Co-learning: Learning from noisy labels with self-supervision

This repository contains a unified framework for co-training-based noisy label learning methods.

The official implementation of the paper Co-learning: Learning from noisy labels with self-supervision is also included.

Introduction

Supported algorithms

• Decoupling (NeurIPS'2017)
• Co-teaching (NeurIPS'2018)
• Co-teaching+ (ICML'2019)
• JoCoR (CVPR'2020)
• Co-learning (MM'2021)

Supported datasets:

• CIFAR-10
• CIFAR-100

Supported synthetic noise types:

• 'sym' (Symmetric noisy labels)
• 'asym' (Asymmetric noisy labels)
• 'ins' (Instance-dependent noisy labels)

Dependency

• numpy
• torch, torchvision
• scipy
• addict
• matplotlib

Citation

If you are interested in our repository and our paper, please cite the following paper:

@inproceedings{tan2021co,
  title={Co-learning: Learning from noisy labels with self-supervision},
  author={Tan, Cheng and Xia, Jun and Wu, Lirong and Li, Stan Z},
  booktitle={Proceedings of the 29th ACM International Conference on Multimedia},
  pages={1405--1413},
  year={2021}
}