support parallel loss

fastNLP/core/losses.py

@@ -251,7 +251,8 @@ class LossInForward(LossBase):
         if not (isinstance(loss, torch.Tensor) and len(loss.size()) == 0):
             if not isinstance(loss, torch.Tensor):
                 raise TypeError(f"Loss excepted to be a torch.Tensor, got {type(loss)}")
-            raise RuntimeError(f"The size of loss excepts to be torch.Size([]), got {loss.size()}")
+            loss = torch.sum(loss) / (loss.view(-1)).size(0)
+            # raise RuntimeError(f"The size of loss excepts to be torch.Size([]), got {loss.size()}")
 
         return loss