[Feature] Reduce messages with scatter_add in PyTorch #427
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
6 tasks
jermainewang
reviewed
Mar 2, 2019
zheng-da
pushed a commit
to zheng-da/dgl
that referenced
this pull request
Mar 6, 2019
* implement pytorch spmm with gather and scatter add * fix * replace torch take with index_select * comments * comment about pytorch __getitem__ operator pitfall * typo
zheng-da
pushed a commit
to zheng-da/dgl
that referenced
this pull request
Mar 6, 2019
* implement pytorch spmm with gather and scatter add * fix * replace torch take with index_select * comments * comment about pytorch __getitem__ operator pitfall * typo
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Outdated suggestions cannot be applied.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.
Description
Spmm with uncoalesced sparse matrix in PyTorch has bad performance because the coalesce step is costly. Although coalesce sparse matrix ahead of time could make the forward pass much faster, the backward phase is still slow. Given expression y=spmat * x, in order to calculate gradient of x, we need to do spmat^T * grad_y. But transpose sparse matrix and then coalesce is again very costly.
This PR replaces PyTorch spmm with index_select and scatter_add. This solution is motivated by pytorch_geometric.
Checklist
Please feel free to remove inapplicable items for your PR.
or have been fixed to be compatible with this change
Changes