Skip to content

Latest commit

 

History

History
62 lines (35 loc) · 2.25 KB

README.md

File metadata and controls

62 lines (35 loc) · 2.25 KB

ProMix: Combating Label Noise via Maximizing Clean Sample Utility

This is the PyTorch implementation of our IJCAI 2023 paper ProMix. A previous version at the LMNL_challenge branch won the 1st Learning and Mining with Noisy Labels Challenge in IJCAI-ECAI 2022.

Title: ProMix: Combating Label Noise via Maximizing Clean Sample Utility

Authors: Ruixuan Xiao, Dong Yiwen, Haobo Wang, Lei Feng, Runze Wu, Gang Chen, Junbo Zhao

Affliations: Zhejiang University, Nanyang Technological University, NetEase Fuxi AI Lab

@inproceedings{ijcai2023p494,
  title     = {ProMix: Combating Label Noise via Maximizing Clean Sample Utility},
  author    = {Xiao, Ruixuan and Dong, Yiwen and Wang, Haobo and Feng, Lei and Wu, Runze and Chen, Gang and Zhao, Junbo},
  booktitle = {Proceedings of the Thirty-Second International Joint Conference on
               Artificial Intelligence, {IJCAI-23}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Edith Elkind},
  pages     = {4442--4450},
  year      = {2023},
  month     = {8},
  note      = {Main Track},
  doi       = {10.24963/ijcai.2023/494},
  url       = {https://doi.org/10.24963/ijcai.2023/494},
}

Framework

Framework

Main Results on CIFAR-10/100

result_cf

Main Results on CIFAR-N

result_cfn

Usage

After creating a virtual environment, run

pip install -r requirements.txt

We provide the shell codes for model training in the run.sh file. Please download the source data of CIFAR-10/100 and the noise file of CIFAR-N following Learning with Noisy Labels Revisited: A Study Using Real-World Human Annotations and put them under the data folder.

Acknowledgement

This paper is supported by Netease Youling Crowdsourcing Platform. As the importance of data continues rising, Netease Youling Crowdsourcing Platform is dedicated to utilizing various advanced algorithms to provide high-quality, low-noise labeled samples. Feel free to contact us for more information.