$\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise

This repository is the official pytorch implementation of the eps-softmax [NeurIPS2024].

How to use

We simplify $\epsilon$-softmax with CE and FL by ECE and EFL in the code.

Benchmark Datasets: The running file is main.py.

• --dataset: cifar10 | cifar100, etc.
• --loss: ECEandMAE, EFLandMAE, CE, GCE, etc.
• --noise_type: symmetric | asymmetric | dependent (instance-dependent noise), etc.

CE $_\epsilon$+MAE (Semi): The running file is main_semi.py.

• --dataset: cifar10 | cifar100.
• --noise_type: human (cifar-n dataset), etc.

Real-World Datasets: The running file is main_real_world.py.

• --dataset: webvision | clothing1m.
• --loss: ECEandMAE, EFLandMAE, CE, GCE, etc.

Examples

ECEandMAE for cifar10 0.8 symmetric noise:

$ python3 main.py --dataset cifar10 --noise_type symmetric --noise_rate 0.8 --loss ECEandMAE    

ECEandMAE(Semi) for cifar10 human (cifar-n dataset) worst:

$ python3 main_semi.py --dataset cifar10 --noise_type human --noise_rate worst  

ECEandMAE for webvision:

$ python3 main_real_world.py --dataset webvision --loss ECEandMAE

Reference

For technical details and full experimental results, please check the paper. If you have used our method or code in your own, please consider citing:

@inproceedings{wang2024epsilonsoftmax,
  title={\${\textbackslash}epsilon\$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise},
  author={Jialiang, Wang and Xiong, Zhou and Deming, Zhai and Junjun, Jiang and Xiangyang, Ji and Xianming, Liu},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year={2024}
}