SCALE + ASH methods #88
Open
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
Added safeguard for preventing react and scale to be used simultaneously
Note: SCALE has poor performances on CIFAR10 vs SVHN because the penultimate layer has dimension 64 only on the tested resnet18
+ safeguards to prevent using ash with react and scale
☂️ Python Coverage
Overall Coverage
New FilesNo new covered files... Modified Files
|
cofri
requested changes
Aug 30, 2024
| ) | ||
| def test_ash(detector_name, auroc_thr, fpr95_thr): | ||
| """ | ||
| Test ASH + [MLS, MSP, Energy, ODIN, Entropy] on toy blobs OOD dataset-wise task |
There was a problem hiding this comment.
Why don't you test Entropy and MSP, as suggested in the docstring here? BTW, we could now add GEN.
There was a problem hiding this comment.
Because Entropy and MSP relies on softmax and seems to induce numerical instabilities and Nans. It is the same for GEN. Maybe we should put in the doc that ASH should not be used with Entropy MSP and GEN.
There was a problem hiding this comment.
Ok, indeed it can be good to add a warning message somewhere to warn users about ASH
Co-authored-by: cofri <cofri@users.noreply.github.com>
Co-authored-by: cofri <cofri@users.noreply.github.com>
cofri
approved these changes
Aug 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Outdated suggestions cannot be applied.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.
Implements both ASH and SCALE baselines at the same time because they are very similar.
SCALE method
SCALE method basically consists of re-using existing logit-based OOD methods, but with penultimate layer activations scaled. Let$a$ be the activation vector, and $P_p(a)$ the $p$ -th percentile of $a$ 's values. The scaling is computed using the formula
$s = \exp(\frac{\sum_{i} a_i}{\sum_{a_i > P_p(a)} a_i})$
Here, we focus on a Resnet trained on CIFAR10, challenged on SVHN.
Reference
Scaling for Training Time and Post-hoc Out-of-distribution Detection Enhancement, ICLR 2024
https://arxiv.org/abs/2111.12797
ASH method
This notebook aims at evaluating the ASH method.
ASH method basically consists of re-using existing logit-based OOD methods, but with penultimate layer activations scaled and pruned. Let$a$ be the activation vector, and $P_p(a)$ the $p$ -th percentile of $a$ 's values. The scaling is computed using the same formula as above. The activation vector is pruned for values $a_i \leq P_p(a)$ .
Here, we focus on a Resnet trained on CIFAR10, challenged on SVHN.
Reference
Extremely Simple Activation Shaping for Out-of-Distribution Detection, ICLR 2023
http://arxiv.org/abs/2209.09858