Dynamic Loss for Robust Learning

This repository contains the source code for the research paper 'Dynamic Loss for Robust Learning,' built upon the mmclassification framework.

Requirements

• Python 3
• yapf (version ====0.40.1)
• PyTorch (version == 1.7.1)
• yaml
• higher(version == 0.2.1)
• mmcv-full (version == 1.5.0)

Installation

The installation process is similar to mmclassification. Please follow the same steps.

Training

python tools/train_meta.py configs/metadynamic/metadynamic_resnet32_cifar10_cor0.2_imb0.1.py 

Visualization

After training, you can visualize the rank weight and margin.

python python tools/visualize_tools/vis_rank_margin.py --config configs/metadynamic/metadynamic_resnet32_cifar10_cor0.2_imb0.1.py --checkpoint work_dirs/metadynamic_resnet32_cifar10_cor0.2_imb0.1/latest.pth

The image will save in directory 'work_dir/metadynamic_resnet32_cifar10_cor0.2_imb0.1/'.

The label correct weight of each rank in each class.

Per class margin.

Citation

If you find Dynamic Loss useful, please cite the following paper

@ARTICLE{10238823,
  author={Jiang, Shenwang and Li, Jianan and Zhang, Jizhou and Wang, Ying and Xu, Tingfa},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Dynamic Loss for Robust Learning}, 
  year={2023},
  volume={45},
  number={12},
  pages={14420-14434},
  doi={10.1109/TPAMI.2023.3311636}}

Acknowledgement

The code is based on mmclassification and BalancedMetaSoftmax.Thanks for their great contributions on the computer vision community.