All You Need Is a Few Shifts: Designing Efficient Convolutional Neural Networks for Image Classification

This repo is the official implementation of CVPR2019 paper "All You Need Is a Few Shifts: Designing Efficient Convolutional Neural Networks for Image Classification".

Paper introduction

The comparison of Sparse Shift Layer (SSL) with other shift operations applied to feature maps.

Preparation

Prerequistes

• Python 3.6
• Pytorch 1.0
• CUDA 8.0 or higher

Usage

Data folds structure

.ImageNetOrigin/
  └── train/
  |      ├── n01440764
  |      |      ├── n01440764_18.JPEG
  |      |      ├── n01440764_36.JPEG
  |      |      └── ...
  |      ├── n01443537
  |      └── ...
  └── val2/
         ├── n01440764
         |      ├── n01440764_18.JPEG
         |      ├── n01440764_36.JPEG
         |      └── ...
         ├── n01443537
         └── ...

Train

bash train_imagenet.sh

Eval

bash test_imagenet.sh

Main Results

Networks	MAdds	Prarms	Model Weights	Top-1	TrainLog
FE-Net 1.0x	301M	3.7M	To-Be-Released	73.6 %	log-file
FE-Net 1.0x + Autoaugment	301M	3.7M	To-Be-Released	73.9 %	log-file
FE-Net 1.375x	563M	5.9M	To-Be-Released	75.7 %	log-file
FE-Net 1.375x + Autoaugment	563M	5.9M	To-Be-Released	76.6 %	log-file

Citation

Please consider citing our paper if you find it useful for your research.

@inproceedings{Chen2019ShiftNet,
  author = {Weijie Chen, Di Xie, Yuan Zhang, Shiliang Pu},
  title = {All You Need is a Few Shifts: Designing Efficient Convolutional Neural Networks for Image Classification},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2019},
}

License

This project is released under the Apache 2.0 license. Other codes from open source repository follows the original distributive licenses.

Acknowledgement

We appreciate Xianfeng Li for transforming the source code from Caffe to Pytorch. If you have any problem about this work, please feel free to contact Xianfeng Li (lixianfeng6-at-hikvision.com) and Weijie Chen (chenweijie5-at-hikvision.com).