Group-wise Inhibition based Feature Regularization for Robust Classification (ICCV 2021)

arXiv Paper

Requirements

• numpy>=1.17.2
• Pillow>=6.2.0
• pytorch>=1.6.0
• torchvision>=0.7.0

Datasets

• CIFAR-10 and CIFAR-100 datasets can be downloaded by torchvision. Note that you can set torchvision.datasets.CIFAR10(..., download=True) in ./Data/get_data.py to download the corresponding dataset and keep the directory path.

• CIFAR-10-C and CIFAR-100-C datasets can be downloaded with:

    mkdir -p ./data/cifar
    curl -O https://zenodo.org/record/2535967/files/CIFAR-10-C.tar
    curl -O https://zenodo.org/record/3555552/files/CIFAR-100-C.tar
    tar -xvf CIFAR-100-C.tar -C data/cifar/
    tar -xvf CIFAR-10-C.tar -C data/cifar/

Usage

• CIFAR-10 using TENET Training with GPU 0, /dataset/CIFAR10/ is your directory path of dataset.

python main.py --dataset CIFAR10 --data_root /dataset/CIFAR10/ --gpu_id 0

• CIFAR-100 using TENET Training with GPU 0, /dataset/CIFAR100/ is your directory path of dataset.

python main.py --dataset CIFAR100 --data_root /dataset/CIFAR100/ --gpu_id 0

Trained Weights

The trained model weights are available in current directory.

• The model trained with TENET Training on CIFAR-10, which has a 3.50% Top-1 error rate on clean CIFAR-10 dataset and a 12.31% Top-1 mean corruption error rate on CIFAR-10-C dataset.

./trained_weights/CIFAR-10_resnext29.pth

• The model trained with TENET Training on CIFAR-100, which has a 19.46% Top-1 error rate on clean CIFAR-100 dataset and a 35.73% Top-1 mean corruption error rate on CIFAR-100-C dataset.

./trained_weights/CIFAR-100_resnext29.pth

A more complete repo is coming soon

The current repository is the version originally submitted to CMT as supplementary materials.

We are working to update the code (DDP training on ImageNet, adversarial training, etc.) and model parameters, making them easy-to-use.

Citation

Please cite our work if it's useful for your research.

@inproceedings{liu2021group,
  title={Group-wise Inhibition based Feature Regularization for Robust Classification},
  author={Liu, Haozhe and Wu, Haoqian and Xie, Weicheng and Liu, Feng and Shen, Linlin},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={478--486},
  year={2021}
}