!6272 fix bug in bert TrainOneStepCell

Merge pull request !6272 from wangnan39/fix_bug_in_bert

model_zoo/official/nlp/bert/src/bert_for_pre_training.py

@@ -258,7 +258,7 @@ class BertNetworkWithLoss(nn.Cell):
         return self.cast(total_loss, mstype.float32)
 
 
-class BertTrainOneStepCell(nn.Cell):
+class BertTrainOneStepCell(nn.TrainOneStepCell):
     """
     Encapsulation class of bert network training.
 

model_zoo/official/nlp/bert_thor/src/bert_for_pre_training.py

@@ -275,7 +275,7 @@ class BertNetworkWithLoss(nn.Cell):
         return self.cast(total_loss, mstype.float32)
 
 
-class BertTrainOneStepCell(nn.Cell):
+class BertTrainOneStepCell(nn.TrainOneStepCell):
     """
     Encapsulation class of bert network training.
 
@@ -287,7 +287,6 @@ class BertTrainOneStepCell(nn.Cell):
         optimizer (Optimizer): Optimizer for updating the weights.
         sens (Number): The adjust parameter. Default: 1.0.
     """
-
     def __init__(self, network, optimizer, sens=1.0):
         super(BertTrainOneStepCell, self).__init__(network, optimizer, sens)
         self.cast = P.Cast()