This repository contains the code to reproduce the results of our paper Feature Perturbation Augmentation (FPA): arXiv.
Requirements for running the notebooks are PyTorch, PyTorch Lightning and Numba.
We provide code for training the models with FPA, but pre-trained weights can also be downloaded here and have to be put in a folder called "weights".
As a first step, importance estimates are generated using the get_estimators.ipynb notebook and the perturbation-based evaluation can then be performed using the perturb.ipynb notebook.