A few bugs in our implementation were brought to our attention by Zongtao (Github) and Tim Beyer (Google Scholar)

The bugs are

• We were using the transpose jacobian instead of the jacobian.
• We take a dot product when computing the expected gradient by multiplying the perturbed_output and noise_gradient, whereas an outer product needs to be computed.

A bug fixed pytorch implementation is available at https://github.com/benbergner/dps . We apologies for any inconvenience caused.

This is a reference implementation for "Differentiable Patch Selection for Image Recognition" by Jean-Baptiste Cordonnier, Aravindh Mahendran, Alexey Dosovitskiy, Dirk Weissenborn, Jakob Uszkoreit, and Thomas Unterthiner.

Contact: [email protected]

When using this code, please cite the paper:

@article{cordonnier2021differentiable,
    title={Differentiable Patch Selection for Image Recognition},
    author={Cordonnier, Jean-Baptiste and Mahendran, Aravindh and Dosovitskiy, Alexey and Weissenborn, Dirk and Uszkoreit, Jakob and Unterthiner, Thomas}
    journal={CVPR},
    year={2021}
}

This code was developed using JAX and FLAX. It also requires the following packages:

chex
clu
einops
numpy
ml_collections
optax

This release includes the data for the Billiard Experiments (see paper). To reproduce these (or re-use the data for your own purposes) please download the data from http://storage.googleapis.com/gresearch/ptokp_patch_selection/billiard.tar.xz and extract it to "libs/datasets/billiard".

To run the data, pick a config file and run it, for example:

python3 image_classification.py --config configs/billiard/topk.py --workdir /tmp/topk

This repository is licensed under the Apache License, Version 2.0. See LICENSE for details.

This is not an official Google product.