Add graph validation and statistics logging for debugging graph construction#519
Add graph validation and statistics logging for debugging graph construction#519princekumarlahon wants to merge 4 commits intomllam:mainfrom
Conversation
|
Happy to iterate on this if any changes are needed! |
kshirajahere
left a comment
kshirajahere
left a comment
There was a problem hiding this comment.
Hey, thanks for your work on this. I found a couple of things that are worth clearing up, and I’ve left inline comments with the details.
The PR also includes an unrelated CustomMLFlowLogger type-hint cleanup in custom_loggers.py.
neural_lam/create_graph.py
Outdated
|
|
||
| import torch | ||
|
|
||
| degrees = torch.bincount(edge_index[1], minlength=num_nodes) |
There was a problem hiding this comment.
These stats are using only in-degree (edge_index[1]), but the log messages say "degree" and "isolated nodes" as if they describe the whole graph. For directed graphs like g2m, source-only nodes get counted as isolated even when they have outgoing edges, so the debug output is misleading.
| if edge_index.min() < 0: | ||
| raise ValueError(f"[{name}] found negative node indices") | ||
|
|
||
| if edge_index.max() >= num_nodes: |
There was a problem hiding this comment.
This check is incompatible with the hierarchical m2m graphs below, because from_networkx_with_start_index() intentionally keeps globally offset node ids. For level 1+ edge_index.max() can be much larger than num_nodes even though the graph is valid, so this now breaks hierarchical graph generation.
There was a problem hiding this comment.
Thanks for the feedback I've updated the stats to use total degree (in + out), and added separate logging for in-degree and out-degree. Also added a small guard for empty graphs to avoid edge-case issues. Let me know if this looks good!
2 participants
Describe your changes
This PR introduces lightweight graph validation and diagnostic utilities to improve debugging during graph construction.
It adds two helper functions:
validate_graphPerforms sanity checks on graph structure (shape, empty edges, invalid indices) and fails early with clear error messages.
compute_graph_statsLogs useful statistics about the graph, including number of nodes, edges, degree distribution, and isolated nodes.
These utilities are integrated into the graph creation pipeline for:
g2m)m2g)m2m, per level)Motivation and context
While working with graph construction, it can be difficult to quickly verify whether a generated graph is valid or understand its structure.
This change helps by:
Dependencies
No new dependencies are introduced.
Issue Link
closes #518
Type of change
Checklist before requesting a review
Checklist for reviewers
Author checklist after completed review
Checklist for assignee