MSEloss by default uses 'mean' as reduction method

[akiori]

MSEloss by default uses 'mean' as reduction method, so I think epoch_loss += (loss.detach().item() / batchsize) is incorrect

[Alex-177]

I came in to write the same thing. Besides, now the loss is accumulating on every butch. More precisely, the loss / batchsize value is accumulated. It seems to me more logical to output the loss value on the last batch of training.