ADJUSTING PRETRAINED BACKBONES FOR PERFORMATIVITY

We provide our implementation for PaP (Performativity-aware Predictor).

Install Environment

Create a conda environment

    conda create -n PaP python=3.10
    conda activate PaP

Please install packages with pip install -r requirements.txt

Setup tllib

Clone the repository tllib Put the additional datasets files provided under PaP/datasets/ folder into Transfer-Learning-Library/tllib/vision/datasets directory (together with the init.py, also please add the dataset names into the dictionary, and include the imports)

Run Scripts

You can reproduce our experiments in our paper with commands ./scripts/<exp_type>/<exp_name>.

Directory Structure

The directory structure should look like below

PaP/
|–– datasets/
|   |–– __init__.py
|   |–– agnews.py
|   |–– amazon.py
|   |–– cifar.py
|   |–– imagenet100.py
|   |–– terraincognita.py
|–– scripts/
|   |–– additional_domain_shift
|   |–– language
|   |–– model_switching
|   |–– vision
|–– architectures.py
|–– main.py
|–– nlp_utils.py
|–– performative_util.py
|–– utils.py
|–– requirements.txt

@misc{demirel2024adjusting,
      title={Adjusting Pretrained Backbones for Performativity}, 
      author={Berker Demirel and Lingjing Kong and Kun Zhang and Theofanis Karaletsos and Celestine Mendler-Dünner and Francesco Locatello},
      year={2024},
      eprint={2410.04499},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.04499}, 
}