The DirectML average pooling operator with setting IncludePadding true seems to have a strange behavior.
When calculating edges of input tensors, DirectML uses directly the window sizes for divisors of averages.
But the divisors should be truncated to the paddings if the kernel windows are larger than the paddings.
I have confirmed that Caffe and CuDnn behave to truncate such divisors.
If there is a same function as Caffe or CuDNN, please let me know how to use it.1 vote
- Don't see your idea?