diff --git a/research/nlp/textrcnn/src/textrcnn.py b/research/nlp/textrcnn/src/textrcnn.py
index 6fb032778973461bf180614cae05dbd02ba44467..32d71ce99e6fad4f0911dd1d9f7ef1893ea50421 100644
--- a/research/nlp/textrcnn/src/textrcnn.py
+++ b/research/nlp/textrcnn/src/textrcnn.py
@@ -48,7 +48,7 @@ class textrcnn(nn.Cell):
         if cell == "lstm":
             if self.gpu_flag:
                 self.lstm = nn.LSTM(self.embed_size, self.num_hiddens)
-                self.lstm.to_float(mstype.float16)
+                self.lstm.to_float(mstype.float32)
             else:
                 self.lstm = P.DynamicRNN(forget_bias=0.0)
                 self.w1_fw = Parameter(
@@ -216,5 +216,5 @@ class textrcnn(nn.Cell):
         output_dense = self.tanh(output_dense)  # sl*bs, num_hidden
         output = self.reshape(output_dense, (F.shape(x)[0], self.batch_size, self.num_hiddens))  # sl, bs, num_hidden
         output = self.reduce_max(output, 0)  # bs, num_hidden
-        outputs = self.cast(self.mydense(output), mstype.float16)  # bs, num_classes
+        outputs = self.cast(self.mydense(output), mstype.float32)  # bs, num_classes
         return outputs
diff --git a/research/nlp/textrcnn/train.py b/research/nlp/textrcnn/train.py
index c2113aa6736f996361dafba82eed6457153d7027..987b92d87c9f405435dd3c191e7bdb3235093253 100644
--- a/research/nlp/textrcnn/train.py
+++ b/research/nlp/textrcnn/train.py
@@ -84,8 +84,7 @@ def run_train():
     loss_cb = LossMonitor()
     time_cb = TimeMonitor()
     if cfg.cell == "lstm" and cfg.device_target == "GPU":
-        model = Model(network, loss_fn=loss, optimizer=opt, metrics={'acc': Accuracy()}, amp_level="O3",
-                      loss_scale_manager=loss_scale)
+        model = Model(network, loss_fn=loss, optimizer=opt, metrics={'acc': Accuracy()}, loss_scale_manager=loss_scale)
     else:
         model = Model(network, loss_fn=loss, optimizer=opt, metrics={'acc': Accuracy()}, amp_level="O3")