In this repo, we show how to train a self-supervised model by using Robust Global Contrastive Loss (RGCL) on several unimodal image datasets (e.g., CIFAR10/100, ImageNet100, etc.) and a widely used bimodal image-text dataset CC3M. The code and scripts for reproducing the unimodal and bimodal experimental results in our paper are provided in unimodal_exp and bimodal_exp folder, respectively.
If you find this tutorial helpful, please cite our paper:
@inproceedings{qiu2023not,
title={Not All Semantics are Created Equal: Contrastive Self-supervised Learning with Automatic Temperature Individualization},
author={Zi-Hao Qiu, Quanqi Hu, Zhuoning Yuan, Denny Zhou, Lijun Zhang, and Tianbao Yang},
booktitle={International Conference on Machine Learning},
pages={TBD},
year={2023},
organization={PMLR}
}