Fourier Sensitivity and Regularization of Computer Vision Models

This repository contains code for work presented in Fourier Sensitivity and Regularization of Computer Vision Models, published at TMLR 2022. Fourier Sensitivity of computer vision models is based on a rigorously defined measure of sensitivity to input frequencies. Please see the paper for details.

Installing libraries

The code was run with Python3.8

pip install -r requirements.txt

Generating Fourier Sensitivity plots

Please see the jupyter notebook "Fourier-Sensitivity.ipynb" for examples of plotting the Fourier Sensitivity of pre-trained models. Run all cells to re-generate the plots (be sure to change the path to the dataset, i.e., PATH_TO_IMAGENET)

Fourier Regularized training

We have provided a reference implementation of Fourier-regularized training on CIFAR10 (train.py). Please use the commands below.

# standard training (CIFAR10)
python train.py

# Fourier-regularized training (CIFAR10)
python train.py --regularizer {LSF,MSF,HSF,ASF} --regularier_lambda 0.5

Acknowledgement

CIFAR10 training code is based on https://github.com/kuangliu/pytorch-cifar