We wrote a custom loss with PyTorch. The problem is that even if we write:
... loss = Variable(loss, requires_grad = True) return loss
the g
