Official PyTorch Implementation · NeurIPS 2024
This repository provides training scripts for benchmark, semi-supervised, and real-world noisy-label learning.
-
Losses:
$\epsilon$ -softmax with CE or FL, combined with MAE. - Noise types: symmetric, asymmetric, instance-dependent, human, and real-world.
- Datasets: CIFAR-10, CIFAR-100, CIFAR-N, WebVision, and Clothing1M.
Note
In the code, ECE and EFL denote CE and FL with ECEandMAE and EFLandMAE.
git clone https://github.com/cswjl/eps-softmax.git && cd eps-softmax| Setting | Entry point | Datasets | Noise types |
|---|---|---|---|
| Benchmark | main.py |
cifar10, cifar100 |
symmetric, asymmetric, dependent, human |
| Semi-supervised | main_semi.py |
cifar10, cifar100 |
symmetric, asymmetric, dependent, human |
| Real-world | main_real_world.py |
webvision, clothing1m |
Real-world label noise |
| Argument | Description | Examples |
|---|---|---|
--dataset |
Dataset | cifar10, cifar100, webvision, clothing1m |
--loss |
Loss function (benchmark and real-world) | ECEandMAE, EFLandMAE, CE, GCE |
--noise_type |
Noise type (benchmark and semi-supervised) | symmetric, asymmetric, dependent, human |
--noise_rate |
Noise rate, or CIFAR-N label set with human |
0.8, worst, noisy100 |
--root |
Dataset root directory | ../data (default) |
📂 Repository structure
eps-softmax/
├── main.py # Benchmark training
├── main_semi.py # Semi-supervised training
├── main_real_world.py # Real-world dataset training
├── losses.py # Loss function implementations
├── config.py # Loss and regularization configurations
├── models.py # Network architectures
├── utils.py # Training and evaluation utilities
└── datasets/ # Data loaders, noise annotations, and dataset lists
# CIFAR-10 · 80% symmetric noise · ECE + MAE
python3 main.py --dataset cifar10 --noise_type symmetric --noise_rate 0.8 --loss ECEandMAE
# CIFAR-10N · worst human labels · ECE + MAE (Semi)
python3 main_semi.py --dataset cifar10 --noise_type human --noise_rate worst
# WebVision · real-world noise · ECE + MAE
python3 main_real_world.py --dataset webvision --loss ECEandMAEIf you find this work useful, please cite our paper:
@article{wang2024epsilon,
title={$\epsilon$-Softmax: Approximating One-Hot Vectors for Mitigating Label Noise},
author={Wang, Jialiang and Zhou, Xiong and Zhai, Deming and Jiang, Junjun and Ji, Xiangyang and Liu, Xianming},
journal={Advances in Neural Information Processing Systems},
volume={37},
pages={32012--32038},
year={2024}
}Questions about the paper or code? Contact Jialiang Wang at cswjl@stu.hit.edu.cn.
⭐ Star us on GitHub — it motivates us a lot!
