Matrix-SSL

Official implementation of ICML 2024 paper 'Matrix Information Theory for Self-Supervised Learning' (https://arxiv.org/abs/2305.17326).

Introduction

PyTorch Implementation for Matrix-SSL, Modified from MEC Codes.

Installation

Requirement:

• Conda

Once installed conda, you can create the ssl environment using conda env create -f environment_ssl.yaml.

Pretraining

In order to get the best checkpoint over 100 epochs, run

$ python main_pretrain.py \
  -a resnet50 \
  --dist-url 'tcp://localhost:10501' --multiprocessing-distributed --world-size 1 --rank 0 \
  --mce_mu 1.0 --mce_lambd 1.0 --gamma 1.0 \ 
  /path/to/imagenet/dataset/

Linear Probing

To evaluate the pre-trained models by linear probing, simply run:

$ python main_lincls.py \
  -a resnet50 \
  --dist-url 'tcp://localhost:10051' --multiprocessing-distributed --world-size 1 --rank 0 \
  --pretrained /path/to/pretrained/checkpoint \
  --lars \
  /path/to/imagenet/dataset

Transfer learning on COCO

We use exactly the same code as MoCo. One can just follow its instructions, firstly convert the pre-trained weights to detectron2's format, then do transfer learning.

Acknowledgement

This repo is mainly based on MEC and MoCo. Many thanks to their wonderful work!

Citations

Please cite the paper and star this repo if you use Matrix-SSL and find it interesting/useful, thanks! Feel free to contact [email protected] | [email protected] or open an issue if you have any questions.

@article{zhang2023matrix,
  title={Matrix Information Theory for Self-Supervised Learning},
  author={Zhang, Yifan and Tan, Zhiquan and Yang, Jingqin and Weiran, Huang and Yuan, Yang},
  journal={arXiv preprint arXiv:2305.17326},
  year={2023}
}