Uh oh!
There was an error while loading. Please reload this page.
[ONNX] Make the ONNX Importer More Static - #7429
Merged
Merged
Conversation
mbrookhartforce-pushed
the
more_static_onnx_importer
branch
from
February 9, 2021 21:25
9303222 to
52f09eeComparembrookhartforce-pushed
the
more_static_onnx_importer
branch
from
February 10, 2021 20:40
70894a1 to
6e7b71cCompareelectriclilies
approved these changes
Feb 11, 2021
electriclilies
left a comment
Contributor
There was a problem hiding this comment.
Overall, looks good to me -- I just left one comment about code style.
Uh oh!
There was an error while loading. Please reload this page.
mbrookhartforce-pushed
the
more_static_onnx_importer
branch
from
February 12, 2021 18:03
6e7b71c to
9c2c5a6Comparembrookhart
commented
Feb 12, 2021
ContributorAuthor
@masahi@jwfromm@tmoreau89 I think this is ready for review |
junrushao
approved these changes
Feb 12, 2021
junrushao
left a comment
Member
There was a problem hiding this comment.
Really nice work @mbrookhart!!
tmoreau89
commented
Feb 13, 2021
Contributor
Thank you @mbrookhart@electriclilies@junrushao1994 the PR has been merged |
Lokiiiiii pushed a commit
to Lokiiiiii/tvm
that referenced
this pull request
Mar 2, 2021
* Construct static Ops if inputs are Constant * Expose FoldConstant as a function in addition to the pass * refactor onnx importer to do more static imports by constant folding fix pylint * fix test regressions * fix style, two bugs * pipe freeze_params through sub_graphs when importing loops and control flow
trevor-m pushed a commit
to neo-ai/tvm
that referenced
this pull request
Mar 2, 2021
* Construct static Ops if inputs are Constant * Expose FoldConstant as a function in addition to the pass * refactor onnx importer to do more static imports by constant folding fix pylint * fix test regressions * fix style, two bugs * pipe freeze_params through sub_graphs when importing loops and control flow
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 freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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.
We've been hitting a lot of issues in more complicated ONNX models where results of something like dynamic strided slice feed into an op like dynamic reshape, causing dynamic ranks to appear and crash the importer. To fix this, I decided it might be good to try to progagate constant information through the importer during import instead of trying to retroactively remove dynamic ops in a pass. To do that, I changed the freeze_params behavior of the importer to create constants before import, and made the importer fold constants through the importer as they are seen. Additionally, I edited the ops to only construct dynamic ops if the inputs are not constant.
In sum, this ends up creating many fewer dynamic ops in the ONNX importer, and so we see many fewer dynamic rank issues when importing models.
cc @masahi@jwfromm@electriclilies
@tmoreau89 found a serious performance regression on #7368, this change takes the import + dynamic_to_static of ONNX Bert Large from 20+ minutes to < 30s.