Scale Matters: Attribution Meets the Wavelet Domain to Explain Model Sensitivity to Image Corruptions

This repository contains the supplementary material to replicate the results of the paper Scale Matters: Attribution Meets the Wavelet Domain to Explain Model Sensitivity to Image Corruptions.

The repository is organized as follows:

* models/ contains the models weights used to evaluate the WCAM.

* data/ contains the complementary data to ImageNet datasets (i.e., label files) and the subset of ImageNet images used to make the benchmarks on the Xplique benchmark library.

Note that the ImageNet data for the Xplique benchmark is in the ImageNet-Xplique folder, and that the validation images from ImageNet should be placed in ImageNet-val.

* outputs/ contains raw data outputs

Usage

You'll need to download the ImageNet validation set, ImageNet-A and ImageNet-R. Put the images of ImageNet in the folder data/ImageNet, and the content of the ImageNet-A and ImageNet-R archives in the folders data/ImageNet-A and data/ImageNet-R.

Download ImageNet here :https://www.image-net.org/download.php, ImageNet-A here : https://github.com/hendrycks/natural-adv-examples and ImageNet-R here : https://github.com/hendrycks/imagenet-r