Fix AlphaDropout implementation and add tests #1781
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Behaviour and outputs are adapted from the PyTorch and TF implementations
|
Unexpected bonus: this is now 100% GPU compatible and leaves us with one less broken test :) |
|
looks good! |
CarloLucibello
previously approved these changes
Nov 24, 2021
|
Looks good, I think this deserves an entry in NEWS.md. Especially the GPU bit. |
CarloLucibello
approved these changes
Nov 26, 2021
|
bors r+ |
| return x | ||
| p = a.p | ||
| iszero(p) && return x | ||
| isone(p) && return sign.(x) .* T(0) |
There was a problem hiding this comment.
Is this correct? Are we intentionally creating -0.0 values here.
There was a problem hiding this comment.
Yes, see point 2 in #1781 (comment).
| iszero(p) && return x | ||
| isone(p) && return sign.(x) .* T(0) | ||
|
|
||
| α′ = T(-1.7580993408473766) # selu(-Inf) == -λα |
There was a problem hiding this comment.
oftype requires a value of type T, but the only way to get that in this function would be to extract an element of x (which would fail if x is empty and might also trigger the getindex adjoint unnecessarily). Since this function should be specializing on eltype anyhow, it makes sense to take advantage of that for casting.
|
Build succeeded: |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
AFAICT, the original implementation never behaved as expected even pre-Zygote. This was likely not caught because the original PR didn't come with tests, so this PR should remedy that.
Behaviour and outputs are adapted from the PyTorch and TF implementations. Some points of note:
p = 0to avoid propagating NaNs when calcuatingAandB.p = 1. TF just returns the input, but I think the PyTorch approach of returning all zeros (+/- depending on the input sign) is more in line withDropout.ifelseis used instead of something like https://github.com/keras-team/keras/blob/v2.7.0/keras/layers/noise.py#L200. I think it better reflects the conditional nature of the operation and it was also faster in local benchmarking.PR Checklist