This repository contains the training code of Quantization Networks introduced in the CVPR 2019 paper: Quantization Networks.

Our implementation is a modified version of the original implementation. The main changes are:

1. Support of CIFAR-10 and MNIST dataset, besides ImageNet.
2. A unified training script for full-precision, weight quantization and activation quantization.
3. Training curve visualization using Tensorboard.
4. Clustering before quantization training, needless of manual clustering.
5. Freezing of BN layer during activation quantization training (idea borrowed from FQN).
6. Removal of outlier in clustering using 3-sigma rule (also borrowed from FQN).

First, please create a compatible Python environment.

Here are our environment configurations during development.

• Ubuntu 18.04
• Python 3.7
• opencv-python
• numpy 1.17.4
• pytorch 1.3.0
• torchvision 0.4.2
• Tensorboard 2.0.0
• argparse 1.1
• logging 0.5.1.2

Please refer to PREPARE_DATA.md.

Please refer to GET_STARTED.md.

Please refer to EXPERIMENTS.md.

This repository is forked from aliyun/alibabacloud-quantization-networks and keep its Apache 2.0 license.

Please cite the paper if it helps your research:

@inproceedings{yang2019quantization,
  title={Quantization Networks},
  author={Yang Jiwei, Shen Xu, Xing Jun, Tian Xinmei, Li Houqiang, Deng Bing, Huang Jianqiang and Hua Xian-sheng},
  booktitle={CVPR},
  year={2019}
}