Skip to content

Conversation

jjsjann123
Copy link
Collaborator

This reverts the 9a9bb44

Fixing the broken case which reverts the previous commit.
details about fix:
modified: aten/src/ATen/native/Convolution.cpp

called contiguous on 3D input tensor. This avoids the code path to accidentally
recognize the input as channel_last stride, due to unsqueezing of permuted 3d
tensor.

This reverts the 9a9bb44

Fixing the broken case which reverts the previous commit.
details about fix:
	modified:   aten/src/ATen/native/Convolution.cpp

called contiguous on 3D input tensor. This avoids the code path to accidentally
recognize the input as channel_last stride, due to unsqueezing of permuted 3d
tensor.
@jjsjann123 jjsjann123 requested a review from apaszke as a code owner November 7, 2019 05:20
@facebook-github-bot facebook-github-bot added the oncall: jit Add this issue/PR to JIT oncall triage queue label Nov 7, 2019
@jjsjann123
Copy link
Collaborator Author

The 3 line fix is here: https://github.com/pytorch/pytorch/pull/29361/files#diff-57ac615408468d3c7a461e505581bea3R568-R571

The previous failing case is actually still a legit nhwc case in cudnn. It's failing some trivial dimension checks. I'm working with cudnn team on fixing that.
Nevertheless it's undesired behavior due to my oversight earlier. The proper behavior would be to not accidentally use NHWC on 3d tensor (until we extend channel_last support to arbitrary dimension tensors)

Copy link
Contributor

@facebook-github-bot facebook-github-bot left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

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

zdevito pushed a commit to zdevito/ATen that referenced this pull request Nov 7, 2019
Summary:
This reverts the 9a9bb448ee49a1493f22bbbeed4af92b1364fce9

Fixing the broken case which reverts the previous commit.
details about fix:
	modified:   aten/src/ATen/native/Convolution.cpp

called contiguous on 3D input tensor. This avoids the code path to accidentally
recognize the input as channel_last stride, due to unsqueezing of permuted 3d
tensor.
Pull Request resolved: pytorch/pytorch#29361

Differential Revision: D18371964

Pulled By: VitalyFedyunin

fbshipit-source-id: a5985f4687b37e183649fa35b8ccdb50368ebfdf
@facebook-github-bot
Copy link
Contributor

@VitalyFedyunin merged this pull request in fdab1cf.

VitalyFedyunin added a commit to VitalyFedyunin/pytorch that referenced this pull request Nov 18, 2019
This PR was bisected as the reason of HRNet becomeing unstable. We are keeping inversigation, meanwhile I will revert it.

This reverts commit fdab1cf.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Merged oncall: jit Add this issue/PR to JIT oncall triage queue

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants