Joint Asymmetric Loss for Learning with Noisy Labels

This repository is the official pytorch code of the Joint Asymmetric Loss (JAL).

How to use

🔵 We use NCEandAMSE as JAL-CE and NFLandAMSE as JAL-FL in the code.

Benchmark Datasets: The running file is main.py

• dataset: cifar10 | cifar100.
• loss: NCEandAMSE, NFLandAMSE, CE, GCE, etc.
• noise_type: symmetric | asymmetric | dependent (instance-dependent noise) | human (cifar-n).

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

• dataset: webvision | clothing1m.
• loss: NCEandAMSE, NFLandAMSE, CE, GCE, etc.

Examples

NCEandAMSE for cifar10 with 0.8 symmetric noise:

python main.py --dataset cifar10 --noise_type symmetric --noise_rate 0.8 --loss NCEandAMSE    

NCEandAMSE for webvision:

python main_real_world.py --dataset webvision --loss NCEandAMSE

If you have any question, you can contact [email protected]