Lrp for review#900
Open
Nimanui wants to merge 14 commits intosunlabuiuc:masterfrom
Open
Conversation
Nimanui
commented
Mar 20, 2026
Contributor
Author
There was a problem hiding this comment.
This is just some refactoring to pull out the visualization components for use with other saliency methods, I don't know how we may want to formalize visualization going forward or otherwise evolve that.
Logiquo
reviewed
Apr 18, 2026
Comment on lines
+405
to
+431
| if mode == "binary": | ||
| # Binary logits have shape [B] or [B, 1]. | ||
| # target_class_idx=1 (positive) → use logit as-is. | ||
| # target_class_idx=0 (negative) → negate so LRP attributes for the | ||
| # "not-positive" direction (relevance conservation still holds for -logit). | ||
| if target_class_idx is not None: | ||
| idx = ( | ||
| target_class_idx | ||
| if isinstance(target_class_idx, int) | ||
| else int(target_class_idx.item()) | ||
| ) | ||
| if idx == 0: | ||
| return -logits | ||
| return logits | ||
| if mode in ("multiclass", "multilabel"): | ||
| if target_class_idx is None: | ||
| target_class_idx = torch.argmax(logits, dim=-1) | ||
| elif not isinstance(target_class_idx, torch.Tensor): | ||
| target_class_idx = torch.tensor( | ||
| target_class_idx, device=logits.device | ||
| ) | ||
| batch_size = logits.size(0) | ||
| output_relevance = torch.zeros_like(logits) | ||
| output_relevance[range(batch_size), target_class_idx] = logits[ | ||
| range(batch_size), target_class_idx | ||
| ] | ||
| return output_relevance |
Collaborator
There was a problem hiding this comment.
The target_class_index hanlding has been unified in #926. With the following semantics
- binary will always have targret_class_index 0
- the attribution is always towards positive, and up to the evaluator to flip the sign for negative class contribution
- if not specified, both multi-class and multi-label uses argmax
Logiquo
reviewed
Apr 18, 2026
Comment on lines
+26
to
+34
| ReLULRPHandler, | ||
| Conv2dLRPHandler, | ||
| MaxPool2dLRPHandler, | ||
| AvgPool2dLRPHandler, | ||
| AdaptiveAvgPool2dLRPHandler, | ||
| BatchNorm2dLRPHandler, | ||
| FlattenLRPHandler, | ||
| DropoutLRPHandler, | ||
| RNNLRPHandler, |
Collaborator
There was a problem hiding this comment.
Do we need special handling for Conv1d, Sigmoid, Tanh, Softmax in StageNet?
Logiquo
reviewed
Apr 18, 2026
Comment on lines
+446
to
+452
| def _propagate_through_residual_block( | ||
| self, | ||
| block_name: str, | ||
| block_layer_names_rev: list, | ||
| r_in: torch.Tensor, | ||
| ctx: _LRPHookContext, | ||
| ) -> torch.Tensor: |
Collaborator
There was a problem hiding this comment.
this seems to ignore rule args for LRP and hard code to epsilon rule regardless if it is epsilon rule or alpha beta?
Logiquo
reviewed
Apr 18, 2026
Comment on lines
+395
to
+396
| c_neg = crop_spatial(F.conv_transpose2d(s, W_neg, None, **trans_kw), x.shape) | ||
| return x * (alpha * c_pos + beta * c_neg) |
Collaborator
There was a problem hiding this comment.
The linear uses x_p * (alpha * c_pos - beta * c_neg). Is it expected for them to be different?
Logiquo
requested changes
Apr 18, 2026
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
Looking to add some alpha, beta, and epsilon LRP saliency options run across a few different models (both the MIMIC IV stagenet and the chest x-ray CNN)
Original implementation is based off of https://arxiv.org/abs/1604.00825 and is very vibey so I'm expecting some comments
I may prune down the example options a bit as this ended up a bit bloated, and I'll need to add some more documentation