This is the official code for the paper "Hard Negative Sampling via Regularized Optimal Transport for Contrastive Representation Learning". This repository contains the implementation of HCL-OT and related experiments described in the paper.

For instance, to run the code on the "cifar100" dataset using the entropy Optimal Transport (OT) method with a regularization parameter epsilon of 0.7:

python main.py --dataset_name "cifar100" --reg 0.7

If you find this repo useful for your research, please consider citing the paper:

@article{jiang2021hard,
  title={Hard Negative Sampling via Regularized Optimal Transport for Contrastive Representation Learning},
  author={Jiang, Ruijie and Ishwar, Prakash and Aeron, Shuchin},
  journal={arXiv preprint arXiv:2111.03169},
  year={2021}
}

For any questions, please contact Ruijie Jiang ([email protected])

This code is a modified version of the HCL implementation by Josh/HCL. The only difference from the their code is a minor alteration in the hard negative sampling approach, we change it from function "criterion" in their code to "OT_hard". To ensure a fair comparison, we have maintained all hyperparameters in Josh's implementation as they were in the original code.

Part of this code is inspired by leftthomas/SimCLR, by Josh/HCL, and by fanyun-sun/InfoGraph.