[ONNX] Add a post-pass for If folding #49410

KsenijaS · 2020-12-15T17:49:58Z

Currently ONNX Runtime is doing Shape and Type Inference on both branches of the If operator, regardless of which branch is executing in runtime. This can cause Runtime errors in some cases:

Condition of the If node is based on shape / size of the input
then and else branch have different return types

This pass is folding an If node and it enables following tests:

test_embedding_sequential_1
test_embedding_sequential_2
test_nllloss
test_crossentropy
test_embedding_bag_1d_per_sample_weights
test_embedding_bag_2d_per_sample_weights

facebook-github-bot · 2020-12-15T17:50:22Z

💊 CI failures summary and remediations

As of commit c652fee (more details on the Dr. CI page):

💚 💚 Looks good so far! There are no failures yet. 💚 💚

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions to the (internal) Dr. CI Users group.

This comment has been revised 101 times.

…f_fold

torch/csrc/jit/passes/onnx/shape_type_inference.cpp

torch/csrc/jit/passes/onnx/fold_if_node.h

torch/csrc/jit/passes/onnx/fold_if_node.cpp

…f_fold

torch/csrc/jit/passes/onnx/shape_type_inference.cpp

torch/csrc/jit/passes/onnx/scalar_type_analysis.cpp

torch/csrc/jit/passes/onnx/fold_if_node.cpp

torch/csrc/jit/passes/onnx/scalar_type_analysis.h

BowenBao

LGTM, minor suggestions.

torch/csrc/jit/passes/onnx/fold_if_node.cpp

torch/csrc/jit/passes/onnx/shape_type_inference.cpp

jiafatom

LGTM!

…f_fold

* add pass * add tests * enable tests * update dtype symbolic * update pass * clang format * update utils file * fix clang_tidy errors * update pass * update pass * clang format * update pass * update pass * update pass * update pass * update pass * update pass * update pass * update pass * fix mypy tests * update comment * update pass * add new line * update pass * empty commit * empty commit * fix merge conflict

Summary: [ONNX] ONNX dev branch merge 01-06-2021 - [ONNX] Support onnx if/loop sequence output in opset 13 - (#49270) - Symbolic function for torch.square (#49446) - [ONNX] Add checks in ONNXSetDynamicInputShape (#49783) … - [ONNX] Enable export af aten::__derive_index (#49514) … - [ONNX] Update symbolic for unfold (#49378) … - [ONNX] Update the sequence of initializers in exported graph so that it is as same as inputs. (#49798) - [ONNX] Enable opset 13 ops (#49612) … - [ONNX] Improve error message for supported model input types in ONNX export API. (#50119) - [ONNX] Add a post-pass for If folding (#49410) Pull Request resolved: #50163 Reviewed By: pbelevich Differential Revision: D25821059 Pulled By: SplitInfinity fbshipit-source-id: 9f511a93d9d5812d0ab0a49d61ed0fa5f8066948

…50582) Fixing pytorch/vision#3251 (PR #49410 triggers the torch vision test build failure, on three tests test_faster_rcnn, test_mask_rcnn, test_keypoint_rcnn. ) The offending PR is fine on pytorch UT, because the torchvision and pytorch test has a gap when we merge them - we are using different test API on two sides, therefore causing some discrepancy. This PR bridge the gap for the above three tests, and disable _jit_pass_onnx_fold_if pass until it gets fixed. Allow _jit_pass_onnx_fold_if only when dynamic_axes is None.

…50582) Fixing pytorch/vision#3251 (PR #49410 triggers the torch vision test build failure, on three tests test_faster_rcnn, test_mask_rcnn, test_keypoint_rcnn. ) The offending PR is fine on pytorch UT, because the torchvision and pytorch test has a gap when we merge them - we are using different test API on two sides, therefore causing some discrepancy. This PR bridge the gap for the above three tests, and disable _jit_pass_onnx_fold_if pass until it gets fixed. Allow _jit_pass_onnx_fold_if only when dynamic_axes is None. [ghstack-poisoned]

…s is None (#50582)" Fixing pytorch/vision#3251 (PR #49410 triggers the torch vision test build failure, on three tests test_faster_rcnn, test_mask_rcnn, test_keypoint_rcnn. ) The offending PR is fine on pytorch UT, because the torchvision and pytorch test has a gap when we merge them - we are using different test API on two sides, therefore causing some discrepancy. This PR bridge the gap for the above three tests, and disable _jit_pass_onnx_fold_if pass until it gets fixed. Allow _jit_pass_onnx_fold_if only when dynamic_axes is None. [ghstack-poisoned]

…s is None (#50582)" Fixing pytorch/vision#3251 (PR #49410 triggers the torch vision test build failure, on three tests test_faster_rcnn, test_mask_rcnn, test_keypoint_rcnn. ) The offending PR is fine on pytorch UT, because the torchvision and pytorch test has a gap when we merge them - we are using different test API on two sides, therefore causing some discrepancy. This PR bridge the gap for the above three tests, and disable _jit_pass_onnx_fold_if pass until it gets fixed. Allow _jit_pass_onnx_fold_if only when dynamic_axes is None. Differential Revision: [D26023934](https://our.internmc.facebook.com/intern/diff/D26023934) [ghstack-poisoned]

…s is None (#50582)" Fixing pytorch/vision#3251 (PR #49410 triggers the torch vision test build failure, on three tests test_faster_rcnn, test_mask_rcnn, test_keypoint_rcnn. ) The offending PR is fine on pytorch UT, because the torchvision and pytorch test has a gap when we merge them - we are using different test API on two sides, therefore causing some discrepancy. This PR bridge the gap for the above three tests, and disable _jit_pass_onnx_fold_if pass until it gets fixed. Allow _jit_pass_onnx_fold_if only when dynamic_axes is None. [ghstack-poisoned]

…s is None (#50582)" Fixing pytorch/vision#3251 (PR #49410 triggers the torch vision test build failure, on three tests test_faster_rcnn, test_mask_rcnn, test_keypoint_rcnn. ) The offending PR is fine on pytorch UT, because the torchvision and pytorch test has a gap when we merge them - we are using different test API on two sides, therefore causing some discrepancy. This PR bridge the gap for the above three tests, and disable _jit_pass_onnx_fold_if pass until it gets fixed. Allow _jit_pass_onnx_fold_if only when dynamic_axes is None. Differential Revision: [D26050886](https://our.internmc.facebook.com/intern/diff/D26050886) [ghstack-poisoned]

…50582) (#50910) Summary: Pull Request resolved: #50910 Fixing pytorch/vision#3251 (PR #49410 triggers the torch vision test build failure, on three tests test_faster_rcnn, test_mask_rcnn, test_keypoint_rcnn. ) The offending PR is fine on pytorch UT, because the torchvision and pytorch test has a gap when we merge them - we are using different test API on two sides, therefore causing some discrepancy. This PR bridge the gap for the above three tests, and disable _jit_pass_onnx_fold_if pass until it gets fixed. Allow _jit_pass_onnx_fold_if only when dynamic_axes is None. Test Plan: Imported from OSS Reviewed By: pbelevich Differential Revision: D26050886 Pulled By: SplitInfinity fbshipit-source-id: b765ffe30914261866dcc761f0d0999fd16169e3

…ytorch#50582) Fixing pytorch/vision#3251 (PR pytorch#49410 triggers the torch vision test build failure, on three tests test_faster_rcnn, test_mask_rcnn, test_keypoint_rcnn. ) The offending PR is fine on pytorch UT, because the torchvision and pytorch test has a gap when we merge them - we are using different test API on two sides, therefore causing some discrepancy. This PR bridge the gap for the above three tests, and disable _jit_pass_onnx_fold_if pass until it gets fixed. Allow _jit_pass_onnx_fold_if only when dynamic_axes is None. ghstack-source-id: 53b04161a00c3f7ae959dada3780be360e3071b7 Pull Request resolved: pytorch#50910

KsenijaS added 18 commits December 9, 2020 21:49

add pass

26760a8

add tests

2a0383f

enable tests

93eac3e

Merge https://github.com/pytorch/pytorch into if_fold

4da0ff3

update dtype symbolic

620cdb1

update pass

8ae4093

clang format

4a91187

update utils file

aab9944

fix clang_tidy errors

f891155

update pass

ae9e93c

update pass

dde7cc9

clang format

8770782

update pass

b7cfda8

Merge https://github.com/pytorch/pytorch into if_fold

469aff8

update pass

34eb993

update pass

7358893

update pass

3d8f513

update pass

e16bb5a

facebook-github-bot added the cla signed label Dec 15, 2020

facebook-github-bot added the oncall: jit Add this issue/PR to JIT oncall triage queue label Dec 15, 2020

pytorchbot added the open source label Dec 15, 2020

KsenijaS added 2 commits December 15, 2020 10:33

Merge branch 'onnx_ms_1' of https://github.com/pytorch/pytorch into i…

f392f86

…f_fold

update pass

51b4614

BowenBao reviewed Dec 16, 2020

View reviewed changes

KsenijaS added 4 commits December 15, 2020 17:00

Merge branch 'onnx_ms_1' of https://github.com/pytorch/pytorch into i…

d0b0640

…f_fold

update pass

283ee2f

update pass

9e189d4

fix mypy tests

95a4458

jiafatom reviewed Dec 16, 2020

View reviewed changes

BowenBao reviewed Dec 17, 2020

View reviewed changes

torch/csrc/jit/passes/onnx/fold_if_node.cpp Outdated Show resolved Hide resolved

BowenBao reviewed Dec 17, 2020

View reviewed changes

torch/csrc/jit/passes/onnx/scalar_type_analysis.h Outdated Show resolved Hide resolved

KsenijaS added 2 commits December 16, 2020 17:55

update pass

7b71e0e

add new line

e8fc7aa

BowenBao approved these changes Dec 17, 2020

View reviewed changes

torch/csrc/jit/passes/onnx/fold_if_node.cpp Outdated Show resolved Hide resolved

torch/csrc/jit/passes/onnx/shape_type_inference.cpp Outdated Show resolved Hide resolved

update pass

ae24547

jiafatom approved these changes Dec 17, 2020

View reviewed changes

BowenBao force-pushed the onnx_ms_1 branch from ab060a1 to 1baebbb Compare January 4, 2021 18:29

KsenijaS added 5 commits January 4, 2021 11:19

Merge branch 'onnx_ms_1' of https://github.com/pytorch/pytorch into i…

303a5f3

…f_fold

empty commit

a822e55

empty commit

59f9d5d

fix merge conflict

569e7a4

fix merge conflict

c652fee

BowenBao merged commit 33f26ed into pytorch:onnx_ms_1 Jan 6, 2021

spandantiwari mentioned this pull request Jan 12, 2021

[ONNX] ONNX dev branch merge 01-06-2021 #50163

Closed

This was referenced Jan 15, 2021

ONNX tests failing on master pytorch/vision#3251

Closed

[ONNX] Enable _jit_pass_onnx_fold_if only when dynamic_axes is None #50582

Merged

BowenBao mentioned this pull request Jan 21, 2021

[ONNX] Enable _jit_pass_onnx_fold_if only when dynamic_axes is None (#50582) #50910

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[ONNX] Add a post-pass for If folding #49410

[ONNX] Add a post-pass for If folding #49410

KsenijaS commented Dec 15, 2020 •

edited

facebook-github-bot commented Dec 15, 2020 •

edited

BowenBao left a comment

jiafatom left a comment

[ONNX] Add a post-pass for If folding #49410

[ONNX] Add a post-pass for If folding #49410

Conversation

KsenijaS commented Dec 15, 2020 • edited

facebook-github-bot commented Dec 15, 2020 • edited

💊 CI failures summary and remediations

BowenBao left a comment

Choose a reason for hiding this comment

jiafatom left a comment

Choose a reason for hiding this comment

KsenijaS commented Dec 15, 2020 •

edited

facebook-github-bot commented Dec 15, 2020 •

edited