[ONNX] export group_norm #27071

neginraoof · 2019-09-30T17:19:36Z

Updated group_norm symbolic

test/onnx/test_operators.py

torch/onnx/symbolic_opset9.py

BowenBao · 2019-09-30T18:28:38Z

torch/onnx/symbolic_opset9.py

+@parse_args('v', 'i', 'v', 'v', 'f', 'i')
+def group_norm(g, input, num_groups, weight, bias, eps, cudnn_enabled):
+    input_sizes = input.type().sizes()
+    shape = [1, input_sizes[0] * num_groups, input_sizes[1] / num_groups, -1]


In fact it might be possible to remove the dependency on explicit input_sizes.

For the first reshape, reshape to [0, num_groups, -1], the effect will be [batch, channel, h, w] => [batch, num_groups, channel / num_groups * h * w].

For weight and bias, consider the below comment regarding broadcasting.

For the second reshape, just reshape back to the original input shape: g.op('Reshape', norm_reshaped, g.op("Shape", input)).

For the last weight and bias, it seems we still need the rank information of input. We can remove this dependency as well if we add two more reshapes to the norm tensor, but that might not worth the effort as rank info is usually static.

The motivation is that we won't be able to support dynamic input sizes if we depend on static input_sizes.

Thanks @BowenBao, apparently broadcasting is not supported for InstanceNormalization.
So I don't think we can avoid accessing the size for weight_ and bias_

Ok, but I think we can still construct weight_ and bias_ with just num_groups, if in the first reshape we keep batch and channel axes separated.

True, I gonna update the first two parts anyways. But I agree that it might not be worth it to add two reshapes since rank here could be N (it's not limited in spec). I'll leave a note in case anyone needed this op with dynamic rank in future.

neginraoof · 2019-10-03T03:38:17Z

@pytorchbot rebase this please

BowenBao

please always remember to update test operator expect files if symbolic fn is updated.

BowenBao · 2019-10-03T22:36:04Z

torch/onnx/symbolic_opset9.py

+
+@parse_args('v', 'i', 'v', 'v', 'f', 'i')
+def group_norm(g, input, num_groups, weight, bias, eps, cudnn_enabled):
+    if input.isCompleteTensor():


isCompleteTensor can be relaxed. We only require its scalar type and rank to be known.

Sure. I'm on console access to the environment and build is slow, so I updated the files here.

test/onnx/test_pytorch_onnx_caffe2.py

facebook-github-bot

@houseroad has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

neginraoof · 2019-10-07T23:33:09Z

cc @houseroad tests are fixed. Please re-import. Thanks.

neginraoof · 2019-10-08T16:26:38Z

cc @houseroad for review

neginraoof · 2019-10-09T18:15:58Z

cc @houseroad for review

neginraoof · 2019-10-10T06:58:04Z

@houseroad please review updates.

neginraoof · 2019-10-10T16:41:13Z

@pytorchbot retest this please

neginraoof · 2019-10-11T16:27:34Z

cc @houseroad for review

neginraoof · 2019-10-14T15:05:05Z

@houseroad for review

neginraoof · 2019-10-15T01:01:56Z

@pytorchbot rebase this please

…oof/groupNorm

neginraoof · 2019-10-22T16:24:15Z

@houseroad please review the updates

facebook-github-bot

@houseroad has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

houseroad

This broke some internal use cases, I need further investigation.

neginraoof · 2019-10-22T20:50:35Z

Thanks @houseroad for looking into this. Is there a failure regarding group_norm?

houseroad

Let's check in this for now. However, this translation is really inefficient. We should introduce GroupNorm to ONNX, so we can use one op to represent the model.

houseroad · 2019-10-23T18:00:15Z

@pytorchbot rebase this please

facebook-github-bot · 2019-10-23T23:42:12Z

@houseroad merged this pull request in 76d262d.

Summary: Updated group_norm symbolic Pull Request resolved: pytorch#27071 Reviewed By: hl475 Differential Revision: D17792249 Pulled By: houseroad fbshipit-source-id: 08be6071952ca2c256d2c6a0a6bbc19a8442f1fe

group norm export

b91b51a

pytorchbot added the module: onnx Related to torch.onnx label Sep 30, 2019

clean up

1142be0

ezyang added the open source label Sep 30, 2019

spandantiwari mentioned this pull request Sep 30, 2019

torch.nn.GroupNorm is not supported in ONNX export #26753

Closed

BowenBao reviewed Sep 30, 2019

View reviewed changes

test/onnx/test_operators.py Outdated Show resolved Hide resolved

spandantiwari mentioned this pull request Sep 30, 2019

Using Experimental operators (such as ATen) over C/C++ API. microsoft/onnxruntime#1645

Closed

BowenBao reviewed Sep 30, 2019

View reviewed changes

torch/onnx/symbolic_opset9.py Show resolved Hide resolved

neginraoof added 2 commits September 30, 2019 11:17

fix for input types

e26c264

fixed expect tests

e61e448

BowenBao reviewed Sep 30, 2019

View reviewed changes

neginraoof and others added 6 commits September 30, 2019 13:01

fix for comments

70bcac5

fix for comments

51c0c23

clean up

083005c

skip if embed for instance norm

50abb7c

fixing for comments

3bb899b

fixed expect files

be9799e

pytorchbot and others added 3 commits October 3, 2019 03:38

Merge remote-tracking branch 'origin/master' into HEAD

1ab0292

fixed expect files

cd291cc

fixed expect files

7e2a5c3

BowenBao reviewed Oct 3, 2019

View reviewed changes

neginraoof added 3 commits October 3, 2019 23:17

Merge branch 'master' into neraoof/groupNorm

9a6be15

Merge branch 'master' into neraoof/groupNorm

c25e393

fix for comments

227290d

facebook-github-bot reviewed Oct 7, 2019

View reviewed changes

Update TestOperators.test_groupnorm_noaffine.expect

aea1405

neginraoof added 4 commits October 9, 2019 23:29

added aten fallback

14941e0

flake8

e4e6cbd

fix for comments

5cbd451

fix for comments

ce1fccb

flake8 fix

e22c516

neginraoof added 4 commits October 17, 2019 12:17

Merge branch 'master' of https://github.com/pytorch/pytorch into nera…

70b99e8

…oof/groupNorm

added more tests for group_norm

99e836a

removed expected operators tests

2092933

added test cases

5cd70c1

facebook-github-bot reviewed Oct 22, 2019

View reviewed changes

houseroad requested a review from BowenBao October 22, 2019 17:25

houseroad requested changes Oct 22, 2019

View reviewed changes

houseroad approved these changes Oct 23, 2019

View reviewed changes

facebook-github-bot closed this in 76d262d Oct 23, 2019

facebook-github-bot added the merged label Oct 23, 2019

mruberry added the Merged label Oct 28, 2020

garymm mentioned this pull request Aug 25, 2021

Onnx export does not support Group Normalization. #14315

Closed

jcwchen mentioned this pull request Mar 24, 2022

ONNX support LayerNorm and GroupNorm onnx/onnx#4085

Closed

[ONNX] export group_norm #27071

[ONNX] export group_norm #27071

Uh oh!

Conversation

neginraoof commented Sep 30, 2019

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

BowenBao Sep 30, 2019

Choose a reason for hiding this comment

Uh oh!

neginraoof Sep 30, 2019

Choose a reason for hiding this comment

Uh oh!

BowenBao Sep 30, 2019

Choose a reason for hiding this comment

Uh oh!

neginraoof Oct 1, 2019

Choose a reason for hiding this comment

Uh oh!

neginraoof commented Oct 3, 2019

Uh oh!

BowenBao left a comment

Choose a reason for hiding this comment

Uh oh!

BowenBao Oct 3, 2019

Choose a reason for hiding this comment

Uh oh!

neginraoof Oct 4, 2019

Choose a reason for hiding this comment

Uh oh!

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

neginraoof commented Oct 7, 2019

Uh oh!

neginraoof commented Oct 8, 2019

Uh oh!

neginraoof commented Oct 9, 2019

Uh oh!

neginraoof commented Oct 10, 2019

Uh oh!

neginraoof commented Oct 10, 2019

Uh oh!

neginraoof commented Oct 11, 2019

Uh oh!

neginraoof commented Oct 14, 2019

Uh oh!

neginraoof commented Oct 15, 2019

Uh oh!

neginraoof commented Oct 22, 2019

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

houseroad left a comment

Choose a reason for hiding this comment

Uh oh!

neginraoof commented Oct 22, 2019

Uh oh!

houseroad left a comment

Choose a reason for hiding this comment

Uh oh!

houseroad commented Oct 23, 2019

Uh oh!

facebook-github-bot commented Oct 23, 2019

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants