[Torch] Add aten::roll support for Swin Transformer #9371
Merged
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Outdated suggestions cannot be applied.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.
Closes #9368
This PR adds missing converters necessary for importing the recent state of the art model Swin-Transoformer.
In particular,
aten::roll
is an interesting op to implement and generally useful op to have. It can be implemented viagather
but the encoding is not obvious. Here are the references:https://pytorch.org/docs/stable/generated/torch.roll.html
https://numpy.org/doc/stable/reference/generated/numpy.roll.html
please review @comaniac @jcf94 @junrushao1994
@Kyrie-Zhao
aten::rand
is not needed, it is used in dropout but if you trace the model witheval
mode, dropout is gone.Now the following script works with the error
1.4081733e-07
.