Entropy-based selection criterion

Official PyTorch Implementation of paper "A Joint Training Framework for Learning with Noisy Labels".

Paper "A Joint Training Framework for Learning with Noisy Labels" is accepted to SCIENTIA SINICA Informationis 2023.

论文 “面向标签噪声的联合训练框架” 被 中国科学-信息科学 2023 接收.

Training

For CIFAR-10, warm_up = 10,model = resnet18

For CIFAR-100, warm_up = 30,model = resnet34

Noise Labels Settings : CIFAR-10 & CIFAR-100 (sym 0.2, sym 0.5, pair 0.4, instance 0.2, instance 0.4)

Run

python main.py --dataset cifar10 --noise_mode sym --r 0.2 --penal_coeff 0.3 --T 3 --threshold 0.3 --main_type base --gpuid 0

More information can be found in "/code/run.sh"

Note that the code refers to Dividemix (ICLR 2020) and Self-Filtering (ECCV 2022).

More results

1. hyperparameter selection of T and \epsilon

2. Training curve where "Base" denotes the Entropy-based selection criterion.

Cite

Please kindly cite our work if this work is helpful for your research.

@article{:/publisher/Science China Press/journal/SCIENTIA SINICA Informationis///10.1360/SSI-2022-0395,
  author = "wei qi,sun haoliang,yin yilong,ma yuling",
  title = "面向标签噪声的联合训练框架",
  journal = "SCIENTIA SINICA Informationis",
  year = "2023",
  pages = "-",
  url = "http://www.sciengine.com/publisher/Science China Press/journal/SCIENTIA SINICA Informationis///10.1360/SSI-2022-0395,
  doi = "https://doi.org/10.1360/SSI-2022-0395"
}