Why the mini-batch size are always set as 218 or 512, with the train-data size can't be evenly divisible by the input size? Thanks