Fix atom self-image detection bug in Torch-based graph construction#164
Open
PascalSalzbrenner wants to merge 1 commit into
Open
Fix atom self-image detection bug in Torch-based graph construction#164PascalSalzbrenner wants to merge 1 commit into
PascalSalzbrenner wants to merge 1 commit into
Conversation
…e, resulting in failed self-interaction testing and causing the model to return NaN predictions
Author
|
@microsoft-github-policy-service agree company="Microsoft" |
Author
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Diagnosis: On the GPU / Torch / TorchSim path, torch.cdist sometimes returns the distance of an atom with itself (in the same unit cell, not with periodic images) incorrectly as ~0.011 instead of exactly 0. This happens for atoms far from the origin in large (>20 A) unit cells due to an accumulation of errors resulting from the float32 precision. Such an atom would incorrectly be marked as
Falsein theis_self_looptest in line 250 ofradius_graph_pbc.py. Then, when m3gnet computes the distances exactly, we get 0/0 divisions in them3gnet.pygraph construction. Mattersim then returnsNaNfor the structure's energy, forces, and stresses.Fix: The simple fix is to base the
is_self_looptest on quantities we know exactly (same index + offset == 0, ie same cell). I made this fix and added tests against regression as well as verifying periodic images are picked up correctly.