PyTorch_Quantization

all methods of pytorch quantization based on resnet50 with cifar-10

Method

User should run

python3 quantization.py --tq [BOOL] --sq [BOOL] --qat [BOOL]

Each argument parser means

tq : tutorial qauntization, which imports quantized model where pytorch official page offers

sq : static quantization, manually defines resnet 50 models and quantize

qat : quantization aware training, train with illusive transformer (fp32 -> int8) while training

Future work

• quantization aware training

  need more training epochs for training code

• speed issue in CPU

  currently quantized model is more slower than expected, need to test in mobile devices

• various backbone model

  currently supporting resnet only

• customized dataset loading