Coalesce and strip explicit zeros before CAN index access - #376
Merged
Conversation
LiftLayer and the CAN attention modules call indices() on user-supplied neighborhood matrices. Sparse matrices built with toponetx's from_sparse preserve explicit zero entries (e.g. zero-valued diagonal self-loops emitted by CellComplex.adjacency_matrix), so the number of stored entries does not match the number of true neighborhood relations: - uncoalesced input raised RuntimeError on indices()/values() - explicit zeros created phantom neighbors in attention softmax - LiftLayer produced one signal row per stored entry instead of per edge, mismatching x_1 when concatenating Add a sanitize_neighborhood helper that coalesces and removes explicit zero entries, and apply it before all index access. References #236, references #242
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #376 +/- ##
==========================================
- Coverage 96.43% 96.36% -0.08%
==========================================
Files 58 58
Lines 2078 2090 +12
==========================================
+ Hits 2004 2014 +10
- Misses 74 76 +2 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
There was a problem hiding this comment.
Pull request overview
This PR addresses correctness issues in CAN modules when consuming user-supplied sparse neighborhood matrices that may be uncoalesced and/or contain explicit stored zeros (e.g., produced by toponetx.utils.sparse.from_sparse), which previously led to runtime errors on indices()/values() and phantom neighborhood relations during attention computation.
Changes:
- Adds a
sanitize_neighborhoodhelper to coalesce sparse COO neighborhoods and drop explicit zero entries. - Applies neighborhood sanitization prior to index-based access in
LiftLayerand the CAN attention modules to align stored entries with true neighborhood relations.
Suppressed comments (1)
topomodelx/nn/cell/can_layer.py:771
- Same as above: once
sanitize_neighborhood()has stripped explicit zeros, checking emptiness via_nnz()avoids two full scans overvalues().nonzero()and keeps the early-return logic simpler.
neighborhood = sanitize_neighborhood(neighborhood)
# If there are no non-zero values in the neighborhood, then the neighborhood is empty. -> return zero tensor
if not neighborhood.values().nonzero().size(0) > 0 and self.concat:
return torch.zeros(
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Comment on lines
+37
to
+45
| neighborhood = neighborhood.coalesce() | ||
| values = neighborhood.values() | ||
| keep = values != 0 | ||
| if bool(keep.all()): | ||
| return neighborhood | ||
| indices = neighborhood.indices()[:, keep] | ||
| return torch.sparse_coo_tensor( | ||
| indices, values[keep], neighborhood.shape, device=neighborhood.device | ||
| ).coalesce() |
Comment on lines
+16
to
+24
| def sanitize_neighborhood(neighborhood: torch.Tensor) -> torch.Tensor: | ||
| r"""Return a coalesced neighborhood matrix without explicit zero entries. | ||
|
|
||
| The number of stored entries of the returned tensor equals its number of | ||
| true non-zero entries, so that iterating over stored entries (e.g. via | ||
| ``indices()``) corresponds to actual neighborhood relations. This matters | ||
| because sparse matrices built with ``toponetx.utils.sparse.from_sparse`` | ||
| preserve explicit zeros that other construction paths (e.g. dense casting) | ||
| drop. |
Comment on lines
+567
to
571
| neighborhood = sanitize_neighborhood(neighborhood) | ||
|
|
||
| # If there are no non-zero values in the neighborhood, then the neighborhood is empty. -> return zero tensor | ||
| if not neighborhood.values().nonzero().size(0) > 0 and self.concat: | ||
| return torch.zeros( |
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
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.
LiftLayer and the CAN attention modules call indices() on user-supplied neighborhood matrices. Sparse matrices built with toponetx's from_sparse preserve explicit zero entries (e.g. zero-valued diagonal self-loops emitted by CellComplex.adjacency_matrix), so the number of stored entries does not match the number of true neighborhood relations:
Add a sanitize_neighborhood helper that coalesces and removes explicit zero entries, and apply it before all index access.
Closes #236 and closes #242