Feature Visualization: grayscale images #163
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…properties output
…adapt the parametrization to grayscale
Agustin-Picard
added
enhancement
|
I had to limit tensorflow to versions <2.16 as the latest updates break a lot of things. Edit: But I'm still getting what seems to be compatibility errors, so it's not due to the tf version... |
AntoninPoche
approved these changes
Sep 25, 2024
lucashervier
approved these changes
Sep 30, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Feature visualization for grayscale images
This PR adapts the code for generating feature visualizations (via Fourier, in pixel space, and using MaCo) for models trained on grayscale images.
In the case of grayscale images, no re-correlation is needed when optimizing in the Fourier domain (standard and MaCo). However, there is no standardized dataset that would generally work on any grayscale application (like ImageNet for natural images), so a dataset must be provided to compute the average magnitude for MaCo.
Disclaimer: Depending on the model under study, the optimization process can provide images that are really difficult to interpret. It seems to work much better on 1-Lipschitz models (trained using DEEL-LIP) and sometimes MaCo helps to generate better visualizations.
It responds to the feature requested in #148
Fix CI
The release of TF 2.16 integrates Keras 3.X, which seems to break some functionality. I've limited the installation of TF to <2.16.
Also, the release of NumPy 2.X breaks packages that were compiled using 1.X, and this includes TF. Thus, I've limited NumPy to 1.X.