As a first step, requirements.txt have to be installed, and N24News and MS-COCO 2014 have to be downloaded from their official sources in a data folder, and extracted.

Then, representations for all models can be computed by running the extract_representations.py script, specifying a model and a dataset. After that, $I_d$Corr can be computed using correlation.py, while baselines are computed by running baselines.py. In both cases, a dataset parameter can be specified.

On ImageNet, we also compute correlation between coarsely aligned representations. The code for this experiment is found in coarse.py.

Finally, the example on MNIST (section 4.1) can be reproduced running mnist.py.