Keras2TFLiteQuantizer

Keras2TFLiteQuantizer is a utility to convert Keras models to the TFLite format, with support for various quantization options. This class helps convert models like MobileNetV2 into lightweight versions using formats like fp32 (floating point 32-bit), int8, float16, and more.

Features

• Converts Keras models to TFLite format
• Includes a utility function load_keras_model to load pre-configured Keras models with preprocessing functions
• Supports multiple quantization options:
  • Dynamic Range Quantization (int8, float16)
  • Full Integer Quantization (int8, int16)
  • No Quantization (FP32)
• Offers customizable settings for file saving and data type conversions, with exception handling for invalid configurations

Directory Structure

TensorFlowQuantization/
├── quantizer/
│   ├── __init__.py
│   ├── Keras2TFLiteQuantizer.py        # Keras2TFLiteQuantizer class definition
│   └── KerasModelLoader.py             # KerasModel and it's preprocess function
├── tests/
│   ├── __init__.py
│   └── test_quantizer.py               # Test file
└── requirements.txt                    # Dependency file
└── README.md                           # Project description file

Installation

1. Install required libraries: Install tensorflow and other dependencies.

pip install -r requirements.txt

2. Organize directory structure: Place the file containing the Keras2TFLiteQuantizer class in src/, and the test files in `tests/.

Usage Example

import tensorflow as tf
from src.your_module_name import Keras2TFLiteQuantizer

# Load a Keras model (e.g., MobileNetV2)
model = tf.keras.applications.MobileNetV2(weights=None, input_shape=(224, 224, 3))
preproc_fn = tf.keras.applications.mobilenet_v2.preprocess_input

# Instantiate the Keras2TFLiteQuantizer
quantizer = Keras2TFLiteQuantizer(
    model=model, 
    model_name="mobilenetv2", 
    preproc_fn=preproc_fn, 
    saved_dir="./tflite_models"
)

# Generate TFLite models with various quantization options
tflite_fp32 = quantizer.tfl_no_quant()
tflite_dynamic_int8 = quantizer.tfl_dynamic_range_quant(weight_type="int8")
tflite_full_int8 = quantizer.tfl_full_integer_quant(signal_type="int8", weight_type="int8")

Key Methods

• tfl_no_quant: Converts a Keras model to TFLite format without quantization
• tfl_dynamic_range_quant(weight_type="int8"): Applies dynamic range quantization to int8 or float16
• tfl_full_integer_quant: Applies full integer quantization
  • Parameters:
    • signal_type: Type of quantization for activations (int8 or int16)
    • weight_type: Type of quantization for weights (int8)
    • if_input_type: Input tensor type (float32, int8, uint8)
    • if_output_type: Output tensor type (float32, int8, uint8)

Exception Handling

• InvalidQuantizationError: Raised for invalid quantization configurations or unsupported data types
• NotImplementedError: Raised when unsupported weight_type or data types are specified

Testing

Run tests using pytest from the project root directory:

pytest

This command executes all tests in the tests/ directory to verify that the class methods work as expected.