nikokks / promix

PyTorch Code for ProMix: Combating Label Noise via Maximizing Clean Sample Utility

ProMix: Combating Label Noise via Maximizing Clean Sample Utility

Winner of the 1st Learning and Mining with Noisy Labels Challenge (link) in IJCAI-ECAI 2022.

Ranks:

• 1st under Task1 (Image Classification, CIFAR-10N);
• 1st under Task1 (Image Classification, CIFAR-100N);
• 1st under Task2 (Label Noise Detection, CIFAR-10N);
• 1st under Task2 (Label Noise Detection, CIFAR-100N);

The technical report of our work can be found here.

Title: ProMix: Combating Label Noise via Maximizing Clean Sample Utility
Authors: Haobo Wang*, Ruixuan Xiao*, Yiwen Dong, Lei Feng, Junbo Zhao
Institute: Zhejiang University, Chongqing Unveristy

The shell codes are provided in run.sh. Please correctly set the data path arguments.

Framework

Main Results

Accuracy comparisons on CIFAR-10N and CIFAR-100N under different noise types.

Detection ability of ProMix on CIFAR-10N datasets. Noise detection indicates the wrongly-labeled data are regarded as positive, and clean detection regards correctly-labeled ones as positive.