[ONNX][Relay] Add dynamic unsqueeze / expand_dims op #9039
Merged
Conversation
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
AndrewZhaoLuo
requested review from
Huyuwei,
MarisaKirisame,
ZihengJiang,
anijain2305,
areusch,
comaniac,
jroesch,
junrushao,
jwfromm,
kazum,
mbrookhart,
merrymercy,
siju-samuel,
slyubomirsky,
srkreddy1238,
tqchen,
vinx13,
wweic,
yzhliu and
zhiics
as code owners
* main: [3/10] Moved TIR generation from Python to C++ for CMSIS-NN (apache#8951) Support match pvar with dtype constraint (apache#9016)
mbrookhart
requested changes
Sep 20, 2021
Contributor
mbrookhart
left a comment
mbrookhart
left a comment
There was a problem hiding this comment.
I'm not sure I get the multiple back to back expand_dims calls.
Comment on lines
5019
to
5021
| "test_unsqueeze_three_axes", | ||
| "test_unsqueeze_two_axes", | ||
| "test_unsqueeze_unsorted_axes", |
Contributor
There was a problem hiding this comment.
Can you add these to device-specific skips below?
Contributor
Author
There was a problem hiding this comment.
Hmm actually all targets are really slow it seems. The LLVM target takes > 3 minutes for test_unsqueeze_two_axes.
Comment on lines
113
to
123
| if isinstance(axis, int): | ||
| return _make.expand_dims(data, axis, num_newaxis) | ||
| if isinstance(axis, Expr): | ||
| # TODO (AndrewZhaoLuo): investigate performance issues with consecutive | ||
| # dynamic expand_dims on non-llvm targets. | ||
| for _ in range(num_newaxis): | ||
| # Dynamic rank is not well supported so we can only increase rank | ||
| # by a static amount (e.g. 1) so we have to do this | ||
| data = _dyn_make.expand_dims(data, axis) | ||
| return data | ||
| raise ValueError(f"Unknown type for axis: {type(axis)}") |
Contributor
There was a problem hiding this comment.
Why can't we do it all at once if we know that num_newaxis is static?
Contributor
Author
There was a problem hiding this comment.
We can, done.
Comment on lines
+1474
to
+1484
| # dynamic expand_dims on non-llvm targets. | ||
| for i in range(num_new_axis): | ||
| axis = relay.TupleGetItem(axes, i) | ||
| # Unpack scalar | ||
| axis = relay.reshape(axis, []) | ||
| axis = relay.If( | ||
| axis >= relay.const(0, "int64"), axis, axis + relay.const(rank_input, "int64") | ||
| ) | ||
| result = _op.expand_dims(result, axis) | ||
| return result | ||
|
|
Contributor
There was a problem hiding this comment.
Again, I think this should be doable as one call?
Contributor
Author
There was a problem hiding this comment.
This might be doable if we change the interface to accept a list of sorted axis but it'll be quite a bit more complicated.
Contributor
Author
|
PTAL @mbrookhart |
* main: Add back missing __init__.py to unbreak CI. (apache#9052) [Meta Schedule][M3b] Builder (apache#9044)
mbrookhart
approved these changes
Sep 21, 2021
ylc
pushed a commit
to ylc/tvm
that referenced
this pull request
Sep 29, 2021
* initial dyn unsqueeze example * simplify, properly unpack scalar * basic tests * squish bugs -- assign proper types * working topi * fix things * temp work * fix casting to int64 * shape encoding method for axis * working shape encoding metric * add comment * move to non-rank encoded axis * failing regime * fix * it works! * add test * add comment on shape func * remove unused topi * undo some file changes * more cleanup * newline * clean up * clean up * enable multiple axis tests * move tests to dynamic op * Update docs * add converter * initial dyn unsqueeze example * simplify, properly unpack scalar * basic tests * squish bugs -- assign proper types * working topi * fix things * temp work * fix casting to int64 * shape encoding method for axis * working shape encoding metric * add comment * move to non-rank encoded axis * failing regime * fix * it works! * add test * add comment on shape func * remove unused topi * undo some file changes * more cleanup * newline * clean up * clean up * enable multiple axis tests * move tests to dynamic op * Update docs * add converter * working tests * add test, remove unneeded file * fix things * more lint * more lint * pick things * disable opencl tests * unsqueeze tests * clean up * dyn stuff * add num_newaxis Co-authored-by: Andrew Zhao Luo <andrewzhaoluo@system76-pc.localdomain>
ylc
pushed a commit
to ylc/tvm
that referenced
this pull request
Jan 13, 2022
* initial dyn unsqueeze example * simplify, properly unpack scalar * basic tests * squish bugs -- assign proper types * working topi * fix things * temp work * fix casting to int64 * shape encoding method for axis * working shape encoding metric * add comment * move to non-rank encoded axis * failing regime * fix * it works! * add test * add comment on shape func * remove unused topi * undo some file changes * more cleanup * newline * clean up * clean up * enable multiple axis tests * move tests to dynamic op * Update docs * add converter * initial dyn unsqueeze example * simplify, properly unpack scalar * basic tests * squish bugs -- assign proper types * working topi * fix things * temp work * fix casting to int64 * shape encoding method for axis * working shape encoding metric * add comment * move to non-rank encoded axis * failing regime * fix * it works! * add test * add comment on shape func * remove unused topi * undo some file changes * more cleanup * newline * clean up * clean up * enable multiple axis tests * move tests to dynamic op * Update docs * add converter * working tests * add test, remove unneeded file * fix things * more lint * more lint * pick things * disable opencl tests * unsqueeze tests * clean up * dyn stuff * add num_newaxis Co-authored-by: Andrew Zhao Luo <andrewzhaoluo@system76-pc.localdomain>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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.
Adds new dynamic unsqueeze op. There is an issue with performance however. On targets
metalandopenclwhich I tested the ONNX and unit tests take multiple hours to run. For now, I've disabled offending tests on such targets. The performance issues seem to happen when we have multiple consecutive dynamic reshape ops.