We still changing the batch size during learning given the change in the number of samples per forward pass.