chenmc1996 / Robust-LR

Code for our conference paper"Two Wrongs Don’t Make a Right: Combating Confirmation Bias in Learning with Label Noise" and journal submission "Robust and Class-balanced Refurbishment of Noisy Labels"

Robust-LR

Experiments

To obtain the results on standard learning with noisy labels benchmarks,run RoLR.py, such as

 $ python RoLR.py --data_path path-to-your-dataset --num_class 10 --dataset cifar10 --num_epochs 500 --lambda_p 2 --T 1 --r 0.8

To obtain the results on learning with long-tailed noisy labels benchmarks,run RoLR_LT.py, such as

 $ python RoLR_LT.py --data_path path-to-your-dataset --num_class 10 --dataset cifar10 --num_epochs 500 --lambda_p 2 --T 1 --imbalance 0.02 --r 0.2 --conf_mode W --class_weight freq 

Requirements

• Python >= 3.6
• PyTorch >= 1.6
• CUDA
• Numpy

Reference

We thank the implementation of DivideMix and FixMatch in:

• "https://github.com/LiJunnan1992/DivideMix",
• "https://github.com/LeeDoYup/FixMatch-pytorch".