-
-
Notifications
You must be signed in to change notification settings - Fork 611
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
Tidy up Maxout
#1794
Tidy up Maxout
#1794
Conversation
In fact, the improvement is less clear at larger sizes. But I think we can optimise the gradient of
|
The stranger result to me is the forward pass. It's not clear why the before version would be faster, and it appears both versions are unrolling the per-element comparisons. |
I have been surprised before by
|
Come to think of it though, |
I was going to suggest reverting to I wrote some gradients for
|
Co-authored-by: Brian Chen <ToucheSir@users.noreply.github.com>
Co-authored-by: Brian Chen <ToucheSir@users.noreply.github.com>
bors r+ |
Build succeeded: |
Maxout is from #698 . This:
Maxout(layer, layer, layer)
, rather than providing a tuple, to be more like other layers (with deprecation)mapreduce
. I see now this was a performance choice at the time, discussed here Add MaxOut layer #647 (comment) , but with Zygote this is much slower.Before:
After: