Learning Decorrelated Representations Efficiently Using Fast Fourier Transform

This is the official implementation of the following paper:

Yutaro Shigeto*, Masashi Shimbo*, Yuya Yoshikawa, Akikazu Takeuchi. Learning Decorrelated Representations Efficiently Using Fast Fourier Transform. CVPR 2023.

* Equal contribution.

[ arXiv | CVF | Short presentation (YouTube) ]

Setup

1. Clone this repository, including the submodule (solo-learn)

   git clone --recurse-submodules https://github.com/yutaro-s/scalable-decorrelation-ssl.git
   

2. Add some files to the solo-learn repository

   make solo-main
   

3. Build a Docker image

   make docker-build
   

4. Set your API key and username if you intend to use W&B

   export WANDB_API_KEY=[API key]
   export WANDB_ENTITY=[username]
   

5. Launch a Docker container

   make docker-run
   

Training and Evaluation

project

1. Self-supervised learning on ImageNet

   WANDB_PROJECT=[project name] bash script/in1k-r50-d8192/pretrain/sbarlow.sh
   

2. Linear evaluation on ImageNet

   WANDB_PROJECT=[project name] bash ./script/in1k-r50-d8192/linear/sbarlow.sh [path to the checkpoint]
   

Citation

If you use this code, please cite our paper:

@InProceedings{Shigeto_2023_CVPR,
    author    = {Shigeto, Yutaro and Shimbo, Masashi and Yoshikawa, Yuya and Takeuchi, Akikazu},
    title     = {Learning Decorrelated Representations Efficiently Using Fast Fourier Transform},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {2052-2060}
}

Acknowledgments

This repository is built using solo-learn. I would like to express my gratitude to the authors of solo-learn.

This work is based on results obtained from Project JPNP20006, commissioned by the New Energy and Industrial Technology Development Organization (NEDO).