Does BN care about input dimensionality? The 1d vs 2d vs 3d just assert input dims, all else is same. Is below valid for the N-dim (or at least 4D) case?
class BatchNormNd(nn.modules.batchnorm._BatchNorm):
def _check_input_dim(self, input):
pass
I verified this much but I wonder then what’s the point of having 1d etc instead of just one BatchNormNd? An API thing maybe, but as long as forward and backward passes are same within a network, I’m fine