Awesome

Generalist: Decoupling Natural and Robust Generalization

Official implementation for CVPR'23 paper "Generalist: Decoupling Natural and Robust Generalization"

Prerequisites

• Python (>=3.7)
• Pytorch (>=1.5)
• Torchvision
• CUDA
• Numpy
• AutoAttack

Training and Testing

• Train ResNet-18 on CIFAR10:

  $ CUDA_VISIBLE_DEVICES={your GPU number} python3 main.py 

• Train WRN-32-10 on CIFAR10

  $ CUDA_VISIBLE_DEVICES={your GPU number} python3 main.py --arch 'WRN32'

Then, it will automatically run all the robustness evaluation in our paper, including NAT, PGD20/100, MIM, CW, APGD_ce, APGD_dlr, APGD_t, FAB_t, Square and AutoAttack.

Pretrained model

Citation

If you are interested in our work, please consider citing the related paper:

@inproceedings{wang2023simple,
  title={Generalist: Decoupling Natural and Robust Generalization},
  author={Hongjun Wang and Yisen Wang},
  booktitle={CVPR},
  year={2023}
}

@inproceedings{wang2022selfensemble,
  title={Self-ensemble Adversarial Training for Improved Robustness},
  author={Hongjun Wang and Yisen Wang},
  booktitle={ICLR},
  year={2022}
}