diff --git a/official/nlp/bert/src/finetune_eval_model.py b/official/nlp/bert/src/finetune_eval_model.py
index 09c524101bf008d3bd10c6dfea15b0c2b7c3bce0..1dd21ae760fd38102fc85a5c4cadb9b694ed56a5 100644
--- a/official/nlp/bert/src/finetune_eval_model.py
+++ b/official/nlp/bert/src/finetune_eval_model.py
@@ -16,17 +16,13 @@
 '''
 Bert finetune and evaluation model script.
 '''
-
-import numpy as np
-
 import mindspore.nn as nn
 from mindspore.common.initializer import TruncatedNormal
 from mindspore.ops import operations as P
-from mindspore import context, Tensor
+from mindspore import context
 from .bert_model import BertModel
 
 
-
 class BertCLSModel(nn.Cell):
     """
     This class is responsible for classification task evaluation, i.e. XNLI(num_labels=3),
@@ -142,8 +138,8 @@ class BertNERModel(nn.Cell):
             batch_size = input_ids.shape[0]
             data_type = self.dtype
             hidden_size = self.lstm_hidden_size
-            h0 = Tensor(np.zeros((2, batch_size, hidden_size)), data_type)
-            c0 = Tensor(np.zeros((2, batch_size, hidden_size)), data_type)
+            h0 = P.Zeros()((2, batch_size, hidden_size), data_type)
+            c0 = P.Zeros()((2, batch_size, hidden_size), data_type)
             seq, _ = self.lstm(seq, (h0, c0))
         seq = self.reshape(seq, self.shape)
         logits = self.dense_1(seq)