Skip to content

lenbrocki/Feature-Perturbation-Augmentation

Repository files navigation

Feature-Perturbation-Augmentation

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.

About

This repository contains the code to reproduce the results of our paper Feature Perturbation Augmentation (FPA)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published