Noise Attention Learning: Enhancing Noise Robustness by Gradient Scaling

This repository is the implementation of Noise Attention Learning: Enhancing Noise Robustness by Gradient Scaling (NeurIPS 2022).

Requirements

This codebase is written for python3, other necessary python packages are including

• numpy=1.20.1
• torch=1.8.1
• torchvision=0.2.2
• Pillow=8.2.0

Training

Usage

To train the model in the paper, run the following commands:

Datasets with synthetic label noise

Train the network on the CIFAR-10 dataset with 60% symmmetric label noise:

python train_cifar.py --root_dir <path_to_save_data> --noise_mode 'sym' --r 0.6 

Train the network on the CIFAR-10 dataset with 40% asymmmetric label noise :

python train_cifar.py --root_dir <path_to_save_data> --noise_mode 'asym' --r 0.4 --num_epochs 300 

Please refer to the Appendix for more information of hyperparameters.

hyperparameter options:
--batch_size            batch size
--lr                    learning rate
--num_epochs            number of epochs
--noise_mode            label noise model(e.g. sym, asym)
--r                     noise level (e.g. 0.4)
--lam                   penalty loss coefficient
--delta                 confidence threshold (optional)
--beta                  auxiliary regularization coefficient (optional for RNAL)
--m                     momentum for target estimation
--es                    epoch that starts to perform target estimation

Datasets with real-world label noise

Before running the code, please download the datasets ANIMAL-10N, Clothing1M and Webvision.

Citing this work

If you use this code in your work, please cite our paper.