This is an official implementation of "Contrast to Divide: self-supervised pre-training for learning with noisy labels". The code is based on DivideMix implementation.

Following tables summarize main results of the paper:

CIFAR-10:

CIFAR-100:

Clothing1M:

mini-WebVision:

First you need to install dependencies by running pip install -r requirements.txt.

You can download pretrained self-supervised models from Google Drive. Alternatively, you can train them by yourself, using SimCLR implementation. Put them into ./pretrained folder.

Then you can run the code for CIFAR

for Clothing1M

or for mini-WebVision

To run C2D with ELR+ just use the self-suprevised pretrained models with the original code.

This project is licensed under the terms of the MIT license.