Merge remote-tracking branch 'origin/dev' into dev

fastNLP/core/trainer.py

@@ -265,7 +265,7 @@ class Trainer(object):
 
                     # edit prediction
                     self.callback_manager.on_loss_begin(batch_y, prediction)
-                    loss = self._compute_loss(prediction, batch_y)
+                    loss = self._compute_loss(prediction, batch_y).mean()
                     avg_loss += loss.item()
                     loss = loss/self.update_every