Fix batchnorm layer numerics by replacing powx()
#5136
Closed
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
shelhamer
changed the title
powx()
jeffdonahue
added a commit
that referenced
this pull request
Apr 13, 2017
|
Thanks for this fix @pfollmann! Merged in c560658 with the corresponding change for CPU batch norm. I left off the explicit casts and comments as I don't think these are needed for this stabilization. |
beniz
pushed a commit
to jolibrain/caffe
that referenced
this pull request
Jul 10, 2017
This was referenced Jan 2, 2018
Closed
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Outdated suggestions cannot be applied.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.
I had problems to train ResNet on ImageNet, by using the solver-setting debug_info: true, I saw that there were nan's produced by the BatchNorm layer.
The problem was that the square inside the variance-term was computed via the caffe_gpu_powx function. This is slow on the one hand, but also unstable, especially for negative numbers. Therefore it was replaced by a caffe_gpu_mul call.
In the same way, I added a caffe_gpu_sqrt function in order to replace the caffe_gpu_powx(..., ..., 0.5, ...), that could lead to the same problems.
Additionally I added Dtype() castings to make sure, where appropriate.
Note: I didn't change the CPU version as it seams to be working. Nevertheless, using caffe_sqr instead of caffe_powx (..., 2) could speed up the layer... of course, also here a sqrt-function could be better than a caffe_powx (..., 0.5).