mkolod commented on issue #11380: Add ability to query cuDNN BatchNorm min. 
epsilon. Allow ONNX importer to use cuDNN BN if chosen eps >= cuDNN min. eps.
URL: https://github.com/apache/incubator-mxnet/pull/11380#issuecomment-401122235
 
 
   @piiswrong Do you have other suggestions? `CUDNN_BN_MIN_EPSILON ` is a 
macro, not a constant that can be called from ctypes? Models imported via ONNX 
will work much slower when using the framework's BatchNorm. This will get more 
severe when more optimized kernels are added, e.g. for NHWC. If not adding the 
C API is overkill, then I see 2 options:
   
   (1) Ignore the feature altogether. This results in ONNX imported models 
working more slowly than the MxNet checkpointed models.
   
   (2) We can hard-code the value to 1e-5 in Python, and not check what cuDNN 
provides. In that case, we have no C API, but we risk getting the wrong value 
in case cuDNN changes the constant. It's extremely unlikely that cuDNN will 
ever change it, but it's not a guarantee.
   
   What should I do next?

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
[email protected]


With regards,
Apache Git Services

Reply via email to