scale-adjusted-training

PyTorch implementation of Towards Efficient Training for Neural Network Quantization

Introduction

This repo implement the Scale-Adjusted Training from Towards Efficient Training for Neural Network Quantization including:

1. Constant rescaling Dorefa-quantize
2. Calibrated gradient PACT

TODO

• constant rescaling DoReFaQuantize layer
• CGPACT layer
• test with mobilenetv1
• test with mobilenetv2
• test with resnet50

Acknowledgement

• https://github.com/marvis/pytorch-mobilenet
• https://github.com/tonylins/pytorch-mobilenet-v2
• https://github.com/ricky40403/PyTransformer/tree/hotfix