Distilling Out-of-Distribution Robustness from Vision-Language Foundation Models [NeurIPS 2023]

Installation

Install from Source:

$ git clone https://github.com/andyz245/RobustKD.git
$ cd easyrobust
$ pip install -e .
$ pip install PyWavelets
$ pip install matplotlib
$ pip install tensorboard

Evaluation Data and Models

$ sh download_data.sh

Teacher can be found at: https://drive.google.com/file/d/1I7h2oe-LP4djuMwFAwhD4kmvOrsgoDEN/view?usp=sharing

Example Checkpoints

Base ResNet50 https://drive.google.com/file/d/1TGvOW6vit4wA1PlzOFomRnKUf1ZjxuWu/view?usp=sharing

Base ViT-B-16/224 https://drive.google.com/file/d/1ly0j_2nfphNgFd2NOJ0NhIea9o3JXqZ_/view?usp=sharing

ResNet50 + DAD https://drive.google.com/file/d/1lbYzO7dp9xqwVMx2NOjPS_X9B3jh-LHv/view?usp=sharing

Running

Scripts can be found in easyrobust/scripts

Please update the paths with your actual system path. To reproduce paper results, please initialize the model with the base pretrained ViT or ResNet50 models.

Citation

If you found our paper or repo interesting, thanks! Please feel free to give us a star or cite our paper!

@inproceedings{
    zhou2023distilling,
    title={Distilling Out-of-Distribution Robustness from Vision-Language Foundation Models},
    author={Andy Zhou and Jindong Wang and Yu-Xiong Wang and Haohan Wang},
    booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
    year={2023},
    url={https://openreview.net/forum?id=iwp3H8uSeK}
}