Loss update wrongly for variable batch_size

[The-Anh Vu-Le]

Last batch might be of different sizes so loss_meter.add(loss) is not correct, should be loss_meter.add(loss*bs, bs).