PyTorch Implementation of PACT: Parameterized Clipping Activation for Quantized Neural Networks

Paper : PACT I have implemented to reproduce quantization paper PACT on CIFAR10 with ResNet20 I have kept all experiment with first layer and last layer in 8bitwidth Please look at Table 2 in the paper

Bitwidth	Original Paper ERR	Reproduced result ERR
5 bitwidth	8.4%	8.4%
4 bitwidth	9.2%	9.11%
3 bitwidth	9.4%	9.91%
2 bitwidth	11.1%	12.93%

Note

3 bitwidth and 2 bitwidth has slightly lower performance than actual paper.

Details of Pipeline

I have adopted training pipeline and pre-trained model from Akamaster's repo(https://github.com/akamaster/pytorch_resnet_cifar10)