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
[SR-14218] [AutoDiff] Gradients not propagating out of member function #55999
Comments
Ping @rxwei |
I'll take a look some time later. I'll need to convert all these AD bugs starting with "TF-" to SR ones. Some help with that would be appreciated too! |
Sure! There might be a minimally manual way to do it via the "Export" option at the top right of the page... I don't see an import option however. |
Any preferences about converting them? With or without comments? Is order important? Close all the converted TF bugs I assume? Preserve blocking/related/subtask relationships? Etc. |
You can click "More" -> "Move" to convert these issues to SR. It will keep all the existing info and just change the bug number. |
🙂 that sounds more sensical. |
How many/which of the TF issues do you want to convert? |
All of AD-related bugs. |
If it's not too much trouble, please add an "AutoDiff" label if there isn't one. Thanks so much, @porterchild! |
Glad to help! |
Including Closed issues, @rxwei? |
Just the open ones. |
Include ones like https://bugs.swift.org/browse/TF-4? Not a bug, but directly to do with autodiff... Does order matter? I was going to start back at the beginning. |
I'll go ahead and close TF-4 since it's complete. |
Let's sync issues that are related specifically to the AD feature, not the S4TF-specific ones such as planning/process. Thanks again! |
Looks like there some of the earlier issues may be stale, so I'll start around TF-700. I'm just going to convert the ones that are obviously bugs/crashes (and leave out everything else). Let me know if you want to be more/less inclusive than that. |
Sounds good. Thank you! |
👍 |
OK I should have gotten most of them from TF-700 on. I can go back further than that if you like. Let me know if you want to convert any other types of tickets |
Should be resolved by PR #37861. |
Additional Detail from JIRA
md5: de69ff0ec8ab8408330b2613882418fa
Issue Description:
In the below code, the gradient values propagate unexpectedly. Specifically, the self.number tangent vector disappears after leaving the scope of inner() (on the reverse pass). Removing the useless loop results in correct gradients.
The text was updated successfully, but these errors were encountered: