zeroQiaoba / IRNet

Official code of "IRNet: Iterative Refinement Network for Noisy Partial Label Learning"

IRNet: Iterative Refinement Network for Noisy Partial Label Learning

Correspondence to:

• Zheng Lian ([e-mail])
• Mingyu Xu ([e-mail])

Paper

IRNet: Iterative Refinement Network for Noisy Partial Label Learning
Zheng Lian, Mingyu Xu, Lan Chen, Licai Sun, Bin Liu, Jianhua Tao

Please cite our paper if you find our work useful for your research:

@article{lian2022arnet,
  title={ARNet: Automatic Refinement Network for Noisy Partial Label Learning},
  author={Lian, Zheng and Xu, Mingyu and Chen, Lan and Sun, Licai and Liu, Bin and Tao, Jianhua},
  journal={arXiv preprint arXiv:2211.04774},
  year={2022}
}

Usage

Datasets

# download dataset and put it into ./dataset (or you can download it via torchvision)
https://drive.google.com/file/d/18YrX6JFzOpG2a0OW1jyG65DFgG6r1Seg/view   ->   ./dataset

Run IRNet

cd irnet
python -u train_merge.py --dataset='cifar10' --partial_rate=0.3  --noise_rate=0.3 --epochs=1000 --encoder='resnet' --lr=0.01 --lr_adjust='Case1' --optimizer='sgd' --weight_decay=1e-3 --gpu=0 --correct_auto --correct_autowin=100 --correct_threshold_range='0.008,0.008' --correct_type='cluster' --correct_update='case3' --loss_type='SCE' --sce_alpha=6.0  --sce_beta=1.0

Run PiCO Baseline

cd irnet
python -u train_merge.py --dataset='cifar10' --partial_rate=0.3  --noise_rate=0.3 --epochs=1000 --encoder='resnet' --lr=0.01 --lr_adjust='Case1' --optimizer='sgd' --weight_decay=1e-3 --gpu=0

For other datasets and other settings, please refer to run.sh

Acknowledgement

Thanks to PiCO, RC&CC, PRODEN, LWC&LWS, and LOG.