QPyTorch implementation for the ICLR 2018 paper, WAGE. This is replicate from the Tensorflow repo by the paper's authors. We modify this implementation based on our previous pytorch implementation.
With QPyTorch the simulation overhead is much smaller. Results are obtained on a GTX 1080 Ti
| Seeting | Training Time per epoch | Simulation Overhead |
|---|---|---|
| No Quantization | 13.60 | 0 |
| QPyTorch | 17.50 | 3.9 |
| PyTorch | 21.16 | 7.56(1.9x) |
- NVIDIA GPU + CUDA + CuDNN
- PyTorch
- QPyTorch
- TensorboardX
- Tabulate
Please follow the official instruction to install PyTorch and NVIDIA related prerequisites. Other things should be handled by
pip install -r requirements.txtStart training using the following scripts:
./reproduce.shIf you find this paper or this repository helpful, please cite the original paper:
@inproceedings{
wu2018training,
title={Training and Inference with Integers in Deep Neural Networks},
author={Shuang Wu and Guoqi Li and Feng Chen and Luping Shi},
booktitle={International Conference on Learning Representations},
year={2018},
url={https://openreview.net/forum?id=HJGXzmspb},
}