Authors: Shengbang Tong*, Yubei Chen*, Yi Ma, Yann LeCun

This repository contains the implementation for the paper "EMP-SSL: Towards Self-Supervised Learning in One Training Epoch." The paper introduces a simplistic but efficient self-supervised learning method called Extreme-Multi-Patch Self-Supervised-Learning (EMP-SSL). EMP-SSL significantly reduces the training epochs required for convergence by increasing the number of fix size image patches from each image instance.

Cifar10 and Cifar100 can be downloaded automatically in the script. ImageNet100 is a special subset of ImageNet. Details can be found in this link.

Current code implementation supports Cifar10, Cifar100 and ImageNet100.

To get started with the EMP-SSL implementation, follow these instructions:

git clone https://github.com/tsb0601/emp-ssl.git
cd emp-ssl

pip install -r requirements.txt

For CIFAR10 or CIFAR100

python main.py --data cifar10 --epoch 2 --patch_sim 200 --arch 'resnet18-cifar' --num_patches 20 --lr 0.3

For ImageNet100

python main.py --data imagenet100 --epoch 2 --patch_sim 200 --arch 'resnet18-imagenet' --num_patches 20 --lr 0.3

Change num_patches here to change the number of patches used in EMP-SSL training.

python main.py --data cifar10 --epoch 30 --patch_sim 200 --arch 'resnet18-cifar' --num_patches 20 --lr 0.3

Because our model is trained with only fixed size image patches. To evaluate the performance, we adopt bag-of-features model from intra-instance VICReg paper. Change test_patches here to adjust number of patches used in bag-of-feature model for different GPUs.

python evaluate.py --model_path 'path to your evaluated model' --test_patches 128

This repo is inspired by MCR2, solo-learn and NMCE repo.

If you find this repository useful, please consider giving a star ⭐ and citation:

@article{tong2023empssl,
title={EMP-SSL: Towards Self-Supervised Learning in One Training Epoch},
author={Shengbang Tong and Yubei Chen and Yi Ma and Yann Lecun},
journal={arXiv preprint arXiv:2304.03977},
year={2023}
}