ResNeSt速度问题跟踪

[HiHippie]

该issue记录ResNeSt速度问题跟踪进展。

单卡

OneFlow resnest50 time: 118.3ms (= 11832.9ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnest50 time: 79.6ms (= 7961.8ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 0.67 (= 79.6ms / 118.3ms)

OneFlow resnest101 time: 254.0ms (= 25399.3ms / 100, input_shape=[16, 3, 256, 256])
PyTorch resnest101 time: 162.6ms (= 16264.7ms / 100, input_shape=[16, 3, 256, 256])
❌ Relative speed: 0.64 (= 162.6ms / 254.0ms)

OneFlow resnest200 time: 372.0ms (= 37202.1ms / 100, input_shape=[4, 3, 320, 320])
PyTorch resnest200 time: 197.7ms (= 19765.0ms / 100, input_shape=[4, 3, 320, 320])
❌ Relative speed: 0.53 (= 197.7ms / 372.0ms)

OneFlow resnest269 time: 686.4ms (= 68636.2ms / 100, input_shape=[4, 3, 416, 416])
PyTorch resnest269 time: 383.0ms (= 38296.1ms / 100, input_shape=[4, 3, 416, 416])
❌ Relative speed: 0.56 (= 383.0ms / 686.4ms)

多卡

OneFlow resnest50 time: 32.2ms (= 3216.4ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
PyTorch resnest50 time: 25.4ms (= 2536.2ms / 100, input_shape=[16, 3, 224, 224], backward is disabled, ddp, world size=2)
❌ Relative speed: 0.79 (= 25.4ms / 32.2ms)

OneFlow resnest101 time: 66.6ms (= 6661.4ms / 100, input_shape=[16, 3, 256, 256], backward is disabled, ddp, world size=2)
PyTorch resnest101 time: 50.5ms (= 5047.7ms / 100, input_shape=[16, 3, 256, 256], backward is disabled, ddp, world size=2)
❌ Relative speed: 0.76 (= 50.5ms / 66.6ms)

OneFlow resnest200 time: 176.6ms (= 17656.3ms / 100, input_shape=[16, 3, 320, 320], backward is disabled, ddp, world size=2)
PyTorch resnest200 time: 140.3ms (= 14034.6ms / 100, input_shape=[16, 3, 320, 320], backward is disabled, ddp, world size=2)
❌ Relative speed: 0.79 (= 140.3ms / 176.6ms)

OneFlow resnest269 time: 341.0ms (= 34099.7ms / 100, input_shape=[16, 3, 416, 416], backward is disabled, ddp, world size=2)
PyTorch resnest269 time: 295.3ms (= 29532.5ms / 100, input_shape=[16, 3, 416, 416], backward is disabled, ddp, world size=2)
❌ Relative speed: 0.87 (= 295.3ms / 341.0ms)

OneFlow resnest50 time: 132.9ms (= 13292.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnest50 time: 88.1ms (= 8809.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 0.66 (= 88.1ms / 132.9ms)

OneFlow resnest101 time: 293.8ms (= 29382.8ms / 100, input_shape=[16, 3, 256, 256], ddp, world size=2)
PyTorch resnest101 time: 182.2ms (= 18217.1ms / 100, input_shape=[16, 3, 256, 256], ddp, world size=2)
❌ Relative speed: 0.62 (= 182.2ms / 293.8ms)