This directory contains implementation of various techniques for quantizing convolutional neural network models into low bit-precision.
Work in progress.
The quantization codes are implemented on top of tf_cnn_benchmarks,
which requires the latest nightly version of TensorFlow.
To run ResNet50 4-bit fixed-point quantized training with synthetic data with a single GPU (from the google_research directory), run
python -m cnn_quantization.tf_cnn_benchmarks.tf_cnn_benchmarks --num_gpus=1 \
--batch_size=32 --model=resnet50_v2 --variable_update=parameter_server \
--quant_weight=true --quant_weight_bits=4 --use_relu_x=true --quant_act=true \
--quant_act_bits=4
To run on CPU, run
python -m cnn_quantization.tf_cnn_benchmarks.tf_cnn_benchmarks \
--data_format=NHWC --batch_size=32 --model=resnet50_v2 \
--variable_update=parameter_server --quant_weight=true --quant_weight_bits=4 \
--use_relu_x=true --quant_act=true --quant_act_bits=4
To see the full list of flags, run
python -m cnn_quantization.tf_cnn_benchmarks.tf_cnn_benchmarks --help