-
Notifications
You must be signed in to change notification settings - Fork 961
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
BatchNorm layer failed with width==height==1 #14
Comments
Thanks. Reproduced. Will try to get this fixed this week. |
Thanks. |
Sorry for a delay. I've not committed a fix, but I have some news. This looks like a test issue. The bnorm kernel precomputes some values as an optimization while the test does not. Also, depending on the compiler, the test may or may not use FMA instructions. I'm looking into finding a way to change the test to allow such behavior while not relaxing the accuracy requirements too much. |
I calculated this iw==ih==1 case with ATLAS outside, but got different result with same input data. |
Not that I'm proud of this, but I had to resort to relaxing the precision requirements... I've added some 1x1 cases which now pass. Closing... (note to self: close this via a commit message next time) |
Found the gtest failed with BN layer when manually setting h==w==1
Line 406 at tests/gtests/test_batch_normalization.cpp
The text was updated successfully, but these errors were encountered: