Skip to content

Fix batch norm defaults #83

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

Closed
albertz opened this issue Nov 20, 2021 · 4 comments
Closed

Fix batch norm defaults #83

albertz opened this issue Nov 20, 2021 · 4 comments
Milestone

Comments

@albertz
Copy link
Member

albertz commented Nov 20, 2021

They are wrong in RETURNN. But we can fix them here.

@albertz albertz added this to the first-release milestone Nov 20, 2021
@albertz
Copy link
Member Author

albertz commented Nov 23, 2021

We can also fix them on RETURNN side: rwth-i6/returnn#522
Edit This is done now on RETURNN side.

And then maybe via #84. Edit Also done.

@albertz
Copy link
Member Author

albertz commented Jan 5, 2022

When we just reimplement batch norm here in returnn-common (#89), of course we also should fix them as well. The question is, what defaults to use then. Maybe the new defaults as discussed in rwth-i6/returnn#522.

@albertz
Copy link
Member Author

albertz commented Jan 5, 2022

The question is also about masked_time (or whatever equivalent option we would have here). masked_time=False could have a big impact on the runtime performance efficiency, and is even the default in all other frameworks, even though it is a bit incorrect.

I would say that masked_time maybe should even be mandatory when there are dynamic dims, to be explicit about this behavior.

See also here for some discussion on RETURNN side.

@albertz
Copy link
Member Author

albertz commented Jan 6, 2022

We wrapped the RETURNN BatchNormLayer now (see #89), and that one has the new defaults now (rwth-i6/returnn#522).

@albertz albertz closed this as completed Jan 6, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant