prabhuteja12 / pest

Code release for paper Test time Adaptation through Perturbation Robustness

Code for Test time Adaptation through PErturbation robuSTness (PEST)

Code release for paper Test time Adaptation through Perturbation Robustness by Prabhu Teja S, and François Fleuret to be published at NeurIPS 2021 Workshop on Distribution Shifts.

Installation

The requirements are pytorch, pytorch-lightning, and RobustBench package.

Running the code

The code heavily relies on yaml formatted configuration files. These are read in main.py and passed onto the appropriate function. A sample configuration file is given. The code can be run with

python main.py sample_config.yaml

Downloading the datasets

The current codebase supports CIFAR-10-C, CIFAR-100-C, and VisDA-C datasets. They have to be downloaded and placed in the datasets folder.

Citing

If you find me work or code useful, consider citing us using

@misc{sivaprasad2021test,
      title={Test time Adaptation through Perturbation Robustness}, 
      author={Prabhu Teja Sivaprasad and François Fleuret},
      year={2021},
      eprint={2110.10232},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}