Why convolutions always use odd-numbers as filter_size