The difference between SAME and VALID are discussed here but not which one delivers a better prediction model.
In nn.conv2d, 'SAME' seems to be preferred in order to reduce data loss but 'VALID' is faster.
Ignoring the time lost on 'SAME' is there any reason to use 'VALID' over 'SAME' for nn.maxpool? Won't 'SAME' deliver a better prediction model?
