hux

So the experimental results in your paper are obtained by training with batch size 256 or 1024? And if my training log matches yours in terms of the loss value?...