[bugfix]修复了trainer在update_every大于1时loss的显示错误

fastNLP/core/trainer.py

@@ -666,8 +666,8 @@ class Trainer(object):
                     # edit prediction
                     self.callback_manager.on_loss_begin(batch_y, prediction)
                     loss = self._compute_loss(prediction, batch_y).mean()
-                    avg_loss += loss.item()
                     loss = loss / self.update_every
+                    avg_loss += loss.item()
 
                     # Is loss NaN or inf? requires_grad = False
                     self.callback_manager.on_backward_begin(loss)