-
Notifications
You must be signed in to change notification settings - Fork 22.2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Inconsistent gradient from different backend of CTCLoss #26797
Labels
module: cudnn
Related to torch.backends.cudnn, and CuDNN support
module: loss
Problem is related to loss function
triaged
This issue has been looked at a team member, and triaged and prioritized into an appropriate module
Comments
Alexander-H-Liu
changed the title
Inconsistent gradient from the different backend of CTCLoss
Inconsistent gradient from different backend of CTCLoss
Sep 25, 2019
I think this and #25833 are the same, the others I'm not so sure about. |
vincentqb
added
module: cudnn
Related to torch.backends.cudnn, and CuDNN support
and removed
triage review
labels
Sep 30, 2019
ailzhang
added
high priority
triaged
This issue has been looked at a team member, and triaged and prioritized into an appropriate module
and removed
high priority
labels
Oct 1, 2019
zdevito
pushed a commit
to zdevito/ATen
that referenced
this issue
Oct 15, 2019
Summary: Using grad_out for CuDNN CTC loss fixes: pytorch/pytorch#26797, pytorch/pytorch#25833. We also fix a cudnn incompatible change that surfaced during the testing: As of CuDNN 7.6 the semantics of the CTC loss gradients are different. This leads us to disable CuDNN CTC for CuDNN < 7.6. To mitigate the impact on users, we convert the parameters for the native implementation if CuDNN isn't applicable (previously this would give an error.) Pull Request resolved: pytorch/pytorch#27039 Differential Revision: D17910815 Pulled By: ngimel fbshipit-source-id: 465b33612d3402f10c355aa7026a7e1ffaef3073
thiagocrepaldi
pushed a commit
to thiagocrepaldi/pytorch
that referenced
this issue
Feb 4, 2020
Summary: Using grad_out for CuDNN CTC loss fixes: pytorch#26797, pytorch#25833. We also fix a cudnn incompatible change that surfaced during the testing: As of CuDNN 7.6 the semantics of the CTC loss gradients are different. This leads us to disable CuDNN CTC for CuDNN < 7.6. To mitigate the impact on users, we convert the parameters for the native implementation if CuDNN isn't applicable (previously this would give an error.) Pull Request resolved: pytorch#27039 Differential Revision: D17910815 Pulled By: ngimel fbshipit-source-id: 465b33612d3402f10c355aa7026a7e1ffaef3073
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Labels
module: cudnn
Related to torch.backends.cudnn, and CuDNN support
module: loss
Problem is related to loss function
triaged
This issue has been looked at a team member, and triaged and prioritized into an appropriate module
🐛 Bug
I'm switching to the cudnn backend of CTCLoss since the other is not fully reproducible,
however, it turns out that the exact same model that used to work with pytorch's cuda backend ctc loss now failed.
With some simple example, I found that there's a huge difference in gradient (both direction and magnitude) between two backends.
I'm not sure if the bug is on pytorch or cudnn, but as far as I know, TensorFlow also used CTC from cudnn and there is no similar issue.
Thanks in advance.
To Reproduce
Behavior
Environment
cc @ezyang @gchanan @zou3519
The text was updated successfully, but these errors were encountered: