DID: "Measuring dissimilarity with diffeomorphism invariance" (ICML 2020)

What is DID?

In a nutshell, DID is a generic dissimilarity that can be used to compare two points (images, time series, ...) in such a way that is invariant to diffeomorphisms between the points. For example: DID(x, rotate(x)) gives a small value.

This repository provides an implementation of DID for images. We hope providing our code will help spur enthusiasm around the method, and allow researchers to reproduce the experiments in our paper.

For more complete motivation (and all the details), see the paper:

Measuring dissimilarity with diffeomorphism invariance, Théophile Cantelobre, Carlo Ciliberto, Benjamin Guedj, Alessandro Rudi (2022). ICML.

Wait... why is the library called did and not diffy like the repo? We have not decided on a name yet... but you can open an issue if you have an opinion.

Documentation

This README presents installation and getting started information. See the documentation for more details.

To make the documentation locally, run make html in the docs directory.

Getting started with DID

$ git clone

Install dependencies

With conda:

$ conda create -n did_env -f environment.yml
$ conda activate did_env

With pip:

$ python3 -m venv did_env
$ source did_env/bin/activate
(did_env) $ pip install -r requirements.txt
(did_env) $ pip install -r docs_requirements.txt  # optional

Using DID (without installing)

(did_env) $ python appendix_peppers_match.py

Installing DID (optional)

(did_env) $ python setup.py install

Hardware acceleration (optional)

DID can be hardware accelerated thanks to torch and CUDA. If you do not have a GPU, set device to cpu.

Playing with DID

The best place to start is in the demo_... .py files in this repository. They contain simple examples you can play around with... think of them as minimal working examples. In fact, they reproduce Figure 1 in the paper.

Once you have an understanding of the way DID works, you can experiment on your own, or use the exp_... .py files. These are for experiments that concern the other figures in the paper.

Feeding DID (data)

The demo_... .py scripts depend on data in the perspective and scenes directories. Other provided scripts require Imagenette (ImageNet will work too). You can set a path to Imagenette/Imagenet in imagenet.py.

Citing this work

If you use DID, we'd be happy to read about it if you cite the pre-print:

 @article{cantelobre2022measuring,
        title={Measuring dissimilarity with diffeomorphism invariance}, 
        author={Théophile Cantelobre and Carlo Ciliberto and Benjamin Guedj and Alessandro Rudi},
        year={2022}, 
        journal={arXiv 2202.05614}, 
        volume={stat.ML},
        url={https://arxiv.org/abs/2202.05614}
 }

theophilec/diffy