel", + "_view_count": null, + "_view_module": "@jupyter-widgets/base", + "_view_module_version": "1.2.0", + "_view_name": "LayoutView", + "align_content": null, + "align_items": null, + "align_self": null, + "border": null, + "bottom": null, + "display": null, + "flex": null, + "flex_flow": null, + "grid_area": null, + "grid_auto_columns": null, + "grid_auto_flow": null, + "grid_auto_rows": null, + "grid_column": null, + "grid_gap": null, + "grid_row": null, + "grid_template_areas": null, + "grid_template_columns": null, + "grid_template_rows": null, + "height": null, + "justify_content": null, + "justify_items": null, + "left": null, + "margin": null, + "max_height": null, + "max_width": null, + "min_height": null, + "min_width": null, + "object_fit": null, + "object_position": null, + "order": null, + "overflow": null, + "overflow_x": null, + "overflow_y": null, + "padding": null, + "right": null, + "top": null, + "visibility": null, + "width": null + } + } + } + } + }, + "nbformat": 4, + "nbformat_minor": 0 +} diff --git a/application/yolov5_example/utils/__init__.py b/application/yolov5_example/utils/__init__.py new file mode 100644 index 00000000..da53a4d2 --- /dev/null +++ b/application/yolov5_example/utils/__init__.py @@ -0,0 +1,36 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +utils/initialization +""" + + +def notebook_init(verbose=True): + # Check system software and hardware + print('Checking setup...') + + import os + import shutil + + from utils.general import check_requirements, emojis, is_colab + from utils.torch_utils import select_device # imports + + check_requirements(('psutil', 'IPython')) + import psutil + from IPython import display # to display images and clear console output + + if is_colab(): + shutil.rmtree('/content/sample_data', ignore_errors=True) # remove colab /sample_data directory + + # System info + if verbose: + gb = 1 << 30 # bytes to GiB (1024 ** 3) + ram = psutil.virtual_memory().total + total, used, free = shutil.disk_usage("/") + display.clear_output() + s = f'({os.cpu_count()} CPUs, {ram / gb:.1f} GB RAM, {(total - free) / gb:.1f}/{total / gb:.1f} GB disk)' + else: + s = '' + + select_device(newline=False) + print(emojis(f'Setup complete ✅ {s}')) + return display diff --git a/application/yolov5_example/utils/activations.py b/application/yolov5_example/utils/activations.py new file mode 100644 index 00000000..084ce8c4 --- /dev/null +++ b/application/yolov5_example/utils/activations.py @@ -0,0 +1,103 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Activation functions +""" + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class SiLU(nn.Module): + # SiLU activation https://arxiv.org/pdf/1606.08415.pdf + @staticmethod + def forward(x): + return x * torch.sigmoid(x) + + +class Hardswish(nn.Module): + # Hard-SiLU activation + @staticmethod + def forward(x): + # return x * F.hardsigmoid(x) # for TorchScript and CoreML + return x * F.hardtanh(x + 3, 0.0, 6.0) / 6.0 # for TorchScript, CoreML and ONNX + + +class Mish(nn.Module): + # Mish activation https://github.com/digantamisra98/Mish + @staticmethod + def forward(x): + return x * F.softplus(x).tanh() + + +class MemoryEfficientMish(nn.Module): + # Mish activation memory-efficient + class F(torch.autograd.Function): + + @staticmethod + def forward(ctx, x): + ctx.save_for_backward(x) + return x.mul(torch.tanh(F.softplus(x))) # x * tanh(ln(1 + exp(x))) + + @staticmethod + def backward(ctx, grad_output): + x = ctx.saved_tensors[0] + sx = torch.sigmoid(x) + fx = F.softplus(x).tanh() + return grad_output * (fx + x * sx * (1 - fx * fx)) + + def forward(self, x): + return self.F.apply(x) + + +class FReLU(nn.Module): + # FReLU activation https://arxiv.org/abs/2007.11824 + def __init__(self, c1, k=3): # ch_in, kernel + super().__init__() + self.conv = nn.Conv2d(c1, c1, k, 1, 1, groups=c1, bias=False) + self.bn = nn.BatchNorm2d(c1) + + def forward(self, x): + return torch.max(x, self.bn(self.conv(x))) + + +class AconC(nn.Module): + r""" ACON activation (activate or not) + AconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is a learnable parameter + according to "Activate or Not: Learning Customized Activation" . + """ + + def __init__(self, c1): + super().__init__() + self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1)) + self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1)) + self.beta = nn.Parameter(torch.ones(1, c1, 1, 1)) + + def forward(self, x): + dpx = (self.p1 - self.p2) * x + return dpx * torch.sigmoid(self.beta * dpx) + self.p2 * x + + +class MetaAconC(nn.Module): + r""" ACON activation (activate or not) + MetaAconC: (p1*x-p2*x) * sigmoid(beta*(p1*x-p2*x)) + p2*x, beta is generated by a small network + according to "Activate or Not: Learning Customized Activation" . + """ + + def __init__(self, c1, k=1, s=1, r=16): # ch_in, kernel, stride, r + super().__init__() + c2 = max(r, c1 // r) + self.p1 = nn.Parameter(torch.randn(1, c1, 1, 1)) + self.p2 = nn.Parameter(torch.randn(1, c1, 1, 1)) + self.fc1 = nn.Conv2d(c1, c2, k, s, bias=True) + self.fc2 = nn.Conv2d(c2, c1, k, s, bias=True) + # self.bn1 = nn.BatchNorm2d(c2) + # self.bn2 = nn.BatchNorm2d(c1) + + def forward(self, x): + y = x.mean(dim=2, keepdims=True).mean(dim=3, keepdims=True) + # batch-size 1 bug/instabilities https://github.com/ultralytics/yolov5/issues/2891 + # beta = torch.sigmoid(self.bn2(self.fc2(self.bn1(self.fc1(y))))) # bug/unstable + beta = torch.sigmoid(self.fc2(self.fc1(y))) # bug patch BN layers removed + dpx = (self.p1 - self.p2) * x + return dpx * torch.sigmoid(beta * dpx) + self.p2 * x diff --git a/application/yolov5_example/utils/augmentations.py b/application/yolov5_example/utils/augmentations.py new file mode 100644 index 00000000..b00519ae --- /dev/null +++ b/application/yolov5_example/utils/augmentations.py @@ -0,0 +1,350 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Image augmentation functions +""" + +import math +import random + +import cv2 +import numpy as np +import torchvision.transforms as T +import torchvision.transforms.functional as TF + +from utils.general import LOGGER, check_version, colorstr, resample_segments, segment2box +from utils.metrics import bbox_ioa + +IMAGENET_MEAN = 0.485, 0.456, 0.406 # RGB mean +IMAGENET_STD = 0.229, 0.224, 0.225 # RGB standard deviation + + +class Albumentations: + # YOLOv5 Albumentations class (optional, only used if package is installed) + def __init__(self): + self.transform = None + prefix = colorstr('albumentations: ') + try: + import albumentations as A + check_version(A.__version__, '1.0.3', hard=True) # version requirement + + T = [ + A.Blur(p=0.01), + A.MedianBlur(p=0.01), + A.ToGray(p=0.01), + A.CLAHE(p=0.01), + A.RandomBrightnessContrast(p=0.0), + A.RandomGamma(p=0.0), + A.ImageCompression(quality_lower=75, p=0.0)] # transforms + self.transform = A.Compose(T, bbox_params=A.BboxParams(format='yolo', label_fields=['class_labels'])) + + LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p)) + except ImportError: # package not installed, skip + pass + except Exception as e: + LOGGER.info(f'{prefix}{e}') + + def __call__(self, im, labels, p=1.0): + if self.transform and random.random() < p: + new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0]) # transformed + im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])]) + return im, labels + + +def normalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD, inplace=False): + # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = (x - mean) / std + return TF.normalize(x, mean, std, inplace=inplace) + + +def denormalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD): + # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = x * std + mean + for i in range(3): + x[:, i] = x[:, i] * std[i] + mean[i] + return x + + +def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5): + # HSV color-space augmentation + if hgain or sgain or vgain: + r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1 # random gains + hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV)) + dtype = im.dtype # uint8 + + x = np.arange(0, 256, dtype=r.dtype) + lut_hue = ((x * r[0]) % 180).astype(dtype) + lut_sat = np.clip(x * r[1], 0, 255).astype(dtype) + lut_val = np.clip(x * r[2], 0, 255).astype(dtype) + + im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val))) + cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im) # no return needed + + +def hist_equalize(im, clahe=True, bgr=False): + # Equalize histogram on BGR image 'im' with im.shape(n,m,3) and range 0-255 + yuv = cv2.cvtColor(im, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV) + if clahe: + c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8)) + yuv[:, :, 0] = c.apply(yuv[:, :, 0]) + else: + yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0]) # equalize Y channel histogram + return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB) # convert YUV image to RGB + + +def replicate(im, labels): + # Replicate labels + h, w = im.shape[:2] + boxes = labels[:, 1:].astype(int) + x1, y1, x2, y2 = boxes.T + s = ((x2 - x1) + (y2 - y1)) / 2 # side length (pixels) + for i in s.argsort()[:round(s.size * 0.5)]: # smallest indices + x1b, y1b, x2b, y2b = boxes[i] + bh, bw = y2b - y1b, x2b - x1b + yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw)) # offset x, y + x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh] + im[y1a:y2a, x1a:x2a] = im[y1b:y2b, x1b:x2b] # im4[ymin:ymax, xmin:xmax] + labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0) + + return im, labels + + +def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32): + # Resize and pad image while meeting stride-multiple constraints + shape = im.shape[:2] # current shape [height, width] + if isinstance(new_shape, int): + new_shape = (new_shape, new_shape) + + # Scale ratio (new / old) + r = min(new_shape[0] / shape[0], new_shape[1] / shape[1]) + if not scaleup: # only scale down, do not scale up (for better val mAP) + r = min(r, 1.0) + + # Compute padding + ratio = r, r # width, height ratios + new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r)) + dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1] # wh padding + if auto: # minimum rectangle + print('auto') + dw, dh = np.mod(dw, stride), np.mod(dh, stride) # wh padding + elif scaleFill: # stretch + print('stretch') + dw, dh = 0.0, 0.0 + new_unpad = (new_shape[1], new_shape[0]) + ratio = new_shape[1] / shape[1], new_shape[0] / shape[0] # width, height ratios + + dw /= 2 # divide padding into 2 sides + dh /= 2 + + if shape[::-1] != new_unpad: # resize + print('resize2') + im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR) + top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1)) + left, right = int(round(dw - 0.1)), int(round(dw + 0.1)) + im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # add border + return im, ratio, (dw, dh) + + +def random_perspective(im, + targets=(), + segments=(), + degrees=10, + translate=.1, + scale=.1, + shear=10, + perspective=0.0, + border=(0, 0)): + # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(0.1, 0.1), scale=(0.9, 1.1), shear=(-10, 10)) + # targets = [cls, xyxy] + + height = im.shape[0] + border[0] * 2 # shape(h,w,c) + width = im.shape[1] + border[1] * 2 + + # Center + C = np.eye(3) + C[0, 2] = -im.shape[1] / 2 # x translation (pixels) + C[1, 2] = -im.shape[0] / 2 # y translation (pixels) + + # Perspective + P = np.eye(3) + P[2, 0] = random.uniform(-perspective, perspective) # x perspective (about y) + P[2, 1] = random.uniform(-perspective, perspective) # y perspective (about x) + + # Rotation and Scale + R = np.eye(3) + a = random.uniform(-degrees, degrees) + # a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations + s = random.uniform(1 - scale, 1 + scale) + # s = 2 ** random.uniform(-scale, scale) + R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s) + + # Shear + S = np.eye(3) + S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg) + S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg) + + # Translation + T = np.eye(3) + T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width # x translation (pixels) + T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height # y translation (pixels) + + # Combined rotation matrix + M = T @ S @ R @ P @ C # order of operations (right to left) is IMPORTANT + if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any(): # image changed + if perspective: + im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114)) + else: # affine + im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114)) + + # Visualize + # import matplotlib.pyplot as plt + # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel() + # ax[0].imshow(im[:, :, ::-1]) # base + # ax[1].imshow(im2[:, :, ::-1]) # warped + + # Transform label coordinates + n = len(targets) + if n: + use_segments = any(x.any() for x in segments) + new = np.zeros((n, 4)) + if use_segments: # warp segments + segments = resample_segments(segments) # upsample + for i, segment in enumerate(segments): + xy = np.ones((len(segment), 3)) + xy[:, :2] = segment + xy = xy @ M.T # transform + xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2] # perspective rescale or affine + + # clip + new[i] = segment2box(xy, width, height) + + else: # warp boxes + xy = np.ones((n * 4, 3)) + xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2) # x1y1, x2y2, x1y2, x2y1 + xy = xy @ M.T # transform + xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8) # perspective rescale or affine + + # create new boxes + x = xy[:, [0, 2, 4, 6]] + y = xy[:, [1, 3, 5, 7]] + new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T + + # clip + new[:, [0, 2]] = new[:, [0, 2]].clip(0, width) + new[:, [1, 3]] = new[:, [1, 3]].clip(0, height) + + # filter candidates + i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10) + targets = targets[i] + targets[:, 1:5] = new[i] + + return im, targets + + +def copy_paste(im, labels, segments, p=0.5): + # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy) + n = len(segments) + if p and n: + h, w, c = im.shape # height, width, channels + im_new = np.zeros(im.shape, np.uint8) + for j in random.sample(range(n), k=round(p * n)): + l, s = labels[j], segments[j] + box = w - l[3], l[2], w - l[1], l[4] + ioa = bbox_ioa(box, labels[:, 1:5]) # intersection over area + if (ioa < 0.30).all(): # allow 30% obscuration of existing labels + labels = np.concatenate((labels, [[l[0], *box]]), 0) + segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1)) + cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED) + + result = cv2.bitwise_and(src1=im, src2=im_new) + result = cv2.flip(result, 1) # augment segments (flip left-right) + i = result > 0 # pixels to replace + # i[:, :] = result.max(2).reshape(h, w, 1) # act over ch + im[i] = result[i] # cv2.imwrite('debug.jpg', im) # debug + + return im, labels, segments + + +def cutout(im, labels, p=0.5): + # Applies image cutout augmentation https://arxiv.org/abs/1708.04552 + if random.random() < p: + h, w = im.shape[:2] + scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16 # image size fraction + for s in scales: + mask_h = random.randint(1, int(h * s)) # create random masks + mask_w = random.randint(1, int(w * s)) + + # box + xmin = max(0, random.randint(0, w) - mask_w // 2) + ymin = max(0, random.randint(0, h) - mask_h // 2) + xmax = min(w, xmin + mask_w) + ymax = min(h, ymin + mask_h) + + # apply random color mask + im[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)] + + # return unobscured labels + if len(labels) and s > 0.03: + box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32) + ioa = bbox_ioa(box, labels[:, 1:5]) # intersection over area + labels = labels[ioa < 0.60] # remove >60% obscured labels + + return labels + + +def mixup(im, labels, im2, labels2): + # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf + r = np.random.beta(32.0, 32.0) # mixup ratio, alpha=beta=32.0 + im = (im * r + im2 * (1 - r)).astype(np.uint8) + labels = np.concatenate((labels, labels2), 0) + return im, labels + + +def box_candidates(box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16): # box1(4,n), box2(4,n) + # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio + w1, h1 = box1[2] - box1[0], box1[3] - box1[1] + w2, h2 = box2[2] - box2[0], box2[3] - box2[1] + ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps)) # aspect ratio + return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr) # candidates + + +def classify_albumentations(augment=True, + size=224, + scale=(0.08, 1.0), + hflip=0.5, + vflip=0.0, + jitter=0.4, + mean=IMAGENET_MEAN, + std=IMAGENET_STD, + auto_aug=False): + # YOLOv5 classification Albumentations (optional, only used if package is installed) + prefix = colorstr('albumentations: ') + try: + import albumentations as A + from albumentations.pytorch import ToTensorV2 + check_version(A.__version__, '1.0.3', hard=True) # version requirement + if augment: # Resize and crop + T = [A.RandomResizedCrop(height=size, width=size, scale=scale)] + if auto_aug: + # TODO: implement AugMix, AutoAug & RandAug in albumentation + LOGGER.info(f'{prefix}auto augmentations are currently not supported') + else: + if hflip > 0: + T += [A.HorizontalFlip(p=hflip)] + if vflip > 0: + T += [A.VerticalFlip(p=vflip)] + if jitter > 0: + color_jitter = (float(jitter),) * 3 # repeat value for brightness, contrast, satuaration, 0 hue + T += [A.ColorJitter(*color_jitter, 0)] + else: # Use fixed crop for eval set (reproducibility) + T = [A.SmallestMaxSize(max_size=size), A.CenterCrop(height=size, width=size)] + T += [A.Normalize(mean=mean, std=std), ToTensorV2()] # Normalize and convert to Tensor + LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p)) + return A.Compose(T) + + except ImportError: # package not installed, skip + pass + except Exception as e: + LOGGER.info(f'{prefix}{e}') + + +def classify_transforms(size=224): + # Transforms to apply if albumentations not installed + return T.Compose([T.ToTensor(), T.Resize(size), T.CenterCrop(size), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)]) diff --git a/application/yolov5_example/utils/autoanchor.py b/application/yolov5_example/utils/autoanchor.py new file mode 100644 index 00000000..f2222203 --- /dev/null +++ b/application/yolov5_example/utils/autoanchor.py @@ -0,0 +1,170 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +AutoAnchor utils +""" + +import random + +import numpy as np +import torch +import yaml +from tqdm import tqdm + +from utils.general import LOGGER, colorstr + +PREFIX = colorstr('AutoAnchor: ') + + +def check_anchor_order(m): + # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary + a = m.anchors.prod(-1).mean(-1).view(-1) # mean anchor area per output layer + da = a[-1] - a[0] # delta a + ds = m.stride[-1] - m.stride[0] # delta s + if da and (da.sign() != ds.sign()): # same order + LOGGER.info(f'{PREFIX}Reversing anchor order') + m.anchors[:] = m.anchors.flip(0) + + +def check_anchors(dataset, model, thr=4.0, imgsz=640): + # Check anchor fit to data, recompute if necessary + m = model.module.model[-1] if hasattr(model, 'module') else model.model[-1] # Detect() + shapes = imgsz * dataset.shapes / dataset.shapes.max(1, keepdims=True) + scale = np.random.uniform(0.9, 1.1, size=(shapes.shape[0], 1)) # augment scale + wh = torch.tensor(np.concatenate([l[:, 3:5] * s for s, l in zip(shapes * scale, dataset.labels)])).float() # wh + + def metric(k): # compute metric + r = wh[:, None] / k[None] + x = torch.min(r, 1 / r).min(2)[0] # ratio metric + best = x.max(1)[0] # best_x + aat = (x > 1 / thr).float().sum(1).mean() # anchors above threshold + bpr = (best > 1 / thr).float().mean() # best possible recall + return bpr, aat + + stride = m.stride.to(m.anchors.device).view(-1, 1, 1) # model strides + anchors = m.anchors.clone() * stride # current anchors + bpr, aat = metric(anchors.cpu().view(-1, 2)) + s = f'\n{PREFIX}{aat:.2f} anchors/target, {bpr:.3f} Best Possible Recall (BPR). ' + if bpr > 0.98: # threshold to recompute + LOGGER.info(f'{s}Current anchors are a good fit to dataset ✅') + else: + LOGGER.info(f'{s}Anchors are a poor fit to dataset ⚠️, attempting to improve...') + na = m.anchors.numel() // 2 # number of anchors + try: + anchors = kmean_anchors(dataset, n=na, img_size=imgsz, thr=thr, gen=1000, verbose=False) + except Exception as e: + LOGGER.info(f'{PREFIX}ERROR: {e}') + new_bpr = metric(anchors)[0] + if new_bpr > bpr: # replace anchors + anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors) + m.anchors[:] = anchors.clone().view_as(m.anchors) + check_anchor_order(m) # must be in pixel-space (not grid-space) + m.anchors /= stride + s = f'{PREFIX}Done ✅ (optional: update model *.yaml to use these anchors in the future)' + else: + s = f'{PREFIX}Done ⚠️ (original anchors better than new anchors, proceeding with original anchors)' + LOGGER.info(s) + + +def kmean_anchors(dataset='./data/coco128.yaml', n=9, img_size=640, thr=4.0, gen=1000, verbose=True): + """ Creates kmeans-evolved anchors from training dataset + + Arguments: + dataset: path to data.yaml, or a loaded dataset + n: number of anchors + img_size: image size used for training + thr: anchor-label wh ratio threshold hyperparameter hyp['anchor_t'] used for training, default=4.0 + gen: generations to evolve anchors using genetic algorithm + verbose: print all results + + Return: + k: kmeans evolved anchors + + Usage: + from utils.autoanchor import *; _ = kmean_anchors() + """ + from scipy.cluster.vq import kmeans + + npr = np.random + thr = 1 / thr + + def metric(k, wh): # compute metrics + r = wh[:, None] / k[None] + x = torch.min(r, 1 / r).min(2)[0] # ratio metric + # x = wh_iou(wh, torch.tensor(k)) # iou metric + return x, x.max(1)[0] # x, best_x + + def anchor_fitness(k): # mutation fitness + _, best = metric(torch.tensor(k, dtype=torch.float32), wh) + return (best * (best > thr).float()).mean() # fitness + + def print_results(k, verbose=True): + k = k[np.argsort(k.prod(1))] # sort small to large + x, best = metric(k, wh0) + bpr, aat = (best > thr).float().mean(), (x > thr).float().mean() * n # best possible recall, anch > thr + s = f'{PREFIX}thr={thr:.2f}: {bpr:.4f} best possible recall, {aat:.2f} anchors past thr\n' \ + f'{PREFIX}n={n}, img_size={img_size}, metric_all={x.mean():.3f}/{best.mean():.3f}-mean/best, ' \ + f'past_thr={x[x > thr].mean():.3f}-mean: ' + for x in k: + s += '%i,%i, ' % (round(x[0]), round(x[1])) + if verbose: + LOGGER.info(s[:-2]) + return k + + if isinstance(dataset, str): # *.yaml file + with open(dataset, errors='ignore') as f: + data_dict = yaml.safe_load(f) # model dict + from utils.dataloaders import LoadImagesAndLabels + dataset = LoadImagesAndLabels(data_dict['train'], augment=True, rect=True) + + # Get label wh + shapes = img_size * dataset.shapes / dataset.shapes.max(1, keepdims=True) + wh0 = np.concatenate([l[:, 3:5] * s for s, l in zip(shapes, dataset.labels)]) # wh + + # Filter + i = (wh0 < 3.0).any(1).sum() + if i: + LOGGER.info(f'{PREFIX}WARNING: Extremely small objects found: {i} of {len(wh0)} labels are < 3 pixels in size') + wh = wh0[(wh0 >= 2.0).any(1)] # filter > 2 pixels + # wh = wh * (npr.rand(wh.shape[0], 1) * 0.9 + 0.1) # multiply by random scale 0-1 + + # Kmeans init + try: + LOGGER.info(f'{PREFIX}Running kmeans for {n} anchors on {len(wh)} points...') + assert n <= len(wh) # apply overdetermined constraint + s = wh.std(0) # sigmas for whitening + k = kmeans(wh / s, n, iter=30)[0] * s # points + assert n == len(k) # kmeans may return fewer points than requested if wh is insufficient or too similar + except Exception: + LOGGER.warning(f'{PREFIX}WARNING: switching strategies from kmeans to random init') + k = np.sort(npr.rand(n * 2)).reshape(n, 2) * img_size # random init + wh, wh0 = (torch.tensor(x, dtype=torch.float32) for x in (wh, wh0)) + k = print_results(k, verbose=False) + + # Plot + # k, d = [None] * 20, [None] * 20 + # for i in tqdm(range(1, 21)): + # k[i-1], d[i-1] = kmeans(wh / s, i) # points, mean distance + # fig, ax = plt.subplots(1, 2, figsize=(14, 7), tight_layout=True) + # ax = ax.ravel() + # ax[0].plot(np.arange(1, 21), np.array(d) ** 2, marker='.') + # fig, ax = plt.subplots(1, 2, figsize=(14, 7)) # plot wh + # ax[0].hist(wh[wh[:, 0]<100, 0],400) + # ax[1].hist(wh[wh[:, 1]<100, 1],400) + # fig.savefig('wh.png', dpi=200) + + # Evolve + f, sh, mp, s = anchor_fitness(k), k.shape, 0.9, 0.1 # fitness, generations, mutation prob, sigma + pbar = tqdm(range(gen), bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar + for _ in pbar: + v = np.ones(sh) + while (v == 1).all(): # mutate until a change occurs (prevent duplicates) + v = ((npr.random(sh) < mp) * random.random() * npr.randn(*sh) * s + 1).clip(0.3, 3.0) + kg = (k.copy() * v).clip(min=2.0) + fg = anchor_fitness(kg) + if fg > f: + f, k = fg, kg.copy() + pbar.desc = f'{PREFIX}Evolving anchors with Genetic Algorithm: fitness = {f:.4f}' + if verbose: + print_results(k, verbose) + + return print_results(k) diff --git a/application/yolov5_example/utils/autobatch.py b/application/yolov5_example/utils/autobatch.py new file mode 100644 index 00000000..c231d24c --- /dev/null +++ b/application/yolov5_example/utils/autobatch.py @@ -0,0 +1,66 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Auto-batch utils +""" + +from copy import deepcopy + +import numpy as np +import torch + +from utils.general import LOGGER, colorstr +from utils.torch_utils import profile + + +def check_train_batch_size(model, imgsz=640, amp=True): + # Check YOLOv5 training batch size + with torch.cuda.amp.autocast(amp): + return autobatch(deepcopy(model).train(), imgsz) # compute optimal batch size + + +def autobatch(model, imgsz=640, fraction=0.9, batch_size=16): + # Automatically estimate best batch size to use `fraction` of available CUDA memory + # Usage: + # import torch + # from utils.autobatch import autobatch + # model = torch.hub.load('ultralytics/yolov5', 'yolov5s', autoshape=False) + # print(autobatch(model)) + + # Check device + prefix = colorstr('AutoBatch: ') + LOGGER.info(f'{prefix}Computing optimal batch size for --imgsz {imgsz}') + device = next(model.parameters()).device # get model device + if device.type == 'cpu': + LOGGER.info(f'{prefix}CUDA not detected, using default CPU batch-size {batch_size}') + return batch_size + + # Inspect CUDA memory + gb = 1 << 30 # bytes to GiB (1024 ** 3) + d = str(device).upper() # 'CUDA:0' + properties = torch.cuda.get_device_properties(device) # device properties + t = properties.total_memory / gb # GiB total + r = torch.cuda.memory_reserved(device) / gb # GiB reserved + a = torch.cuda.memory_allocated(device) / gb # GiB allocated + f = t - (r + a) # GiB free + LOGGER.info(f'{prefix}{d} ({properties.name}) {t:.2f}G total, {r:.2f}G reserved, {a:.2f}G allocated, {f:.2f}G free') + + # Profile batch sizes + batch_sizes = [1, 2, 4, 8, 16] + try: + img = [torch.zeros(b, 3, imgsz, imgsz) for b in batch_sizes] + results = profile(img, model, n=3, device=device) + except Exception as e: + LOGGER.warning(f'{prefix}{e}') + + # Fit a solution + y = [x[2] for x in results if x] # memory [2] + p = np.polyfit(batch_sizes[:len(y)], y, deg=1) # first degree polynomial fit + b = int((f * fraction - p[1]) / p[0]) # y intercept (optimal batch size) + if None in results: # some sizes failed + i = results.index(None) # first fail index + if b >= batch_sizes[i]: # y intercept above failure point + b = batch_sizes[max(i - 1, 0)] # select prior safe point + + fraction = np.polyval(p, b) / t # actual fraction predicted + LOGGER.info(f'{prefix}Using batch-size {b} for {d} {t * fraction:.2f}G/{t:.2f}G ({fraction * 100:.0f}%) ✅') + return b diff --git a/application/yolov5_example/utils/aws/__init__.py b/application/yolov5_example/utils/aws/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/application/yolov5_example/utils/aws/mime.sh b/application/yolov5_example/utils/aws/mime.sh new file mode 100644 index 00000000..c319a83c --- /dev/null +++ b/application/yolov5_example/utils/aws/mime.sh @@ -0,0 +1,26 @@ +# AWS EC2 instance startup 'MIME' script https://aws.amazon.com/premiumsupport/knowledge-center/execute-user-data-ec2/ +# This script will run on every instance restart, not only on first start +# --- DO NOT COPY ABOVE COMMENTS WHEN PASTING INTO USERDATA --- + +Content-Type: multipart/mixed; boundary="//" +MIME-Version: 1.0 + +--// +Content-Type: text/cloud-config; charset="us-ascii" +MIME-Version: 1.0 +Content-Transfer-Encoding: 7bit +Content-Disposition: attachment; filename="cloud-config.txt" + +#cloud-config +cloud_final_modules: +- [scripts-user, always] + +--// +Content-Type: text/x-shellscript; charset="us-ascii" +MIME-Version: 1.0 +Content-Transfer-Encoding: 7bit +Content-Disposition: attachment; filename="userdata.txt" + +#!/bin/bash +# --- paste contents of userdata.sh here --- +--// diff --git a/application/yolov5_example/utils/aws/resume.py b/application/yolov5_example/utils/aws/resume.py new file mode 100644 index 00000000..b21731c9 --- /dev/null +++ b/application/yolov5_example/utils/aws/resume.py @@ -0,0 +1,40 @@ +# Resume all interrupted trainings in yolov5/ dir including DDP trainings +# Usage: $ python utils/aws/resume.py + +import os +import sys +from pathlib import Path + +import torch +import yaml + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[2] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH + +port = 0 # --master_port +path = Path('').resolve() +for last in path.rglob('*/**/last.pt'): + ckpt = torch.load(last) + if ckpt['optimizer'] is None: + continue + + # Load opt.yaml + with open(last.parent.parent / 'opt.yaml', errors='ignore') as f: + opt = yaml.safe_load(f) + + # Get device count + d = opt['device'].split(',') # devices + nd = len(d) # number of devices + ddp = nd > 1 or (nd == 0 and torch.cuda.device_count() > 1) # distributed data parallel + + if ddp: # multi-GPU + port += 1 + cmd = f'python -m torch.distributed.run --nproc_per_node {nd} --master_port {port} train.py --resume {last}' + else: # single-GPU + cmd = f'python train.py --resume {last}' + + cmd += ' > /dev/null 2>&1 &' # redirect output to dev/null and run in daemon thread + print(cmd) + os.system(cmd) diff --git a/application/yolov5_example/utils/aws/userdata.sh b/application/yolov5_example/utils/aws/userdata.sh new file mode 100644 index 00000000..5fc1332a --- /dev/null +++ b/application/yolov5_example/utils/aws/userdata.sh @@ -0,0 +1,27 @@ +#!/bin/bash +# AWS EC2 instance startup script https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/user-data.html +# This script will run only once on first instance start (for a re-start script see mime.sh) +# /home/ubuntu (ubuntu) or /home/ec2-user (amazon-linux) is working dir +# Use >300 GB SSD + +cd home/ubuntu +if [ ! -d yolov5 ]; then + echo "Running first-time script." # install dependencies, download COCO, pull Docker + git clone https://github.com/ultralytics/yolov5 -b master && sudo chmod -R 777 yolov5 + cd yolov5 + bash data/scripts/get_coco.sh && echo "COCO done." & + sudo docker pull ultralytics/yolov5:latest && echo "Docker done." & + python -m pip install --upgrade pip && pip install -r requirements.txt && python detect.py && echo "Requirements done." & + wait && echo "All tasks done." # finish background tasks +else + echo "Running re-start script." # resume interrupted runs + i=0 + list=$(sudo docker ps -qa) # container list i.e. $'one\ntwo\nthree\nfour' + while IFS= read -r id; do + ((i++)) + echo "restarting container $i: $id" + sudo docker start $id + # sudo docker exec -it $id python train.py --resume # single-GPU + sudo docker exec -d $id python utils/aws/resume.py # multi-scenario + done <<<"$list" +fi diff --git a/application/yolov5_example/utils/benchmarks.py b/application/yolov5_example/utils/benchmarks.py new file mode 100644 index 00000000..d412653c --- /dev/null +++ b/application/yolov5_example/utils/benchmarks.py @@ -0,0 +1,157 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Run YOLOv5 benchmarks on all supported export formats + +Format | `export.py --include` | Model +--- | --- | --- +PyTorch | - | yolov5s.pt +TorchScript | `torchscript` | yolov5s.torchscript +ONNX | `onnx` | yolov5s.onnx +OpenVINO | `openvino` | yolov5s_openvino_model/ +TensorRT | `engine` | yolov5s.engine +CoreML | `coreml` | yolov5s.mlmodel +TensorFlow SavedModel | `saved_model` | yolov5s_saved_model/ +TensorFlow GraphDef | `pb` | yolov5s.pb +TensorFlow Lite | `tflite` | yolov5s.tflite +TensorFlow Edge TPU | `edgetpu` | yolov5s_edgetpu.tflite +TensorFlow.js | `tfjs` | yolov5s_web_model/ + +Requirements: + $ pip install -r requirements.txt coremltools onnx onnx-simplifier onnxruntime openvino-dev tensorflow-cpu # CPU + $ pip install -r requirements.txt coremltools onnx onnx-simplifier onnxruntime-gpu openvino-dev tensorflow # GPU + $ pip install -U nvidia-tensorrt --index-url https://pypi.ngc.nvidia.com # TensorRT + +Usage: + $ python utils/benchmarks.py --weights yolov5s.pt --img 640 +""" + +import argparse +import platform +import sys +import time +from pathlib import Path + +import pandas as pd + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +# ROOT = ROOT.relative_to(Path.cwd()) # relative + +import export +import val +from utils import notebook_init +from utils.general import LOGGER, check_yaml, file_size, print_args +from utils.torch_utils import select_device + + +def run( + weights=ROOT / 'yolov5s.pt', # weights path + imgsz=640, # inference size (pixels) + batch_size=1, # batch size + data=ROOT / 'data/coco128.yaml', # dataset.yaml path + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + half=False, # use FP16 half-precision inference + test=False, # test exports only + pt_only=False, # test PyTorch only + hard_fail=False, # throw error on benchmark failure +): + y, t = [], time.time() + device = select_device(device) + for i, (name, f, suffix, cpu, gpu) in export.export_formats().iterrows(): # index, (name, file, suffix, CPU, GPU) + try: + assert i not in (9, 10), 'inference not supported' # Edge TPU and TF.js are unsupported + assert i != 5 or platform.system() == 'Darwin', 'inference only supported on macOS>=10.13' # CoreML + if 'cpu' in device.type: + assert cpu, 'inference not supported on CPU' + if 'cuda' in device.type: + assert gpu, 'inference not supported on GPU' + + # Export + if f == '-': + w = weights # PyTorch format + else: + w = export.run(weights=weights, imgsz=[imgsz], include=[f], device=device, half=half)[-1] # all others + assert suffix in str(w), 'export failed' + + # Validate + result = val.run(data, w, batch_size, imgsz, plots=False, device=device, task='benchmark', half=half) + metrics = result[0] # metrics (mp, mr, map50, map, *losses(box, obj, cls)) + speeds = result[2] # times (preprocess, inference, postprocess) + y.append([name, round(file_size(w), 1), round(metrics[3], 4), round(speeds[1], 2)]) # MB, mAP, t_inference + except Exception as e: + if hard_fail: + assert type(e) is AssertionError, f'Benchmark --hard-fail for {name}: {e}' + LOGGER.warning(f'WARNING: Benchmark failure for {name}: {e}') + y.append([name, None, None, None]) # mAP, t_inference + if pt_only and i == 0: + break # break after PyTorch + + # Print results + LOGGER.info('\n') + parse_opt() + notebook_init() # print system info + c = ['Format', 'Size (MB)', 'mAP@0.5:0.95', 'Inference time (ms)'] if map else ['Format', 'Export', '', ''] + py = pd.DataFrame(y, columns=c) + LOGGER.info(f'\nBenchmarks complete ({time.time() - t:.2f}s)') + LOGGER.info(str(py if map else py.iloc[:, :2])) + return py + + +def test( + weights=ROOT / 'yolov5s.pt', # weights path + imgsz=640, # inference size (pixels) + batch_size=1, # batch size + data=ROOT / 'data/coco128.yaml', # dataset.yaml path + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + half=False, # use FP16 half-precision inference + test=False, # test exports only + pt_only=False, # test PyTorch only + hard_fail=False, # throw error on benchmark failure +): + y, t = [], time.time() + device = select_device(device) + for i, (name, f, suffix, gpu) in export.export_formats().iterrows(): # index, (name, file, suffix, gpu-capable) + try: + w = weights if f == '-' else \ + export.run(weights=weights, imgsz=[imgsz], include=[f], device=device, half=half)[-1] # weights + assert suffix in str(w), 'export failed' + y.append([name, True]) + except Exception: + y.append([name, False]) # mAP, t_inference + + # Print results + LOGGER.info('\n') + parse_opt() + notebook_init() # print system info + py = pd.DataFrame(y, columns=['Format', 'Export']) + LOGGER.info(f'\nExports complete ({time.time() - t:.2f}s)') + LOGGER.info(str(py)) + return py + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--weights', type=str, default=ROOT / 'yolov5s.pt', help='weights path') + parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)') + parser.add_argument('--batch-size', type=int, default=1, help='batch size') + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--test', action='store_true', help='test exports only') + parser.add_argument('--pt-only', action='store_true', help='test PyTorch only') + parser.add_argument('--hard-fail', action='store_true', help='throw error on benchmark failure') + opt = parser.parse_args() + opt.data = check_yaml(opt.data) # check YAML + print_args(vars(opt)) + return opt + + +def main(opt): + test(**vars(opt)) if opt.test else run(**vars(opt)) + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/application/yolov5_example/utils/callbacks.py b/application/yolov5_example/utils/callbacks.py new file mode 100644 index 00000000..2b32df0b --- /dev/null +++ b/application/yolov5_example/utils/callbacks.py @@ -0,0 +1,71 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Callback utils +""" + + +class Callbacks: + """" + Handles all registered callbacks for YOLOv5 Hooks + """ + + def __init__(self): + # Define the available callbacks + self._callbacks = { + 'on_pretrain_routine_start': [], + 'on_pretrain_routine_end': [], + 'on_train_start': [], + 'on_train_epoch_start': [], + 'on_train_batch_start': [], + 'optimizer_step': [], + 'on_before_zero_grad': [], + 'on_train_batch_end': [], + 'on_train_epoch_end': [], + 'on_val_start': [], + 'on_val_batch_start': [], + 'on_val_image_end': [], + 'on_val_batch_end': [], + 'on_val_end': [], + 'on_fit_epoch_end': [], # fit = train + val + 'on_model_save': [], + 'on_train_end': [], + 'on_params_update': [], + 'teardown': [],} + self.stop_training = False # set True to interrupt training + + def register_action(self, hook, name='', callback=None): + """ + Register a new action to a callback hook + + Args: + hook: The callback hook name to register the action to + name: The name of the action for later reference + callback: The callback to fire + """ + assert hook in self._callbacks, f"hook '{hook}' not found in callbacks {self._callbacks}" + assert callable(callback), f"callback '{callback}' is not callable" + self._callbacks[hook].append({'name': name, 'callback': callback}) + + def get_registered_actions(self, hook=None): + """" + Returns all the registered actions by callback hook + + Args: + hook: The name of the hook to check, defaults to all + """ + return self._callbacks[hook] if hook else self._callbacks + + def run(self, hook, *args, **kwargs): + """ + Loop through the registered actions and fire all callbacks + + Args: + hook: The name of the hook to check, defaults to all + args: Arguments to receive from YOLOv5 + kwargs: Keyword Arguments to receive from YOLOv5 + """ + + assert hook in self._callbacks, f"hook '{hook}' not found in callbacks {self._callbacks}" + + for logger in self._callbacks[hook]: + logger['callback'](*args, **kwargs) diff --git a/application/yolov5_example/utils/dataloaders.py b/application/yolov5_example/utils/dataloaders.py new file mode 100644 index 00000000..184558fd --- /dev/null +++ b/application/yolov5_example/utils/dataloaders.py @@ -0,0 +1,1158 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Dataloaders and dataset utils +""" + +import contextlib +import glob +import hashlib +import json +import math +import os +import random +import shutil +import time +from itertools import repeat +from multiprocessing.pool import Pool, ThreadPool +from pathlib import Path +from threading import Thread +from urllib.parse import urlparse +from zipfile import ZipFile + +import numpy as np +import torch +import torch.nn.functional as F +import torchvision +import yaml +from PIL import ExifTags, Image, ImageOps +from torch.utils.data import DataLoader, Dataset, dataloader, distributed +from tqdm import tqdm + +from utils.augmentations import (Albumentations, augment_hsv, classify_albumentations, classify_transforms, copy_paste, + letterbox, mixup, random_perspective) +from utils.general import (DATASETS_DIR, LOGGER, NUM_THREADS, check_dataset, check_requirements, check_yaml, clean_str, + cv2, is_colab, is_kaggle, segments2boxes, xyn2xy, xywh2xyxy, xywhn2xyxy, xyxy2xywhn) +from utils.torch_utils import torch_distributed_zero_first + +# Parameters +HELP_URL = 'https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data' +IMG_FORMATS = 'bmp', 'dng', 'jpeg', 'jpg', 'mpo', 'png', 'tif', 'tiff', 'webp' # include image suffixes +VID_FORMATS = 'asf', 'avi', 'gif', 'm4v', 'mkv', 'mov', 'mp4', 'mpeg', 'mpg', 'ts', 'wmv' # include video suffixes +BAR_FORMAT = '{l_bar}{bar:10}{r_bar}{bar:-10b}' # tqdm bar format +LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1)) # https://pytorch.org/docs/stable/elastic/run.html + +# Get orientation exif tag +for orientation in ExifTags.TAGS.keys(): + if ExifTags.TAGS[orientation] == 'Orientation': + break + + +def get_hash(paths): + # Returns a single hash value of a list of paths (files or dirs) + size = sum(os.path.getsize(p) for p in paths if os.path.exists(p)) # sizes + h = hashlib.md5(str(size).encode()) # hash sizes + h.update(''.join(paths).encode()) # hash paths + return h.hexdigest() # return hash + + +def exif_size(img): + # Returns exif-corrected PIL size + s = img.size # (width, height) + with contextlib.suppress(Exception): + rotation = dict(img._getexif().items())[orientation] + if rotation in [6, 8]: # rotation 270 or 90 + s = (s[1], s[0]) + return s + + +def exif_transpose(image): + """ + Transpose a PIL image accordingly if it has an EXIF Orientation tag. + Inplace version of https://github.com/python-pillow/Pillow/blob/master/src/PIL/ImageOps.py exif_transpose() + + :param image: The image to transpose. + :return: An image. + """ + exif = image.getexif() + orientation = exif.get(0x0112, 1) # default 1 + if orientation > 1: + method = { + 2: Image.FLIP_LEFT_RIGHT, + 3: Image.ROTATE_180, + 4: Image.FLIP_TOP_BOTTOM, + 5: Image.TRANSPOSE, + 6: Image.ROTATE_270, + 7: Image.TRANSVERSE, + 8: Image.ROTATE_90,}.get(orientation) + if method is not None: + image = image.transpose(method) + del exif[0x0112] + image.info["exif"] = exif.tobytes() + return image + + +def seed_worker(worker_id): + # Set dataloader worker seed https://pytorch.org/docs/stable/notes/randomness.html#dataloader + worker_seed = torch.initial_seed() % 2 ** 32 + np.random.seed(worker_seed) + random.seed(worker_seed) + + +def create_dataloader(path, + imgsz, + batch_size, + stride, + single_cls=False, + hyp=None, + augment=False, + cache=False, + pad=0.0, + rect=False, + rank=-1, + workers=8, + image_weights=False, + quad=False, + prefix='', + shuffle=False): + if rect and shuffle: + LOGGER.warning('WARNING: --rect is incompatible with DataLoader shuffle, setting shuffle=False') + shuffle = False + with torch_distributed_zero_first(rank): # init dataset *.cache only once if DDP + dataset = LoadImagesAndLabels( + path, + imgsz, + batch_size, + augment=augment, # augmentation + hyp=hyp, # hyperparameters + rect=rect, # rectangular batches + cache_images=cache, + single_cls=single_cls, + stride=int(stride), + pad=pad, + image_weights=image_weights, + prefix=prefix) + + batch_size = min(batch_size, len(dataset)) + nd = torch.cuda.device_count() # number of CUDA devices + nw = min([os.cpu_count() // max(nd, 1), batch_size if batch_size > 1 else 0, workers]) # number of workers + sampler = None if rank == -1 else distributed.DistributedSampler(dataset, shuffle=shuffle) + loader = DataLoader if image_weights else InfiniteDataLoader # only DataLoader allows for attribute updates + generator = torch.Generator() + generator.manual_seed(0) + return loader(dataset, + batch_size=batch_size, + shuffle=shuffle and sampler is None, + num_workers=nw, + sampler=sampler, + pin_memory=True, + collate_fn=LoadImagesAndLabels.collate_fn4 if quad else LoadImagesAndLabels.collate_fn, + worker_init_fn=seed_worker, + generator=generator), dataset + + +class InfiniteDataLoader(dataloader.DataLoader): + """ Dataloader that reuses workers + + Uses same syntax as vanilla DataLoader + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + object.__setattr__(self, 'batch_sampler', _RepeatSampler(self.batch_sampler)) + self.iterator = super().__iter__() + + def __len__(self): + return len(self.batch_sampler.sampler) + + def __iter__(self): + for _ in range(len(self)): + yield next(self.iterator) + + +class _RepeatSampler: + """ Sampler that repeats forever + + Args: + sampler (Sampler) + """ + + def __init__(self, sampler): + self.sampler = sampler + + def __iter__(self): + while True: + yield from iter(self.sampler) + + +class LoadImages: + # YOLOv5 image/video dataloader, i.e. `python detect.py --source image.jpg/vid.mp4` + def __init__(self, path, img_size=640, stride=32, auto=True): + files = [] + for p in sorted(path) if isinstance(path, (list, tuple)) else [path]: + p = str(Path(p).resolve()) + if '*' in p: + files.extend(sorted(glob.glob(p, recursive=True))) # glob + elif os.path.isdir(p): + files.extend(sorted(glob.glob(os.path.join(p, '*.*')))) # dir + elif os.path.isfile(p): + files.append(p) # files + else: + raise FileNotFoundError(f'{p} does not exist') + + images = [x for x in files if x.split('.')[-1].lower() in IMG_FORMATS] + videos = [x for x in files if x.split('.')[-1].lower() in VID_FORMATS] + ni, nv = len(images), len(videos) + + self.img_size = img_size + self.stride = stride + self.files = images + videos + self.nf = ni + nv # number of files + self.video_flag = [False] * ni + [True] * nv + self.mode = 'image' + self.auto = auto + if any(videos): + self.new_video(videos[0]) # new video + else: + self.cap = None + assert self.nf > 0, f'No images or videos found in {p}. ' \ + f'Supported formats are:\nimages: {IMG_FORMATS}\nvideos: {VID_FORMATS}' + + def __iter__(self): + self.count = 0 + return self + + def __next__(self): + if self.count == self.nf: + raise StopIteration + path = self.files[self.count] + + if self.video_flag[self.count]: + # Read video + self.mode = 'video' + ret_val, img0 = self.cap.read() + while not ret_val: + self.count += 1 + self.cap.release() + if self.count == self.nf: # last video + raise StopIteration + path = self.files[self.count] + self.new_video(path) + ret_val, img0 = self.cap.read() + + self.frame += 1 + s = f'video {self.count + 1}/{self.nf} ({self.frame}/{self.frames}) {path}: ' + + else: + # Read image + self.count += 1 + img0 = cv2.imread(path) # BGR + assert img0 is not None, f'Image Not Found {path}' + s = f'image {self.count}/{self.nf} {path}: ' + + # Padded resize + img = letterbox(img0, self.img_size, stride=self.stride, auto=self.auto)[0] + + # Convert + img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB + img = np.ascontiguousarray(img) + + return path, img, img0, self.cap, s + + def new_video(self, path): + self.frame = 0 + self.cap = cv2.VideoCapture(path) + self.frames = int(self.cap.get(cv2.CAP_PROP_FRAME_COUNT)) + + def __len__(self): + return self.nf # number of files + + +class LoadWebcam: # for inference + # YOLOv5 local webcam dataloader, i.e. `python detect.py --source 0` + def __init__(self, pipe='0', img_size=640, stride=32): + self.img_size = img_size + self.stride = stride + self.pipe = eval(pipe) if pipe.isnumeric() else pipe + self.cap = cv2.VideoCapture(self.pipe) # video capture object + self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 3) # set buffer size + + def __iter__(self): + self.count = -1 + return self + + def __next__(self): + self.count += 1 + if cv2.waitKey(1) == ord('q'): # q to quit + self.cap.release() + cv2.destroyAllWindows() + raise StopIteration + + # Read frame + ret_val, img0 = self.cap.read() + img0 = cv2.flip(img0, 1) # flip left-right + + # Print + assert ret_val, f'Camera Error {self.pipe}' + img_path = 'webcam.jpg' + s = f'webcam {self.count}: ' + + # Padded resize + img = letterbox(img0, self.img_size, stride=self.stride)[0] + + # Convert + img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB + img = np.ascontiguousarray(img) + + return img_path, img, img0, None, s + + def __len__(self): + return 0 + + +class LoadStreams: + # YOLOv5 streamloader, i.e. `python detect.py --source 'rtsp://example.com/media.mp4' # RTSP, RTMP, HTTP streams` + def __init__(self, sources='streams.txt', img_size=640, stride=32, auto=True): + self.mode = 'stream' + self.img_size = img_size + self.stride = stride + + if os.path.isfile(sources): + with open(sources) as f: + sources = [x.strip() for x in f.read().strip().splitlines() if len(x.strip())] + else: + sources = [sources] + + n = len(sources) + self.imgs, self.fps, self.frames, self.threads = [None] * n, [0] * n, [0] * n, [None] * n + self.sources = [clean_str(x) for x in sources] # clean source names for later + self.auto = auto + for i, s in enumerate(sources): # index, source + # Start thread to read frames from video stream + st = f'{i + 1}/{n}: {s}... ' + if urlparse(s).hostname in ('www.youtube.com', 'youtube.com', 'youtu.be'): # if source is YouTube video + check_requirements(('pafy', 'youtube_dl==2020.12.2')) + import pafy + s = pafy.new(s).getbest(preftype="mp4").url # YouTube URL + s = eval(s) if s.isnumeric() else s # i.e. s = '0' local webcam + if s == 0: + assert not is_colab(), '--source 0 webcam unsupported on Colab. Rerun command in a local environment.' + assert not is_kaggle(), '--source 0 webcam unsupported on Kaggle. Rerun command in a local environment.' + cap = cv2.VideoCapture(s) + assert cap.isOpened(), f'{st}Failed to open {s}' + w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)) + h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) + fps = cap.get(cv2.CAP_PROP_FPS) # warning: may return 0 or nan + self.frames[i] = max(int(cap.get(cv2.CAP_PROP_FRAME_COUNT)), 0) or float('inf') # infinite stream fallback + self.fps[i] = max((fps if math.isfinite(fps) else 0) % 100, 0) or 30 # 30 FPS fallback + + _, self.imgs[i] = cap.read() # guarantee first frame + self.threads[i] = Thread(target=self.update, args=([i, cap, s]), daemon=True) + LOGGER.info(f"{st} Success ({self.frames[i]} frames {w}x{h} at {self.fps[i]:.2f} FPS)") + self.threads[i].start() + LOGGER.info('') # newline + + # check for common shapes + s = np.stack([letterbox(x, self.img_size, stride=self.stride, auto=self.auto)[0].shape for x in self.imgs]) + self.rect = np.unique(s, axis=0).shape[0] == 1 # rect inference if all shapes equal + if not self.rect: + LOGGER.warning('WARNING: Stream shapes differ. For optimal performance supply similarly-shaped streams.') + + def update(self, i, cap, stream): + # Read stream `i` frames in daemon thread + n, f, read = 0, self.frames[i], 1 # frame number, frame array, inference every 'read' frame + while cap.isOpened() and n < f: + n += 1 + # _, self.imgs[index] = cap.read() + cap.grab() + if n % read == 0: + success, im = cap.retrieve() + if success: + self.imgs[i] = im + else: + LOGGER.warning('WARNING: Video stream unresponsive, please check your IP camera connection.') + self.imgs[i] = np.zeros_like(self.imgs[i]) + cap.open(stream) # re-open stream if signal was lost + time.sleep(0.0) # wait time + + def __iter__(self): + self.count = -1 + return self + + def __next__(self): + self.count += 1 + if not all(x.is_alive() for x in self.threads) or cv2.waitKey(1) == ord('q'): # q to quit + cv2.destroyAllWindows() + raise StopIteration + + # Letterbox + img0 = self.imgs.copy() + img = [letterbox(x, self.img_size, stride=self.stride, auto=self.rect and self.auto)[0] for x in img0] + + # Stack + img = np.stack(img, 0) + + # Convert + img = img[..., ::-1].transpose((0, 3, 1, 2)) # BGR to RGB, BHWC to BCHW + img = np.ascontiguousarray(img) + + return self.sources, img, img0, None, '' + + def __len__(self): + return len(self.sources) # 1E12 frames = 32 streams at 30 FPS for 30 years + + +def img2label_paths(img_paths): + # Define label paths as a function of image paths + sa, sb = f'{os.sep}images{os.sep}', f'{os.sep}labels{os.sep}' # /images/, /labels/ substrings + return [sb.join(x.rsplit(sa, 1)).rsplit('.', 1)[0] + '.txt' for x in img_paths] + + +class LoadImagesAndLabels(Dataset): + # YOLOv5 train_loader/val_loader, loads images and labels for training and validation + cache_version = 0.6 # dataset labels *.cache version + rand_interp_methods = [cv2.INTER_NEAREST, cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_LANCZOS4] + + def __init__(self, + path, + img_size=640, + batch_size=16, + augment=False, + hyp=None, + rect=False, + image_weights=False, + cache_images=False, + single_cls=False, + stride=32, + pad=0.0, + prefix=''): + self.img_size = img_size + self.augment = augment + self.hyp = hyp + self.image_weights = image_weights + self.rect = False if image_weights else rect + self.mosaic = self.augment and not self.rect # load 4 images at a time into a mosaic (only during training) + self.mosaic_border = [-img_size // 2, -img_size // 2] + self.stride = stride + self.path = path + self.albumentations = Albumentations() if augment else None + + try: + f = [] # image files + for p in path if isinstance(path, list) else [path]: + p = Path(p) # os-agnostic + if p.is_dir(): # dir + f += glob.glob(str(p / '**' / '*.*'), recursive=True) + # f = list(p.rglob('*.*')) # pathlib + elif p.is_file(): # file + with open(p) as t: + t = t.read().strip().splitlines() + parent = str(p.parent) + os.sep + f += [x.replace('./', parent) if x.startswith('./') else x for x in t] # local to global path + # f += [p.parent / x.lstrip(os.sep) for x in t] # local to global path (pathlib) + else: + raise FileNotFoundError(f'{prefix}{p} does not exist') + self.im_files = sorted(x.replace('/', os.sep) for x in f if x.split('.')[-1].lower() in IMG_FORMATS) + # self.img_files = sorted([x for x in f if x.suffix[1:].lower() in IMG_FORMATS]) # pathlib + assert self.im_files, f'{prefix}No images found' + except Exception as e: + raise Exception(f'{prefix}Error loading data from {path}: {e}\nSee {HELP_URL}') + + # Check cache + self.label_files = img2label_paths(self.im_files) # labels + cache_path = (p if p.is_file() else Path(self.label_files[0]).parent).with_suffix('.cache') + try: + cache, exists = np.load(cache_path, allow_pickle=True).item(), True # load dict + assert cache['version'] == self.cache_version # matches current version + assert cache['hash'] == get_hash(self.label_files + self.im_files) # identical hash + except Exception: + cache, exists = self.cache_labels(cache_path, prefix), False # run cache ops + + # Display cache + nf, nm, ne, nc, n = cache.pop('results') # found, missing, empty, corrupt, total + if exists and LOCAL_RANK in {-1, 0}: + d = f"Scanning '{cache_path}' images and labels... {nf} found, {nm} missing, {ne} empty, {nc} corrupt" + tqdm(None, desc=prefix + d, total=n, initial=n, bar_format=BAR_FORMAT) # display cache results + if cache['msgs']: + LOGGER.info('\n'.join(cache['msgs'])) # display warnings + assert nf > 0 or not augment, f'{prefix}No labels in {cache_path}. Can not train without labels. See {HELP_URL}' + + # Read cache + [cache.pop(k) for k in ('hash', 'version', 'msgs')] # remove items + labels, shapes, self.segments = zip(*cache.values()) + self.labels = list(labels) + self.shapes = np.array(shapes) + self.im_files = list(cache.keys()) # update + self.label_files = img2label_paths(cache.keys()) # update + n = len(shapes) # number of images + bi = np.floor(np.arange(n) / batch_size).astype(np.int) # batch index + nb = bi[-1] + 1 # number of batches + self.batch = bi # batch index of image + self.n = n + self.indices = range(n) + + # Update labels + include_class = [] # filter labels to include only these classes (optional) + include_class_array = np.array(include_class).reshape(1, -1) + for i, (label, segment) in enumerate(zip(self.labels, self.segments)): + if include_class: + j = (label[:, 0:1] == include_class_array).any(1) + self.labels[i] = label[j] + if segment: + self.segments[i] = segment[j] + if single_cls: # single-class training, merge all classes into 0 + self.labels[i][:, 0] = 0 + if segment: + self.segments[i][:, 0] = 0 + + # Rectangular Training + if self.rect: + # Sort by aspect ratio + s = self.shapes # wh + ar = s[:, 1] / s[:, 0] # aspect ratio + irect = ar.argsort() + self.im_files = [self.im_files[i] for i in irect] + self.label_files = [self.label_files[i] for i in irect] + self.labels = [self.labels[i] for i in irect] + self.shapes = s[irect] # wh + ar = ar[irect] + + # Set training image shapes + shapes = [[1, 1]] * nb + for i in range(nb): + ari = ar[bi == i] + mini, maxi = ari.min(), ari.max() + if maxi < 1: + shapes[i] = [maxi, 1] + elif mini > 1: + shapes[i] = [1, 1 / mini] + + self.batch_shapes = np.ceil(np.array(shapes) * img_size / stride + pad).astype(np.int) * stride + + # Cache images into RAM/disk for faster training (WARNING: large datasets may exceed system resources) + self.ims = [None] * n + self.npy_files = [Path(f).with_suffix('.npy') for f in self.im_files] + if cache_images: + gb = 0 # Gigabytes of cached images + self.im_hw0, self.im_hw = [None] * n, [None] * n + fcn = self.cache_images_to_disk if cache_images == 'disk' else self.load_image + results = ThreadPool(NUM_THREADS).imap(fcn, range(n)) + pbar = tqdm(enumerate(results), total=n, bar_format=BAR_FORMAT, disable=LOCAL_RANK > 0) + for i, x in pbar: + if cache_images == 'disk': + gb += self.npy_files[i].stat().st_size + else: # 'ram' + self.ims[i], self.im_hw0[i], self.im_hw[i] = x # im, hw_orig, hw_resized = load_image(self, i) + gb += self.ims[i].nbytes + pbar.desc = f'{prefix}Caching images ({gb / 1E9:.1f}GB {cache_images})' + pbar.close() + + def cache_labels(self, path=Path('./labels.cache'), prefix=''): + # Cache dataset labels, check images and read shapes + x = {} # dict + nm, nf, ne, nc, msgs = 0, 0, 0, 0, [] # number missing, found, empty, corrupt, messages + desc = f"{prefix}Scanning '{path.parent / path.stem}' images and labels..." + with Pool(NUM_THREADS) as pool: + pbar = tqdm(pool.imap(verify_image_label, zip(self.im_files, self.label_files, repeat(prefix))), + desc=desc, + total=len(self.im_files), + bar_format=BAR_FORMAT) + for im_file, lb, shape, segments, nm_f, nf_f, ne_f, nc_f, msg in pbar: + nm += nm_f + nf += nf_f + ne += ne_f + nc += nc_f + if im_file: + x[im_file] = [lb, shape, segments] + if msg: + msgs.append(msg) + pbar.desc = f"{desc}{nf} found, {nm} missing, {ne} empty, {nc} corrupt" + + pbar.close() + if msgs: + LOGGER.info('\n'.join(msgs)) + if nf == 0: + LOGGER.warning(f'{prefix}WARNING: No labels found in {path}. See {HELP_URL}') + x['hash'] = get_hash(self.label_files + self.im_files) + x['results'] = nf, nm, ne, nc, len(self.im_files) + x['msgs'] = msgs # warnings + x['version'] = self.cache_version # cache version + try: + np.save(path, x) # save cache for next time + path.with_suffix('.cache.npy').rename(path) # remove .npy suffix + LOGGER.info(f'{prefix}New cache created: {path}') + except Exception as e: + LOGGER.warning(f'{prefix}WARNING: Cache directory {path.parent} is not writeable: {e}') # not writeable + return x + + def __len__(self): + return len(self.im_files) + + # def __iter__(self): + # self.count = -1 + # print('ran dataset iter') + # #self.shuffled_vector = np.random.permutation(self.nF) if self.augment else np.arange(self.nF) + # return self + + def __getitem__(self, index): + index = self.indices[index] # linear, shuffled, or image_weights + + hyp = self.hyp + mosaic = self.mosaic and random.random() < hyp['mosaic'] + if mosaic: + # Load mosaic + img, labels = self.load_mosaic(index) + shapes = None + + # MixUp augmentation + if random.random() < hyp['mixup']: + img, labels = mixup(img, labels, *self.load_mosaic(random.randint(0, self.n - 1))) + + else: + # Load image + img, (h0, w0), (h, w) = self.load_image(index) + + # Letterbox + shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size # final letterboxed shape + img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment) + shapes = (h0, w0), ((h / h0, w / w0), pad) # for COCO mAP rescaling + + labels = self.labels[index].copy() + if labels.size: # normalized xywh to pixel xyxy format + labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1]) + + if self.augment: + img, labels = random_perspective(img, + labels, + degrees=hyp['degrees'], + translate=hyp['translate'], + scale=hyp['scale'], + shear=hyp['shear'], + perspective=hyp['perspective']) + + nl = len(labels) # number of labels + if nl: + labels[:, 1:5] = xyxy2xywhn(labels[:, 1:5], w=img.shape[1], h=img.shape[0], clip=True, eps=1E-3) + + if self.augment: + # Albumentations + img, labels = self.albumentations(img, labels) + nl = len(labels) # update after albumentations + + # HSV color-space + augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v']) + + # Flip up-down + if random.random() < hyp['flipud']: + img = np.flipud(img) + if nl: + labels[:, 2] = 1 - labels[:, 2] + + # Flip left-right + if random.random() < hyp['fliplr']: + img = np.fliplr(img) + if nl: + labels[:, 1] = 1 - labels[:, 1] + + # Cutouts + # labels = cutout(img, labels, p=0.5) + # nl = len(labels) # update after cutout + + labels_out = torch.zeros((nl, 6)) + if nl: + labels_out[:, 1:] = torch.from_numpy(labels) + + # Convert + img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB + img = np.ascontiguousarray(img) + + return torch.from_numpy(img), labels_out, self.im_files[index], shapes + + def load_image(self, i): + # Loads 1 image from dataset index 'i', returns (im, original hw, resized hw) + im, f, fn = self.ims[i], self.im_files[i], self.npy_files[i], + if im is None: # not cached in RAM + if fn.exists(): # load npy + im = np.load(fn) + else: # read image + # print('idx:{} file:{}'.format(i, f)) + im = cv2.imread(f) # BGR + assert im is not None, f'Image Not Found {f}' + h0, w0 = im.shape[:2] # orig hw + r = self.img_size / max(h0, w0) # ratio + if r != 1: # if sizes are not equal + interp = cv2.INTER_LINEAR if (self.augment or r > 1) else cv2.INTER_AREA + # print('idx:{} r:{} interp:{}'.format(i, r, interp)) + im = cv2.resize(im, (int(w0 * r), int(h0 * r)), interpolation=interp) + return im, (h0, w0), im.shape[:2] # im, hw_original, hw_resized + return self.ims[i], self.im_hw0[i], self.im_hw[i] # im, hw_original, hw_resized + + def cache_images_to_disk(self, i): + # Saves an image as an *.npy file for faster loading + f = self.npy_files[i] + if not f.exists(): + np.save(f.as_posix(), cv2.imread(self.im_files[i])) + + def load_mosaic(self, index): + # YOLOv5 4-mosaic loader. Loads 1 image + 3 random images into a 4-image mosaic + labels4, segments4 = [], [] + s = self.img_size + yc, xc = (int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border) # mosaic center x, y + indices = [index] + random.choices(self.indices, k=3) # 3 additional image indices + random.shuffle(indices) + for i, index in enumerate(indices): + # Load image + img, _, (h, w) = self.load_image(index) + + # place img in img4 + if i == 0: # top left + img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles + x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc # xmin, ymin, xmax, ymax (large image) + x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h # xmin, ymin, xmax, ymax (small image) + elif i == 1: # top right + x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc + x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h + elif i == 2: # bottom left + x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h) + x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h) + elif i == 3: # bottom right + x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h) + x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h) + + img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b] # img4[ymin:ymax, xmin:xmax] + padw = x1a - x1b + padh = y1a - y1b + + # Labels + labels, segments = self.labels[index].copy(), self.segments[index].copy() + if labels.size: + labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh) # normalized xywh to pixel xyxy format + segments = [xyn2xy(x, w, h, padw, padh) for x in segments] + labels4.append(labels) + segments4.extend(segments) + + # Concat/clip labels + labels4 = np.concatenate(labels4, 0) + for x in (labels4[:, 1:], *segments4): + np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective() + # img4, labels4 = replicate(img4, labels4) # replicate + + # Augment + img4, labels4, segments4 = copy_paste(img4, labels4, segments4, p=self.hyp['copy_paste']) + img4, labels4 = random_perspective(img4, + labels4, + segments4, + degrees=self.hyp['degrees'], + translate=self.hyp['translate'], + scale=self.hyp['scale'], + shear=self.hyp['shear'], + perspective=self.hyp['perspective'], + border=self.mosaic_border) # border to remove + + return img4, labels4 + + def load_mosaic9(self, index): + # YOLOv5 9-mosaic loader. Loads 1 image + 8 random images into a 9-image mosaic + labels9, segments9 = [], [] + s = self.img_size + indices = [index] + random.choices(self.indices, k=8) # 8 additional image indices + random.shuffle(indices) + hp, wp = -1, -1 # height, width previous + for i, index in enumerate(indices): + # Load image + img, _, (h, w) = self.load_image(index) + + # place img in img9 + if i == 0: # center + img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8) # base image with 4 tiles + h0, w0 = h, w + c = s, s, s + w, s + h # xmin, ymin, xmax, ymax (base) coordinates + elif i == 1: # top + c = s, s - h, s + w, s + elif i == 2: # top right + c = s + wp, s - h, s + wp + w, s + elif i == 3: # right + c = s + w0, s, s + w0 + w, s + h + elif i == 4: # bottom right + c = s + w0, s + hp, s + w0 + w, s + hp + h + elif i == 5: # bottom + c = s + w0 - w, s + h0, s + w0, s + h0 + h + elif i == 6: # bottom left + c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h + elif i == 7: # left + c = s - w, s + h0 - h, s, s + h0 + elif i == 8: # top left + c = s - w, s + h0 - hp - h, s, s + h0 - hp + + padx, pady = c[:2] + x1, y1, x2, y2 = (max(x, 0) for x in c) # allocate coords + + # Labels + labels, segments = self.labels[index].copy(), self.segments[index].copy() + if labels.size: + labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padx, pady) # normalized xywh to pixel xyxy format + segments = [xyn2xy(x, w, h, padx, pady) for x in segments] + labels9.append(labels) + segments9.extend(segments) + + # Image + img9[y1:y2, x1:x2] = img[y1 - pady:, x1 - padx:] # img9[ymin:ymax, xmin:xmax] + hp, wp = h, w # height, width previous + + # Offset + yc, xc = (int(random.uniform(0, s)) for _ in self.mosaic_border) # mosaic center x, y + img9 = img9[yc:yc + 2 * s, xc:xc + 2 * s] + + # Concat/clip labels + labels9 = np.concatenate(labels9, 0) + labels9[:, [1, 3]] -= xc + labels9[:, [2, 4]] -= yc + c = np.array([xc, yc]) # centers + segments9 = [x - c for x in segments9] + + for x in (labels9[:, 1:], *segments9): + np.clip(x, 0, 2 * s, out=x) # clip when using random_perspective() + # img9, labels9 = replicate(img9, labels9) # replicate + + # Augment + img9, labels9 = random_perspective(img9, + labels9, + segments9, + degrees=self.hyp['degrees'], + translate=self.hyp['translate'], + scale=self.hyp['scale'], + shear=self.hyp['shear'], + perspective=self.hyp['perspective'], + border=self.mosaic_border) # border to remove + + return img9, labels9 + + @staticmethod + def collate_fn(batch): + im, label, path, shapes = zip(*batch) # transposed + for i, lb in enumerate(label): + lb[:, 0] = i # add target image index for build_targets() + return torch.stack(im, 0), torch.cat(label, 0), path, shapes + + @staticmethod + def collate_fn4(batch): + img, label, path, shapes = zip(*batch) # transposed + n = len(shapes) // 4 + im4, label4, path4, shapes4 = [], [], path[:n], shapes[:n] + + ho = torch.tensor([[0.0, 0, 0, 1, 0, 0]]) + wo = torch.tensor([[0.0, 0, 1, 0, 0, 0]]) + s = torch.tensor([[1, 1, 0.5, 0.5, 0.5, 0.5]]) # scale + for i in range(n): # zidane torch.zeros(16,3,720,1280) # BCHW + i *= 4 + if random.random() < 0.5: + im = F.interpolate(img[i].unsqueeze(0).float(), scale_factor=2.0, mode='bilinear', + align_corners=False)[0].type(img[i].type()) + lb = label[i] + else: + im = torch.cat((torch.cat((img[i], img[i + 1]), 1), torch.cat((img[i + 2], img[i + 3]), 1)), 2) + lb = torch.cat((label[i], label[i + 1] + ho, label[i + 2] + wo, label[i + 3] + ho + wo), 0) * s + im4.append(im) + label4.append(lb) + + for i, lb in enumerate(label4): + lb[:, 0] = i # add target image index for build_targets() + + return torch.stack(im4, 0), torch.cat(label4, 0), path4, shapes4 + + +# Ancillary functions -------------------------------------------------------------------------------------------------- +def flatten_recursive(path=DATASETS_DIR / 'coco128'): + # Flatten a recursive directory by bringing all files to top level + new_path = Path(f'{str(path)}_flat') + if os.path.exists(new_path): + shutil.rmtree(new_path) # delete output folder + os.makedirs(new_path) # make new output folder + for file in tqdm(glob.glob(f'{str(Path(path))}/**/*.*', recursive=True)): + shutil.copyfile(file, new_path / Path(file).name) + + +def extract_boxes(path=DATASETS_DIR / 'coco128'): # from utils.dataloaders import *; extract_boxes() + # Convert detection dataset into classification dataset, with one directory per class + path = Path(path) # images dir + shutil.rmtree(path / 'classification') if (path / 'classification').is_dir() else None # remove existing + files = list(path.rglob('*.*')) + n = len(files) # number of files + for im_file in tqdm(files, total=n): + if im_file.suffix[1:] in IMG_FORMATS: + # image + im = cv2.imread(str(im_file))[..., ::-1] # BGR to RGB + h, w = im.shape[:2] + + # labels + lb_file = Path(img2label_paths([str(im_file)])[0]) + if Path(lb_file).exists(): + with open(lb_file) as f: + lb = np.array([x.split() for x in f.read().strip().splitlines()], dtype=np.float32) # labels + + for j, x in enumerate(lb): + c = int(x[0]) # class + f = (path / 'classifier') / f'{c}' / f'{path.stem}_{im_file.stem}_{j}.jpg' # new filename + if not f.parent.is_dir(): + f.parent.mkdir(parents=True) + + b = x[1:] * [w, h, w, h] # box + # b[2:] = b[2:].max() # rectangle to square + b[2:] = b[2:] * 1.2 + 3 # pad + b = xywh2xyxy(b.reshape(-1, 4)).ravel().astype(np.int) + + b[[0, 2]] = np.clip(b[[0, 2]], 0, w) # clip boxes outside of image + b[[1, 3]] = np.clip(b[[1, 3]], 0, h) + assert cv2.imwrite(str(f), im[b[1]:b[3], b[0]:b[2]]), f'box failure in {f}' + + +def autosplit(path=DATASETS_DIR / 'coco128/images', weights=(0.9, 0.1, 0.0), annotated_only=False): + """ Autosplit a dataset into train/val/test splits and save path/autosplit_*.txt files + Usage: from utils.dataloaders import *; autosplit() + Arguments + path: Path to images directory + weights: Train, val, test weights (list, tuple) + annotated_only: Only use images with an annotated txt file + """ + path = Path(path) # images dir + files = sorted(x for x in path.rglob('*.*') if x.suffix[1:].lower() in IMG_FORMATS) # image files only + n = len(files) # number of files + random.seed(0) # for reproducibility + indices = random.choices([0, 1, 2], weights=weights, k=n) # assign each image to a split + + txt = ['autosplit_train.txt', 'autosplit_val.txt', 'autosplit_test.txt'] # 3 txt files + [(path.parent / x).unlink(missing_ok=True) for x in txt] # remove existing + + print(f'Autosplitting images from {path}' + ', using *.txt labeled images only' * annotated_only) + for i, img in tqdm(zip(indices, files), total=n): + if not annotated_only or Path(img2label_paths([str(img)])[0]).exists(): # check label + with open(path.parent / txt[i], 'a') as f: + f.write(f'./{img.relative_to(path.parent).as_posix()}' + '\n') # add image to txt file + + +def verify_image_label(args): + # Verify one image-label pair + im_file, lb_file, prefix = args + nm, nf, ne, nc, msg, segments = 0, 0, 0, 0, '', [] # number (missing, found, empty, corrupt), message, segments + try: + # verify images + im = Image.open(im_file) + im.verify() # PIL verify + shape = exif_size(im) # image size + assert (shape[0] > 9) & (shape[1] > 9), f'image size {shape} <10 pixels' + assert im.format.lower() in IMG_FORMATS, f'invalid image format {im.format}' + if im.format.lower() in ('jpg', 'jpeg'): + with open(im_file, 'rb') as f: + f.seek(-2, 2) + if f.read() != b'\xff\xd9': # corrupt JPEG + ImageOps.exif_transpose(Image.open(im_file)).save(im_file, 'JPEG', subsampling=0, quality=100) + msg = f'{prefix}WARNING: {im_file}: corrupt JPEG restored and saved' + + # verify labels + if os.path.isfile(lb_file): + nf = 1 # label found + with open(lb_file) as f: + lb = [x.split() for x in f.read().strip().splitlines() if len(x)] + if any(len(x) > 6 for x in lb): # is segment + classes = np.array([x[0] for x in lb], dtype=np.float32) + segments = [np.array(x[1:], dtype=np.float32).reshape(-1, 2) for x in lb] # (cls, xy1...) + lb = np.concatenate((classes.reshape(-1, 1), segments2boxes(segments)), 1) # (cls, xywh) + lb = np.array(lb, dtype=np.float32) + nl = len(lb) + if nl: + assert lb.shape[1] == 5, f'labels require 5 columns, {lb.shape[1]} columns detected' + assert (lb >= 0).all(), f'negative label values {lb[lb < 0]}' + assert (lb[:, 1:] <= 1).all(), f'non-normalized or out of bounds coordinates {lb[:, 1:][lb[:, 1:] > 1]}' + _, i = np.unique(lb, axis=0, return_index=True) + if len(i) < nl: # duplicate row check + lb = lb[i] # remove duplicates + if segments: + segments = segments[i] + msg = f'{prefix}WARNING: {im_file}: {nl - len(i)} duplicate labels removed' + else: + ne = 1 # label empty + lb = np.zeros((0, 5), dtype=np.float32) + else: + nm = 1 # label missing + lb = np.zeros((0, 5), dtype=np.float32) + return im_file, lb, shape, segments, nm, nf, ne, nc, msg + except Exception as e: + nc = 1 + msg = f'{prefix}WARNING: {im_file}: ignoring corrupt image/label: {e}' + return [None, None, None, None, nm, nf, ne, nc, msg] + + +class HUBDatasetStats(): + """ Return dataset statistics dictionary with images and instances counts per split per class + To run in parent directory: export PYTHONPATH="$PWD/yolov5" + Usage1: from utils.dataloaders import *; HUBDatasetStats('coco128.yaml', autodownload=True) + Usage2: from utils.dataloaders import *; HUBDatasetStats('path/to/coco128_with_yaml.zip') + Arguments + path: Path to data.yaml or data.zip (with data.yaml inside data.zip) + autodownload: Attempt to download dataset if not found locally + """ + + def __init__(self, path='coco128.yaml', autodownload=False): + # Initialize class + zipped, data_dir, yaml_path = self._unzip(Path(path)) + try: + with open(check_yaml(yaml_path), errors='ignore') as f: + data = yaml.safe_load(f) # data dict + if zipped: + data['path'] = data_dir + except Exception as e: + raise Exception("error/HUB/dataset_stats/yaml_load") from e + + check_dataset(data, autodownload) # download dataset if missing + self.hub_dir = Path(data['path'] + '-hub') + self.im_dir = self.hub_dir / 'images' + self.im_dir.mkdir(parents=True, exist_ok=True) # makes /images + self.stats = {'nc': data['nc'], 'names': data['names']} # statistics dictionary + self.data = data + + @staticmethod + def _find_yaml(dir): + # Return data.yaml file + files = list(dir.glob('*.yaml')) or list(dir.rglob('*.yaml')) # try root level first and then recursive + assert files, f'No *.yaml file found in {dir}' + if len(files) > 1: + files = [f for f in files if f.stem == dir.stem] # prefer *.yaml files that match dir name + assert files, f'Multiple *.yaml files found in {dir}, only 1 *.yaml file allowed' + assert len(files) == 1, f'Multiple *.yaml files found: {files}, only 1 *.yaml file allowed in {dir}' + return files[0] + + def _unzip(self, path): + # Unzip data.zip + if not str(path).endswith('.zip'): # path is data.yaml + return False, None, path + assert Path(path).is_file(), f'Error unzipping {path}, file not found' + ZipFile(path).extractall(path=path.parent) # unzip + dir = path.with_suffix('') # dataset directory == zip name + assert dir.is_dir(), f'Error unzipping {path}, {dir} not found. path/to/abc.zip MUST unzip to path/to/abc/' + return True, str(dir), self._find_yaml(dir) # zipped, data_dir, yaml_path + + def _hub_ops(self, f, max_dim=1920): + # HUB ops for 1 image 'f': resize and save at reduced quality in /dataset-hub for web/app viewing + f_new = self.im_dir / Path(f).name # dataset-hub image filename + try: # use PIL + im = Image.open(f) + r = max_dim / max(im.height, im.width) # ratio + if r < 1.0: # image too large + im = im.resize((int(im.width * r), int(im.height * r))) + im.save(f_new, 'JPEG', quality=50, optimize=True) # save + except Exception as e: # use OpenCV + print(f'WARNING: HUB ops PIL failure {f}: {e}') + im = cv2.imread(f) + im_height, im_width = im.shape[:2] + r = max_dim / max(im_height, im_width) # ratio + if r < 1.0: # image too large + im = cv2.resize(im, (int(im_width * r), int(im_height * r)), interpolation=cv2.INTER_AREA) + cv2.imwrite(str(f_new), im) + + def get_json(self, save=False, verbose=False): + # Return dataset JSON for Ultralytics HUB + def _round(labels): + # Update labels to integer class and 6 decimal place floats + return [[int(c), *(round(x, 4) for x in points)] for c, *points in labels] + + for split in 'train', 'val', 'test': + if self.data.get(split) is None: + self.stats[split] = None # i.e. no test set + continue + dataset = LoadImagesAndLabels(self.data[split]) # load dataset + x = np.array([ + np.bincount(label[:, 0].astype(int), minlength=self.data['nc']) + for label in tqdm(dataset.labels, total=dataset.n, desc='Statistics')]) # shape(128x80) + self.stats[split] = { + 'instance_stats': { + 'total': int(x.sum()), + 'per_class': x.sum(0).tolist()}, + 'image_stats': { + 'total': dataset.n, + 'unlabelled': int(np.all(x == 0, 1).sum()), + 'per_class': (x > 0).sum(0).tolist()}, + 'labels': [{ + str(Path(k).name): _round(v.tolist())} for k, v in zip(dataset.im_files, dataset.labels)]} + + # Save, print and return + if save: + stats_path = self.hub_dir / 'stats.json' + print(f'Saving {stats_path.resolve()}...') + with open(stats_path, 'w') as f: + json.dump(self.stats, f) # save stats.json + if verbose: + print(json.dumps(self.stats, indent=2, sort_keys=False)) + return self.stats + + def process_images(self): + # Compress images for Ultralytics HUB + for split in 'train', 'val', 'test': + if self.data.get(split) is None: + continue + dataset = LoadImagesAndLabels(self.data[split]) # load dataset + desc = f'{split} images' + for _ in tqdm(ThreadPool(NUM_THREADS).imap(self._hub_ops, dataset.im_files), total=dataset.n, desc=desc): + pass + print(f'Done. All images saved to {self.im_dir}') + return self.im_dir + + +# Classification dataloaders ------------------------------------------------------------------------------------------- +class ClassificationDataset(torchvision.datasets.ImageFolder): + """ + YOLOv5 Classification Dataset. + Arguments + root: Dataset path + transform: torchvision transforms, used by default + album_transform: Albumentations transforms, used if installed + """ + + def __init__(self, root, augment, imgsz, cache=False): + super().__init__(root=root) + self.torch_transforms = classify_transforms(imgsz) + self.album_transforms = classify_albumentations(augment, imgsz) if augment else None + self.cache_ram = cache is True or cache == 'ram' + self.cache_disk = cache == 'disk' + self.samples = [list(x) + [Path(x[0]).with_suffix('.npy'), None] for x in self.samples] # file, index, npy, im + + def __getitem__(self, i): + f, j, fn, im = self.samples[i] # filename, index, filename.with_suffix('.npy'), image + if self.album_transforms: + if self.cache_ram and im is None: + im = self.samples[i][3] = cv2.imread(f) + elif self.cache_disk: + if not fn.exists(): # load npy + np.save(fn.as_posix(), cv2.imread(f)) + im = np.load(fn) + else: # read image + im = cv2.imread(f) # BGR + sample = self.album_transforms(image=cv2.cvtColor(im, cv2.COLOR_BGR2RGB))["image"] + else: + sample = self.torch_transforms(self.loader(f)) + return sample, j + + +def create_classification_dataloader(path, + imgsz=224, + batch_size=16, + augment=True, + cache=False, + rank=-1, + workers=8, + shuffle=True): + # Returns Dataloader object to be used with YOLOv5 Classifier + with torch_distributed_zero_first(rank): # init dataset *.cache only once if DDP + dataset = ClassificationDataset(root=path, imgsz=imgsz, augment=augment, cache=cache) + batch_size = min(batch_size, len(dataset)) + nd = torch.cuda.device_count() + nw = min([os.cpu_count() // max(nd, 1), batch_size if batch_size > 1 else 0, workers]) + sampler = None if rank == -1 else distributed.DistributedSampler(dataset, shuffle=shuffle) + generator = torch.Generator() + generator.manual_seed(0) + return InfiniteDataLoader(dataset, + batch_size=batch_size, + shuffle=shuffle and sampler is None, + num_workers=nw, + sampler=sampler, + pin_memory=True, + worker_init_fn=seed_worker, + generator=generator) # or DataLoader(persistent_workers=True) diff --git a/application/yolov5_example/utils/docker/Dockerfile b/application/yolov5_example/utils/docker/Dockerfile new file mode 100644 index 00000000..2280f209 --- /dev/null +++ b/application/yolov5_example/utils/docker/Dockerfile @@ -0,0 +1,68 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Builds ultralytics/yolov5:latest image on DockerHub https://hub.docker.com/r/ultralytics/yolov5 +# Image is CUDA-optimized for YOLOv5 single/multi-GPU training and inference + +# Start FROM NVIDIA PyTorch image https://ngc.nvidia.com/catalog/containers/nvidia:pytorch +FROM nvcr.io/nvidia/pytorch:22.07-py3 +RUN rm -rf /opt/pytorch # remove 1.2GB dir + +# Downloads to user config dir +ADD https://ultralytics.com/assets/Arial.ttf https://ultralytics.com/assets/Arial.Unicode.ttf /root/.config/Ultralytics/ + +# Install linux packages +RUN apt update && apt install --no-install-recommends -y zip htop screen libgl1-mesa-glx + +# Install pip packages +COPY requirements.txt . +RUN python -m pip install --upgrade pip wheel +RUN pip uninstall -y Pillow torchtext # torch torchvision +RUN pip install --no-cache -r requirements.txt albumentations wandb gsutil notebook Pillow>=9.1.0 \ + 'opencv-python<4.6.0.66' \ + --extra-index-url https://download.pytorch.org/whl/cu113 + +# Create working directory +RUN mkdir -p /usr/src/app +WORKDIR /usr/src/app + +# Copy contents +# COPY . /usr/src/app (issues as not a .git directory) +RUN git clone https://github.com/ultralytics/yolov5 /usr/src/app + +# Set environment variables +ENV OMP_NUM_THREADS=8 + + +# Usage Examples ------------------------------------------------------------------------------------------------------- + +# Build and Push +# t=ultralytics/yolov5:latest && sudo docker build -f utils/docker/Dockerfile -t $t . && sudo docker push $t + +# Pull and Run +# t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all $t + +# Pull and Run with local directory access +# t=ultralytics/yolov5:latest && sudo docker pull $t && sudo docker run -it --ipc=host --gpus all -v "$(pwd)"/datasets:/usr/src/datasets $t + +# Kill all +# sudo docker kill $(sudo docker ps -q) + +# Kill all image-based +# sudo docker kill $(sudo docker ps -qa --filter ancestor=ultralytics/yolov5:latest) + +# Bash into running container +# sudo docker exec -it 5a9b5863d93d bash + +# Bash into stopped container +# id=$(sudo docker ps -qa) && sudo docker start $id && sudo docker exec -it $id bash + +# Clean up +# docker system prune -a --volumes + +# Update Ubuntu drivers +# https://www.maketecheasier.com/install-nvidia-drivers-ubuntu/ + +# DDP test +# python -m torch.distributed.run --nproc_per_node 2 --master_port 1 train.py --epochs 3 + +# GCP VM from Image +# docker.io/ultralytics/yolov5:latest diff --git a/application/yolov5_example/utils/docker/Dockerfile-arm64 b/application/yolov5_example/utils/docker/Dockerfile-arm64 new file mode 100644 index 00000000..fe92c8d5 --- /dev/null +++ b/application/yolov5_example/utils/docker/Dockerfile-arm64 @@ -0,0 +1,42 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Builds ultralytics/yolov5:latest-arm64 image on DockerHub https://hub.docker.com/r/ultralytics/yolov5 +# Image is aarch64-compatible for Apple M1 and other ARM architectures i.e. Jetson Nano and Raspberry Pi + +# Start FROM Ubuntu image https://hub.docker.com/_/ubuntu +FROM arm64v8/ubuntu:20.04 + +# Downloads to user config dir +ADD https://ultralytics.com/assets/Arial.ttf https://ultralytics.com/assets/Arial.Unicode.ttf /root/.config/Ultralytics/ + +# Install linux packages +RUN apt update +RUN DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt install -y tzdata +RUN apt install --no-install-recommends -y python3-pip git zip curl htop gcc \ + libgl1-mesa-glx libglib2.0-0 libpython3.8-dev +# RUN alias python=python3 + +# Install pip packages +COPY requirements.txt . +RUN python3 -m pip install --upgrade pip wheel +RUN pip install --no-cache -r requirements.txt gsutil notebook \ + tensorflow-aarch64 + # tensorflowjs \ + # onnx onnx-simplifier onnxruntime \ + # coremltools openvino-dev \ + +# Create working directory +RUN mkdir -p /usr/src/app +WORKDIR /usr/src/app + +# Copy contents +# COPY . /usr/src/app (issues as not a .git directory) +RUN git clone https://github.com/ultralytics/yolov5 /usr/src/app + + +# Usage Examples ------------------------------------------------------------------------------------------------------- + +# Build and Push +# t=ultralytics/yolov5:latest-M1 && sudo docker build --platform linux/arm64 -f utils/docker/Dockerfile-arm64 -t $t . && sudo docker push $t + +# Pull and Run +# t=ultralytics/yolov5:latest-M1 && sudo docker pull $t && sudo docker run -it --ipc=host -v "$(pwd)"/datasets:/usr/src/datasets $t diff --git a/application/yolov5_example/utils/docker/Dockerfile-cpu b/application/yolov5_example/utils/docker/Dockerfile-cpu new file mode 100644 index 00000000..d61dfeff --- /dev/null +++ b/application/yolov5_example/utils/docker/Dockerfile-cpu @@ -0,0 +1,39 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +# Builds ultralytics/yolov5:latest-cpu image on DockerHub https://hub.docker.com/r/ultralytics/yolov5 +# Image is CPU-optimized for ONNX, OpenVINO and PyTorch YOLOv5 deployments + +# Start FROM Ubuntu image https://hub.docker.com/_/ubuntu +FROM ubuntu:20.04 + +# Downloads to user config dir +ADD https://ultralytics.com/assets/Arial.ttf https://ultralytics.com/assets/Arial.Unicode.ttf /root/.config/Ultralytics/ + +# Install linux packages +RUN apt update +RUN DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC apt install -y tzdata +RUN apt install --no-install-recommends -y python3-pip git zip curl htop libgl1-mesa-glx libglib2.0-0 libpython3.8-dev +# RUN alias python=python3 + +# Install pip packages +COPY requirements.txt . +RUN python3 -m pip install --upgrade pip wheel +RUN pip install --no-cache -r requirements.txt albumentations gsutil notebook \ + coremltools onnx onnx-simplifier onnxruntime openvino-dev tensorflow-cpu tensorflowjs \ + --extra-index-url https://download.pytorch.org/whl/cpu + +# Create working directory +RUN mkdir -p /usr/src/app +WORKDIR /usr/src/app + +# Copy contents +# COPY . /usr/src/app (issues as not a .git directory) +RUN git clone https://github.com/ultralytics/yolov5 /usr/src/app + + +# Usage Examples ------------------------------------------------------------------------------------------------------- + +# Build and Push +# t=ultralytics/yolov5:latest-cpu && sudo docker build -f utils/docker/Dockerfile-cpu -t $t . && sudo docker push $t + +# Pull and Run +# t=ultralytics/yolov5:latest-cpu && sudo docker pull $t && sudo docker run -it --ipc=host -v "$(pwd)"/datasets:/usr/src/datasets $t diff --git a/application/yolov5_example/utils/downloads.py b/application/yolov5_example/utils/downloads.py new file mode 100644 index 00000000..9d4780ad --- /dev/null +++ b/application/yolov5_example/utils/downloads.py @@ -0,0 +1,180 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Download utils +""" + +import logging +import os +import platform +import subprocess +import time +import urllib +from pathlib import Path +from zipfile import ZipFile + +import requests +import torch + + +def is_url(url, check_online=True): + # Check if online file exists + try: + url = str(url) + result = urllib.parse.urlparse(url) + assert all([result.scheme, result.netloc, result.path]) # check if is url + return (urllib.request.urlopen(url).getcode() == 200) if check_online else True # check if exists online + except (AssertionError, urllib.request.HTTPError): + return False + + +def gsutil_getsize(url=''): + # gs://bucket/file size https://cloud.google.com/storage/docs/gsutil/commands/du + s = subprocess.check_output(f'gsutil du {url}', shell=True).decode('utf-8') + return eval(s.split(' ')[0]) if len(s) else 0 # bytes + + +def safe_download(file, url, url2=None, min_bytes=1E0, error_msg=''): + # Attempts to download file from url or url2, checks and removes incomplete downloads < min_bytes + from utils.general import LOGGER + + file = Path(file) + assert_msg = f"Downloaded file '{file}' does not exist or size is < min_bytes={min_bytes}" + try: # url1 + LOGGER.info(f'Downloading {url} to {file}...') + torch.hub.download_url_to_file(url, str(file), progress=LOGGER.level <= logging.INFO) + assert file.exists() and file.stat().st_size > min_bytes, assert_msg # check + except Exception as e: # url2 + file.unlink(missing_ok=True) # remove partial downloads + LOGGER.info(f'ERROR: {e}\nRe-attempting {url2 or url} to {file}...') + os.system(f"curl -L '{url2 or url}' -o '{file}' --retry 3 -C -") # curl download, retry and resume on fail + finally: + if not file.exists() or file.stat().st_size < min_bytes: # check + file.unlink(missing_ok=True) # remove partial downloads + LOGGER.info(f"ERROR: {assert_msg}\n{error_msg}") + LOGGER.info('') + + +def attempt_download(file, repo='ultralytics/yolov5', release='v6.1'): + # Attempt file download from GitHub release assets if not found locally. release = 'latest', 'v6.1', etc. + from utils.general import LOGGER + + def github_assets(repository, version='latest'): + # Return GitHub repo tag (i.e. 'v6.1') and assets (i.e. ['yolov5s.pt', 'yolov5m.pt', ...]) + if version != 'latest': + version = f'tags/{version}' # i.e. tags/v6.1 + response = requests.get(f'https://api.github.com/repos/{repository}/releases/{version}').json() # github api + return response['tag_name'], [x['name'] for x in response['assets']] # tag, assets + + file = Path(str(file).strip().replace("'", '')) + if not file.exists(): + # URL specified + name = Path(urllib.parse.unquote(str(file))).name # decode '%2F' to '/' etc. + if str(file).startswith(('http:/', 'https:/')): # download + url = str(file).replace(':/', '://') # Pathlib turns :// -> :/ + file = name.split('?')[0] # parse authentication https://url.com/file.txt?auth... + if Path(file).is_file(): + LOGGER.info(f'Found {url} locally at {file}') # file already exists + else: + safe_download(file=file, url=url, min_bytes=1E5) + return file + + # GitHub assets + assets = [ + 'yolov5n.pt', 'yolov5s.pt', 'yolov5m.pt', 'yolov5l.pt', 'yolov5x.pt', 'yolov5n6.pt', 'yolov5s6.pt', + 'yolov5m6.pt', 'yolov5l6.pt', 'yolov5x6.pt'] + try: + tag, assets = github_assets(repo, release) + except Exception: + try: + tag, assets = github_assets(repo) # latest release + except Exception: + try: + tag = subprocess.check_output('git tag', shell=True, stderr=subprocess.STDOUT).decode().split()[-1] + except Exception: + tag = release + + file.parent.mkdir(parents=True, exist_ok=True) # make parent dir (if required) + if name in assets: + url3 = 'https://drive.google.com/drive/folders/1EFQTEUeXWSFww0luse2jB9M1QNZQGwNl' # backup gdrive mirror + safe_download( + file, + url=f'https://github.com/{repo}/releases/download/{tag}/{name}', + url2=f'https://storage.googleapis.com/{repo}/{tag}/{name}', # backup url (optional) + min_bytes=1E5, + error_msg=f'{file} missing, try downloading from https://github.com/{repo}/releases/{tag} or {url3}') + + return str(file) + + +def gdrive_download(id='16TiPfZj7htmTyhntwcZyEEAejOUxuT6m', file='tmp.zip'): + # Downloads a file from Google Drive. from yolov5.utils.downloads import *; gdrive_download() + t = time.time() + file = Path(file) + cookie = Path('cookie') # gdrive cookie + print(f'Downloading https://drive.google.com/uc?export=download&id={id} as {file}... ', end='') + file.unlink(missing_ok=True) # remove existing file + cookie.unlink(missing_ok=True) # remove existing cookie + + # Attempt file download + out = "NUL" if platform.system() == "Windows" else "/dev/null" + os.system(f'curl -c ./cookie -s -L "drive.google.com/uc?export=download&id={id}" > {out}') + if os.path.exists('cookie'): # large file + s = f'curl -Lb ./cookie "drive.google.com/uc?export=download&confirm={get_token()}&id={id}" -o {file}' + else: # small file + s = f'curl -s -L -o {file} "drive.google.com/uc?export=download&id={id}"' + r = os.system(s) # execute, capture return + cookie.unlink(missing_ok=True) # remove existing cookie + + # Error check + if r != 0: + file.unlink(missing_ok=True) # remove partial + print('Download error ') # raise Exception('Download error') + return r + + # Unzip if archive + if file.suffix == '.zip': + print('unzipping... ', end='') + ZipFile(file).extractall(path=file.parent) # unzip + file.unlink() # remove zip + + print(f'Done ({time.time() - t:.1f}s)') + return r + + +def get_token(cookie="./cookie"): + with open(cookie) as f: + for line in f: + if "download" in line: + return line.split()[-1] + return "" + + +# Google utils: https://cloud.google.com/storage/docs/reference/libraries ---------------------------------------------- +# +# +# def upload_blob(bucket_name, source_file_name, destination_blob_name): +# # Uploads a file to a bucket +# # https://cloud.google.com/storage/docs/uploading-objects#storage-upload-object-python +# +# storage_client = storage.Client() +# bucket = storage_client.get_bucket(bucket_name) +# blob = bucket.blob(destination_blob_name) +# +# blob.upload_from_filename(source_file_name) +# +# print('File {} uploaded to {}.'.format( +# source_file_name, +# destination_blob_name)) +# +# +# def download_blob(bucket_name, source_blob_name, destination_file_name): +# # Uploads a blob from a bucket +# storage_client = storage.Client() +# bucket = storage_client.get_bucket(bucket_name) +# blob = bucket.blob(source_blob_name) +# +# blob.download_to_filename(destination_file_name) +# +# print('Blob {} downloaded to {}.'.format( +# source_blob_name, +# destination_file_name)) diff --git a/application/yolov5_example/utils/flask_rest_api/README.md b/application/yolov5_example/utils/flask_rest_api/README.md new file mode 100644 index 00000000..a726acbd --- /dev/null +++ b/application/yolov5_example/utils/flask_rest_api/README.md @@ -0,0 +1,73 @@ +# Flask REST API + +[REST](https://en.wikipedia.org/wiki/Representational_state_transfer) [API](https://en.wikipedia.org/wiki/API)s are +commonly used to expose Machine Learning (ML) models to other services. This folder contains an example REST API +created using Flask to expose the YOLOv5s model from [PyTorch Hub](https://pytorch.org/hub/ultralytics_yolov5/). + +## Requirements + +[Flask](https://palletsprojects.com/p/flask/) is required. Install with: + +```shell +$ pip install Flask +``` + +## Run + +After Flask installation run: + +```shell +$ python3 restapi.py --port 5000 +``` + +Then use [curl](https://curl.se/) to perform a request: + +```shell +$ curl -X POST -F image=@zidane.jpg 'http://localhost:5000/v1/object-detection/yolov5s' +``` + +The model inference results are returned as a JSON response: + +```json +[ + { + "class": 0, + "confidence": 0.8900438547, + "height": 0.9318675399, + "name": "person", + "width": 0.3264600933, + "xcenter": 0.7438579798, + "ycenter": 0.5207948685 + }, + { + "class": 0, + "confidence": 0.8440024257, + "height": 0.7155083418, + "name": "person", + "width": 0.6546785235, + "xcenter": 0.427829951, + "ycenter": 0.6334488392 + }, + { + "class": 27, + "confidence": 0.3771208823, + "height": 0.3902671337, + "name": "tie", + "width": 0.0696444362, + "xcenter": 0.3675483763, + "ycenter": 0.7991207838 + }, + { + "class": 27, + "confidence": 0.3527112305, + "height": 0.1540903747, + "name": "tie", + "width": 0.0336618312, + "xcenter": 0.7814827561, + "ycenter": 0.5065554976 + } +] +``` + +An example python script to perform inference using [requests](https://docs.python-requests.org/en/master/) is given +in `example_request.py` diff --git a/application/yolov5_example/utils/flask_rest_api/example_request.py b/application/yolov5_example/utils/flask_rest_api/example_request.py new file mode 100644 index 00000000..773ad893 --- /dev/null +++ b/application/yolov5_example/utils/flask_rest_api/example_request.py @@ -0,0 +1,19 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Perform test request +""" + +import pprint + +import requests + +DETECTION_URL = "http://localhost:5000/v1/object-detection/yolov5s" +IMAGE = "zidane.jpg" + +# Read image +with open(IMAGE, "rb") as f: + image_data = f.read() + +response = requests.post(DETECTION_URL, files={"image": image_data}).json() + +pprint.pprint(response) diff --git a/application/yolov5_example/utils/flask_rest_api/restapi.py b/application/yolov5_example/utils/flask_rest_api/restapi.py new file mode 100644 index 00000000..8482435c --- /dev/null +++ b/application/yolov5_example/utils/flask_rest_api/restapi.py @@ -0,0 +1,48 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Run a Flask REST API exposing one or more YOLOv5s models +""" + +import argparse +import io + +import torch +from flask import Flask, request +from PIL import Image + +app = Flask(__name__) +models = {} + +DETECTION_URL = "/v1/object-detection/" + + +@app.route(DETECTION_URL, methods=["POST"]) +def predict(model): + if request.method != "POST": + return + + if request.files.get("image"): + # Method 1 + # with request.files["image"] as f: + # im = Image.open(io.BytesIO(f.read())) + + # Method 2 + im_file = request.files["image"] + im_bytes = im_file.read() + im = Image.open(io.BytesIO(im_bytes)) + + if model in models: + results = models[model](im, size=640) # reduce size=320 for faster inference + return results.pandas().xyxy[0].to_json(orient="records") + + +if __name__ == "__main__": + parser = argparse.ArgumentParser(description="Flask API exposing YOLOv5 model") + parser.add_argument("--port", default=5000, type=int, help="port number") + parser.add_argument('--model', nargs='+', default=['yolov5s'], help='model(s) to run, i.e. --model yolov5n yolov5s') + opt = parser.parse_args() + + for m in opt.model: + models[m] = torch.hub.load("ultralytics/yolov5", m, force_reload=True, skip_validation=True) + + app.run(host="0.0.0.0", port=opt.port) # debug=True causes Restarting with stat diff --git a/application/yolov5_example/utils/general.py b/application/yolov5_example/utils/general.py new file mode 100644 index 00000000..a9463ddf --- /dev/null +++ b/application/yolov5_example/utils/general.py @@ -0,0 +1,1051 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +General utils +""" + +import contextlib +import glob +import inspect +import logging +import math +import os +import platform +import random +import re +import shutil +import signal +import sys +import threading +import time +import urllib +from datetime import datetime +from itertools import repeat +from multiprocessing.pool import ThreadPool +from pathlib import Path +from subprocess import check_output +from typing import Optional +from zipfile import ZipFile + +import cv2 +import numpy as np +import pandas as pd +import pkg_resources as pkg +import torch +import torchvision +import yaml + +from utils.downloads import gsutil_getsize +from utils.metrics import box_iou, fitness + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[1] # YOLOv5 root directory +RANK = int(os.getenv('RANK', -1)) + +# Settings +DATASETS_DIR = ROOT.parent / 'datasets' # YOLOv5 datasets directory +NUM_THREADS = min(8, max(1, os.cpu_count() - 1)) # number of YOLOv5 multiprocessing threads +AUTOINSTALL = str(os.getenv('YOLOv5_AUTOINSTALL', True)).lower() == 'true' # global auto-install mode +VERBOSE = str(os.getenv('YOLOv5_VERBOSE', True)).lower() == 'true' # global verbose mode +FONT = 'Arial.ttf' # https://ultralytics.com/assets/Arial.ttf + +torch.set_printoptions(linewidth=320, precision=5, profile='long') +np.set_printoptions(linewidth=320, formatter={'float_kind': '{:11.5g}'.format}) # format short g, %precision=5 +pd.options.display.max_columns = 10 +cv2.setNumThreads(0) # prevent OpenCV from multithreading (incompatible with PyTorch DataLoader) +os.environ['NUMEXPR_MAX_THREADS'] = str(NUM_THREADS) # NumExpr max threads +os.environ['OMP_NUM_THREADS'] = '1' if platform.system() == 'darwin' else str(NUM_THREADS) # OpenMP (PyTorch and SciPy) + + +def is_ascii(s=''): + # Is string composed of all ASCII (no UTF) characters? (note str().isascii() introduced in python 3.7) + s = str(s) # convert list, tuple, None, etc. to str + return len(s.encode().decode('ascii', 'ignore')) == len(s) + + +def is_chinese(s='人工智能'): + # Is string composed of any Chinese characters? + return bool(re.search('[\u4e00-\u9fff]', str(s))) + + +def is_colab(): + # Is environment a Google Colab instance? + return 'COLAB_GPU' in os.environ + + +def is_kaggle(): + # Is environment a Kaggle Notebook? + return os.environ.get('PWD') == '/kaggle/working' and os.environ.get('KAGGLE_URL_BASE') == 'https://www.kaggle.com' + + +def is_docker() -> bool: + """Check if the process runs inside a docker container.""" + if Path("/.dockerenv").exists(): + return True + try: # check if docker is in control groups + with open("/proc/self/cgroup") as file: + return any("docker" in line for line in file) + except OSError: + return False + + +def is_writeable(dir, test=False): + # Return True if directory has write permissions, test opening a file with write permissions if test=True + if not test: + return os.access(dir, os.W_OK) # possible issues on Windows + file = Path(dir) / 'tmp.txt' + try: + with open(file, 'w'): # open file with write permissions + pass + file.unlink() # remove file + return True + except OSError: + return False + + +def set_logging(name=None, verbose=VERBOSE): + # Sets level and returns logger + if is_kaggle() or is_colab(): + for h in logging.root.handlers: + logging.root.removeHandler(h) # remove all handlers associated with the root logger object + rank = int(os.getenv('RANK', -1)) # rank in world for Multi-GPU trainings + level = logging.INFO if verbose and rank in {-1, 0} else logging.ERROR + log = logging.getLogger(name) + log.setLevel(level) + handler = logging.StreamHandler() + handler.setFormatter(logging.Formatter("%(message)s")) + handler.setLevel(level) + log.addHandler(handler) + + +set_logging() # run before defining LOGGER +LOGGER = logging.getLogger("yolov5") # define globally (used in train.py, val.py, detect.py, etc.) +if platform.system() == 'Windows': + for fn in LOGGER.info, LOGGER.warning: + setattr(LOGGER, fn.__name__, lambda x: fn(emojis(x))) # emoji safe logging + + +def user_config_dir(dir='Ultralytics', env_var='YOLOV5_CONFIG_DIR'): + # Return path of user configuration directory. Prefer environment variable if exists. Make dir if required. + env = os.getenv(env_var) + if env: + path = Path(env) # use environment variable + else: + cfg = {'Windows': 'AppData/Roaming', 'Linux': '.config', 'Darwin': 'Library/Application Support'} # 3 OS dirs + path = Path.home() / cfg.get(platform.system(), '') # OS-specific config dir + path = (path if is_writeable(path) else Path('/tmp')) / dir # GCP and AWS lambda fix, only /tmp is writeable + path.mkdir(exist_ok=True) # make if required + return path + + +CONFIG_DIR = user_config_dir() # Ultralytics settings dir + + +class Profile(contextlib.ContextDecorator): + # Usage: @Profile() decorator or 'with Profile():' context manager + def __enter__(self): + self.start = time.time() + + def __exit__(self, type, value, traceback): + print(f'Profile results: {time.time() - self.start:.5f}s') + + +class Timeout(contextlib.ContextDecorator): + # Usage: @Timeout(seconds) decorator or 'with Timeout(seconds):' context manager + def __init__(self, seconds, *, timeout_msg='', suppress_timeout_errors=True): + self.seconds = int(seconds) + self.timeout_message = timeout_msg + self.suppress = bool(suppress_timeout_errors) + + def _timeout_handler(self, signum, frame): + raise TimeoutError(self.timeout_message) + + def __enter__(self): + if platform.system() != 'Windows': # not supported on Windows + signal.signal(signal.SIGALRM, self._timeout_handler) # Set handler for SIGALRM + signal.alarm(self.seconds) # start countdown for SIGALRM to be raised + + def __exit__(self, exc_type, exc_val, exc_tb): + if platform.system() != 'Windows': + signal.alarm(0) # Cancel SIGALRM if it's scheduled + if self.suppress and exc_type is TimeoutError: # Suppress TimeoutError + return True + + +class WorkingDirectory(contextlib.ContextDecorator): + # Usage: @WorkingDirectory(dir) decorator or 'with WorkingDirectory(dir):' context manager + def __init__(self, new_dir): + self.dir = new_dir # new dir + self.cwd = Path.cwd().resolve() # current dir + + def __enter__(self): + os.chdir(self.dir) + + def __exit__(self, exc_type, exc_val, exc_tb): + os.chdir(self.cwd) + + +def try_except(func): + # try-except function. Usage: @try_except decorator + def handler(*args, **kwargs): + try: + func(*args, **kwargs) + except Exception as e: + print(e) + + return handler + + +def threaded(func): + # Multi-threads a target function and returns thread. Usage: @threaded decorator + def wrapper(*args, **kwargs): + thread = threading.Thread(target=func, args=args, kwargs=kwargs, daemon=True) + thread.start() + return thread + + return wrapper + + +def methods(instance): + # Get class/instance methods + return [f for f in dir(instance) if callable(getattr(instance, f)) and not f.startswith("__")] + + +def print_args(args: Optional[dict] = None, show_file=True, show_fcn=False): + # Print function arguments (optional args dict) + x = inspect.currentframe().f_back # previous frame + file, _, fcn, _, _ = inspect.getframeinfo(x) + if args is None: # get args automatically + args, _, _, frm = inspect.getargvalues(x) + args = {k: v for k, v in frm.items() if k in args} + try: + file = Path(file).resolve().relative_to(ROOT).with_suffix('') + except ValueError: + file = Path(file).stem + s = (f'{file}: ' if show_file else '') + (f'{fcn}: ' if show_fcn else '') + LOGGER.info(colorstr(s) + ', '.join(f'{k}={v}' for k, v in args.items())) + + +def init_seeds(seed=0, deterministic=False): + # Initialize random number generator (RNG) seeds https://pytorch.org/docs/stable/notes/randomness.html + # cudnn seed 0 settings are slower and more reproducible, else faster and less reproducible + import torch.backends.cudnn as cudnn + + if deterministic and check_version(torch.__version__, '1.12.0'): # https://github.com/ultralytics/yolov5/pull/8213 + torch.use_deterministic_algorithms(True) + os.environ['CUBLAS_WORKSPACE_CONFIG'] = ':4096:8' + os.environ['PYTHONHASHSEED'] = str(seed) + + random.seed(seed) + np.random.seed(seed) + torch.manual_seed(seed) + cudnn.benchmark, cudnn.deterministic = (False, True) if seed == 0 else (True, False) + torch.cuda.manual_seed(seed) + torch.cuda.manual_seed_all(seed) # for Multi-GPU, exception safe + + +def intersect_dicts(da, db, exclude=()): + # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values + return {k: v for k, v in da.items() if k in db and not any(x in k for x in exclude) and v.shape == db[k].shape} + + +def get_latest_run(search_dir='.'): + # Return path to most recent 'last.pt' in /runs (i.e. to --resume from) + last_list = glob.glob(f'{search_dir}/**/last*.pt', recursive=True) + return max(last_list, key=os.path.getctime) if last_list else '' + + +def emojis(str=''): + # Return platform-dependent emoji-safe version of string + return str.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else str + + +def file_age(path=__file__): + # Return days since last file update + dt = (datetime.now() - datetime.fromtimestamp(Path(path).stat().st_mtime)) # delta + return dt.days # + dt.seconds / 86400 # fractional days + + +def file_date(path=__file__): + # Return human-readable file modification date, i.e. '2021-3-26' + t = datetime.fromtimestamp(Path(path).stat().st_mtime) + return f'{t.year}-{t.month}-{t.day}' + + +def file_size(path): + # Return file/dir size (MB) + mb = 1 << 20 # bytes to MiB (1024 ** 2) + path = Path(path) + if path.is_file(): + return path.stat().st_size / mb + elif path.is_dir(): + return sum(f.stat().st_size for f in path.glob('**/*') if f.is_file()) / mb + else: + return 0.0 + + +def check_online(): + # Check internet connectivity + import socket + try: + socket.create_connection(("1.1.1.1", 443), 5) # check host accessibility + return True + except OSError: + return False + + +def git_describe(path=ROOT): # path must be a directory + # Return human-readable git description, i.e. v5.0-5-g3e25f1e https://git-scm.com/docs/git-describe + try: + assert (Path(path) / '.git').is_dir() + return check_output(f'git -C {path} describe --tags --long --always', shell=True).decode()[:-1] + except Exception: + return '' + + +@try_except +@WorkingDirectory(ROOT) +def check_git_status(repo='ultralytics/yolov5'): + # YOLOv5 status check, recommend 'git pull' if code is out of date + url = f'https://github.com/{repo}' + msg = f', for updates see {url}' + s = colorstr('github: ') # string + assert Path('.git').exists(), s + 'skipping check (not a git repository)' + msg + assert check_online(), s + 'skipping check (offline)' + msg + + splits = re.split(pattern=r'\s', string=check_output('git remote -v', shell=True).decode()) + matches = [repo in s for s in splits] + if any(matches): + remote = splits[matches.index(True) - 1] + else: + remote = 'ultralytics' + check_output(f'git remote add {remote} {url}', shell=True) + check_output(f'git fetch {remote}', shell=True, timeout=5) # git fetch + branch = check_output('git rev-parse --abbrev-ref HEAD', shell=True).decode().strip() # checked out + n = int(check_output(f'git rev-list {branch}..{remote}/master --count', shell=True)) # commits behind + if n > 0: + pull = 'git pull' if remote == 'origin' else f'git pull {remote} master' + s += f"⚠️ YOLOv5 is out of date by {n} commit{'s' * (n > 1)}. Use `{pull}` or `git clone {url}` to update." + else: + s += f'up to date with {url} ✅' + LOGGER.info(s) + + +def check_python(minimum='3.7.0'): + # Check current python version vs. required python version + check_version(platform.python_version(), minimum, name='Python ', hard=True) + + +def check_version(current='0.0.0', minimum='0.0.0', name='version ', pinned=False, hard=False, verbose=False): + # Check version vs. required version + current, minimum = (pkg.parse_version(x) for x in (current, minimum)) + result = (current == minimum) if pinned else (current >= minimum) # bool + s = f'{name}{minimum} required by YOLOv5, but {name}{current} is currently installed' # string + if hard: + assert result, s # assert min requirements met + if verbose and not result: + LOGGER.warning(s) + return result + + +@try_except +def check_requirements(requirements=ROOT / 'requirements.txt', exclude=(), install=True, cmds=()): + # Check installed dependencies meet YOLOv5 requirements (pass *.txt file or list of packages) + prefix = colorstr('red', 'bold', 'requirements:') + check_python() # check python version + if isinstance(requirements, (str, Path)): # requirements.txt file + file = Path(requirements) + assert file.exists(), f"{prefix} {file.resolve()} not found, check failed." + with file.open() as f: + requirements = [f'{x.name}{x.specifier}' for x in pkg.parse_requirements(f) if x.name not in exclude] + else: # list or tuple of packages + requirements = [x for x in requirements if x not in exclude] + + n = 0 # number of packages updates + for i, r in enumerate(requirements): + try: + pkg.require(r) + except Exception: # DistributionNotFound or VersionConflict if requirements not met + s = f"{prefix} {r} not found and is required by YOLOv5" + if install and AUTOINSTALL: # check environment variable + LOGGER.info(f"{s}, attempting auto-update...") + try: + assert check_online(), f"'pip install {r}' skipped (offline)" + LOGGER.info(check_output(f'pip install "{r}" {cmds[i] if cmds else ""}', shell=True).decode()) + n += 1 + except Exception as e: + LOGGER.warning(f'{prefix} {e}') + else: + LOGGER.info(f'{s}. Please install and rerun your command.') + + if n: # if packages updated + source = file.resolve() if 'file' in locals() else requirements + s = f"{prefix} {n} package{'s' * (n > 1)} updated per {source}\n" \ + f"{prefix} ⚠️ {colorstr('bold', 'Restart runtime or rerun command for updates to take effect')}\n" + LOGGER.info(s) + + +def check_img_size(imgsz, s=32, floor=0): + # Verify image size is a multiple of stride s in each dimension + if isinstance(imgsz, int): # integer i.e. img_size=640 + new_size = max(make_divisible(imgsz, int(s)), floor) + else: # list i.e. img_size=[640, 480] + imgsz = list(imgsz) # convert to list if tuple + new_size = [max(make_divisible(x, int(s)), floor) for x in imgsz] + if new_size != imgsz: + LOGGER.warning(f'WARNING: --img-size {imgsz} must be multiple of max stride {s}, updating to {new_size}') + return new_size + + +def check_imshow(): + # Check if environment supports image displays + try: + assert not is_docker(), 'cv2.imshow() is disabled in Docker environments' + assert not is_colab(), 'cv2.imshow() is disabled in Google Colab environments' + cv2.imshow('test', np.zeros((1, 1, 3))) + cv2.waitKey(1) + cv2.destroyAllWindows() + cv2.waitKey(1) + return True + except Exception as e: + LOGGER.warning(f'WARNING: Environment does not support cv2.imshow() or PIL Image.show() image displays\n{e}') + return False + + +def check_suffix(file='yolov5s.pt', suffix=('.pt',), msg=''): + # Check file(s) for acceptable suffix + if file and suffix: + if isinstance(suffix, str): + suffix = [suffix] + for f in file if isinstance(file, (list, tuple)) else [file]: + s = Path(f).suffix.lower() # file suffix + if len(s): + assert s in suffix, f"{msg}{f} acceptable suffix is {suffix}" + + +def check_yaml(file, suffix=('.yaml', '.yml')): + # Search/download YAML file (if necessary) and return path, checking suffix + return check_file(file, suffix) + + +def check_file(file, suffix=''): + # Search/download file (if necessary) and return path + check_suffix(file, suffix) # optional + file = str(file) # convert to str() + if Path(file).is_file() or not file: # exists + return file + elif file.startswith(('http:/', 'https:/')): # download + url = file # warning: Pathlib turns :// -> :/ + file = Path(urllib.parse.unquote(file).split('?')[0]).name # '%2F' to '/', split https://url.com/file.txt?auth + if Path(file).is_file(): + LOGGER.info(f'Found {url} locally at {file}') # file already exists + else: + LOGGER.info(f'Downloading {url} to {file}...') + torch.hub.download_url_to_file(url, file) + assert Path(file).exists() and Path(file).stat().st_size > 0, f'File download failed: {url}' # check + return file + elif file.startswith('clearml://'): # ClearML Dataset ID + assert 'clearml' in sys.modules, "ClearML is not installed, so cannot use ClearML dataset. Try running 'pip install clearml'." + return file + else: # search + files = [] + for d in 'data', 'models', 'utils': # search directories + files.extend(glob.glob(str(ROOT / d / '**' / file), recursive=True)) # find file + assert len(files), f'File not found: {file}' # assert file was found + assert len(files) == 1, f"Multiple files match '{file}', specify exact path: {files}" # assert unique + return files[0] # return file + + +def check_font(font=FONT, progress=False): + # Download font to CONFIG_DIR if necessary + font = Path(font) + file = CONFIG_DIR / font.name + if not font.exists() and not file.exists(): + url = "https://ultralytics.com/assets/" + font.name + LOGGER.info(f'Downloading {url} to {file}...') + torch.hub.download_url_to_file(url, str(file), progress=progress) + + +def check_dataset(data, autodownload=True): + # Download, check and/or unzip dataset if not found locally + + # Download (optional) + extract_dir = '' + if isinstance(data, (str, Path)) and str(data).endswith('.zip'): # i.e. gs://bucket/dir/coco128.zip + download(data, dir=f'{DATASETS_DIR}/{Path(data).stem}', unzip=True, delete=False, curl=False, threads=1) + data = next((DATASETS_DIR / Path(data).stem).rglob('*.yaml')) + extract_dir, autodownload = data.parent, False + + # Read yaml (optional) + if isinstance(data, (str, Path)): + with open(data, errors='ignore') as f: + data = yaml.safe_load(f) # dictionary + + # Checks + for k in 'train', 'val', 'nc': + assert k in data, f"data.yaml '{k}:' field missing ❌" + if 'names' not in data: + LOGGER.warning("data.yaml 'names:' field missing ⚠️, assigning default names 'class0', 'class1', etc.") + data['names'] = [f'class{i}' for i in range(data['nc'])] # default names + + # Resolve paths + path = Path(extract_dir or data.get('path') or '') # optional 'path' default to '.' + if not path.is_absolute(): + path = (ROOT / path).resolve() + data['path'] = str(path) + for k in 'train', 'val', 'test': + if data.get(k): # prepend path + data[k] = str(path / data[k]) if isinstance(data[k], str) else [str(path / x) for x in data[k]] + + # Parse yaml + train, val, test, s = (data.get(x) for x in ('train', 'val', 'test', 'download')) + if val: + val = [Path(x).resolve() for x in (val if isinstance(val, list) else [val])] # val path + if not all(x.exists() for x in val): + LOGGER.info('\nDataset not found ⚠️, missing paths %s' % [str(x) for x in val if not x.exists()]) + if not s or not autodownload: + raise Exception('Dataset not found ❌') + t = time.time() + root = path.parent if 'path' in data else '..' # unzip directory i.e. '../' + if s.startswith('http') and s.endswith('.zip'): # URL + f = Path(s).name # filename + LOGGER.info(f'Downloading {s} to {f}...') + torch.hub.download_url_to_file(s, f) + Path(root).mkdir(parents=True, exist_ok=True) # create root + ZipFile(f).extractall(path=root) # unzip + Path(f).unlink() # remove zip + r = None # success + elif s.startswith('bash '): # bash script + LOGGER.info(f'Running {s} ...') + r = os.system(s) + else: # python script + r = exec(s, {'yaml': data}) # return None + dt = f'({round(time.time() - t, 1)}s)' + s = f"success ✅ {dt}, saved to {colorstr('bold', root)}" if r in (0, None) else f"failure {dt} ❌" + LOGGER.info(f"Dataset download {s}") + check_font('Arial.ttf' if is_ascii(data['names']) else 'Arial.Unicode.ttf', progress=True) # download fonts + return data # dictionary + + +def check_amp(model): + # Check PyTorch Automatic Mixed Precision (AMP) functionality. Return True on correct operation + from models.common import AutoShape, DetectMultiBackend + + def amp_allclose(model, im): + # All close FP32 vs AMP results + m = AutoShape(model, verbose=False) # model + a = m(im).xywhn[0] # FP32 inference + m.amp = True + b = m(im).xywhn[0] # AMP inference + return a.shape == b.shape and torch.allclose(a, b, atol=0.1) # close to 10% absolute tolerance + + prefix = colorstr('AMP: ') + device = next(model.parameters()).device # get model device + if device.type == 'cpu': + return False # AMP disabled on CPU + f = ROOT / 'data' / 'images' / 'bus.jpg' # image to check + im = f if f.exists() else 'https://ultralytics.com/images/bus.jpg' if check_online() else np.ones((640, 640, 3)) + try: + assert amp_allclose(model, im) or amp_allclose(DetectMultiBackend('yolov5n.pt', device), im) + LOGGER.info(f'{prefix}checks passed ✅') + return True + except Exception: + help_url = 'https://github.com/ultralytics/yolov5/issues/7908' + LOGGER.warning(f'{prefix}checks failed ❌, disabling Automatic Mixed Precision. See {help_url}') + return False + + +def yaml_load(file='data.yaml'): + # Single-line safe yaml loading + with open(file, errors='ignore') as f: + return yaml.safe_load(f) + + +def yaml_save(file='data.yaml', data={}): + # Single-line safe yaml saving + with open(file, 'w') as f: + yaml.safe_dump({k: str(v) if isinstance(v, Path) else v for k, v in data.items()}, f, sort_keys=False) + + +def url2file(url): + # Convert URL to filename, i.e. https://url.com/file.txt?auth -> file.txt + url = str(Path(url)).replace(':/', '://') # Pathlib turns :// -> :/ + return Path(urllib.parse.unquote(url)).name.split('?')[0] # '%2F' to '/', split https://url.com/file.txt?auth + + +def download(url, dir='.', unzip=True, delete=True, curl=False, threads=1, retry=3): + # Multi-threaded file download and unzip function, used in data.yaml for autodownload + def download_one(url, dir): + # Download 1 file + success = True + f = dir / Path(url).name # filename + if Path(url).is_file(): # exists in current path + Path(url).rename(f) # move to dir + elif not f.exists(): + LOGGER.info(f'Downloading {url} to {f}...') + for i in range(retry + 1): + if curl: + s = 'sS' if threads > 1 else '' # silent + r = os.system(f'curl -{s}L "{url}" -o "{f}" --retry 9 -C -') # curl download with retry, continue + success = r == 0 + else: + torch.hub.download_url_to_file(url, f, progress=threads == 1) # torch download + success = f.is_file() + if success: + break + elif i < retry: + LOGGER.warning(f'Download failure, retrying {i + 1}/{retry} {url}...') + else: + LOGGER.warning(f'Failed to download {url}...') + + if unzip and success and f.suffix in ('.zip', '.tar', '.gz'): + LOGGER.info(f'Unzipping {f}...') + if f.suffix == '.zip': + ZipFile(f).extractall(path=dir) # unzip + elif f.suffix == '.tar': + os.system(f'tar xf {f} --directory {f.parent}') # unzip + elif f.suffix == '.gz': + os.system(f'tar xfz {f} --directory {f.parent}') # unzip + if delete: + f.unlink() # remove zip + + dir = Path(dir) + dir.mkdir(parents=True, exist_ok=True) # make directory + if threads > 1: + pool = ThreadPool(threads) + pool.imap(lambda x: download_one(*x), zip(url, repeat(dir))) # multi-threaded + pool.close() + pool.join() + else: + for u in [url] if isinstance(url, (str, Path)) else url: + download_one(u, dir) + + +def make_divisible(x, divisor): + # Returns nearest x divisible by divisor + if isinstance(divisor, torch.Tensor): + divisor = int(divisor.max()) # to int + return math.ceil(x / divisor) * divisor + + +def clean_str(s): + # Cleans a string by replacing special characters with underscore _ + return re.sub(pattern="[|@#!¡·$€%&()=?¿^*;:,¨´><+]", repl="_", string=s) + + +def one_cycle(y1=0.0, y2=1.0, steps=100): + # lambda function for sinusoidal ramp from y1 to y2 https://arxiv.org/pdf/1812.01187.pdf + return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1 + + +def colorstr(*input): + # Colors a string https://en.wikipedia.org/wiki/ANSI_escape_code, i.e. colorstr('blue', 'hello world') + *args, string = input if len(input) > 1 else ('blue', 'bold', input[0]) # color arguments, string + colors = { + 'black': '\033[30m', # basic colors + 'red': '\033[31m', + 'green': '\033[32m', + 'yellow': '\033[33m', + 'blue': '\033[34m', + 'magenta': '\033[35m', + 'cyan': '\033[36m', + 'white': '\033[37m', + 'bright_black': '\033[90m', # bright colors + 'bright_red': '\033[91m', + 'bright_green': '\033[92m', + 'bright_yellow': '\033[93m', + 'bright_blue': '\033[94m', + 'bright_magenta': '\033[95m', + 'bright_cyan': '\033[96m', + 'bright_white': '\033[97m', + 'end': '\033[0m', # misc + 'bold': '\033[1m', + 'underline': '\033[4m'} + return ''.join(colors[x] for x in args) + f'{string}' + colors['end'] + + +def labels_to_class_weights(labels, nc=80): + # Get class weights (inverse frequency) from training labels + if labels[0] is None: # no labels loaded + return torch.Tensor() + + labels = np.concatenate(labels, 0) # labels.shape = (866643, 5) for COCO + classes = labels[:, 0].astype(int) # labels = [class xywh] + weights = np.bincount(classes, minlength=nc) # occurrences per class + + # Prepend gridpoint count (for uCE training) + # gpi = ((320 / 32 * np.array([1, 2, 4])) ** 2 * 3).sum() # gridpoints per image + # weights = np.hstack([gpi * len(labels) - weights.sum() * 9, weights * 9]) ** 0.5 # prepend gridpoints to start + + weights[weights == 0] = 1 # replace empty bins with 1 + weights = 1 / weights # number of targets per class + weights /= weights.sum() # normalize + return torch.from_numpy(weights).float() + + +def labels_to_image_weights(labels, nc=80, class_weights=np.ones(80)): + # Produces image weights based on class_weights and image contents + # Usage: index = random.choices(range(n), weights=image_weights, k=1) # weighted image sample + class_counts = np.array([np.bincount(x[:, 0].astype(int), minlength=nc) for x in labels]) + return (class_weights.reshape(1, nc) * class_counts).sum(1) + + +def coco80_to_coco91_class(): # converts 80-index (val2014) to 91-index (paper) + # https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/ + # a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n') + # b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n') + # x1 = [list(a[i] == b).index(True) + 1 for i in range(80)] # darknet to coco + # x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)] # coco to darknet + return [ + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, + 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, + 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90] + + +def xyxy2xywh(x): + # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] where xy1=top-left, xy2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = (x[:, 0] + x[:, 2]) / 2 # x center + y[:, 1] = (x[:, 1] + x[:, 3]) / 2 # y center + y[:, 2] = x[:, 2] - x[:, 0] # width + y[:, 3] = x[:, 3] - x[:, 1] # height + return y + + +def xywh2xyxy(x): + # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = x[:, 0] - x[:, 2] / 2 # top left x + y[:, 1] = x[:, 1] - x[:, 3] / 2 # top left y + y[:, 2] = x[:, 0] + x[:, 2] / 2 # bottom right x + y[:, 3] = x[:, 1] + x[:, 3] / 2 # bottom right y + return y + + +def xywhn2xyxy(x, w=640, h=640, padw=0, padh=0): + # Convert nx4 boxes from [x, y, w, h] normalized to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = w * (x[:, 0] - x[:, 2] / 2) + padw # top left x + y[:, 1] = h * (x[:, 1] - x[:, 3] / 2) + padh # top left y + y[:, 2] = w * (x[:, 0] + x[:, 2] / 2) + padw # bottom right x + y[:, 3] = h * (x[:, 1] + x[:, 3] / 2) + padh # bottom right y + return y + + +def xyxy2xywhn(x, w=640, h=640, clip=False, eps=0.0): + # Convert nx4 boxes from [x1, y1, x2, y2] to [x, y, w, h] normalized where xy1=top-left, xy2=bottom-right + if clip: + clip_coords(x, (h - eps, w - eps)) # warning: inplace clip + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = ((x[:, 0] + x[:, 2]) / 2) / w # x center + y[:, 1] = ((x[:, 1] + x[:, 3]) / 2) / h # y center + y[:, 2] = (x[:, 2] - x[:, 0]) / w # width + y[:, 3] = (x[:, 3] - x[:, 1]) / h # height + return y + + +def xyn2xy(x, w=640, h=640, padw=0, padh=0): + # Convert normalized segments into pixel segments, shape (n,2) + y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x) + y[:, 0] = w * x[:, 0] + padw # top left x + y[:, 1] = h * x[:, 1] + padh # top left y + return y + + +def segment2box(segment, width=640, height=640): + # Convert 1 segment label to 1 box label, applying inside-image constraint, i.e. (xy1, xy2, ...) to (xyxy) + x, y = segment.T # segment xy + inside = (x >= 0) & (y >= 0) & (x <= width) & (y <= height) + x, y, = x[inside], y[inside] + return np.array([x.min(), y.min(), x.max(), y.max()]) if any(x) else np.zeros((1, 4)) # xyxy + + +def segments2boxes(segments): + # Convert segment labels to box labels, i.e. (cls, xy1, xy2, ...) to (cls, xywh) + boxes = [] + for s in segments: + x, y = s.T # segment xy + boxes.append([x.min(), y.min(), x.max(), y.max()]) # cls, xyxy + return xyxy2xywh(np.array(boxes)) # cls, xywh + + +def resample_segments(segments, n=1000): + # Up-sample an (n,2) segment + for i, s in enumerate(segments): + s = np.concatenate((s, s[0:1, :]), axis=0) + x = np.linspace(0, len(s) - 1, n) + xp = np.arange(len(s)) + segments[i] = np.concatenate([np.interp(x, xp, s[:, i]) for i in range(2)]).reshape(2, -1).T # segment xy + return segments + + +def scale_coords(img1_shape, coords, img0_shape, ratio_pad=None): + # Rescale coords (xyxy) from img1_shape to img0_shape + if ratio_pad is None: # calculate from img0_shape + gain = min(img1_shape[0] / img0_shape[0], img1_shape[1] / img0_shape[1]) # gain = old / new + pad = (img1_shape[1] - img0_shape[1] * gain) / 2, (img1_shape[0] - img0_shape[0] * gain) / 2 # wh padding + else: + gain = ratio_pad[0][0] + pad = ratio_pad[1] + + coords[:, [0, 2]] -= pad[0] # x padding + coords[:, [1, 3]] -= pad[1] # y padding + coords[:, :4] /= gain + clip_coords(coords, img0_shape) + return coords + + +def clip_coords(boxes, shape): + # Clip bounding xyxy bounding boxes to image shape (height, width) + if isinstance(boxes, torch.Tensor): # faster individually + boxes[:, 0].clamp_(0, shape[1]) # x1 + boxes[:, 1].clamp_(0, shape[0]) # y1 + boxes[:, 2].clamp_(0, shape[1]) # x2 + boxes[:, 3].clamp_(0, shape[0]) # y2 + else: # np.array (faster grouped) + boxes[:, [0, 2]] = boxes[:, [0, 2]].clip(0, shape[1]) # x1, x2 + boxes[:, [1, 3]] = boxes[:, [1, 3]].clip(0, shape[0]) # y1, y2 + + +def non_max_suppression(prediction, + conf_thres=0.25, + iou_thres=0.45, + classes=None, + agnostic=False, + multi_label=False, + labels=(), + max_det=300): + """Non-Maximum Suppression (NMS) on inference results to reject overlapping bounding boxes + + Returns: + list of detections, on (n,6) tensor per image [xyxy, conf, cls] + """ + + bs = prediction.shape[0] # batch size + nc = prediction.shape[2] - 5 # number of classes + xc = prediction[..., 4] > conf_thres # candidates + + # Checks + assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0' + assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0' + + # Settings + # min_wh = 2 # (pixels) minimum box width and height + max_wh = 7680 # (pixels) maximum box width and height + max_nms = 30000 # maximum number of boxes into torchvision.ops.nms() + time_limit = 0.3 + 0.03 * bs # seconds to quit after + redundant = True # require redundant detections + multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img) + merge = False # use merge-NMS + + t = time.time() + output = [torch.zeros((0, 6), device=prediction.device)] * bs + for xi, x in enumerate(prediction): # image index, image inference + # Apply constraints + # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height + x = x[xc[xi]] # confidence + + # Cat apriori labels if autolabelling + if labels and len(labels[xi]): + lb = labels[xi] + v = torch.zeros((len(lb), nc + 5), device=x.device) + v[:, :4] = lb[:, 1:5] # box + v[:, 4] = 1.0 # conf + v[range(len(lb)), lb[:, 0].long() + 5] = 1.0 # cls + x = torch.cat((x, v), 0) + + # If none remain process next image + if not x.shape[0]: + continue + + # Compute conf + x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf + + # Box (center x, center y, width, height) to (x1, y1, x2, y2) + box = xywh2xyxy(x[:, :4]) + + # Detections matrix nx6 (xyxy, conf, cls) + if multi_label: + i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T + x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1) + else: # best class only + conf, j = x[:, 5:].max(1, keepdim=True) + x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres] + + # Filter by class + if classes is not None: + x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)] + + # Apply finite constraint + # if not torch.isfinite(x).all(): + # x = x[torch.isfinite(x).all(1)] + + # Check shape + n = x.shape[0] # number of boxes + if not n: # no boxes + continue + elif n > max_nms: # excess boxes + x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence + + # Batched NMS + c = x[:, 5:6] * (0 if agnostic else max_wh) # classes + boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores + i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS + if i.shape[0] > max_det: # limit detections + i = i[:max_det] + if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean) + # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4) + iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix + weights = iou * scores[None] # box weights + x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes + if redundant: + i = i[iou.sum(1) > 1] # require redundancy + + output[xi] = x[i] + if (time.time() - t) > time_limit: + LOGGER.warning(f'WARNING: NMS time limit {time_limit:.3f}s exceeded') + break # time limit exceeded + + return output + + +def strip_optimizer(f='best.pt', s=''): # from utils.general import *; strip_optimizer() + # Strip optimizer from 'f' to finalize training, optionally save as 's' + x = torch.load(f, map_location=torch.device('cpu')) + if x.get('ema'): + x['model'] = x['ema'] # replace model with ema + for k in 'optimizer', 'best_fitness', 'wandb_id', 'ema', 'updates': # keys + x[k] = None + x['epoch'] = -1 + x['model'].half() # to FP16 + for p in x['model'].parameters(): + p.requires_grad = False + torch.save(x, s or f) + mb = os.path.getsize(s or f) / 1E6 # filesize + LOGGER.info(f"Optimizer stripped from {f},{f' saved as {s},' if s else ''} {mb:.1f}MB") + + +def print_mutation(results, hyp, save_dir, bucket, prefix=colorstr('evolve: ')): + evolve_csv = save_dir / 'evolve.csv' + evolve_yaml = save_dir / 'hyp_evolve.yaml' + keys = ('metrics/precision', 'metrics/recall', 'metrics/mAP_0.5', 'metrics/mAP_0.5:0.95', 'val/box_loss', + 'val/obj_loss', 'val/cls_loss') + tuple(hyp.keys()) # [results + hyps] + keys = tuple(x.strip() for x in keys) + vals = results + tuple(hyp.values()) + n = len(keys) + + # Download (optional) + if bucket: + url = f'gs://{bucket}/evolve.csv' + if gsutil_getsize(url) > (evolve_csv.stat().st_size if evolve_csv.exists() else 0): + os.system(f'gsutil cp {url} {save_dir}') # download evolve.csv if larger than local + + # Log to evolve.csv + s = '' if evolve_csv.exists() else (('%20s,' * n % keys).rstrip(',') + '\n') # add header + with open(evolve_csv, 'a') as f: + f.write(s + ('%20.5g,' * n % vals).rstrip(',') + '\n') + + # Save yaml + with open(evolve_yaml, 'w') as f: + data = pd.read_csv(evolve_csv) + data = data.rename(columns=lambda x: x.strip()) # strip keys + i = np.argmax(fitness(data.values[:, :4])) # + generations = len(data) + f.write('# YOLOv5 Hyperparameter Evolution Results\n' + f'# Best generation: {i}\n' + + f'# Last generation: {generations - 1}\n' + '# ' + ', '.join(f'{x.strip():>20s}' for x in keys[:7]) + + '\n' + '# ' + ', '.join(f'{x:>20.5g}' for x in data.values[i, :7]) + '\n\n') + yaml.safe_dump(data.loc[i][7:].to_dict(), f, sort_keys=False) + + # Print to screen + LOGGER.info(prefix + f'{generations} generations finished, current result:\n' + prefix + + ', '.join(f'{x.strip():>20s}' for x in keys) + '\n' + prefix + ', '.join(f'{x:20.5g}' + for x in vals) + '\n\n') + + if bucket: + os.system(f'gsutil cp {evolve_csv} {evolve_yaml} gs://{bucket}') # upload + + +def apply_classifier(x, model, img, im0): + # Apply a second stage classifier to YOLO outputs + # Example model = torchvision.models.__dict__['efficientnet_b0'](pretrained=True).to(device).eval() + im0 = [im0] if isinstance(im0, np.ndarray) else im0 + for i, d in enumerate(x): # per image + if d is not None and len(d): + d = d.clone() + + # Reshape and pad cutouts + b = xyxy2xywh(d[:, :4]) # boxes + b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1) # rectangle to square + b[:, 2:] = b[:, 2:] * 1.3 + 30 # pad + d[:, :4] = xywh2xyxy(b).long() + + # Rescale boxes from img_size to im0 size + scale_coords(img.shape[2:], d[:, :4], im0[i].shape) + + # Classes + pred_cls1 = d[:, 5].long() + ims = [] + for a in d: + cutout = im0[i][int(a[1]):int(a[3]), int(a[0]):int(a[2])] + im = cv2.resize(cutout, (224, 224)) # BGR + + im = im[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 + im = np.ascontiguousarray(im, dtype=np.float32) # uint8 to float32 + im /= 255 # 0 - 255 to 0.0 - 1.0 + ims.append(im) + + pred_cls2 = model(torch.Tensor(ims).to(d.device)).argmax(1) # classifier prediction + x[i] = x[i][pred_cls1 == pred_cls2] # retain matching class detections + + return x + + +def increment_path(path, exist_ok=False, sep='', mkdir=False): + # Increment file or directory path, i.e. runs/exp --> runs/exp{sep}2, runs/exp{sep}3, ... etc. + path = Path(path) # os-agnostic + if path.exists() and not exist_ok: + path, suffix = (path.with_suffix(''), path.suffix) if path.is_file() else (path, '') + + # Method 1 + for n in range(2, 9999): + p = f'{path}{sep}{n}{suffix}' # increment path + if not os.path.exists(p): # + break + path = Path(p) + + # Method 2 (deprecated) + # dirs = glob.glob(f"{path}{sep}*") # similar paths + # matches = [re.search(rf"{path.stem}{sep}(\d+)", d) for d in dirs] + # i = [int(m.groups()[0]) for m in matches if m] # indices + # n = max(i) + 1 if i else 2 # increment number + # path = Path(f"{path}{sep}{n}{suffix}") # increment path + + if mkdir: + path.mkdir(parents=True, exist_ok=True) # make directory + + return path + + +# OpenCV Chinese-friendly functions ------------------------------------------------------------------------------------ +imshow_ = cv2.imshow # copy to avoid recursion errors + + +def imread(path, flags=cv2.IMREAD_COLOR): + return cv2.imdecode(np.fromfile(path, np.uint8), flags) + + +def imwrite(path, im): + try: + cv2.imencode(Path(path).suffix, im)[1].tofile(path) + return True + except Exception: + return False + + +def imshow(path, im): + imshow_(path.encode('unicode_escape').decode(), im) + + +cv2.imread, cv2.imwrite, cv2.imshow = imread, imwrite, imshow # redefine + +# Variables ------------------------------------------------------------------------------------------------------------ +NCOLS = 0 if is_docker() else shutil.get_terminal_size().columns # terminal window size for tqdm diff --git a/application/yolov5_example/utils/google_app_engine/Dockerfile b/application/yolov5_example/utils/google_app_engine/Dockerfile new file mode 100644 index 00000000..0155618f --- /dev/null +++ b/application/yolov5_example/utils/google_app_engine/Dockerfile @@ -0,0 +1,25 @@ +FROM gcr.io/google-appengine/python + +# Create a virtualenv for dependencies. This isolates these packages from +# system-level packages. +# Use -p python3 or -p python3.7 to select python version. Default is version 2. +RUN virtualenv /env -p python3 + +# Setting these environment variables are the same as running +# source /env/bin/activate. +ENV VIRTUAL_ENV /env +ENV PATH /env/bin:$PATH + +RUN apt-get update && apt-get install -y python-opencv + +# Copy the application's requirements.txt and run pip to install all +# dependencies into the virtualenv. +ADD requirements.txt /app/requirements.txt +RUN pip install -r /app/requirements.txt + +# Add the application source code. +ADD . /app + +# Run a WSGI server to serve the application. gunicorn must be declared as +# a dependency in requirements.txt. +CMD gunicorn -b :$PORT main:app diff --git a/application/yolov5_example/utils/google_app_engine/additional_requirements.txt b/application/yolov5_example/utils/google_app_engine/additional_requirements.txt new file mode 100644 index 00000000..42d7ffc0 --- /dev/null +++ b/application/yolov5_example/utils/google_app_engine/additional_requirements.txt @@ -0,0 +1,4 @@ +# add these requirements in your app on top of the existing ones +pip==21.1 +Flask==1.0.2 +gunicorn==19.9.0 diff --git a/application/yolov5_example/utils/google_app_engine/app.yaml b/application/yolov5_example/utils/google_app_engine/app.yaml new file mode 100644 index 00000000..5056b7c1 --- /dev/null +++ b/application/yolov5_example/utils/google_app_engine/app.yaml @@ -0,0 +1,14 @@ +runtime: custom +env: flex + +service: yolov5app + +liveness_check: + initial_delay_sec: 600 + +manual_scaling: + instances: 1 +resources: + cpu: 1 + memory_gb: 4 + disk_size_gb: 20 diff --git a/application/yolov5_example/utils/loggers/__init__.py b/application/yolov5_example/utils/loggers/__init__.py new file mode 100644 index 00000000..8ec846f8 --- /dev/null +++ b/application/yolov5_example/utils/loggers/__init__.py @@ -0,0 +1,308 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Logging utils +""" + +import os +import warnings +from pathlib import Path + +import pkg_resources as pkg +import torch +from torch.utils.tensorboard import SummaryWriter + +from utils.general import colorstr, cv2 +from utils.loggers.clearml.clearml_utils import ClearmlLogger +from utils.loggers.wandb.wandb_utils import WandbLogger +from utils.plots import plot_images, plot_results +from utils.torch_utils import de_parallel + +LOGGERS = ('csv', 'tb', 'wandb', 'clearml') # *.csv, TensorBoard, Weights & Biases, ClearML +RANK = int(os.getenv('RANK', -1)) + +try: + import wandb + + assert hasattr(wandb, '__version__') # verify package import not local dir + if pkg.parse_version(wandb.__version__) >= pkg.parse_version('0.12.2') and RANK in {0, -1}: + try: + wandb_login_success = wandb.login(timeout=30) + except wandb.errors.UsageError: # known non-TTY terminal issue + wandb_login_success = False + if not wandb_login_success: + wandb = None +except (ImportError, AssertionError): + wandb = None + +try: + import clearml + + assert hasattr(clearml, '__version__') # verify package import not local dir +except (ImportError, AssertionError): + clearml = None + + +class Loggers(): + # YOLOv5 Loggers class + def __init__(self, save_dir=None, weights=None, opt=None, hyp=None, logger=None, include=LOGGERS): + self.save_dir = save_dir + self.weights = weights + self.opt = opt + self.hyp = hyp + self.logger = logger # for printing results to console + self.include = include + self.keys = [ + 'train/box_loss', + 'train/obj_loss', + 'train/cls_loss', # train loss + 'metrics/precision', + 'metrics/recall', + 'metrics/mAP_0.5', + 'metrics/mAP_0.5:0.95', # metrics + 'val/box_loss', + 'val/obj_loss', + 'val/cls_loss', # val loss + 'x/lr0', + 'x/lr1', + 'x/lr2'] # params + self.best_keys = ['best/epoch', 'best/precision', 'best/recall', 'best/mAP_0.5', 'best/mAP_0.5:0.95'] + for k in LOGGERS: + setattr(self, k, None) # init empty logger dictionary + self.csv = True # always log to csv + + # Messages + if not wandb: + prefix = colorstr('Weights & Biases: ') + s = f"{prefix}run 'pip install wandb' to automatically track and visualize YOLOv5 🚀 runs in Weights & Biases" + self.logger.info(s) + if not clearml: + prefix = colorstr('ClearML: ') + s = f"{prefix}run 'pip install clearml' to automatically track, visualize and remotely train YOLOv5 🚀 in ClearML" + self.logger.info(s) + + # TensorBoard + s = self.save_dir + if 'tb' in self.include and not self.opt.evolve: + prefix = colorstr('TensorBoard: ') + self.logger.info(f"{prefix}Start with 'tensorboard --logdir {s.parent}', view at http://localhost:6006/") + self.tb = SummaryWriter(str(s)) + + # W&B + if wandb and 'wandb' in self.include: + wandb_artifact_resume = isinstance(self.opt.resume, str) and self.opt.resume.startswith('wandb-artifact://') + run_id = torch.load(self.weights).get('wandb_id') if self.opt.resume and not wandb_artifact_resume else None + self.opt.hyp = self.hyp # add hyperparameters + self.wandb = WandbLogger(self.opt, run_id) + # temp warn. because nested artifacts not supported after 0.12.10 + if pkg.parse_version(wandb.__version__) >= pkg.parse_version('0.12.11'): + s = "YOLOv5 temporarily requires wandb version 0.12.10 or below. Some features may not work as expected." + self.logger.warning(s) + else: + self.wandb = None + + # ClearML + if clearml and 'clearml' in self.include: + self.clearml = ClearmlLogger(self.opt, self.hyp) + else: + self.clearml = None + + def on_train_start(self): + # Callback runs on train start + pass + + def on_pretrain_routine_end(self): + # Callback runs on pre-train routine end + paths = self.save_dir.glob('*labels*.jpg') # training labels + if self.wandb: + self.wandb.log({"Labels": [wandb.Image(str(x), caption=x.name) for x in paths]}) + if self.clearml: + pass # ClearML saves these images automatically using hooks + + def on_train_batch_end(self, ni, model, imgs, targets, paths, plots): + # Callback runs on train batch end + # ni: number integrated batches (since train start) + if plots: + if ni == 0 and not self.opt.sync_bn and self.tb: + log_tensorboard_graph(self.tb, model, imgsz=list(imgs.shape[2:4])) + if ni < 3: + f = self.save_dir / f'train_batch{ni}.jpg' # filename + plot_images(imgs, targets, paths, f) + if (self.wandb or self.clearml) and ni == 10: + files = sorted(self.save_dir.glob('train*.jpg')) + if self.wandb: + self.wandb.log({'Mosaics': [wandb.Image(str(f), caption=f.name) for f in files if f.exists()]}) + if self.clearml: + self.clearml.log_debug_samples(files, title='Mosaics') + + def on_train_epoch_end(self, epoch): + # Callback runs on train epoch end + if self.wandb: + self.wandb.current_epoch = epoch + 1 + + def on_val_image_end(self, pred, predn, path, names, im): + # Callback runs on val image end + if self.wandb: + self.wandb.val_one_image(pred, predn, path, names, im) + if self.clearml: + self.clearml.log_image_with_boxes(path, pred, names, im) + + def on_val_end(self): + # Callback runs on val end + if self.wandb or self.clearml: + files = sorted(self.save_dir.glob('val*.jpg')) + if self.wandb: + self.wandb.log({"Validation": [wandb.Image(str(f), caption=f.name) for f in files]}) + if self.clearml: + self.clearml.log_debug_samples(files, title='Validation') + + def on_fit_epoch_end(self, vals, epoch, best_fitness, fi): + # Callback runs at the end of each fit (train+val) epoch + x = dict(zip(self.keys, vals)) + if self.csv: + file = self.save_dir / 'results.csv' + n = len(x) + 1 # number of cols + s = '' if file.exists() else (('%20s,' * n % tuple(['epoch'] + self.keys)).rstrip(',') + '\n') # add header + with open(file, 'a') as f: + f.write(s + ('%20.5g,' * n % tuple([epoch] + vals)).rstrip(',') + '\n') + + if self.tb: + for k, v in x.items(): + self.tb.add_scalar(k, v, epoch) + elif self.clearml: # log to ClearML if TensorBoard not used + for k, v in x.items(): + title, series = k.split('/') + self.clearml.task.get_logger().report_scalar(title, series, v, epoch) + + if self.wandb: + if best_fitness == fi: + best_results = [epoch] + vals[3:7] + for i, name in enumerate(self.best_keys): + self.wandb.wandb_run.summary[name] = best_results[i] # log best results in the summary + self.wandb.log(x) + self.wandb.end_epoch(best_result=best_fitness == fi) + + if self.clearml: + self.clearml.current_epoch_logged_images = set() # reset epoch image limit + self.clearml.current_epoch += 1 + + def on_model_save(self, last, epoch, final_epoch, best_fitness, fi): + # Callback runs on model save event + if self.wandb: + if ((epoch + 1) % self.opt.save_period == 0 and not final_epoch) and self.opt.save_period != -1: + self.wandb.log_model(last.parent, self.opt, epoch, fi, best_model=best_fitness == fi) + + if self.clearml: + if ((epoch + 1) % self.opt.save_period == 0 and not final_epoch) and self.opt.save_period != -1: + self.clearml.task.update_output_model(model_path=str(last), + model_name='Latest Model', + auto_delete_file=False) + + def on_train_end(self, last, best, plots, epoch, results): + # Callback runs on training end + if plots: + plot_results(file=self.save_dir / 'results.csv') # save results.png + files = ['results.png', 'confusion_matrix.png', *(f'{x}_curve.png' for x in ('F1', 'PR', 'P', 'R'))] + files = [(self.save_dir / f) for f in files if (self.save_dir / f).exists()] # filter + self.logger.info(f"Results saved to {colorstr('bold', self.save_dir)}") + + if self.tb and not self.clearml: # These images are already captured by ClearML by now, we don't want doubles + for f in files: + self.tb.add_image(f.stem, cv2.imread(str(f))[..., ::-1], epoch, dataformats='HWC') + + if self.wandb: + self.wandb.log(dict(zip(self.keys[3:10], results))) + self.wandb.log({"Results": [wandb.Image(str(f), caption=f.name) for f in files]}) + # Calling wandb.log. TODO: Refactor this into WandbLogger.log_model + if not self.opt.evolve: + wandb.log_artifact(str(best if best.exists() else last), + type='model', + name=f'run_{self.wandb.wandb_run.id}_model', + aliases=['latest', 'best', 'stripped']) + self.wandb.finish_run() + + if self.clearml: + # Save the best model here + if not self.opt.evolve: + self.clearml.task.update_output_model(model_path=str(best if best.exists() else last), + name='Best Model') + + def on_params_update(self, params): + # Update hyperparams or configs of the experiment + # params: A dict containing {param: value} pairs + if self.wandb: + self.wandb.wandb_run.config.update(params, allow_val_change=True) + + +class GenericLogger: + """ + YOLOv5 General purpose logger for non-task specific logging + Usage: from utils.loggers import GenericLogger; logger = GenericLogger(...) + Arguments + opt: Run arguments + console_logger: Console logger + include: loggers to include + """ + + def __init__(self, opt, console_logger, include=('tb', 'wandb')): + # init default loggers + self.save_dir = opt.save_dir + self.include = include + self.console_logger = console_logger + if 'tb' in self.include: + prefix = colorstr('TensorBoard: ') + self.console_logger.info( + f"{prefix}Start with 'tensorboard --logdir {self.save_dir.parent}', view at http://localhost:6006/") + self.tb = SummaryWriter(str(self.save_dir)) + + if wandb and 'wandb' in self.include: + self.wandb = wandb.init(project="YOLOv5-Classifier" if opt.project == "runs/train" else opt.project, + name=None if opt.name == "exp" else opt.name, + config=opt) + else: + self.wandb = None + + def log_metrics(self, metrics_dict, epoch): + # Log metrics dictionary to all loggers + if self.tb: + for k, v in metrics_dict.items(): + self.tb.add_scalar(k, v, epoch) + + if self.wandb: + self.wandb.log(metrics_dict, step=epoch) + + def log_images(self, files, name='Images', epoch=0): + # Log images to all loggers + files = [Path(f) for f in (files if isinstance(files, (tuple, list)) else [files])] # to Path + files = [f for f in files if f.exists()] # filter by exists + + if self.tb: + for f in files: + self.tb.add_image(f.stem, cv2.imread(str(f))[..., ::-1], epoch, dataformats='HWC') + + if self.wandb: + self.wandb.log({name: [wandb.Image(str(f), caption=f.name) for f in files]}, step=epoch) + + def log_graph(self, model, imgsz=(640, 640)): + # Log model graph to all loggers + if self.tb: + log_tensorboard_graph(self.tb, model, imgsz) + + def log_model(self, model_path, epoch=0, metadata={}): + # Log model to all loggers + if self.wandb: + art = wandb.Artifact(name=f"run_{wandb.run.id}_model", type="model", metadata=metadata) + art.add_file(str(model_path)) + wandb.log_artifact(art) + + +def log_tensorboard_graph(tb, model, imgsz=(640, 640)): + # Log model graph to TensorBoard + try: + p = next(model.parameters()) # for device, type + imgsz = (imgsz, imgsz) if isinstance(imgsz, int) else imgsz # expand + im = torch.zeros((1, 3, *imgsz)).to(p.device).type_as(p) # input image + with warnings.catch_warnings(): + warnings.simplefilter('ignore') # suppress jit trace warning + tb.add_graph(torch.jit.trace(de_parallel(model), im, strict=False), []) + except Exception: + print('WARNING: TensorBoard graph visualization failure') diff --git a/application/yolov5_example/utils/loggers/clearml/README.md b/application/yolov5_example/utils/loggers/clearml/README.md new file mode 100644 index 00000000..64eef6be --- /dev/null +++ b/application/yolov5_example/utils/loggers/clearml/README.md @@ -0,0 +1,222 @@ +# ClearML Integration + +Clear|MLClear|ML + +## About ClearML + +[ClearML](https://cutt.ly/yolov5-tutorial-clearml) is an [open-source](https://github.com/allegroai/clearml) toolbox designed to save you time ⏱️. + +🔨 Track every YOLOv5 training run in the experiment manager + +🔧 Version and easily access your custom training data with the integrated ClearML Data Versioning Tool + +🔦 Remotely train and monitor your YOLOv5 training runs using ClearML Agent + +🔬 Get the very best mAP using ClearML Hyperparameter Optimization + +🔭 Turn your newly trained YOLOv5 model into an API with just a few commands using ClearML Serving + +
+And so much more. It's up to you how many of these tools you want to use, you can stick to the experiment manager, or chain them all together into an impressive pipeline! +
+
+ +![ClearML scalars dashboard](https://github.com/thepycoder/clearml_screenshots/raw/main/experiment_manager_with_compare.gif) + + +
+
+ +## 🦾 Setting Things Up + +To keep track of your experiments and/or data, ClearML needs to communicate to a server. You have 2 options to get one: + +Either sign up for free to the [ClearML Hosted Service](https://cutt.ly/yolov5-tutorial-clearml) or you can set up your own server, see [here](https://clear.ml/docs/latest/docs/deploying_clearml/clearml_server). Even the server is open-source, so even if you're dealing with sensitive data, you should be good to go! + +1. Install the `clearml` python package: + + ```bash + pip install clearml + ``` + +1. Connect the ClearML SDK to the server by [creating credentials](https://app.clear.ml/settings/workspace-configuration) (go right top to Settings -> Workspace -> Create new credentials), then execute the command below and follow the instructions: + + ```bash + clearml-init + ``` + +That's it! You're done 😎 + +
+ +## 🚀 Training YOLOv5 With ClearML + +To enable ClearML experiment tracking, simply install the ClearML pip package. + +```bash +pip install clearml +``` + +This will enable integration with the YOLOv5 training script. Every training run from now on, will be captured and stored by the ClearML experiment manager. If you want to change the `project_name` or `task_name`, head over to our custom logger, where you can change it: `utils/loggers/clearml/clearml_utils.py` + +```bash +python train.py --img 640 --batch 16 --epochs 3 --data coco128.yaml --weights yolov5s.pt --cache +``` + +This will capture: +- Source code + uncommitted changes +- Installed packages +- (Hyper)parameters +- Model files (use `--save-period n` to save a checkpoint every n epochs) +- Console output +- Scalars (mAP_0.5, mAP_0.5:0.95, precision, recall, losses, learning rates, ...) +- General info such as machine details, runtime, creation date etc. +- All produced plots such as label correlogram and confusion matrix +- Images with bounding boxes per epoch +- Mosaic per epoch +- Validation images per epoch +- ... + +That's a lot right? 🤯 +Now, we can visualize all of this information in the ClearML UI to get an overview of our training progress. Add custom columns to the table view (such as e.g. mAP_0.5) so you can easily sort on the best performing model. Or select multiple experiments and directly compare them! + +There even more we can do with all of this information, like hyperparameter optimization and remote execution, so keep reading if you want to see how that works! + +
+ +## 🔗 Dataset Version Management + +Versioning your data separately from your code is generally a good idea and makes it easy to aqcuire the latest version too. This repository supports supplying a dataset version ID and it will make sure to get the data if it's not there yet. Next to that, this workflow also saves the used dataset ID as part of the task parameters, so you will always know for sure which data was used in which experiment! + +![ClearML Dataset Interface](https://github.com/thepycoder/clearml_screenshots/raw/main/clearml_data.gif) + +### Prepare Your Dataset + +The YOLOv5 repository supports a number of different datasets by using yaml files containing their information. By default datasets are downloaded to the `../datasets` folder in relation to the repository root folder. So if you downloaded the `coco128` dataset using the link in the yaml or with the scripts provided by yolov5, you get this folder structure: + +``` +.. +|_ yolov5 +|_ datasets + |_ coco128 + |_ images + |_ labels + |_ LICENSE + |_ README.txt +``` +But this can be any dataset you wish. Feel free to use your own, as long as you keep to this folder structure. + +Next, ⚠️**copy the corresponding yaml file to the root of the dataset folder**⚠️. This yaml files contains the information ClearML will need to properly use the dataset. You can make this yourself too, of course, just follow the structure of the example yamls. + +Basically we need the following keys: `path`, `train`, `test`, `val`, `nc`, `names`. + +``` +.. +|_ yolov5 +|_ datasets + |_ coco128 + |_ images + |_ labels + |_ coco128.yaml # <---- HERE! + |_ LICENSE + |_ README.txt +``` + +### Upload Your Dataset + +To get this dataset into ClearML as a versionned dataset, go to the dataset root folder and run the following command: +```bash +cd coco128 +clearml-data sync --project YOLOv5 --name coco128 --folder . +``` + +The command `clearml-data sync` is actually a shorthand command. You could also run these commands one after the other: +```bash +# Optionally add --parent if you want to base +# this version on another dataset version, so no duplicate files are uploaded! +clearml-data create --name coco128 --project YOLOv5 +clearml-data add --files . +clearml-data close +``` + +### Run Training Using A ClearML Dataset + +Now that you have a ClearML dataset, you can very simply use it to train custom YOLOv5 🚀 models! + +```bash +python train.py --img 640 --batch 16 --epochs 3 --data clearml:// --weights yolov5s.pt --cache +``` + +
+ +## 👀 Hyperparameter Optimization + +Now that we have our experiments and data versioned, it's time to take a look at what we can build on top! + +Using the code information, installed packages and environment details, the experiment itself is now **completely reproducible**. In fact, ClearML allows you to clone an experiment and even change its parameters. We can then just rerun it with these new parameters automatically, this is basically what HPO does! + +To **run hyperparameter optimization locally**, we've included a pre-made script for you. Just make sure a training task has been run at least once, so it is in the ClearML experiment manager, we will essentially clone it and change its hyperparameters. + +You'll need to fill in the ID of this `template task` in the script found at `utils/loggers/clearml/hpo.py` and then just run it :) You can change `task.execute_locally()` to `task.execute()` to put it in a ClearML queue and have a remote agent work on it instead. + +```bash +# To use optuna, install it first, otherwise you can change the optimizer to just be RandomSearch +pip install optuna +python utils/loggers/clearml/hpo.py +``` + +![HPO](https://github.com/thepycoder/clearml_screenshots/raw/main/hpo.png) + +## 🤯 Remote Execution (advanced) + +Running HPO locally is really handy, but what if we want to run our experiments on a remote machine instead? Maybe you have access to a very powerful GPU machine on-site or you have some budget to use cloud GPUs. +This is where the ClearML Agent comes into play. Check out what the agent can do here: + +- [YouTube video](https://youtu.be/MX3BrXnaULs) +- [Documentation](https://clear.ml/docs/latest/docs/clearml_agent) + +In short: every experiment tracked by the experiment manager contains enough information to reproduce it on a different machine (installed packages, uncommitted changes etc.). So a ClearML agent does just that: it listens to a queue for incoming tasks and when it finds one, it recreates the environment and runs it while still reporting scalars, plots etc. to the experiment manager. + +You can turn any machine (a cloud VM, a local GPU machine, your own laptop ... ) into a ClearML agent by simply running: +```bash +clearml-agent daemon --queue [--docker] +``` + +### Cloning, Editing And Enqueuing + +With our agent running, we can give it some work. Remember from the HPO section that we can clone a task and edit the hyperparameters? We can do that from the interface too! + +🪄 Clone the experiment by right clicking it + +🎯 Edit the hyperparameters to what you wish them to be + +⏳ Enqueue the task to any of the queues by right clicking it + +![Enqueue a task from the UI](https://github.com/thepycoder/clearml_screenshots/raw/main/enqueue.gif) + +### Executing A Task Remotely + +Now you can clone a task like we explained above, or simply mark your current script by adding `task.execute_remotely()` and on execution it will be put into a queue, for the agent to start working on! + +To run the YOLOv5 training script remotely, all you have to do is add this line to the training.py script after the clearml logger has been instatiated: +```python +# ... +# Loggers +data_dict = None +if RANK in {-1, 0}: + loggers = Loggers(save_dir, weights, opt, hyp, LOGGER) # loggers instance + if loggers.clearml: + loggers.clearml.task.execute_remotely(queue='my_queue') # <------ ADD THIS LINE + # Data_dict is either None is user did not choose for ClearML dataset or is filled in by ClearML + data_dict = loggers.clearml.data_dict +# ... +``` +When running the training script after this change, python will run the script up until that line, after which it will package the code and send it to the queue instead! + +### Autoscaling workers + +ClearML comes with autoscalers too! This tool will automatically spin up new remote machines in the cloud of your choice (AWS, GCP, Azure) and turn them into ClearML agents for you whenever there are experiments detected in the queue. Once the tasks are processed, the autoscaler will automatically shut down the remote machines and you stop paying! + +Check out the autoscalers getting started video below. + +[![Watch the video](https://img.youtube.com/vi/j4XVMAaUt3E/0.jpg)](https://youtu.be/j4XVMAaUt3E) diff --git a/application/yolov5_example/utils/loggers/clearml/__init__.py b/application/yolov5_example/utils/loggers/clearml/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/application/yolov5_example/utils/loggers/clearml/clearml_utils.py b/application/yolov5_example/utils/loggers/clearml/clearml_utils.py new file mode 100644 index 00000000..52320c09 --- /dev/null +++ b/application/yolov5_example/utils/loggers/clearml/clearml_utils.py @@ -0,0 +1,156 @@ +"""Main Logger class for ClearML experiment tracking.""" +import glob +import re +from pathlib import Path + +import numpy as np +import yaml + +from utils.plots import Annotator, colors + +try: + import clearml + from clearml import Dataset, Task + assert hasattr(clearml, '__version__') # verify package import not local dir +except (ImportError, AssertionError): + clearml = None + + +def construct_dataset(clearml_info_string): + """Load in a clearml dataset and fill the internal data_dict with its contents. + """ + dataset_id = clearml_info_string.replace('clearml://', '') + dataset = Dataset.get(dataset_id=dataset_id) + dataset_root_path = Path(dataset.get_local_copy()) + + # We'll search for the yaml file definition in the dataset + yaml_filenames = list(glob.glob(str(dataset_root_path / "*.yaml")) + glob.glob(str(dataset_root_path / "*.yml"))) + if len(yaml_filenames) > 1: + raise ValueError('More than one yaml file was found in the dataset root, cannot determine which one contains ' + 'the dataset definition this way.') + elif len(yaml_filenames) == 0: + raise ValueError('No yaml definition found in dataset root path, check that there is a correct yaml file ' + 'inside the dataset root path.') + with open(yaml_filenames[0]) as f: + dataset_definition = yaml.safe_load(f) + + assert set(dataset_definition.keys()).issuperset( + {'train', 'test', 'val', 'nc', 'names'} + ), "The right keys were not found in the yaml file, make sure it at least has the following keys: ('train', 'test', 'val', 'nc', 'names')" + + data_dict = dict() + data_dict['train'] = str( + (dataset_root_path / dataset_definition['train']).resolve()) if dataset_definition['train'] else None + data_dict['test'] = str( + (dataset_root_path / dataset_definition['test']).resolve()) if dataset_definition['test'] else None + data_dict['val'] = str( + (dataset_root_path / dataset_definition['val']).resolve()) if dataset_definition['val'] else None + data_dict['nc'] = dataset_definition['nc'] + data_dict['names'] = dataset_definition['names'] + + return data_dict + + +class ClearmlLogger: + """Log training runs, datasets, models, and predictions to ClearML. + + This logger sends information to ClearML at app.clear.ml or to your own hosted server. By default, + this information includes hyperparameters, system configuration and metrics, model metrics, code information and + basic data metrics and analyses. + + By providing additional command line arguments to train.py, datasets, + models and predictions can also be logged. + """ + + def __init__(self, opt, hyp): + """ + - Initialize ClearML Task, this object will capture the experiment + - Upload dataset version to ClearML Data if opt.upload_dataset is True + + arguments: + opt (namespace) -- Commandline arguments for this run + hyp (dict) -- Hyperparameters for this run + + """ + self.current_epoch = 0 + # Keep tracked of amount of logged images to enforce a limit + self.current_epoch_logged_images = set() + # Maximum number of images to log to clearML per epoch + self.max_imgs_to_log_per_epoch = 16 + # Get the interval of epochs when bounding box images should be logged + self.bbox_interval = opt.bbox_interval + self.clearml = clearml + self.task = None + self.data_dict = None + if self.clearml: + self.task = Task.init( + project_name='YOLOv5', + task_name='training', + tags=['YOLOv5'], + output_uri=True, + auto_connect_frameworks={'pytorch': False} + # We disconnect pytorch auto-detection, because we added manual model save points in the code + ) + # ClearML's hooks will already grab all general parameters + # Only the hyperparameters coming from the yaml config file + # will have to be added manually! + self.task.connect(hyp, name='Hyperparameters') + + # Get ClearML Dataset Version if requested + if opt.data.startswith('clearml://'): + # data_dict should have the following keys: + # names, nc (number of classes), test, train, val (all three relative paths to ../datasets) + self.data_dict = construct_dataset(opt.data) + # Set data to data_dict because wandb will crash without this information and opt is the best way + # to give it to them + opt.data = self.data_dict + + def log_debug_samples(self, files, title='Debug Samples'): + """ + Log files (images) as debug samples in the ClearML task. + + arguments: + files (List(PosixPath)) a list of file paths in PosixPath format + title (str) A title that groups together images with the same values + """ + for f in files: + if f.exists(): + it = re.search(r'_batch(\d+)', f.name) + iteration = int(it.groups()[0]) if it else 0 + self.task.get_logger().report_image(title=title, + series=f.name.replace(it.group(), ''), + local_path=str(f), + iteration=iteration) + + def log_image_with_boxes(self, image_path, boxes, class_names, image, conf_threshold=0.25): + """ + Draw the bounding boxes on a single image and report the result as a ClearML debug sample. + + arguments: + image_path (PosixPath) the path the original image file + boxes (list): list of scaled predictions in the format - [xmin, ymin, xmax, ymax, confidence, class] + class_names (dict): dict containing mapping of class int to class name + image (Tensor): A torch tensor containing the actual image data + """ + if len(self.current_epoch_logged_images) < self.max_imgs_to_log_per_epoch and self.current_epoch >= 0: + # Log every bbox_interval times and deduplicate for any intermittend extra eval runs + if self.current_epoch % self.bbox_interval == 0 and image_path not in self.current_epoch_logged_images: + im = np.ascontiguousarray(np.moveaxis(image.mul(255).clamp(0, 255).byte().cpu().numpy(), 0, 2)) + annotator = Annotator(im=im, pil=True) + for i, (conf, class_nr, box) in enumerate(zip(boxes[:, 4], boxes[:, 5], boxes[:, :4])): + color = colors(i) + + class_name = class_names[int(class_nr)] + confidence = round(float(conf) * 100, 2) + label = f"{class_name}: {confidence}%" + + if confidence > conf_threshold: + annotator.rectangle(box.cpu().numpy(), outline=color) + annotator.box_label(box.cpu().numpy(), label=label, color=color) + + annotated_image = annotator.result() + self.task.get_logger().report_image(title='Bounding Boxes', + series=image_path.name, + iteration=self.current_epoch, + image=annotated_image) + self.current_epoch_logged_images.add(image_path) diff --git a/application/yolov5_example/utils/loggers/clearml/hpo.py b/application/yolov5_example/utils/loggers/clearml/hpo.py new file mode 100644 index 00000000..96c2c544 --- /dev/null +++ b/application/yolov5_example/utils/loggers/clearml/hpo.py @@ -0,0 +1,84 @@ +from clearml import Task +# Connecting ClearML with the current process, +# from here on everything is logged automatically +from clearml.automation import HyperParameterOptimizer, UniformParameterRange +from clearml.automation.optuna import OptimizerOptuna + +task = Task.init(project_name='Hyper-Parameter Optimization', + task_name='YOLOv5', + task_type=Task.TaskTypes.optimizer, + reuse_last_task_id=False) + +# Example use case: +optimizer = HyperParameterOptimizer( + # This is the experiment we want to optimize + base_task_id='', + # here we define the hyper-parameters to optimize + # Notice: The parameter name should exactly match what you see in the UI: / + # For Example, here we see in the base experiment a section Named: "General" + # under it a parameter named "batch_size", this becomes "General/batch_size" + # If you have `argparse` for example, then arguments will appear under the "Args" section, + # and you should instead pass "Args/batch_size" + hyper_parameters=[ + UniformParameterRange('Hyperparameters/lr0', min_value=1e-5, max_value=1e-1), + UniformParameterRange('Hyperparameters/lrf', min_value=0.01, max_value=1.0), + UniformParameterRange('Hyperparameters/momentum', min_value=0.6, max_value=0.98), + UniformParameterRange('Hyperparameters/weight_decay', min_value=0.0, max_value=0.001), + UniformParameterRange('Hyperparameters/warmup_epochs', min_value=0.0, max_value=5.0), + UniformParameterRange('Hyperparameters/warmup_momentum', min_value=0.0, max_value=0.95), + UniformParameterRange('Hyperparameters/warmup_bias_lr', min_value=0.0, max_value=0.2), + UniformParameterRange('Hyperparameters/box', min_value=0.02, max_value=0.2), + UniformParameterRange('Hyperparameters/cls', min_value=0.2, max_value=4.0), + UniformParameterRange('Hyperparameters/cls_pw', min_value=0.5, max_value=2.0), + UniformParameterRange('Hyperparameters/obj', min_value=0.2, max_value=4.0), + UniformParameterRange('Hyperparameters/obj_pw', min_value=0.5, max_value=2.0), + UniformParameterRange('Hyperparameters/iou_t', min_value=0.1, max_value=0.7), + UniformParameterRange('Hyperparameters/anchor_t', min_value=2.0, max_value=8.0), + UniformParameterRange('Hyperparameters/fl_gamma', min_value=0.0, max_value=4.0), + UniformParameterRange('Hyperparameters/hsv_h', min_value=0.0, max_value=0.1), + UniformParameterRange('Hyperparameters/hsv_s', min_value=0.0, max_value=0.9), + UniformParameterRange('Hyperparameters/hsv_v', min_value=0.0, max_value=0.9), + UniformParameterRange('Hyperparameters/degrees', min_value=0.0, max_value=45.0), + UniformParameterRange('Hyperparameters/translate', min_value=0.0, max_value=0.9), + UniformParameterRange('Hyperparameters/scale', min_value=0.0, max_value=0.9), + UniformParameterRange('Hyperparameters/shear', min_value=0.0, max_value=10.0), + UniformParameterRange('Hyperparameters/perspective', min_value=0.0, max_value=0.001), + UniformParameterRange('Hyperparameters/flipud', min_value=0.0, max_value=1.0), + UniformParameterRange('Hyperparameters/fliplr', min_value=0.0, max_value=1.0), + UniformParameterRange('Hyperparameters/mosaic', min_value=0.0, max_value=1.0), + UniformParameterRange('Hyperparameters/mixup', min_value=0.0, max_value=1.0), + UniformParameterRange('Hyperparameters/copy_paste', min_value=0.0, max_value=1.0)], + # this is the objective metric we want to maximize/minimize + objective_metric_title='metrics', + objective_metric_series='mAP_0.5', + # now we decide if we want to maximize it or minimize it (accuracy we maximize) + objective_metric_sign='max', + # let us limit the number of concurrent experiments, + # this in turn will make sure we do dont bombard the scheduler with experiments. + # if we have an auto-scaler connected, this, by proxy, will limit the number of machine + max_number_of_concurrent_tasks=1, + # this is the optimizer class (actually doing the optimization) + # Currently, we can choose from GridSearch, RandomSearch or OptimizerBOHB (Bayesian optimization Hyper-Band) + optimizer_class=OptimizerOptuna, + # If specified only the top K performing Tasks will be kept, the others will be automatically archived + save_top_k_tasks_only=5, # 5, + compute_time_limit=None, + total_max_jobs=20, + min_iteration_per_job=None, + max_iteration_per_job=None, +) + +# report every 10 seconds, this is way too often, but we are testing here +optimizer.set_report_period(10) +# You can also use the line below instead to run all the optimizer tasks locally, without using queues or agent +# an_optimizer.start_locally(job_complete_callback=job_complete_callback) +# set the time limit for the optimization process (2 hours) +optimizer.set_time_limit(in_minutes=120.0) +# Start the optimization process in the local environment +optimizer.start_locally() +# wait until process is done (notice we are controlling the optimization process in the background) +optimizer.wait() +# make sure background optimization stopped +optimizer.stop() + +print('We are done, good bye') diff --git a/application/yolov5_example/utils/loggers/wandb/README.md b/application/yolov5_example/utils/loggers/wandb/README.md new file mode 100644 index 00000000..d78324b4 --- /dev/null +++ b/application/yolov5_example/utils/loggers/wandb/README.md @@ -0,0 +1,162 @@ +📚 This guide explains how to use **Weights & Biases** (W&B) with YOLOv5 🚀. UPDATED 29 September 2021. + +- [About Weights & Biases](#about-weights-&-biases) +- [First-Time Setup](#first-time-setup) +- [Viewing runs](#viewing-runs) +- [Disabling wandb](#disabling-wandb) +- [Advanced Usage: Dataset Versioning and Evaluation](#advanced-usage) +- [Reports: Share your work with the world!](#reports) + +## About Weights & Biases + +Think of [W&B](https://wandb.ai/site?utm_campaign=repo_yolo_wandbtutorial) like GitHub for machine learning models. With a few lines of code, save everything you need to debug, compare and reproduce your models — architecture, hyperparameters, git commits, model weights, GPU usage, and even datasets and predictions. + +Used by top researchers including teams at OpenAI, Lyft, Github, and MILA, W&B is part of the new standard of best practices for machine learning. How W&B can help you optimize your machine learning workflows: + +- [Debug](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#Free-2) model performance in real time +- [GPU usage](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#System-4) visualized automatically +- [Custom charts](https://wandb.ai/wandb/customizable-charts/reports/Powerful-Custom-Charts-To-Debug-Model-Peformance--VmlldzoyNzY4ODI) for powerful, extensible visualization +- [Share insights](https://wandb.ai/wandb/getting-started/reports/Visualize-Debug-Machine-Learning-Models--VmlldzoyNzY5MDk#Share-8) interactively with collaborators +- [Optimize hyperparameters](https://docs.wandb.com/sweeps) efficiently +- [Track](https://docs.wandb.com/artifacts) datasets, pipelines, and production models + +## First-Time Setup + +
+ Toggle Details +When you first train, W&B will prompt you to create a new account and will generate an **API key** for you. If you are an existing user you can retrieve your key from https://wandb.ai/authorize. This key is used to tell W&B where to log your data. You only need to supply your key once, and then it is remembered on the same device. + +W&B will create a cloud **project** (default is 'YOLOv5') for your training runs, and each new training run will be provided a unique run **name** within that project as project/name. You can also manually set your project and run name as: + +```shell +$ python train.py --project ... --name ... +``` + +YOLOv5 notebook example: Open In Colab Open In Kaggle +Screen Shot 2021-09-29 at 10 23 13 PM + +
+ +## Viewing Runs + +
+ Toggle Details +Run information streams from your environment to the W&B cloud console as you train. This allows you to monitor and even cancel runs in realtime . All important information is logged: + +- Training & Validation losses +- Metrics: Precision, Recall, mAP@0.5, mAP@0.5:0.95 +- Learning Rate over time +- A bounding box debugging panel, showing the training progress over time +- GPU: Type, **GPU Utilization**, power, temperature, **CUDA memory usage** +- System: Disk I/0, CPU utilization, RAM memory usage +- Your trained model as W&B Artifact +- Environment: OS and Python types, Git repository and state, **training command** + +

Weights & Biases dashboard

+
+ +## Disabling wandb + +- training after running `wandb disabled` inside that directory creates no wandb run + ![Screenshot (84)](https://user-images.githubusercontent.com/15766192/143441777-c780bdd7-7cb4-4404-9559-b4316030a985.png) + +- To enable wandb again, run `wandb online` + ![Screenshot (85)](https://user-images.githubusercontent.com/15766192/143441866-7191b2cb-22f0-4e0f-ae64-2dc47dc13078.png) + +## Advanced Usage + +You can leverage W&B artifacts and Tables integration to easily visualize and manage your datasets, models and training evaluations. Here are some quick examples to get you started. + +
+

1: Train and Log Evaluation simultaneousy

+ This is an extension of the previous section, but it'll also training after uploading the dataset. This also evaluation Table + Evaluation table compares your predictions and ground truths across the validation set for each epoch. It uses the references to the already uploaded datasets, + so no images will be uploaded from your system more than once. +
+ Usage + Code $ python train.py --upload_data val + +![Screenshot from 2021-11-21 17-40-06](https://user-images.githubusercontent.com/15766192/142761183-c1696d8c-3f38-45ab-991a-bb0dfd98ae7d.png) + +
+ +

2. Visualize and Version Datasets

+ Log, visualize, dynamically query, and understand your data with W&B Tables. You can use the following command to log your dataset as a W&B Table. This will generate a {dataset}_wandb.yaml file which can be used to train from dataset artifact. +
+ Usage + Code $ python utils/logger/wandb/log_dataset.py --project ... --name ... --data .. + +![Screenshot (64)](https://user-images.githubusercontent.com/15766192/128486078-d8433890-98a3-4d12-8986-b6c0e3fc64b9.png) + +
+ +

3: Train using dataset artifact

+ When you upload a dataset as described in the first section, you get a new config file with an added `_wandb` to its name. This file contains the information that + can be used to train a model directly from the dataset artifact. This also logs evaluation +
+ Usage + Code $ python train.py --data {data}_wandb.yaml + +![Screenshot (72)](https://user-images.githubusercontent.com/15766192/128979739-4cf63aeb-a76f-483f-8861-1c0100b938a5.png) + +
+ +

4: Save model checkpoints as artifacts

+ To enable saving and versioning checkpoints of your experiment, pass `--save_period n` with the base cammand, where `n` represents checkpoint interval. + You can also log both the dataset and model checkpoints simultaneously. If not passed, only the final model will be logged + +
+ Usage + Code $ python train.py --save_period 1 + +![Screenshot (68)](https://user-images.githubusercontent.com/15766192/128726138-ec6c1f60-639d-437d-b4ee-3acd9de47ef3.png) + +
+ +
+ +

5: Resume runs from checkpoint artifacts.

+Any run can be resumed using artifacts if the --resume argument starts with wandb-artifact:// prefix followed by the run path, i.e, wandb-artifact://username/project/runid . This doesn't require the model checkpoint to be present on the local system. + +
+ Usage + Code $ python train.py --resume wandb-artifact://{run_path} + +![Screenshot (70)](https://user-images.githubusercontent.com/15766192/128728988-4e84b355-6c87-41ae-a591-14aecf45343e.png) + +
+ +

6: Resume runs from dataset artifact & checkpoint artifacts.

+ Local dataset or model checkpoints are not required. This can be used to resume runs directly on a different device + The syntax is same as the previous section, but you'll need to lof both the dataset and model checkpoints as artifacts, i.e, set bot --upload_dataset or + train from _wandb.yaml file and set --save_period + +
+ Usage + Code $ python train.py --resume wandb-artifact://{run_path} + +![Screenshot (70)](https://user-images.githubusercontent.com/15766192/128728988-4e84b355-6c87-41ae-a591-14aecf45343e.png) + +
+ + + +

Reports

+W&B Reports can be created from your saved runs for sharing online. Once a report is created you will receive a link you can use to publically share your results. Here is an example report created from the COCO128 tutorial trainings of all four YOLOv5 models ([link](https://wandb.ai/glenn-jocher/yolov5_tutorial/reports/YOLOv5-COCO128-Tutorial-Results--VmlldzozMDI5OTY)). + +Weights & Biases Reports + +## Environments + +YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including [CUDA](https://developer.nvidia.com/cuda)/[CUDNN](https://developer.nvidia.com/cudnn), [Python](https://www.python.org/) and [PyTorch](https://pytorch.org/) preinstalled): + +- **Google Colab and Kaggle** notebooks with free GPU: Open In Colab Open In Kaggle +- **Google Cloud** Deep Learning VM. See [GCP Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/GCP-Quickstart) +- **Amazon** Deep Learning AMI. See [AWS Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/AWS-Quickstart) +- **Docker Image**. See [Docker Quickstart Guide](https://github.com/ultralytics/yolov5/wiki/Docker-Quickstart) Docker Pulls + +## Status + +![CI CPU testing](https://github.com/ultralytics/yolov5/workflows/CI%20CPU%20testing/badge.svg) + +If this badge is green, all [YOLOv5 GitHub Actions](https://github.com/ultralytics/yolov5/actions) Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training ([train.py](https://github.com/ultralytics/yolov5/blob/master/train.py)), validation ([val.py](https://github.com/ultralytics/yolov5/blob/master/val.py)), inference ([detect.py](https://github.com/ultralytics/yolov5/blob/master/detect.py)) and export ([export.py](https://github.com/ultralytics/yolov5/blob/master/export.py)) on macOS, Windows, and Ubuntu every 24 hours and on every commit. diff --git a/application/yolov5_example/utils/loggers/wandb/__init__.py b/application/yolov5_example/utils/loggers/wandb/__init__.py new file mode 100644 index 00000000..e69de29b diff --git a/application/yolov5_example/utils/loggers/wandb/log_dataset.py b/application/yolov5_example/utils/loggers/wandb/log_dataset.py new file mode 100644 index 00000000..06e81fb6 --- /dev/null +++ b/application/yolov5_example/utils/loggers/wandb/log_dataset.py @@ -0,0 +1,27 @@ +import argparse + +from wandb_utils import WandbLogger + +from utils.general import LOGGER + +WANDB_ARTIFACT_PREFIX = 'wandb-artifact://' + + +def create_dataset_artifact(opt): + logger = WandbLogger(opt, None, job_type='Dataset Creation') # TODO: return value unused + if not logger.wandb: + LOGGER.info("install wandb using `pip install wandb` to log the dataset") + + +if __name__ == '__main__': + parser = argparse.ArgumentParser() + parser.add_argument('--data', type=str, default='data/coco128.yaml', help='data.yaml path') + parser.add_argument('--single-cls', action='store_true', help='train as single-class dataset') + parser.add_argument('--project', type=str, default='YOLOv5', help='name of W&B Project') + parser.add_argument('--entity', default=None, help='W&B entity') + parser.add_argument('--name', type=str, default='log dataset', help='name of W&B run') + + opt = parser.parse_args() + opt.resume = False # Explicitly disallow resume check for dataset upload job + + create_dataset_artifact(opt) diff --git a/application/yolov5_example/utils/loggers/wandb/sweep.py b/application/yolov5_example/utils/loggers/wandb/sweep.py new file mode 100644 index 00000000..d49ea6f2 --- /dev/null +++ b/application/yolov5_example/utils/loggers/wandb/sweep.py @@ -0,0 +1,41 @@ +import sys +from pathlib import Path + +import wandb + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[3] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH + +from train import parse_opt, train +from utils.callbacks import Callbacks +from utils.general import increment_path +from utils.torch_utils import select_device + + +def sweep(): + wandb.init() + # Get hyp dict from sweep agent. Copy because train() modifies parameters which confused wandb. + hyp_dict = vars(wandb.config).get("_items").copy() + + # Workaround: get necessary opt args + opt = parse_opt(known=True) + opt.batch_size = hyp_dict.get("batch_size") + opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok or opt.evolve)) + opt.epochs = hyp_dict.get("epochs") + opt.nosave = True + opt.data = hyp_dict.get("data") + opt.weights = str(opt.weights) + opt.cfg = str(opt.cfg) + opt.data = str(opt.data) + opt.hyp = str(opt.hyp) + opt.project = str(opt.project) + device = select_device(opt.device, batch_size=opt.batch_size) + + # train + train(hyp_dict, opt, device, callbacks=Callbacks()) + + +if __name__ == "__main__": + sweep() diff --git a/application/yolov5_example/utils/loggers/wandb/sweep.yaml b/application/yolov5_example/utils/loggers/wandb/sweep.yaml new file mode 100644 index 00000000..688b1ea0 --- /dev/null +++ b/application/yolov5_example/utils/loggers/wandb/sweep.yaml @@ -0,0 +1,143 @@ +# Hyperparameters for training +# To set range- +# Provide min and max values as: +# parameter: +# +# min: scalar +# max: scalar +# OR +# +# Set a specific list of search space- +# parameter: +# values: [scalar1, scalar2, scalar3...] +# +# You can use grid, bayesian and hyperopt search strategy +# For more info on configuring sweeps visit - https://docs.wandb.ai/guides/sweeps/configuration + +program: utils/loggers/wandb/sweep.py +method: random +metric: + name: metrics/mAP_0.5 + goal: maximize + +parameters: + # hyperparameters: set either min, max range or values list + data: + value: "data/coco128.yaml" + batch_size: + values: [64] + epochs: + values: [10] + + lr0: + distribution: uniform + min: 1e-5 + max: 1e-1 + lrf: + distribution: uniform + min: 0.01 + max: 1.0 + momentum: + distribution: uniform + min: 0.6 + max: 0.98 + weight_decay: + distribution: uniform + min: 0.0 + max: 0.001 + warmup_epochs: + distribution: uniform + min: 0.0 + max: 5.0 + warmup_momentum: + distribution: uniform + min: 0.0 + max: 0.95 + warmup_bias_lr: + distribution: uniform + min: 0.0 + max: 0.2 + box: + distribution: uniform + min: 0.02 + max: 0.2 + cls: + distribution: uniform + min: 0.2 + max: 4.0 + cls_pw: + distribution: uniform + min: 0.5 + max: 2.0 + obj: + distribution: uniform + min: 0.2 + max: 4.0 + obj_pw: + distribution: uniform + min: 0.5 + max: 2.0 + iou_t: + distribution: uniform + min: 0.1 + max: 0.7 + anchor_t: + distribution: uniform + min: 2.0 + max: 8.0 + fl_gamma: + distribution: uniform + min: 0.0 + max: 4.0 + hsv_h: + distribution: uniform + min: 0.0 + max: 0.1 + hsv_s: + distribution: uniform + min: 0.0 + max: 0.9 + hsv_v: + distribution: uniform + min: 0.0 + max: 0.9 + degrees: + distribution: uniform + min: 0.0 + max: 45.0 + translate: + distribution: uniform + min: 0.0 + max: 0.9 + scale: + distribution: uniform + min: 0.0 + max: 0.9 + shear: + distribution: uniform + min: 0.0 + max: 10.0 + perspective: + distribution: uniform + min: 0.0 + max: 0.001 + flipud: + distribution: uniform + min: 0.0 + max: 1.0 + fliplr: + distribution: uniform + min: 0.0 + max: 1.0 + mosaic: + distribution: uniform + min: 0.0 + max: 1.0 + mixup: + distribution: uniform + min: 0.0 + max: 1.0 + copy_paste: + distribution: uniform + min: 0.0 + max: 1.0 diff --git a/application/yolov5_example/utils/loggers/wandb/wandb_utils.py b/application/yolov5_example/utils/loggers/wandb/wandb_utils.py new file mode 100644 index 00000000..e850d2ac --- /dev/null +++ b/application/yolov5_example/utils/loggers/wandb/wandb_utils.py @@ -0,0 +1,584 @@ +"""Utilities and tools for tracking runs with Weights & Biases.""" + +import logging +import os +import sys +from contextlib import contextmanager +from pathlib import Path +from typing import Dict + +import yaml +from tqdm import tqdm + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[3] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH + +from utils.dataloaders import LoadImagesAndLabels, img2label_paths +from utils.general import LOGGER, check_dataset, check_file + +try: + import wandb + + assert hasattr(wandb, '__version__') # verify package import not local dir +except (ImportError, AssertionError): + wandb = None + +RANK = int(os.getenv('RANK', -1)) +WANDB_ARTIFACT_PREFIX = 'wandb-artifact://' + + +def remove_prefix(from_string, prefix=WANDB_ARTIFACT_PREFIX): + return from_string[len(prefix):] + + +def check_wandb_config_file(data_config_file): + wandb_config = '_wandb.'.join(data_config_file.rsplit('.', 1)) # updated data.yaml path + if Path(wandb_config).is_file(): + return wandb_config + return data_config_file + + +def check_wandb_dataset(data_file): + is_trainset_wandb_artifact = False + is_valset_wandb_artifact = False + if isinstance(data_file, dict): + # In that case another dataset manager has already processed it and we don't have to + return data_file + if check_file(data_file) and data_file.endswith('.yaml'): + with open(data_file, errors='ignore') as f: + data_dict = yaml.safe_load(f) + is_trainset_wandb_artifact = isinstance(data_dict['train'], + str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX) + is_valset_wandb_artifact = isinstance(data_dict['val'], + str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX) + if is_trainset_wandb_artifact or is_valset_wandb_artifact: + return data_dict + else: + return check_dataset(data_file) + + +def get_run_info(run_path): + run_path = Path(remove_prefix(run_path, WANDB_ARTIFACT_PREFIX)) + run_id = run_path.stem + project = run_path.parent.stem + entity = run_path.parent.parent.stem + model_artifact_name = 'run_' + run_id + '_model' + return entity, project, run_id, model_artifact_name + + +def check_wandb_resume(opt): + process_wandb_config_ddp_mode(opt) if RANK not in [-1, 0] else None + if isinstance(opt.resume, str): + if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + if RANK not in [-1, 0]: # For resuming DDP runs + entity, project, run_id, model_artifact_name = get_run_info(opt.resume) + api = wandb.Api() + artifact = api.artifact(entity + '/' + project + '/' + model_artifact_name + ':latest') + modeldir = artifact.download() + opt.weights = str(Path(modeldir) / "last.pt") + return True + return None + + +def process_wandb_config_ddp_mode(opt): + with open(check_file(opt.data), errors='ignore') as f: + data_dict = yaml.safe_load(f) # data dict + train_dir, val_dir = None, None + if isinstance(data_dict['train'], str) and data_dict['train'].startswith(WANDB_ARTIFACT_PREFIX): + api = wandb.Api() + train_artifact = api.artifact(remove_prefix(data_dict['train']) + ':' + opt.artifact_alias) + train_dir = train_artifact.download() + train_path = Path(train_dir) / 'data/images/' + data_dict['train'] = str(train_path) + + if isinstance(data_dict['val'], str) and data_dict['val'].startswith(WANDB_ARTIFACT_PREFIX): + api = wandb.Api() + val_artifact = api.artifact(remove_prefix(data_dict['val']) + ':' + opt.artifact_alias) + val_dir = val_artifact.download() + val_path = Path(val_dir) / 'data/images/' + data_dict['val'] = str(val_path) + if train_dir or val_dir: + ddp_data_path = str(Path(val_dir) / 'wandb_local_data.yaml') + with open(ddp_data_path, 'w') as f: + yaml.safe_dump(data_dict, f) + opt.data = ddp_data_path + + +class WandbLogger(): + """Log training runs, datasets, models, and predictions to Weights & Biases. + + This logger sends information to W&B at wandb.ai. By default, this information + includes hyperparameters, system configuration and metrics, model metrics, + and basic data metrics and analyses. + + By providing additional command line arguments to train.py, datasets, + models and predictions can also be logged. + + For more on how this logger is used, see the Weights & Biases documentation: + https://docs.wandb.com/guides/integrations/yolov5 + """ + + def __init__(self, opt, run_id=None, job_type='Training'): + """ + - Initialize WandbLogger instance + - Upload dataset if opt.upload_dataset is True + - Setup training processes if job_type is 'Training' + + arguments: + opt (namespace) -- Commandline arguments for this run + run_id (str) -- Run ID of W&B run to be resumed + job_type (str) -- To set the job_type for this run + + """ + # Pre-training routine -- + self.job_type = job_type + self.wandb, self.wandb_run = wandb, None if not wandb else wandb.run + self.val_artifact, self.train_artifact = None, None + self.train_artifact_path, self.val_artifact_path = None, None + self.result_artifact = None + self.val_table, self.result_table = None, None + self.bbox_media_panel_images = [] + self.val_table_path_map = None + self.max_imgs_to_log = 16 + self.wandb_artifact_data_dict = None + self.data_dict = None + # It's more elegant to stick to 1 wandb.init call, + # but useful config data is overwritten in the WandbLogger's wandb.init call + if isinstance(opt.resume, str): # checks resume from artifact + if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + entity, project, run_id, model_artifact_name = get_run_info(opt.resume) + model_artifact_name = WANDB_ARTIFACT_PREFIX + model_artifact_name + assert wandb, 'install wandb to resume wandb runs' + # Resume wandb-artifact:// runs here| workaround for not overwriting wandb.config + self.wandb_run = wandb.init(id=run_id, + project=project, + entity=entity, + resume='allow', + allow_val_change=True) + opt.resume = model_artifact_name + elif self.wandb: + self.wandb_run = wandb.init(config=opt, + resume="allow", + project='YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem, + entity=opt.entity, + name=opt.name if opt.name != 'exp' else None, + job_type=job_type, + id=run_id, + allow_val_change=True) if not wandb.run else wandb.run + if self.wandb_run: + if self.job_type == 'Training': + if opt.upload_dataset: + if not opt.resume: + self.wandb_artifact_data_dict = self.check_and_upload_dataset(opt) + + if isinstance(opt.data, dict): + # This means another dataset manager has already processed the dataset info (e.g. ClearML) + # and they will have stored the already processed dict in opt.data + self.data_dict = opt.data + elif opt.resume: + # resume from artifact + if isinstance(opt.resume, str) and opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + self.data_dict = dict(self.wandb_run.config.data_dict) + else: # local resume + self.data_dict = check_wandb_dataset(opt.data) + else: + self.data_dict = check_wandb_dataset(opt.data) + self.wandb_artifact_data_dict = self.wandb_artifact_data_dict or self.data_dict + + # write data_dict to config. useful for resuming from artifacts. Do this only when not resuming. + self.wandb_run.config.update({'data_dict': self.wandb_artifact_data_dict}, allow_val_change=True) + self.setup_training(opt) + + if self.job_type == 'Dataset Creation': + self.wandb_run.config.update({"upload_dataset": True}) + self.data_dict = self.check_and_upload_dataset(opt) + + def check_and_upload_dataset(self, opt): + """ + Check if the dataset format is compatible and upload it as W&B artifact + + arguments: + opt (namespace)-- Commandline arguments for current run + + returns: + Updated dataset info dictionary where local dataset paths are replaced by WAND_ARFACT_PREFIX links. + """ + assert wandb, 'Install wandb to upload dataset' + config_path = self.log_dataset_artifact(opt.data, opt.single_cls, + 'YOLOv5' if opt.project == 'runs/train' else Path(opt.project).stem) + with open(config_path, errors='ignore') as f: + wandb_data_dict = yaml.safe_load(f) + return wandb_data_dict + + def setup_training(self, opt): + """ + Setup the necessary processes for training YOLO models: + - Attempt to download model checkpoint and dataset artifacts if opt.resume stats with WANDB_ARTIFACT_PREFIX + - Update data_dict, to contain info of previous run if resumed and the paths of dataset artifact if downloaded + - Setup log_dict, initialize bbox_interval + + arguments: + opt (namespace) -- commandline arguments for this run + + """ + self.log_dict, self.current_epoch = {}, 0 + self.bbox_interval = opt.bbox_interval + if isinstance(opt.resume, str): + modeldir, _ = self.download_model_artifact(opt) + if modeldir: + self.weights = Path(modeldir) / "last.pt" + config = self.wandb_run.config + opt.weights, opt.save_period, opt.batch_size, opt.bbox_interval, opt.epochs, opt.hyp, opt.imgsz = str( + self.weights), config.save_period, config.batch_size, config.bbox_interval, config.epochs,\ + config.hyp, config.imgsz + data_dict = self.data_dict + if self.val_artifact is None: # If --upload_dataset is set, use the existing artifact, don't download + self.train_artifact_path, self.train_artifact = self.download_dataset_artifact( + data_dict.get('train'), opt.artifact_alias) + self.val_artifact_path, self.val_artifact = self.download_dataset_artifact( + data_dict.get('val'), opt.artifact_alias) + + if self.train_artifact_path is not None: + train_path = Path(self.train_artifact_path) / 'data/images/' + data_dict['train'] = str(train_path) + if self.val_artifact_path is not None: + val_path = Path(self.val_artifact_path) / 'data/images/' + data_dict['val'] = str(val_path) + + if self.val_artifact is not None: + self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation") + columns = ["epoch", "id", "ground truth", "prediction"] + columns.extend(self.data_dict['names']) + self.result_table = wandb.Table(columns) + self.val_table = self.val_artifact.get("val") + if self.val_table_path_map is None: + self.map_val_table_path() + if opt.bbox_interval == -1: + self.bbox_interval = opt.bbox_interval = (opt.epochs // 10) if opt.epochs > 10 else 1 + if opt.evolve or opt.noplots: + self.bbox_interval = opt.bbox_interval = opt.epochs + 1 # disable bbox_interval + train_from_artifact = self.train_artifact_path is not None and self.val_artifact_path is not None + # Update the the data_dict to point to local artifacts dir + if train_from_artifact: + self.data_dict = data_dict + + def download_dataset_artifact(self, path, alias): + """ + download the model checkpoint artifact if the path starts with WANDB_ARTIFACT_PREFIX + + arguments: + path -- path of the dataset to be used for training + alias (str)-- alias of the artifact to be download/used for training + + returns: + (str, wandb.Artifact) -- path of the downladed dataset and it's corresponding artifact object if dataset + is found otherwise returns (None, None) + """ + if isinstance(path, str) and path.startswith(WANDB_ARTIFACT_PREFIX): + artifact_path = Path(remove_prefix(path, WANDB_ARTIFACT_PREFIX) + ":" + alias) + dataset_artifact = wandb.use_artifact(artifact_path.as_posix().replace("\\", "/")) + assert dataset_artifact is not None, "'Error: W&B dataset artifact doesn\'t exist'" + datadir = dataset_artifact.download() + return datadir, dataset_artifact + return None, None + + def download_model_artifact(self, opt): + """ + download the model checkpoint artifact if the resume path starts with WANDB_ARTIFACT_PREFIX + + arguments: + opt (namespace) -- Commandline arguments for this run + """ + if opt.resume.startswith(WANDB_ARTIFACT_PREFIX): + model_artifact = wandb.use_artifact(remove_prefix(opt.resume, WANDB_ARTIFACT_PREFIX) + ":latest") + assert model_artifact is not None, 'Error: W&B model artifact doesn\'t exist' + modeldir = model_artifact.download() + # epochs_trained = model_artifact.metadata.get('epochs_trained') + total_epochs = model_artifact.metadata.get('total_epochs') + is_finished = total_epochs is None + assert not is_finished, 'training is finished, can only resume incomplete runs.' + return modeldir, model_artifact + return None, None + + def log_model(self, path, opt, epoch, fitness_score, best_model=False): + """ + Log the model checkpoint as W&B artifact + + arguments: + path (Path) -- Path of directory containing the checkpoints + opt (namespace) -- Command line arguments for this run + epoch (int) -- Current epoch number + fitness_score (float) -- fitness score for current epoch + best_model (boolean) -- Boolean representing if the current checkpoint is the best yet. + """ + model_artifact = wandb.Artifact('run_' + wandb.run.id + '_model', + type='model', + metadata={ + 'original_url': str(path), + 'epochs_trained': epoch + 1, + 'save period': opt.save_period, + 'project': opt.project, + 'total_epochs': opt.epochs, + 'fitness_score': fitness_score}) + model_artifact.add_file(str(path / 'last.pt'), name='last.pt') + wandb.log_artifact(model_artifact, + aliases=['latest', 'last', 'epoch ' + str(self.current_epoch), 'best' if best_model else '']) + LOGGER.info(f"Saving model artifact on epoch {epoch + 1}") + + def log_dataset_artifact(self, data_file, single_cls, project, overwrite_config=False): + """ + Log the dataset as W&B artifact and return the new data file with W&B links + + arguments: + data_file (str) -- the .yaml file with information about the dataset like - path, classes etc. + single_class (boolean) -- train multi-class data as single-class + project (str) -- project name. Used to construct the artifact path + overwrite_config (boolean) -- overwrites the data.yaml file if set to true otherwise creates a new + file with _wandb postfix. Eg -> data_wandb.yaml + + returns: + the new .yaml file with artifact links. it can be used to start training directly from artifacts + """ + upload_dataset = self.wandb_run.config.upload_dataset + log_val_only = isinstance(upload_dataset, str) and upload_dataset == 'val' + self.data_dict = check_dataset(data_file) # parse and check + data = dict(self.data_dict) + nc, names = (1, ['item']) if single_cls else (int(data['nc']), data['names']) + names = {k: v for k, v in enumerate(names)} # to index dictionary + + # log train set + if not log_val_only: + self.train_artifact = self.create_dataset_table(LoadImagesAndLabels(data['train'], rect=True, batch_size=1), + names, + name='train') if data.get('train') else None + if data.get('train'): + data['train'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'train') + + self.val_artifact = self.create_dataset_table( + LoadImagesAndLabels(data['val'], rect=True, batch_size=1), names, name='val') if data.get('val') else None + if data.get('val'): + data['val'] = WANDB_ARTIFACT_PREFIX + str(Path(project) / 'val') + + path = Path(data_file) + # create a _wandb.yaml file with artifacts links if both train and test set are logged + if not log_val_only: + path = (path.stem if overwrite_config else path.stem + '_wandb') + '.yaml' # updated data.yaml path + path = ROOT / 'data' / path + data.pop('download', None) + data.pop('path', None) + with open(path, 'w') as f: + yaml.safe_dump(data, f) + LOGGER.info(f"Created dataset config file {path}") + + if self.job_type == 'Training': # builds correct artifact pipeline graph + if not log_val_only: + self.wandb_run.log_artifact( + self.train_artifact) # calling use_artifact downloads the dataset. NOT NEEDED! + self.wandb_run.use_artifact(self.val_artifact) + self.val_artifact.wait() + self.val_table = self.val_artifact.get('val') + self.map_val_table_path() + else: + self.wandb_run.log_artifact(self.train_artifact) + self.wandb_run.log_artifact(self.val_artifact) + return path + + def map_val_table_path(self): + """ + Map the validation dataset Table like name of file -> it's id in the W&B Table. + Useful for - referencing artifacts for evaluation. + """ + self.val_table_path_map = {} + LOGGER.info("Mapping dataset") + for i, data in enumerate(tqdm(self.val_table.data)): + self.val_table_path_map[data[3]] = data[0] + + def create_dataset_table(self, dataset: LoadImagesAndLabels, class_to_id: Dict[int, str], name: str = 'dataset'): + """ + Create and return W&B artifact containing W&B Table of the dataset. + + arguments: + dataset -- instance of LoadImagesAndLabels class used to iterate over the data to build Table + class_to_id -- hash map that maps class ids to labels + name -- name of the artifact + + returns: + dataset artifact to be logged or used + """ + # TODO: Explore multiprocessing to slpit this loop parallely| This is essential for speeding up the the logging + artifact = wandb.Artifact(name=name, type="dataset") + img_files = tqdm([dataset.path]) if isinstance(dataset.path, str) and Path(dataset.path).is_dir() else None + img_files = tqdm(dataset.im_files) if not img_files else img_files + for img_file in img_files: + if Path(img_file).is_dir(): + artifact.add_dir(img_file, name='data/images') + labels_path = 'labels'.join(dataset.path.rsplit('images', 1)) + artifact.add_dir(labels_path, name='data/labels') + else: + artifact.add_file(img_file, name='data/images/' + Path(img_file).name) + label_file = Path(img2label_paths([img_file])[0]) + artifact.add_file(str(label_file), name='data/labels/' + + label_file.name) if label_file.exists() else None + table = wandb.Table(columns=["id", "train_image", "Classes", "name"]) + class_set = wandb.Classes([{'id': id, 'name': name} for id, name in class_to_id.items()]) + for si, (img, labels, paths, shapes) in enumerate(tqdm(dataset)): + box_data, img_classes = [], {} + for cls, *xywh in labels[:, 1:].tolist(): + cls = int(cls) + box_data.append({ + "position": { + "middle": [xywh[0], xywh[1]], + "width": xywh[2], + "height": xywh[3]}, + "class_id": cls, + "box_caption": "%s" % (class_to_id[cls])}) + img_classes[cls] = class_to_id[cls] + boxes = {"ground_truth": {"box_data": box_data, "class_labels": class_to_id}} # inference-space + table.add_data(si, wandb.Image(paths, classes=class_set, boxes=boxes), list(img_classes.values()), + Path(paths).name) + artifact.add(table, name) + return artifact + + def log_training_progress(self, predn, path, names): + """ + Build evaluation Table. Uses reference from validation dataset table. + + arguments: + predn (list): list of predictions in the native space in the format - [xmin, ymin, xmax, ymax, confidence, class] + path (str): local path of the current evaluation image + names (dict(int, str)): hash map that maps class ids to labels + """ + class_set = wandb.Classes([{'id': id, 'name': name} for id, name in names.items()]) + box_data = [] + avg_conf_per_class = [0] * len(self.data_dict['names']) + pred_class_count = {} + for *xyxy, conf, cls in predn.tolist(): + if conf >= 0.25: + cls = int(cls) + box_data.append({ + "position": { + "minX": xyxy[0], + "minY": xyxy[1], + "maxX": xyxy[2], + "maxY": xyxy[3]}, + "class_id": cls, + "box_caption": f"{names[cls]} {conf:.3f}", + "scores": { + "class_score": conf}, + "domain": "pixel"}) + avg_conf_per_class[cls] += conf + + if cls in pred_class_count: + pred_class_count[cls] += 1 + else: + pred_class_count[cls] = 1 + + for pred_class in pred_class_count.keys(): + avg_conf_per_class[pred_class] = avg_conf_per_class[pred_class] / pred_class_count[pred_class] + + boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space + id = self.val_table_path_map[Path(path).name] + self.result_table.add_data(self.current_epoch, id, self.val_table.data[id][1], + wandb.Image(self.val_table.data[id][1], boxes=boxes, classes=class_set), + *avg_conf_per_class) + + def val_one_image(self, pred, predn, path, names, im): + """ + Log validation data for one image. updates the result Table if validation dataset is uploaded and log bbox media panel + + arguments: + pred (list): list of scaled predictions in the format - [xmin, ymin, xmax, ymax, confidence, class] + predn (list): list of predictions in the native space - [xmin, ymin, xmax, ymax, confidence, class] + path (str): local path of the current evaluation image + """ + if self.val_table and self.result_table: # Log Table if Val dataset is uploaded as artifact + self.log_training_progress(predn, path, names) + + if len(self.bbox_media_panel_images) < self.max_imgs_to_log and self.current_epoch > 0: + if self.current_epoch % self.bbox_interval == 0: + box_data = [{ + "position": { + "minX": xyxy[0], + "minY": xyxy[1], + "maxX": xyxy[2], + "maxY": xyxy[3]}, + "class_id": int(cls), + "box_caption": f"{names[int(cls)]} {conf:.3f}", + "scores": { + "class_score": conf}, + "domain": "pixel"} for *xyxy, conf, cls in pred.tolist()] + boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space + self.bbox_media_panel_images.append(wandb.Image(im, boxes=boxes, caption=path.name)) + + def log(self, log_dict): + """ + save the metrics to the logging dictionary + + arguments: + log_dict (Dict) -- metrics/media to be logged in current step + """ + if self.wandb_run: + for key, value in log_dict.items(): + self.log_dict[key] = value + + def end_epoch(self, best_result=False): + """ + commit the log_dict, model artifacts and Tables to W&B and flush the log_dict. + + arguments: + best_result (boolean): Boolean representing if the result of this evaluation is best or not + """ + if self.wandb_run: + with all_logging_disabled(): + if self.bbox_media_panel_images: + self.log_dict["BoundingBoxDebugger"] = self.bbox_media_panel_images + try: + wandb.log(self.log_dict) + except BaseException as e: + LOGGER.info( + f"An error occurred in wandb logger. The training will proceed without interruption. More info\n{e}" + ) + self.wandb_run.finish() + self.wandb_run = None + + self.log_dict = {} + self.bbox_media_panel_images = [] + if self.result_artifact: + self.result_artifact.add(self.result_table, 'result') + wandb.log_artifact(self.result_artifact, + aliases=[ + 'latest', 'last', 'epoch ' + str(self.current_epoch), + ('best' if best_result else '')]) + + wandb.log({"evaluation": self.result_table}) + columns = ["epoch", "id", "ground truth", "prediction"] + columns.extend(self.data_dict['names']) + self.result_table = wandb.Table(columns) + self.result_artifact = wandb.Artifact("run_" + wandb.run.id + "_progress", "evaluation") + + def finish_run(self): + """ + Log metrics if any and finish the current W&B run + """ + if self.wandb_run: + if self.log_dict: + with all_logging_disabled(): + wandb.log(self.log_dict) + wandb.run.finish() + + +@contextmanager +def all_logging_disabled(highest_level=logging.CRITICAL): + """ source - https://gist.github.com/simon-weber/7853144 + A context manager that will prevent any logging messages triggered during the body from being processed. + :param highest_level: the maximum logging level in use. + This would only need to be changed if a custom level greater than CRITICAL is defined. + """ + previous_level = logging.root.manager.disable + logging.disable(highest_level) + try: + yield + finally: + logging.disable(previous_level) diff --git a/application/yolov5_example/utils/loss.py b/application/yolov5_example/utils/loss.py new file mode 100644 index 00000000..9b9c3d9f --- /dev/null +++ b/application/yolov5_example/utils/loss.py @@ -0,0 +1,234 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Loss functions +""" + +import torch +import torch.nn as nn + +from utils.metrics import bbox_iou +from utils.torch_utils import de_parallel + + +def smooth_BCE(eps=0.1): # https://github.com/ultralytics/yolov3/issues/238#issuecomment-598028441 + # return positive, negative label smoothing BCE targets + return 1.0 - 0.5 * eps, 0.5 * eps + + +class BCEBlurWithLogitsLoss(nn.Module): + # BCEwithLogitLoss() with reduced missing label effects. + def __init__(self, alpha=0.05): + super().__init__() + self.loss_fcn = nn.BCEWithLogitsLoss(reduction='none') # must be nn.BCEWithLogitsLoss() + self.alpha = alpha + + def forward(self, pred, true): + loss = self.loss_fcn(pred, true) + pred = torch.sigmoid(pred) # prob from logits + dx = pred - true # reduce only missing label effects + # dx = (pred - true).abs() # reduce missing label and false label effects + alpha_factor = 1 - torch.exp((dx - 1) / (self.alpha + 1e-4)) + loss *= alpha_factor + return loss.mean() + + +class FocalLoss(nn.Module): + # Wraps focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) + def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): + super().__init__() + self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() + self.gamma = gamma + self.alpha = alpha + self.reduction = loss_fcn.reduction + self.loss_fcn.reduction = 'none' # required to apply FL to each element + + def forward(self, pred, true): + loss = self.loss_fcn(pred, true) + # p_t = torch.exp(-loss) + # loss *= self.alpha * (1.000001 - p_t) ** self.gamma # non-zero power for gradient stability + + # TF implementation https://github.com/tensorflow/addons/blob/v0.7.1/tensorflow_addons/losses/focal_loss.py + pred_prob = torch.sigmoid(pred) # prob from logits + p_t = true * pred_prob + (1 - true) * (1 - pred_prob) + alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) + modulating_factor = (1.0 - p_t) ** self.gamma + loss *= alpha_factor * modulating_factor + + if self.reduction == 'mean': + return loss.mean() + elif self.reduction == 'sum': + return loss.sum() + else: # 'none' + return loss + + +class QFocalLoss(nn.Module): + # Wraps Quality focal loss around existing loss_fcn(), i.e. criteria = FocalLoss(nn.BCEWithLogitsLoss(), gamma=1.5) + def __init__(self, loss_fcn, gamma=1.5, alpha=0.25): + super().__init__() + self.loss_fcn = loss_fcn # must be nn.BCEWithLogitsLoss() + self.gamma = gamma + self.alpha = alpha + self.reduction = loss_fcn.reduction + self.loss_fcn.reduction = 'none' # required to apply FL to each element + + def forward(self, pred, true): + loss = self.loss_fcn(pred, true) + + pred_prob = torch.sigmoid(pred) # prob from logits + alpha_factor = true * self.alpha + (1 - true) * (1 - self.alpha) + modulating_factor = torch.abs(true - pred_prob) ** self.gamma + loss *= alpha_factor * modulating_factor + + if self.reduction == 'mean': + return loss.mean() + elif self.reduction == 'sum': + return loss.sum() + else: # 'none' + return loss + + +class ComputeLoss: + sort_obj_iou = False + + # Compute losses + def __init__(self, model, autobalance=False): + device = next(model.parameters()).device # get model device + h = model.hyp # hyperparameters + + # Define criteria + BCEcls = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['cls_pw']], device=device)) + BCEobj = nn.BCEWithLogitsLoss(pos_weight=torch.tensor([h['obj_pw']], device=device)) + + # Class label smoothing https://arxiv.org/pdf/1902.04103.pdf eqn 3 + self.cp, self.cn = smooth_BCE(eps=h.get('label_smoothing', 0.0)) # positive, negative BCE targets + + # Focal loss + g = h['fl_gamma'] # focal loss gamma + if g > 0: + BCEcls, BCEobj = FocalLoss(BCEcls, g), FocalLoss(BCEobj, g) + + m = de_parallel(model).model[-1] # Detect() module + self.balance = {3: [4.0, 1.0, 0.4]}.get(m.nl, [4.0, 1.0, 0.25, 0.06, 0.02]) # P3-P7 + self.ssi = list(m.stride).index(16) if autobalance else 0 # stride 16 index + self.BCEcls, self.BCEobj, self.gr, self.hyp, self.autobalance = BCEcls, BCEobj, 1.0, h, autobalance + self.na = m.na # number of anchors + self.nc = m.nc # number of classes + self.nl = m.nl # number of layers + self.anchors = m.anchors + self.device = device + + def __call__(self, p, targets): # predictions, targets + lcls = torch.zeros(1, device=self.device) # class loss + lbox = torch.zeros(1, device=self.device) # box loss + lobj = torch.zeros(1, device=self.device) # object loss + tcls, tbox, indices, anchors = self.build_targets(p, targets) # targets + + # Losses + for i, pi in enumerate(p): # layer index, layer predictions + b, a, gj, gi = indices[i] # image, anchor, gridy, gridx + tobj = torch.zeros(pi.shape[:4], dtype=pi.dtype, device=self.device) # target obj + + n = b.shape[0] # number of targets + if n: + # pxy, pwh, _, pcls = pi[b, a, gj, gi].tensor_split((2, 4, 5), dim=1) # faster, requires torch 1.8.0 + pxy, pwh, _, pcls = pi[b, a, gj, gi].split((2, 2, 1, self.nc), 1) # target-subset of predictions + + # Regression + pxy = pxy.sigmoid() * 2 - 0.5 + pwh = (pwh.sigmoid() * 2) ** 2 * anchors[i] + pbox = torch.cat((pxy, pwh), 1) # predicted box + iou = bbox_iou(pbox, tbox[i], CIoU=True).squeeze() # iou(prediction, target) + lbox += (1.0 - iou).mean() # iou loss + + # Objectness + iou = iou.detach().clamp(0).type(tobj.dtype) + if self.sort_obj_iou: + j = iou.argsort() + b, a, gj, gi, iou = b[j], a[j], gj[j], gi[j], iou[j] + if self.gr < 1: + iou = (1.0 - self.gr) + self.gr * iou + tobj[b, a, gj, gi] = iou # iou ratio + + # Classification + if self.nc > 1: # cls loss (only if multiple classes) + t = torch.full_like(pcls, self.cn, device=self.device) # targets + t[range(n), tcls[i]] = self.cp + lcls += self.BCEcls(pcls, t) # BCE + + # Append targets to text file + # with open('targets.txt', 'a') as file: + # [file.write('%11.5g ' * 4 % tuple(x) + '\n') for x in torch.cat((txy[i], twh[i]), 1)] + + obji = self.BCEobj(pi[..., 4], tobj) + lobj += obji * self.balance[i] # obj loss + if self.autobalance: + self.balance[i] = self.balance[i] * 0.9999 + 0.0001 / obji.detach().item() + + if self.autobalance: + self.balance = [x / self.balance[self.ssi] for x in self.balance] + lbox *= self.hyp['box'] + lobj *= self.hyp['obj'] + lcls *= self.hyp['cls'] + bs = tobj.shape[0] # batch size + + return (lbox + lobj + lcls) * bs, torch.cat((lbox, lobj, lcls)).detach() + + def build_targets(self, p, targets): + # Build targets for compute_loss(), input targets(image,class,x,y,w,h) + na, nt = self.na, targets.shape[0] # number of anchors, targets + tcls, tbox, indices, anch = [], [], [], [] + gain = torch.ones(7, device=self.device) # normalized to gridspace gain + ai = torch.arange(na, device=self.device).float().view(na, 1).repeat(1, nt) # same as .repeat_interleave(nt) + targets = torch.cat((targets.repeat(na, 1, 1), ai[..., None]), 2) # append anchor indices + + g = 0.5 # bias + off = torch.tensor( + [ + [0, 0], + [1, 0], + [0, 1], + [-1, 0], + [0, -1], # j,k,l,m + # [1, 1], [1, -1], [-1, 1], [-1, -1], # jk,jm,lk,lm + ], + device=self.device).float() * g # offsets + + for i in range(self.nl): + anchors, shape = self.anchors[i], p[i].shape + gain[2:6] = torch.tensor(shape)[[3, 2, 3, 2]] # xyxy gain + + # Match targets to anchors + t = targets * gain # shape(3,n,7) + if nt: + # Matches + r = t[..., 4:6] / anchors[:, None] # wh ratio + j = torch.max(r, 1 / r).max(2)[0] < self.hyp['anchor_t'] # compare + # j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2)) + t = t[j] # filter + + # Offsets + gxy = t[:, 2:4] # grid xy + gxi = gain[[2, 3]] - gxy # inverse + j, k = ((gxy % 1 < g) & (gxy > 1)).T + l, m = ((gxi % 1 < g) & (gxi > 1)).T + j = torch.stack((torch.ones_like(j), j, k, l, m)) + t = t.repeat((5, 1, 1))[j] + offsets = (torch.zeros_like(gxy)[None] + off[:, None])[j] + else: + t = targets[0] + offsets = 0 + + # Define + bc, gxy, gwh, a = t.chunk(4, 1) # (image, class), grid xy, grid wh, anchors + a, (b, c) = a.long().view(-1), bc.long().T # anchors, image, class + gij = (gxy - offsets).long() + gi, gj = gij.T # grid indices + + # Append + indices.append((b, a, gj.clamp_(0, shape[2] - 1), gi.clamp_(0, shape[3] - 1))) # image, anchor, grid + tbox.append(torch.cat((gxy - gij, gwh), 1)) # box + anch.append(anchors[a]) # anchors + tcls.append(c) # class + + return tcls, tbox, indices, anch diff --git a/application/yolov5_example/utils/metrics.py b/application/yolov5_example/utils/metrics.py new file mode 100644 index 00000000..08880cd3 --- /dev/null +++ b/application/yolov5_example/utils/metrics.py @@ -0,0 +1,364 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Model validation metrics +""" + +import math +import warnings +from pathlib import Path + +import matplotlib.pyplot as plt +import numpy as np +import torch + + +def fitness(x): + # Model fitness as a weighted combination of metrics + w = [0.0, 0.0, 0.1, 0.9] # weights for [P, R, mAP@0.5, mAP@0.5:0.95] + return (x[:, :4] * w).sum(1) + + +def smooth(y, f=0.05): + # Box filter of fraction f + nf = round(len(y) * f * 2) // 2 + 1 # number of filter elements (must be odd) + p = np.ones(nf // 2) # ones padding + yp = np.concatenate((p * y[0], y, p * y[-1]), 0) # y padded + return np.convolve(yp, np.ones(nf) / nf, mode='valid') # y-smoothed + + +def ap_per_class(tp, conf, pred_cls, target_cls, plot=False, save_dir='.', names=(), eps=1e-16): + """ Compute the average precision, given the recall and precision curves. + Source: https://github.com/rafaelpadilla/Object-Detection-Metrics. + # Arguments + tp: True positives (nparray, nx1 or nx10). + conf: Objectness value from 0-1 (nparray). + pred_cls: Predicted object classes (nparray). + target_cls: True object classes (nparray). + plot: Plot precision-recall curve at mAP@0.5 + save_dir: Plot save directory + # Returns + The average precision as computed in py-faster-rcnn. + """ + + # Sort by objectness + i = np.argsort(-conf) + tp, conf, pred_cls = tp[i], conf[i], pred_cls[i] + + # Find unique classes + unique_classes, nt = np.unique(target_cls, return_counts=True) + nc = unique_classes.shape[0] # number of classes, number of detections + + # Create Precision-Recall curve and compute AP for each class + px, py = np.linspace(0, 1, 1000), [] # for plotting + ap, p, r = np.zeros((nc, tp.shape[1])), np.zeros((nc, 1000)), np.zeros((nc, 1000)) + for ci, c in enumerate(unique_classes): + i = pred_cls == c + n_l = nt[ci] # number of labels + n_p = i.sum() # number of predictions + if n_p == 0 or n_l == 0: + continue + + # Accumulate FPs and TPs + fpc = (1 - tp[i]).cumsum(0) + tpc = tp[i].cumsum(0) + + # Recall + recall = tpc / (n_l + eps) # recall curve + r[ci] = np.interp(-px, -conf[i], recall[:, 0], left=0) # negative x, xp because xp decreases + + # Precision + precision = tpc / (tpc + fpc) # precision curve + p[ci] = np.interp(-px, -conf[i], precision[:, 0], left=1) # p at pr_score + + # AP from recall-precision curve + for j in range(tp.shape[1]): + ap[ci, j], mpre, mrec = compute_ap(recall[:, j], precision[:, j]) + if plot and j == 0: + py.append(np.interp(px, mrec, mpre)) # precision at mAP@0.5 + + # Compute F1 (harmonic mean of precision and recall) + f1 = 2 * p * r / (p + r + eps) + names = [v for k, v in names.items() if k in unique_classes] # list: only classes that have data + names = dict(enumerate(names)) # to dict + if plot: + plot_pr_curve(px, py, ap, Path(save_dir) / 'PR_curve.png', names) + plot_mc_curve(px, f1, Path(save_dir) / 'F1_curve.png', names, ylabel='F1') + plot_mc_curve(px, p, Path(save_dir) / 'P_curve.png', names, ylabel='Precision') + plot_mc_curve(px, r, Path(save_dir) / 'R_curve.png', names, ylabel='Recall') + + i = smooth(f1.mean(0), 0.1).argmax() # max F1 index + p, r, f1 = p[:, i], r[:, i], f1[:, i] + tp = (r * nt).round() # true positives + fp = (tp / (p + eps) - tp).round() # false positives + return tp, fp, p, r, f1, ap, unique_classes.astype(int) + + +def compute_ap(recall, precision): + """ Compute the average precision, given the recall and precision curves + # Arguments + recall: The recall curve (list) + precision: The precision curve (list) + # Returns + Average precision, precision curve, recall curve + """ + + # Append sentinel values to beginning and end + mrec = np.concatenate(([0.0], recall, [1.0])) + mpre = np.concatenate(([1.0], precision, [0.0])) + + # Compute the precision envelope + mpre = np.flip(np.maximum.accumulate(np.flip(mpre))) + + # Integrate area under curve + method = 'interp' # methods: 'continuous', 'interp' + if method == 'interp': + x = np.linspace(0, 1, 101) # 101-point interp (COCO) + ap = np.trapz(np.interp(x, mrec, mpre), x) # integrate + else: # 'continuous' + i = np.where(mrec[1:] != mrec[:-1])[0] # points where x axis (recall) changes + ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) # area under curve + + return ap, mpre, mrec + + +class ConfusionMatrix: + # Updated version of https://github.com/kaanakan/object_detection_confusion_matrix + def __init__(self, nc, conf=0.25, iou_thres=0.45): + self.matrix = np.zeros((nc + 1, nc + 1)) + self.nc = nc # number of classes + self.conf = conf + self.iou_thres = iou_thres + + def process_batch(self, detections, labels): + """ + Return intersection-over-union (Jaccard index) of boxes. + Both sets of boxes are expected to be in (x1, y1, x2, y2) format. + Arguments: + detections (Array[N, 6]), x1, y1, x2, y2, conf, class + labels (Array[M, 5]), class, x1, y1, x2, y2 + Returns: + None, updates confusion matrix accordingly + """ + if detections is None: + gt_classes = labels.int() + for i, gc in enumerate(gt_classes): + self.matrix[self.nc, gc] += 1 # background FN + return + + detections = detections[detections[:, 4] > self.conf] + gt_classes = labels[:, 0].int() + detection_classes = detections[:, 5].int() + iou = box_iou(labels[:, 1:], detections[:, :4]) + + x = torch.where(iou > self.iou_thres) + if x[0].shape[0]: + matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() + if x[0].shape[0] > 1: + matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 1], return_index=True)[1]] + matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 0], return_index=True)[1]] + else: + matches = np.zeros((0, 3)) + + n = matches.shape[0] > 0 + m0, m1, _ = matches.transpose().astype(int) + for i, gc in enumerate(gt_classes): + j = m0 == i + if n and sum(j) == 1: + self.matrix[detection_classes[m1[j]], gc] += 1 # correct + else: + self.matrix[self.nc, gc] += 1 # background FP + + if n: + for i, dc in enumerate(detection_classes): + if not any(m1 == i): + self.matrix[dc, self.nc] += 1 # background FN + + def matrix(self): + return self.matrix + + def tp_fp(self): + tp = self.matrix.diagonal() # true positives + fp = self.matrix.sum(1) - tp # false positives + # fn = self.matrix.sum(0) - tp # false negatives (missed detections) + return tp[:-1], fp[:-1] # remove background class + + def plot(self, normalize=True, save_dir='', names=()): + try: + import seaborn as sn + + array = self.matrix / ((self.matrix.sum(0).reshape(1, -1) + 1E-9) if normalize else 1) # normalize columns + array[array < 0.005] = np.nan # don't annotate (would appear as 0.00) + + fig = plt.figure(figsize=(12, 9), tight_layout=True) + nc, nn = self.nc, len(names) # number of classes, names + sn.set(font_scale=1.0 if nc < 50 else 0.8) # for label size + labels = (0 < nn < 99) and (nn == nc) # apply names to ticklabels + with warnings.catch_warnings(): + warnings.simplefilter('ignore') # suppress empty matrix RuntimeWarning: All-NaN slice encountered + sn.heatmap(array, + annot=nc < 30, + annot_kws={ + "size": 8}, + cmap='Blues', + fmt='.2f', + square=True, + vmin=0.0, + xticklabels=names + ['background FP'] if labels else "auto", + yticklabels=names + ['background FN'] if labels else "auto").set_facecolor((1, 1, 1)) + fig.axes[0].set_xlabel('True') + fig.axes[0].set_ylabel('Predicted') + plt.title('Confusion Matrix') + fig.savefig(Path(save_dir) / 'confusion_matrix.png', dpi=250) + plt.close() + except Exception as e: + print(f'WARNING: ConfusionMatrix plot failure: {e}') + + def print(self): + for i in range(self.nc + 1): + print(' '.join(map(str, self.matrix[i]))) + + +def bbox_iou(box1, box2, xywh=True, GIoU=False, DIoU=False, CIoU=False, eps=1e-7): + # Returns Intersection over Union (IoU) of box1(1,4) to box2(n,4) + + # Get the coordinates of bounding boxes + if xywh: # transform from xywh to xyxy + (x1, y1, w1, h1), (x2, y2, w2, h2) = box1.chunk(4, 1), box2.chunk(4, 1) + w1_, h1_, w2_, h2_ = w1 / 2, h1 / 2, w2 / 2, h2 / 2 + b1_x1, b1_x2, b1_y1, b1_y2 = x1 - w1_, x1 + w1_, y1 - h1_, y1 + h1_ + b2_x1, b2_x2, b2_y1, b2_y2 = x2 - w2_, x2 + w2_, y2 - h2_, y2 + h2_ + else: # x1, y1, x2, y2 = box1 + b1_x1, b1_y1, b1_x2, b1_y2 = box1.chunk(4, 1) + b2_x1, b2_y1, b2_x2, b2_y2 = box2.chunk(4, 1) + w1, h1 = b1_x2 - b1_x1, b1_y2 - b1_y1 + w2, h2 = b2_x2 - b2_x1, b2_y2 - b2_y1 + + # Intersection area + inter = (torch.min(b1_x2, b2_x2) - torch.max(b1_x1, b2_x1)).clamp(0) * \ + (torch.min(b1_y2, b2_y2) - torch.max(b1_y1, b2_y1)).clamp(0) + + # Union Area + union = w1 * h1 + w2 * h2 - inter + eps + + # IoU + iou = inter / union + if CIoU or DIoU or GIoU: + cw = torch.max(b1_x2, b2_x2) - torch.min(b1_x1, b2_x1) # convex (smallest enclosing box) width + ch = torch.max(b1_y2, b2_y2) - torch.min(b1_y1, b2_y1) # convex height + if CIoU or DIoU: # Distance or Complete IoU https://arxiv.org/abs/1911.08287v1 + c2 = cw ** 2 + ch ** 2 + eps # convex diagonal squared + rho2 = ((b2_x1 + b2_x2 - b1_x1 - b1_x2) ** 2 + (b2_y1 + b2_y2 - b1_y1 - b1_y2) ** 2) / 4 # center dist ** 2 + if CIoU: # https://github.com/Zzh-tju/DIoU-SSD-pytorch/blob/master/utils/box/box_utils.py#L47 + v = (4 / math.pi ** 2) * torch.pow(torch.atan(w2 / (h2 + eps)) - torch.atan(w1 / (h1 + eps)), 2) + with torch.no_grad(): + alpha = v / (v - iou + (1 + eps)) + return iou - (rho2 / c2 + v * alpha) # CIoU + return iou - rho2 / c2 # DIoU + c_area = cw * ch + eps # convex area + return iou - (c_area - union) / c_area # GIoU https://arxiv.org/pdf/1902.09630.pdf + return iou # IoU + + +def box_area(box): + # box = xyxy(4,n) + return (box[2] - box[0]) * (box[3] - box[1]) + + +def box_iou(box1, box2, eps=1e-7): + # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py + """ + Return intersection-over-union (Jaccard index) of boxes. + Both sets of boxes are expected to be in (x1, y1, x2, y2) format. + Arguments: + box1 (Tensor[N, 4]) + box2 (Tensor[M, 4]) + Returns: + iou (Tensor[N, M]): the NxM matrix containing the pairwise + IoU values for every element in boxes1 and boxes2 + """ + + # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2) + (a1, a2), (b1, b2) = box1[:, None].chunk(2, 2), box2.chunk(2, 1) + inter = (torch.min(a2, b2) - torch.max(a1, b1)).clamp(0).prod(2) + + # IoU = inter / (area1 + area2 - inter) + return inter / (box_area(box1.T)[:, None] + box_area(box2.T) - inter + eps) + + +def bbox_ioa(box1, box2, eps=1e-7): + """ Returns the intersection over box2 area given box1, box2. Boxes are x1y1x2y2 + box1: np.array of shape(4) + box2: np.array of shape(nx4) + returns: np.array of shape(n) + """ + + # Get the coordinates of bounding boxes + b1_x1, b1_y1, b1_x2, b1_y2 = box1 + b2_x1, b2_y1, b2_x2, b2_y2 = box2.T + + # Intersection area + inter_area = (np.minimum(b1_x2, b2_x2) - np.maximum(b1_x1, b2_x1)).clip(0) * \ + (np.minimum(b1_y2, b2_y2) - np.maximum(b1_y1, b2_y1)).clip(0) + + # box2 area + box2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1) + eps + + # Intersection over box2 area + return inter_area / box2_area + + +def wh_iou(wh1, wh2, eps=1e-7): + # Returns the nxm IoU matrix. wh1 is nx2, wh2 is mx2 + wh1 = wh1[:, None] # [N,1,2] + wh2 = wh2[None] # [1,M,2] + inter = torch.min(wh1, wh2).prod(2) # [N,M] + return inter / (wh1.prod(2) + wh2.prod(2) - inter + eps) # iou = inter / (area1 + area2 - inter) + + +# Plots ---------------------------------------------------------------------------------------------------------------- + + +def plot_pr_curve(px, py, ap, save_dir=Path('pr_curve.png'), names=()): + # Precision-recall curve + fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) + py = np.stack(py, axis=1) + + if 0 < len(names) < 21: # display per-class legend if < 21 classes + for i, y in enumerate(py.T): + ax.plot(px, y, linewidth=1, label=f'{names[i]} {ap[i, 0]:.3f}') # plot(recall, precision) + else: + ax.plot(px, py, linewidth=1, color='grey') # plot(recall, precision) + + ax.plot(px, py.mean(1), linewidth=3, color='blue', label='all classes %.3f mAP@0.5' % ap[:, 0].mean()) + ax.set_xlabel('Recall') + ax.set_ylabel('Precision') + ax.set_xlim(0, 1) + ax.set_ylim(0, 1) + plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") + plt.title('Precision-Recall Curve') + fig.savefig(save_dir, dpi=250) + plt.close() + + +def plot_mc_curve(px, py, save_dir=Path('mc_curve.png'), names=(), xlabel='Confidence', ylabel='Metric'): + # Metric-confidence curve + fig, ax = plt.subplots(1, 1, figsize=(9, 6), tight_layout=True) + + if 0 < len(names) < 21: # display per-class legend if < 21 classes + for i, y in enumerate(py): + ax.plot(px, y, linewidth=1, label=f'{names[i]}') # plot(confidence, metric) + else: + ax.plot(px, py.T, linewidth=1, color='grey') # plot(confidence, metric) + + y = smooth(py.mean(0), 0.05) + ax.plot(px, y, linewidth=3, color='blue', label=f'all classes {y.max():.2f} at {px[y.argmax()]:.3f}') + ax.set_xlabel(xlabel) + ax.set_ylabel(ylabel) + ax.set_xlim(0, 1) + ax.set_ylim(0, 1) + plt.legend(bbox_to_anchor=(1.04, 1), loc="upper left") + plt.title(f'{ylabel}-Confidence Curve') + fig.savefig(save_dir, dpi=250) + plt.close() diff --git a/application/yolov5_example/utils/plots.py b/application/yolov5_example/utils/plots.py new file mode 100644 index 00000000..7417308c --- /dev/null +++ b/application/yolov5_example/utils/plots.py @@ -0,0 +1,519 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Plotting utils +""" + +import math +import os +from copy import copy +from pathlib import Path +from urllib.error import URLError + +import cv2 +import matplotlib +import matplotlib.pyplot as plt +import numpy as np +import pandas as pd +import seaborn as sn +import torch +from PIL import Image, ImageDraw, ImageFont + +from utils.general import (CONFIG_DIR, FONT, LOGGER, Timeout, check_font, check_requirements, clip_coords, + increment_path, is_ascii, threaded, try_except, xywh2xyxy, xyxy2xywh) +from utils.metrics import fitness + +# Settings +RANK = int(os.getenv('RANK', -1)) +matplotlib.rc('font', **{'size': 11}) +matplotlib.use('Agg') # for writing to files only + + +class Colors: + # Ultralytics color palette https://ultralytics.com/ + def __init__(self): + # hex = matplotlib.colors.TABLEAU_COLORS.values() + hexs = ('FF3838', 'FF9D97', 'FF701F', 'FFB21D', 'CFD231', '48F90A', '92CC17', '3DDB86', '1A9334', '00D4BB', + '2C99A8', '00C2FF', '344593', '6473FF', '0018EC', '8438FF', '520085', 'CB38FF', 'FF95C8', 'FF37C7') + self.palette = [self.hex2rgb(f'#{c}') for c in hexs] + self.n = len(self.palette) + + def __call__(self, i, bgr=False): + c = self.palette[int(i) % self.n] + return (c[2], c[1], c[0]) if bgr else c + + @staticmethod + def hex2rgb(h): # rgb order (PIL) + return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4)) + + +colors = Colors() # create instance for 'from utils.plots import colors' + + +def check_pil_font(font=FONT, size=10): + # Return a PIL TrueType Font, downloading to CONFIG_DIR if necessary + font = Path(font) + font = font if font.exists() else (CONFIG_DIR / font.name) + try: + return ImageFont.truetype(str(font) if font.exists() else font.name, size) + except Exception: # download if missing + try: + check_font(font) + return ImageFont.truetype(str(font), size) + except TypeError: + check_requirements('Pillow>=8.4.0') # known issue https://github.com/ultralytics/yolov5/issues/5374 + except URLError: # not online + return ImageFont.load_default() + + +class Annotator: + # YOLOv5 Annotator for train/val mosaics and jpgs and detect/hub inference annotations + def __init__(self, im, line_width=None, font_size=None, font='Arial.ttf', pil=False, example='abc'): + assert im.data.contiguous, 'Image not contiguous. Apply np.ascontiguousarray(im) to Annotator() input images.' + non_ascii = not is_ascii(example) # non-latin labels, i.e. asian, arabic, cyrillic + self.pil = pil or non_ascii + if self.pil: # use PIL + self.im = im if isinstance(im, Image.Image) else Image.fromarray(im) + self.draw = ImageDraw.Draw(self.im) + self.font = check_pil_font(font='Arial.Unicode.ttf' if non_ascii else font, + size=font_size or max(round(sum(self.im.size) / 2 * 0.035), 12)) + else: # use cv2 + self.im = im + self.lw = line_width or max(round(sum(im.shape) / 2 * 0.003), 2) # line width + + def box_label(self, box, label='', color=(128, 128, 128), txt_color=(255, 255, 255)): + # Add one xyxy box to image with label + if self.pil or not is_ascii(label): + self.draw.rectangle(box, width=self.lw, outline=color) # box + if label: + w, h = self.font.getsize(label) # text width, height + outside = box[1] - h >= 0 # label fits outside box + self.draw.rectangle( + (box[0], box[1] - h if outside else box[1], box[0] + w + 1, + box[1] + 1 if outside else box[1] + h + 1), + fill=color, + ) + # self.draw.text((box[0], box[1]), label, fill=txt_color, font=self.font, anchor='ls') # for PIL>8.0 + self.draw.text((box[0], box[1] - h if outside else box[1]), label, fill=txt_color, font=self.font) + else: # cv2 + p1, p2 = (int(box[0]), int(box[1])), (int(box[2]), int(box[3])) + cv2.rectangle(self.im, p1, p2, color, thickness=self.lw, lineType=cv2.LINE_AA) + if label: + tf = max(self.lw - 1, 1) # font thickness + w, h = cv2.getTextSize(label, 0, fontScale=self.lw / 3, thickness=tf)[0] # text width, height + outside = p1[1] - h >= 3 + p2 = p1[0] + w, p1[1] - h - 3 if outside else p1[1] + h + 3 + cv2.rectangle(self.im, p1, p2, color, -1, cv2.LINE_AA) # filled + cv2.putText(self.im, + label, (p1[0], p1[1] - 2 if outside else p1[1] + h + 2), + 0, + self.lw / 3, + txt_color, + thickness=tf, + lineType=cv2.LINE_AA) + + def rectangle(self, xy, fill=None, outline=None, width=1): + # Add rectangle to image (PIL-only) + self.draw.rectangle(xy, fill, outline, width) + + def text(self, xy, text, txt_color=(255, 255, 255)): + # Add text to image (PIL-only) + w, h = self.font.getsize(text) # text width, height + self.draw.text((xy[0], xy[1] - h + 1), text, fill=txt_color, font=self.font) + + def result(self): + # Return annotated image as array + return np.asarray(self.im) + + +def feature_visualization(x, module_type, stage, n=32, save_dir=Path('runs/detect/exp')): + """ + x: Features to be visualized + module_type: Module type + stage: Module stage within model + n: Maximum number of feature maps to plot + save_dir: Directory to save results + """ + if 'Detect' not in module_type: + batch, channels, height, width = x.shape # batch, channels, height, width + if height > 1 and width > 1: + f = save_dir / f"stage{stage}_{module_type.split('.')[-1]}_features.png" # filename + + blocks = torch.chunk(x[0].cpu(), channels, dim=0) # select batch index 0, block by channels + n = min(n, channels) # number of plots + fig, ax = plt.subplots(math.ceil(n / 8), 8, tight_layout=True) # 8 rows x n/8 cols + ax = ax.ravel() + plt.subplots_adjust(wspace=0.05, hspace=0.05) + for i in range(n): + ax[i].imshow(blocks[i].squeeze()) # cmap='gray' + ax[i].axis('off') + + LOGGER.info(f'Saving {f}... ({n}/{channels})') + plt.title('Features') + plt.savefig(f, dpi=300, bbox_inches='tight') + plt.close() + np.save(str(f.with_suffix('.npy')), x[0].cpu().numpy()) # npy save + + +def hist2d(x, y, n=100): + # 2d histogram used in labels.png and evolve.png + xedges, yedges = np.linspace(x.min(), x.max(), n), np.linspace(y.min(), y.max(), n) + hist, xedges, yedges = np.histogram2d(x, y, (xedges, yedges)) + xidx = np.clip(np.digitize(x, xedges) - 1, 0, hist.shape[0] - 1) + yidx = np.clip(np.digitize(y, yedges) - 1, 0, hist.shape[1] - 1) + return np.log(hist[xidx, yidx]) + + +def butter_lowpass_filtfilt(data, cutoff=1500, fs=50000, order=5): + from scipy.signal import butter, filtfilt + + # https://stackoverflow.com/questions/28536191/how-to-filter-smooth-with-scipy-numpy + def butter_lowpass(cutoff, fs, order): + nyq = 0.5 * fs + normal_cutoff = cutoff / nyq + return butter(order, normal_cutoff, btype='low', analog=False) + + b, a = butter_lowpass(cutoff, fs, order=order) + return filtfilt(b, a, data) # forward-backward filter + + +def output_to_target(output): + # Convert model output to target format [batch_id, class_id, x, y, w, h, conf] + targets = [] + for i, o in enumerate(output): + for *box, conf, cls in o.cpu().numpy(): + targets.append([i, cls, *list(*xyxy2xywh(np.array(box)[None])), conf]) + return np.array(targets) + + +@threaded +def plot_images(images, targets, paths=None, fname='images.jpg', names=None, max_size=1920, max_subplots=16): + # Plot image grid with labels + if isinstance(images, torch.Tensor): + images = images.cpu().float().numpy() + if isinstance(targets, torch.Tensor): + targets = targets.cpu().numpy() + if np.max(images[0]) <= 1: + images *= 255 # de-normalise (optional) + bs, _, h, w = images.shape # batch size, _, height, width + bs = min(bs, max_subplots) # limit plot images + ns = np.ceil(bs ** 0.5) # number of subplots (square) + + # Build Image + mosaic = np.full((int(ns * h), int(ns * w), 3), 255, dtype=np.uint8) # init + for i, im in enumerate(images): + if i == max_subplots: # if last batch has fewer images than we expect + break + x, y = int(w * (i // ns)), int(h * (i % ns)) # block origin + im = im.transpose(1, 2, 0) + mosaic[y:y + h, x:x + w, :] = im + + # Resize (optional) + scale = max_size / ns / max(h, w) + if scale < 1: + h = math.ceil(scale * h) + w = math.ceil(scale * w) + mosaic = cv2.resize(mosaic, tuple(int(x * ns) for x in (w, h))) + + # Annotate + fs = int((h + w) * ns * 0.01) # font size + annotator = Annotator(mosaic, line_width=round(fs / 10), font_size=fs, pil=True, example=names) + for i in range(i + 1): + x, y = int(w * (i // ns)), int(h * (i % ns)) # block origin + annotator.rectangle([x, y, x + w, y + h], None, (255, 255, 255), width=2) # borders + if paths: + annotator.text((x + 5, y + 5 + h), text=Path(paths[i]).name[:40], txt_color=(220, 220, 220)) # filenames + if len(targets) > 0: + ti = targets[targets[:, 0] == i] # image targets + boxes = xywh2xyxy(ti[:, 2:6]).T + classes = ti[:, 1].astype('int') + labels = ti.shape[1] == 6 # labels if no conf column + conf = None if labels else ti[:, 6] # check for confidence presence (label vs pred) + + if boxes.shape[1]: + if boxes.max() <= 1.01: # if normalized with tolerance 0.01 + boxes[[0, 2]] *= w # scale to pixels + boxes[[1, 3]] *= h + elif scale < 1: # absolute coords need scale if image scales + boxes *= scale + boxes[[0, 2]] += x + boxes[[1, 3]] += y + for j, box in enumerate(boxes.T.tolist()): + cls = classes[j] + color = colors(cls) + cls = names[cls] if names else cls + if labels or conf[j] > 0.25: # 0.25 conf thresh + label = f'{cls}' if labels else f'{cls} {conf[j]:.1f}' + annotator.box_label(box, label, color=color) + annotator.im.save(fname) # save + + +def plot_lr_scheduler(optimizer, scheduler, epochs=300, save_dir=''): + # Plot LR simulating training for full epochs + optimizer, scheduler = copy(optimizer), copy(scheduler) # do not modify originals + y = [] + for _ in range(epochs): + scheduler.step() + y.append(optimizer.param_groups[0]['lr']) + plt.plot(y, '.-', label='LR') + plt.xlabel('epoch') + plt.ylabel('LR') + plt.grid() + plt.xlim(0, epochs) + plt.ylim(0) + plt.savefig(Path(save_dir) / 'LR.png', dpi=200) + plt.close() + + +def plot_val_txt(): # from utils.plots import *; plot_val() + # Plot val.txt histograms + x = np.loadtxt('val.txt', dtype=np.float32) + box = xyxy2xywh(x[:, :4]) + cx, cy = box[:, 0], box[:, 1] + + fig, ax = plt.subplots(1, 1, figsize=(6, 6), tight_layout=True) + ax.hist2d(cx, cy, bins=600, cmax=10, cmin=0) + ax.set_aspect('equal') + plt.savefig('hist2d.png', dpi=300) + + fig, ax = plt.subplots(1, 2, figsize=(12, 6), tight_layout=True) + ax[0].hist(cx, bins=600) + ax[1].hist(cy, bins=600) + plt.savefig('hist1d.png', dpi=200) + + +def plot_targets_txt(): # from utils.plots import *; plot_targets_txt() + # Plot targets.txt histograms + x = np.loadtxt('targets.txt', dtype=np.float32).T + s = ['x targets', 'y targets', 'width targets', 'height targets'] + fig, ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True) + ax = ax.ravel() + for i in range(4): + ax[i].hist(x[i], bins=100, label=f'{x[i].mean():.3g} +/- {x[i].std():.3g}') + ax[i].legend() + ax[i].set_title(s[i]) + plt.savefig('targets.jpg', dpi=200) + + +def plot_val_study(file='', dir='', x=None): # from utils.plots import *; plot_val_study() + # Plot file=study.txt generated by val.py (or plot all study*.txt in dir) + save_dir = Path(file).parent if file else Path(dir) + plot2 = False # plot additional results + if plot2: + ax = plt.subplots(2, 4, figsize=(10, 6), tight_layout=True)[1].ravel() + + fig2, ax2 = plt.subplots(1, 1, figsize=(8, 4), tight_layout=True) + # for f in [save_dir / f'study_coco_{x}.txt' for x in ['yolov5n6', 'yolov5s6', 'yolov5m6', 'yolov5l6', 'yolov5x6']]: + for f in sorted(save_dir.glob('study*.txt')): + y = np.loadtxt(f, dtype=np.float32, usecols=[0, 1, 2, 3, 7, 8, 9], ndmin=2).T + x = np.arange(y.shape[1]) if x is None else np.array(x) + if plot2: + s = ['P', 'R', 'mAP@.5', 'mAP@.5:.95', 't_preprocess (ms/img)', 't_inference (ms/img)', 't_NMS (ms/img)'] + for i in range(7): + ax[i].plot(x, y[i], '.-', linewidth=2, markersize=8) + ax[i].set_title(s[i]) + + j = y[3].argmax() + 1 + ax2.plot(y[5, 1:j], + y[3, 1:j] * 1E2, + '.-', + linewidth=2, + markersize=8, + label=f.stem.replace('study_coco_', '').replace('yolo', 'YOLO')) + + ax2.plot(1E3 / np.array([209, 140, 97, 58, 35, 18]), [34.6, 40.5, 43.0, 47.5, 49.7, 51.5], + 'k.-', + linewidth=2, + markersize=8, + alpha=.25, + label='EfficientDet') + + ax2.grid(alpha=0.2) + ax2.set_yticks(np.arange(20, 60, 5)) + ax2.set_xlim(0, 57) + ax2.set_ylim(25, 55) + ax2.set_xlabel('GPU Speed (ms/img)') + ax2.set_ylabel('COCO AP val') + ax2.legend(loc='lower right') + f = save_dir / 'study.png' + print(f'Saving {f}...') + plt.savefig(f, dpi=300) + + +@try_except # known issue https://github.com/ultralytics/yolov5/issues/5395 +@Timeout(30) # known issue https://github.com/ultralytics/yolov5/issues/5611 +def plot_labels(labels, names=(), save_dir=Path('')): + # plot dataset labels + LOGGER.info(f"Plotting labels to {save_dir / 'labels.jpg'}... ") + c, b = labels[:, 0], labels[:, 1:].transpose() # classes, boxes + nc = int(c.max() + 1) # number of classes + x = pd.DataFrame(b.transpose(), columns=['x', 'y', 'width', 'height']) + + # seaborn correlogram + sn.pairplot(x, corner=True, diag_kind='auto', kind='hist', diag_kws=dict(bins=50), plot_kws=dict(pmax=0.9)) + plt.savefig(save_dir / 'labels_correlogram.jpg', dpi=200) + plt.close() + + # matplotlib labels + matplotlib.use('svg') # faster + ax = plt.subplots(2, 2, figsize=(8, 8), tight_layout=True)[1].ravel() + y = ax[0].hist(c, bins=np.linspace(0, nc, nc + 1) - 0.5, rwidth=0.8) + try: # color histogram bars by class + [y[2].patches[i].set_color([x / 255 for x in colors(i)]) for i in range(nc)] # known issue #3195 + except Exception: + pass + ax[0].set_ylabel('instances') + if 0 < len(names) < 30: + ax[0].set_xticks(range(len(names))) + ax[0].set_xticklabels(names, rotation=90, fontsize=10) + else: + ax[0].set_xlabel('classes') + sn.histplot(x, x='x', y='y', ax=ax[2], bins=50, pmax=0.9) + sn.histplot(x, x='width', y='height', ax=ax[3], bins=50, pmax=0.9) + + # rectangles + labels[:, 1:3] = 0.5 # center + labels[:, 1:] = xywh2xyxy(labels[:, 1:]) * 2000 + img = Image.fromarray(np.ones((2000, 2000, 3), dtype=np.uint8) * 255) + for cls, *box in labels[:1000]: + ImageDraw.Draw(img).rectangle(box, width=1, outline=colors(cls)) # plot + ax[1].imshow(img) + ax[1].axis('off') + + for a in [0, 1, 2, 3]: + for s in ['top', 'right', 'left', 'bottom']: + ax[a].spines[s].set_visible(False) + + plt.savefig(save_dir / 'labels.jpg', dpi=200) + matplotlib.use('Agg') + plt.close() + + +def imshow_cls(im, labels=None, pred=None, names=None, nmax=25, verbose=False, f=Path('images.jpg')): + # Show classification image grid with labels (optional) and predictions (optional) + from utils.augmentations import denormalize + + names = names or [f'class{i}' for i in range(1000)] + blocks = torch.chunk(denormalize(im.clone()).cpu().float(), len(im), + dim=0) # select batch index 0, block by channels + n = min(len(blocks), nmax) # number of plots + m = min(8, round(n ** 0.5)) # 8 x 8 default + fig, ax = plt.subplots(math.ceil(n / m), m) # 8 rows x n/8 cols + ax = ax.ravel() if m > 1 else [ax] + # plt.subplots_adjust(wspace=0.05, hspace=0.05) + for i in range(n): + ax[i].imshow(blocks[i].squeeze().permute((1, 2, 0)).numpy().clip(0.0, 1.0)) + ax[i].axis('off') + if labels is not None: + s = names[labels[i]] + (f'—{names[pred[i]]}' if pred is not None else '') + ax[i].set_title(s, fontsize=8, verticalalignment='top') + plt.savefig(f, dpi=300, bbox_inches='tight') + plt.close() + if verbose: + LOGGER.info(f"Saving {f}") + if labels is not None: + LOGGER.info('True: ' + ' '.join(f'{names[i]:3s}' for i in labels[:nmax])) + if pred is not None: + LOGGER.info('Predicted:' + ' '.join(f'{names[i]:3s}' for i in pred[:nmax])) + return f + + +def plot_evolve(evolve_csv='path/to/evolve.csv'): # from utils.plots import *; plot_evolve() + # Plot evolve.csv hyp evolution results + evolve_csv = Path(evolve_csv) + data = pd.read_csv(evolve_csv) + keys = [x.strip() for x in data.columns] + x = data.values + f = fitness(x) + j = np.argmax(f) # max fitness index + plt.figure(figsize=(10, 12), tight_layout=True) + matplotlib.rc('font', **{'size': 8}) + print(f'Best results from row {j} of {evolve_csv}:') + for i, k in enumerate(keys[7:]): + v = x[:, 7 + i] + mu = v[j] # best single result + plt.subplot(6, 5, i + 1) + plt.scatter(v, f, c=hist2d(v, f, 20), cmap='viridis', alpha=.8, edgecolors='none') + plt.plot(mu, f.max(), 'k+', markersize=15) + plt.title(f'{k} = {mu:.3g}', fontdict={'size': 9}) # limit to 40 characters + if i % 5 != 0: + plt.yticks([]) + print(f'{k:>15}: {mu:.3g}') + f = evolve_csv.with_suffix('.png') # filename + plt.savefig(f, dpi=200) + plt.close() + print(f'Saved {f}') + + +def plot_results(file='path/to/results.csv', dir=''): + # Plot training results.csv. Usage: from utils.plots import *; plot_results('path/to/results.csv') + save_dir = Path(file).parent if file else Path(dir) + fig, ax = plt.subplots(2, 5, figsize=(12, 6), tight_layout=True) + ax = ax.ravel() + files = list(save_dir.glob('results*.csv')) + assert len(files), f'No results.csv files found in {save_dir.resolve()}, nothing to plot.' + for f in files: + try: + data = pd.read_csv(f) + s = [x.strip() for x in data.columns] + x = data.values[:, 0] + for i, j in enumerate([1, 2, 3, 4, 5, 8, 9, 10, 6, 7]): + y = data.values[:, j].astype('float') + # y[y == 0] = np.nan # don't show zero values + ax[i].plot(x, y, marker='.', label=f.stem, linewidth=2, markersize=8) + ax[i].set_title(s[j], fontsize=12) + # if j in [8, 9, 10]: # share train and val loss y axes + # ax[i].get_shared_y_axes().join(ax[i], ax[i - 5]) + except Exception as e: + LOGGER.info(f'Warning: Plotting error for {f}: {e}') + ax[1].legend() + fig.savefig(save_dir / 'results.png', dpi=200) + plt.close() + + +def profile_idetection(start=0, stop=0, labels=(), save_dir=''): + # Plot iDetection '*.txt' per-image logs. from utils.plots import *; profile_idetection() + ax = plt.subplots(2, 4, figsize=(12, 6), tight_layout=True)[1].ravel() + s = ['Images', 'Free Storage (GB)', 'RAM Usage (GB)', 'Battery', 'dt_raw (ms)', 'dt_smooth (ms)', 'real-world FPS'] + files = list(Path(save_dir).glob('frames*.txt')) + for fi, f in enumerate(files): + try: + results = np.loadtxt(f, ndmin=2).T[:, 90:-30] # clip first and last rows + n = results.shape[1] # number of rows + x = np.arange(start, min(stop, n) if stop else n) + results = results[:, x] + t = (results[0] - results[0].min()) # set t0=0s + results[0] = x + for i, a in enumerate(ax): + if i < len(results): + label = labels[fi] if len(labels) else f.stem.replace('frames_', '') + a.plot(t, results[i], marker='.', label=label, linewidth=1, markersize=5) + a.set_title(s[i]) + a.set_xlabel('time (s)') + # if fi == len(files) - 1: + # a.set_ylim(bottom=0) + for side in ['top', 'right']: + a.spines[side].set_visible(False) + else: + a.remove() + except Exception as e: + print(f'Warning: Plotting error for {f}; {e}') + ax[1].legend() + plt.savefig(Path(save_dir) / 'idetection_profile.png', dpi=200) + + +def save_one_box(xyxy, im, file=Path('im.jpg'), gain=1.02, pad=10, square=False, BGR=False, save=True): + # Save image crop as {file} with crop size multiple {gain} and {pad} pixels. Save and/or return crop + xyxy = torch.tensor(xyxy).view(-1, 4) + b = xyxy2xywh(xyxy) # boxes + if square: + b[:, 2:] = b[:, 2:].max(1)[0].unsqueeze(1) # attempt rectangle to square + b[:, 2:] = b[:, 2:] * gain + pad # box wh * gain + pad + xyxy = xywh2xyxy(b).long() + clip_coords(xyxy, im.shape) + crop = im[int(xyxy[0, 1]):int(xyxy[0, 3]), int(xyxy[0, 0]):int(xyxy[0, 2]), ::(1 if BGR else -1)] + if save: + file.parent.mkdir(parents=True, exist_ok=True) # make directory + f = str(increment_path(file).with_suffix('.jpg')) + # cv2.imwrite(f, crop) # save BGR, https://github.com/ultralytics/yolov5/issues/7007 chroma subsampling issue + Image.fromarray(crop[..., ::-1]).save(f, quality=95, subsampling=0) # save RGB + return crop diff --git a/application/yolov5_example/utils/torch_utils.py b/application/yolov5_example/utils/torch_utils.py new file mode 100644 index 00000000..350c506e --- /dev/null +++ b/application/yolov5_example/utils/torch_utils.py @@ -0,0 +1,433 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +PyTorch utils +""" + +import math +import os +import platform +import subprocess +import time +import warnings +from contextlib import contextmanager +from copy import deepcopy +from pathlib import Path + +import torch +import torch.distributed as dist +import torch.nn as nn +import torch.nn.functional as F +from torch.nn.parallel import DistributedDataParallel as DDP + +from utils.general import LOGGER, check_version, colorstr, file_date, git_describe + +LOCAL_RANK = int(os.getenv('LOCAL_RANK', -1)) # https://pytorch.org/docs/stable/elastic/run.html +RANK = int(os.getenv('RANK', -1)) +WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1)) + +try: + import thop # for FLOPs computation +except ImportError: + thop = None + +# Suppress PyTorch warnings +warnings.filterwarnings('ignore', message='User provided device_type of \'cuda\', but CUDA is not available. Disabling') + + +def smart_inference_mode(torch_1_9=check_version(torch.__version__, '1.9.0')): + # Applies torch.inference_mode() decorator if torch>=1.9.0 else torch.no_grad() decorator + def decorate(fn): + return (torch.inference_mode if torch_1_9 else torch.no_grad)()(fn) + + return decorate + + +def smartCrossEntropyLoss(label_smoothing=0.0): + # Returns nn.CrossEntropyLoss with label smoothing enabled for torch>=1.10.0 + if check_version(torch.__version__, '1.10.0'): + return nn.CrossEntropyLoss(label_smoothing=label_smoothing) # loss function + else: + if label_smoothing > 0: + LOGGER.warning(f'WARNING: label smoothing {label_smoothing} requires torch>=1.10.0') + return nn.CrossEntropyLoss() # loss function + + +def smart_DDP(model): + # Model DDP creation with checks + assert not check_version(torch.__version__, '1.12.0', pinned=True), \ + 'torch==1.12.0 torchvision==0.13.0 DDP training is not supported due to a known issue. ' \ + 'Please upgrade or downgrade torch to use DDP. See https://github.com/ultralytics/yolov5/issues/8395' + if check_version(torch.__version__, '1.11.0'): + return DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK, static_graph=True) + else: + return DDP(model, device_ids=[LOCAL_RANK], output_device=LOCAL_RANK) + + +def reshape_classifier_output(model, n=1000): + # Update a TorchVision classification model to class count 'n' if required + from models.common import Classify + name, m = list((model.model if hasattr(model, 'model') else model).named_children())[-1] # last module + if isinstance(m, Classify): # YOLOv5 Classify() head + if m.linear.out_features != n: + m.linear = nn.Linear(m.linear.in_features, n) + elif isinstance(m, nn.Linear): # ResNet, EfficientNet + if m.out_features != n: + setattr(model, name, nn.Linear(m.in_features, n)) + elif isinstance(m, nn.Sequential): + types = [type(x) for x in m] + if nn.Linear in types: + i = types.index(nn.Linear) # nn.Linear index + if m[i].out_features != n: + m[i] = nn.Linear(m[i].in_features, n) + elif nn.Conv2d in types: + i = types.index(nn.Conv2d) # nn.Conv2d index + if m[i].out_channels != n: + m[i] = nn.Conv2d(m[i].in_channels, n, m[i].kernel_size, m[i].stride, bias=m[i].bias) + + +@contextmanager +def torch_distributed_zero_first(local_rank: int): + # Decorator to make all processes in distributed training wait for each local_master to do something + if local_rank not in [-1, 0]: + dist.barrier(device_ids=[local_rank]) + yield + if local_rank == 0: + dist.barrier(device_ids=[0]) + + +def device_count(): + # Returns number of CUDA devices available. Safe version of torch.cuda.device_count(). Supports Linux and Windows + assert platform.system() in ('Linux', 'Windows'), 'device_count() only supported on Linux or Windows' + try: + cmd = 'nvidia-smi -L | wc -l' if platform.system() == 'Linux' else 'nvidia-smi -L | find /c /v ""' # Windows + return int(subprocess.run(cmd, shell=True, capture_output=True, check=True).stdout.decode().split()[-1]) + except Exception: + return 0 + + +def select_device(device='', batch_size=0, newline=True): + # device = None or 'cpu' or 0 or '0' or '0,1,2,3' + s = f'YOLOv5 🚀 {git_describe() or file_date()} Python-{platform.python_version()} torch-{torch.__version__} ' + device = str(device).strip().lower().replace('cuda:', '').replace('none', '') # to string, 'cuda:0' to '0' + cpu = device == 'cpu' + mps = device == 'mps' # Apple Metal Performance Shaders (MPS) + if cpu or mps: + os.environ['CUDA_VISIBLE_DEVICES'] = '-1' # force torch.cuda.is_available() = False + elif device: # non-cpu device requested + os.environ['CUDA_VISIBLE_DEVICES'] = device # set environment variable - must be before assert is_available() + assert torch.cuda.is_available() and torch.cuda.device_count() >= len(device.replace(',', '')), \ + f"Invalid CUDA '--device {device}' requested, use '--device cpu' or pass valid CUDA device(s)" + + if not (cpu or mps) and torch.cuda.is_available(): # prefer GPU if available + devices = device.split(',') if device else '0' # range(torch.cuda.device_count()) # i.e. 0,1,6,7 + n = len(devices) # device count + if n > 1 and batch_size > 0: # check batch_size is divisible by device_count + assert batch_size % n == 0, f'batch-size {batch_size} not multiple of GPU count {n}' + space = ' ' * (len(s) + 1) + for i, d in enumerate(devices): + p = torch.cuda.get_device_properties(i) + s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / (1 << 20):.0f}MiB)\n" # bytes to MB + arg = 'cuda:0' + elif mps and getattr(torch, 'has_mps', False) and torch.backends.mps.is_available(): # prefer MPS if available + s += 'MPS\n' + arg = 'mps' + else: # revert to CPU + s += 'CPU\n' + arg = 'cpu' + + if not newline: + s = s.rstrip() + LOGGER.info(s) + return torch.device(arg) + + +def time_sync(): + # PyTorch-accurate time + if torch.cuda.is_available(): + torch.cuda.synchronize() + return time.time() + + +def profile(input, ops, n=10, device=None): + """ YOLOv5 speed/memory/FLOPs profiler + Usage: + input = torch.randn(16, 3, 640, 640) + m1 = lambda x: x * torch.sigmoid(x) + m2 = nn.SiLU() + profile(input, [m1, m2], n=100) # profile over 100 iterations + """ + results = [] + if not isinstance(device, torch.device): + device = select_device(device) + print(f"{'Params':>12s}{'GFLOPs':>12s}{'GPU_mem (GB)':>14s}{'forward (ms)':>14s}{'backward (ms)':>14s}" + f"{'input':>24s}{'output':>24s}") + + for x in input if isinstance(input, list) else [input]: + x = x.to(device) + x.requires_grad = True + for m in ops if isinstance(ops, list) else [ops]: + m = m.to(device) if hasattr(m, 'to') else m # device + m = m.half() if hasattr(m, 'half') and isinstance(x, torch.Tensor) and x.dtype is torch.float16 else m + tf, tb, t = 0, 0, [0, 0, 0] # dt forward, backward + try: + flops = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 # GFLOPs + except Exception: + flops = 0 + + try: + for _ in range(n): + t[0] = time_sync() + y = m(x) + t[1] = time_sync() + try: + _ = (sum(yi.sum() for yi in y) if isinstance(y, list) else y).sum().backward() + t[2] = time_sync() + except Exception: # no backward method + # print(e) # for debug + t[2] = float('nan') + tf += (t[1] - t[0]) * 1000 / n # ms per op forward + tb += (t[2] - t[1]) * 1000 / n # ms per op backward + mem = torch.cuda.memory_reserved() / 1E9 if torch.cuda.is_available() else 0 # (GB) + s_in, s_out = (tuple(x.shape) if isinstance(x, torch.Tensor) else 'list' for x in (x, y)) # shapes + p = sum(x.numel() for x in m.parameters()) if isinstance(m, nn.Module) else 0 # parameters + print(f'{p:12}{flops:12.4g}{mem:>14.3f}{tf:14.4g}{tb:14.4g}{str(s_in):>24s}{str(s_out):>24s}') + results.append([p, flops, mem, tf, tb, s_in, s_out]) + except Exception as e: + print(e) + results.append(None) + torch.cuda.empty_cache() + return results + + +def is_parallel(model): + # Returns True if model is of type DP or DDP + return type(model) in (nn.parallel.DataParallel, nn.parallel.DistributedDataParallel) + + +def de_parallel(model): + # De-parallelize a model: returns single-GPU model if model is of type DP or DDP + return model.module if is_parallel(model) else model + + +def initialize_weights(model): + for m in model.modules(): + t = type(m) + if t is nn.Conv2d: + pass # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') + elif t is nn.BatchNorm2d: + m.eps = 1e-3 + m.momentum = 0.03 + elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]: + m.inplace = True + + +def find_modules(model, mclass=nn.Conv2d): + # Finds layer indices matching module class 'mclass' + return [i for i, m in enumerate(model.module_list) if isinstance(m, mclass)] + + +def sparsity(model): + # Return global model sparsity + a, b = 0, 0 + for p in model.parameters(): + a += p.numel() + b += (p == 0).sum() + return b / a + + +def prune(model, amount=0.3): + # Prune model to requested global sparsity + import torch.nn.utils.prune as prune + for name, m in model.named_modules(): + if isinstance(m, nn.Conv2d): + prune.l1_unstructured(m, name='weight', amount=amount) # prune + prune.remove(m, 'weight') # make permanent + LOGGER.info(f'Model pruned to {sparsity(model):.3g} global sparsity') + + +def fuse_conv_and_bn(conv, bn): + # Fuse Conv2d() and BatchNorm2d() layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/ + fusedconv = nn.Conv2d(conv.in_channels, + conv.out_channels, + kernel_size=conv.kernel_size, + stride=conv.stride, + padding=conv.padding, + groups=conv.groups, + bias=True).requires_grad_(False).to(conv.weight.device) + + # Prepare filters + w_conv = conv.weight.clone().view(conv.out_channels, -1) + w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var))) + fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape)) + + # Prepare spatial bias + b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias + b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps)) + fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn) + + return fusedconv + + +def model_info(model, verbose=False, imgsz=640): + # Model information. img_size may be int or list, i.e. img_size=640 or img_size=[640, 320] + n_p = sum(x.numel() for x in model.parameters()) # number parameters + n_g = sum(x.numel() for x in model.parameters() if x.requires_grad) # number gradients + if verbose: + print(f"{'layer':>5} {'name':>40} {'gradient':>9} {'parameters':>12} {'shape':>20} {'mu':>10} {'sigma':>10}") + for i, (name, p) in enumerate(model.named_parameters()): + name = name.replace('module_list.', '') + print('%5g %40s %9s %12g %20s %10.3g %10.3g' % + (i, name, p.requires_grad, p.numel(), list(p.shape), p.mean(), p.std())) + + try: # FLOPs + p = next(model.parameters()) + stride = max(int(model.stride.max()), 32) if hasattr(model, 'stride') else 32 # max stride + im = torch.zeros((1, p.shape[1], stride, stride), device=p.device) # input image in BCHW format + flops = thop.profile(deepcopy(model), inputs=(im,), verbose=False)[0] / 1E9 * 2 # stride GFLOPs + imgsz = imgsz if isinstance(imgsz, list) else [imgsz, imgsz] # expand if int/float + fs = f', {flops * imgsz[0] / stride * imgsz[1] / stride:.1f} GFLOPs' # 640x640 GFLOPs + except Exception: + fs = '' + + name = Path(model.yaml_file).stem.replace('yolov5', 'YOLOv5') if hasattr(model, 'yaml_file') else 'Model' + LOGGER.info(f"{name} summary: {len(list(model.modules()))} layers, {n_p} parameters, {n_g} gradients{fs}") + + +def scale_img(img, ratio=1.0, same_shape=False, gs=32): # img(16,3,256,416) + # Scales img(bs,3,y,x) by ratio constrained to gs-multiple + if ratio == 1.0: + return img + h, w = img.shape[2:] + s = (int(h * ratio), int(w * ratio)) # new size + img = F.interpolate(img, size=s, mode='bilinear', align_corners=False) # resize + if not same_shape: # pad/crop img + h, w = (math.ceil(x * ratio / gs) * gs for x in (h, w)) + return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447) # value = imagenet mean + + +def copy_attr(a, b, include=(), exclude=()): + # Copy attributes from b to a, options to only include [...] and to exclude [...] + for k, v in b.__dict__.items(): + if (len(include) and k not in include) or k.startswith('_') or k in exclude: + continue + else: + setattr(a, k, v) + + +def smart_optimizer(model, name='Adam', lr=0.001, momentum=0.9, decay=1e-5): + # YOLOv5 3-param group optimizer: 0) weights with decay, 1) weights no decay, 2) biases no decay + g = [], [], [] # optimizer parameter groups + bn = tuple(v for k, v in nn.__dict__.items() if 'Norm' in k) # normalization layers, i.e. BatchNorm2d() + for v in model.modules(): + if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): # bias (no decay) + g[2].append(v.bias) + if isinstance(v, bn): # weight (no decay) + g[1].append(v.weight) + elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): # weight (with decay) + g[0].append(v.weight) + + if name == 'Adam': + optimizer = torch.optim.Adam(g[2], lr=lr, betas=(momentum, 0.999)) # adjust beta1 to momentum + elif name == 'AdamW': + optimizer = torch.optim.AdamW(g[2], lr=lr, betas=(momentum, 0.999), weight_decay=0.0) + elif name == 'RMSProp': + optimizer = torch.optim.RMSprop(g[2], lr=lr, momentum=momentum) + elif name == 'SGD': + optimizer = torch.optim.SGD(g[2], lr=lr, momentum=momentum, nesterov=True) + else: + raise NotImplementedError(f'Optimizer {name} not implemented.') + + optimizer.add_param_group({'params': g[0], 'weight_decay': decay}) # add g0 with weight_decay + optimizer.add_param_group({'params': g[1], 'weight_decay': 0.0}) # add g1 (BatchNorm2d weights) + LOGGER.info(f"{colorstr('optimizer:')} {type(optimizer).__name__}(lr={lr}) with parameter groups " + f"{len(g[1])} weight(decay=0.0), {len(g[0])} weight(decay={decay}), {len(g[2])} bias") + return optimizer + + +def smart_hub_load(repo='ultralytics/yolov5', model='yolov5s', **kwargs): + # YOLOv5 torch.hub.load() wrapper with smart error/issue handling + if check_version(torch.__version__, '1.9.1'): + kwargs['skip_validation'] = True # validation causes GitHub API rate limit errors + if check_version(torch.__version__, '1.12.0'): + kwargs['trust_repo'] = True # argument required starting in torch 0.12 + try: + return torch.hub.load(repo, model, **kwargs) + except Exception: + return torch.hub.load(repo, model, force_reload=True, **kwargs) + + +def smart_resume(ckpt, optimizer, ema=None, weights='yolov5s.pt', epochs=300, resume=True): + # Resume training from a partially trained checkpoint + best_fitness = 0.0 + start_epoch = ckpt['epoch'] + 1 + if ckpt['optimizer'] is not None: + optimizer.load_state_dict(ckpt['optimizer']) # optimizer + best_fitness = ckpt['best_fitness'] + if ema and ckpt.get('ema'): + ema.ema.load_state_dict(ckpt['ema'].float().state_dict()) # EMA + ema.updates = ckpt['updates'] + if resume: + assert start_epoch > 0, f'{weights} training to {epochs} epochs is finished, nothing to resume.\n' \ + f"Start a new training without --resume, i.e. 'python train.py --weights {weights}'" + LOGGER.info(f'Resuming training from {weights} from epoch {start_epoch} to {epochs} total epochs') + if epochs < start_epoch: + LOGGER.info(f"{weights} has been trained for {ckpt['epoch']} epochs. Fine-tuning for {epochs} more epochs.") + epochs += ckpt['epoch'] # finetune additional epochs + return best_fitness, start_epoch, epochs + + +class EarlyStopping: + # YOLOv5 simple early stopper + def __init__(self, patience=30): + self.best_fitness = 0.0 # i.e. mAP + self.best_epoch = 0 + self.patience = patience or float('inf') # epochs to wait after fitness stops improving to stop + self.possible_stop = False # possible stop may occur next epoch + + def __call__(self, epoch, fitness): + if fitness >= self.best_fitness: # >= 0 to allow for early zero-fitness stage of training + self.best_epoch = epoch + self.best_fitness = fitness + delta = epoch - self.best_epoch # epochs without improvement + self.possible_stop = delta >= (self.patience - 1) # possible stop may occur next epoch + stop = delta >= self.patience # stop training if patience exceeded + if stop: + LOGGER.info(f'Stopping training early as no improvement observed in last {self.patience} epochs. ' + f'Best results observed at epoch {self.best_epoch}, best model saved as best.pt.\n' + f'To update EarlyStopping(patience={self.patience}) pass a new patience value, ' + f'i.e. `python train.py --patience 300` or use `--patience 0` to disable EarlyStopping.') + return stop + + +class ModelEMA: + """ Updated Exponential Moving Average (EMA) from https://github.com/rwightman/pytorch-image-models + Keeps a moving average of everything in the model state_dict (parameters and buffers) + For EMA details see https://www.tensorflow.org/api_docs/python/tf/train/ExponentialMovingAverage + """ + + def __init__(self, model, decay=0.9999, tau=2000, updates=0): + # Create EMA + self.ema = deepcopy(de_parallel(model)).eval() # FP32 EMA + # if next(model.parameters()).device.type != 'cpu': + # self.ema.half() # FP16 EMA + self.updates = updates # number of EMA updates + self.decay = lambda x: decay * (1 - math.exp(-x / tau)) # decay exponential ramp (to help early epochs) + for p in self.ema.parameters(): + p.requires_grad_(False) + + @smart_inference_mode() + def update(self, model): + return + # Update EMA parameters + self.updates += 1 + d = self.decay(self.updates) + + msd = de_parallel(model).state_dict() # model state_dict + for k, v in self.ema.state_dict().items(): + if v.dtype.is_floating_point: + v *= d + v += (1 - d) * msd[k].detach() + + def update_attr(self, model, include=(), exclude=('process_group', 'reducer')): + # Update EMA attributes + copy_attr(self.ema, model, include, exclude) diff --git a/application/yolov5_example/val.py b/application/yolov5_example/val.py new file mode 100644 index 00000000..13049623 --- /dev/null +++ b/application/yolov5_example/val.py @@ -0,0 +1,396 @@ +# YOLOv5 🚀 by Ultralytics, GPL-3.0 license +""" +Validate a trained YOLOv5 model accuracy on a custom dataset + +Usage: + $ python path/to/val.py --weights yolov5s.pt --data coco128.yaml --img 640 + +Usage - formats: + $ python path/to/val.py --weights yolov5s.pt # PyTorch + yolov5s.torchscript # TorchScript + yolov5s.onnx # ONNX Runtime or OpenCV DNN with --dnn + yolov5s.xml # OpenVINO + yolov5s.engine # TensorRT + yolov5s.mlmodel # CoreML (macOS-only) + yolov5s_saved_model # TensorFlow SavedModel + yolov5s.pb # TensorFlow GraphDef + yolov5s.tflite # TensorFlow Lite + yolov5s_edgetpu.tflite # TensorFlow Edge TPU +""" + +import argparse +import json +import os +import sys +from pathlib import Path + +import numpy as np +import torch +from tqdm import tqdm + +FILE = Path(__file__).resolve() +ROOT = FILE.parents[0] # YOLOv5 root directory +if str(ROOT) not in sys.path: + sys.path.append(str(ROOT)) # add ROOT to PATH +ROOT = Path(os.path.relpath(ROOT, Path.cwd())) # relative + +from models.common import DetectMultiBackend +from utils.callbacks import Callbacks +from utils.dataloaders import create_dataloader +from utils.general import (LOGGER, check_dataset, check_img_size, check_requirements, check_yaml, + coco80_to_coco91_class, colorstr, increment_path, non_max_suppression, print_args, + scale_coords, xywh2xyxy, xyxy2xywh) +from utils.metrics import ConfusionMatrix, ap_per_class, box_iou +from utils.plots import output_to_target, plot_images, plot_val_study +from utils.torch_utils import select_device, smart_inference_mode, time_sync + + +def save_one_txt(predn, save_conf, shape, file): + # Save one txt result + gn = torch.tensor(shape)[[1, 0, 1, 0]] # normalization gain whwh + for *xyxy, conf, cls in predn.tolist(): + xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh + line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format + with open(file, 'a') as f: + f.write(('%g ' * len(line)).rstrip() % line + '\n') + + +def save_one_json(predn, jdict, path, class_map): + # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236} + image_id = int(path.stem) if path.stem.isnumeric() else path.stem + box = xyxy2xywh(predn[:, :4]) # xywh + box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner + for p, b in zip(predn.tolist(), box.tolist()): + jdict.append({ + 'image_id': image_id, + 'category_id': class_map[int(p[5])], + 'bbox': [round(x, 3) for x in b], + 'score': round(p[4], 5)}) + + +def process_batch(detections, labels, iouv): + """ + Return correct predictions matrix. Both sets of boxes are in (x1, y1, x2, y2) format. + Arguments: + detections (Array[N, 6]), x1, y1, x2, y2, conf, class + labels (Array[M, 5]), class, x1, y1, x2, y2 + Returns: + correct (Array[N, 10]), for 10 IoU levels + """ + correct = np.zeros((detections.shape[0], iouv.shape[0])).astype(bool) + iou = box_iou(labels[:, 1:], detections[:, :4]) + correct_class = labels[:, 0:1] == detections[:, 5] + for i in range(len(iouv)): + x = torch.where((iou >= iouv[i]) & correct_class) # IoU > threshold and classes match + if x[0].shape[0]: + matches = torch.cat((torch.stack(x, 1), iou[x[0], x[1]][:, None]), 1).cpu().numpy() # [label, detect, iou] + if x[0].shape[0] > 1: + matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 1], return_index=True)[1]] + # matches = matches[matches[:, 2].argsort()[::-1]] + matches = matches[np.unique(matches[:, 0], return_index=True)[1]] + correct[matches[:, 1].astype(int), i] = True + return torch.tensor(correct, dtype=torch.bool, device=iouv.device) + + +@smart_inference_mode() +def run( + data, + weights=None, # model.pt path(s) + batch_size=32, # batch size + imgsz=640, # inference size (pixels) + conf_thres=0.001, # confidence threshold + iou_thres=0.6, # NMS IoU threshold + task='val', # train, val, test, speed or study + device='', # cuda device, i.e. 0 or 0,1,2,3 or cpu + workers=8, # max dataloader workers (per RANK in DDP mode) + single_cls=False, # treat as single-class dataset + augment=False, # augmented inference + verbose=False, # verbose output + save_txt=False, # save results to *.txt + save_hybrid=False, # save label+prediction hybrid results to *.txt + save_conf=False, # save confidences in --save-txt labels + save_json=False, # save a COCO-JSON results file + project=ROOT / 'runs/val', # save to project/name + name='exp', # save to project/name + exist_ok=False, # existing project/name ok, do not increment + half=True, # use FP16 half-precision inference + dnn=False, # use OpenCV DNN for ONNX inference + model=None, + dataloader=None, + save_dir=Path(''), + plots=True, + callbacks=Callbacks(), + compute_loss=None, +): + # Initialize/load model and set device + training = model is not None + if training: # called by train.py + device, pt, jit, engine = next(model.parameters()).device, True, False, False # get model device, PyTorch model + half &= device.type != 'cpu' # half precision only supported on CUDA + model.half() if half else model.float() + else: # called directly + device = select_device(device, batch_size=batch_size) + + # Directories + save_dir = increment_path(Path(project) / name, exist_ok=exist_ok) # increment run + (save_dir / 'labels' if save_txt else save_dir).mkdir(parents=True, exist_ok=True) # make dir + + # Load model + model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half) + stride, pt, jit, engine = model.stride, model.pt, model.jit, model.engine + imgsz = check_img_size(imgsz, s=stride) # check image size + half = model.fp16 # FP16 supported on limited backends with CUDA + if engine: + batch_size = model.batch_size + else: + device = model.device + if not (pt or jit): + batch_size = 1 # export.py models default to batch-size 1 + LOGGER.info(f'Forcing --batch-size 1 square inference (1,3,{imgsz},{imgsz}) for non-PyTorch models') + + # Data + data = check_dataset(data) # check + + # Configure + model.eval() + cuda = device.type != 'cpu' + is_coco = isinstance(data.get('val'), str) and data['val'].endswith(f'coco{os.sep}val2017.txt') # COCO dataset + nc = 1 if single_cls else int(data['nc']) # number of classes + iouv = torch.linspace(0.5, 0.95, 10, device=device) # iou vector for mAP@0.5:0.95 + niou = iouv.numel() + + # Dataloader + if not training: + if pt and not single_cls: # check --weights are trained on --data + ncm = model.model.nc + assert ncm == nc, f'{weights} ({ncm} classes) trained on different --data than what you passed ({nc} ' \ + f'classes). Pass correct combination of --weights and --data that are trained together.' + model.warmup(imgsz=(1 if pt else batch_size, 3, imgsz, imgsz)) # warmup + pad = 0.0 if task in ('speed', 'benchmark') else 0.5 + rect = False if task == 'benchmark' else pt # square inference for benchmarks + task = task if task in ('train', 'val', 'test') else 'val' # path to train/val/test images + dataloader = create_dataloader(data[task], + imgsz, + batch_size, + stride, + single_cls, + pad=pad, + rect=rect, + workers=workers, + prefix=colorstr(f'{task}: '))[0] + + seen = 0 + confusion_matrix = ConfusionMatrix(nc=nc) + names = dict(enumerate(model.names if hasattr(model, 'names') else model.module.names)) + class_map = coco80_to_coco91_class() if is_coco else list(range(1000)) + s = ('%20s' + '%11s' * 6) % ('Class', 'Images', 'Labels', 'P', 'R', 'mAP@.5', 'mAP@.5:.95') + dt, p, r, f1, mp, mr, map50, map = [0.0, 0.0, 0.0], 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 + loss = torch.zeros(3, device=device) + jdict, stats, ap, ap_class = [], [], [], [] + callbacks.run('on_val_start') + pbar = tqdm(dataloader, desc=s, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}') # progress bar + for batch_i, (im, targets, paths, shapes) in enumerate(pbar): + callbacks.run('on_val_batch_start') + t1 = time_sync() + if cuda: + im = im.to(device, non_blocking=True) + targets = targets.to(device) + im = im.half() if half else im.float() # uint8 to fp16/32 + im /= 255 # 0 - 255 to 0.0 - 1.0 + nb, _, height, width = im.shape # batch size, channels, height, width + t2 = time_sync() + dt[0] += t2 - t1 + + # Inference + out, train_out = model(im) if training else model(im, augment=augment, val=True) # inference, loss outputs + dt[1] += time_sync() - t2 + + # Loss + if compute_loss: + loss += compute_loss([x.float() for x in train_out], targets)[1] # box, obj, cls + + # NMS + targets[:, 2:] *= torch.tensor((width, height, width, height), device=device) # to pixels + lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)] if save_hybrid else [] # for autolabelling + t3 = time_sync() + out = non_max_suppression(out, conf_thres, iou_thres, labels=lb, multi_label=True, agnostic=single_cls) + dt[2] += time_sync() - t3 + + # Metrics + for si, pred in enumerate(out): + labels = targets[targets[:, 0] == si, 1:] + nl, npr = labels.shape[0], pred.shape[0] # number of labels, predictions + path, shape = Path(paths[si]), shapes[si][0] + correct = torch.zeros(npr, niou, dtype=torch.bool, device=device) # init + seen += 1 + + if npr == 0: + if nl: + stats.append((correct, *torch.zeros((2, 0), device=device), labels[:, 0])) + if plots: + confusion_matrix.process_batch(detections=None, labels=labels[:, 0]) + continue + + # Predictions + if single_cls: + pred[:, 5] = 0 + predn = pred.clone() + scale_coords(im[si].shape[1:], predn[:, :4], shape, shapes[si][1]) # native-space pred + + # Evaluate + if nl: + tbox = xywh2xyxy(labels[:, 1:5]) # target boxes + scale_coords(im[si].shape[1:], tbox, shape, shapes[si][1]) # native-space labels + labelsn = torch.cat((labels[:, 0:1], tbox), 1) # native-space labels + correct = process_batch(predn, labelsn, iouv) + if plots: + confusion_matrix.process_batch(predn, labelsn) + stats.append((correct, pred[:, 4], pred[:, 5], labels[:, 0])) # (correct, conf, pcls, tcls) + + # Save/log + if save_txt: + save_one_txt(predn, save_conf, shape, file=save_dir / 'labels' / f'{path.stem}.txt') + if save_json: + save_one_json(predn, jdict, path, class_map) # append to COCO-JSON dictionary + callbacks.run('on_val_image_end', pred, predn, path, names, im[si]) + + # Plot images + if plots and batch_i < 3: + plot_images(im, targets, paths, save_dir / f'val_batch{batch_i}_labels.jpg', names) # labels + plot_images(im, output_to_target(out), paths, save_dir / f'val_batch{batch_i}_pred.jpg', names) # pred + + callbacks.run('on_val_batch_end') + + # Compute metrics + stats = [torch.cat(x, 0).cpu().numpy() for x in zip(*stats)] # to numpy + if len(stats) and stats[0].any(): + tp, fp, p, r, f1, ap, ap_class = ap_per_class(*stats, plot=plots, save_dir=save_dir, names=names) + ap50, ap = ap[:, 0], ap.mean(1) # AP@0.5, AP@0.5:0.95 + mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean() + nt = np.bincount(stats[3].astype(int), minlength=nc) # number of targets per class + + # Print results + pf = '%20s' + '%11i' * 2 + '%11.3g' * 4 # print format + LOGGER.info(pf % ('all', seen, nt.sum(), mp, mr, map50, map)) + if nt.sum() == 0: + LOGGER.warning(f'WARNING: no labels found in {task} set, can not compute metrics without labels ⚠️') + + # Print results per class + if (verbose or (nc < 50 and not training)) and nc > 1 and len(stats): + for i, c in enumerate(ap_class): + LOGGER.info(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i])) + + # Print speeds + t = tuple(x / seen * 1E3 for x in dt) # speeds per image + if not training: + shape = (batch_size, 3, imgsz, imgsz) + LOGGER.info(f'Speed: %.1fms pre-process, %.1fms inference, %.1fms NMS per image at shape {shape}' % t) + + # Plots + if plots: + confusion_matrix.plot(save_dir=save_dir, names=list(names.values())) + callbacks.run('on_val_end') + + # Save JSON + if save_json and len(jdict): + w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else '' # weights + anno_json = str(Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json') # annotations json + pred_json = str(save_dir / f"{w}_predictions.json") # predictions json + LOGGER.info(f'\nEvaluating pycocotools mAP... saving {pred_json}...') + with open(pred_json, 'w') as f: + json.dump(jdict, f) + + try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb + check_requirements(['pycocotools']) + from pycocotools.coco import COCO + from pycocotools.cocoeval import COCOeval + + anno = COCO(anno_json) # init annotations api + pred = anno.loadRes(pred_json) # init predictions api + eval = COCOeval(anno, pred, 'bbox') + if is_coco: + eval.params.imgIds = [int(Path(x).stem) for x in dataloader.dataset.im_files] # image IDs to evaluate + eval.evaluate() + eval.accumulate() + eval.summarize() + map, map50 = eval.stats[:2] # update results (mAP@0.5:0.95, mAP@0.5) + except Exception as e: + LOGGER.info(f'pycocotools unable to run: {e}') + + # Return results + model.float() # for training + if not training: + s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else '' + LOGGER.info(f"Results saved to {colorstr('bold', save_dir)}{s}") + maps = np.zeros(nc) + map + for i, c in enumerate(ap_class): + maps[c] = ap[i] + return (mp, mr, map50, map, *(loss.cpu() / len(dataloader)).tolist()), maps, t + + +def parse_opt(): + parser = argparse.ArgumentParser() + parser.add_argument('--data', type=str, default=ROOT / 'data/coco128.yaml', help='dataset.yaml path') + parser.add_argument('--weights', nargs='+', type=str, default=ROOT / 'yolov5s.pt', help='model.pt path(s)') + parser.add_argument('--batch-size', type=int, default=32, help='batch size') + parser.add_argument('--imgsz', '--img', '--img-size', type=int, default=640, help='inference size (pixels)') + parser.add_argument('--conf-thres', type=float, default=0.001, help='confidence threshold') + parser.add_argument('--iou-thres', type=float, default=0.6, help='NMS IoU threshold') + parser.add_argument('--task', default='val', help='train, val, test, speed or study') + parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu') + parser.add_argument('--workers', type=int, default=8, help='max dataloader workers (per RANK in DDP mode)') + parser.add_argument('--single-cls', action='store_true', help='treat as single-class dataset') + parser.add_argument('--augment', action='store_true', help='augmented inference') + parser.add_argument('--verbose', action='store_true', help='report mAP by class') + parser.add_argument('--save-txt', action='store_true', help='save results to *.txt') + parser.add_argument('--save-hybrid', action='store_true', help='save label+prediction hybrid results to *.txt') + parser.add_argument('--save-conf', action='store_true', help='save confidences in --save-txt labels') + parser.add_argument('--save-json', action='store_true', help='save a COCO-JSON results file') + parser.add_argument('--project', default=ROOT / 'runs/val', help='save to project/name') + parser.add_argument('--name', default='exp', help='save to project/name') + parser.add_argument('--exist-ok', action='store_true', help='existing project/name ok, do not increment') + parser.add_argument('--half', action='store_true', help='use FP16 half-precision inference') + parser.add_argument('--dnn', action='store_true', help='use OpenCV DNN for ONNX inference') + opt = parser.parse_args() + opt.data = check_yaml(opt.data) # check YAML + opt.save_json |= opt.data.endswith('coco.yaml') + opt.save_txt |= opt.save_hybrid + print_args(vars(opt)) + return opt + + +def main(opt): + check_requirements(requirements=ROOT / 'requirements.txt', exclude=('tensorboard', 'thop')) + + if opt.task in ('train', 'val', 'test'): # run normally + if opt.conf_thres > 0.001: # https://github.com/ultralytics/yolov5/issues/1466 + LOGGER.info(f'WARNING: confidence threshold {opt.conf_thres} > 0.001 produces invalid results ⚠️') + run(**vars(opt)) + + else: + weights = opt.weights if isinstance(opt.weights, list) else [opt.weights] + opt.half = True # FP16 for fastest results + if opt.task == 'speed': # speed benchmarks + # python val.py --task speed --data coco.yaml --batch 1 --weights yolov5n.pt yolov5s.pt... + opt.conf_thres, opt.iou_thres, opt.save_json = 0.25, 0.45, False + for opt.weights in weights: + run(**vars(opt), plots=False) + + elif opt.task == 'study': # speed vs mAP benchmarks + # python val.py --task study --data coco.yaml --iou 0.7 --weights yolov5n.pt yolov5s.pt... + for opt.weights in weights: + f = f'study_{Path(opt.data).stem}_{Path(opt.weights).stem}.txt' # filename to save to + x, y = list(range(256, 1536 + 128, 128)), [] # x axis (image sizes), y axis + for opt.imgsz in x: # img-size + LOGGER.info(f'\nRunning {f} --imgsz {opt.imgsz}...') + r, _, t = run(**vars(opt), plots=False) + y.append(r + t) # results and times + np.savetxt(f, y, fmt='%10.4g') # save + os.system('zip -r study.zip study_*.txt') + plot_val_study(x=x) # plot + + +if __name__ == "__main__": + opt = parse_opt() + main(opt) diff --git a/fpemu_cpp.cpython-38-x86_64-linux-gnu.so b/fpemu_cpp.cpython-38-x86_64-linux-gnu.so new file mode 100755 index 00000000..b3cd17de Binary files /dev/null and b/fpemu_cpp.cpython-38-x86_64-linux-gnu.so differ diff --git a/fpemu_cuda.cpython-38-x86_64-linux-gnu.so b/fpemu_cuda.cpython-38-x86_64-linux-gnu.so new file mode 100755 index 00000000..a28e9719 Binary files /dev/null and b/fpemu_cuda.cpython-38-x86_64-linux-gnu.so differ diff --git a/mqbench/convert_deploy.py b/mqbench/convert_deploy.py index c5fdb0b7..68e7c0df 100644 --- a/mqbench/convert_deploy.py +++ b/mqbench/convert_deploy.py @@ -1,8 +1,10 @@ +import json import os.path as osp - +import os import torch from torch.fx import GraphModule - +import onnx +from onnxsim import simplify import mqbench.custom_symbolic_opset # noqa: F401 import mqbench.fusion_method # noqa: F401 from mqbench.prepare_by_platform import BackendType @@ -18,6 +20,8 @@ remove_fakequantize_and_collect_params, replace_fakequantize_and_collect_params_openvino, remove_fakequantize_and_collect_params_tengine, + remove_fakequantize_and_collect_params_sophgo, + # remove_fakequantize_and_collect_params_academic, ONNXQLinearPass, ONNXQNNPass ) @@ -32,15 +36,19 @@ @register_deploy_function(BackendType.NNIE) @register_deploy_function(BackendType.Vitis) @register_deploy_function(BackendType.OPENVINO) +@register_deploy_function(BackendType.Sophgo_TPU) def convert_merge_bn(model: GraphModule, **kwargs): + # print('wlog before convert_merge_bn, model.named_modules:', dict(model.named_modules())['']) + # print('wlog before convert_merge_bn, model.graph:', model.graph) logger.info("Merge BN for deploy.") nodes = list(model.graph.nodes) modules = dict(model.named_modules()) for node in nodes: if node.op == 'call_module': - if type(modules[node.target]) in FUSED_MODULE_CONVERT_FUNCTION: + if node.target in modules and type(modules[node.target]) in FUSED_MODULE_CONVERT_FUNCTION: FUSED_MODULE_CONVERT_FUNCTION[type(modules[node.target])](model, node) - + # print('wlog after convert_merge_bn, model.named_modules:', dict(model.named_modules())['']) + # print('wlog after convert_merge_bn, model.graph:', model.graph) @register_deploy_function(BackendType.Academic_NLP) @register_deploy_function(BackendType.Tensorrt_NLP) @@ -54,8 +62,13 @@ def convert_merge_bn(model: GraphModule, **kwargs): @register_deploy_function(BackendType.NNIE) @register_deploy_function(BackendType.Vitis) @register_deploy_function(BackendType.OPENVINO) +@register_deploy_function(BackendType.Sophgo_TPU) def convert_onnx(model: GraphModule, input_shape_dict, dummy_input, onnx_model_path, **kwargs): - logger.info("Export to onnx.") + pt_file_name = onnx_model_path.split('.') + pt_file_name[-1] = 'pt' + #torch.save(model, '.'.join(pt_file_name)) + logger.info("Export to onnx, onnx_model_path:{}".format(onnx_model_path)) + model = model.cpu() output_names = kwargs.get('output_names', []) dynamic_axes = kwargs.get('dynamic_axes', {}) input_names = kwargs.get('input_names', []) @@ -76,18 +89,26 @@ def convert_onnx(model: GraphModule, input_shape_dict, dummy_input, onnx_model_p opset_version=opset_version, dynamic_axes=dynamic_axes, do_constant_folding=True, - custom_opsets={'' : opset_version}) + custom_opsets={'MQBench_custom' : opset_version}) except ONNXCheckerError: pass except ImportError: + ### torch 1.13 and 2.0.1 torch.onnx.export(model, dummy_input, onnx_model_path, - input_names=input_names, - output_names=output_names, - opset_version=opset_version, - do_constant_folding=True, - custom_opsets={'' : opset_version}, - enable_onnx_checker=False) + input_names=input_names, + output_names=output_names, + opset_version=opset_version, + do_constant_folding=True, + custom_opsets={'MQBench_custom' : opset_version}) + tmp_model = onnx.load(onnx_model_path) + simplified_model, check = simplify(tmp_model) + onnx.save_model(simplified_model, onnx_model_path) + model_onnx = onnx.load(onnx_model_path) + onnx.checker.check_model(model_onnx) + model_onnx = onnx.shape_inference.infer_shapes(model_onnx) + os.system(f"rm -f {onnx_model_path}") + onnx.save(model_onnx, onnx_model_path) @register_deploy_function(BackendType.Tensorrt) def convert_onnx_qlinear(model: GraphModule, onnx_model_path, model_name, **kwargs): @@ -112,6 +133,100 @@ def deploy_qparams_openvino(model: GraphModule, onnx_model_path, model_name, **k def deploy_qparams_tensorrt(model: GraphModule, onnx_model_path, model_name, **kwargs): logger.info("Extract qparams for TensorRT.") remove_fakequantize_and_collect_params(onnx_model_path, model_name, backend='tensorrt') +#2023.7.27修改 +@register_deploy_function(BackendType.Academic_NLP) +def deploy_qparams_Academic_NLP(model: GraphModule, onnx_model_path, model_name, **kwargs): + logger.info("Extract qparams for Academic_NLP.") + remove_fakequantize_and_collect_params(onnx_model_path, model_name, backend='Academic_NLP') + print("导出calitable") + output_path = osp.dirname(onnx_model_path) + context_filename = osp.join(output_path, '{}_clip_ranges.json'.format(model_name)) + file_h = open(context_filename, "r") + blob_range = json.loads(file_h.read())["Academic_NLP"] + file_h.close() + cali_table = osp.join(output_path, '{}_cali_table_from_mqbench_Academic_NLP'.format(model_name)) + work_mode = kwargs.get('work_mode', 'QAT_all_int8') + if work_mode not in ['QAT_all_int8', 'int4_and_int8_mix', 'int4_and_int8_mix_no_fc']: + print('QAT_all_int8 not in [QAT_all_int8, int4_and_int8_mix, int4_and_int8_mix_no_fc],set to QAT_all_int8') + work_mode = 'QAT_all_int8' + with open(cali_table, 'w') as f: + f.write(f"# work_mode:{work_mode} #Automatically generated, do not modify, work_mode choice:[QAT_all_int8, int4_and_int8_mix, int4_and_int8_mix_no_fc]\n") + f.write("# op_name threshold min max\n") + weight_scale = [] + int4_th = [] + for name,value in blob_range.items(): + if 'threshold' in value: + tmpstr = "{} {:.7f} {:.7f} {:.7f}\n".format(name[:-2], value['threshold'], value['min'], value['max']) + if name.endswith('_4'): + int4_th.append(tmpstr) + elif name.endswith('_8'): + f.write(tmpstr) + else: + f.write("{} {:.7f} {:.7f} {:.7f}\n".format(name, value['threshold'], value['min'], value['max'])) + else: + tmpstr = "{} {} {} {} {}\n".format(name, len(value['step']), ' '.join([str(i) for i in value['step']]), + len(value['zero_point']), ' '.join([str(i) for i in value['zero_point']])) + if name.endswith('_weight') or name.endswith('_bias'): + weight_scale.append(tmpstr) + else: + f.write(tmpstr) + f.write('#int4_th\n') + for i in int4_th: + f.write(i) + f.write('#weight_scale\n') + for i in weight_scale: + f.write(i) + print("导出qtable") + file_h = open(context_filename, "r") + blob_range = json.loads(file_h.read())["Academic_NLP"] + file_h.close() + q_table = osp.join(output_path, '{}_q_table_from_mqbench_Academic_NLP'.format(model_name)) + with open(q_table, 'w') as f: + f.write("# op_name bit type\n") + for name,value in blob_range.items(): + f.write("{} {} {} \n".format(name, value['bit'], value['type'])) + +@register_deploy_function(BackendType.Sophgo_TPU) +def deploy_qparams_sophgo_tpu(model: GraphModule, onnx_model_path, model_name, **kwargs): + logger.info("Extract qparams for sophgo_tpu.") + remove_fakequantize_and_collect_params_sophgo(onnx_model_path, model_name) + output_path = osp.dirname(onnx_model_path) + context_filename = osp.join(output_path, '{}_clip_ranges.json'.format(model_name)) + file_h = open(context_filename, "r") + blob_range = json.loads(file_h.read())["sophgo_tpu"] + file_h.close() + cali_table = osp.join(output_path, '{}_cali_table_from_mqbench_sophgo_tpu'.format(model_name)) + work_mode = kwargs.get('work_mode', 'QAT_all_int8') + if work_mode not in ['QAT_all_int8', 'int4_and_int8_mix', 'int4_and_int8_mix_no_fc']: + print('QAT_all_int8 not in [QAT_all_int8, int4_and_int8_mix, int4_and_int8_mix_no_fc],set to QAT_all_int8') + work_mode = 'QAT_all_int8' + with open(cali_table, 'w') as f: + f.write(f"# work_mode:{work_mode} #Automatically generated, do not modify, work_mode choice:[QAT_all_int8, int4_and_int8_mix, int4_and_int8_mix_no_fc]\n") + f.write("# op_name threshold min max\n") + weight_scale = [] + int4_th = [] + for name,value in blob_range.items(): + if 'threshold' in value: + tmpstr = "{} {:.7f} {:.7f} {:.7f}\n".format(name[:-2], value['threshold'], value['min'], value['max']) + if name.endswith('_4'): + int4_th.append(tmpstr) + elif name.endswith('_8'): + f.write(tmpstr) + else: + f.write("{} {:.7f} {:.7f} {:.7f}\n".format(name, value['threshold'], value['min'], value['max'])) + else: + tmpstr = "{} {} {} {} {}\n".format(name, len(value['step']), ' '.join([str(i) for i in value['step']]), + len(value['zero_point']), ' '.join([str(i) for i in value['zero_point']])) + if name.endswith('_weight') or name.endswith('_bias'): + weight_scale.append(tmpstr) + else: + f.write(tmpstr) + f.write('#int4_th\n') + for i in int4_th: + f.write(i) + f.write('#weight_scale\n') + for i in weight_scale: + f.write(i) @register_deploy_function(BackendType.Vitis) diff --git a/mqbench/custom_quantizer/__init__.py b/mqbench/custom_quantizer/__init__.py index 94646a04..a8723d27 100644 --- a/mqbench/custom_quantizer/__init__.py +++ b/mqbench/custom_quantizer/__init__.py @@ -6,4 +6,5 @@ from .tensorrt_quantizer import TRTModelQuantizer, TensorrtNLPQuantizer from .tengine_u8_quantizer import TengineQuantizer from .onnx_qnn_quantizer import ONNXQNNQuantizer -from .nlp_quantizer import AcademicNLPQuantizer \ No newline at end of file +from .nlp_quantizer import AcademicNLPQuantizer +from .sophgo_tpu_quantizer import SophgoTpuQuantizer diff --git a/mqbench/custom_quantizer/academic_quantizer.py b/mqbench/custom_quantizer/academic_quantizer.py index 923d7834..30a50a52 100644 --- a/mqbench/custom_quantizer/academic_quantizer.py +++ b/mqbench/custom_quantizer/academic_quantizer.py @@ -6,7 +6,7 @@ import torch from torch.fx import GraphModule from torch.quantization import propagate_qconfig_ -from torch.quantization.fx.qconfig_utils import get_flattened_qconfig_dict +from mqbench.utils import get_flattened_qconfig_dict from mqbench.utils import is_symmetric_quant, getitem2node from mqbench.utils.logger import logger diff --git a/mqbench/custom_quantizer/model_quantizer.py b/mqbench/custom_quantizer/model_quantizer.py index 8a4a64bb..a42180a6 100644 --- a/mqbench/custom_quantizer/model_quantizer.py +++ b/mqbench/custom_quantizer/model_quantizer.py @@ -23,18 +23,20 @@ from torch.quantization.utils import ( get_combined_dict ) -from torch.quantization.fx.qconfig_utils import ( +from torch.quantization.fx.qconfig_utils import( get_flattened_qconfig_dict ) from torch.quantization.quantize_fx import ( _fuse_fx ) +from mqbench.utils import get_flattened_qconfig_dict from mqbench.utils import getitem2node from mqbench.utils.logger import logger from mqbench.utils.registry import register_model_quantizer from mqbench.prepare_by_platform import BackendType - +import mqbench.nn.intrinsic.qat as qnniqat +import mqbench.nn.qat as qnnqat @register_model_quantizer(BackendType.Tensorrt) @register_model_quantizer(BackendType.NNIE) @@ -76,7 +78,7 @@ def _insert_fake_quantize_for_act_quant( quantizer_prefix = "_post_act_fake_quantizer" node_to_quantize_output = self._find_act_quants(model) node_to_quantize_output = OrderedDict.fromkeys(node_to_quantize_output).keys() - + qconfig = qconfig[''] for node in node_to_quantize_output: fake_quantizer = qconfig.activation() quantizer_name = node.name + quantizer_prefix @@ -121,7 +123,8 @@ def _fix_succ_recursivly(self, args, target_node, inserted_node): def _weight_quant(self, model: GraphModule, qconfig): logger.info("Replace module to qat module.") - flattened_qconfig_dict = get_flattened_qconfig_dict({'': qconfig}) + flattened_qconfig_dict = get_flattened_qconfig_dict(qconfig)#torch�ӿ� + print('flattened_qconfig_dict:', flattened_qconfig_dict) propagate_qconfig_(model, flattened_qconfig_dict) self._qat_swap_modules(model, self.additional_qat_module_mapping) return model @@ -232,7 +235,7 @@ def _find_act_quants(self, model: GraphModule) -> List: if (node.op == "call_module" and isinstance(modules[node.target], self.module_type_to_quant_input)) or \ ((node.op == 'call_function' or node.op == 'call_method') and node.target in self.function_type_to_quant_input) or node.name in self.additional_node_name: - input_node_list = self._flatten_args(node.args) + input_node_list = self._flatten_args(node.all_input_nodes) # Means this is not Tensor + Tensor. if not all([isinstance(_node, torch.fx.node.Node) for _node in input_node_list]): continue @@ -240,6 +243,9 @@ def _find_act_quants(self, model: GraphModule) -> List: if self._is_implicit_merge(modules, (node, _node)): logger.info("Implicit merge: {} + {}".format(_node.name, node.name)) continue + if _node.op == "placeholder" and 'tensor_meta' in node.meta: + if len(_node.meta['tensor_meta'].shape) == 1: + continue if _node in node_need_to_quantize_output: continue if _node in g2node: diff --git a/mqbench/custom_quantizer/nlp_quantizer.py b/mqbench/custom_quantizer/nlp_quantizer.py index 69251f99..1040b070 100644 --- a/mqbench/custom_quantizer/nlp_quantizer.py +++ b/mqbench/custom_quantizer/nlp_quantizer.py @@ -10,7 +10,7 @@ class AcademicNLPQuantizer(ModelQuantizer): """ NLP model quantizer for Academic settings. Should not de 8bit for first / last layer. - We should uantize Linear / Embedding weights. + We should quantize Linear / Embedding weights. Linear / Matmul layer inputs(activations). """ @property diff --git a/mqbench/custom_quantizer/openvino_quantizer.py b/mqbench/custom_quantizer/openvino_quantizer.py index 1509b832..46453d93 100644 --- a/mqbench/custom_quantizer/openvino_quantizer.py +++ b/mqbench/custom_quantizer/openvino_quantizer.py @@ -6,9 +6,9 @@ import torch from torch.fx import GraphModule from torch.quantization import propagate_qconfig_ -from torch.quantization.fx.qconfig_utils import get_flattened_qconfig_dict from torch.quantization.quantize_fx import _fuse_fx +from mqbench.utils import get_flattened_qconfig_dict from mqbench.utils import is_symmetric_quant from mqbench.utils.logger import logger from mqbench.utils.registry import register_model_quantizer diff --git a/mqbench/custom_quantizer/sophgo_tpu_quantizer.py b/mqbench/custom_quantizer/sophgo_tpu_quantizer.py new file mode 100644 index 00000000..d60b8557 --- /dev/null +++ b/mqbench/custom_quantizer/sophgo_tpu_quantizer.py @@ -0,0 +1,334 @@ +import operator +import torch +from torch.fx import GraphModule +import torch.nn.intrinsic as nni +import mqbench.nn.intrinsic.qat as qnniqat +import mqbench.nn.intrinsic as qnni +from mqbench.utils.registry import register_model_quantizer +from mqbench.prepare_by_platform import BackendType +from mqbench.custom_quantizer import ModelQuantizer +import torch.nn as nn +import mqbench.nn.qat as qnnqat +from collections import OrderedDict +from mqbench.utils.logger import logger + +from typing import ( + List, Dict, Any, Callable +) + +from mqbench.utils import get_flattened_qconfig_dict + +@register_model_quantizer(BackendType.Sophgo_TPU) +class SophgoTpuQuantizer(ModelQuantizer): + """There is only INT8 calculations in the model. + We quantize the input tensors and output tensors of all layers, + except those in _passed_func_type and _passed_module_type. + For example add + relu pattern, there is no need to insert fake + quantize node between them. + """ + + def __init__(self, extra_quantizer_dict, extra_fuse_dict): + super().__init__(extra_quantizer_dict, extra_fuse_dict) + self.additional_qat_module_mapping = { + # Intrinsic modules: + nni.ConvBn2d: qnniqat.ConvBn2d_sophgo, + nni.ConvBnReLU2d: qnniqat.ConvBnReLU2d_sophgo, + nn.Conv2d: qnnqat.Conv2d_sophgo, + nni.ConvReLU2d: qnniqat.ConvReLU2d_sophgo, + nni.LinearReLU: qnniqat.LinearReLU_sophgo, + nn.Linear: qnniqat.Linear_sophgo, + qnni.LinearBn1d: qnniqat.LinearBn1d_sophgo, + qnni.ConvTransposeBnReLU2d:qnniqat.ConvTransposeBnReLU2d_sophgo, + qnni.ConvTransposeReLU2d:qnniqat.ConvTransposeReLU2d_sophgo, + qnni.ConvTransposeBn2d:qnniqat.ConvTransposeBn2d_sophgo, + } + self.exclude_module_name.append(nn.modules.dropout.Dropout) + + @property + def module_type_to_quant_input(self) -> tuple: + return ( + torch.nn.Hardswish, + torch.nn.Sigmoid, + torch.nn.SiLU, + torch.nn.Tanh, + torch.nn.SELU, + torch.nn.LogSigmoid, + torch.nn.GELU, + torch.nn.GLU, + torch.nn.Mish, + torch.nn.Hardsigmoid, + torch.nn.Softshrink, + torch.nn.Softplus, + torch.nn.ELU, + torch.nn.CELU, + ) + super().module_type_to_quant_input + self._layers_need_scale_form_input_fake_quantizer + + @property + def function_type_to_quant_input(self) -> tuple: + return super().function_type_to_quant_input + [ + operator.sub, + operator.abs, + torch.cat, + torch.sub, + torch.clamp, + torch.nn.functional.hardswish, + torch.nn.functional.sigmoid, + torch.nn.functional.silu, + torch.nn.functional.tanh, + torch.nn.functional.selu, + torch.nn.functional.logsigmoid, + torch.nn.functional.gelu, + torch.nn.functional.glu, + torch.nn.functional.mish, + torch.nn.functional.hardsigmoid, + torch.nn.functional.gumbel_softmax, + torch.nn.functional.softshrink, + torch.nn.functional.softplus, + torch.nn.functional.elu, + torch.nn.functional.celu, + ] + + @property + def _passed_func_type(self): + return ( + torch.nn.functional.relu, + # torch.nn.functional.relu6, + torch.flatten + ) + + @property + def _passed_module_type(self): + return ( + torch.nn.ReLU, + torch.nn.ReLU6 + ) + + @property + def _layers_need_scale_form_input_fake_quantizer(self): + return ( + qnniqat.ConvBnReLU2d_sophgo, #todo:add transposeConv support + qnniqat.ConvBn2d_sophgo, + qnniqat.ConvReLU2d_sophgo, + qnnqat.Conv2d_sophgo, + qnniqat.LinearReLU_sophgo, + qnniqat.Linear_sophgo, + qnniqat.LinearBn1d_sophgo, + qnniqat.ConvTransposeBnReLU2d_sophgo, + qnniqat.ConvTransposeReLU2d_sophgo, + qnniqat.ConvTransposeBn2d_sophgo, + ) + + @property + def _layers_need_check_is_dw(self): + return ( + qnniqat.ConvBnReLU2d_sophgo, + qnniqat.ConvBn2d_sophgo, + qnniqat.ConvReLU2d_sophgo, + qnnqat.Conv2d_sophgo, + ) + + def _insert_fake_quantizer(self, model, graph, modules, flattened_qconfig_dict, node, int84_layers): + if len(int84_layers) > 0: + layer = int84_layers[0] #多个int4或int8节点选其中第1个就能正确决定节点类型 + qconfig1 = flattened_qconfig_dict.get(layer.target, None) #首先根据层名去取，优先级最高 + if qconfig1 is None and layer.target in modules: + print(f'layer.target:{layer.target}, type:',type(modules[layer.target])) + qconfig1 = flattened_qconfig_dict.get(type(modules[layer.target]), None) #其次根据type去取 + if isinstance(modules[layer.target], self._layers_need_check_is_dw): + if modules[layer.target].groups > 1: + qconfig1 = None #深度卷积使用int8计算 + if qconfig1 is None: + qconfig1 = flattened_qconfig_dict.get('', None) #最后找全局qconfig，优先级最低 + fake_quantizer = qconfig1.activation() + quantizer_name = layer.name + self.quantizer_prefix + if hasattr(model, quantizer_name): + quantizer_name = layer.name +'_n2_br'+ self.quantizer_prefix + setattr(model, quantizer_name, fake_quantizer) + logger.info("Insert act quant {}".format(quantizer_name)) + with graph.inserting_after(node): + inserted_node = graph.create_node("call_module", quantizer_name, (node,), {}) + for int84_layer in int84_layers: + int84_layer.args = self._fix_succ_recursivly(int84_layer.args, node, inserted_node) + + def _insert_fake_quantize_for_act_quant( + self, + model: GraphModule, + qconfig: Any): + graph = model.graph + nodes = list(model.graph.nodes) + modules = dict(model.named_modules()) + + self.quantizer_prefix = "_input_act_fake_quantizer" + node_to_quantize_output = self._find_act_quants(model) + node_to_quantize_output = OrderedDict.fromkeys(node_to_quantize_output).keys() + flattened_qconfig_dict = get_flattened_qconfig_dict(qconfig) + print('node_to_quantize_output:', node_to_quantize_output, 'flattened_qconfig_dict:', flattened_qconfig_dict) + int4_and_int8_mix = False + for m in flattened_qconfig_dict: + if not isinstance(flattened_qconfig_dict[m].activation(), torch.nn.Identity) and flattened_qconfig_dict[m].activation().bitwidth == 4: + int4_and_int8_mix = True + break + print('int4_and_int8_mix:', int4_and_int8_mix) + if int4_and_int8_mix: + def all_next_layers_is_trivial(node): + if len(node.users) == 0: + return True + for user in node.users: + if not all_next_layers_is_trivial(user): + return False + if node.op != 'call_method' or node.target not in ['view', 'permute', 'contiguous']: + return False + return True + + def find_next_int4_and_int8_layers(node, int8_layers, int4_layers): + for user in node.users: #若后继有1个层或多个不同类型的层，则插入多个input量化节点 todo:多个节点，部分相同，部分不同 + if user.target in modules and type(modules[user.target]) in self.exclude_module_name: + print(f'user:{user.name} is excluded') + user.replace_all_uses_with(node) + graph.erase_node(user) + del modules[user.target] + find_next_int4_and_int8_layers(user, int8_layers, int4_layers) + continue #dropout等层前不要插入伪量化节点 + if user.op == "call_module" and isinstance(modules[user.target], self._layers_need_scale_form_input_fake_quantizer): + if isinstance(modules[user.target], self._layers_need_check_is_dw): + if modules[user.target].groups > 1: + int8_layers.append(user) + continue + int4_layers.append(user) + else: + if not all_next_layers_is_trivial(user): + int8_layers.append(user) + + for node in node_to_quantize_output: + int8_layers, int4_layers = [],[] #找到node后的多个int4后继节点和多个int8后继节点，然后这多个int8或int4后继节点共享1个输入量化节点 + find_next_int4_and_int8_layers(node, int8_layers, int4_layers) + print(f'node:{node}, int4_layers:', int4_layers, 'int8_layers:', int8_layers) + self._insert_fake_quantizer(model, graph, modules, flattened_qconfig_dict, node, int4_layers) + self._insert_fake_quantizer(model, graph, modules, flattened_qconfig_dict, node, int8_layers) + model.recompile() + model.graph.lint() + graph = model.graph + nodes = list(model.graph.nodes) + modules = dict(model.named_modules()) + + for node in node_to_quantize_output: + if node.op == 'placeholder' and int4_and_int8_mix: + continue + qconfig2 = flattened_qconfig_dict.get(node.target, None) #首先根据层名去取，优先级最高 + if qconfig2 is None and node.target in modules: + qconfig2 = flattened_qconfig_dict.get(type(modules[node.target]), None) #其次根据type去取 + if isinstance(modules[node.target], self._layers_need_check_is_dw): + if modules[node.target].groups > 1: + qconfig2 = None #深度卷积使用int8计算 + if qconfig2 is None: + qconfig2 = flattened_qconfig_dict.get('', None) #最后找全局qconfig，优先级最低 + if int4_and_int8_mix: + if node.target in modules and type(modules[node.target]) == torch.nn.ReLU6 and node.args[0].target in modules: + qconfig2 = flattened_qconfig_dict.get(type(modules[node.args[0].target]), None) + if isinstance(modules[node.args[0].target], self._layers_need_check_is_dw): + if modules[node.args[0].target].groups > 1: + qconfig2 = flattened_qconfig_dict.get('', None) ##深度卷积使用int8计算 + node_fake_quantizer = qconfig2.activation() + # node_fake_quantizer.enable_only_observer() + quantizer_name2 = node.name + "_post_act_fake_quantizer" + setattr(model, quantizer_name2, node_fake_quantizer) + with graph.inserting_after(node): + inserted_node = graph.create_node("call_module", quantizer_name2, (node,), {}) + for _node in nodes: + _node.args = self._fix_succ_recursivly(_node.args, node, inserted_node) + + if int4_and_int8_mix: + model.recompile() + model.graph.lint() + graph = model.graph + modules = dict(model.named_modules()) + nodes = list(model.graph.nodes) + for node in nodes: + if "_post_act_fake_quantizer" in node.name: + #post量化节点的下一个是input量化节点，且input量化节点的后面只有1个节点，且该节点后再无后继节点，此时删除这个input量化节点，用于删除网络输出的最后1个多余的量化节点 + for user in list(node.users.keys()): + if "_input_act_fake_quantizer" in user.name: + users = list(user.users.keys()) + if len(users) == 1 and len(users[0].users) == 0: + user.replace_all_uses_with(node) + graph.erase_node(user) + del modules[user.target] + if modules[node.target].bitwidth == 8: #2个相邻的量化节点都为8bit量化，则删除冗余的后1个节点 + for user in list(node.users.keys()): + if "_input_act_fake_quantizer" in user.name: + if modules[user.target].bitwidth == 8: + user.replace_all_uses_with(node) + graph.erase_node(user) + del modules[user.target] + elif modules[node.target].bitwidth == 4: #2个相邻的量化节点都为4bit量化，则删除冗余的后1个节点 + for user in list(node.users.keys()): + if "_input_act_fake_quantizer" in user.name: + if modules[user.target].bitwidth == 4: + user.replace_all_uses_with(node) + graph.erase_node(user) + del modules[user.target] + model.recompile() + model.graph.lint() + return model + + def prepare(self, model: GraphModule, qconfig): + model = super().prepare(model, qconfig) + model = self._set_fake_quantizer_to_next_weight_layer(model) + return model + + def _find_act_quants(self, model: GraphModule) -> list: + nodes = list(model.graph.nodes) + modules = dict(model.named_modules()) + node_need_to_quantize_output = super()._find_act_quants(model) + self.only_enable_ob = [] + for node in nodes: + if (node.op == "call_module" and node.target in self.exclude_module_name) or \ + ((node.op == 'call_function' or node.op == 'call_method') and + node.target in self.exclude_function_type) or \ + node.name in self.exclude_node_name: + continue + if (node.op == "call_module" and isinstance(modules[node.target], self.module_type_to_quant_input)) or \ + ((node.op == 'call_function' or node.op == 'call_method') and + node.target in self.function_type_to_quant_input): + for next_node in node.users: + if ((next_node.op == 'call_function' and next_node.target in self._passed_func_type) or + (next_node.op == 'call_module' and isinstance(modules[next_node.target], self._passed_module_type))): + if next_node not in node_need_to_quantize_output: + node_need_to_quantize_output.append(next_node) + self.only_enable_ob.append(next_node.name) + else: + if node not in node_need_to_quantize_output: + node_need_to_quantize_output.append(node) + self.only_enable_ob.append(node.name) + for node in nodes: + if node.target in modules and type(modules[node.target]) in self.exclude_module_name: + print(f'{type(modules[node.target])} is excluded') + node_need_to_quantize_output.remove(node) + if node.op == "placeholder": + if 'tensor_meta' in node.meta: + if len(node.meta['tensor_meta'].shape) > 1: + print(f'add placeholder {node.target} to node_need_to_quantize_output by tensor_meta') + node_need_to_quantize_output.append(node) + else: + print(f'no tensor_meta, add placeholder {node.target} to node_need_to_quantize_output') + node_need_to_quantize_output.append(node) + + return node_need_to_quantize_output + + def _set_fake_quantizer_to_next_weight_layer(self, model: GraphModule): + nodes = list(model.graph.nodes) + modules = dict(model.named_modules()) + for node in nodes: + if node.target in modules and (self.quantizer_prefix in node.target or "_post_act_fake_quantizer" in node.target): + fake_quantizer = getattr(model, node.target) + for user in node.users: + if (user.op == "call_module" and isinstance(modules[user.target], self._layers_need_scale_form_input_fake_quantizer)): + setattr(modules[user.target], "input_fake_quantizer", fake_quantizer) + print('wlog:', user.target,'\'type is:', type(modules[user.target]), "add input_fake_quantizer") + if user.target in modules and type(modules[user.target]) in self.exclude_module_name: + for user2 in user.users: + if (user2.op == "call_module" and isinstance(modules[user2.target], self._layers_need_scale_form_input_fake_quantizer)): + setattr(modules[user2.target], "input_fake_quantizer", fake_quantizer) + print('wlog:', user2.target,'\'type is:', type(modules[user2.target]), "add input_fake_quantizer") + + return model diff --git a/mqbench/custom_symbolic_opset.py b/mqbench/custom_symbolic_opset.py index 6fcb1f28..480c08b2 100644 --- a/mqbench/custom_symbolic_opset.py +++ b/mqbench/custom_symbolic_opset.py @@ -1,23 +1,38 @@ from torch.onnx import register_custom_op_symbolic +from torch.onnx import symbolic_helper -# Register symbolic op for torch.quantize_function op. - +get_size = symbolic_helper._get_tensor_sizes def _fake_quantize_learnable_per_tensor_affine(g, x, scale, zero_point, quant_min, quant_max, grad_factor): - return g.op("::LearnablePerTensorAffine", x, scale, zero_point, quant_min, quant_max) - + output = g.op("MQBench_custom::LearnablePerTensorAffine", x, scale, zero_point, quant_min, quant_max) + input_shape = symbolic_helper._get_tensor_sizes(x) -register_custom_op_symbolic('::_fake_quantize_learnable_per_tensor_affine', _fake_quantize_learnable_per_tensor_affine, 11) + if input_shape is not None and hasattr(x.type(), 'with_sizes'): + output_type = x.type().with_sizes(input_shape) + output.setType(output_type) + return output def fake_quantize_per_channel_affine(g, x, scale, zero_point, ch_axis, quant_min, quant_max): - return g.op("::FixedPerChannelAffine", x, scale, zero_point, ch_axis, quant_min, quant_max) + output = g.op("MQBench_custom::FixedPerChannelAffine", x, scale, zero_point, ch_axis, quant_min, quant_max) + input_shape = symbolic_helper._get_tensor_sizes(x) + if input_shape is not None and hasattr(x.type(), 'with_sizes'): + output_type = x.type().with_sizes(input_shape) + output.setType(output_type) -register_custom_op_symbolic('::fake_quantize_per_channel_affine', fake_quantize_per_channel_affine, 11) - + return output def fake_quantize_per_tensor_affine(g, x, scale, zero_point, quant_min, quant_max): - return g.op("::FixedPerTensorAffine", x, scale, zero_point, quant_min, quant_max) + output = g.op("MQBench_custom::FixedPerTensorAffine", x, scale, zero_point, quant_min, quant_max) + input_shape = symbolic_helper._get_tensor_sizes(x) + + if input_shape is not None and hasattr(x.type(), 'with_sizes'): + output_type = x.type().with_sizes(input_shape) + output.setType(output_type) + return output -register_custom_op_symbolic('::fake_quantize_per_tensor_affine', fake_quantize_per_tensor_affine, 11) \ No newline at end of file +for i in range(1, 16): + register_custom_op_symbolic('::_fake_quantize_learnable_per_tensor_affine', _fake_quantize_learnable_per_tensor_affine, i) + register_custom_op_symbolic('::fake_quantize_per_channel_affine', fake_quantize_per_channel_affine, i) + register_custom_op_symbolic('::fake_quantize_per_tensor_affine', fake_quantize_per_tensor_affine, i) \ No newline at end of file diff --git a/mqbench/deploy/__init__.py b/mqbench/deploy/__init__.py index 9aed37ea..94137001 100644 --- a/mqbench/deploy/__init__.py +++ b/mqbench/deploy/__init__.py @@ -4,3 +4,5 @@ from .deploy_onnx_qnn import ONNXQNNPass from .deploy_openvino import replace_fakequantize_and_collect_params_openvino from .deploy_tengine import remove_fakequantize_and_collect_params_tengine +from .deploy_sophgo import remove_fakequantize_and_collect_params_sophgo +#from .deploy_academicnlp import remove_fakequantize_and_collect_params_academic \ No newline at end of file diff --git a/mqbench/deploy/common.py b/mqbench/deploy/common.py index 54dab486..ec1518d7 100644 --- a/mqbench/deploy/common.py +++ b/mqbench/deploy/common.py @@ -190,7 +190,7 @@ def remove_fake_pad_op(self, graph, name2data, inp2node, out2node): return - +#输出tensor到节点，输入tensor到节点及该输入序号 def update_inp2node_out2node(graph): out2node = {} inp2node = {} @@ -230,18 +230,25 @@ def parse_attrs(node_attrs): for attr in node_attrs: if attr.type == onnx.AttributeProto.AttributeType.INTS: attrs[attr.name] = tuple(attr.ints) + attrs['dtype']='ints' elif attr.type == onnx.AttributeProto.AttributeType.INT: attrs[attr.name] = attr.i + attrs['dtype']='int' elif attr.type == onnx.AttributeProto.AttributeType.FLOATS: attrs[attr.name] = tuple(attr.floats) + attrs['dtype']='floats' elif attr.type == onnx.AttributeProto.AttributeType.FLOAT: attrs[attr.name] = attr.f + attrs['dtype']='float' elif attr.type == onnx.AttributeProto.AttributeType.TENSOR: attrs[attr.name] = numpy_helper.to_array(attr.t) + attrs['dtype']='t' elif attr.type == onnx.AttributeProto.AttributeType.STRING: attrs[attr.name] = str(attr.s) + attrs['dtype']='st' elif attr.type == onnx.AttributeProto.AttributeType.STRINGS: attrs[attr.name] = tuple([str(x) for x in attr.strings]) + attrs['dtype']='none' else: raise Exception("ATTR Type [{}] Not Supported!".format(attr.type)) return attrs diff --git a/mqbench/deploy/deploy_linear.py b/mqbench/deploy/deploy_linear.py index 8dd4deb3..f3dc4cfd 100644 --- a/mqbench/deploy/deploy_linear.py +++ b/mqbench/deploy/deploy_linear.py @@ -1,5 +1,6 @@ import json import os +import copy import onnx @@ -83,16 +84,22 @@ def parse_qparams(self, node, name2data): if len(node.input) > 3: qmin, qmax = node.input[-2:] qmin, qmax = name2data[qmin], name2data[qmax] + if len(node.attribute) > 0: + qparams = parse_attrs(node.attribute) + dtype1=qparams['dtype'] + else: + dtype1='None' elif len(node.attribute) > 0: qparams = parse_attrs(node.attribute) qmin = qparams['quant_min'] qmax = qparams['quant_max'] + dtype1=qparams['dtype'] else: logger.info(f'qmin and qmax are not found for <{node.name}>!') - return tensor_name, scale, zero_point, qmin, qmax + return tensor_name, scale, zero_point, qmin, qmax,dtype1 def clip_weight(self, node, name2data, inp2node, named_initializer): - tensor_name, scale, zero_point, qmin, qmax = self.parse_qparams(node, name2data) + tensor_name, scale, zero_point, qmin, qmax, dtype1= self.parse_qparams(node, name2data) data = name2data[tensor_name] clip_range_min = ((qmin - zero_point) * scale).astype(data.dtype) clip_range_max = ((qmax - zero_point) * scale).astype(data.dtype) @@ -114,11 +121,19 @@ def clip_weight(self, node, name2data, inp2node, named_initializer): logger.info(f'Clip weights <{tensor_name}> to range [{clip_range_min}, {clip_range_max}].') new_data = numpy_helper.from_array(new_data) named_initializer[tensor_name].raw_data = new_data.raw_data + def get_correct_sophgo_tpu_input_tensor_name(self, node, out2node): + input_0 = node.input[0] + tensor_name = '{}_{}'.format(input_0, out2node[input_0].op_type if input_0 in out2node else '') + if tensor_name[-1] == '_': + tensor_name = tensor_name[:-1] + return tensor_name - def post_process_clip_ranges(self, clip_ranges, graph, inp2node): + def post_process_clip_ranges(self, clip_ranges, graph, inp2node, out2node): def find_the_closest_clip_range(node): - if node.input[0] in clip_ranges: - return node.input[0] + tensor_name = self.get_correct_sophgo_tpu_input_tensor_name(node, out2node) + + if tensor_name in clip_ranges: + return tensor_name elif node.op_type in ['Flatten', 'Resize'] and node.output[0] in inp2node: return find_the_closest_clip_range(inp2node[node.output[0]][0][0]) else: @@ -128,9 +143,50 @@ def find_the_closest_clip_range(node): if node.op_type in ['Flatten', 'Resize']: tensor_name = find_the_closest_clip_range(node) if tensor_name: - clip_ranges[node.input[0]] = clip_ranges[tensor_name] + new_name = self.get_correct_sophgo_tpu_input_tensor_name(node, out2node) + clip_ranges[new_name] = copy.deepcopy(clip_ranges[tensor_name]) + clip_ranges[new_name]['ori_name'] = new_name logger.info(f'Pass <{tensor_name}> clip range to <{node.name}> input <{node.input[0]}>.') return clip_ranges + def post_process_clip_ranges2(self, clip_ranges, graph, inp2node, out2node): + op_type_inAndOutShouldSameClipRange = ['Flatten', 'Resize', 'Reshape', 'Transpose'] + for node in graph.node: + tensor_name = f'{node.output[0]}_{node.op_type}' + if tensor_name not in clip_ranges: + pre_op = node + finded = False + while pre_op.op_type in op_type_inAndOutShouldSameClipRange: + tensor_name2 = self.get_correct_sophgo_tpu_input_tensor_name(pre_op, out2node) + if tensor_name2 in clip_ranges: + finded = True + clip_ranges[tensor_name] = clip_ranges[tensor_name2] + print(f'pre_op finded, transfer {tensor_name2} to {tensor_name}') + break + if pre_op.input[0] in out2node: + pre_op = out2node[pre_op.input[0]] + else: + print(f'{pre_op.name}\'s pre_node not exist') + break + if not finded: + if node.output[0] in inp2node: + next_op = inp2node[node.output[0]][0][0] + while next_op.op_type in op_type_inAndOutShouldSameClipRange: + tensor_name2 = f'{next_op.output[0]}_{next_op.op_type}' + if tensor_name2 in clip_ranges: + finded = True + clip_ranges[tensor_name] = clip_ranges[tensor_name2] + print(f'next_op finded, transfer {tensor_name2} to {tensor_name}') + break + if next_op.output[0] in inp2node: + next_op = inp2node[next_op.output[0]][0][0] + else: + print(f'{next_op.name}\'s next_op not exist') + break + else: + print(f'{node.name}\'s next_op not exist') + # if not finded: + # print(f'Waring:{node.name}\'s clip_ranges not exist, maybe have some error') + return clip_ranges def remove_fakequantize_and_collect_params(self, onnx_path, model_name, backend): model = onnx.load(onnx_path) @@ -150,13 +206,17 @@ def remove_fakequantize_and_collect_params(self, onnx_path, model_name, backend) nodes_to_be_removed.append(node) nodes_to_be_removed.extend(get_constant_inputs(node, out2node)) + if node.output[0] not in inp2node: + assert node.output[0] in [l.name for l in graph.output] + inp2node[node.output[0]] = [] + next_nodes = inp2node[node.output[0]] if node.op_type in PERCHANNEL_FAKEQUANTIZER: # fake quantize for weights, suppose per-channel quantize only for weight redundant_nodes = self.deal_with_weight_fakequant(node, out2node, inp2node, named_initializer) nodes_to_be_removed.extend(redundant_nodes) self.clip_weight(node, name2data, inp2node, named_initializer) + tensor_name, scale, zero_point, qmin, qmax,dtype1 = self.parse_qparams(node, name2data) if backend == 'ppl': - tensor_name, scale, zero_point, qmin, qmax = self.parse_qparams(node, name2data) clip_ranges[tensor_name] = {'step': [float(x) for x in scale], 'zero_point': [int(x) for x in zero_point], 'min': [float(x) for x in scale * (qmin - zero_point)], @@ -167,30 +227,54 @@ def remove_fakequantize_and_collect_params(self, onnx_path, model_name, backend) elif backend == 'vitis': logger.info("Vitis-DPU does not support per-channel quatization.") raise NotImplementedError("Vitis-DPU does not support per-channel quatization.") - - + elif backend == 'sophgo_tpu': + #卷积权重per-channel量化参数，bias的per-chan量化参数没有去调优 + if len(next_nodes) == 1 and next_nodes[0][0].op_type in ['Gemm', 'Conv']:#当前伪量化节点只有1个后继，且第1个后继节点为conv类型 + next_node_output = next_nodes[0][0].output[0] + if inp2node[next_node_output][0][0].op_type == 'Relu':##伪量化节点的第1个后继conv节点的第1个后继节点为Relu(fake->conv->relu) + #若是fake->conv->relu,因为relu会融合到前面conv，故用relu的输出tensor名+Relu作为量化参数保存tensor名 + tensor_name = '{}_{}'.format(inp2node[next_node_output][0][0].output[0], 'Relu') + else: + #若是fake->conv->not_relu_type,直接用conv的输出tensor名+conv作为量化参数保存tensor名 + tensor_name = '{}_{}'.format(next_node_output, next_nodes[0][0].op_type) + tensor_name += '_{}'.format('weight' if next_nodes[0][1] == 1 else 'bias' ) + clip_ranges[tensor_name] = {'step': [float(x) for x in scale], + 'zero_point': [int(x) for x in zero_point] + } elif node.op_type in PERTENSOR_FAKEQUANTIZER: - if node.output[0] not in inp2node: - assert node.output[0] in [l.name for l in graph.output] - inp2node[node.output[0]] = [] - next_nodes = inp2node[node.output[0]] if len(next_nodes) == 1 and next_nodes[0][1] == 1 and next_nodes[0][0].op_type in ['Gemm', 'Conv']: # fake quantize for weights redundant_nodes = self.deal_with_weight_fakequant(node, out2node, inp2node, named_initializer) - tensor_name, scale, zero_point, qmin, qmax = self.parse_qparams(node, name2data) + tensor_name, scale, zero_point, qmin, qmax,dtype = self.parse_qparams(node, name2data) nodes_to_be_removed.extend(redundant_nodes) self.clip_weight(node, name2data, inp2node, named_initializer) + if backend == 'sophgo_tpu': + assert next_nodes[0][0].op_type == 'Gemm' + tensor_name += '{}_{}_weight'.format(inp2node[node.output[0]][0][0].output[0], inp2node[node.output[0]][0][0].op_type) + clip_ranges[tensor_name] = {'threshold':float(scale * max(-qmin, qmax)), #对称量化时这个参数生效 + 'min': float(scale * (qmin - zero_point)), + 'max': float(scale * (qmax - zero_point)), + 'ori_name': 'none'} + if backend == 'Academic_NLP': + assert next_nodes[0][0].op_type == 'Gemm' + tensor_name += '{}_{}_weight'.format(inp2node[node.output[0]][0][0].output[0], inp2node[node.output[0]][0][0].op_type) + clip_ranges[tensor_name] = {'threshold':float(scale * max(-qmin, qmax)), #对称量化时这个参数生效 + 'min': float(scale * (qmin - zero_point)), + 'max': float(scale * (qmax - zero_point)), + 'bit': int(np.log2(qmax - qmin + 1)), + 'type': "int" if int(np.log2(qmax - qmin + 1))==4 else dtype1, + 'ori_name': 'none'} elif len(next_nodes) == 1 and next_nodes[0][1] == 2 and next_nodes[0][0].op_type in ['Gemm', 'Conv']: # fake quantize for bias assert backend == 'vitis' redundant_nodes = self.deal_with_weight_fakequant(node, out2node, inp2node, named_initializer) - tensor_name, scale, zero_point, qmin, qmax = self.parse_qparams(node, name2data) + tensor_name, scale, zero_point, qmin, qmax,dtype1 = self.parse_qparams(node, name2data) nodes_to_be_removed.extend(redundant_nodes) self.clip_weight(node, name2data, inp2node, named_initializer) else: # fake quantize for activations self.deal_with_activation_fakequant(node, inp2node) - tensor_name, scale, zero_point, qmin, qmax = self.parse_qparams(node, name2data) + tensor_name, scale, zero_point, qmin, qmax,dtype1 = self.parse_qparams(node, name2data) for out in graph.output: if out.name == node.output[0]: out.name = tensor_name @@ -203,6 +287,26 @@ def remove_fakequantize_and_collect_params(self, onnx_path, model_name, backend) 'min': float(scale * (qmin - zero_point)), 'max': float(scale * (qmax - zero_point))} ] + elif backend == 'sophgo_tpu': + scale_name = node.input[1] + post_str = '_post_act_fake_quantizer.scale' + if tensor_name in out2node: + tensor_name += '_{}'.format(out2node[tensor_name].op_type) + clip_ranges[tensor_name] = {'threshold':float(scale * max(-qmin, qmax)), #对称量化时这个参数生效 + 'min': float(scale * (qmin - zero_point)), + 'max': float(scale * (qmax - zero_point)), + 'ori_name': scale_name[:len(scale_name)-len(post_str)] if scale_name.endswith(post_str) else 'none'} + elif backend == 'Academic_NLP': + scale_name = node.input[1] + post_str = '_post_act_fake_quantizer.scale' + if tensor_name in out2node: + tensor_name += '_{}'.format(out2node[tensor_name].op_type) + clip_ranges[tensor_name] = {'threshold':float(scale * max(-qmin, qmax)), #对称量化时这个参数生效 + 'min': float(scale * (qmin - zero_point)), + 'max': float(scale * (qmax - zero_point)), + 'bit': int(np.log2(qmax - qmin + 1)), + 'type': "int" if int(np.log2(qmax - qmin + 1))==4 else dtype1, + 'ori_name': scale_name[:len(scale_name)-len(post_str)] if scale_name.endswith(post_str) else 'none'} if backend == 'ppl': clip_ranges[tensor_name] = {'step': float(scale), 'zero_point': int(zero_point), @@ -225,8 +329,8 @@ def remove_fakequantize_and_collect_params(self, onnx_path, model_name, backend) if name in (out2node.keys() | inp2node.keys()): continue graph.initializer.remove(initial_data) - - clip_ranges = self.post_process_clip_ranges(clip_ranges, graph, inp2node) + + clip_ranges = self.post_process_clip_ranges(clip_ranges, graph, inp2node, out2node) if backend == 'tensorrt': context = {"tensorrt": {"blob_range": clip_ranges}} elif backend == 'snpe': @@ -237,6 +341,14 @@ def remove_fakequantize_and_collect_params(self, onnx_path, model_name, backend) context = {'vitis': clip_ranges} elif backend == 'ppl-cuda': context = {'ppl-cuda': clip_ranges} + elif backend == 'sophgo_tpu': + clip_ranges = self.post_process_clip_ranges2(clip_ranges, graph, inp2node, out2node) + context = {'sophgo_tpu': clip_ranges} + context['w_qscheme'] = '' + context['a_qscheme'] = '' + elif backend == 'Academic_NLP': + clip_ranges = self.post_process_clip_ranges2(clip_ranges, graph, inp2node, out2node) + context = {'Academic_NLP': clip_ranges} output_path = os.path.dirname(onnx_path) context_filename = os.path.join(output_path, '{}_clip_ranges.json'.format(model_name)) with open(context_filename, 'w') as f: diff --git a/mqbench/deploy/deploy_sophgo.py b/mqbench/deploy/deploy_sophgo.py new file mode 100644 index 00000000..6852393b --- /dev/null +++ b/mqbench/deploy/deploy_sophgo.py @@ -0,0 +1,355 @@ +import json +import os +import copy + + +import onnx +import numpy as np +from onnx import numpy_helper + +from mqbench.utils.logger import logger +from mqbench.deploy.common import ( + update_inp2node_out2node, + prepare_initializer, + prepare_data, + OnnxPreprocess, + get_constant_inputs, + parse_attrs +) + + +PERCHANNEL_FAKEQUANTIZER = ['FakeQuantizeLearnablePerchannelAffine', + 'FixedPerChannelAffine', + 'FakeQuantizeDSQPerchannel'] +PERTENSOR_FAKEQUANTIZER = ['LearnablePerTensorAffine', + 'FixedPerTensorAffine', + 'FakeQuantizeDSQPertensor', + 'FakeQuantizeTqtAffine'] +ALL_FAKEQUANTIZER = PERCHANNEL_FAKEQUANTIZER + PERTENSOR_FAKEQUANTIZER + + +class LinearQuantizer_process(object): + # some method like dorefa need pre-compute weights + def weight_preprocess(self, target_tensor, out2node, inp2node, named_initializer): + def find_weight(tensor): + if tensor not in named_initializer: + _node = out2node[tensor] + for inp in _node.input: + return find_weight(inp) + return tensor + weight = find_weight(target_tensor) + + # TODO need more general method, like onnxruntime infer + data = numpy_helper.to_array(named_initializer[weight]) + data = np.tanh(data) + data = data / (np.max(np.abs(data)) + 1e-5) + data = numpy_helper.from_array(data) + named_initializer[weight].raw_data = data.raw_data + + redundant_nodes = [] + + def find_redundant_nodes(tensor): + if tensor == target_tensor: + return + nodes = inp2node[tensor] + for node, idx in nodes: + if node not in redundant_nodes: + redundant_nodes.append(node) + redundant_nodes.extend(get_constant_inputs(node, out2node)) + find_redundant_nodes(node.output[0]) + find_redundant_nodes(weight) + return weight, redundant_nodes + + def deal_with_weight_fakequant(self, node, out2node, inp2node, named_initializer): + next_nodes = inp2node[node.output[0]] + assert len(next_nodes) == 1 + next_node, idx = next_nodes[0] + assert next_node.op_type in ['Conv', 'Gemm', 'ConvTranspose'] + redundant_nodes = [] + if node.input[0] not in named_initializer: + node.input[0], redundant_nodes = \ + self.weight_preprocess(node.input[0], out2node, inp2node, named_initializer) + next_node.input[idx] = node.input[0] + return redundant_nodes + + def deal_with_activation_fakequant(self, node, inp2node): + next_nodes = inp2node[node.output[0]] + for next_node, idx in next_nodes: + # if next_node.op_type == 'Add' and next_node.output[0] in inp2node: #将observer qdq的输入写到下一个layer的输入 + # nextnode, i = inp2node[next_node.output[0]][0] + # nextnode.input[i] = node.input[0] + # else: + next_node.input[idx] = node.input[0] + return + + def parse_qparams(self, node, name2data): + tensor_name, scale, zero_point = node.input[:3] + scale, zero_point = name2data[scale], name2data[zero_point] + if len(node.input) > 3: + qmin, qmax = node.input[-2:] + qmin, qmax = name2data[qmin], name2data[qmax] + elif len(node.attribute) > 0: + qparams = parse_attrs(node.attribute) + qmin = qparams['quant_min'] + qmax = qparams['quant_max'] + else: + logger.info(f'qmin and qmax are not found for <{node.name}>!') + return tensor_name, scale, zero_point, qmin, qmax + + def clip_weight(self, node, name2data, inp2node, named_initializer): + tensor_name, scale, zero_point, qmin, qmax = self.parse_qparams(node, name2data) + data = name2data[tensor_name] + clip_range_min = ((qmin - zero_point) * scale).astype(data.dtype) + clip_range_max = ((qmax - zero_point) * scale).astype(data.dtype) + if len(scale.shape) > 0 and scale.shape[0] > 1: + new_data = [] + transposed = False + next_node = inp2node[node.output[0]] + if len(next_node) == 1 and next_node[0][0].op_type == 'ConvTranspose': + transposed = True + data = data.transpose(1, 0, 2, 3) + for c in range(data.shape[0]): + new_data.append(np.clip(data[c], clip_range_min[c], clip_range_max[c])) + new_data = np.array(new_data) + if transposed: + new_data = new_data.transpose(1, 0, 2, 3) + logger.info(f'Clip weights <{tensor_name}> to per-channel ranges.') + else: + new_data = np.clip(data, clip_range_min, clip_range_max) + logger.info(f'Clip weights <{tensor_name}> to range [{clip_range_min}, {clip_range_max}].') + new_data = numpy_helper.from_array(new_data) + named_initializer[tensor_name].raw_data = new_data.raw_data + + def get_correct_sophgo_tpu_input_tensor_name(self, node, out2node, have_int4=False): #和tpu-mlir的命名风格一致 + input_0 = node.input[0] + op_type = '' + if input_0 in out2node: + op_type = out2node[input_0].op_type + tensor_name = '{}_{}'.format(input_0, op_type) + else: + tensor_name = input_0 + if have_int4 and op_type in ['Conv', "Gemm"]: + tensor_name += '_4' + else: + tensor_name += '_8' + return tensor_name + + def post_process_clip_ranges(self, clip_ranges, graph, inp2node, out2node): + def find_the_closest_clip_range(node): + tensor_name = self.get_correct_sophgo_tpu_input_tensor_name(node, out2node) + if tensor_name in clip_ranges: + return tensor_name + elif node.op_type in ['Flatten', 'Resize'] and node.output[0] in inp2node: + return find_the_closest_clip_range(inp2node[node.output[0]][0][0]) + else: + return None + + for node in graph.node: + if node.op_type in ['Flatten', 'Resize']: + tensor_name = find_the_closest_clip_range(node) + if tensor_name: + new_name = self.get_correct_sophgo_tpu_input_tensor_name(node, out2node) + clip_ranges[new_name] = copy.deepcopy(clip_ranges[tensor_name]) + logger.info(f'Pass <{tensor_name}> clip range to <{node.name}> input <{node.input[0]}>.') + return clip_ranges + + def post_process_clip_ranges2(self, clip_ranges, graph, inp2node, out2node, have_int4 = False): + op_type_inAndOutShouldSameClipRange = ['Flatten', 'Resize', 'Reshape', 'Transpose'] + for node in graph.node: + tensor_name = f'{node.output[0]}_{node.op_type}' + tensor_name += '_4' if have_int4 and node.op_type in ['Conv', "Gemm"] else '_8' + if tensor_name not in clip_ranges: + pre_op = node + finded = False + while pre_op.op_type in op_type_inAndOutShouldSameClipRange: + tensor_name2 = self.get_correct_sophgo_tpu_input_tensor_name(pre_op, out2node, have_int4) + if tensor_name2 in clip_ranges: + finded = True + clip_ranges[tensor_name] = clip_ranges[tensor_name2] + print(f'pre_op finded, transfer {tensor_name2} to {tensor_name}') + break + if pre_op.input[0] in out2node: + pre_op = out2node[pre_op.input[0]] + else: + print(f'{pre_op.name}\'s pre_node not exist') + break + if not finded: + if node.output[0] in inp2node: + next_op = inp2node[node.output[0]][0][0] + while next_op.op_type in op_type_inAndOutShouldSameClipRange: + tensor_name2 = f'{next_op.output[0]}_{next_op.op_type}' + tensor_name2 += '_4' if have_int4 and next_op.op_type in ['Conv', "Gemm"] else '_8' + if tensor_name2 in clip_ranges: + finded = True + clip_ranges[tensor_name] = clip_ranges[tensor_name2] + print(f'next_op finded, transfer {tensor_name2} to {tensor_name}') + break + if next_op.output[0] in inp2node: + next_op = inp2node[next_op.output[0]][0][0] + else: + print(f'{next_op.name}\'s next_op not exist') + break + else: + print(f'{node.name}\'s next_op not exist') + # if not finded: + # print(f'Waring:{node.name}\'s clip_ranges not exist, maybe have some error') + return clip_ranges + + def isQdqAdd(self, tensor_name, out2node): + pre_name = None + for i in out2node[tensor_name].input: + if out2node[i].op_type == 'LearnablePerTensorAffine': + pre_name = out2node[i].input[0] + if pre_name in out2node: + pre_name = '{}_{}'.format(pre_name, out2node[pre_name].op_type) + if out2node[i].op_type not in ['LearnablePerTensorAffine', 'Sub']: + return False, pre_name + return True, pre_name + + def remove_Qdq_add_sub(self, graph, inp2node, out2node): + nodes_to_be_removed = [] + for idx, node in enumerate(graph.node): + if node.op_type == 'Add': + isQdqAdd, _ = self.isQdqAdd(node.output[0], out2node) + if isQdqAdd: + for i in node.input: + if i.op_type == 'Sub': + nodes_to_be_removed.append(out2node[i]) + else: + fake_quant_output = i + next_nodes = inp2node[node.output[0]] + for next_node, idx in next_nodes: + next_node.input[idx] = fake_quant_output + nodes_to_be_removed.append(node) + nodes_to_be_removed.extend(get_constant_inputs(node, out2node)) + for node in nodes_to_be_removed: + graph.node.remove(node) + return + + def remove_fakequantize_and_collect_params(self, onnx_path, model_name): + model = onnx.load(onnx_path) + graph = model.graph + out2node, inp2node = update_inp2node_out2node(graph) + name2data = prepare_data(graph) + named_initializer = prepare_initializer(graph) + + preprocess = OnnxPreprocess() + preprocess.remove_fake_pad_op(graph, name2data, inp2node, out2node) + out2node, inp2node = update_inp2node_out2node(graph) + # self.remove_Qdq_add_sub(graph, inp2node, out2node) + # out2node, inp2node = update_inp2node_out2node(graph) + + clip_ranges = {} + nodes_to_be_removed = [] + output_path = os.path.dirname(onnx_path) + file_name = os.path.join(output_path, 'layer_outputs.npz') + layer_out_tensor = None + have_int4 = False + layer_out_tensor2 = {} + if os.path.exists(file_name): + layer_out_tensor = np.load(file_name) + for node in graph.node: + print(f'process node:{node.name}, type:{node.op_type}') + if node.op_type in ALL_FAKEQUANTIZER: + nodes_to_be_removed.append(node) + nodes_to_be_removed.extend(get_constant_inputs(node, out2node)) + if node.output[0] not in inp2node: + assert node.output[0] in [l.name for l in graph.output] + inp2node[node.output[0]] = [] + next_nodes = inp2node[node.output[0]] + if node.op_type in PERCHANNEL_FAKEQUANTIZER: + # fake quantize for weights, suppose per-channel quantize only for weight + redundant_nodes = self.deal_with_weight_fakequant(node, out2node, inp2node, named_initializer) + nodes_to_be_removed.extend(redundant_nodes) + self.clip_weight(node, name2data, inp2node, named_initializer) + tensor_name, scale, zero_point, qmin, qmax = self.parse_qparams(node, name2data) + #卷积权重per-channel量化参数，bias的per-chan量化参数没有去调优 + if len(next_nodes) == 1 and next_nodes[0][0].op_type in ['Gemm', 'Conv']:#当前伪量化节点只有1个后继，且第1个后继节点为conv类型 + next_node_output = next_nodes[0][0].output[0] + if inp2node[next_node_output][0][0].op_type == 'Relu':##伪量化节点的第1个后继conv节点的第1个后继节点为Relu(fake->conv->relu) + #若是fake->conv->relu,因为relu会融合到前面conv，故用relu的输出tensor名+Relu作为量化参数保存tensor名 + tensor_name = '{}_{}'.format(inp2node[next_node_output][0][0].output[0], 'Relu') + else: + #若是fake->conv->not_relu_type,直接用conv的输出tensor名+conv作为量化参数保存tensor名 + tensor_name = '{}_{}'.format(next_node_output, next_nodes[0][0].op_type) + tensor_name += '_{}'.format('weight' if next_nodes[0][1] == 1 else 'bias' ) + clip_ranges[tensor_name] = {'step': [float(x) for x in scale], + 'zero_point': [int(x) for x in zero_point]} + elif node.op_type in PERTENSOR_FAKEQUANTIZER: + if len(next_nodes) == 1 and next_nodes[0][1] == 1 and next_nodes[0][0].op_type in ['Gemm', 'Conv']: + # fake quantize for weights + redundant_nodes = self.deal_with_weight_fakequant(node, out2node, inp2node, named_initializer) + tensor_name, scale, zero_point, qmin, qmax = self.parse_qparams(node, name2data) + nodes_to_be_removed.extend(redundant_nodes) + self.clip_weight(node, name2data, inp2node, named_initializer) + assert next_nodes[0][0].op_type == 'Gemm' + tensor_name_new = '{}_{}_weight'.format(next_nodes[0][0].output[0], next_nodes[0][0].op_type) + clip_ranges[tensor_name_new] = {'step': [float(x) for x in scale], + 'zero_point': [int(x) for x in zero_point]} + else: + # fake quantize for activations + self.deal_with_activation_fakequant(node, inp2node) + tensor_name, scale, zero_point, qmin, qmax = self.parse_qparams(node, name2data) + bits = 4 if qmax == 7 else 8 + if bits == 4: + have_int4 = True + for out in graph.output: + output_name = node.output[0] + # if inp2node[node.output[0]][0][0].op_type == 'Add': + # output_name = inp2node[node.output[0]][0][0].output[0] + if out.name == output_name: + out.name = tensor_name + + scale_name = node.input[1] + post_str = '_post_act_fake_quantizer.scale' + tensor_name_new = tensor_name + if tensor_name in out2node: + tensor_name_new += '_{}'.format(out2node[tensor_name].op_type) + # if out2node[tensor_name].op_type == 'Add': + # isQdqAdd, pre_name = self.isQdqAdd(tensor_name, out2node) + # if isQdqAdd: + # tensor_name_new = pre_name+'_qdq' + + # if out2node[tensor_name].op_type in ['Conv', 'Gemm']: + # if bits == 8: + # tensor_name_new += '_8' + # else: + # if bits == 4 and out2node[out2node[tensor_name].input[0]].op_type not in ['Conv', 'Gemm']: + # tensor_name_new += '_4' + if layer_out_tensor is not None and scale_name.endswith(post_str): + torch_name = scale_name[:len(scale_name)-len(post_str)] + if torch_name in layer_out_tensor.files: + layer_out_tensor2[tensor_name_new] = layer_out_tensor[torch_name] + clip_ranges[tensor_name_new+f'_{bits}'] = {'threshold':float(scale * max(-qmin, qmax)), #对称量化时这个参数生效 + 'min': float(scale * (qmin - zero_point)), + 'max': float(scale * (qmax - zero_point))} + if layer_out_tensor is not None and len(layer_out_tensor2) > 0: + if 'data' in layer_out_tensor.files: + layer_out_tensor2['data'] = layer_out_tensor['data'] + os.system(f'rm -f {file_name}') + np.savez(file_name, **layer_out_tensor2) + for node in nodes_to_be_removed: + graph.node.remove(node) + # delete initializer + out2node, inp2node = update_inp2node_out2node(graph) + named_initializer = prepare_initializer(graph) + for name, initial_data in named_initializer.items(): + if name in (out2node.keys() | inp2node.keys()): + continue + graph.initializer.remove(initial_data) + + # clip_ranges = self.post_process_clip_ranges(clip_ranges, graph, inp2node, out2node) + clip_ranges = self.post_process_clip_ranges2(clip_ranges, graph, inp2node, out2node, have_int4) + context = {'sophgo_tpu': clip_ranges} + context['w_qscheme'] = '' + context['a_qscheme'] = '' + context_filename = os.path.join(output_path, '{}_clip_ranges.json'.format(model_name)) + with open(context_filename, 'w') as f: + json.dump(context, f, indent=4) + onnx_filename = os.path.join(output_path, '{}_deploy_model.onnx'.format(model_name)) + model_onnx = onnx.shape_inference.infer_shapes(model) + os.system(f"rm -f {onnx_filename}") + onnx.save(model_onnx, onnx_filename) + logger.info("Finish deploy process.") + +remove_fakequantize_and_collect_params_sophgo = LinearQuantizer_process().remove_fakequantize_and_collect_params diff --git a/mqbench/fake_quantize/__init__.py b/mqbench/fake_quantize/__init__.py index 846bb6ac..8525a623 100644 --- a/mqbench/fake_quantize/__init__.py +++ b/mqbench/fake_quantize/__init__.py @@ -6,4 +6,6 @@ from .pact import PACTFakeQuantize from .tqt import TqtFakeQuantize from .adaround_quantizer import AdaRoundFakeQuantize -from .qdrop_quantizer import QDropFakeQuantize \ No newline at end of file +from .qdrop_quantizer import QDropFakeQuantize +from .e4m3 import E4M3FakeQuantize +from .e5m2 import E5M2FakeQuantize \ No newline at end of file diff --git a/mqbench/fake_quantize/dsq.py b/mqbench/fake_quantize/dsq.py index 316f1be6..c37e40a0 100644 --- a/mqbench/fake_quantize/dsq.py +++ b/mqbench/fake_quantize/dsq.py @@ -1,7 +1,7 @@ import math import torch - +from torch.onnx import symbolic_helper from mqbench.fake_quantize.quantize_base import QuantizeBase from mqbench.utils import is_tracing_state from mqbench.utils.hook import PerChannelLoadHook @@ -84,8 +84,14 @@ def forward(ctx, x, scale, zero_point, quant_min, quant_max, ch_axis, alpha): @staticmethod def symbolic(g, x, scale, zero_point, quant_min, quant_max, ch_axis, alpha): - return g.op("::FakeQuantizeDSQPerchannel", x, scale, zero_point, quant_min_i=quant_min, quant_max_i=quant_max, alpha_f=alpha) + output = g.op("MQBench_custom::FakeQuantizeDSQPerchannel", x, scale, zero_point, quant_min_i=quant_min, quant_max_i=quant_max, alpha_f=alpha) + + input_shape = symbolic_helper._get_tensor_sizes(x) + if input_shape is not None and hasattr(x.type(), 'with_sizes'): + output_type = x.type().with_sizes(input_shape) + output.setType(output_type) + return output class FakeQuantizeDSQPertensor(torch.autograd.Function): @staticmethod @@ -94,4 +100,11 @@ def forward(ctx, x, scale, zero_point, quant_min, quant_max, alpha): @staticmethod def symbolic(g, x, scale, zero_point, quant_min, quant_max, alpha): - return g.op("::FakeQuantizeDSQPertensor", x, scale, zero_point, quant_min_i=quant_min, quant_max_i=quant_max, alpha_f=alpha) + output = g.op("MQBench_custom::FakeQuantizeDSQPertensor", x, scale, zero_point, quant_min_i=quant_min, quant_max_i=quant_max, alpha_f=alpha) + + input_shape = symbolic_helper._get_tensor_sizes(x) + if input_shape is not None and hasattr(x.type(), 'with_sizes'): + output_type = x.type().with_sizes(input_shape) + output.setType(output_type) + + return output \ No newline at end of file diff --git a/mqbench/fake_quantize/e4m3.py b/mqbench/fake_quantize/e4m3.py new file mode 100644 index 00000000..7f8f7ab2 --- /dev/null +++ b/mqbench/fake_quantize/e4m3.py @@ -0,0 +1,217 @@ +# E4M3的Fake Quantize + +import os +import yaml +from easydict import EasyDict +import torch + +from .quantize_base import QuantizeBase +from ..utils.hook import PerChannelLoadHook + +_version_under_1100 = int(torch.__version__.split('.')[1]) < 10 + +mode_list = ["E4M3_IEEE_RNE", "E4M3_IEEE_STOCHASTIC", "E4M3_RNE", "E4M3_STOCHASTIC"] # 可选择的E4M3量化手段 + +# 用parse_config函数获取config文件中的mode，用于之后具体量化方法的选择： +def parse_config(config_file): + with open(config_file) as f: + config = yaml.load(f, Loader=yaml.FullLoader) + cur_config = config + cur_path = config_file + while 'root' in cur_config: + root_path = os.path.dirname(cur_path) + cur_path = os.path.join(root_path, cur_config['root']) + with open(cur_path) as r: + root_config = yaml.load(r, Loader=yaml.FullLoader) + for k, v in root_config.items(): + if k not in config: + config[k] = v + cur_config = root_config + config = EasyDict(config) + return config + +# 写一个获取FP8量化所能表示的最大值函数： +def get_flt_max(mode): + if mode.lower() == "e5m2": + return float(57344.0) # E5M2所能表示的最大值 + elif mode.lower() == "e4m3": + return float(448.0) # E4M3所能表示的最大值 + +# 写一个获取FP8量化所能表示的最小值函数： +def get_flt_min(mode): + if mode.lower() =="e5m2": + return float(1.5258789E-05) # E5M2所能表示的最小值 + elif mode.lower() == "e4m3": + return float(1.9531250E-03) #E4M3所能表示的最小值 + +# 写一个Int量化的转化函数（以支持INT8/INT4量化）： +def quantize_to_integer(tensor, mode, inplace=False): + # compute tensor min and max values + min_val = torch.min(tensor) + max_val = torch.max(tensor) + # int8 quantization range + + nbits = int(mode.split("INT")[1])-1 + q_min = -1*2**nbits + q_max = (2**nbits)-1 + + """ + q_min = -128 + q_max = 127 + if mode == "INT4": + q_min = -8 + q_max = 7 + """ + # compute scale and zero_point + scale = (max_val - min_val) / (q_max - q_min) + zero_point = q_min - (min_val / scale) + # Quantize the input tensor using int8 representation + qtensor = torch.round((tensor / scale) + zero_point) + # Clamp the values to the int8 range + qtensor = torch.clamp(qtensor, q_min, q_max) + # Dequantize the tensor + dqtensor = scale * (qtensor - zero_point) + + if inplace is True: + tensor.data.copy_(dqtensor) + return tensor + + return dqtensor + +#调用emulator函数计算量化后的权重： +def fpemu_device_fn(tensor, mode, inplace=True, scale=1.0): + #if "INT8" in mode or "INT4" in mode: + if "INT" in mode: # 如果输入的mode是INT类型，走这个循环进行整数的量化 + return quantize_to_integer(tensor, mode.split("_")[0], inplace=inplace) + + if tensor.is_cuda : # 如果使用CUDA走这个循环，调用了pytquant中的CUDA函数 + from FP8_Emulator.pytquant.cuda import fpemu_cuda + X = fpemu_cuda.FPEmuOp.apply(tensor, mode, inplace, scale) + + else : # 如果使用CPU走这个循环，调用了pytquant中的CPP函数 + from FP8_Emulator.pytquant.cpp import fpemu_cpp + X = fpemu_cpp.FPEmuOp.apply(tensor, mode, inplace, scale) + + return X + +class E4M3FakeQuantize(QuantizeBase): + """This is fp8 E4M3 Quantization Emulator.""" + + def __init__(self, observer, **observer_kwargs): + super(E4M3FakeQuantize, self).__init__(observer, **observer_kwargs) + self.register_buffer('scale', torch.tensor([1.0], dtype=torch.float)) # 首先定义一个scale，初始值为1.0 + self.register_buffer('zero_point', torch.tensor([0], dtype=torch.int)) # 首先定义一个zero point，初始值为0.0（用于之后输出量化表） + self.load_state_dict_hook = PerChannelLoadHook(self) + + def forward(self, X): + tensor_q = torch.zeros_like(X) + #scaling_method = parse_config('config.yaml').quant.scaling_method + scaling_method = "max" # 如果采取config文件定义args的方式可以选用上面的代码，否则需要手动更换round和scale模式 + if self.observer_enabled[0] == 1: # 如果使用observer + self.activation_post_process(X.detach()) + _scale, _zero_point = self.calculate_qparams() # 通过原本的函数获得zp（fp8量化中不需要zp，但可以输出量化表作为参考） + if scaling_method.lower() == "mean": + mean = torch.mean(abs(torch.flatten(X.detach()))) + mean = abs(mean) if abs(mean) > 1e-5 else get_flt_min("e4m3") #将mean的绝对值与1e-5比较 + if abs(mean) > 0.0: + _scale = get_flt_min("e4m3") / abs(mean) # 取得e4m3的最小值，与mean的绝对值做比值求得scale + elif scaling_method.lower() == "max": + vmax = torch.max(abs(torch.flatten(X.detach()))) #求出权重的max + _scale = get_flt_max("e4m3") / vmax + _scale = torch.tensor(6.55e+04) if _scale.item() > 3.275e+04 else _scale + else: + _scale = torch.tensor(1.0) + _scale, _zero_point = _scale.to(self.scale.device), _zero_point.to(self.zero_point.device) + if self.scale.shape != _scale.shape: + self.scale.resize_(_scale.shape) + self.zero_point.resize_(_zero_point.shape) + self.scale.copy_(_scale) + self.zero_point.copy_(_zero_point) + + if self.fake_quant_enabled[0] == 1: # 如果使用fake quantize + #work_mode = 'E4M3_' + parse_config('config.yaml').quant.mode.upper() + work_mode = 'E4M3_RNE' + if self.is_per_channel: # 按照per channel的方式计算scale + channels = X.shape[1] + for c in range(channels): + sub_tensor = X.select(1, c).detach() + if scaling_method.lower() == "mean": + mean = torch.mean(abs(torch.flatten(X.detach()))) + mean = abs(mean) if abs(mean) > 1e-5 else get_flt_min("e4m3") + if abs(mean) > 0.0: + _scale = get_flt_min("e4m3") / abs(mean) + elif scaling_method.lower() == "max": + vmax = torch.max(abs(torch.flatten(X.detach()))) + _scale = get_flt_max("e4m3") / vmax + _scale = torch.tensor(6.55e+04) if _scale.item() > 3.275e+04 else _scale + else: + _scale = torch.tensor(1.0) + self.scale.copy_(_scale) + sub_tensor = fpemu_device_fn(sub_tensor, mode=work_mode, inplace=False, scale=self.scale.item()) + tensor_q.select(1, c).data.copy_(sub_tensor) + X = tensor_q # per channel方式计算后的量化权重 + else: # 按照per tensor的方法计算scale + if scaling_method.lower() == "mean": + mean = torch.mean(abs(torch.flatten(X.detach()))) + mean = abs(mean) if abs(mean) > 1e-5 else get_flt_min("e4m3") #将mean的绝对值与1e-5比较 + if abs(mean) > 0.0: + _scale = get_flt_min("e4m3") / abs(mean) # 取得e4m3的最小值，与mean的绝对值做比值求得scale + elif scaling_method.lower() == "max": + vmax = torch.max(abs(torch.flatten(X.detach()))) #求出权重的max + _scale = get_flt_max("e4m3") / vmax + _scale = torch.tensor(6.55e+04) if _scale.item() > 3.275e+04 else _scale + else: + _scale = torch.tensor(1.0) + _scale = _scale.to(self.scale.device) + self.scale.copy_(_scale) + X = fpemu_device_fn(X, mode=work_mode, inplace=False, scale=self.scale.item()) #返回per tensor方式计算的量化权重 + return X + + @torch.jit.export + def extra_repr(self): + return 'fake_quant_enabled={}, observer_enabled={}, ' \ + 'quant_min={}, quant_max={}, dtype={}, qscheme={}, ch_axis={}, ' \ + 'scale={}, zero_point={}'.format( + self.fake_quant_enabled, self.observer_enabled, + self.quant_min, self.quant_max, + self.dtype, self.qscheme, self.ch_axis, self.scale if self.ch_axis == -1 else 'List', + self.zero_point if self.ch_axis == -1 else 'List') + + def _save_to_state_dict(self, destination, prefix, keep_vars): + # We cannot currently register scalar values as buffers, so need to manually + # specify serialization here. + super(E4M3FakeQuantize, self)._save_to_state_dict(destination, prefix, keep_vars) + destination[prefix + 'scale'] = self.scale + destination[prefix + 'zero_point'] = self.zero_point + + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, + missing_keys, unexpected_keys, error_msgs): + # Removing this function throws an error that the the size of the loaded tensor does not match the original size + # i.e., These buffers start out with numel 0 and become numel 1 once they have their first forward pass. + local_state = ['scale', 'zero_point'] + for name in local_state: + key = prefix + name + if key in state_dict: + val = state_dict[key] + # Custom handling to allow loading scale and zero_point + # of size N into uninitialized buffers of size 0. The + # buffers are resized here, and the values are copied in + # the default state_dict loading code of the parent. + if name == 'scale': + self.scale.resize_(val.shape) + else: + assert name == 'zero_point' + self.zero_point.resize_(val.shape) + # For torchscript module we need to update the attributes here since we do not + # call the `_load_from_state_dict` function defined module.py + if torch.jit.is_scripting(): + if name == 'scale': + self.scale.copy_(val) + else: + assert name == 'zero_point' + self.zero_point.copy_(val) + elif strict: + missing_keys.append(key) + + super(E4M3FakeQuantize, self)._load_from_state_dict(state_dict, prefix, local_metadata, strict, + missing_keys, unexpected_keys, error_msgs) diff --git a/mqbench/fake_quantize/e5m2.py b/mqbench/fake_quantize/e5m2.py new file mode 100644 index 00000000..6fe6bdfd --- /dev/null +++ b/mqbench/fake_quantize/e5m2.py @@ -0,0 +1,219 @@ +# E5M2的Fake Quantize + +import os +import yaml +from easydict import EasyDict +import torch + +from .quantize_base import QuantizeBase +from ..utils.hook import PerChannelLoadHook + +_version_under_1100 = int(torch.__version__.split('.')[1]) < 10 + +mode_list = ["E5M2_RTE", "E5M2_RNE", "E5M2_STOCHASTIC", "E5M2_RNAZ", + "E5M2_RNTZ", "E5M2_RPINF", "E5M2_RNINF", "E5M2_DAZ_RNE", + "E5M2_DAZ_STOCHASITC", "E5M2_DAZ_RNAZ", "E5M2_DAZ_RNTZ"] # 可以选择的E5M2量化方法 + +# 用parse_config函数获取config文件中的mode，用于之后具体量化方法的选择： +def parse_config(config_file): + with open(config_file) as f: + config = yaml.load(f, Loader=yaml.FullLoader) + cur_config = config + cur_path = config_file + while 'root' in cur_config: + root_path = os.path.dirname(cur_path) + cur_path = os.path.join(root_path, cur_config['root']) + with open(cur_path) as r: + root_config = yaml.load(r, Loader=yaml.FullLoader) + for k, v in root_config.items(): + if k not in config: + config[k] = v + cur_config = root_config + config = EasyDict(config) + return config + +# 写一个获取FP8量化所能表示的最大值函数： +def get_flt_max(mode): + if mode.lower() == "e5m2": + return float(57344.0) # E5M2所能表示的最大值 + elif mode.lower() == "e4m3": + return float(448.0) # E4M3所能表示的最大值 + +# 写一个获取FP8量化所能表示的最小值函数： +def get_flt_min(mode): + if mode.lower() =="e5m2": + return float(1.5258789E-05) # E5M2所能表示的最小值 + elif mode.lower() == "e4m3": + return float(1.9531250E-03) #E4M3所能表示的最小值 + +# 写一个Int量化的转化函数（以支持INT8/INT4量化）： +def quantize_to_integer(tensor, mode, inplace=False): + # compute tensor min and max values + min_val = torch.min(tensor) + max_val = torch.max(tensor) + # int8 quantization range + + nbits = int(mode.split("INT")[1])-1 + q_min = -1*2**nbits + q_max = (2**nbits)-1 + + """ + q_min = -128 + q_max = 127 + if mode == "INT4": + q_min = -8 + q_max = 7 + """ + # compute scale and zero_point + scale = (max_val - min_val) / (q_max - q_min) + zero_point = q_min - (min_val / scale) + # Quantize the input tensor using int8 representation + qtensor = torch.round((tensor / scale) + zero_point) + # Clamp the values to the int8 range + qtensor = torch.clamp(qtensor, q_min, q_max) + # Dequantize the tensor + dqtensor = scale * (qtensor - zero_point) + + if inplace is True: + tensor.data.copy_(dqtensor) + return tensor + + return dqtensor + +#调用emulator函数计算量化后的权重： +def fpemu_device_fn(tensor, mode, inplace=True, scale=1.0): + #if "INT8" in mode or "INT4" in mode: + if "INT" in mode: # 如果输入的mode是INT类型，走这个循环进行整数的量化 + return quantize_to_integer(tensor, mode.split("_")[0], inplace=inplace) + + if tensor.is_cuda : # 如果使用CUDA走这个循环，调用了pytquant中的CUDA函数 + from FP8_Emulator.pytquant.cuda import fpemu_cuda + X = fpemu_cuda.FPEmuOp.apply(tensor, mode, inplace, scale) + + else : # 如果使用CPU走这个循环，调用了pytquant中的CPP函数 + from FP8_Emulator.pytquant.cpp import fpemu_cpp + X = fpemu_cpp.FPEmuOp.apply(tensor, mode, inplace, scale) + + return X + +class E5M2FakeQuantize(QuantizeBase): + """This is fp8 E5M2 Quantization Emulator.""" + + def __init__(self, observer, **observer_kwargs): + super(E5M2FakeQuantize, self).__init__(observer, **observer_kwargs) + self.register_buffer('scale', torch.tensor([1.0], dtype=torch.float)) # 首先定义一个scale，初始值为1.0 + self.register_buffer('zero_point', torch.tensor([0], dtype=torch.int)) # 首先定义一个zero point，初始值为0.0（用于之后输出量化表） + self.load_state_dict_hook = PerChannelLoadHook(self) + + def forward(self, X): + tensor_q = torch.zeros_like(X) + #scaling_method = parse_config('config.yaml').quant.scaling_method + scaling_method = "no_scaling" + if self.observer_enabled[0] == 1: # 如果使用observer + self.activation_post_process(X.detach()) + _scale, _zero_point = self.calculate_qparams() # 通过原本的函数获得zp（fp8量化中不需要zp，但可以输出量化表作为参考） + if scaling_method.lower() == "mean": + mean = torch.mean(abs(torch.flatten(X.detach()))) + mean = abs(mean) if abs(mean) > 1e-5 else get_flt_min("e5m2") #将mean的绝对值与1e-5比较 + if abs(mean) > 0.0: + _scale = get_flt_min("e5m2") / abs(mean) # 取得e5m2的最小值，与mean的绝对值做比值求得scale + elif scaling_method.lower() == "max": + vmax = torch.max(abs(torch.flatten(X.detach()))) #求出权重的max + _scale = get_flt_max("e5m2") / vmax + _scale = torch.tensor(6.55e+04) if _scale.item() > 3.275e+04 else _scale + else: + _scale = torch.tensor(1.0) + _scale, _zero_point = _scale.to(self.scale.device), _zero_point.to(self.zero_point.device) + if self.scale.shape != _scale.shape: + self.scale.resize_(_scale.shape) + self.zero_point.resize_(_zero_point.shape) + self.scale.copy_(_scale) + self.zero_point.copy_(_zero_point) + + if self.fake_quant_enabled[0] == 1: # 如果使用fake quantize + #work_mode = 'E5M2_' + parse_config('config.yaml').quant.mode.upper() + work_mode = "E5M2_RNE" + if self.is_per_channel: # 按照per channel的方式计算scale + channels = X.shape[1] + for c in range(channels): + sub_tensor = X.select(1, c).detach() + if scaling_method.lower() == "mean": + mean = torch.mean(abs(torch.flatten(X.detach()))) + mean = abs(mean) if abs(mean) > 1e-5 else get_flt_min("e5m2") + if abs(mean) > 0.0: + _scale = get_flt_min("e5m2") / abs(mean) + elif scaling_method.lower() == "max": + vmax = torch.max(abs(torch.flatten(X.detach()))) + _scale = get_flt_max("e5m2") / vmax + _scale = torch.tensor(6.55e+04) if _scale.item() > 3.275e+04 else _scale + else: + _scale = torch.tensor(1.0) + self.scale.copy_(_scale) + sub_tensor = fpemu_device_fn(sub_tensor, mode=work_mode, inplace=False, scale=self.scale.item()) + tensor_q.select(1, c).data.copy_(sub_tensor) + X = tensor_q # per channel方式计算后的量化权重 + else: # 按照per tensor的方法计算scale + if scaling_method.lower() == "mean": + mean = torch.mean(abs(torch.flatten(X.detach()))) + mean = abs(mean) if abs(mean) > 1e-5 else get_flt_min("e5m2") #将mean的绝对值与1e-5比较 + if abs(mean) > 0.0: + _scale = get_flt_min("e5m2") / abs(mean) # 取得e5m2的最小值，与mean的绝对值做比值求得scale + elif scaling_method.lower() == "max": + vmax = torch.max(abs(torch.flatten(X.detach()))) #求出权重的max + _scale = get_flt_max("e5m2") / vmax + _scale = torch.tensor(6.55e+04) if _scale.item() > 3.275e+04 else _scale + else: + _scale = torch.tensor(1.0) + _scale = _scale.to(self.scale.device) + self.scale.copy_(_scale) + X = fpemu_device_fn(X, mode=work_mode, inplace=False, scale=self.scale.item()) #返回per tensor方式计算的量化权重 + return X + + @torch.jit.export + def extra_repr(self): + return 'fake_quant_enabled={}, observer_enabled={}, ' \ + 'quant_min={}, quant_max={}, dtype={}, qscheme={}, ch_axis={}, ' \ + 'scale={}, zero_point={}'.format( + self.fake_quant_enabled, self.observer_enabled, + self.quant_min, self.quant_max, + self.dtype, self.qscheme, self.ch_axis, self.scale if self.ch_axis == -1 else 'List', + self.zero_point if self.ch_axis == -1 else 'List') + + def _save_to_state_dict(self, destination, prefix, keep_vars): + # We cannot currently register scalar values as buffers, so need to manually + # specify serialization here. + super(E5M2FakeQuantize, self)._save_to_state_dict(destination, prefix, keep_vars) + destination[prefix + 'scale'] = self.scale + destination[prefix + 'zero_point'] = self.zero_point + + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, + missing_keys, unexpected_keys, error_msgs): + # Removing this function throws an error that the the size of the loaded tensor does not match the original size + # i.e., These buffers start out with numel 0 and become numel 1 once they have their first forward pass. + local_state = ['scale', 'zero_point'] + for name in local_state: + key = prefix + name + if key in state_dict: + val = state_dict[key] + # Custom handling to allow loading scale and zero_point + # of size N into uninitialized buffers of size 0. The + # buffers are resized here, and the values are copied in + # the default state_dict loading code of the parent. + if name == 'scale': + self.scale.resize_(val.shape) + else: + assert name == 'zero_point' + self.zero_point.resize_(val.shape) + # For torchscript module we need to update the attributes here since we do not + # call the `_load_from_state_dict` function defined module.py + if torch.jit.is_scripting(): + if name == 'scale': + self.scale.copy_(val) + else: + assert name == 'zero_point' + self.zero_point.copy_(val) + elif strict: + missing_keys.append(key) + + super(E5M2FakeQuantize, self)._load_from_state_dict(state_dict, prefix, local_metadata, strict, + missing_keys, unexpected_keys, error_msgs) diff --git a/mqbench/fake_quantize/fixed.py b/mqbench/fake_quantize/fixed.py index 1fd2ae2b..46c494fc 100644 --- a/mqbench/fake_quantize/fixed.py +++ b/mqbench/fake_quantize/fixed.py @@ -3,8 +3,7 @@ from mqbench.fake_quantize.quantize_base import QuantizeBase from mqbench.utils.hook import PerChannelLoadHook - -_version_under_1100 = int(torch.__version__.split('.')[1]) < 10 +from mqbench.fake_quantize.quantize_base import _version_under_1100 class FixedFakeQuantize(QuantizeBase): """This is actually torch.quantization.FakeQuantize. diff --git a/mqbench/fake_quantize/fp16.py b/mqbench/fake_quantize/fp16.py new file mode 100644 index 00000000..829fb5b2 --- /dev/null +++ b/mqbench/fake_quantize/fp16.py @@ -0,0 +1,24 @@ +import torch + + +class Fp16FakeQuantize(): + def __init__(self): + pass + + @torch.jit.export + def extra_repr(self): + return 'Fp16FakeQuantize' + + def forward(self, X): + #调用自定义torch c++ op将fp32的X转为fp16后再转会fp32，引入误差 + +class BF16FakeQuantize(): + def __init__(self): + pass + + @torch.jit.export + def extra_repr(self): + return 'BF16FakeQuantize' + + def forward(self, X): + #fp32的X转为bf16后再转会fp32，引入误差 \ No newline at end of file diff --git a/mqbench/fake_quantize/global_var.py b/mqbench/fake_quantize/global_var.py new file mode 100644 index 00000000..52cc930d --- /dev/null +++ b/mqbench/fake_quantize/global_var.py @@ -0,0 +1,17 @@ +import torch + +def _init(): + global all_data_dict + all_data_dict = {} + +def set_value(key, value): + all_data_dict[key] = value + +def get_value(key): + try: + return all_data_dict[key] + except: + print('all_data_dict has no', key) + +def get_var(): + return all_data_dict \ No newline at end of file diff --git a/mqbench/fake_quantize/gptq.py b/mqbench/fake_quantize/gptq.py new file mode 100644 index 00000000..df8eaf17 --- /dev/null +++ b/mqbench/fake_quantize/gptq.py @@ -0,0 +1,260 @@ +import torch +from torch.nn.parameter import Parameter +from torch.quantization.observer import MovingAverageMinMaxObserver + +from mqbench.fake_quantize.quantize_base import QuantizeBase, _version_under_1100 +from mqbench.utils.hook import PerChannelLoadHook +from mqbench.fake_quantize import global_var + +import transformers +import math +import time +import numpy as np + +DEBUG = False + +def quantize(x, scale, zero, maxq): + if maxq < 0: + return (x > scale / 2).float() * scale + (x < zero / 2).float() * zero + q = torch.clamp(torch.round(x / scale) + zero, 0, maxq) + return scale * (q - zero) + +class GPTQFakeQuantize(QuantizeBase): + ''' + This is gptq method mqbench version. + ''' + + def __init__(self, observer, **observer_kwargs): + super(GPTQFakeQuantize, self).__init__(observer, **observer_kwargs) + self.register_buffer('scale', torch.tensor([1.0], dtype=torch.float)) + self.register_buffer('zero_point', torch.tensor([0], dtype=torch.int)) + self.is_gptq_valid = False + self.layer_name = None + self.rows = None + self.columns = None + self.H = torch.tensor([0], dtype=torch.float) + self.nsamples = 0 + self.layer_module = None + self.dev = torch.device('cpu') + # self.register_buffer('H', torch.tensor([0], dtype=torch.float)) + self.is_gptq_done = False + self.is_add_batch = False + self.actorder = False + + def forward(self, X): + # print('This FakeQuantizer is "', self.layer_name) + self.dev = X.device + x_ori = X + # print("GPTQFakeQuantize - forward") + if self.observer_enabled[0] == 1: + if (self.H.shape == torch.Size([1])): + W = X.data.clone() + if isinstance(self.layer_module, torch.nn.Conv2d): + W = W.flatten(1) + if isinstance(self.layer_module, transformers.Conv1D): + W = W.t() + self.rows = W.shape[0] + self.columns = W.shape[1] + self.H = torch.zeros((W.shape[1], W.shape[1]), device=self.dev) + + if (self.is_gptq_valid): + if (self.layer_name+'.inp' in global_var.get_var().keys()): + self.input = global_var.get_value(self.layer_name+'.inp') + self.add_batch() + + self.activation_post_process(X.detach()) + # All is per layer + _scale, _zero_point = self.activation_post_process.calculate_qparams() + _scale = _scale.to(self.scale.device) + _zero_point = _zero_point.to(self.zero_point.device) + + if self.scale.shape != _scale.shape: + self.scale.resize_(_scale.shape) + self.zero_point.resize_(_zero_point.shape) + + self.scale.data.copy_(_scale) + self.zero_point.data.copy_(_zero_point.float()) + if self.fake_quant_enabled[0] == 1: + # # Use GPTQ + if (self.is_gptq_valid): + if (not self.is_gptq_done): + with torch.no_grad(): + self.input = global_var.get_value(self.layer_name+'.inp') + self.output = global_var.get_value(self.layer_name+'.out') + if (self.input.device != self.dev or self.output.device != self.dev): + self.input = self.input.to(self.dev) + self.output = self.output.to(self.dev) + self.weight = X + # self.add_batch() + X = self.fasterquant(X) + else: + return X + # # Use FixedFakeQuantize per Tensor + else: + X = torch.fake_quantize_per_tensor_affine( + X, self.scale.item(), int(self.zero_point.item()), + self.quant_min, self.quant_max) + # X = torch.fake_quantize_per_tensor_affine( + # X, self.scale.item(), int(self.zero_point.item()), + # self.quant_min, self.quant_max) + return X + + @torch.jit.export + def extra_repr(self): + return 'fake_quant_enabled={}, observer_enabled={}, ' \ + 'quant_min={}, quant_max={}, dtype={}, qscheme={}, ch_axis={}, ' \ + 'scale={}, zero_point={}'.format( + self.fake_quant_enabled, self.observer_enabled, + self.quant_min, self.quant_max, + self.dtype, self.qscheme, self.ch_axis, self.scale if self.ch_axis == -1 else 'List[%s]' % str(self.scale.shape), + self.zero_point if self.ch_axis == -1 else 'List') + + def add_batch(self): + inp = self.input + if len(inp.shape) == 2: + inp = inp.unsqueeze(0) + tmp = inp.shape[0] + + if isinstance(self.layer_module, torch.nn.Linear) or isinstance(self.layer_module, transformers.Conv1D): + if len(inp.shape) == 3: + inp = inp.reshape((-1, inp.shape[-1])) + inp = inp.t() + if isinstance(self.layer_module, torch.nn.Conv2d): + unfold = torch.nn.Unfold( + self.layer_module.kernel_size, + dilation=self.layer_module.dilation, + padding=self.layer_module.padding, + stride=self.layer_module.stride + ) + inp = unfold(inp) + inp = inp.permute([1, 0, 2]) + inp = inp.flatten(1) + + self.H *= self.nsamples / (self.nsamples + tmp) # always zero ? + self.nsamples += tmp + # inp = inp.float() + inp = math.sqrt(2 / self.nsamples) * inp.float() + # self.H += 2 / self.nsamples * inp.matmul(inp.t()) + inp_inpt = inp.matmul(inp.t()) + self.H += inp_inpt + # inp_inpt.to(torch.device('cpu')) + # torch.cuda.empty_cache() + del inp_inpt + # print(self.H.size()) + + def fasterquant(self, X, blocksize=128, percdamp=.01, groupsize=-1, actorder=False): + # W = self.layer.weight.data.clone() + # if isinstance(self.layer,torch.nn.Conv2d): + # W = self.weight.data.clone() + W = X.data.clone() + + if isinstance(self.layer_module, torch.nn.Conv2d): + W = W.flatten(1) + if isinstance(self.layer_module, transformers.Conv1D): + W = W.t() + W = W.float() + + # # Calculate by Observer + # # if not self.quantizer.ready(): + # # self.quantizer.find_params(W, weight=True) + + H = self.H + # # del self.H + + # dead = torch.diag(H) == 0 + dead = [] + h, w = H.shape + H_diag = torch.zeros(w) + for i in range(h): + if (H[i, i] == 0): + dead.append(i) + H_diag[i] = H[i, i] + + H[dead, dead] = 1 + W[:, dead] = 0 + + if actorder: + perm = torch.argsort(torch.diag(H), descending=True) + W = W[:, perm] + H = H[perm][:, perm] + + Losses = torch.zeros_like(W) + Q = torch.zeros_like(W) + + + damp = percdamp * torch.mean(H_diag) + diag = torch.arange(self.columns, device=self.dev) + H[diag, diag] += damp # 使 H 转变为全正数，保证正定 + + H_c = H.clone() + H_cpu = H_c.cpu() + del H_c + + # H1 = torch.linalg.cholesky(H) + # H2 = torch.cholesky_inverse(H1) + # H3 = torch.linalg.cholesky(H2, upper=True) + # Hinv = H + + H_cpu = np.linalg.inv(H_cpu) + H_cpu = np.linalg.cholesky(H_cpu) + Hinv = torch.tensor(H_cpu, device=self.dev) + + for i1 in range(0, self.columns, blocksize): # 以步长 blocksize (128) 循环到 columns + i2 = min(i1 + blocksize, self.columns) + count = i2 - i1 + + W1 = W[:, i1:i2].clone() + Q1 = torch.zeros_like(W1) + Err1 = torch.zeros_like(W1) + Losses1 = torch.zeros_like(W1) + Hinv1 = Hinv[i1:i2, i1:i2] + + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + + # Find by observer + # if groupsize != -1: + # if (i1 + i) % groupsize == 0: + # self.quantizer.find_params(W[:, (i1 + i):(i1 + i + groupsize)], weight=True) + + # q = quantize( + # w, self.scale, self.zero_point, self.quant_max + # ).flatten() + q = torch.fake_quantize_per_tensor_affine( + w, self.scale.item(), int(self.zero_point.item()), + self.quant_min, self.quant_max) + Q1[:, i] = q + Losses1[:, i] = (w - q) ** 2 / d ** 2 + + err1 = (w - q) / d + W1[:, i:] -= err1.unsqueeze(1).matmul(Hinv1[i, i:].unsqueeze(0)) + Err1[:, i] = err1 + + Q[:, i1:i2] = Q1 + Losses[:, i1:i2] = Losses1 / 2 + + W[:, i2:] -= Err1.matmul(Hinv[i1:i2, i2:]) + + + torch.cuda.synchronize() + # print('time %.2f' % (time.time() - tick)) + print('error', torch.sum(Losses).item()) + + if isinstance(self.layer_module, transformers.Conv1D): + Q = Q.t() + # self.weight.data = Q.reshape(self.weight.shape).to(self.weight.data.dtype) + W = Q.reshape(self.weight.shape).to(self.weight.data.dtype) # fixed fakequantize 并没有重置 parameter weight + + + # if DEBUG: + # print(torch.sum((self.layer(self.inp1) - self.out1) ** 2)) + + # self.layer_module.weight.data = W + # if ('Conv2d' in self.layer_type): + # print(torch.sum((self.layer_module(self.input) - self.layer_out) ** 2)) + # if ('Linear' in self.layer_type): + # print(torch.sum((self.layer_module(self.input) - self.output) ** 2)) + # del self.layer_module + + return W \ No newline at end of file diff --git a/mqbench/fake_quantize/lsq.py b/mqbench/fake_quantize/lsq.py index c133b0d6..39fe7032 100644 --- a/mqbench/fake_quantize/lsq.py +++ b/mqbench/fake_quantize/lsq.py @@ -1,5 +1,6 @@ import torch from torch.nn.parameter import Parameter +from torch.onnx import symbolic_helper from mqbench.fake_quantize.quantize_base import QuantizeBase from mqbench.utils import is_symmetric_quant, is_tracing_state @@ -36,8 +37,9 @@ def extra_repr(self): self.zero_point if self.ch_axis == -1 else 'List') def forward(self, X): + x_ori = X # Learnable fake quantize have to zero_point.float() to make it learnable. - if self.observer_enabled[0] == 1: + if self.observer_enabled[0] == 1:# or self.only_enable_observer: self.activation_post_process(X.detach()) _scale, _zero_point = self.activation_post_process.calculate_qparams() _scale = _scale.to(self.scale.device) @@ -53,7 +55,10 @@ def forward(self, X): self.scale.data.abs_() self.scale.data.clamp_(min=self.eps.item()) - if self.fake_quant_enabled[0] == 1: + if self.fake_quant_enabled[0] == 1:# and (not self.only_enable_observer or self.run_fquant_time > 0): + # if self.run_fquant_time > 0: + # print('wxc1 run_fquant_time') + # self.run_fquant_time -= 1 if is_symmetric_quant(self.qscheme): self.zero_point.data.zero_() else: @@ -72,15 +77,25 @@ def forward(self, X): X = _fake_quantize_learnable_per_channel_affine_training( X, self.scale, self.zero_point, self.ch_axis, self.quant_min, self.quant_max, grad_factor) + x_ori = X else: if self.use_grad_scaling: grad_factor = 1.0 / (X.numel() * self.quant_max) ** 0.5 else: grad_factor = 1.0 + scale, zero_point = self.scale, self.zero_point + # if self.only_enable_observer: + # scale, zero_point = 1, 0 X = torch._fake_quantize_learnable_per_tensor_affine( - X, self.scale, self.zero_point, + x_ori, scale, zero_point, self.quant_min, self.quant_max, grad_factor) - return X + diff = x_ori - X + if self.only_enable_observer: + x_ori = X + diff.detach() + else: + x_ori = X + + return x_ori def _fake_quantize_learnable_per_channel_affine_training(x, scale, zero_point, ch_axis, quant_min, quant_max, grad_factor): @@ -107,4 +122,11 @@ def forward(ctx, x, scale, zero_point, ch_axis, quant_min, quant_max, grad_facto @staticmethod def symbolic(g, x, scale, zero_point, ch_axis, quant_min, quant_max, grad_factor): - return g.op("::FakeQuantizeLearnablePerchannelAffine", x, scale, zero_point, quant_min_i=quant_min, quant_max_i=quant_max) + output = g.op("MQBench_custom::FakeQuantizeLearnablePerchannelAffine", x, scale, zero_point, quant_min_i=quant_min, quant_max_i=quant_max) + + input_shape = symbolic_helper._get_tensor_sizes(x) + if input_shape is not None and hasattr(x.type(), 'with_sizes'): + output_type = x.type().with_sizes(input_shape) + output.setType(output_type) + + return output \ No newline at end of file diff --git a/mqbench/fake_quantize/quantize_base.py b/mqbench/fake_quantize/quantize_base.py index 8fe8fc18..36e34e26 100644 --- a/mqbench/fake_quantize/quantize_base.py +++ b/mqbench/fake_quantize/quantize_base.py @@ -5,7 +5,7 @@ from mqbench.utils import is_symmetric_quant -_version_under_1100 = int(torch.__version__.split('.')[1]) < 10 +_version_under_1100 = (int(torch.__version__.split('.')[1]) < 10) and (int(torch.__version__.split('.')[0]) == 1) class QuantizeBase(FakeQuantizeBase): r""" This is an extension of the FakeQuantize module in fake_quantize.py, which @@ -35,6 +35,12 @@ def __init__(self, observer=MovingAverageMinMaxObserver, **observer_kwargs): bitrange = torch.tensor(self.quant_max - self.quant_min + 1).double() self.bitwidth = int(torch.log2(bitrange).item()) self.is_symmetric_quant = is_symmetric_quant(self.qscheme) + self.only_enable_observer = False + self.run_fquant_time = 0 + + def enable_only_observer(self, enable = True): + self.only_enable_observer = enable + self.run_fquant_time = 1 @torch.jit.export def calculate_qparams(self): diff --git a/mqbench/fuser_method_mappings.py b/mqbench/fuser_method_mappings.py index eb1a56af..535d93f7 100644 --- a/mqbench/fuser_method_mappings.py +++ b/mqbench/fuser_method_mappings.py @@ -2,8 +2,6 @@ import torch import torch.nn as nn -from torch.quantization.fx.fusion_patterns import ConvBNReLUFusion, ModuleReLUFusion -from torch.quantization.fx.quantization_types import QuantizerCls from torch.fx.graph import Node import mqbench.nn as qnn @@ -12,6 +10,277 @@ from mqbench.utils.fusion import fuse_deconv_bn_eval from mqbench.nn.modules import FrozenBatchNorm2d +from collections import OrderedDict +from torch.ao.quantization.fuser_method_mappings import get_fuser_method +from abc import ABC, abstractmethod +from typing import Any, Callable, Dict +QuantizerCls = Any + +# pattern for conv bn fusion +DEFAULT_FUSION_PATTERNS = OrderedDict() +def register_fusion_pattern(pattern): + def insert(fn): + DEFAULT_FUSION_PATTERNS[pattern] = fn + return fn + return insert + +# turn foo.bar -> ['foo', 'bar'] +def _parent_name(target): + r = target.rsplit('.', 1) + if len(r) == 1: + return '', r[0] + else: + return r[0], r[1] + +# --------------------- +# Fusion Pattern Registrations +# --------------------- + +# Base Pattern Handler +class FuseHandler(ABC): + """ Base handler class for the fusion patterns + """ + def __init__(self, quantizer: QuantizerCls, node: Node): + pass + + @abstractmethod + def fuse(self, quantizer: QuantizerCls, load_arg: Callable, + fuse_custom_config_dict: Dict[str, Any] = None) -> Node: + pass + +@register_fusion_pattern((torch.nn.ReLU, torch.nn.Conv1d)) +@register_fusion_pattern((torch.nn.ReLU, torch.nn.Conv2d)) +@register_fusion_pattern((torch.nn.ReLU, torch.nn.Conv3d)) +@register_fusion_pattern((torch.nn.functional.relu, torch.nn.Conv1d)) +@register_fusion_pattern((torch.nn.functional.relu, torch.nn.Conv2d)) +@register_fusion_pattern((torch.nn.functional.relu, torch.nn.Conv3d)) +@register_fusion_pattern((torch.nn.BatchNorm1d, torch.nn.Conv1d)) +@register_fusion_pattern((torch.nn.BatchNorm2d, torch.nn.Conv2d)) +@register_fusion_pattern((torch.nn.BatchNorm3d, torch.nn.Conv3d)) +@register_fusion_pattern((torch.nn.ReLU, (torch.nn.BatchNorm1d, torch.nn.Conv1d))) +@register_fusion_pattern((torch.nn.ReLU, (torch.nn.BatchNorm2d, torch.nn.Conv2d))) +@register_fusion_pattern((torch.nn.ReLU, (torch.nn.BatchNorm3d, torch.nn.Conv3d))) +@register_fusion_pattern((torch.nn.functional.relu, (torch.nn.BatchNorm1d, torch.nn.Conv1d))) +@register_fusion_pattern((torch.nn.functional.relu, (torch.nn.BatchNorm2d, torch.nn.Conv2d))) +@register_fusion_pattern((torch.nn.functional.relu, (torch.nn.BatchNorm3d, torch.nn.Conv3d))) +class ConvBNReLUFusion(FuseHandler): + def __init__(self, quantizer: QuantizerCls, node: Node): + super().__init__(quantizer, node) + self.relu_node = None + self.bn_node = None + if (node.op == 'call_function' and node.target is torch.nn.functional.relu) or \ + (node.op == 'call_module' and type(quantizer.modules[node.target]) == torch.nn.ReLU): + self.relu_node = node + assert isinstance(node.args[0], Node) + node = node.args[0] + assert node.op == 'call_module' + if type(quantizer.modules[node.target]) in [torch.nn.BatchNorm1d, torch.nn.BatchNorm2d, torch.nn.BatchNorm3d]: + self.bn_node = node + self.bn = quantizer.modules[self.bn_node.target] + assert isinstance(node.args[0], Node) + node = node.args[0] + assert node.op == 'call_module' + self.conv_node = node + self.conv = quantizer.modules[self.conv_node.target] + + def fuse(self, quantizer: QuantizerCls, load_arg: Callable, + fuse_custom_config_dict: Dict[str, Any] = None) -> Node: + if fuse_custom_config_dict is None: + fuse_custom_config_dict = {} + additional_fuser_method_mapping = fuse_custom_config_dict.get("additional_fuser_method_mapping", {}) + op_list = [] + if self.relu_node is not None: + # since relu can be used multiple times, we'll need to create a relu module for each match + if self.relu_node.op == 'call_module': + relu = torch.nn.ReLU(quantizer.modules[self.relu_node.target].inplace) + else: + # TODO: get inplace argument from functional + relu = torch.nn.ReLU() + op_list.append(relu) + relu.training = self.conv.training + if self.bn_node is not None: + op_list.append(self.bn) + op_list.append(self.conv) + else: + assert self.bn_node is not None + op_list.append(self.bn) + op_list.append(self.conv) + + # the modules are added in order of relu - bn - conv + # so we need to correct it + op_list.reverse() + op_type_list = tuple(type(m) for m in op_list) + conv_parent_name, conv_name = _parent_name(self.conv_node.target) + fuser_method = get_fuser_method(op_type_list, additional_fuser_method_mapping) + if fuser_method is None: + raise NotImplementedError("Cannot fuse modules: {}".format(op_type_list)) + fused = fuser_method(*op_list) + setattr(quantizer.modules[conv_parent_name], conv_name, fused) + + # TODO: do we need to make sure bn is only used once? + if self.bn_node is not None: + parent_name, name = _parent_name(self.bn_node.target) + setattr(quantizer.modules[parent_name], name, torch.nn.Identity()) + # relu may be used multiple times, so we don't set relu to identity + return quantizer.fused_graph.node_copy(self.conv_node, load_arg) + +@register_fusion_pattern((torch.nn.functional.relu, torch.nn.Linear)) +@register_fusion_pattern((torch.nn.ReLU, torch.nn.Linear)) +@register_fusion_pattern((torch.nn.functional.relu, torch.nn.BatchNorm2d)) +@register_fusion_pattern((torch.nn.ReLU, torch.nn.BatchNorm2d)) +@register_fusion_pattern((torch.nn.functional.relu, torch.nn.BatchNorm3d)) +@register_fusion_pattern((torch.nn.ReLU, torch.nn.BatchNorm3d)) +class ModuleReLUFusion(FuseHandler): + def __init__(self, quantizer: QuantizerCls, node: Node): + super().__init__(quantizer, node) + self.relu_node = node + assert isinstance(node.args[0], Node) + node = node.args[0] + assert node.op == 'call_module' + self.module_node = node + self.module = quantizer.modules[self.module_node.target] + + def fuse(self, quantizer: QuantizerCls, load_arg: Callable, + fuse_custom_config_dict: Dict[str, Any] = None) -> Node: + if fuse_custom_config_dict is None: + fuse_custom_config_dict = {} + additional_fuser_method_mapping = fuse_custom_config_dict.get("additional_fuser_method_mapping", {}) + op_list = [] + # since relu can be used multiple times, we'll need to create a relu module for each match + if self.relu_node.op == 'call_module': + relu = torch.nn.ReLU(quantizer.modules[self.relu_node.target].inplace) + else: + # TODO: get inplace argument from functional + relu = torch.nn.ReLU() + relu.training = self.module.training + op_list.append(relu) + op_list.append(self.module) + + op_list.reverse() + op_type_list = tuple(type(m) for m in op_list) + module_parent_name, module_name = _parent_name(self.module_node.target) + fuser_method = get_fuser_method(op_type_list, additional_fuser_method_mapping) + setattr(quantizer.modules[module_parent_name], module_name, fuser_method(*op_list)) + return quantizer.fused_graph.node_copy(self.module_node, load_arg) + +import torch.nn.intrinsic as nni +from typing import Union, Callable, Tuple, Dict, Optional, Type +from torch.ao.quantization.utils import get_combined_dict + +def fuse_conv_bn(conv, bn): + r"""Given the conv and bn modules, fuses them and returns the fused module + + Args: + conv: Module instance of type conv2d/conv3d + bn: Spatial BN instance that needs to be fused with the conv + + Examples:: + + >>> m1 = nn.Conv2d(10, 20, 3) + >>> b1 = nn.BatchNorm2d(20) + >>> m2 = fuse_conv_bn(m1, b1) + """ + assert(conv.training == bn.training),\ + "Conv and BN both must be in the same mode (train or eval)." + + fused_module_class_map = { + nn.Conv1d: nni.ConvBn1d, + nn.Conv2d: nni.ConvBn2d, + nn.Conv3d: nni.ConvBn3d, + } + + if conv.training: + assert bn.num_features == conv.out_channels, 'Output channel of Conv2d must match num_features of BatchNorm2d' + assert bn.affine, 'Only support fusing BatchNorm2d with affine set to True' + assert bn.track_running_stats, 'Only support fusing BatchNorm2d with tracking_running_stats set to True' + fused_module_class = fused_module_class_map.get((type(conv)), None) + if fused_module_class is not None: + return fused_module_class(conv, bn) + else: + raise NotImplementedError("Cannot fuse train modules: {}".format((conv, bn))) + else: + return nn.utils.fuse_conv_bn_eval(conv, bn) + +def fuse_conv_bn_relu(conv, bn, relu): + r"""Given the conv and bn modules, fuses them and returns the fused module + + Args: + conv: Module instance of type conv2d/conv3d + bn: Spatial BN instance that needs to be fused with the conv + + Examples:: + + >>> m1 = nn.Conv2d(10, 20, 3) + >>> b1 = nn.BatchNorm2d(20) + >>> r1 = nn.ReLU(inplace=False) + >>> m2 = fuse_conv_bn_relu(m1, b1, r1) + """ + assert(conv.training == bn.training == relu.training),\ + "Conv and BN both must be in the same mode (train or eval)." + fused_module : Optional[Type[nn.Sequential]] = None + if conv.training: + map_to_fused_module_train = { + nn.Conv1d: nni.ConvBnReLU1d, + nn.Conv2d: nni.ConvBnReLU2d, + nn.Conv3d: nni.ConvBnReLU3d, + } + assert bn.num_features == conv.out_channels, 'Output channel of Conv must match num_features of BatchNorm' + assert bn.affine, 'Only support fusing BatchNorm with affine set to True' + assert bn.track_running_stats, 'Only support fusing BatchNorm with tracking_running_stats set to True' + fused_module = map_to_fused_module_train.get(type(conv), None) + if fused_module is not None: + return fused_module(conv, bn, relu) + else: + raise NotImplementedError("Cannot fuse train modules: {}".format((conv, bn, relu))) + else: + map_to_fused_module_eval = { + nn.Conv1d: nni.ConvReLU1d, + nn.Conv2d: nni.ConvReLU2d, + nn.Conv3d: nni.ConvReLU3d, + } + fused_module = map_to_fused_module_eval.get(type(conv), None) + if fused_module is not None: + fused_conv = nn.utils.fusion.fuse_conv_bn_eval(conv, bn) + return fused_module(fused_conv, relu) + else: + raise NotImplementedError("Cannot fuse eval modules: {}".format((conv, bn, relu))) + +def fuse_linear_bn(linear, bn): + r"""Given the linear and bn modules, fuses them and returns the fused module + + Args: + linear: Module instance of type Linear + bn: BatchNorm1d instance that needs to be fused with the linear layer + + Examples:: + + >>> m1 = nn.Linear(20, 10) + >>> b1 = nn.BatchNorm1d(10) + >>> m2 = fuse_linear_bn(m1, b1) + """ + assert(linear.training == bn.training),\ + "Linear and BN both must be in the same mode (train or eval)." + + if linear.training: + raise Exception("Fusing Linear+BatchNorm not yet supported in training.") + else: + return nn.utils.fusion.fuse_linear_bn_eval(linear, bn) + +DEFAULT_OP_LIST_TO_FUSER_METHOD : Dict[Tuple, Union[nn.Sequential, Callable]] = { + (nn.Conv1d, nn.BatchNorm1d): fuse_conv_bn, + (nn.Conv1d, nn.BatchNorm1d, nn.ReLU): fuse_conv_bn_relu, + (nn.Conv2d, nn.BatchNorm2d): fuse_conv_bn, + (nn.Conv2d, nn.BatchNorm2d, nn.ReLU): fuse_conv_bn_relu, + (nn.Conv3d, nn.BatchNorm3d): fuse_conv_bn, + (nn.Conv3d, nn.BatchNorm3d, nn.ReLU): fuse_conv_bn_relu, + (nn.Conv1d, nn.ReLU): nni.ConvReLU1d, + (nn.Conv2d, nn.ReLU): nni.ConvReLU2d, + (nn.Conv3d, nn.ReLU): nni.ConvReLU3d, + (nn.Linear, nn.BatchNorm1d): fuse_linear_bn, + (nn.Linear, nn.ReLU): nni.LinearReLU, + (nn.BatchNorm2d, nn.ReLU): nni.BNReLU2d, + (nn.BatchNorm3d, nn.ReLU): nni.BNReLU3d, +} class ConvFreezebnReLUFusion(ConvBNReLUFusion): def __init__(self, quantizer: QuantizerCls, node: Node): @@ -202,11 +471,8 @@ def _sort_fusion_patterns(pats): for key in keys: pats.move_to_end(key) - # Sinse additional_fuser_method_mapping will not be set because fuser.py:54 # do not pass this dict. -from torch.quantization.fuser_method_mappings import DEFAULT_OP_LIST_TO_FUSER_METHOD -from torch.quantization.fx.pattern_utils import DEFAULT_FUSION_PATTERNS from torch.quantization.quantization_mappings import DEFAULT_QAT_MODULE_MAPPINGS DEFAULT_OP_LIST_TO_FUSER_METHOD.update( diff --git a/mqbench/fusion_method.py b/mqbench/fusion_method.py index 2bbf693c..30326a1a 100644 --- a/mqbench/fusion_method.py +++ b/mqbench/fusion_method.py @@ -1,7 +1,14 @@ import torch import torch.nn.intrinsic.qat as nniqat from torch.nn.utils.fusion import fuse_conv_bn_eval, fuse_linear_bn_eval -from torch.quantization.fx.utils import _parent_name + +# turn foo.bar -> ['foo', 'bar'] +def _parent_name(target): + r = target.rsplit('.', 1) + if len(r) == 1: + return '', r[0] + else: + return r[0], r[1] import mqbench.nn.intrinsic as qnni import mqbench.nn.intrinsic.qat as qnniqat @@ -11,6 +18,90 @@ from mqbench.quantization.default_bias_fake_quant import bias_fake_quantizer +@register_convert_function(qnniqat.Linear_sophgo) +def convert_qnniqat_linear(model, fused_node): + modules = dict(model.named_modules()) + fused_module = modules[fused_node.target] + # Create a Linear from FusedModule. + linear = torch.nn.Linear(fused_module.in_features, fused_module.out_features, fused_module.bias is not None) + linear.weight = fused_module.weight + if fused_module.bias is not None: + linear.bias = fused_module.bias + # We need nn.qat.linear here to export weight quantize node. + linear.qconfig = fused_module.qconfig + linear = torch.nn.qat.Linear.from_float(linear) + # Attach weight fake quantize params. + linear.weight_fake_quant = fused_module.weight_fake_quant + linear_parent_name, linear_name = _parent_name(fused_node.target) + setattr(modules[linear_parent_name], linear_name, linear) + +@register_convert_function(qnnqat.Conv2d_sophgo) +def convert_qnnqat_conv2d(model, fused_node): + modules = dict(model.named_modules()) + fused_module = modules[fused_node.target] + # Create a Conv2d from FusedModule. + conv = torch.nn.Conv2d(fused_module.in_channels, fused_module.out_channels, fused_module.kernel_size, + fused_module.stride, fused_module.padding, fused_module.dilation, + fused_module.groups, fused_module.bias is not None, fused_module.padding_mode) + conv.weight = fused_module.weight + if fused_module.bias is not None: + conv.bias = fused_module.bias + # We need nn.qat.conv here to export weight quantize node. + conv.qconfig = fused_module.qconfig + conv = torch.nn.qat.Conv2d.from_float(conv) + # Attach weight fake quantize params. + conv.weight_fake_quant = fused_module.weight_fake_quant + conv_parent_name, conv_name = _parent_name(fused_node.target) + setattr(modules[conv_parent_name], conv_name, conv) + +@register_convert_function(qnnqat.ConvTranspose2d_sophgo) +def convert_qnnqat_deconv2d(model, fused_node): + modules = dict(model.named_modules()) + fused_module = modules[fused_node.target] + deconv = torch.nn.ConvTranspose2d(fused_module.in_channels, fused_module.out_channels, fused_module.kernel_size, + stride=fused_module.stride, padding=fused_module.padding, output_padding=fused_module.output_padding, + groups=fused_module.groups, bias=fused_module.bias is not None, + dilation=fused_module.dilation, + padding_mode=fused_module.padding_mode) + deconv.weight = fused_module.weight + if fused_module.bias is not None: + deconv.bias = fused_module.bias + fused_deconv = fuse_deconv_bn_eval(deconv.eval(), fused_module.bn) + fused_deconv.qconfig = fused_module.qconfig + fused_deconv = qnnqat.ConvTranspose2d.from_float(fused_deconv) + fused_deconv.weight_fake_quant = fused_module.weight_fake_quant + deconv_parent_name, deconv_name = _parent_name(fused_node.target) + setattr(modules[deconv_parent_name], deconv_name, fused_deconv) + +@register_convert_function(qnniqat.LinearReLU_sophgo) +def linearert_qnniqat_linearrelu(model, fused_node): + convert_qnniqat_linear(model, fused_node) + modules = dict(model.named_modules()) + fused_module = modules[fused_node.target] + # We need to Insert Relu after Merged linear. + linear_parent_name, linear_name = _parent_name(fused_node.target) + relu_name = 'relu' + # Maybe has another name, but we cannot know for now. + if not hasattr(modules[linear_parent_name], relu_name): + setattr(modules[linear_parent_name], relu_name, + torch.nn.ReLU(inplace=True).train(fused_module.training)) + # Update modules. + modules = dict(model.named_modules()) + graph = model.graph + nodes = list(model.graph.nodes) + with graph.inserting_after(fused_node): + relu_node_name = relu_name if linear_parent_name == "" else "{}.{}".format(linear_parent_name, relu_name) + assert relu_node_name in modules and isinstance(modules[relu_node_name], torch.nn.ReLU) + inserted_node = graph.create_node("call_module", relu_node_name, (fused_node,), {}) + for _node in nodes: + for i, _arg in enumerate(_node.args): + if _arg == fused_node: + _tmp = list(_node.args) + _tmp[i] = inserted_node + _node.args = tuple(_tmp) + model.recompile() + model.graph.lint() + @register_convert_function(qnni.LinearBn1d) def convert_qnni_linearbn(model, fused_node): modules = dict(model.named_modules()) @@ -21,6 +112,7 @@ def convert_qnni_linearbn(model, fused_node): @register_convert_function(qnniqat.LinearBn1d) +@register_convert_function(qnniqat.LinearBn1d_sophgo) def convert_qnniqat_linearbn(model, fused_node): modules = dict(model.named_modules()) fused_module = modules[fused_node.target] @@ -32,10 +124,10 @@ def convert_qnniqat_linearbn(model, fused_node): # Merge Linear + BN fused_linear = fuse_linear_bn_eval(linear.eval(), fused_module.bn) # We need nn.qat.linear here to export weight quantize node. - linear.qconfig = fused_module.qconfig - linear = torch.nn.qat.Linear.from_float(linear) + fused_linear.qconfig = fused_module.qconfig + fused_linear = torch.nn.qat.Linear.from_float(fused_linear) # Attach weight fake quantize params. - linear.weight_fake_quant = fused_module.weight_fake_quant + fused_linear.weight_fake_quant = fused_module.weight_fake_quant linear_parent_name, linear_name = _parent_name(fused_node.target) setattr(modules[linear_parent_name], linear_name, fused_linear) @@ -43,6 +135,7 @@ def convert_qnniqat_linearbn(model, fused_node): @register_convert_function(qnniqat.ConvFreezebn2d) @register_convert_function(nniqat.ConvBn2d) @register_convert_function(nniqat.ConvBn3d) +@register_convert_function(qnniqat.ConvBn2d_sophgo) def convert_nniqat_convbn(model, fused_node): """nniqat.ConvBn2d ----> nn.Conv2d ----> nniqat.Conv2d """ @@ -53,6 +146,8 @@ def convert_nniqat_convbn(model, fused_node): nniqat.ConvBnReLU2d: torch.nn.Conv2d, nniqat.ConvBn3d: torch.nn.Conv3d, nniqat.ConvBnReLU3d: torch.nn.Conv3d, + qnniqat.ConvBn2d_sophgo: torch.nn.Conv2d, + qnniqat.ConvBnReLU2d_sophgo: torch.nn.Conv2d, } fused_qat_module_class_map = { torch.nn.Conv2d: torch.nn.qat.Conv2d, @@ -82,6 +177,7 @@ def convert_nniqat_convbn(model, fused_node): @register_convert_function(qnniqat.ConvFreezebnReLU2d) @register_convert_function(nniqat.ConvBnReLU2d) @register_convert_function(nniqat.ConvBnReLU3d) +@register_convert_function(qnniqat.ConvBnReLU2d_sophgo) def convert_nniqat_convbnrelu(model, fused_node): convert_nniqat_convbn(model, fused_node) modules = dict(model.named_modules()) @@ -110,6 +206,34 @@ def convert_nniqat_convbnrelu(model, fused_node): model.recompile() model.graph.lint() +@register_convert_function(qnniqat.ConvReLU2d_sophgo) +def convert_nniqat_convrelu(model, fused_node): + convert_qnnqat_conv2d(model, fused_node) + modules = dict(model.named_modules()) + fused_module = modules[fused_node.target] + # We need to Insert Relu after Merged conv. + conv_parent_name, conv_name = _parent_name(fused_node.target) + relu_name = 'relu' + # Maybe has another name, but we cannot know for now. + if not hasattr(modules[conv_parent_name], relu_name): + setattr(modules[conv_parent_name], relu_name, + torch.nn.ReLU(inplace=True).train(fused_module.training)) + # Update modules. + modules = dict(model.named_modules()) + graph = model.graph + nodes = list(model.graph.nodes) + with graph.inserting_after(fused_node): + relu_node_name = relu_name if conv_parent_name == "" else "{}.{}".format(conv_parent_name, relu_name) + assert relu_node_name in modules and isinstance(modules[relu_node_name], torch.nn.ReLU) + inserted_node = graph.create_node("call_module", relu_node_name, (fused_node,), {}) + for _node in nodes: + for i, _arg in enumerate(_node.args): + if _arg == fused_node: + _tmp = list(_node.args) + _tmp[i] = inserted_node + _node.args = tuple(_tmp) + model.recompile() + model.graph.lint() @register_convert_function(qnni.ConvTransposeFreezebn2d) @register_convert_function(qnni.ConvTransposeBn2d) @@ -134,6 +258,7 @@ def convert_qnni_deconvbn(model, fused_node): @register_convert_function(qnniqat.ConvTransposeFreezebn2d) @register_convert_function(qnniqat.ConvTransposeBn2d) +@register_convert_function(qnniqat.ConvTransposeBn2d_sophgo) def convert_qnniqat_deconvbn(model, fused_node): modules = dict(model.named_modules()) fused_module = modules[fused_node.target] @@ -158,6 +283,7 @@ def convert_qnniqat_deconvbn(model, fused_node): @register_convert_function(qnni.ConvTransposeFreezebnReLU2d) @register_convert_function(qnni.ConvTransposeBnReLU2d) +@register_convert_function(qnniqat.ConvTransposeBnReLU2d_sophgo) def convert_qnni_deconvbnrelu(model, fused_node): convert_qnni_deconvbn(model, fused_node) modules = dict(model.named_modules()) @@ -186,6 +312,7 @@ def convert_qnni_deconvbnrelu(model, fused_node): @register_convert_function(qnniqat.ConvTransposeFreezebnReLU2d) @register_convert_function(qnniqat.ConvTransposeBnReLU2d) +@register_convert_function(qnniqat.ConvTransposeBnReLU2d_sophgo) def convert_qnniqat_deconvbnrelu(model, fused_node): convert_qnniqat_deconvbn(model, fused_node) modules = dict(model.named_modules()) @@ -263,7 +390,7 @@ def convert_qnniqat_convbnrelu(model, fused_node): with graph.inserting_after(fused_node): relu_node_name = relu_name if conv_parent_name == "" else "{}.{}".format(conv_parent_name, relu_name) assert relu_node_name in modules and isinstance(modules[relu_node_name], torch.nn.ReLU) - inserted_node = graph.create_node("call_module", relu_node_name, (fused_node,), {}) + inserted_node = graph.create_node("call_module", relu_node_name, (fused_node,), {}, name = fused_node.name) for _node in nodes: for i, _arg in enumerate(_node.args): if _arg == fused_node: diff --git a/mqbench/nn/intrinsic/qat/modules/__init__.py b/mqbench/nn/intrinsic/qat/modules/__init__.py index 3118f9bf..7b13120a 100644 --- a/mqbench/nn/intrinsic/qat/modules/__init__.py +++ b/mqbench/nn/intrinsic/qat/modules/__init__.py @@ -2,3 +2,7 @@ from .deconv_fused import ConvTransposeBnReLU2d, ConvTransposeBn2d, ConvTransposeReLU2d from .conv_fused import ConvBnReLU2d, ConvBn2d, ConvReLU2d from .freezebn import ConvFreezebn2d, ConvFreezebnReLU2d, ConvTransposeFreezebn2d, ConvTransposeFreezebnReLU2d + +from .conv_fused_sophgo_tpu import ConvBnReLU2d_sophgo, ConvBn2d_sophgo, ConvReLU2d_sophgo +from .linear_fused_sophgo_tpu import LinearBn1d_sophgo, LinearReLU_sophgo, Linear_sophgo +from .deconv_fused_sophgo_tpu import ConvTransposeBnReLU2d_sophgo, ConvTransposeBn2d_sophgo, ConvTransposeReLU2d_sophgo diff --git a/mqbench/nn/intrinsic/qat/modules/conv_fused_sophgo_tpu.py b/mqbench/nn/intrinsic/qat/modules/conv_fused_sophgo_tpu.py new file mode 100644 index 00000000..25179d23 --- /dev/null +++ b/mqbench/nn/intrinsic/qat/modules/conv_fused_sophgo_tpu.py @@ -0,0 +1,487 @@ +import math + +import torch +import torch.nn as nn +import torch.nn.intrinsic as nni +import torch.nn.functional as F +from torch.nn import init +from torch.nn.intrinsic import _FusedModule +from torch.nn.parameter import Parameter +from torch.nn.modules.utils import _pair + +from typing import TypeVar + + +import mqbench.nn.qat as qnnqat +import torch.nn.qat.modules as nnqat +# from mqbench.quantization.default_bias_fake_quant import bias_fake_quantizer + +_BN_CLASS_MAP = { + 1: nn.BatchNorm1d, + 2: nn.BatchNorm2d, + 3: nn.BatchNorm3d, +} + +MOD = TypeVar('MOD', bound=nn.modules.conv._ConvNd) + + +class _ConvBnNd(nn.modules.conv._ConvNd, _FusedModule): + + _version = 2 + _FLOAT_MODULE = MOD + + def __init__(self, + # ConvNd args + in_channels, out_channels, kernel_size, stride, + padding, dilation, transposed, output_padding, + groups, + bias, + padding_mode, + # BatchNormNd args + # num_features: out_channels + eps=1e-05, momentum=0.1, + # affine: True + # track_running_stats: True + # Args for this module + freeze_bn=False, + qconfig=None, + dim=2): + nn.modules.conv._ConvNd.__init__(self, in_channels, out_channels, kernel_size, + stride, padding, dilation, transposed, + output_padding, groups, False, padding_mode) + assert qconfig, 'qconfig must be provided for QAT module' + self.qconfig = qconfig + self.freeze_bn = freeze_bn if self.training else True + self.bn = _BN_CLASS_MAP[dim](out_channels, eps, momentum, True, True) + self.weight_fake_quant = self.qconfig.weight() + if bias: + self.bias = Parameter(torch.empty(out_channels)) + else: + self.register_parameter('bias', None) + # self.bias_fake_quant = bias_fake_quantizer() + self.reset_bn_parameters() + + # this needs to be called after reset_bn_parameters, + # as they modify the same state + if self.training: + if freeze_bn: + self.freeze_bn_stats() + else: + self.update_bn_stats() + else: + self.freeze_bn_stats() + + def reset_running_stats(self): + self.bn.reset_running_stats() + + def reset_bn_parameters(self): + self.bn.reset_running_stats() + init.uniform_(self.bn.weight) + init.zeros_(self.bn.bias) + # note: below is actully for conv, not BN + if self.bias is not None: + fan_in, _ = init._calculate_fan_in_and_fan_out(self.weight) + bound = 1 / math.sqrt(fan_in) + init.uniform_(self.bias, -bound, bound) + + def reset_parameters(self): + super(_ConvBnNd, self).reset_parameters() + + def update_bn_stats(self): + self.freeze_bn = False + self.bn.training = True + return self + + def freeze_bn_stats(self): + self.freeze_bn = True + self.bn.training = False + return self + +# def _forward(self, input): +# assert self.bn.running_var is not None +# running_std = torch.sqrt(self.bn.running_var + self.bn.eps) +# scale_factor = self.bn.weight / running_std +# weight_shape = [1] * len(self.weight.shape) +# weight_shape[0] = -1 +# bias_shape = [1] * len(self.weight.shape) +# bias_shape[1] = -1 +# scaled_weight = self.weight_fake_quant(self.weight * scale_factor.reshape(weight_shape)) +# # using zero bias here since the bias for original conv +# # will be added later +# if self.bias is not None: +# zero_bias = torch.zeros_like(self.bias) +# conv_bias = self.bias +# else: +# zero_bias = torch.zeros(self.out_channels, device=scaled_weight.device) +# conv_bias = torch.zeros_like(zero_bias, device=scaled_weight.device) +# if self.bn.affine: +# full_bias = (conv_bias - self.bn.running_mean) / running_std * self.bn.weight + self.bn.bias +# else: +# full_bias = (conv_bias - self.bn.running_mean) / running_std +# quant_bias = self.bias_fake_quant(full_bias) +# conv_with_bias = self._conv_forward(input, scaled_weight, quant_bias) +# conv_orig = (conv_with_bias - full_bias.reshape(bias_shape)) / scale_factor.reshape(bias_shape) + conv_bias.reshape(bias_shape) +# conv = self.bn(conv_orig) +# return conv + + + def bias_fake_quant_proc(self, bias, scale_w, in_scale): + scale = scale_w*in_scale + if torch.nonzero(scale).size()[0] != scale.numel(): + print('error! scale has 0, scale:', scale) + bias_q = bias/scale + bias = (bias_q.round()-bias_q).detach() + bias_q + bias = bias*scale + return bias + + # def _forward(self, input): + # # print('xxx2') + # assert self.bn.running_var is not None + # running_std = torch.sqrt(self.bn.running_var + self.bn.eps) + # scale_factor = self.bn.weight / running_std + # weight_shape = [1] * len(self.weight.shape) + # weight_shape[0] = -1 + # bias_shape = [1] * len(self.weight.shape) + # bias_shape[1] = -1 + # if torch.isnan(self.weight).any(): + # print('weight have nan') + # if self.input_fake_quantizer is not None and torch.isnan(self.input_fake_quantizer.scale).any(): + # print('input_fake_quantizer.scale have nan') + # if self.bias is not None and torch.isnan(self.bias).any(): + # print('weight have nan') + # scaled_weight = self.weight_fake_quant(self.weight * scale_factor.reshape(weight_shape)) + # #bias伪量化 + # if self.weight_fake_quant.fake_quant_enabled[0] == 1: + # _, fused_bias = nn.utils.fuse_conv_bn_weights(self.weight, self.bias, + # self.bn.running_mean, self.bn.running_var, self.bn.eps, self.bn.weight, self.bn.bias) + # in_scale = self.input_fake_quantizer.scale #从上一个activation_fake_quant节点获取scale + # scale_fused_bias = self.bias_fake_quant_proc(fused_bias, self.weight_fake_quant.scale, in_scale) + # diff_fused_bias = fused_bias - scale_fused_bias + # # using zero bias here since the bias for original conv + # # will be added later + # if self.bias is not None: + # zero_bias = torch.zeros_like(self.bias) + # else: + # zero_bias = torch.zeros(self.out_channels, device=scaled_weight.device) + # conv = self._conv_forward(input, scaled_weight, zero_bias) + # conv_orig = conv / scale_factor.reshape(bias_shape) + # if self.bias is not None: + # conv_orig = conv_orig + self.bias.reshape(bias_shape) + # conv = self.bn(conv_orig) + # if self.weight_fake_quant.fake_quant_enabled[0] == 1: + # conv -= diff_fused_bias.reshape(bias_shape) #这里从推导看应该是减 + # return conv + + def _forward(self, input): + assert self.bn.running_var is not None + running_std = torch.sqrt(self.bn.running_var + self.bn.eps) + scale_factor = self.bn.weight / running_std + weight_shape = [1] * len(self.weight.shape) + weight_shape[0] = -1 + bias_shape = [1] * len(self.weight.shape) + bias_shape[1] = -1 + scaled_weight = self.weight_fake_quant(self.weight * scale_factor.reshape(weight_shape)) + # using zero bias here since the bias for original conv + # will be added later + if self.bias is not None: + zero_bias = torch.zeros_like(self.bias) + conv_bias = self.bias + else: + zero_bias = torch.zeros(self.out_channels, device=scaled_weight.device) + conv_bias = torch.zeros_like(zero_bias, device=scaled_weight.device) + if self.bn.affine: + full_bias = (conv_bias - self.bn.running_mean) / running_std * self.bn.weight + self.bn.bias + else: + full_bias = (conv_bias - self.bn.running_mean) / running_std + + if full_bias is not None and self.weight_fake_quant.fake_quant_enabled[0] == 1: + quant_bias = self.bias_fake_quant_proc(full_bias, self.weight_fake_quant.scale, self.input_fake_quantizer.scale) + else: + quant_bias = full_bias + + conv_with_bias = self._conv_forward(input, scaled_weight, quant_bias) + conv_orig = (conv_with_bias - full_bias.reshape(bias_shape)) / scale_factor.reshape(bias_shape) + conv_bias.reshape(bias_shape) + conv = self.bn(conv_orig) + return conv + + def extra_repr(self): + # TODO(jerryzh): extend + return super(_ConvBnNd, self).extra_repr() + + def forward(self, input): + return self._forward(input) + + def train(self, mode=True): + """ + Batchnorm's training behavior is using the self.training flag. Prevent + changing it if BN is frozen. This makes sure that calling `model.train()` + on a model with a frozen BN will behave properly. + """ + self.training = mode + if not self.freeze_bn: + for module in self.children(): + module.train(mode) + return self + + # ===== Serialization version history ===== + # + # Version 1/None + # self + # |--- weight : Tensor + # |--- bias : Tensor + # |--- gamma : Tensor + # |--- beta : Tensor + # |--- running_mean : Tensor + # |--- running_var : Tensor + # |--- num_batches_tracked : Tensor + # + # Version 2 + # self + # |--- weight : Tensor + # |--- bias : Tensor + # |--- bn : Module + # |--- weight : Tensor (moved from v1.self.gamma) + # |--- bias : Tensor (moved from v1.self.beta) + # |--- running_mean : Tensor (moved from v1.self.running_mean) + # |--- running_var : Tensor (moved from v1.self.running_var) + # |--- num_batches_tracked : Tensor (moved from v1.self.num_batches_tracked) + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): + version = local_metadata.get('version', None) + if version is None or version == 1: + # BN related parameters and buffers were moved into the BN module for v2 + v2_to_v1_names = { + 'bn.weight': 'gamma', + 'bn.bias': 'beta', + 'bn.running_mean': 'running_mean', + 'bn.running_var': 'running_var', + 'bn.num_batches_tracked': 'num_batches_tracked', + } + for v2_name, v1_name in v2_to_v1_names.items(): + if prefix + v1_name in state_dict: + state_dict[prefix + v2_name] = state_dict[prefix + v1_name] + state_dict.pop(prefix + v1_name) + elif prefix + v2_name in state_dict: + # there was a brief period where forward compatibility + # for this module was broken (between + # https://github.com/pytorch/pytorch/pull/38478 + # and https://github.com/pytorch/pytorch/pull/38820) + # and modules emitted the v2 state_dict format while + # specifying that version == 1. This patches the forward + # compatibility issue by allowing the v2 style entries to + # be used. + pass + elif strict: + missing_keys.append(prefix + v2_name) + + super(_ConvBnNd, self)._load_from_state_dict( + state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + + @classmethod + def from_float(cls, mod): + r"""Create a qat module from a float module or qparams_dict + + Args: `mod` a float module, either produced by torch.ao.quantization utilities + or directly from user + """ + # The ignore is because _FLOAT_MODULE is a TypeVar here where the bound + # has no __name__ (code is fine though) + assert type(mod) == cls._FLOAT_MODULE, 'qat.' + cls.__name__ + '.from_float only works for ' + \ + cls._FLOAT_MODULE.__name__ # type: ignore[attr-defined] + assert hasattr(mod, 'qconfig'), 'Input float module must have qconfig defined' + assert mod.qconfig, 'Input float module must have a valid qconfig' + qconfig = mod.qconfig + conv, bn = mod[0], mod[1] + qat_convbn = cls(conv.in_channels, conv.out_channels, conv.kernel_size, + conv.stride, conv.padding, conv.dilation, + conv.groups, conv.bias is not None, + conv.padding_mode, + bn.eps, bn.momentum, + False, + qconfig) + qat_convbn.weight = conv.weight + qat_convbn.bias = conv.bias + qat_convbn.bn.weight = bn.weight + qat_convbn.bn.bias = bn.bias + qat_convbn.bn.running_mean = bn.running_mean + qat_convbn.bn.running_var = bn.running_var + # mypy error: Cannot determine type of 'num_batches_tracked' + qat_convbn.bn.num_batches_tracked = bn.num_batches_tracked # type: ignore[has-type] + return qat_convbn + + def to_float(self): + modules = [] + cls = type(self) + conv = cls._FLOAT_CONV_MODULE( # type: ignore[attr-defined] + self.in_channels, + self.out_channels, + self.kernel_size, + self.stride, + self.padding, + self.dilation, + self.groups, + self.bias is not None, + self.padding_mode) + conv.weight = torch.nn.Parameter(self.weight.detach()) + if self.bias is not None: + conv.bias = torch.nn.Parameter(self.bias.detach()) + modules.append(conv) + + if cls._FLOAT_BN_MODULE: # type: ignore[attr-defined] + bn = cls._FLOAT_BN_MODULE( # type: ignore[attr-defined] + self.bn.num_features, + self.bn.eps, + self.bn.momentum, + self.bn.affine, + self.bn.track_running_stats) + bn.weight = Parameter(self.bn.weight.detach()) + if self.bn.affine: + bn.bias = Parameter(self.bn.bias.detach()) + modules.append(bn) + + if cls._FLOAT_RELU_MODULE: # type: ignore[attr-defined] + relu = cls._FLOAT_RELU_MODULE() # type: ignore[attr-defined] + modules.append(relu) + + result = cls._FLOAT_MODULE(*modules) # type: ignore[operator] + result.train(self.training) + return result + + + +class ConvBn2d_sophgo(_ConvBnNd, nn.Conv2d): + r""" + A ConvBn2d module is a module fused from Conv2d and BatchNorm2d, + attached with FakeQuantize modules for weight, + used in quantization aware training. + + We combined the interface of :class:`torch.nn.Conv2d` and + :class:`torch.nn.BatchNorm2d`. + + Similar to :class:`torch.nn.Conv2d`, with FakeQuantize modules initialized + to default. + + Attributes: + freeze_bn: + weight_fake_quant: fake quant module for weight + + """ + _FLOAT_MODULE = nni.ConvBn2d + _FLOAT_CONV_MODULE = nn.Conv2d + _FLOAT_BN_MODULE = nn.BatchNorm2d + _FLOAT_RELU_MODULE = None + + def __init__(self, + # ConvNd args + in_channels, out_channels, kernel_size, stride=1, + padding=0, dilation=1, groups=1, + bias=None, + padding_mode='zeros', + # BatchNorm2d args + # num_features: out_channels + eps=1e-05, momentum=0.1, + # affine: True + # track_running_stats: True + # Args for this module + freeze_bn=False, + qconfig=None): + kernel_size = _pair(kernel_size) + stride = _pair(stride) + padding = _pair(padding) + dilation = _pair(dilation) + _ConvBnNd.__init__(self, in_channels, out_channels, kernel_size, stride, + padding, dilation, False, _pair(0), groups, bias, padding_mode, + eps, momentum, freeze_bn, qconfig, dim=2) + +class ConvBnReLU2d_sophgo(ConvBn2d_sophgo): + r""" + A ConvBnReLU2d module is a module fused from Conv2d, BatchNorm2d and ReLU, + attached with FakeQuantize modules for weight, + used in quantization aware training. + + We combined the interface of :class:`torch.nn.Conv2d` and + :class:`torch.nn.BatchNorm2d` and :class:`torch.nn.ReLU`. + + Similar to `torch.nn.Conv2d`, with FakeQuantize modules initialized to + default. + + Attributes: + weight_fake_quant: fake quant module for weight + + """ + # base class defines _FLOAT_MODULE as "ConvBn2d" + _FLOAT_MODULE = nni.ConvBnReLU2d # type: ignore[assignment] + _FLOAT_CONV_MODULE = nn.Conv2d + _FLOAT_BN_MODULE = nn.BatchNorm2d + _FLOAT_RELU_MODULE = nn.ReLU # type: ignore[assignment] + + def __init__(self, + # Conv2d args + in_channels, out_channels, kernel_size, stride=1, + padding=0, dilation=1, groups=1, + bias=None, + padding_mode='zeros', + # BatchNorm2d args + # num_features: out_channels + eps=1e-05, momentum=0.1, + # affine: True + # track_running_stats: True + # Args for this module + freeze_bn=False, + qconfig=None): + super(ConvBnReLU2d_sophgo, self).__init__(in_channels, out_channels, kernel_size, stride, + padding, dilation, groups, bias, + padding_mode, eps, momentum, + freeze_bn, + qconfig) + + def forward(self, input): + return F.relu(ConvBn2d_sophgo._forward(self, input)) + + @classmethod + def from_float(cls, mod): + return super(ConvBnReLU2d_sophgo, cls).from_float(mod) + +class ConvReLU2d_sophgo(qnnqat.Conv2d_sophgo, _FusedModule): + r"""A ConvReLU2d module is a fused module of Conv2d and ReLU, attached with + FakeQuantize modules for weight for + quantization aware training. + + We combined the interface of :class:`~torch.nn.Conv2d` and + :class:`~torch.nn.BatchNorm2d`. + + Attributes: + weight_fake_quant: fake quant module for weight + + """ + _FLOAT_MODULE = nni.ConvReLU2d + _FLOAT_CONV_MODULE = nn.Conv2d + _FLOAT_BN_MODULE = None + _FLOAT_RELU_MODULE = nn.ReLU + + def __init__(self, in_channels, out_channels, kernel_size, stride=1, + padding=0, dilation=1, groups=1, + bias=True, padding_mode='zeros', + qconfig=None): + super(ConvReLU2d_sophgo, self).__init__(in_channels, out_channels, kernel_size, + stride=stride, padding=padding, dilation=dilation, + groups=groups, bias=bias, padding_mode=padding_mode, + qconfig=qconfig) + + def forward(self, input): + return F.relu(qnnqat.Conv2d_sophgo.forward(self, input)) + + @classmethod + def from_float(cls, mod): + assert type(mod) == cls._FLOAT_MODULE, 'qat.' + cls.__name__ + '.from_float only works for ' + cls._FLOAT_MODULE.__name__ + assert hasattr(mod, 'qconfig'), 'Input float module must have qconfig defined' + assert mod.qconfig, 'Input float module must have a valid qconfig' + if type(mod) == cls._FLOAT_MODULE: + mod = mod[0] + qconfig = mod.qconfig + qat_conv = cls(mod.in_channels, mod.out_channels, mod.kernel_size, + stride=mod.stride, padding=mod.padding, dilation=mod.dilation, + groups=mod.groups, bias=mod.bias is not None, + padding_mode=mod.padding_mode, qconfig=qconfig) + qat_conv.weight = mod.weight + qat_conv.bias = mod.bias + return qat_conv diff --git a/mqbench/nn/intrinsic/qat/modules/deconv_fused_sophgo_tpu.py b/mqbench/nn/intrinsic/qat/modules/deconv_fused_sophgo_tpu.py new file mode 100644 index 00000000..50d77a59 --- /dev/null +++ b/mqbench/nn/intrinsic/qat/modules/deconv_fused_sophgo_tpu.py @@ -0,0 +1,455 @@ +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch.nn import init +from torch.nn.intrinsic import _FusedModule +from torch.nn.parameter import Parameter +from torch.nn.modules.utils import _pair, _single + +from typing import TypeVar + +import mqbench.nn.intrinsic as qnni +import mqbench.nn.qat as qnnqat +from mqbench.utils.fusion import fuse_deconv_bn_weights + + +_BN_CLASS_MAP = { + 1: nn.BatchNorm1d, + 2: nn.BatchNorm2d, + 3: nn.BatchNorm3d, +} + +MOD = TypeVar('MOD', bound=nn.modules.conv._ConvTransposeNd) + + +class _ConvTransposeBnNd(nn.modules.conv._ConvTransposeNd, _FusedModule): + + _version = 2 + _FLOAT_MODULE = MOD + + def __init__( + self, + # ConvTransposeBnNd args + in_channels, + out_channels, + kernel_size, + stride, + bias, + transposed, + padding, + output_padding, + groups, + dilation, + padding_mode, + # bn args + # BatchNormNd args + # num_features: out_channels + eps=1e-05, + momentum=0.1, + # affine: True + # track_running_stats: True + # Args for this module + freeze_bn=False, + qconfig=None, + dim=2): + kernel_size = _single(kernel_size) + stride = _single(stride) + padding = _single(padding) + dilation = _single(dilation) + output_padding = _single(output_padding) + nn.modules.conv._ConvTransposeNd.__init__(self, in_channels, + out_channels, kernel_size, + stride, padding, dilation, + transposed, output_padding, + groups, False, padding_mode) + assert qconfig, 'qconfig must be provided for a QAT module' + self.qconfig = qconfig + self.freeze_bn = freeze_bn if self.training else True + self.bn = _BN_CLASS_MAP[dim](out_channels, eps, momentum, True, True) + self.weight_fake_quant = self.qconfig.weight() + # ConvTranspose do per-channel quantize on output channel. + if self.weight_fake_quant.ch_axis != -1: + self.weight_fake_quant.ch_axis = 1 + self.weight_fake_quant.activation_post_process.ch_axis = 1 + if bias: + self.bias = Parameter(torch.Tensor(out_channels)) + else: + self.register_parameter('bias', None) + self.reset_bn_parameters() + + if self.training: + if freeze_bn: + self.freeze_bn_stats() + else: + self.update_bn_stats() + else: + self.freeze_bn_stats() + + def reset_running_stats(self): + self.bn.reset_running_stats() + + def reset_bn_parameters(self): + self.bn.reset_running_stats() + init.uniform_(self.bn.weight) + init.zeros_(self.bn.bias) + # note: below is actully for conv, not BN + if self.bias is not None: + fan_in, _ = init._calculate_fan_in_and_fan_out(self.weight) + bound = 1 / math.sqrt(fan_in) + init.uniform_(self.bias, -bound, bound) + + def reset_parameters(self): + super(_ConvTransposeBnNd, self).reset_parameters() + + def update_bn_stats(self): + self.freeze_bn = False + self.bn.training = True + return self + + def freeze_bn_stats(self): + self.freeze_bn = True + self.bn.training = False + return self + + # def _forward(self, input): + # assert self.bn.running_var is not None + # running_std = torch.sqrt(self.bn.running_var + self.bn.eps) + # scale_factor = self.bn.weight / running_std + # weight_shape = [1] * len(self.weight.shape) + # weight_shape[1] = -1 + # bias_shape = [1] * len(self.weight.shape) + # bias_shape[1] = -1 + # scaled_weight = self.weight_fake_quant(self.weight * scale_factor.reshape(weight_shape)) + # # using zero bias here since the bias for original conv + # # will be added later + # if self.bias is not None: + # zero_bias = torch.zeros_like(self.bias) + # else: + # zero_bias = torch.zeros(self.out_channels, device=scaled_weight.device) + # deconv = self._convtransposed_forward(input, scaled_weight, zero_bias) + # deconv_orig = deconv / scale_factor.reshape(bias_shape) + # if self.bias is not None: + # deconv_orig = deconv_orig + self.bias.reshape(bias_shape) + # deconv = self.bn(deconv_orig) + # return deconv + + def bias_fake_quant_proc(self, bias, scale_w, in_scale): + scale = scale_w*in_scale + if torch.nonzero(scale).size()[0] != scale.numel(): + print('error! scale has 0, scale:', scale) + bias_q = bias/scale + bias = (bias_q.round()-bias_q).detach() + bias_q + bias = bias*scale + return bias + + # def _forward(self, input): + # assert self.bn.running_var is not None + # running_std = torch.sqrt(self.bn.running_var + self.bn.eps) + # scale_factor = self.bn.weight / running_std + # weight_shape = [1] * len(self.weight.shape) + # weight_shape[1] = -1 + # bias_shape = [1] * len(self.weight.shape) + # bias_shape[1] = -1 + # scaled_weight = self.weight_fake_quant(self.weight * scale_factor.reshape(weight_shape)) + # if self.weight_fake_quant.fake_quant_enabled[0] == 1: + # _, fused_bias = fuse_deconv_bn_weights(self.weight, self.bias, + # self.bn.running_mean, self.bn.running_var, self.bn.eps, self.bn.weight, self.bn.bias) + # in_scale = self.input_fake_quantizer.scale #从上一个activation_fake_quant节点获取scale + # scale_fused_bias = self.bias_fake_quant_proc(fused_bias, self.weight_fake_quant.scale, in_scale) + # diff_fused_bias = fused_bias - scale_fused_bias + + # if self.bias is not None: + # zero_bias = torch.zeros_like(self.bias) + # else: + # zero_bias = torch.zeros(self.out_channels, device=scaled_weight.device) + # conv = self._convtransposed_forward(input, scaled_weight, zero_bias) + # conv_orig = conv / scale_factor.reshape(bias_shape) + # if self.bias is not None: + # conv_orig = conv_orig + self.bias.reshape(bias_shape) + # conv = self.bn(conv_orig) + # if self.weight_fake_quant.fake_quant_enabled[0] == 1: + # conv -= diff_fused_bias.reshape(bias_shape) #这里从推导看应该是减 + # return conv + + + def _forward(self, input): + assert self.bn.running_var is not None + running_std = torch.sqrt(self.bn.running_var + self.bn.eps) + scale_factor = self.bn.weight / running_std + weight_shape = [1] * len(self.weight.shape) + weight_shape[1] = -1 + bias_shape = [1] * len(self.weight.shape) + bias_shape[1] = -1 + scaled_weight = self.weight_fake_quant(self.weight * scale_factor.reshape(weight_shape)) + # using zero bias here since the bias for original conv + # will be added later + if self.bias is not None: + zero_bias = torch.zeros_like(self.bias) + conv_bias = self.bias + else: + zero_bias = torch.zeros(self.out_channels, device=scaled_weight.device) + conv_bias = torch.zeros_like(zero_bias, device=scaled_weight.device) + if self.bn.affine: + full_bias = (conv_bias - self.bn.running_mean) / running_std * self.bn.weight + self.bn.bias + else: + full_bias = (conv_bias - self.bn.running_mean) / running_std + # quant_bias = self.bias_fake_quant(full_bias) + quant_bias = self.bias_fake_quant_proc(full_bias, self.weight_fake_quant.scale, self.input_fake_quantizer.scale) + conv_with_bias = self._convtransposed_forward(input, scaled_weight, quant_bias) + deconv_orig = (conv_with_bias - full_bias.reshape(bias_shape)) / scale_factor.reshape(bias_shape) + conv_bias.reshape(bias_shape) + deconv = self.bn(deconv_orig) + return deconv + + def _convtransposed_forward(self, x, w, b): + raise NotImplementedError( + 'The sub-class must implement this function to forward in the needed dim-version!' + ) + + def extra_repr(self): + # TODO(jerryzh): extend + return super(_ConvTransposeBnNd, self).extra_repr() + + def forward(self, input): + return self._forward(input) + + def train(self, mode=True): + """ + Batchnorm's training behavior is using the self.training flag. Prevent + changing it if BN is frozen. This makes sure that calling `model.train()` + on a model with a frozen BN will behave properly. + """ + self.training = mode + if not self.freeze_bn: + for module in self.children(): + module.train(mode) + return self + + # ===== Serialization version history ===== + # + # Version 1/None + # self + # |--- weight : Tensor + # |--- bias : Tensor + # |--- gamma : Tensor + # |--- beta : Tensor + # |--- running_mean : Tensor + # |--- running_var : Tensor + # |--- num_batches_tracked : Tensor + # + # Version 2 + # self + # |--- weight : Tensor + # |--- bias : Tensor + # |--- bn : Module + # |--- weight : Tensor (moved from v1.self.gamma) + # |--- bias : Tensor (moved from v1.self.beta) + # |--- running_mean : Tensor (moved from v1.self.running_mean) + # |--- running_var : Tensor (moved from v1.self.running_var) + # |--- num_batches_tracked : Tensor (moved from v1.self.num_batches_tracked) + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, + missing_keys, unexpected_keys, error_msgs): + version = local_metadata.get('version', None) + if version is None or version == 1: + # BN related parameters and buffers were moved into the BN module for v2 + v2_to_v1_names = { + 'bn.weight': 'gamma', + 'bn.bias': 'beta', + 'bn.running_mean': 'running_mean', + 'bn.running_var': 'running_var', + 'bn.num_batches_tracked': 'num_batches_tracked', + } + for v2_name, v1_name in v2_to_v1_names.items(): + if prefix + v1_name in state_dict: + state_dict[prefix + v2_name] = state_dict[prefix + v1_name] + state_dict.pop(prefix + v1_name) + elif prefix + v2_name in state_dict: + # there was a brief period where forward compatibility + # for this module was broken (between + # https://github.com/pytorch/pytorch/pull/38478 + # and https://github.com/pytorch/pytorch/pull/38820) + # and modules emitted the v2 state_dict format while + # specifying that version == 1. This patches the forward + # compatibility issue by allowing the v2 style entries to + # be used. + pass + elif strict: + missing_keys.append(prefix + v2_name) + + super(_ConvTransposeBnNd, + self)._load_from_state_dict(state_dict, prefix, local_metadata, + strict, missing_keys, + unexpected_keys, error_msgs) + + @classmethod + def from_float(cls, mod): + r"""Create a qat module from a float module or qparams_dict + + Args: `mod` a float module, either produced by torch.quantization utilities + or directly from user + """ + assert type(mod) == cls._FLOAT_MODULE, 'qat.' + cls.__name__ + '.from_float only works for ' + \ + cls._FLOAT_MODULE.__name__ + assert hasattr( + mod, 'qconfig'), 'Input float module must have qconfig defined' + assert mod.qconfig, 'Input float module must have a valid qconfig' + qconfig = mod.qconfig + deconv, bn = mod[0], mod[1] + qat_deconvbn = cls(deconv.in_channels, deconv.out_channels, + deconv.kernel_size, deconv.stride, deconv.bias + is not None, deconv.transposed, deconv.padding, + deconv.output_padding, deconv.groups, + deconv.dilation, deconv.padding_mode, bn.eps, + bn.momentum, False, qconfig) + qat_deconvbn.weight = deconv.weight + qat_deconvbn.bias = deconv.bias + qat_deconvbn.bn.weight = bn.weight + qat_deconvbn.bn.bias = bn.bias + qat_deconvbn.bn.running_mean = bn.running_mean + qat_deconvbn.bn.running_var = bn.running_var + # mypy error: Cannot determine type of 'num_batches_tracked' + qat_deconvbn.bn.num_batches_tracked = bn.num_batches_tracked # type: ignore[has-type] + return qat_deconvbn + + +class ConvTransposeBn2d_sophgo(_ConvTransposeBnNd, nn.ConvTranspose2d): + _FLOAT_MODULE = qnni.ConvTransposeBn2d + + def __init__( + self, + # ConvTransposeBnNd args + in_channels, + out_channels, + kernel_size, + stride=1, + bias=None, + transposed=True, + padding=0, + output_padding=0, + groups=1, + dilation=1, + padding_mode='zeros', + # bn args + # BatchNormNd args + # num_features: out_channels + eps=1e-05, + momentum=0.1, + # affine: True + # track_running_stats: True + # Args for this module + freeze_bn=False, + qconfig=None): + kernel_size = _pair(kernel_size) + stride = _pair(stride) + padding = _pair(padding) + dilation = _pair(dilation) + _ConvTransposeBnNd.__init__(self, in_channels, out_channels, + kernel_size, stride, bias, transposed, + padding, output_padding, groups, dilation, + padding_mode, eps, momentum, freeze_bn, + qconfig) + + def _convtransposed_forward(self, x, w, b): + output_padding = self._output_padding(x, None, self.stride, + self.padding, self.kernel_size, + self.dilation) + return F.conv_transpose2d(x, w, b, self.stride, self.padding, + output_padding, self.groups, self.dilation) + + +class ConvTransposeBnReLU2d_sophgo(ConvTransposeBn2d_sophgo): + _FLOAT_MODULE = qnni.ConvTransposeBnReLU2d + + def __init__( + self, + # ConvTransposeBnNd args + in_channels, + out_channels, + kernel_size, + stride=1, + bias=None, + transposed=True, + padding=0, + output_padding=0, + groups=1, + dilation=1, + padding_mode='zeros', + # bn args + # BatchNormNd args + # num_features: out_channels + eps=1e-05, + momentum=0.1, + # affine: True + # track_running_stats: True + # Args for this module + freeze_bn=False, + qconfig=None): + # super(ConvTransposeBnReLU2d, self).__init__(in_channels, out_channels, kernel_size, stride, + # padding, dilation, groups, bias, + # padding_mode, eps, momentum, + # freeze_bn, + # qconfig) + super(ConvTransposeBnReLU2d_sophgo, + self).__init__(in_channels, + out_channels, + kernel_size, + stride=stride, + bias=bias, + transposed=transposed, + padding=padding, + output_padding=output_padding, + groups=groups, + dilation=dilation, + padding_mode=padding_mode, + eps=eps, + momentum=momentum, + freeze_bn=freeze_bn, + qconfig=qconfig) + + def forward(self, input): + return F.relu(ConvTransposeBn2d_sophgo._forward(self, input)) + + @classmethod + def from_float(cls, mod): + return super(ConvTransposeBnReLU2d_sophgo, cls).from_float(mod) + + +class ConvTransposeReLU2d_sophgo(qnnqat.ConvTranspose2d_sophgo): + _FLOAT_MODULE = qnni.ConvTransposeReLU2d + _FLOAT_DECONV_MODULE = nn.ConvTranspose2d + _FLOAT_BN_MODULE = None + _FLOAT_RELU_MODULE = nn.ReLU + + def __init__( + self, + # ConvTransposeBnNd args + in_channels, + out_channels, + kernel_size, + stride=1, + bias=None, + transposed=True, + padding=0, + output_padding=0, + groups=1, + dilation=1, + padding_mode='zeros', + qconfig=None): + + super(ConvTransposeReLU2d_sophgo, + self).__init__(in_channels, + out_channels, + kernel_size, + stride=stride, + bias=bias, + padding=padding, + output_padding=output_padding, + groups=groups, + dilation=dilation, + padding_mode=padding_mode, + qconfig=qconfig) + assert qconfig, 'qconfig must be provided for QAT module' + + def forward(self, input, output_size=None): + return F.relu(qnnqat.ConvTranspose2d_sophgo.forward(input, output_size)) \ No newline at end of file diff --git a/mqbench/nn/intrinsic/qat/modules/linear_fused_sophgo_tpu.py b/mqbench/nn/intrinsic/qat/modules/linear_fused_sophgo_tpu.py new file mode 100644 index 00000000..d82bc1ac --- /dev/null +++ b/mqbench/nn/intrinsic/qat/modules/linear_fused_sophgo_tpu.py @@ -0,0 +1,337 @@ +import math + +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch.nn import init +from torch.nn import Linear +from torch.nn.intrinsic import _FusedModule +from torch.nn.parameter import Parameter + +from mqbench.nn.intrinsic import LinearBn1d +import torch.nn.intrinsic as nni + + +class LinearBn1d_sophgo(Linear, _FusedModule): + _version = 2 + _FLOAT_MODULE = LinearBn1d + + def __init__(self, + # ConvNd args + in_features, out_features, bias, + # BatchNormNd args + # num_features: out_channels + eps=1e-05, momentum=0.1, + # affine: True + # track_running_stats: True + # Args for this module + freeze_bn=False, + qconfig=None): + Linear.__init__(self, in_features, out_features, False) + assert qconfig, 'qconfig must be provided for QAT module' + self.qconfig = qconfig + self.freeze_bn = freeze_bn if self.training else True + self.bn = nn.BatchNorm1d(out_features, eps, momentum, True, True) + self.weight_fake_quant = self.qconfig.weight() + if bias: + self.bias = Parameter(torch.empty(out_features)) + else: + self.register_parameter('bias', None) + self.reset_bn_parameters() + + # this needs to be called after reset_bn_parameters, + # as they modify the same state + if self.training: + if freeze_bn: + self.freeze_bn_stats() + else: + self.update_bn_stats() + else: + self.freeze_bn_stats() + + def reset_running_stats(self): + self.bn.reset_running_stats() + + def reset_bn_parameters(self): + self.bn.reset_running_stats() + init.uniform_(self.bn.weight) + init.zeros_(self.bn.bias) + # note: below is actully for Linear, not BN + if self.bias is not None: + fan_in, _ = init._calculate_fan_in_and_fan_out(self.weight) + bound = 1 / math.sqrt(fan_in) + init.uniform_(self.bias, -bound, bound) + + def reset_parameters(self): + super(LinearBn1d, self).reset_parameters() + + def update_bn_stats(self): + self.freeze_bn = False + self.bn.training = True + return self + + def freeze_bn_stats(self): + self.freeze_bn = True + self.bn.training = False + return self + + def bias_fake_quant(self, bias, scale_w, in_scale): + if bias is not None: + scale = scale_w*in_scale + if torch.nonzero(scale).size()[0] != scale.numel(): + print('Linear error! scale has 0, scale:', scale) + + bias_q = bias/scale + bias = (bias_q.round()-bias_q).detach() + bias_q + bias = bias*scale + return bias + + # def _forward(self, input): + # in_scale = self.input_fake_quantizer.scale #����һ��activation_fake_quant�ڵ��ȡscale + # conv = F.linear(input, self.weight_fake_quant(self.weight), + # self.bias_fake_quant(self.bias, self.weight_fake_quant.scale, in_scale)) + # return conv + + def _forward(self, input): + assert self.bn.running_var is not None + running_std = torch.sqrt(self.bn.running_var + self.bn.eps) + # input.shape = (batch_size, in_features, *) + # scale_factor.shape = (out_feature, ) + # self.weight.shape = (out_feature, in_feature, *) + # self.bias.shape = (out_feature, *) + # output.shape = (batch_size, out_feature, *) + if self.bn.affine: + scale_factor = self.bn.weight / running_std + else: + scale_factor = 1. / running_std + weight_shape = [1] * len(self.weight.shape) + weight_shape[0] = -1 + bias_shape = [1] * len(input.shape) + bias_shape[1] = -1 + scaled_weight = self.weight_fake_quant(self.weight * scale_factor.reshape(weight_shape)) + # using zero bias here since the bias for original Linear + # will be added later + # Linear layer takes permuted input since the format is (batch_size, *, in_features) + if self.bias is not None: + zero_bias = torch.zeros_like(self.bias) + fc_bias = self.bias + else: + zero_bias = torch.zeros(self.out_channels, device=scaled_weight.device) + fc_bias = torch.zeros_like(zero_bias, device=scaled_weight.device) + if self.bn.affine: + full_bias = (fc_bias - self.bn.running_mean) / running_std * self.bn.weight + self.bn.bias + else: + full_bias = (fc_bias - self.bn.running_mean) / running_std + in_scale = self.input_fake_quantizer.scale #����һ��activation_fake_quant�ڵ��ȡscale + fquant_bias = self.bias_fake_quant(full_bias, self.weight_fake_quant.scale, in_scale) + linear_out = F.linear(input, scaled_weight, fquant_bias) + linear_orig = (linear_out - full_bias) / scale_factor.reshape(bias_shape) + fc_bias.reshape(bias_shape) + linear_out = self.bn(linear_orig) + return linear_out + + def forward(self, input): + # return F.linear(input, self.weight_fake_quant(self.weight), self.bias) + return self._forward(input) + + def extra_repr(self): + return super(LinearBn1d, self).extra_repr() + + def forward(self, input): + return self._forward(input) + + def train(self, mode=True): + """ + Batchnorm's training behavior is using the self.training flag. Prevent + changing it if BN is frozen. This makes sure that calling `model.train()` + on a model with a frozen BN will behave properly. + """ + self.training = mode + if not self.freeze_bn: + for module in self.children(): + module.train(mode) + return self + + def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): + version = local_metadata.get('version', None) + if version is None or version == 1: + # BN related parameters and buffers were moved into the BN module for v2 + v2_to_v1_names = { + 'bn.weight': 'gamma', + 'bn.bias': 'beta', + 'bn.running_mean': 'running_mean', + 'bn.running_var': 'running_var', + 'bn.num_batches_tracked': 'num_batches_tracked', + } + for v2_name, v1_name in v2_to_v1_names.items(): + if prefix + v1_name in state_dict: + state_dict[prefix + v2_name] = state_dict[prefix + v1_name] + state_dict.pop(prefix + v1_name) + elif prefix + v2_name in state_dict: + # there was a brief period where forward compatibility + # for this module was broken (between + # https://github.com/pytorch/pytorch/pull/38478 + # and https://github.com/pytorch/pytorch/pull/38820) + # and modules emitted the v2 state_dict format while + # specifying that version == 1. This patches the forward + # compatibility issue by allowing the v2 style entries to + # be used. + pass + elif strict: + missing_keys.append(prefix + v2_name) + + super(LinearBn1d, self)._load_from_state_dict( + state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) + + @classmethod + def from_float(cls, mod): + r"""Create a qat module from a float module or qparams_dict + + Args: `mod` a float module, either produced by torch.quantization utilities + or directly from user + """ + assert type(mod) == cls._FLOAT_MODULE, 'qat.' + cls.__name__ + '.from_float only works for ' + \ + cls._FLOAT_MODULE.__name__ + assert hasattr(mod, 'qconfig'), 'Input float module must have qconfig defined' + assert mod.qconfig, 'Input float module must have a valid qconfig' + qconfig = mod.qconfig + linear, bn = mod[0], mod[1] + qat_linearbn = cls(linear.in_features, linear.out_features, False, + bn.eps, bn.momentum, + False, + qconfig) + qat_linearbn.weight = linear.weight + qat_linearbn.bias = linear.bias + qat_linearbn.bn.weight = bn.weight + qat_linearbn.bn.bias = bn.bias + qat_linearbn.bn.running_mean = bn.running_mean + qat_linearbn.bn.running_var = bn.running_var + # mypy error: Cannot determine type of 'num_batches_tracked' + qat_linearbn.bn.num_batches_tracked = bn.num_batches_tracked # type: ignore[has-type] + return qat_linearbn + +class Linear_sophgo(nn.Linear): + r""" + A linear module attached with FakeQuantize modules for weight, + used for quantization aware training. + + We adopt the same interface as `torch.nn.Linear`, please see + https://pytorch.org/docs/stable/nn.html#torch.nn.Linear + for documentation. + + Similar to `torch.nn.Linear`, with FakeQuantize modules initialized to + default. + + Attributes: + weight: fake quant module for weight + """ + _FLOAT_MODULE = nn.Linear + + def __init__(self, in_features, out_features, bias=True, + qconfig=None, device=None, dtype=None) -> None: + factory_kwargs = {'device': device, 'dtype': dtype} + super().__init__(in_features, out_features, bias, **factory_kwargs) + assert qconfig, 'qconfig must be provided for QAT module' + self.qconfig = qconfig + self.weight_fake_quant = qconfig.weight(factory_kwargs=factory_kwargs) + + def bias_fake_quant(self, bias, scale_w, in_scale): + if bias is not None: + scale = scale_w*in_scale + if torch.nonzero(scale).size()[0] != scale.numel(): + print('Linear error! scale has 0, scale:', scale) + scale[torch.abs(scale) < 1e-10] = 1e-10 + print('new scale:', scale) + + bias_q = bias/scale + bias = (bias_q.round()-bias_q).detach() + bias_q + bias = bias*scale + return bias + + def _forward(self, input): + assert hasattr(self, 'input_fake_quantizer') + in_scale = self.input_fake_quantizer.scale #����һ��activation_fake_quant�ڵ��ȡscale + # conv = F.linear(input, self.weight_fake_quant(self.weight), + # self.bias_fake_quant(self.bias, self.weight_fake_quant.scale, in_scale)) + if self.bias is not None and self.weight_fake_quant.fake_quant_enabled[0] == 1: + conv = F.linear(input, self.weight_fake_quant(self.weight), + self.bias_fake_quant(self.bias, self.weight_fake_quant.scale, in_scale)) + else: + conv = F.linear(input, self.weight_fake_quant(self.weight), self.bias) + + return conv + + def forward(self, input): + # return F.linear(input, self.weight_fake_quant(self.weight), self.bias) + return self._forward(input) + + @classmethod + def from_float(cls, mod): + r"""Create a qat module from a float module or qparams_dict + + Args: `mod` a float module, either produced by torch.quantization utilities + or directly from user + """ + assert type(mod) == cls._FLOAT_MODULE, ' qat.' + cls.__name__ + '.from_float only works for ' + \ + cls._FLOAT_MODULE.__name__ + assert hasattr(mod, 'qconfig'), 'Input float module must have qconfig defined' + assert mod.qconfig, 'Input float module must have a valid qconfig' + if type(mod) == nni.LinearReLU: + mod = mod[0] + + qconfig = mod.qconfig + qat_linear = cls(mod.in_features, mod.out_features, bias=mod.bias is not None, qconfig=qconfig) + qat_linear.weight = mod.weight + qat_linear.bias = mod.bias + return qat_linear + + def to_float(self): + linear = torch.nn.Linear(self.in_features, self.out_features, self.bias is not None) + linear.weight = torch.nn.Parameter(self.weight.detach()) + if self.bias is not None: + linear.bias = torch.nn.Parameter(self.bias.detach()) + linear.train(self.training) + return linear + + +class LinearReLU_sophgo(Linear_sophgo): + r""" + A LinearReLU module fused from Linear and ReLU modules, attached with + FakeQuantize modules for weight, used in + quantization aware training. + + We adopt the same interface as :class:`torch.nn.Linear`. + + Similar to `torch.nn.intrinsic.LinearReLU`, with FakeQuantize modules initialized to + default. + + Attributes: + weight: fake quant module for weight + + Examples:: + + >>> m = nn.qat.LinearReLU(20, 30) + >>> input = torch.randn(128, 20) + >>> output = m(input) + >>> print(output.size()) + torch.Size([128, 30]) + """ + _FLOAT_MODULE = nni.LinearReLU + + def __init__(self, in_features, out_features, bias=True, + qconfig=None): + super(LinearReLU_sophgo, self).__init__(in_features, out_features, bias, qconfig) + + def forward(self, input): + return F.relu(Linear_sophgo._forward(self, input)) + + @classmethod + def from_float(cls, mod): + return super(LinearReLU_sophgo, cls).from_float(mod) + + def to_float(self): + linear = torch.nn.Linear(self.in_features, self.out_features, self.bias is not None) + linear.weight = torch.nn.Parameter(self.weight.detach()) + if self.bias is not None: + linear.bias = torch.nn.Parameter(self.bias.detach()) + relu = torch.nn.ReLU() + return torch.nn.intrinsic.LinearReLU(linear, relu) diff --git a/mqbench/nn/qat/modules/__init__.py b/mqbench/nn/qat/modules/__init__.py index c0109020..73cde0e8 100644 --- a/mqbench/nn/qat/modules/__init__.py +++ b/mqbench/nn/qat/modules/__init__.py @@ -1,4 +1,4 @@ from .linear import Linear -from .deconv import ConvTranspose2d -from .conv import Conv2d +from .deconv import ConvTranspose2d, ConvTranspose2d_sophgo +from .conv import Conv2d, Conv2d_sophgo from .embedding import Embedding \ No newline at end of file diff --git a/mqbench/nn/qat/modules/conv.py b/mqbench/nn/qat/modules/conv.py index 08a7e0e4..6fed17eb 100644 --- a/mqbench/nn/qat/modules/conv.py +++ b/mqbench/nn/qat/modules/conv.py @@ -1,4 +1,5 @@ import torch.nn.qat.modules as nnqat +import torch.nn as nn from mqbench.quantization.default_bias_fake_quant import bias_fake_quantizer @@ -9,3 +10,22 @@ def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, def forward(self, input): return self._conv_forward(input, self.weight_fake_quant(self.weight), self.bias_fake_quant(self.bias)) + +class Conv2d_sophgo(nnqat.Conv2d): + def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros', qconfig=None, device=None, dtype=None): + super().__init__(in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, padding_mode=padding_mode, qconfig=qconfig) + + def bias_fake_quant_proc(self, bias, scale_w, in_scale): + scale = scale_w*in_scale + bias_q = bias/scale + bias = (bias_q.round()-bias_q).detach() + bias_q + bias = bias*scale + return bias + + def forward(self, input): + #bias伪量化 + bias = self.bias + if self.bias is not None and self.weight_fake_quant.fake_quant_enabled[0] == 1: + in_scale = self.input_fake_quantizer.scale #从上一个activation_fake_quant节点获取scale + bias = self.bias_fake_quant_proc(self.bias, self.weight_fake_quant.scale, in_scale) + return self._conv_forward(input, self.weight_fake_quant(self.weight), bias) \ No newline at end of file diff --git a/mqbench/nn/qat/modules/deconv.py b/mqbench/nn/qat/modules/deconv.py index ec7ef825..e5ae9193 100644 --- a/mqbench/nn/qat/modules/deconv.py +++ b/mqbench/nn/qat/modules/deconv.py @@ -30,6 +30,60 @@ def forward(self, x, output_size=None): x, self.weight_fake_quant(self.weight), self.bias, self.stride, self.padding, output_padding, self.groups, self.dilation) + @classmethod + def from_float(cls, mod): + assert type(mod) == cls._FLOAT_MODULE, 'qat.' + cls.__name__ + '.from_float only works for ' + \ + cls._FLOAT_MODULE.__name__ + assert mod.qconfig, 'Input float module must have a valid qconfig' + if type(mod) == ConvTransposeReLU2d: + mod = mod[0] + qconfig = mod.qconfig + qat_deconv = cls(mod.in_channels, mod.out_channels, mod.kernel_size, + stride=mod.stride, padding=mod.padding, output_padding=mod.output_padding, + groups=mod.groups, bias=mod.bias is not None, dilation=mod.dilation, + padding_mode=mod.padding_mode, qconfig=qconfig) + qat_deconv.weight = mod.weight + qat_deconv.bias = mod.bias + return qat_deconv + +class ConvTranspose2d_sophgo(nn.ConvTranspose2d): + _FLOAT_MODULE = nn.ConvTranspose2d + + def __init__(self, in_channels, out_channels, kernel_size, + stride=1, padding=0, output_padding=0, + groups=1, bias=True, dilation=1, + padding_mode='zeros', qconfig=None): + super().__init__(in_channels, out_channels, kernel_size, + stride=stride, padding=padding, output_padding=output_padding, + groups=groups, bias=bias, dilation=dilation, padding_mode=padding_mode) + assert qconfig, 'qconfig must be provided for QAT module' + self.qconfig = qconfig + self.weight_fake_quant = qconfig.weight() + # ConvTranspose do per-channel quantize on output channel. + if self.weight_fake_quant.ch_axis != -1: + self.weight_fake_quant.ch_axis = 1 + self.weight_fake_quant.activation_post_process.ch_axis = 1 + + def bias_fake_quant_proc(self, bias, scale_w, in_scale): + scale = scale_w*in_scale + bias_q = bias/scale + bias = (bias_q.round()-bias_q).detach() + bias_q + bias = bias*scale + return bias + + def forward(self, x, output_size=None): + output_padding = self._output_padding( + x, output_size, self.stride, self.padding, self.kernel_size, self.dilation + ) + + bias = self.bias + if self.bias is not None and self.weight_fake_quant.fake_quant_enabled[0] == 1: + in_scale = self.input_fake_quantizer.scale #从上一个activation_fake_quant节点获取scale + bias = self.bias_fake_quant_proc(self.bias, self.weight_fake_quant.scale, in_scale) + return F.conv_transpose2d( + x, self.weight_fake_quant(self.weight), bias, self.stride, self.padding, + output_padding, self.groups, self.dilation) + @classmethod def from_float(cls, mod): assert type(mod) == cls._FLOAT_MODULE, 'qat.' + cls.__name__ + '.from_float only works for ' + \ diff --git a/mqbench/observer.py b/mqbench/observer.py index c166ff85..c53113d0 100644 --- a/mqbench/observer.py +++ b/mqbench/observer.py @@ -636,3 +636,21 @@ def forward(self, x_orig): self.min_val = self.min_val * self.ema_ratio + min_val_cur * (1.0 - self.ema_ratio) self.max_val = self.max_val * self.ema_ratio + max_val_cur * (1.0 - self.ema_ratio) return x + + +class KLDObserver(ObserverBase): + ''' + KLD observer, + With the help of HistogramObserver in torch + ''' + def __init__(self, dtype=torch.quint8, qscheme=torch.per_tensor_affine, reduce_range=False, + quant_min=None, quant_max=None, ch_axis=-1, pot_scale=False, factory_kwargs=None): + super(KLDObserver, self).__init__(dtype, qscheme, reduce_range, quant_min, quant_max, + ch_axis, pot_scale, factory_kwargs) + self.histobserver = torch.ao.quantization.observer.HistogramObserver(dtype=dtype, qscheme=qscheme, reduce_range=reduce_range, quant_min=quant_min, quant_max=quant_max) + + def forward(self, x_orig): + return self.histobserver.forward(x_orig) + + def calculate_qparams(self): + return self.histobserver.calculate_qparams() \ No newline at end of file diff --git a/mqbench/prepare_by_platform.py b/mqbench/prepare_by_platform.py index fd850446..92c827a8 100644 --- a/mqbench/prepare_by_platform.py +++ b/mqbench/prepare_by_platform.py @@ -9,7 +9,8 @@ from torch.fx.graph_module import GraphModule from torch.quantization.quantize_fx import _swap_ff_with_fxff from torch.quantization import QConfig - +import torch.nn.intrinsic as nni +import mqbench.nn.intrinsic as qnni from mqbench.fake_quantize import ( LearnableFakeQuantize, @@ -21,6 +22,8 @@ TqtFakeQuantize, AdaRoundFakeQuantize, QDropFakeQuantize, + E4M3FakeQuantize, + E5M2FakeQuantize ) from mqbench.observer import ( ClipStdObserver, @@ -32,11 +35,14 @@ EMAQuantileObserver, MSEObserver, EMAMSEObserver, + KLDObserver, ) +import mqbench from mqbench.fuser_method_mappings import fuse_custom_config_dict from mqbench.utils.logger import logger from mqbench.utils.registry import DEFAULT_MODEL_QUANTIZER from mqbench.scheme import QuantizeScheme +import mqbench.nn.intrinsic.qat as qnniqat __all__ = ['prepare_by_platform'] @@ -53,10 +59,17 @@ class BackendType(Enum): Tengine_u8 = "Tengine_u8" Tensorrt_NLP = "Tensorrt_NLP" Academic_NLP = "Academic_NLP" + Sophgo_TPU = "Sophgo_TPU" ParamsTable = { - BackendType.Academic: dict(qtype='affine'), # noqa: E241 + BackendType.Academic: dict(qtype='affine', + w_qscheme=QuantizeScheme(symmetry=True, per_channel=True, pot_scale=False, bit=8), + a_qscheme=QuantizeScheme(symmetry=True, per_channel=False, pot_scale=False, bit=8), + default_weight_quantize=E4M3FakeQuantize, + default_act_quantize=LearnableFakeQuantize, + default_weight_observer=MinMaxObserver, + default_act_observer=EMAMinMaxObserver), # noqa: E241 BackendType.NNIE: dict(qtype='nnie', # noqa: E241 # NNIE actually do not need w/a qscheme. We add for initialize observer only. w_qscheme=QuantizeScheme(symmetry=True, per_channel=False, pot_scale=False, bit=8), @@ -121,6 +134,13 @@ class BackendType(Enum): default_act_quantize=LearnableFakeQuantize, default_weight_observer=MinMaxObserver, default_act_observer=EMAMinMaxObserver), + BackendType.Sophgo_TPU: dict(qtype='affine', # noqa: E241 + w_qscheme=QuantizeScheme(symmetry=True, per_channel=True, pot_scale=False, bit=8), + a_qscheme=QuantizeScheme(symmetry=True, per_channel=False, pot_scale=False, bit=8), + default_weight_quantize=LearnableFakeQuantize, + default_act_quantize=LearnableFakeQuantize, + default_weight_observer=MinMaxObserver, + default_act_observer=EMAMinMaxObserver) } ParamsTable[BackendType.Tensorrt_NLP] = ParamsTable[BackendType.Tensorrt] ParamsTable[BackendType.Academic_NLP] = ParamsTable[BackendType.Academic] @@ -135,6 +155,7 @@ class BackendType(Enum): 'LSQObserver': LSQObserver, # Usually used for LSQ. # noqa: E241 'MSEObserver': MSEObserver, # noqa: E241 'EMAMSEObserver': EMAMSEObserver, # noqa: E241 + 'KLDObserver': KLDObserver, } FakeQuantizeDict = { @@ -147,10 +168,12 @@ class BackendType(Enum): 'TqtFakeQuantize': TqtFakeQuantize, # TQT # noqa: E241 'AdaRoundFakeQuantize': AdaRoundFakeQuantize, # AdaRound # noqa: E241 'QDropFakeQuantize': QDropFakeQuantize, # BRECQ & QDrop # noqa: E241 + 'E4M3FakeQuantize': E4M3FakeQuantize, + 'E5M2FakeQuantize': E5M2FakeQuantize } -def get_qconfig_by_platform(deploy_backend: BackendType, extra_qparams: Dict): +def get_qconfig_by_platform(deploy_backend: BackendType, extra_qparams: Dict, work_mode: str): """ Args: @@ -174,6 +197,12 @@ def get_qconfig_by_platform(deploy_backend: BackendType, extra_qparams: Dict): 'a_qscheme': { same with w_qscheme. } + "object_type": [ + (torch.add, qconfig) + ], + "module_name": [ + ("conv1", qconfig) + ] } """ w_observer = extra_qparams.get('w_observer', None) @@ -210,7 +239,14 @@ def get_qconfig_by_platform(deploy_backend: BackendType, extra_qparams: Dict): **w_qscheme.to_observer_params()) a_config = backend_params['default_act_quantize'].with_args(observer=a_observer, **a_qscheme.to_observer_params()) - return QConfig(activation=a_config, weight=w_config) + qconfig = {'': QConfig(activation=a_config, weight=w_config)} + object_type = extra_qparams.get('object_type', None) + if object_type is not None: + qconfig["object_type"] = object_type + module_name = extra_qparams.get('module_name', None) + if module_name is not None: + qconfig["module_name"] = module_name + return qconfig # Academic setting should specific quant scheme in config. if deploy_backend in [BackendType.Academic, BackendType.Academic_NLP]: @@ -260,9 +296,128 @@ def get_qconfig_by_platform(deploy_backend: BackendType, extra_qparams: Dict): a_observer.__name__, str(a_qscheme))) if backend_params['qtype'] == 'vitis': logger.info('Bias Qconfig:\n TqtFakeQuantize with MinMaxObserver') + + if deploy_backend in [BackendType.Academic, BackendType.Academic_NLP]: + return {'':QConfig(activation=a_qconfig, weight=w_qconfig)} + + qconfig = {'': QConfig(activation=a_qconfig, weight=w_qconfig)} + if deploy_backend == BackendType.Sophgo_TPU: + qconfig["object_type"] = {torch.nn.Linear:createQConfigForSophgoLiner()} #int8 qat, Sophgo_TPU use sym per-layer + if work_mode == 'all_int4_qat': + qconfig["object_type"][torch.nn.Linear] = createQConfigForSophgoLiner(bit_num = 4) + if work_mode in ['int4_and_int8_mix', 'int4_and_int8_mix_no_fc']: + w_qscheme = { + 'bit': 4, + 'symmetry': True, + 'per_channel': True, + 'pot_scale': False + } + a_qscheme = { + 'bit': 4, + 'symmetry': True, + 'per_channel': False, + 'pot_scale': False + } + int4_qconfig = createQConfig(w_qscheme = w_qscheme, a_qscheme = a_qscheme) + qconfig["object_type"][torch.nn.Conv2d] = int4_qconfig + from mqbench.custom_quantizer.sophgo_tpu_quantizer import SophgoTpuQuantizer + import torch.nn as nn + additional_qat_module_mapping = [ + nni.ConvBn2d, + nni.ConvBnReLU2d, + nn.Conv2d, + nni.ConvReLU2d, + qnni.ConvTransposeBnReLU2d, + qnni.ConvTransposeReLU2d, + qnni.ConvTransposeBn2d + ] + for i in additional_qat_module_mapping: + qconfig["object_type"][i] = int4_qconfig + if work_mode == 'int4_and_int8_mix_no_fc': + for i in SophgoTpuQuantizer({}, {})._layers_need_check_is_dw: + qconfig["object_type"][i] = int4_qconfig + else: + for i in SophgoTpuQuantizer({}, {})._layers_need_scale_form_input_fake_quantizer: + qconfig["object_type"][i] = int4_qconfig + if work_mode != 'int4_and_int8_mix_no_fc': + liner_qconfig = createQConfigForInt4SophgoLiner() + qconfig["object_type"][nni.LinearReLU] = liner_qconfig + qconfig["object_type"][qnni.LinearBn1d] = liner_qconfig + qconfig["object_type"][torch.nn.Linear] = liner_qconfig + object_type = extra_qparams.get('object_type', None) + if object_type is not None: + if "object_type" in qconfig: + qconfig["object_type"].update(object_type) + else: + qconfig["object_type"] = object_type + + module_name = extra_qparams.get('module_name', None) + if module_name is not None: + qconfig["module_name"] = module_name + return qconfig + +#LearnableFakeQuantize, LearnableFakeQuantize, MinMaxObserver, EMAMinMaxObserver +# 'w_qscheme': { +# 'bit': bitwidth, +# 'symmetry': whether quantize scheme is symmetric, +# 'per_channel': whether quantize scheme is perchannel, +# 'pot_scale': whether scale is power of two. +# } +# 'a_qscheme': { +# 'bit': bitwidth, +# 'symmetry': whether quantize scheme is symmetric, +# 'per_channel': whether quantize scheme is perchannel, +# 'pot_scale': whether scale is power of two. +# } + +def createQConfigForSophgoLiner(bit_num = 8, w_fakequantize = 'LearnableFakeQuantize', w_observer = 'MinMaxObserver', w_fakeq_params = {}, w_observer_extra_args = {}): + w_observer = ObserverDict[w_observer] + w_fakequantize = FakeQuantizeDict[w_fakequantize] + w_qscheme = QuantizeScheme(symmetry=True, per_channel=False, pot_scale=False, bit=bit_num) #Sophgo_TPU use sym per-layer + w_qscheme.kwargs.update(w_observer_extra_args) + w_qconfig = w_fakequantize.with_args(observer=w_observer, **w_fakeq_params, **w_qscheme.to_observer_params()) + return QConfig(activation=torch.nn.Identity, weight=w_qconfig) #activation use global quant conifg + +def createQConfig(w_fakequantize = 'LearnableFakeQuantize', a_fakequantize = 'LearnableFakeQuantize', + w_observer = 'MinMaxObserver', a_observer = 'EMAMinMaxObserver', w_qscheme = {}, a_qscheme = {}, + w_fakeq_params = {}, a_fakeq_params = {}, w_observer_extra_args = {}, a_observer_extra_args = {}): + w_observer = ObserverDict[w_observer] + w_fakequantize = FakeQuantizeDict[w_fakequantize] + if w_qscheme is not None: + w_qscheme = QuantizeScheme(**w_qscheme) + else: + w_qscheme = QuantizeScheme(symmetry=True, per_channel=True, pot_scale=False, bit=8) + + w_qscheme.kwargs.update(w_observer_extra_args) + w_qconfig = w_fakequantize.with_args(observer=w_observer, **w_fakeq_params, **w_qscheme.to_observer_params()) + + a_observer = ObserverDict[a_observer] + a_fakequantize = FakeQuantizeDict[a_fakequantize] + if a_qscheme is not None: + a_qscheme = QuantizeScheme(**a_qscheme) + else: + a_qscheme = QuantizeScheme(symmetry=True, per_channel=False, pot_scale=False, bit=8) + a_qscheme.kwargs.update(a_observer_extra_args) + a_qconfig = a_fakequantize.with_args(observer=a_observer, **a_fakeq_params, **a_qscheme.to_observer_params()) return QConfig(activation=a_qconfig, weight=w_qconfig) +def createQConfigForInt4SophgoLiner(w_fakequantize = 'LearnableFakeQuantize', a_fakequantize = 'LearnableFakeQuantize', + w_observer = 'MinMaxObserver', a_observer = 'EMAMinMaxObserver', w_qscheme = {}, a_qscheme = {}, + w_fakeq_params = {}, a_fakeq_params = {}, w_observer_extra_args = {}, a_observer_extra_args = {}): + w_observer = ObserverDict[w_observer] + w_fakequantize = FakeQuantizeDict[w_fakequantize] + w_qscheme = QuantizeScheme(symmetry=True, per_channel=False, pot_scale=False, bit=4) + + w_qscheme.kwargs.update(w_observer_extra_args) + w_qconfig = w_fakequantize.with_args(observer=w_observer, **w_fakeq_params, **w_qscheme.to_observer_params()) + + a_observer = ObserverDict[a_observer] + a_fakequantize = FakeQuantizeDict[a_fakequantize] + a_qscheme = QuantizeScheme(symmetry=True, per_channel=False, pot_scale=False, bit=4) + a_qscheme.kwargs.update(a_observer_extra_args) + a_qconfig = a_fakequantize.with_args(observer=a_observer, **a_fakeq_params, **a_qscheme.to_observer_params()) + return QConfig(activation=a_qconfig, weight=w_qconfig) class CustomedTracer(Tracer): """ @@ -333,6 +488,7 @@ def _get_attrs(target, attrs): def prepare_by_platform( model: torch.nn.Module, deploy_backend: BackendType, + input_shape_dict: list = None, prepare_custom_config_dict: Dict[str, Any] = {}, custom_tracer: Tracer = None): """ @@ -358,7 +514,8 @@ def prepare_by_platform( # Get Qconfig extra_qconfig_dict = prepare_custom_config_dict.get('extra_qconfig_dict', {}) - qconfig = get_qconfig_by_platform(deploy_backend, extra_qconfig_dict) + work_mode = prepare_custom_config_dict.get('work_mode', '') + qconfig = get_qconfig_by_platform(deploy_backend, extra_qconfig_dict, work_mode) _swap_ff_with_fxff(model) # Preserve attr. @@ -379,13 +536,23 @@ def prepare_by_platform( if custom_tracer is not None: tracer = custom_tracer graph = tracer.trace(model, concrete_args) + print('>>>>>trace graph:',graph) name = model.__class__.__name__ if isinstance(model, torch.nn.Module) else model.__name__ modules = dict(model.named_modules()) + print('>>>>>named_modules:',modules['']) graph, duplicated_modules = duplicate_reused_nodes(graph, modules) constant_nodes = prepare_constant_dict(graph, model) modules.update(duplicated_modules) modules.update(constant_nodes) graph_module = GraphModule(modules, graph, name) + if input_shape_dict is not None: + try: + from torch.fx.passes import shape_prop + dev = next(model.parameters()).device + dummy_input = [torch.rand(shape).to(dev) for shape in input_shape_dict] + shape_prop.ShapeProp(graph_module).propagate(*dummy_input) + except: + print('waring, shape_prop fail') # Model fusion. extra_fuse_dict = prepare_custom_config_dict.get('extra_fuse_dict', {}) extra_fuse_dict.update(fuse_custom_config_dict) diff --git a/mqbench/ptq_train_all_model.py b/mqbench/ptq_train_all_model.py new file mode 100644 index 00000000..fa47f829 --- /dev/null +++ b/mqbench/ptq_train_all_model.py @@ -0,0 +1,52 @@ +import sys +import os +import time +sys.path.append(os.path.abspath('.')) +print(sys.path) + +import os +import time +import argparse +from multiprocessing import Pool + +model_list_all=[ + # "--arch=mobilenet_v2 --batch-size=64 --cali-batch-num=16", + # "--arch=resnet50 --batch-size=64 --cali-batch-num=16", + # "--arch=vgg11_bn --batch-size=64 --cali-batch-num=16", + "--arch=resnet18 --batch-size=64 --cali-batch-num=16", + # "--arch=shufflenet_v2_x0_5 --batch-size=64 --cali-batch-num=16", + # "--arch=squeezenet1_1 --batch-size=64 --cali-batch-num=16", + # "--arch=mobilenet_v3_small --batch-size=64 --cali-batch-num=16" +] + +# output_path='/path-of-your-dir/' +output_path='./ptq_test_before_push' + +cmd_str = f"--data_path=/sea/data/imagenet/for_train_val --backend=sophgo_tpu --seed=1005 --pretrained --quantize_type=naive_ptq --deploy\ + --output_path={output_path}" + +def worker(cmd_line): + print('cmd_line:', cmd_line) + os.system(cmd_line) + +if __name__ == "__main__": + + time_start = time.time() + wp = os.getcwd() + po = Pool(1) + for i,model in enumerate(model_list_all): + arch = model.split(' ')[0].split('=')[1].strip() + os.system(f'mkdir -p {output_path}/{arch}') + + cmd_line = f'cd {wp};CUDA_VISIBLE_DEVICES=0 python3 application/imagenet_example/PTQ/ptq/ptq_main.py {model} {cmd_str} > {output_path}/{arch}/{i}_log_train_{arch} 2>&1' + # cmd_line = f'cd {wp};CUDA_VISIBLE_DEVICES=0 python3 application/imagenet_example/PTQ/ptq/ptq_main.py {model} {cmd_str}' + + print('start', model) + po.apply_async(worker, (cmd_line,)) + + po.close() + po.join() + print('all end') + + time_end = time.time() + print('totally time is ', time_end-time_start) \ No newline at end of file diff --git a/mqbench/qat_train_all_model.py b/mqbench/qat_train_all_model.py new file mode 100644 index 00000000..4c5535d4 --- /dev/null +++ b/mqbench/qat_train_all_model.py @@ -0,0 +1,69 @@ +import os +import time +import argparse +from multiprocessing import Pool + +parser_auto_cali = argparse.ArgumentParser(description='uto_cali params.', conflict_handler='resolve') +parser = argparse.ArgumentParser(description='auto_cali_test params.') +parser.add_argument('--debug_cmd', type=str, default='onnx,sym', help='exclude') +opt = parser.parse_args() + +# model_list_all=[ +# # "--arch=shufflenet_v2_x0_5 --batch-size=320 --lr=1e-2", +# # "--arch=mobilenet_v2 --batch-size=128 --lr=1e-3", +# # "--arch=resnet18 --batch-size=256 --lr=1e-2", +# # "--arch=vgg11_bn --batch-size=32 --lr=1e-3", +# # "--arch=resnet50 --batch-size=32 --lr=1e-2", +# # "--arch=squeezenet1_1 --batch-size=128 --lr=1e-3", +# # "--arch=mobilenet_v3_small --batch-size=128 --lr=1e-2" +# ] + +model_list_all=[ + # "--arch=mobilenet_v2 --batch-size=64 --lr=1e-4", + # "--arch=resnet50 --batch-size=32 --lr=1e-4", + # "--arch=vgg11_bn --batch-size=32 --lr=1e-4", + "--arch=resnet18 --batch-size=128 --lr=1e-4", + # "--arch=shufflenet_v2_x0_5 --batch-size=320 --lr=1e-4", + # "--arch=squeezenet1_1 --batch-size=128 --lr=1e-4", + # "--arch=mobilenet_v3_small --batch-size=128 --lr=1e-4" +] + + +epochs = 1 +output_path='./qat_test_before_push' + +# fast_test = '' +fast_test = '--fast_test' +# pre_eval_and_export = '--pre_eval_and_export' +pre_eval_and_export = '' + +cmd_str = f"--epochs={epochs} --deploy_batch_size=10 --gpu=0 --pretrained --evaluate --backend=sophgo_tpu --optim=sgd \ + --train_data=/sea/data/imagenet/for_train_val --val_data=/sea/data/imagenet/for_train_val --output_path={output_path} {fast_test} {pre_eval_and_export}" +# cmd_str = f"--epochs={epochs} --deploy_batch_size=10 --cpu --pretrained --evaluate --backend=sophgo_tpu --optim=sgd \ +# --train_data=/data/imagenet --val_data=/data/imagenet --output_path={output_path} {fast_test} {pre_eval_and_export}" + + +def worker(cmd_line): + print('cmd_line:', cmd_line) + os.system(cmd_line) + +if __name__ == "__main__": + + time_start = time.time() + wp = os.getcwd() + po = Pool(1) + for i,model in enumerate(model_list_all): + arch = model.split(' ')[0].split('=')[1].strip() + os.system(f'mkdir -p {output_path}/{arch}') + + cmd_line = f'cd {wp};python3 application/imagenet_example/main.py {model} {cmd_str} > {output_path}/{arch}/{i}_log_train_{arch} 2>&1' + # cmd_line = f'cd {wp};python3 application/imagenet_example/main.py {model} {cmd_str}' + + print('start', model) + po.apply_async(worker, (cmd_line,)) + + po.close() + po.join() + print('all end') + time_end = time.time() + print('totally time is ', time_end-time_start) \ No newline at end of file diff --git a/mqbench/utils/utils.py b/mqbench/utils/utils.py index 51a0859a..8a0aa2ee 100644 --- a/mqbench/utils/utils.py +++ b/mqbench/utils/utils.py @@ -176,4 +176,41 @@ def topology_order(model): node2idx = {} for idx, node in enumerate(model.graph.nodes): node2idx[node] = idx - return node2idx \ No newline at end of file + return node2idx + +def get_flattened_qconfig_dict(qconfig_dict): + """ flatten the global, object_type and module_name qconfig + to the same qconfig_dict so that it can be used by + propagate_qconfig_ function. + "module_name_regex" is ignored for now since it's not supported + in propagate_qconfig_, but it can be fixed later. + + For example: + Input: { + "": qconfig, + "object_type": [ + (torch.add, qconfig) + ], + "module_name": [ + ("conv", qconfig) + ] + } + + Output: { + "": qconfig, + torch.add: qconfig, + "conv": qconfig + } + """ + flattened = dict() + if '' in qconfig_dict: + flattened[''] = qconfig_dict[''] + + def flatten_key(key): + if key in qconfig_dict: + for (obj, qconfig) in qconfig_dict[key].items(): + flattened[obj] = qconfig + + flatten_key('object_type') + flatten_key('module_name') + return flattened \ No newline at end of file diff --git a/run_test.sh b/run_test.sh new file mode 100644 index 00000000..4c7d9160 --- /dev/null +++ b/run_test.sh @@ -0,0 +1,6 @@ +#!/bin/bash +# PTQ test +python ./mqbench/ptq_train_all_model.py + +# QAT test +python ./mqbench/qat_train_all_model.py \ No newline at end of file diff --git a/setup.py b/setup.py index ef5e32a2..14c4dce9 100644 --- a/setup.py +++ b/setup.py @@ -1,6 +1,11 @@ +import os +import torch import setuptools from mqbench import __version__ +from torch.utils.cpp_extension import BuildExtension, CppExtension +with open("READMEFP8.md", "r") as fh: + long_description = fh.read() def read_requirements(): reqs = [] @@ -9,6 +14,23 @@ def read_requirements(): reqs.append(line.strip()) return reqs +cmdclass = {} +ext_modules = [] + +ext_modules.append( + CppExtension('fpemu_cpp', + ['FP8_Emulator/pytquant/cpp/avx-fpemu.cpp'], #如果机子支持avx-512指令集，可以在下面添加编译512指令集的args，然后将avx-fpemu文件更换为fpemu_impl.cpp文件 + extra_compile_args = ["-mf16c", "-mavx2", "-mlzcnt", "-fopenmp", "-Wdeprecated-declarations"] + ),) + +if torch.cuda.is_available(): + from torch.utils.cpp_extension import BuildExtension, CUDAExtension + ext_modules.append( + CUDAExtension('fpemu_cuda', [ + 'FP8_Emulator/pytquant/cuda/fpemu_impl.cpp', + 'FP8_Emulator/pytquant/cuda/fpemu_kernels.cu'], + ),) +cmdclass['build_ext'] = BuildExtension setuptools.setup( name="MQBench", @@ -16,11 +38,16 @@ def read_requirements(): author="The Great Cold", author_email="", description=("Quantization aware training."), + ext_modules=ext_modules, + cmdclass=cmdclass, + long_description=long_description, + long_description_content_type="text/markdown", + url="", python_requires='>=3.6', packages=setuptools.find_packages(), classifiers=( 'Development Status :: 3 - Alpha', "Programming Language :: Python :: 3", - "Operating System :: POSIX :: Linux"), + "Operating System :: POSIX :: Linux :: OS Independent"), install_requires=read_requirements() -) +) \ No newline at end of file