Skip to content

Learning recognition/segmentation models without end-to-end training. 40%-60% less GPU memory footprint. Same training time. Better performance.

Notifications You must be signed in to change notification settings

blackfeather-wang/InfoPro-Pytorch

Repository files navigation

InfoPro-Pytorch

The Information Propagation algorithm for training deep networks with local supervision.

Update on 2021/01/25: Release Pre-trained models on ImageNet and Cityscapes.

Update on 2021/01/24: Release Code for Image Classification on CIFAR/SVHN/STL10/ImageNet and Semantic Segmentation on Cityscapes.

Introduction

We propose Information Propagation (InfoPro), a locally supervised deep learning algorithm, from the information-theoretic perspective. By splitting the whole deep network into multiple local modules and training them with local InfoPro loss, we reduce the GPU memory footprint by 40-60% without introducing notable extra computational cost or training time, but improve the performance moderately.

Citation

If you find this work valuable or use our code in your own research, please consider citing us with the following bibtex:

@inproceedings{wang2021revisiting,
        title = {Revisiting Locally Supervised Learning: an Alternative to End-to-end Training},
       author = {Yulin Wang and Zanlin Ni and Shiji Song and Le Yang and Gao Huang},
    booktitle = {International Conference on Learning Representations (ICLR)},
         year = {2021},
          url = {https://openreview.net/forum?id=fAbkE6ant2}
}

Get Started

Please go to the folder Experiments on CIFAR-SVHN-STL10, Experiments on ImageNet and Semantic segmentation for specific docs.

Results

  • CIFAR & STL-10

  • ImageNet

  • Semantic Segmentation

GPU Memory Cost

In the paper, we report the minimally required GPU memory to run the InfoPro* algorithm with torch.backends.cudnn.benchmark=True (for practical acceleration). Note that this result is (sometimes largely) different from what is printed by nvidia-smi.

Contact

This repo is a re-implementation of our original code. If you have any question, please feel free to contact the authors. Yulin Wang: wang-yl19@mails.tsinghua.edu.cn.

Acknowledgments

Our code of Semantic Segmentation is from MMSegmentation. We highly appreciate their awesome work!

About

Learning recognition/segmentation models without end-to-end training. 40%-60% less GPU memory footprint. Same training time. Better performance.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published