Fix tanh layer for badly scaled input values #1455
Merged
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Outdated suggestions cannot be applied.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.
The current tanh layer uses an expression that is unstable for large values. It returns NaN if the input is badly scaled (outside the range [-40, 40] for single precision float).
This fix uses the tanh library function, which is much more stable. I also added a unit test (there was previously no test for tanh) that checks the case of large input values. It compares against an exact expression for tanh that also handles overflow. The unit test fails if you use the previous naive implementation.