NARL-Adjuster

This is an official PyTorch implementation of Improve Noise Tolerance of Robust Loss via Noise-Awareness.

Environment

• python 3.7.10
• torch 0.8.1
• torchvision 0.9.1

  Running this example

  ResNet32 on CIFAR10 with 40% unif noise:

  python main.py --dataset cifar10 --corruption_prob 0.4 --corruption_type unif --epochs 120 --warmup_epochs 0 --batch_size 100 --lr 1e-1 --wd 5e-4 --mwd 1e-4

  Result(CIFAR10)

Noise Type	Test Accuracy
Symmetric	88.10%
Asymmetric	88.03%
Instance	87.33%

Acknowledgments

Thanks to the pytorch implementation of Meta-Weight-Net(https://github.com/xjtushujun/meta-weight-net).

Contact: Ding Kehui(dkh19970303@163.com).