I hope someone can clarify this to me, I was checking the implementation of a FCN in caffe from berkeleyvision and I saw the MaxPooling layer doesn’t mention any reference to ceil_mode, I get the use of ceil_mode, as per the documentation:
ceil_mode: when True, will use ceil instead of floor to compute the output shape
And I know how to calculate the output shape if that mode is used, but I’ve many implementations of FCNs in pytorch here and here. Is there any accuracy gain when using ceil_mode that I’m missing out? Should it make a difference vs using ceil_mode=False?
Thanks in advance!