diff --git a/mindspore/lite/CMakeLists.txt b/mindspore/lite/CMakeLists.txt
index 4918355e2a..f4806af9b3 100644
--- a/mindspore/lite/CMakeLists.txt
+++ b/mindspore/lite/CMakeLists.txt
@@ -219,6 +219,9 @@ add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/internal)
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/nnacl)
 if (ENABLE_TOOLS)
     add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/tools/benchmark)
+    if (SUPPORT_TRAIN)
+        add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/tools/net_train)
+    endif()
 endif()
 if (NOT WIN32)
     if (ENABLE_TOOLS)
diff --git a/mindspore/lite/include/train_model.h b/mindspore/lite/include/train_model.h
index c93cd342da..e1a3366761 100644
--- a/mindspore/lite/include/train_model.h
+++ b/mindspore/lite/include/train_model.h
@@ -18,32 +18,36 @@
 #include <vector>
 #include "include/model.h"
 
-namespace mindspore::lite {
+namespace mindspore {
+namespace lite {
+
+/// \brief TrainModel Defines a class that allows to import and export a mindsport trainable model
 struct TrainModel : public lite::Model {
-  /// \brief Static method to create a TrainModel pointer.
-  ///
-  /// \param[in] model_buf Define the buffer read from a model file.
-  /// \param[in] size Define bytes number of model buffer.
+  /// \brief Static method to create a TrainModel object
   ///
-  /// \return Pointer of MindSpore Lite TrainModel.
+  /// \param[in] model_buf A buffer that was read from a MS model file
+  /// \param[in] size Length of the buffer
+  //
+  /// \return Pointer to MindSpore Lite TrainModel
   static TrainModel *Import(const char *model_buf, size_t size);
 
-  /// \brief Free meta graph temporary buffer
+  /// \brief Free meta graph related data
   void Free() override;
 
-  /// \brief TrainModel destruct, free all memory
+  /// \brief Class destructor, free all memory
   virtual ~TrainModel();
 
-  /// \brief Export Model into buf.
+  /// \brief Export Model into a buffer
   ///
-  /// \param[in] buf Define the buffer to Export into. If nullptr, buf will be allocated
-  /// \param[in] len size of the buffer.
+  /// \param[in] buf The buffer to Export into. If equal to nullptr, buf will be allocated
+  /// \param[in,out] len Size of the pre-allocated buffer, and returned size of the exported buffer
   ///
   /// \return Pointer to buffer with exported model
-  char* ExportBuf(char* buf, size_t* len) const;
+  char *ExportBuf(char *buf, size_t *len) const;
 
   size_t buf_size_;
 };
-}  // namespace mindspore::lite
+}  // namespace lite
+}  // namespace mindspore
 
 #endif  // MINDSPORE_LITE_INCLUDE_TRAIN_MODEL_H_
diff --git a/mindspore/lite/include/train_session.h b/mindspore/lite/include/train_session.h
index c3cab39d61..fbe00bc8fb 100644
--- a/mindspore/lite/include/train_session.h
+++ b/mindspore/lite/include/train_session.h
@@ -25,16 +25,59 @@
 namespace mindspore {
 namespace session {
 
+/// \brief TrainSession Defines a class that allows training a MindSpore model
 class TrainSession : public session::LiteSession {
  public:
+  /// \brief Class destructor
   virtual ~TrainSession() = default;
+
+  /// \brief Static method to create a TrainSession object
+  ///
+  /// \param[in] context Defines the context of the session to be created
+  ///
+  /// \return Pointer of MindSpore Lite TrainSession
   static TrainSession *CreateSession(lite::Context *context);
 
+  /// \brief Compile MindSpore Lite train model
+  ///
+  /// \note CompileTrainGraph should be called before RunGraph
+  ///
+  /// \param[in] model Define the model to be compiled
+  ///
+  /// \return STATUS as an error code of compiling graph, STATUS is defined in errorcode.h
   virtual int CompileTrainGraph(lite::TrainModel *model) = 0;
+
+  /// \brief Export the trained model into a buffer
+  ///
+  /// \param[in] buf The buffer to Export into. If equal to nullptr, buf will be allocated
+  /// \param[in,out] len Size of the pre-allocated buffer, and returned size of the exported buffer
+  ///
+  /// \return pointer to the export buffer
   virtual void *ExportToBuf(char *buf, size_t *len) const = 0;
-  virtual void Train() = 0;
+
+  /// \brief Save the trained model into a flatbuffer file
+  ///
+  /// \param[in] filename Filename to save flatbuffer to
+  ///
+  /// \return 0 on success or -1 in case of error
+  virtual int SaveToFile(const std::string &filename) const = 0;
+
+  /// \brief Set model to train mode
+  /// \return STATUS as an error code of compiling graph, STATUS is defined in errorcode.h
+  virtual int Train() = 0;
+
+  /// \brief Check mode of model
+  ///
+  /// \return boolean indication if model is in train mode
   bool IsTrain() { return train_mode_ == true; }
-  virtual void Eval() = 0;
+
+  /// \brief Set model to eval mode
+  /// \return STATUS as an error code of compiling graph, STATUS is defined in errorcode.h
+  virtual int Eval() = 0;
+
+  /// \brief Check mode of model
+  ///
+  /// \return boolean indication if model is in eval mode
   bool IsEval() { return train_mode_ == false; }
 
  protected:
diff --git a/mindspore/lite/minddata/CMakeLists.txt b/mindspore/lite/minddata/CMakeLists.txt
index 2e01d699fd..576c829b35a 100644
--- a/mindspore/lite/minddata/CMakeLists.txt
+++ b/mindspore/lite/minddata/CMakeLists.txt
@@ -270,11 +270,13 @@ if (BUILD_MINDDATA STREQUAL "full")
         ${CORE_DIR}/utils/ms_utils.cc
         )
 
+    find_package(Threads REQUIRED)
     target_link_libraries(minddata-lite
         securec
         jpeg-turbo
         jpeg
         mindspore::json
+        Threads::Threads
         )
 
     # ref: https://github.com/android/ndk/issues/1202
diff --git a/mindspore/lite/nnacl/fp32/batchnorm.c b/mindspore/lite/nnacl/fp32/batchnorm.c
index 0e2449b1b8..a11138daca 100644
--- a/mindspore/lite/nnacl/fp32/batchnorm.c
+++ b/mindspore/lite/nnacl/fp32/batchnorm.c
@@ -55,20 +55,30 @@ void FusedBatchNormFp32(const void *input, const void *scale, const void *offset
 
 void FusedBatchNormFp32MeanVar(const float *input, float *run_mean, float *run_var, BatchNormParameter *param,
                                float *save_mean, float *save_var) {
-  float N = (float)param->unit_;
+  const float N = (float)param->unit_;
+  const float VN = N;
+  const float VNUB = (N > 1.0f) ? (N - 1.0f) : 1.0f;
+  const float momentum = (1.0f - param->momentum_);
+
   for (int i = 0; i < param->unit_; i++) {
     for (int c = 0; c < param->channel_; c++) {
       int idx = i * param->channel_ + c;
       run_mean[c] += input[idx];
-      run_var[c] += input[idx] * input[idx];
     }
   }
-  const float VN = (N > 1.0f) ? (N - 1.0f) : 1.0f;
   for (int c = 0; c < param->channel_; c++) {
-    run_mean[c] = run_mean[c] / N;
-    run_var[c] = run_var[c] / VN - run_mean[c] * run_mean[c];
-    save_mean[c] = param->momentum_ * save_mean[c] + (1 - param->momentum_) * run_mean[c];
-    const float var = run_var[c];
-    save_var[c] = param->momentum_ * save_var[c] + (1 - param->momentum_) * var;
+    run_mean[c] /= N;
+  }
+  for (int i = 0; i < param->unit_; i++) {
+    for (int c = 0; c < param->channel_; c++) {
+      int idx = i * param->channel_ + c;
+      run_var[c] += (input[idx] - run_mean[c]) * (input[idx] - run_mean[c]);
+    }
+  }
+  for (int c = 0; c < param->channel_; c++) {
+    float unbiased_var = (run_var[c] / VNUB);
+    run_var[c] = (run_var[c] / VN);
+    save_mean[c] = momentum * save_mean[c] + (1.0f - momentum) * run_mean[c];
+    save_var[c] = momentum * save_var[c] + (1.0f - momentum) * unbiased_var;
   }
 }
diff --git a/mindspore/lite/nnacl/fp32_grad/activation_grad.c b/mindspore/lite/nnacl/fp32_grad/activation_grad.c
index d7b070473d..df20d9447e 100644
--- a/mindspore/lite/nnacl/fp32_grad/activation_grad.c
+++ b/mindspore/lite/nnacl/fp32_grad/activation_grad.c
@@ -72,7 +72,7 @@ int HSwishGrad(float *src0, float *src1, int length, float *dst) {
 
 int HSigmoidGrad(float *src0, float *src1, int length, float *dst) {
   for (int i = 0; i < length; ++i) {
-    float tmp = (src1[i] > 3.0f ? 1.0f : (src1[i] < -3.0f ? 0.0f : 1.0f / 6.0f));
+    float tmp = (src1[i] > 3.0f ? 0.0f : (src1[i] < -3.0f ? 0.0f : 1.0f / 6.0f));
     dst[i] = tmp * src0[i];
   }
   return NNACL_OK;
diff --git a/mindspore/lite/nnacl/fp32_grad/arithmetic_grad.c b/mindspore/lite/nnacl/fp32_grad/arithmetic_grad.c
index 6e5b56ecf5..4c0ebe628b 100644
--- a/mindspore/lite/nnacl/fp32_grad/arithmetic_grad.c
+++ b/mindspore/lite/nnacl/fp32_grad/arithmetic_grad.c
@@ -15,6 +15,8 @@
  */
 
 #include "nnacl/fp32_grad/arithmetic_grad.h"
+#include <string.h>
+#include "nnacl/fp32_grad/utils.h"
 
 void ElementDivNegSquare(const float *nom, const float *denom, float *output, int element_size) {
   for (int i = 0; i < element_size; i++) {
@@ -27,3 +29,103 @@ void ElementMulAndDivNegSquare(const float *a, const float *b, const float *deno
     output[i] = -a[i] * b[i] / (denom[i] * denom[i]);
   }
 }
+
+void MaximumByAxes(const float *input0, const float *input1, const float *dy, const int *input0_dims,
+                   const int *input1_dims, const int *dy_dims, float *output0, float *output1, int num_dims) {
+  int num_output0 = 1;
+  int num_output1 = 1;
+  int same_shape = 1;
+  for (int idx = 0; idx < num_dims; ++idx) {
+    num_output0 *= input0_dims[idx];
+    num_output1 *= input1_dims[idx];
+    if (input0_dims[idx] != input1_dims[idx]) {
+      same_shape = 0;
+    }
+  }
+
+  if (same_shape) {
+    int input_iter[8] = {0};
+
+    // Iterate through input_data.
+    do {
+      size_t offset = GetInputOffset(num_dims, input0_dims, input_iter);
+      output0[offset] = input0[offset] > input1[offset] ? dy[offset] : 0.;
+      output1[offset] = input1[offset] >= input0[offset] ? dy[offset] : 0.;
+    } while (NextIndex(num_dims, input0_dims, input_iter));
+  } else {
+    memset(output0, 0, num_output0 * sizeof(float));  // zero output
+    memset(output1, 0, num_output1 * sizeof(float));  // zero output
+
+    int input_iter[8] = {0};
+    int axes0[5] = {0};
+    int axes1[5] = {0};
+    int num_axes0 = 0;
+    int num_axes1 = 0;
+    for (int i = 0; i < num_dims; i++) {
+      if (input0_dims[i] == 1) {
+        axes0[num_axes0++] = i;
+      }
+      if (input1_dims[i] == 1) {
+        axes1[num_axes1++] = i;
+      }
+    }
+
+    do {
+      size_t offset0 = GetOutputOffset(num_dims, input0_dims, input_iter, num_axes0, axes0);
+      size_t offset1 = GetOutputOffset(num_dims, input1_dims, input_iter, num_axes1, axes1);
+      size_t yt_offset = GetInputOffset(num_dims, input0_dims, input_iter);
+      output0[offset0] += input0[offset0] > input1[offset1] ? dy[yt_offset] : 0.;
+      output1[offset1] += input1[offset1] >= input0[offset0] ? dy[yt_offset] : 0.;
+    } while (NextIndex(num_dims, dy_dims, input_iter));
+  }
+}
+
+void MinimumByAxes(const float *input0, const float *input1, const float *dy, const int *input0_dims,
+                   const int *input1_dims, const int *dy_dims, float *output0, float *output1, int num_dims) {
+  int num_output0 = 1;
+  int num_output1 = 1;
+  int same_shape = 1;
+  for (int idx = 0; idx < num_dims; ++idx) {
+    num_output0 *= input0_dims[idx];
+    num_output1 *= input1_dims[idx];
+    if (input0_dims[idx] != input1_dims[idx]) {
+      same_shape = 0;
+    }
+  }
+
+  if (same_shape) {
+    int input_iter[8] = {0};
+
+    // Iterate through input_data.
+    do {
+      size_t offset = GetInputOffset(num_dims, input0_dims, input_iter);
+      output0[offset] = input0[offset] < input1[offset] ? dy[offset] : 0.;
+      output1[offset] = input1[offset] <= input0[offset] ? dy[offset] : 0.;
+    } while (NextIndex(num_dims, input0_dims, input_iter));
+  } else {
+    memset(output0, 0, num_output0 * sizeof(float));  // zero output
+    memset(output1, 0, num_output1 * sizeof(float));  // zero output
+
+    int input_iter[8] = {0};
+    int axes0[5] = {0};
+    int axes1[5] = {0};
+    int num_axes0 = 0;
+    int num_axes1 = 0;
+    for (int i = 0; i < num_dims; i++) {
+      if (input0_dims[i] == 1) {
+        axes0[num_axes0++] = i;
+      }
+      if (input1_dims[i] == 1) {
+        axes1[num_axes1++] = i;
+      }
+    }
+
+    do {
+      size_t offset0 = GetOutputOffset(num_dims, input0_dims, input_iter, num_axes0, axes0);
+      size_t offset1 = GetOutputOffset(num_dims, input1_dims, input_iter, num_axes1, axes1);
+      size_t yt_offset = GetInputOffset(num_dims, input0_dims, input_iter);
+      output0[offset0] += input0[offset0] < input1[offset1] ? dy[yt_offset] : 0.;
+      output1[offset1] += input1[offset1] <= input0[offset0] ? dy[yt_offset] : 0.;
+    } while (NextIndex(num_dims, dy_dims, input_iter));
+  }
+}
diff --git a/mindspore/lite/nnacl/fp32_grad/arithmetic_grad.h b/mindspore/lite/nnacl/fp32_grad/arithmetic_grad.h
index 34486ab7fa..da78a314ef 100644
--- a/mindspore/lite/nnacl/fp32_grad/arithmetic_grad.h
+++ b/mindspore/lite/nnacl/fp32_grad/arithmetic_grad.h
@@ -16,11 +16,17 @@
 #ifndef MINDSPORE_LITE_NNACL_FP32_GRAD_ARITHMETIC_GRAD_H_
 #define MINDSPORE_LITE_NNACL_FP32_GRAD_ARITHMETIC_GRAD_H_
 
+#include "nnacl/op_base.h"
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 void ElementDivNegSquare(const float *nom, const float *denom, float *output, int element_size);
 void ElementMulAndDivNegSquare(const float *a, const float *b, const float *denom, float *output, int element_size);
+void MaximumByAxes(const float *input0, const float *input1, const float *dy, const int *input0_dims,
+                   const int *input1_dims, const int *dy_dims, float *output0, float *output1, int num_dims);
+void MinimumByAxes(const float *input0, const float *input1, const float *dy, const int *input0_dims,
+                   const int *input1_dims, const int *dy_dims, float *output0, float *output1, int num_dims);
 #ifdef __cplusplus
 }
 #endif
diff --git a/mindspore/lite/nnacl/fp32_grad/batch_norm.c b/mindspore/lite/nnacl/fp32_grad/batch_norm.c
index a353dbc85f..d2f489dc52 100644
--- a/mindspore/lite/nnacl/fp32_grad/batch_norm.c
+++ b/mindspore/lite/nnacl/fp32_grad/batch_norm.c
@@ -17,66 +17,55 @@
 #include <string.h>
 #include "nnacl/fp32_grad/batch_norm.h"
 
-void sumSpatialBatch(const float *in, int size, int ch, float *out) {
+void sumSpatialBatch(const float *in, size_t size, int ch, float *out) {
   memset(out, 0, ch * sizeof(float));
-  for (int i = 0; i < size; i++) {
-    const float *ptr = in + i * ch;
-    for (int c = 0; c < ch; c++) {
+  for (size_t i = 0; i < size; i++) {
+    const float *ptr = in + (i * ch);
+    for (size_t c = 0; c < ch; c++) {
       out[c] += ptr[c];
     }
   }
 }
 
-static void meanVar(const float *in, int size, int ch, float eps, float *mean, float *invar) {
-  float N = (float)(size);
-  sumSpatialBatch(in, N, ch, mean);
-  for (int f = 0; f < ch; ++f) {
-    mean[f] /= N;
-  }
-  for (int f = 0; f < ch; f++) {
-    float tvar = 0;
-    for (int i = 0; i < N; i++) {
-      float x = in[i * ch + f];
-      tvar += (x - mean[f]) * (x - mean[f]);
-    }
-    invar[f] = 1.0f / (sqrt(tvar / N + eps));
-  }
-}
-
-void backwardX(const float *in, const float *dout, const float *scale, const int size, int channels, float eps,
-               float *mean, float *invar, float *dxhathat_sum, float *dxhat_sum, float *out) {
-  meanVar(in, size, channels, eps, mean, invar);
-  for (int i = 0; i < size; i++) {
-    for (int f = 0; f < channels; f++) {
-      int ix = i * channels + f;
+void backwardX(const float *in, const float *dout, const float *scale, const size_t size, int channels, float *mean,
+               float *invar, float *dxhathat_sum, float *dxhat_sum, float *out) {
+  const float N = (size);
+  for (size_t i = 0; i < size; i++) {
+    for (size_t f = 0; f < channels; f++) {
+      size_t ix = i * channels + f;
       float x_hat = (in[ix] - mean[f]) * invar[f];
-      float dxhat = dout[ix] * scale[f];
-      dxhat_sum[f] += dxhat;
-      dxhathat_sum[f] += dxhat * x_hat;
+      float dx_hat = dout[ix] * scale[f];
+      dxhat_sum[f] += dx_hat;
+      dxhathat_sum[f] += dx_hat * x_hat;
     }
   }
-  for (int i = 0; i < size; i++) {
-    for (int f = 0; f < channels; f++) {
-      int ix = i * channels + f;
+  for (size_t i = 0; i < size; i++) {
+    for (size_t f = 0; f < channels; f++) {
+      size_t ix = i * channels + f;
       float x_hat = (in[ix] - mean[f]) * invar[f];
-      float dxhat = dout[ix] * scale[f];
-      out[ix] = 1.f / size * invar[f] * (size * dxhat - dxhat_sum[f] - x_hat * dxhathat_sum[f]);
+      float dx_hat = dout[ix] * scale[f];
+      out[ix] = 1.0f / N * (invar[f]) * (N * dx_hat - dxhat_sum[f] - x_hat * dxhathat_sum[f]);
     }
   }
 }
 
-void backwardScale(const float *x, const float *mean, const float *invar, const float *delta, int batch,
-                   int n, int size, float *scale_updates) {
-  int i, b, f;
+void backwardScale(const float *x, const float *mean, const float *invar, const float *delta, int batch, int n,
+                   int size, float *scale_updates) {
+  size_t i, b, f;
   memset(scale_updates, 0, n * sizeof(float));
   for (b = 0; b < batch; ++b) {
     for (i = 0; i < size; ++i) {
       for (f = 0; f < n; ++f) {
         int index = (b * size + i) * n + f;
         float x_norm = (x[index] - mean[f]) * invar[f];
-        scale_updates[f] += delta[index] * x_norm;
+        scale_updates[f] += (delta[index] * x_norm);
       }
     }
   }
 }
 
+void var2Invar(float *save_var, size_t size, float eps) {
+  for (size_t i = 0; i < size; i++) {
+    save_var[i] = 1.0f / sqrt(save_var[i] + eps);
+  }
+}
diff --git a/mindspore/lite/nnacl/fp32_grad/batch_norm.h b/mindspore/lite/nnacl/fp32_grad/batch_norm.h
index 307610b724..53cc6437da 100644
--- a/mindspore/lite/nnacl/fp32_grad/batch_norm.h
+++ b/mindspore/lite/nnacl/fp32_grad/batch_norm.h
@@ -29,11 +29,12 @@ typedef struct BNGradParameter {
 extern "C" {
 #endif
 
-void sumSpatialBatch(const float *in, int size, int ch, float *out);
-void backwardX(const float *in, const float *dout, const float *scale, const int size, int channels, float eps,
-               float *mean, float *invar, float *xhat_sum, float *dxhat_sum, float *out);
-void backwardScale(const float *x, const float *mean, const float *invar, const float *delta, int batch,
-                   int n, int size, float *scale_updates);
+void sumSpatialBatch(const float *in, size_t size, int ch, float *out);
+void backwardX(const float *in, const float *dout, const float *scale, const size_t size, int channels, float *mean,
+               float *invar, float *xhat_sum, float *dxhat_sum, float *out);
+void backwardScale(const float *x, const float *mean, const float *invar, const float *delta, int batch, int n,
+                   int size, float *scale_updates);
+void var2Invar(float *save_var, size_t size, float eps);
 
 #ifdef __cplusplus
 }
diff --git a/mindspore/lite/nnacl/fp32_grad/dropout_grad.c b/mindspore/lite/nnacl/fp32_grad/dropout_grad.c
new file mode 100644
index 0000000000..b54924bada
--- /dev/null
+++ b/mindspore/lite/nnacl/fp32_grad/dropout_grad.c
@@ -0,0 +1,23 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "nnacl/fp32_grad/dropout_grad.h"
+
+void DropoutGrad(const float *yt_ptr, const float *mask, float *output_ptr, int length, float scale) {
+  for (int i = 0; i < length; i++) {
+    output_ptr[i] = yt_ptr[i] * mask[i] * scale;
+  }
+}
diff --git a/mindspore/lite/nnacl/fp32_grad/dropout_grad.h b/mindspore/lite/nnacl/fp32_grad/dropout_grad.h
new file mode 100644
index 0000000000..1124eb29f6
--- /dev/null
+++ b/mindspore/lite/nnacl/fp32_grad/dropout_grad.h
@@ -0,0 +1,31 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_NNACL_FP32_GRAD_DROPOUT_GRAD_H_
+#define MINDSPORE_LITE_NNACL_FP32_GRAD_DROPOUT_GRAD_H_
+
+#include "nnacl/op_base.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void DropoutGrad(const float *yt_ptr, const float *mask, float *output_ptr, int length, float ratio);
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // MINDSPORE_LITE_NNACL_FP32_GRAD_DROPOUT_GRAD_H_
diff --git a/mindspore/lite/nnacl/fp32_grad/dropout_parameter.h b/mindspore/lite/nnacl/fp32_grad/dropout_parameter.h
new file mode 100644
index 0000000000..789254abae
--- /dev/null
+++ b/mindspore/lite/nnacl/fp32_grad/dropout_parameter.h
@@ -0,0 +1,27 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_NNACL_FP32_GRAD_DROPOUT_PARAMETER_H_
+#define MINDSPORE_LITE_NNACL_FP32_GRAD_DROPOUT_PARAMETER_H_
+
+#include "nnacl/op_base.h"
+
+typedef struct DropoutParameter {
+  OpParameter op_parameter_;
+  float ratio_;
+} DropoutParameter;
+
+#endif  // MINDSPORE_LITE_NNACL_FP32_GRAD_DROPOUT_PARAMETER_H_
diff --git a/mindspore/lite/nnacl/fp32_grad/gemm.c b/mindspore/lite/nnacl/fp32_grad/gemm.c
index 0ec25c4141..92a718cef6 100644
--- a/mindspore/lite/nnacl/fp32_grad/gemm.c
+++ b/mindspore/lite/nnacl/fp32_grad/gemm.c
@@ -16,182 +16,536 @@
 
 #include "nnacl/fp32_grad/gemm.h"
 #include <string.h>
+#ifdef __ARM_NEON
+#include <arm_neon.h>
+#endif
+#include "nnacl/fp32/matmul.h"
 
-static void gemm_not_trana_not_tranb(int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b, int ldb,
-                                     float *mat_c, int ldc) {
-  const int block_size = 4;
-  int block_mod = N % block_size;
-  int block_c4 = N - block_mod;
-
-  int i, j, k;
-  for (i = 0; i < M; ++i) {
-    for (k = 0; k < K; ++k) {
-      float a = alpha * mat_a[i * lda + k];
-      for (j = 0; j < block_c4; j += block_size) {
-        float *b = &mat_b[k * ldb + j];
-        float *c = &mat_c[i * ldc + j];
-        c[0] += a * b[0];
-        c[1] += a * b[1];
-        c[2] += a * b[2];
-        c[3] += a * b[3];
-      }
-      for (; j < N; ++j) {
-        mat_c[i * ldc + j] += a * mat_b[k * ldb + j];
-      }
+static void addv(const float *restrict v1, float *restrict v2, float beta, int row, int col, int stride) {
+  const float *src_ptr = v1;
+  float *dst_ptr = v2;
+  for (int r = 0; r < row; r++) {
+    for (int c = 0; c < col; c++) {
+      dst_ptr[c] += beta * src_ptr[c];
+    }
+    src_ptr += stride;
+    dst_ptr += stride;
+  }
+}
+
+int MatSize(int row, int col, int round) {
+  int res = UP_ROUND(row, round) * col;
+  return res;
+}
+
+int MatSizeTotal(int row, int col, int deep, int stride) {
+#ifdef ENABLE_ARM32
+  const int num = C4NUM;
+#else
+  const int num = C12NUM;
+#endif
+  int res = MatSize(row, deep, num) + MatSize(col, deep, C8NUM);
+  if (stride > 0) res += row * stride;
+  return res;
+}
+#ifdef ENABLE_ARM32
+static void RowMajor2Row4MajorStride(const float *src_ptr, float *dst_ptr, int row, int col, int lead) {
+  for (int r = 0; r < row; r++) {
+    const float *src = src_ptr + r * lead;
+    for (int c = 0; c < col; c++) {
+      int cd8 = c / 4;
+      int cm8 = c % 4;
+      dst_ptr[cd8 * 4 * row + r * 4 + cm8] = src[c];
+    }
+  }
+}
+#endif
+
+static void RowMajor2Row8MajorStride(const float *src_ptr, float *dst_ptr, int row, int col, int lead) {
+  for (int r = 0; r < row; r++) {
+    const float *src = src_ptr + r * lead;
+    for (int c = 0; c < col; c++) {
+      int cd8 = c / 8;
+      int cm8 = c % 8;
+      dst_ptr[cd8 * 8 * row + r * 8 + cm8] = src[c];
+    }
+  }
+  return;
+}
+
+#ifndef ENABLE_ARM32
+static void RowMajor2Row12MajorStride(const float *src_ptr, float *dst_ptr, int row, int col, int lead) {
+  for (int r = 0; r < row; r++) {
+    const float *src = src_ptr + r * lead;
+    for (int c = 0; c < col; c++) {
+      int cd8 = c / C12NUM;
+      int cm8 = c % C12NUM;
+      dst_ptr[cd8 * C12NUM * row + r * C12NUM + cm8] = src[c];
     }
   }
+  return;
 }
 
-static void gemm_not_trana_tranb(int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b, int ldb,
-                                 float *mat_c, int ldc) {
-  const int block_size = 4;
-  int block_mod = K % block_size;
-  int block_c4 = K - block_mod;
-
-  int i, j, k;
-  for (i = 0; i < M; ++i) {
-    for (j = 0; j < N; ++j) {
-      float sum = 0;
-      for (k = 0; k < block_c4; k += block_size) {
-        float *a = &mat_a[i * lda + k];
-        float *b = &mat_b[j * ldb + k];
-        sum += alpha * a[0] * b[0];
-        sum += alpha * a[1] * b[1];
-        sum += alpha * a[2] * b[2];
-        sum += alpha * a[3] * b[3];
+static void RowMajor2Col12MajorStride(const float *src_ptr, float *dst_ptr, size_t row, size_t col, int lead) {
+  size_t row_up_12 = UP_ROUND(row, C12NUM);
+  size_t row12 = row / C12NUM * C12NUM;
+  size_t col4 = col / C4NUM * C4NUM;
+  const float *src_r = src_ptr;
+  float *dst_r = dst_ptr;
+
+  size_t ri = 0;
+  for (; ri < row12; ri += C12NUM) {
+    size_t ci = 0;
+    for (; ci < col4; ci += C4NUM) {
+      const float *src_c = src_r + ci;
+      float *dst_c = dst_r + ci * C12NUM;
+
+      /* 12x4 row-major to col-major */
+#ifdef ENABLE_ARM64
+      size_t stride = lead * sizeof(float);
+      asm volatile(
+        "mov x10, %[src_c]\n"
+        "mov x11, %[dst_c]\n"
+
+        "ld1 {v0.4s}, [x10], %[stride]\n"
+        "ld1 {v1.4s}, [x10], %[stride]\n"
+        "ld1 {v2.4s}, [x10], %[stride]\n"
+        "ld1 {v3.4s}, [x10], %[stride]\n"
+
+        "ld1 {v4.4s}, [x10], %[stride]\n"
+        "ld1 {v5.4s}, [x10], %[stride]\n"
+        "ld1 {v6.4s}, [x10], %[stride]\n"
+        "ld1 {v7.4s}, [x10], %[stride]\n"
+
+        "zip1 v12.4s, v0.4s, v1.4s\n"
+        "zip2 v13.4s, v0.4s, v1.4s\n"
+        "zip1 v14.4s, v2.4s, v3.4s\n"
+        "zip2 v15.4s, v2.4s, v3.4s\n"
+
+        "ld1 {v8.4s}, [x10], %[stride]\n"
+        "ld1 {v9.4s}, [x10], %[stride]\n"
+        "ld1 {v10.4s}, [x10], %[stride]\n"
+        "ld1 {v11.4s}, [x10], %[stride]\n"
+
+        "zip1 v16.4s, v4.4s, v5.4s\n"
+        "zip2 v17.4s, v4.4s, v5.4s\n"
+        "zip1 v18.4s, v6.4s, v7.4s\n"
+        "zip2 v19.4s, v6.4s, v7.4s\n"
+
+        "trn1 v20.2d, v12.2d, v14.2d\n"
+        "trn2 v23.2d, v12.2d, v14.2d\n"
+        "trn1 v26.2d, v13.2d, v15.2d\n"
+        "trn2 v29.2d, v13.2d, v15.2d\n"
+
+        "trn1 v21.2d, v16.2d, v18.2d\n"
+        "trn2 v24.2d, v16.2d, v18.2d\n"
+        "trn1 v27.2d, v17.2d, v19.2d\n"
+        "trn2 v30.2d, v17.2d, v19.2d\n"
+
+        "zip1 v12.4s, v8.4s, v9.4s\n"
+        "zip2 v13.4s, v8.4s, v9.4s\n"
+        "zip1 v14.4s, v10.4s, v11.4s\n"
+        "zip2 v15.4s, v10.4s, v11.4s\n"
+
+        "trn1 v22.2d, v12.2d, v14.2d\n"
+        "trn2 v25.2d, v12.2d, v14.2d\n"
+        "trn1 v28.2d, v13.2d, v15.2d\n"
+        "trn2 v31.2d, v13.2d, v15.2d\n"
+
+        "st1 {v20.4s, v21.4s, v22.4s, v23.4s}, [x11], #64\n"
+        "st1 {v24.4s, v25.4s, v26.4s, v27.4s}, [x11], #64\n"
+        "st1 {v28.4s, v29.4s, v30.4s, v31.4s}, [x11], #64\n"
+
+        :
+        : [ dst_c ] "r"(dst_c), [ src_c ] "r"(src_c), [ stride ] "r"(stride)
+        : "x10", "x11", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14",
+          "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29",
+          "v30", "v31");
+#elif ENABLE_ARM32
+      size_t stride = lead * sizeof(float);
+      asm volatile(
+        "mov r10, %[src_c]\n"
+        "mov r12, %[dst_c]\n"
+
+        "vld1.32 {q0}, [r10], %[stride]\n"
+        "vld1.32 {q3}, [r10], %[stride]\n"
+        "vld1.32 {q10}, [r10], %[stride]\n"
+        "vld1.32 {q13}, [r10], %[stride]\n"
+
+        "vtrn.32 d0, d6\n"
+        "vtrn.32 d1, d7\n"
+        "vtrn.32 d20, d26\n"
+        "vtrn.32 d21, d27\n"
+
+        "vld1.32 {q1}, [r10], %[stride]\n"
+        "vld1.32 {q8}, [r10], %[stride]\n"
+        "vld1.32 {q11}, [r10], %[stride]\n"
+        "vld1.32 {q14}, [r10], %[stride]\n"
+
+        "vswp d1, d20\n"
+        "vswp d7, d26\n"
+
+        "vld1.32 {q2}, [r10], %[stride]\n"
+        "vld1.32 {q9}, [r10], %[stride]\n"
+        "vld1.32 {q12}, [r10], %[stride]\n"
+        "vld1.32 {q15}, [r10], %[stride]\n"
+
+        "vtrn.32 d2, d16\n"
+        "vtrn.32 d3, d17\n"
+        "vtrn.32 d22, d28\n"
+        "vtrn.32 d23, d29\n"
+
+        "vswp d3, d22\n"
+        "vswp d17, d28\n"
+
+        "vtrn.32 d4, d18\n"
+        "vtrn.32 d5, d19\n"
+        "vtrn.32 d24, d30\n"
+        "vtrn.32 d25, d31\n"
+
+        "vswp d5, d24\n"
+        "vswp d19, d30\n"
+
+        "vst1.32 {q0, q1}, [r12]!\n"
+        "vst1.32 {q2, q3}, [r12]!\n"
+        "vst1.32 {q8, q9}, [r12]!\n"
+        "vst1.32 {q10, q11}, [r12]!\n"
+        "vst1.32 {q12, q13}, [r12]!\n"
+        "vst1.32 {q14, q15}, [r12]!\n"
+
+        :
+        : [ dst_c ] "r"(dst_c), [ src_c ] "r"(src_c), [ stride ] "r"(stride)
+        : "r10", "r12", "q0", "q1", "q2", "q3", "q8", "q9", "q10", "q11", "q12", "q13", "q14", "q15");
+#else
+      for (int tr = 0; tr < C12NUM; tr++) {
+        for (int tc = 0; tc < C4NUM; tc++) {
+          dst_c[tc * C12NUM + tr] = src_c[tr * lead + tc];
+        }
       }
-      for (; k < K; ++k) {
-        sum += alpha * mat_a[i * lda + k] * mat_b[j * ldb + k];
+#endif
+    }
+    for (; ci < col; ci++) {
+      const float *src_c = src_r + ci;
+      float *dst_c = dst_r + ci * C12NUM;
+      for (size_t i = 0; i < C12NUM; i++) {
+        dst_c[i] = src_c[i * lead];
       }
-      mat_c[i * ldc + j] += sum;
     }
+    src_r += C12NUM * lead;
+    dst_r += C12NUM * col;
   }
+
+  for (; ri < row; ri++) {
+    for (size_t i = 0; i < col; i++) {
+      dst_r[i * C12NUM] = src_r[i];
+    }
+    src_r += lead;
+    dst_r += 1;
+  }
+
+  for (; ri < row_up_12; ri++) {
+    for (size_t i = 0; i < col; i++) {
+      dst_r[i * C12NUM] = 0;
+    }
+    dst_r += 1;
+  }
+  return;
 }
+#endif
+
+static void RowMajor2Col8MajorStride(const float *src_ptr, float *dst_ptr, size_t row, size_t col, int lead) {
+  size_t row8 = row / C8NUM * C8NUM;
+#ifdef ENABLE_ARM64
+  size_t col_skip = col / C8NUM * C8NUM;
+  int skip_size = C8NUM;
+#else
+  size_t col_skip = col / C4NUM * C4NUM;
+  int skip_size = C4NUM;
+#endif
+  const float *src_r = src_ptr;
+  float *dst_r = dst_ptr;
+
+  size_t ri = 0;
+  for (; ri < row8; ri += C8NUM) {
+    size_t ci = 0;
+    for (; ci < col_skip; ci += skip_size) {
+      const float *src_c = src_r + ci;
+      float *dst_c = dst_r + ci * C8NUM;
+
+#ifdef ENABLE_ARM64
+      /* 8x8 row-major to col-major */
+      size_t stride = lead * sizeof(float);
+      asm volatile(
+        "mov x10, %[src_c]\n"
+        "mov x11, %[dst_c]\n"
+
+        "ld1 {v0.4s, v1.4s}, [x10], %[stride]\n"
+        "ld1 {v2.4s, v3.4s}, [x10], %[stride]\n"
+        "ld1 {v4.4s, v5.4s}, [x10], %[stride]\n"
+        "ld1 {v6.4s, v7.4s}, [x10], %[stride]\n"
+
+        "zip1 v8.4s, v0.4s, v2.4s\n"
+        "zip2 v9.4s, v0.4s, v2.4s\n"
+        "zip1 v10.4s, v4.4s, v6.4s\n"
+        "zip2 v11.4s, v4.4s, v6.4s\n"
+
+        "ld1 {v16.4s, v17.4s}, [x10], %[stride]\n"
+        "ld1 {v18.4s, v19.4s}, [x10], %[stride]\n"
+        "ld1 {v20.4s, v21.4s}, [x10], %[stride]\n"
+        "ld1 {v22.4s, v23.4s}, [x10], %[stride]\n"
+
+        "zip1 v12.4s, v1.4s, v3.4s\n"
+        "zip2 v13.4s, v1.4s, v3.4s\n"
+        "zip1 v14.4s, v5.4s, v7.4s\n"
+        "zip2 v15.4s, v5.4s, v7.4s\n"
+
+        "trn1 v0.2d, v8.2d, v10.2d\n"
+        "trn2 v1.2d, v8.2d, v10.2d\n"
+        "trn1 v2.2d, v9.2d, v11.2d\n"
+        "trn2 v3.2d, v9.2d, v11.2d\n"
+
+        "zip1 v24.4s, v16.4s, v18.4s\n"
+        "zip2 v25.4s, v16.4s, v18.4s\n"
+        "zip1 v26.4s, v20.4s, v22.4s\n"
+        "zip2 v27.4s, v20.4s, v22.4s\n"
+
+        "trn1 v4.2d, v12.2d, v14.2d\n"
+        "trn2 v5.2d, v12.2d, v14.2d\n"
+        "trn1 v6.2d, v13.2d, v15.2d\n"
+        "trn2 v7.2d, v13.2d, v15.2d\n"
+
+        "zip1 v28.4s, v17.4s, v19.4s\n"
+        "zip2 v29.4s, v17.4s, v19.4s\n"
+        "zip1 v30.4s, v21.4s, v23.4s\n"
+        "zip2 v31.4s, v21.4s, v23.4s\n"
+
+        "trn1 v16.2d, v24.2d, v26.2d\n"
+        "trn2 v17.2d, v24.2d, v26.2d\n"
+        "trn1 v18.2d, v25.2d, v27.2d\n"
+        "trn2 v19.2d, v25.2d, v27.2d\n"
+
+        "trn1 v20.2d, v28.2d, v30.2d\n"
+        "trn2 v21.2d, v28.2d, v30.2d\n"
+        "trn1 v22.2d, v29.2d, v31.2d\n"
+        "trn2 v23.2d, v29.2d, v31.2d\n"
+
+        "st1 {v0.4s}, [x11], #16\n"
+        "st1 {v16.4s}, [x11], #16\n"
+        "st1 {v1.4s}, [x11], #16\n"
+        "st1 {v17.4s}, [x11], #16\n"
+        "st1 {v2.4s}, [x11], #16\n"
+        "st1 {v18.4s}, [x11], #16\n"
+        "st1 {v3.4s}, [x11], #16\n"
+        "st1 {v19.4s}, [x11], #16\n"
+        "st1 {v4.4s}, [x11], #16\n"
+        "st1 {v20.4s}, [x11], #16\n"
+        "st1 {v5.4s}, [x11], #16\n"
+        "st1 {v21.4s}, [x11], #16\n"
+        "st1 {v6.4s}, [x11], #16\n"
+        "st1 {v22.4s}, [x11], #16\n"
+        "st1 {v7.4s}, [x11], #16\n"
+        "st1 {v23.4s}, [x11], #16\n"
+
+        :
+        : [ dst_c ] "r"(dst_c), [ src_c ] "r"(src_c), [ stride ] "r"(stride)
+        : "x10", "x11", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14",
+          "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29",
+          "v30", "v31");
+#elif ENABLE_ARM32
+      /* 8x4 row-major to col-major */
+      size_t stride = col * sizeof(float);
+      asm volatile(
+        "mov r10, %[src_c]\n"
+        "mov r11, %[dst_c]\n"
+
+        "vld1.32 {q0}, [r10], %[stride]\n"
+        "vld1.32 {q2}, [r10], %[stride]\n"
+        "vld1.32 {q4}, [r10], %[stride]\n"
+        "vld1.32 {q6}, [r10], %[stride]\n"
+
+        "vtrn.32 d0, d4\n"
+        "vtrn.32 d1, d5\n"
+        "vtrn.32 d8, d12\n"
+        "vtrn.32 d9, d13\n"
+
+        "vld1.32 {q1}, [r10], %[stride]\n"
+        "vld1.32 {q3}, [r10], %[stride]\n"
+        "vld1.32 {q5}, [r10], %[stride]\n"
+        "vld1.32 {q7}, [r10], %[stride]\n"
 
-static void gemm_trana_not_tranb(int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b, int ldb,
-                                 float *mat_c, int ldc) {
-  const int block_size = 4;
-  int block_mod = N % block_size;
-  int block_c4 = N - block_mod;
-
-  int i, j, k;
-  for (i = 0; i < M; ++i) {
-    for (k = 0; k < K; ++k) {
-      float a = alpha * mat_a[k * lda + i];
-      for (j = 0; j < block_c4; j += block_size) {
-        float *b = &mat_b[k * ldb + j];
-        float *c = &mat_c[i * ldc + j];
-        c[0] += a * b[0];
-        c[1] += a * b[1];
-        c[2] += a * b[2];
-        c[3] += a * b[3];
+        "vswp d1, d8\n"
+        "vswp d5, d12\n"
+
+        "vtrn.32 d2, d6\n"
+        "vtrn.32 d3, d7\n"
+        "vtrn.32 d10, d14\n"
+        "vtrn.32 d11, d15\n"
+
+        "vswp d3, d10\n"
+        "vswp d7, d14\n"
+
+        "vst1.32 {q0, q1}, [r11]!\n"
+        "vst1.32 {q2, q3}, [r11]!\n"
+        "vst1.32 {q4, q5}, [r11]!\n"
+        "vst1.32 {q6, q7}, [r11]!\n"
+
+        :
+        : [ dst_c ] "r"(dst_c), [ src_c ] "r"(src_c), [ stride ] "r"(stride)
+        : "r10", "r11", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7");
+#else
+      for (int tr = 0; tr < 8; tr++) {
+        for (int tc = 0; tc < 4; tc++) {
+          dst_c[tc * 8 + tr] = src_c[tr * lead + tc];
+        }
       }
-      for (; j < N; ++j) {
-        mat_c[i * ldc + j] += a * mat_b[k * ldb + j];
+#endif
+    }
+    for (; ci < col; ci++) {
+      const float *src_c = src_r + ci;
+      float *dst_c = dst_r + ci * C8NUM;
+      for (size_t i = 0; i < C8NUM; i++) {
+        dst_c[i] = src_c[i * lead];
       }
     }
+    src_r += C8NUM * lead;
+    dst_r += C8NUM * col;
   }
+  for (; ri < row; ri++) {
+    for (size_t i = 0; i < col; i++) {
+      dst_r[i * C8NUM] = src_r[i];
+    }
+    src_r += lead;
+    dst_r += 1;
+  }
+  return;
 }
+#ifdef ENABLE_ARM32
+static void RowMajor2Col4MajorStride(const float *src_ptr, float *dst_ptr, size_t row, size_t col, int lead) {
+  size_t row8 = row / C4NUM * C4NUM;
+  size_t col4 = col / C4NUM * C4NUM;
+  const float *src_r = src_ptr;
+  float *dst_r = dst_ptr;
+
+  size_t ri = 0;
+  for (; ri < row8; ri += C4NUM) {
+    size_t ci = 0;
+    for (; ci < col4; ci += C4NUM) {
+      const float *src_c = src_r + ci;
+      float *dst_c = dst_r + ci * C4NUM;
+
+      /* 4x4 row-major to col-major */
+#ifdef ENABLE_ARM32
+      size_t stride = col * 4;
+      asm volatile(
+        "mov r10, %[src_c]\n"
+        "mov r12, %[dst_c]\n"
+
+        "vld1.32 {q0}, [r10], %[stride]\n"
+        "vld1.32 {q1}, [r10], %[stride]\n"
+        "vld1.32 {q2}, [r10], %[stride]\n"
+        "vld1.32 {q3}, [r10], %[stride]\n"
+
+        "vtrn.32 d0, d2\n"
+        "vtrn.32 d1, d3\n"
+        "vtrn.32 d4, d6\n"
+        "vtrn.32 d5, d7\n"
 
-static void gemm_trana_tranb(int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b, int ldb,
-                             float *mat_c, int ldc) {
-  int i, j, k;
-  const int block_size = 4;
-  int k_block_mod = K % block_size;
-  int k_block_c4 = K - k_block_mod;
-
-  int m_block_mod = M % block_size;
-  int m_block_c4 = M - m_block_mod;
-
-  for (i = 0; i < m_block_c4; i += block_size) {
-    for (j = 0; j < N; ++j) {
-      float sum0 = 0;
-      float sum1 = 0;
-      float sum2 = 0;
-      float sum3 = 0;
-
-      for (k = 0; k < k_block_c4; k += block_size) {
-        float *b = &mat_b[j * ldb + k];
-        sum0 += alpha * mat_a[i + k * lda] * b[0];
-        sum0 += alpha * mat_a[i + (k + 1) * lda] * b[1];
-        sum0 += alpha * mat_a[i + (k + 2) * lda] * b[2];
-        sum0 += alpha * mat_a[i + (k + 3) * lda] * b[3];
-
-        sum1 += alpha * mat_a[i + 1 + k * lda] * b[0];
-        sum1 += alpha * mat_a[i + 1 + (k + 1) * lda] * b[1];
-        sum1 += alpha * mat_a[i + 1 + (k + 2) * lda] * b[2];
-        sum1 += alpha * mat_a[i + 1 + (k + 3) * lda] * b[3];
-
-        sum2 += alpha * mat_a[i + 2 + k * lda] * b[0];
-        sum2 += alpha * mat_a[i + 2 + (k + 1) * lda] * b[1];
-        sum2 += alpha * mat_a[i + 2 + (k + 2) * lda] * b[2];
-        sum2 += alpha * mat_a[i + 2 + (k + 3) * lda] * b[3];
-
-        sum3 += alpha * mat_a[i + 3 + k * lda] * b[0];
-        sum3 += alpha * mat_a[i + 3 + (k + 1) * lda] * b[1];
-        sum3 += alpha * mat_a[i + 3 + (k + 2) * lda] * b[2];
-        sum3 += alpha * mat_a[i + 3 + (k + 3) * lda] * b[3];
+        "vswp d1, d4\n"
+        "vswp d3, d6\n"
+
+        "vst1.32 {q0}, [r12]!\n"
+        "vst1.32 {q1}, [r12]!\n"
+        "vst1.32 {q2}, [r12]!\n"
+        "vst1.32 {q3}, [r12]!\n"
+
+        :
+        : [ dst_c ] "r"(dst_c), [ src_c ] "r"(src_c), [ stride ] "r"(stride)
+        : "r10", "r12", "q0", "q1", "q2", "q3");
+#else
+      for (int tr = 0; tr < C4NUM; tr++) {
+        for (int tc = 0; tc < C4NUM; tc++) {
+          dst_c[tc * C4NUM + tr] = src_c[tr * lead + tc];
+        }
       }
-      for (; k < K; ++k) {
-        float *b = &mat_b[j * ldb + k];
-        sum0 += alpha * mat_a[i + (k * lda)] * b[0];
-        sum1 += alpha * mat_a[i + 1 + (k * lda)] * b[0];
-        sum2 += alpha * mat_a[i + 2 + (k * lda)] * b[0];
-        sum3 += alpha * mat_a[i + 3 + (k * lda)] * b[0];
+#endif
+    }
+    for (; ci < col; ci++) {
+      const float *src_c = src_r + ci;
+      float *dst_c = dst_r + ci * C4NUM;
+      for (size_t i = 0; i < C4NUM; i++) {
+        dst_c[i] = src_c[i * lead];
       }
-      mat_c[i * ldc + j] += sum0;
-      mat_c[(i + 1) * ldc + j] += sum1;
-      mat_c[(i + 2) * ldc + j] += sum2;
-      mat_c[(i + 3) * ldc + j] += sum3;
     }
+    src_r += C4NUM * col;
+    dst_r += C4NUM * col;
   }
-  // no more block of 4x4
-  for (; i < M; ++i) {
-    for (j = 0; j < N; ++j) {
-      float sum = 0;
-      for (k = 0; k < K; ++k) {
-        sum += alpha * mat_a[i + k * lda] * mat_b[k + j * ldb];
-      }
-      mat_c[i * ldc + j] += sum;
+  for (; ri < row; ri++) {
+    for (size_t i = 0; i < col; i++) {
+      dst_r[i * C4NUM] = src_r[i];
     }
+    src_r += lead;
+    dst_r += 1;
   }
+  return;
+}
+#endif
+
+void GemmMatmul(int ta, int tb, int M, int N, int K, float alpha, const float *mat_a, int lda, const float *mat_b,
+                int ldb, float beta, float *mat_c, int ldc, float *workspace) {
+  GemmCb gcb;
+  gcb.atype = ActType_No;
+  gcb.ca = 0;
+  gcb.cb = 0;
+  gcb.bias = NULL;
+  GemmMatmulPlus(ta, tb, M, N, K, alpha, mat_a, lda, mat_b, ldb, beta, mat_c, ldc, workspace, &gcb);
 }
 
-// mat_c = alpha*op( mat_a )*op( mat_b ) + beta*C
-// M - number of rows of matrix a
-// N - number of cols of matrix b
-// K - number of cols of matrix a
-// lda - fast dim of matrix a
-// ldb - fast dim of matrix b
-// ldc - fast dim of matrix c
-void gemm(int transpose_a, int transpose_b, int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b,
-          int ldb, float beta, float *mat_c, int ldc) {
-  if (beta >= 0.f && beta <= 0.f) {
-    memset(mat_c, 0, M * N * sizeof(float));
-  } else if (beta < 1.f || beta > 1.f) {
-    const int block_size = 4;
-    const int size = M * N;
-    int block_mod = size % block_size;
-    int block_c4 = size - block_mod;
-    int i;
-    for (i = 0; i < block_c4; i += block_size) {
-      float *c = &mat_c[i];
-      c[0] *= beta;
-      c[1] *= beta;
-      c[2] *= beta;
-      c[3] *= beta;
-    }
-    for (; i < size; ++i) {
-      mat_c[i] *= beta;
+void GemmMatmulPlus(int ta, int tb, int M, int N, int K, float alpha, const float *mat_a, int lda, const float *mat_b,
+                    int ldb, float beta, float *mat_c, int ldc, float *workspace, GemmCb *gcb) {
+#ifdef ENABLE_ARM32
+  const int num = C4NUM;
+#else
+  const int num = C12NUM;
+#endif
+  float *output = mat_c;
+  float *fworkspace = workspace;
+  int incremental = (beta < 0.f) || (beta > 0.f);
+  float *mat_a_input = (float *)mat_a;
+  float *mat_b_input = (float *)mat_b;
+
+#ifdef ENABLE_ARM32
+  if (!gcb->ca) {
+    mat_a_input = fworkspace;
+    fworkspace += MatSize(M, K, num);
+    if (ta) {
+      RowMajor2Row4MajorStride(mat_a, mat_a_input, K, M, lda);
+    } else {
+      RowMajor2Col4MajorStride(mat_a, mat_a_input, M, K, lda);
     }
   }
-  if (transpose_a && transpose_b) {
-    gemm_trana_tranb(M, N, K, alpha, mat_a, lda, mat_b, ldb, mat_c, ldc);
-  } else if (!transpose_a && !transpose_b) {
-    gemm_not_trana_not_tranb(M, N, K, alpha, mat_a, lda, mat_b, ldb, mat_c, ldc);
-  } else if (!transpose_a && transpose_b) {
-    gemm_not_trana_tranb(M, N, K, alpha, mat_a, lda, mat_b, ldb, mat_c, ldc);
-  } else {
-    gemm_trana_not_tranb(M, N, K, alpha, mat_a, lda, mat_b, ldb, mat_c, ldc);
+#else
+  if (!gcb->ca) {
+    mat_a_input = fworkspace;
+    fworkspace += MatSize(M, K, num);
+    if (ta) {
+      RowMajor2Row12MajorStride(mat_a, mat_a_input, K, M, lda);
+    } else {
+      RowMajor2Col12MajorStride(mat_a, mat_a_input, M, K, lda);
+    }
+  }
+#endif
+  if (!gcb->cb) {
+    mat_b_input = fworkspace;
+    fworkspace += MatSize(N, K, C8NUM);
+    if (tb) {
+      RowMajor2Col8MajorStride(mat_b, mat_b_input, N, K, ldb);
+    } else {
+      RowMajor2Row8MajorStride(mat_b, mat_b_input, K, N, ldb);
+    }
   }
+  if (incremental) output = fworkspace;
+  MatMulOpt(mat_a_input, mat_b_input, output, gcb->bias, gcb->atype, K, M, N, ldc, OutType_Nhwc);
+  if (incremental) addv(output, mat_c, beta, M, N, ldc);
+  gcb->mat_a = mat_a_input;
+  gcb->mat_b = mat_b_input;
 }
diff --git a/mindspore/lite/nnacl/fp32_grad/gemm.h b/mindspore/lite/nnacl/fp32_grad/gemm.h
index ab7007675c..b1da9b8288 100644
--- a/mindspore/lite/nnacl/fp32_grad/gemm.h
+++ b/mindspore/lite/nnacl/fp32_grad/gemm.h
@@ -17,11 +17,26 @@
 #ifndef MINDSPORE_LITE_NNACL_FP32_GRAD_GEMM_H_
 #define MINDSPORE_LITE_NNACL_FP32_GRAD_GEMM_H_
 
+#include <stdlib.h>
+#include "nnacl/op_base.h"
 #ifdef __cplusplus
 extern "C" {
 #endif
-void gemm(int transpose_a, int transpose_b, int M, int N, int K, float alpha, float *mat_a, int lda, float *mat_b,
-          int ldb, float beta, float *mat_c, int ldc);
+typedef struct {
+  int ca;
+  int cb;
+  ActType atype;
+  float *bias;
+  float *mat_a;
+  float *mat_b;
+} GemmCb;
+
+void GemmMatmulPlus(int ta, int tb, int M, int N, int K, float alpha, const float *mat_a, int lda, const float *mat_b,
+                    int ldb, float beta, float *mat_c, int ldc, float *workspace, GemmCb *cb);
+void GemmMatmul(int ta, int tb, int M, int N, int K, float alpha, const float *mat_a, int lda, const float *mat_b,
+                int ldb, float beta, float *mat_c, int ldc, float *workspace);
+int MatSize(int row, int col, int round);
+int MatSizeTotal(int row, int col, int deep, int inc);
 #ifdef __cplusplus
 }
 #endif
diff --git a/mindspore/lite/nnacl/fp32_grad/pack_ext.c b/mindspore/lite/nnacl/fp32_grad/pack_ext.c
index 645ad8dc2d..1406656ecb 100644
--- a/mindspore/lite/nnacl/fp32_grad/pack_ext.c
+++ b/mindspore/lite/nnacl/fp32_grad/pack_ext.c
@@ -16,10 +16,11 @@
 
 #include <string.h>
 #include "nnacl/fp32_grad/pack_ext.h"
+#include "nnacl/pack.h"
 
 static int is_a_ge_zero_and_a_lt_b(int a, int b) { return (unsigned)(a) < (unsigned)(b); }
 
-void im2col_hwc(const float *in_data, float *data_col, ConvParameter *conv_param) {
+void rolling_im2col_hwc(const float *in_data, float *data_col, const ConvParameter *conv_param, int rows, int start) {
   const int pad_left = conv_param->pad_l_;
   const int pad_up = conv_param->pad_u_;
 
@@ -35,42 +36,42 @@ void im2col_hwc(const float *in_data, float *data_col, ConvParameter *conv_param
   const int in_height = conv_param->input_h_;
   const int in_width = conv_param->input_w_;
 
-  const int output_h = conv_param->output_h_;
   const int output_w = conv_param->output_w_;
 
   const int channels = conv_param->input_channel_ / conv_param->group_;
   const int tot_channels = conv_param->input_channel_;
 
-  int kernel_row, kernel_col, output_rows, output_col;
-
-  int row_stride_offset = 0;
+  int kernel_row, kernel_col;
 
-  for (output_rows = output_h; output_rows; output_rows--) {
-    int col_stride_offset = 0;
-    for (output_col = output_w; output_col; output_col--) {
-      for (kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
-        int input_row = -pad_up + kernel_row * dilation_h + row_stride_offset;
-        for (kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
-          int input_col = -pad_left + kernel_col * dilation_w + col_stride_offset;
+  for (int i = 0; i < rows; i++) {
+    int block_start = start + i;
+    int input_h = block_start / output_w * stride_h;
+    int input_w = block_start % output_w * stride_w;
+    for (kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
+      int input_row = -pad_up + kernel_row * dilation_h + input_h;
+      for (kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
+        int input_col = -pad_left + kernel_col * dilation_w + input_w;
 
-          if (is_a_ge_zero_and_a_lt_b(input_row, in_height) && is_a_ge_zero_and_a_lt_b(input_col, in_width)) {
-            const int offset = (input_row * in_width + input_col) * tot_channels;
-            memcpy(data_col, in_data + offset, sizeof(float) * channels);
-            data_col += channels;
-          } else {
-            memset(data_col, 0, sizeof(float) * channels);
-            data_col += channels;
-          }
+        if (is_a_ge_zero_and_a_lt_b(input_row, in_height) && is_a_ge_zero_and_a_lt_b(input_col, in_width)) {
+          const int offset = (input_row * in_width + input_col) * tot_channels;
+          memcpy(data_col, in_data + offset, sizeof(float) * channels);
+          data_col += channels;
+        } else {
+          memset(data_col, 0, sizeof(float) * channels);
+          data_col += channels;
         }
       }
-      col_stride_offset += stride_w;
     }
-    row_stride_offset += stride_h;
   }
 }
 
+void RollingIm2ColPackUnitFp32(const float *input_data, const ConvParameter *conv_param, float *packed_input,
+                               int real_cal_num, int block_index) {
+  rolling_im2col_hwc(input_data, packed_input, conv_param, real_cal_num, block_index);
+}
+
 // output matrix is (kernel_h*kernel_w*channels)X(output_h*output_w)
-void im2row_hwc(const float *in_data, float *data_row, ConvParameter *conv_param, bool transpose) {
+void im2row_hwc(const float *in_data, float *data_row, const ConvParameter *conv_param, bool transpose) {
   const int pad_left = conv_param->pad_l_;
   const int pad_up = conv_param->pad_u_;
 
@@ -150,7 +151,56 @@ void im2row_hwc(const float *in_data, float *data_row, ConvParameter *conv_param
   }
 }
 
-void col2im_hwc(const float *data_col, float *data_im, ConvParameter *conv_param) {
+void rolling_im2row_hwc(const float *in_data, float *data_row, const ConvParameter *conv_param, int rows, int start) {
+  const int pad_left = conv_param->pad_l_;
+  const int pad_up = conv_param->pad_u_;
+
+  const int stride_h = conv_param->stride_h_;
+  const int stride_w = conv_param->stride_w_;
+
+  const int dilation_h = conv_param->dilation_h_;
+  const int dilation_w = conv_param->dilation_w_;
+
+  const int kernel_h = conv_param->kernel_h_;
+  const int kernel_w = conv_param->kernel_w_;
+
+  const int in_height = conv_param->output_h_;
+  const int in_width = conv_param->output_w_;
+
+  const int output_w = conv_param->input_w_;
+
+  const int tot_channels = conv_param->output_channel_;
+  const int channels = tot_channels / conv_param->group_;
+  int channel, kernel_row, kernel_col, output_rows, output_col;
+  for (channel = 0; channel < channels; channel++) {
+    for (kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
+      for (kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
+        for (output_rows = start; output_rows < start + rows; output_rows++) {
+          int input_row = -pad_up + kernel_row * dilation_h + output_rows * stride_h;
+          if (!is_a_ge_zero_and_a_lt_b(input_row, in_height)) {
+            for (output_col = output_w; output_col; output_col--) {
+              *(data_row++) = 0;
+            }
+          } else {
+            int input_col = -pad_left + kernel_col * dilation_w;
+            for (output_col = output_w; output_col; output_col--) {
+              if (is_a_ge_zero_and_a_lt_b(input_col, in_width)) {
+                const int offset = (input_row * in_width + input_col) * tot_channels + channel;
+                *(data_row++) = in_data[offset];
+              } else {
+                *(data_row++) = 0;
+              }
+              input_col += stride_w;
+            }
+          }
+          // input_row += stride_h;
+        }
+      }
+    }
+  }
+}
+
+void col2im_hwc(const float *data_col, float *data_im, const ConvParameter *conv_param) {
   const int pad_left = conv_param->pad_l_;
   const int pad_up = conv_param->pad_u_;
 
@@ -198,3 +248,52 @@ void col2im_hwc(const float *data_col, float *data_im, ConvParameter *conv_param
     row_stride_offset += stride_h;
   }
 }
+
+void rolling_col2im_hwc(const float *data_col, float *data_im, const ConvParameter *conv_param, int rows, int start) {
+  const int pad_left = conv_param->pad_l_;
+  const int pad_up = conv_param->pad_u_;
+
+  const int stride_h = conv_param->stride_h_;
+  const int stride_w = conv_param->stride_w_;
+
+  const int dilation_h = conv_param->dilation_h_;
+  const int dilation_w = conv_param->dilation_w_;
+
+  const int kernel_h = conv_param->kernel_h_;
+  const int kernel_w = conv_param->kernel_w_;
+
+  const int in_height = conv_param->input_h_;
+  const int in_width = conv_param->input_w_;
+
+  const int output_w = conv_param->output_w_;
+  const int channels = conv_param->input_channel_ / conv_param->group_;
+  const int tot_channels = conv_param->input_channel_;
+
+  int kernel_row, kernel_col;
+
+  for (int r = 0; r < rows; r++) {
+    int output_col = (start + r) % output_w;
+    int output_row = (start + r) / output_w;
+    int row_stride_offset = output_row * stride_h;
+    int col_stride_offset = output_col * stride_w;
+
+    // for (output_col = 0; output_col < output_w; output_col++)
+    {
+      for (kernel_row = 0; kernel_row < kernel_h; kernel_row++) {
+        int input_row = -pad_up + kernel_row * dilation_h + row_stride_offset;
+        for (kernel_col = 0; kernel_col < kernel_w; kernel_col++) {
+          int input_col = -pad_left + kernel_col * dilation_w + col_stride_offset;
+
+          if (is_a_ge_zero_and_a_lt_b(input_row, in_height) && is_a_ge_zero_and_a_lt_b(input_col, in_width)) {
+            int offset = (input_row * in_width + input_col) * tot_channels;
+            float *data_im_ptr = &data_im[offset];
+            for (int i = 0; i < channels; i++) {
+              data_im_ptr[i] += data_col[i];
+            }
+          }
+          data_col += channels;
+        }
+      }
+    }
+  }
+}
diff --git a/mindspore/lite/nnacl/fp32_grad/pack_ext.h b/mindspore/lite/nnacl/fp32_grad/pack_ext.h
index aa5f33faa7..c2095a7503 100644
--- a/mindspore/lite/nnacl/fp32_grad/pack_ext.h
+++ b/mindspore/lite/nnacl/fp32_grad/pack_ext.h
@@ -17,14 +17,18 @@
 #ifndef MINDSPORE_LITE_NNACL_FP32_GRAD_PACK_EXT_H_
 #define MINDSPORE_LITE_NNACL_FP32_GRAD_PACK_EXT_H_
 
+#include <stddef.h>
 #include "nnacl/conv_parameter.h"
 
 #ifdef __cplusplus
 extern "C" {
 #endif
-void im2col_hwc(const float *in_data, float *data_col, ConvParameter *conv_param);
-void im2row_hwc(const float *in_data, float *data_row, ConvParameter *conv_param, bool transpose);
-void col2im_hwc(const float *data_col, float *data_im, ConvParameter *conv_param);
+
+void RollingIm2ColPackUnitFp32(const float *input_data, const ConvParameter *conv_param, float *packed_input,
+                               int real_cal_num, int block_index);
+void rolling_im2col_hwc(const float *in_data, float *data_col, const ConvParameter *conv_param, int rows, int start);
+void rolling_im2row_hwc(const float *in_data, float *data_row, const ConvParameter *conv_param, int rows, int start);
+void rolling_col2im_hwc(const float *data_col, float *data_im, const ConvParameter *conv_param, int rows, int start);
 #ifdef __cplusplus
 }
 #endif
diff --git a/mindspore/lite/nnacl/fp32_grad/pooling_grad.c b/mindspore/lite/nnacl/fp32_grad/pooling_grad.c
index ddf1b197cc..a5d702cf38 100644
--- a/mindspore/lite/nnacl/fp32_grad/pooling_grad.c
+++ b/mindspore/lite/nnacl/fp32_grad/pooling_grad.c
@@ -14,6 +14,7 @@
  * limitations under the License.
  */
 #include <stdint.h>
+#include <string.h>
 #include <float.h>
 #include "nnacl/fp32_grad/pooling_grad.h"
 
@@ -31,8 +32,7 @@ void AvgPoolingGrad(const float *input_ptr, float *output_ptr, PoolingParameter
   int output_h = pooling_param->output_h_;
   int output_batch = pooling_param->output_batch_;
 
-  for (int i = 0; i < in_h * in_w * channel * output_batch; i++) output_ptr[i] = 0.0;
-
+  memset(output_ptr, 0, in_h * in_w * channel * output_batch * sizeof(float));
   float kk = (float)(win_h * win_w);
   for (uint16_t ib = 0; ib < output_batch; ib++) {
     float *out = &output_ptr[(ib * in_h * in_w * channel)];
@@ -77,8 +77,7 @@ void MaxPoolingGrad(const float *input_ptr, const float *dx_ptr, const float *dy
   int output_h = pooling_param->output_h_;
   int output_batch = pooling_param->output_batch_;
 
-  for (int i = 0; i < in_h * in_w * channel * output_batch; i++) output_ptr[i] = 0.0;
-
+  memset(output_ptr, 0, in_h * in_w * channel * output_batch * sizeof(float));
   for (uint16_t ib = 0; ib < output_batch; ib++) {
     float *out = &output_ptr[(ib * in_h * in_w * channel)];
     const float *inPtr = (const float *)(&input_ptr[(ib * in_h * in_w * channel)]);
diff --git a/mindspore/lite/nnacl/fp32_grad/reduce_grad.c b/mindspore/lite/nnacl/fp32_grad/reduce_grad.c
index 6963969817..d682342a7a 100644
--- a/mindspore/lite/nnacl/fp32_grad/reduce_grad.c
+++ b/mindspore/lite/nnacl/fp32_grad/reduce_grad.c
@@ -15,50 +15,7 @@
  */
 #include <string.h>
 #include "nnacl/fp32_grad/reduce_grad.h"
-
-static inline int NextIndex(const int num_dims, const int *dims, int *current) {
-  int carry = 1;
-  for (int idx = num_dims - 1; idx >= 0; --idx) {
-    int current_val = current[idx] + carry;
-    if (dims[idx] == current_val) {
-      current[idx] = 0;
-    } else {
-      current[idx] = current_val;
-      carry = 0;
-      break;
-    }
-  }
-  return (carry == 0);
-}
-
-static inline size_t GetInputOffset(const int num_dims, const int *dims, const int *iter) {
-  size_t offset = 0;
-  for (int idx = 0; idx < num_dims; ++idx) {
-    offset = offset * (size_t)(dims[idx]) + (size_t)(iter[idx]);
-  }
-
-  return offset;
-}
-
-static inline size_t GetOutputOffset(const int num_dims, const int *dims, const int *iter, const int num_axis,
-                                     const int *axes) {
-  size_t offset = 0;
-  for (int idx = 0; idx < num_dims; ++idx) {
-    // if we need to skip this axis
-    int is_axis = 0;
-    for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) {
-      if (idx == axes[axis_idx]) {
-        is_axis = 1;
-        break;
-      }
-    }
-
-    if (!is_axis) {
-      offset = offset * (size_t)(dims[idx]) + (size_t)(iter[idx]);
-    }
-  }
-  return offset;
-}
+#include "nnacl/fp32_grad/utils.h"
 
 void ReduceMeanByAxes(const float *input_data, int *input_iter, const int *input_dims, int input_num_dims,
                       const int *axes, int num_axes, float *output_data, const int *output_dims, int output_num_dims) {
@@ -111,7 +68,7 @@ void ReduceSumByAxes(const float *input, const int *input_dims, float *output, c
     return;
   }
 
-  for (int idx = 0; idx < num_outputs; ++idx) output[idx] = 0;  // zero output
+  memset(output, 0, num_outputs * sizeof(float));  // zero output
 
   int input_iter[8] = {0};
   int axes[5] = {0};
diff --git a/mindspore/lite/nnacl/fp32_grad/softmax_grad.c b/mindspore/lite/nnacl/fp32_grad/softmax_grad.c
index 21bcc14188..c863e705dd 100644
--- a/mindspore/lite/nnacl/fp32_grad/softmax_grad.c
+++ b/mindspore/lite/nnacl/fp32_grad/softmax_grad.c
@@ -41,7 +41,6 @@ void SoftmaxGrad(const float *input_ptr, const float *yt_ptr, float *output_ptr,
 
   const int M = input_shape[axis];
   const int N = inner_size;
-  const int K = 1;
   for (int i = 0; i < outter_size; i++) {
     int outter_offset = i * dim;
     memset(sum_data, 0.0f, inner_size * sizeof(float));
@@ -52,7 +51,14 @@ void SoftmaxGrad(const float *input_ptr, const float *yt_ptr, float *output_ptr,
         sum_data[k] += output_ptr[offset] * input_ptr[offset];
       }
     }
-    gemm(0, 0, M, N, K, -1, sum_mul, K, sum_data, N, 1, &output_ptr[outter_offset], N);
+    for (int k = 0; k < M; ++k) {
+      float a = -sum_mul[k];
+      for (int j = 0; j < N; ++j) {
+        *(output_ptr + outter_offset + k * N + j) += a * sum_data[j];
+      }
+    }
+
+    // gemm(0, 0, M, N, K, -1, sum_mul, K, sum_data, N, 1, &output_ptr[outter_offset], N);
   }
 
   for (int i = 0; i < ele_size; i++) {
diff --git a/mindspore/lite/nnacl/fp32_grad/utils.h b/mindspore/lite/nnacl/fp32_grad/utils.h
new file mode 100644
index 0000000000..f7895aa917
--- /dev/null
+++ b/mindspore/lite/nnacl/fp32_grad/utils.h
@@ -0,0 +1,72 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_LITE_NNACL_FP32_GRAD_UTILS_H_
+#define MINDSPORE_LITE_NNACL_FP32_GRAD_UTILS_H_
+
+#include "nnacl/op_base.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static inline size_t GetInputOffset(int num_dims, const int *dims, const int *iter) {
+  size_t offset = 0;
+  for (int idx = 0; idx < num_dims; ++idx) {
+    offset = offset * (size_t)(dims[idx]) + (size_t)(iter[idx]);
+  }
+
+  return offset;
+}
+
+static inline size_t GetOutputOffset(int num_dims, const int *dims, const int *iter, int num_axis, const int *axes) {
+  size_t offset = 0;
+  for (int idx = 0; idx < num_dims; ++idx) {
+    // if we need to skip this axis
+    int is_axis = 0;
+    for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) {
+      if (idx == axes[axis_idx]) {
+        is_axis = 1;
+        break;
+      }
+    }
+
+    if (is_axis == 0) {
+      offset = offset * (size_t)(dims[idx]) + (size_t)(iter[idx]);
+    }
+  }
+  return offset;
+}
+
+static inline int NextIndex(int num_dims, const int *dims, int *current) {
+  int carry = 1;
+  for (int idx = num_dims - 1; idx >= 0; --idx) {
+    int current_val = current[idx] + carry;
+    if (dims[idx] == current_val) {
+      current[idx] = 0;
+    } else {
+      current[idx] = current_val;
+      carry = 0;
+      break;
+    }
+  }
+  return (carry == 0);
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // MINDSPORE_LITE_NNACL_FP32_GRAD_UTILS_H_
diff --git a/mindspore/lite/schema/model.fbs b/mindspore/lite/schema/model.fbs
index e2ccc1bc41..26c31f51ea 100644
--- a/mindspore/lite/schema/model.fbs
+++ b/mindspore/lite/schema/model.fbs
@@ -234,6 +234,9 @@ union PrimitiveType {
     BinaryCrossEntropyGrad,
     BinaryCrossEntropy,
     LpNormalization,
+    DropoutGrad,
+    MaximumGrad,
+    MinimumGrad
 }
 
 enum QuantType: int {
diff --git a/mindspore/lite/schema/ops.fbs b/mindspore/lite/schema/ops.fbs
index f3d7903333..5a768af145 100644
--- a/mindspore/lite/schema/ops.fbs
+++ b/mindspore/lite/schema/ops.fbs
@@ -224,6 +224,7 @@ table Conv2DGradFilter {
     dilateW: int;
     dilateH: int;
     hasBias: bool = false;
+    filter_shape: [int];
     activationType: ActivationType = 0;
 }
 
@@ -244,6 +245,7 @@ table Conv2DGradInput {
     dilateW: int;
     dilateH: int;
     hasBias: bool = false;
+    input_shape: [int];
     activationType: ActivationType = 0;
 }
 
@@ -264,6 +266,7 @@ table GroupConv2DGradInput {
     dilateW: int;
     dilateH: int;
     hasBias: bool = false;
+    input_shape: [int];
     activationType: ActivationType = 0;
 }
 
@@ -478,13 +481,10 @@ table DeConv2DGradFilter {
 }
 
 table BNGrad {
-    eps : float;
-    momentum: float;
-}
-table BNGradInput {
-    eps : float;
+    eps: float;
     momentum: float;
 }
+
 table Scale {
     axis: int;
     activationType: ActivationType = 0;
@@ -1087,6 +1087,16 @@ table FftReal {
 table FftImag {
 }
 
+table DropoutGrad {
+    ratio : float = 0.5;
+}
+
+table MaximumGrad {
+}
+
+table MinimumGrad {
+}
+
 table NonMaxSuppression {
     centerPointBox : int = 0;
 }
diff --git a/mindspore/lite/src/lite_kernel.h b/mindspore/lite/src/lite_kernel.h
index 8defe1be0d..1e2877b514 100644
--- a/mindspore/lite/src/lite_kernel.h
+++ b/mindspore/lite/src/lite_kernel.h
@@ -95,13 +95,23 @@ class LiteKernel {
 
   std::string name() const { return this->name_; }
 
-  virtual void train() { train_mode_ = true; }
+  virtual int Train() {
+    this->train_mode_ = true;
+    return mindspore::lite::RET_OK;
+  }
+
+  virtual bool IsTrain() const { return this->train_mode_; }
+
+  virtual int Eval() {
+    this->train_mode_ = false;
+    return mindspore::lite::RET_OK;
+  }
 
-  virtual bool is_train() { return train_mode_; }
+  virtual bool IsEval() const { return !this->train_mode_; }
 
-  virtual void eval() { train_mode_ = false; }
+  virtual void SetTrainable(bool trainable = true) { this->trainable_ = trainable; }
 
-  virtual bool is_eval() { return !train_mode_; }
+  virtual bool IsTrainable() const { return this->trainable_; }
 
   void set_name(const std::string &name) { this->name_ = name; }
 
@@ -179,6 +189,7 @@ class LiteKernel {
   std::vector<LiteKernel *> in_kernels_;
   std::vector<LiteKernel *> out_kernels_;
   bool train_mode_ = false;
+  bool trainable_ = false;  // paramaters of this Kernel are trained in Train Session
   bool is_model_output_ = false;
   size_t workspace_size_ = 0;
   static void *workspace_;
diff --git a/mindspore/lite/src/ops/adam.cc b/mindspore/lite/src/ops/adam.cc
index 33b09ae829..45c28be1a1 100644
--- a/mindspore/lite/src/ops/adam.cc
+++ b/mindspore/lite/src/ops/adam.cc
@@ -73,7 +73,7 @@ Registry AdamRegistry(schema::PrimitiveType_Adam, AdamCreator);
 
 int Adam::InferShape(std::vector<lite::Tensor *> inputs, std::vector<lite::Tensor *> outputs) {
   if (10 != inputs.size()) {
-    MS_LOG(ERROR) << "Adam should have at 10 input tensors";
+    MS_LOG(ERROR) << "Adam should have 10 input tensors";
     return RET_ERROR;
   }
 
diff --git a/mindspore/lite/src/ops/arithmetic_grad.cc b/mindspore/lite/src/ops/arithmetic_grad.cc
index c4138444b2..58be418faa 100644
--- a/mindspore/lite/src/ops/arithmetic_grad.cc
+++ b/mindspore/lite/src/ops/arithmetic_grad.cc
@@ -42,11 +42,18 @@ int ArithmeticGrad::InferShape(std::vector<lite::Tensor *> inputs_, std::vector<
   MS_ASSERT(dx1 != nullptr);
   MS_ASSERT(dx2 != nullptr);
 
+  if ((Type() == schema::PrimitiveType_MaximumGrad) || (Type() == schema::PrimitiveType_MinimumGrad)) {
+    x1 = inputs_[0];
+    x2 = inputs_[1];
+    dy = inputs_[2];
+  }
+
   auto inShape0 = x1->shape();
   auto inShape1 = x2->shape();
   auto outShape = dy->shape();
 
-  if ((Type() == schema::PrimitiveType_AddGrad) || (Type() == schema::PrimitiveType_SubGrad)) {
+  if ((Type() == schema::PrimitiveType_AddGrad) || (Type() == schema::PrimitiveType_SubGrad) ||
+      (Type() == schema::PrimitiveType_MaximumGrad) || (Type() == schema::PrimitiveType_MinimumGrad)) {
     ndim_ = outShape.size();
     x1_shape_.resize(ndim_);
     x2_shape_.resize(ndim_);
@@ -61,7 +68,6 @@ int ArithmeticGrad::InferShape(std::vector<lite::Tensor *> inputs_, std::vector<
       dy_shape_[i] = outShape[i];
     }
   } else {
-    // if (inShape0.size() < inShape1.size())
     if (dx1->ElementsNum() < dx2->ElementsNum()) {
       ndim_ = inShape1.size();
       x1_shape_.resize(ndim_);
diff --git a/mindspore/lite/src/ops/bias_grad.cc b/mindspore/lite/src/ops/bias_grad.cc
index dfe57ec218..27cc25bca0 100644
--- a/mindspore/lite/src/ops/bias_grad.cc
+++ b/mindspore/lite/src/ops/bias_grad.cc
@@ -45,7 +45,12 @@ int BiasGrad::UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &i
       MS_LOG(ERROR) << "new primitiveT value failed";
       return RET_ERROR;
     }
-    attr->axis = {0};  // GetValue<std::vector<int>>(prim.GetAttr("axis"));
+    if (prim.GetAttr("axis") == nullptr) {
+      MS_LOG(WARNING) << "get axis failed";
+      attr->axis = {0};
+    } else {
+      attr->axis = GetValue<std::vector<int>>(prim.GetAttr("axis"));
+    }
     this->primitive_->value.value = attr;
     if (this->primitive_->value.value == nullptr) {
       MS_LOG(ERROR) << "primitive value is nullptr";
diff --git a/mindspore/lite/src/ops/bn_grad.cc b/mindspore/lite/src/ops/bn_grad.cc
index 5770a3a9e9..c054911b51 100644
--- a/mindspore/lite/src/ops/bn_grad.cc
+++ b/mindspore/lite/src/ops/bn_grad.cc
@@ -42,13 +42,16 @@ int BNGrad::UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inp
     return RET_ERROR;
   }
   if (this->primitive_->value.value == nullptr) {
-    auto attr = new (std::nothrow) schema::BNGradInputT();
+    auto attr = new (std::nothrow) schema::BNGradT();
     if (attr == nullptr) {
       MS_LOG(ERROR) << "new primitiveT value failed";
       return RET_ERROR;
     }
-    attr->momentum = GetValue<float>(prim.GetAttr("momentum"));
-    // FusedBatchNormGrad dows not get this attribute
+    attr->momentum = 0.1f;
+    if (prim.GetAttr("momentum") != nullptr) {
+      attr->momentum = GetValue<float>(prim.GetAttr("momentum"));
+    }
+    attr->eps = 1e-5;
     if (prim.GetAttr("epsilon") != nullptr) {
       attr->eps = GetValue<float>(prim.GetAttr("epsilon"));
     }
@@ -75,6 +78,9 @@ int BNGrad::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers:
   return RET_OK;
 }
 
+PrimitiveC *BNGradCreator(const schema::Primitive *primitive) { return PrimitiveC::NewPrimitiveC<BNGrad>(primitive); }
+Registry BNGradRegistry(schema::PrimitiveType_BNGrad, BNGradCreator);
+
 float BNGrad::GetEps() const { return this->primitive_->value_as_BNGrad()->eps(); }
 float BNGrad::GetMomentum() const { return this->primitive_->value_as_BNGrad()->momentum(); }
 #endif
@@ -90,6 +96,10 @@ int BNGrad::InferShape(std::vector<lite::Tensor *> inputs, std::vector<lite::Ten
   auto in = inputs[1];
   auto scale = inputs[2];
 
+  if (in->shape().size() != 4) {
+    MS_LOG(ERROR) << "Grad Fused batchnorm only support nhwc input!";
+  }
+
   outputs[0]->set_shape(in->shape());
   outputs[1]->set_shape(scale->shape());
   outputs[2]->set_shape(scale->shape());
diff --git a/mindspore/lite/src/ops/conv2d_grad_filter.cc b/mindspore/lite/src/ops/conv2d_grad_filter.cc
index 33a161bcb3..2199c9f0b4 100644
--- a/mindspore/lite/src/ops/conv2d_grad_filter.cc
+++ b/mindspore/lite/src/ops/conv2d_grad_filter.cc
@@ -38,6 +38,7 @@ int Conv2DGradFilter::GetPadRight() const { return this->primitive_->value.AsCon
 int Conv2DGradFilter::GetDilateW() const { return this->primitive_->value.AsConv2DGradFilter()->dilateW; }
 int Conv2DGradFilter::GetDilateH() const { return this->primitive_->value.AsConv2DGradFilter()->dilateH; }
 bool Conv2DGradFilter::GetHasBias() const { return this->primitive_->value.AsConv2DGradFilter()->hasBias; }
+
 int Conv2DGradFilter::GetActivationType() const { return this->primitive_->value.AsConv2DGradFilter()->activationType; }
 
 void Conv2DGradFilter::SetFormat(int format) {
@@ -66,6 +67,9 @@ void Conv2DGradFilter::SetPadRight(int pad_right) {
 void Conv2DGradFilter::SetDilateW(int dilate_w) { this->primitive_->value.AsConv2DGradFilter()->dilateW = dilate_w; }
 void Conv2DGradFilter::SetDilateH(int dilate_h) { this->primitive_->value.AsConv2DGradFilter()->dilateH = dilate_h; }
 void Conv2DGradFilter::SetHasBias(bool has_bias) { this->primitive_->value.AsConv2DGradFilter()->hasBias = has_bias; }
+std::vector<int> Conv2DGradFilter::GetFilterShape() const {
+  return this->primitive_->value.AsConv2DGradFilter()->filter_shape;
+}
 void Conv2DGradFilter::SetActivationType(int activation_type) {
   this->primitive_->value.AsConv2DGradFilter()->activationType = (schema::ActivationType)activation_type;
 }
@@ -134,6 +138,28 @@ int Conv2DGradFilter::UnPackAttr(const Primitive &prim, const std::vector<AnfNod
       attr->activationType = schema::ActivationType_NO_ACTIVATION;
     }
 
+    if (inputs.size() >= kAnfPopulaterThree) {
+      auto filter_shape = inputs[kAnfPopulaterTwo];
+      MS_ASSERT(filter_shape != nullptr);
+      if (filter_shape->isa<ValueNode>()) {
+        auto valueNode = filter_shape->cast<ValueNodePtr>();
+        MS_ASSERT(valueNode != nullptr);
+        auto value = valueNode->value();
+        MS_ASSERT(value != nullptr);
+        if (value->isa<ValueTuple>()) {
+          auto valTuplPtr = dyn_cast<ValueTuple>(value);
+          MS_ASSERT(valTuplPtr != nullptr);
+          const int nchw2nhwc[] = {0, 3, 1, 2};
+          attr->filter_shape.resize(valTuplPtr->size());
+          for (size_t i = 0; i < valTuplPtr->size(); i++) {
+            auto elem = dyn_cast<Int32Imm>((*valTuplPtr)[i]);
+            MS_ASSERT(elem != nullptr);
+            attr->filter_shape[nchw2nhwc[i]] = elem->value();
+          }
+        }
+      }
+    }
+
     this->primitive_->value.value = attr;
     if (this->primitive_->value.value == nullptr) {
       MS_LOG(ERROR) << "primitive value is nullptr";
@@ -151,10 +177,16 @@ int Conv2DGradFilter::UnPackToFlatBuilder(const schema::Primitive *primitive, fl
     MS_LOG(ERROR) << "value_as_Conv2DGradFilter return nullptr";
     return RET_ERROR;
   }
-  auto val_offset = schema::CreateConv2DGradFilter(
+  std::vector<int32_t> filter_shape;
+  if (attr->filter_shape() != nullptr) {
+    for (int i = 0; i < static_cast<int>(attr->filter_shape()->size()); i++) {
+      filter_shape.push_back(attr->filter_shape()->data()[i]);
+    }
+  }
+  auto val_offset = schema::CreateConv2DGradFilterDirect(
     *fbb, attr->format(), attr->group(), attr->channelIn(), attr->channelOut(), attr->kernelW(), attr->kernelH(),
     attr->strideW(), attr->strideH(), attr->padMode(), attr->padUp(), attr->padDown(), attr->padLeft(),
-    attr->padRight(), attr->dilateW(), attr->dilateH(), attr->hasBias(), attr->activationType());
+    attr->padRight(), attr->dilateW(), attr->dilateH(), attr->hasBias(), &filter_shape, attr->activationType());
   auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_Conv2DGradFilter, val_offset.o);
   fbb->Finish(prim_offset);
   return RET_OK;
@@ -175,6 +207,10 @@ int Conv2DGradFilter::GetPadRight() const { return this->primitive_->value_as_Co
 int Conv2DGradFilter::GetDilateW() const { return this->primitive_->value_as_Conv2DGradFilter()->dilateW(); }
 int Conv2DGradFilter::GetDilateH() const { return this->primitive_->value_as_Conv2DGradFilter()->dilateH(); }
 bool Conv2DGradFilter::GetHasBias() const { return this->primitive_->value_as_Conv2DGradFilter()->hasBias(); }
+std::vector<int> Conv2DGradFilter::GetFilterShape() const {
+  auto fb_vector = this->primitive_->value_as_Conv2DGradFilter()->filter_shape();
+  return std::vector<int>(fb_vector->begin(), fb_vector->end());
+}
 int Conv2DGradFilter::GetActivationType() const {
   return this->primitive_->value_as_Conv2DGradFilter()->activationType();
 }
@@ -186,41 +222,22 @@ Registry conv2DGradFilterRegistry(schema::PrimitiveType_Conv2DGradFilter, Conv2D
 #endif
 
 int Conv2DGradFilter::InferShape(std::vector<Tensor *> inputs, std::vector<Tensor *> outputs) {
-  if (3 != inputs.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad Filter should have 3 inputs";
+  if (2 != inputs.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad Filter should have 2 inputs, but it got " << inputs.size();
     return RET_ERROR;
   }
   if (1 != outputs.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad Filter should have one output";
+    MS_LOG(ERROR) << "Conv2d Grad Filter should have one output but it got " << outputs.size();
     return RET_ERROR;
   }
 
   auto *in0 = inputs.at(0);
-  auto *in = inputs.at(2);
   MS_ASSERT(in0 != nullptr);
-  MS_ASSERT(in != nullptr);
-
-  std::vector<int> output_shape;
-  int *out_shape = reinterpret_cast<int *>(in->MutableData());
-  int new_size = in->ElementsNum();
-  if (in0->GetFormat() == in->GetFormat()) {
-    for (int i = 0; i < new_size; i++) output_shape.push_back(out_shape[i]);
-  } else {
-    if ((in0->GetFormat() == schema::Format_NHWC) && (in->GetFormat() == schema::Format_NCHW)) {
-      output_shape.push_back(out_shape[0]);
-      output_shape.push_back(out_shape[2]);
-      output_shape.push_back(out_shape[3]);
-      output_shape.push_back(out_shape[1]);
-    } else {
-      MS_LOG(ERROR) << "Shape covnert is not supported";
-      return RET_ERROR;
-    }
-  }
 
   auto *out = outputs.at(0);
   MS_ASSERT(out != nullptr);
 
-  out->set_shape(output_shape);
+  out->set_shape(GetFilterShape());
   out->set_data_type(in0->data_type());
   out->SetFormat(in0->GetFormat());
 
diff --git a/mindspore/lite/src/ops/conv2d_grad_filter.h b/mindspore/lite/src/ops/conv2d_grad_filter.h
index 93cd81c245..ec56829e51 100644
--- a/mindspore/lite/src/ops/conv2d_grad_filter.h
+++ b/mindspore/lite/src/ops/conv2d_grad_filter.h
@@ -72,6 +72,7 @@ class Conv2DGradFilter : public PrimitiveC {
   int GetDilateH() const;
   bool GetHasBias() const;
   int GetActivationType() const;
+  std::vector<int> GetFilterShape() const;
 };
 }  // namespace lite
 }  // namespace mindspore
diff --git a/mindspore/lite/src/ops/conv2d_grad_input.cc b/mindspore/lite/src/ops/conv2d_grad_input.cc
index 562ae6a326..37889e5bb7 100644
--- a/mindspore/lite/src/ops/conv2d_grad_input.cc
+++ b/mindspore/lite/src/ops/conv2d_grad_input.cc
@@ -39,6 +39,9 @@ int Conv2DGradInput::GetPadRight() const { return this->primitive_->value.AsConv
 int Conv2DGradInput::GetDilateW() const { return this->primitive_->value.AsConv2DGradInput()->dilateW; }
 int Conv2DGradInput::GetDilateH() const { return this->primitive_->value.AsConv2DGradInput()->dilateH; }
 bool Conv2DGradInput::GetHasBias() const { return this->primitive_->value.AsConv2DGradInput()->hasBias; }
+std::vector<int> Conv2DGradInput::GetInputShape() const {
+  return this->primitive_->value.AsConv2DGradInput()->input_shape;
+}
 int Conv2DGradInput::GetActivationType() const { return this->primitive_->value.AsConv2DGradInput()->activationType; }
 
 void Conv2DGradInput::SetFormat(int format) {
@@ -137,6 +140,27 @@ int Conv2DGradInput::UnPackAttr(const Primitive &prim, const std::vector<AnfNode
       attr->activationType = schema::ActivationType_NO_ACTIVATION;
     }
 
+    if (inputs.size() >= kAnfPopulaterThree) {
+      auto input_shape = inputs[kAnfPopulaterTwo];
+      MS_ASSERT(input_shape != nullptr);
+      if (input_shape->isa<ValueNode>()) {
+        auto valueNode = input_shape->cast<ValueNodePtr>();
+        MS_ASSERT(valueNode != nullptr);
+        auto value = valueNode->value();
+        MS_ASSERT(value != nullptr);
+        if (value->isa<ValueTuple>()) {
+          auto valTuplPtr = dyn_cast<ValueTuple>(value);
+          MS_ASSERT(valTuplPtr != nullptr);
+          const int nchw2nhwc[] = {0, 3, 1, 2};
+          attr->input_shape.resize(valTuplPtr->size());
+          for (size_t i = 0; i < valTuplPtr->size(); i++) {
+            auto elem = dyn_cast<Int32Imm>((*valTuplPtr)[i]);
+            MS_ASSERT(elem != nullptr);
+            attr->input_shape[nchw2nhwc[i]] = elem->value();
+          }
+        }
+      }
+    }
     this->primitive_->value.value = attr;
     if (this->primitive_->value.value == nullptr) {
       MS_LOG(ERROR) << "primitive value is nullptr";
@@ -154,10 +178,16 @@ int Conv2DGradInput::UnPackToFlatBuilder(const schema::Primitive *primitive, fla
     MS_LOG(ERROR) << "value_as_Conv2DGradInput return nullptr";
     return RET_ERROR;
   }
-  auto val_offset = schema::CreateConv2DGradInput(
+  std::vector<int32_t> input_shape;
+  if (attr->input_shape() != nullptr) {
+    for (int i = 0; i < static_cast<int>(attr->input_shape()->size()); i++) {
+      input_shape.push_back(attr->input_shape()->data()[i]);
+    }
+  }
+  auto val_offset = schema::CreateConv2DGradInputDirect(
     *fbb, attr->format(), attr->group(), attr->channelIn(), attr->channelOut(), attr->kernelW(), attr->kernelH(),
     attr->strideW(), attr->strideH(), attr->padMode(), attr->padUp(), attr->padDown(), attr->padLeft(),
-    attr->padRight(), attr->dilateW(), attr->dilateH(), attr->hasBias(), attr->activationType());
+    attr->padRight(), attr->dilateW(), attr->dilateH(), attr->hasBias(), &input_shape, attr->activationType());
   auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_Conv2DGradInput, val_offset.o);
   fbb->Finish(prim_offset);
   return RET_OK;
@@ -178,6 +208,10 @@ int Conv2DGradInput::GetPadRight() const { return this->primitive_->value_as_Con
 int Conv2DGradInput::GetDilateW() const { return this->primitive_->value_as_Conv2DGradInput()->dilateW(); }
 int Conv2DGradInput::GetDilateH() const { return this->primitive_->value_as_Conv2DGradInput()->dilateH(); }
 bool Conv2DGradInput::GetHasBias() const { return this->primitive_->value_as_Conv2DGradInput()->hasBias(); }
+std::vector<int> Conv2DGradInput::GetInputShape() const {
+  auto fb_vector = this->primitive_->value_as_Conv2DGradInput()->input_shape();
+  return std::vector<int>(fb_vector->begin(), fb_vector->end());
+}
 int Conv2DGradInput::GetActivationType() const {
   return this->primitive_->value_as_Conv2DGradInput()->activationType();
 }
@@ -189,40 +223,21 @@ Registry Conv2DGradInputRegistry(schema::PrimitiveType_Conv2DGradInput, Conv2DGr
 #endif
 
 int Conv2DGradInput::InferShape(std::vector<Tensor *> inputs, std::vector<Tensor *> outputs) {
-  if (3 != inputs.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad Input should have 3 inputs";
+  if (2 != inputs.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad Input should have 2 inputs";
     return RET_ERROR;
   }
   if (1 != outputs.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad input should have one output";
+    MS_LOG(ERROR) << "Conv2d Grad output should have one output";
     return RET_ERROR;
   }
 
   auto *in0 = inputs.at(0);
-  auto *in = inputs.at(2);
   MS_ASSERT(in0 != nullptr);
-  MS_ASSERT(in != nullptr);
-
-  std::vector<int> output_shape;
-  int *out_shape = reinterpret_cast<int *>(in->MutableData());
-  int new_size = in->ElementsNum();
-  if (in0->GetFormat() == in->GetFormat()) {
-    for (int i = 0; i < new_size; i++) output_shape.push_back(out_shape[i]);
-  } else {
-    if ((in0->GetFormat() == schema::Format_NHWC) && (in->GetFormat() == schema::Format_NCHW)) {
-      output_shape.push_back(out_shape[0]);
-      output_shape.push_back(out_shape[2]);
-      output_shape.push_back(out_shape[3]);
-      output_shape.push_back(out_shape[1]);
-    } else {
-      MS_LOG(ERROR) << "Shape covnert is not supported";
-      return RET_ERROR;
-    }
-  }
 
   auto *out = outputs.at(0);
   MS_ASSERT(out != nullptr);
-  out->set_shape(output_shape);
+  out->set_shape(GetInputShape());
   out->set_data_type(in0->data_type());
   out->SetFormat(in0->GetFormat());
 
diff --git a/mindspore/lite/src/ops/conv2d_grad_input.h b/mindspore/lite/src/ops/conv2d_grad_input.h
index 73e8f39402..53816be7fd 100644
--- a/mindspore/lite/src/ops/conv2d_grad_input.h
+++ b/mindspore/lite/src/ops/conv2d_grad_input.h
@@ -72,6 +72,7 @@ class Conv2DGradInput : public PrimitiveC {
   int GetDilateH() const;
   bool GetHasBias() const;
   int GetActivationType() const;
+  std::vector<int> GetInputShape() const;
 };
 }  // namespace lite
 }  // namespace mindspore
diff --git a/mindspore/lite/src/ops/dropout.cc b/mindspore/lite/src/ops/dropout.cc
index 4470f55e9b..dd74b24506 100644
--- a/mindspore/lite/src/ops/dropout.cc
+++ b/mindspore/lite/src/ops/dropout.cc
@@ -27,6 +27,37 @@ float Dropout::GetRatio() const { return this->primitive_->value.AsDropout()->ra
 
 void Dropout::SetRatio(float ratio) { this->primitive_->value.AsDropout()->ratio = ratio; }
 
+int Dropout::UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) {
+  if (this->primitive_ == nullptr) {
+    this->primitive_ = new (std::nothrow) schema::PrimitiveT;
+    if (this->primitive_ == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT failed";
+      return RET_ERROR;
+    }
+    this->primitive_->value.type = schema::PrimitiveType_Dropout;
+  }
+  if (this->primitive_->value.type != schema::PrimitiveType_Dropout) {
+    MS_LOG(ERROR) << "Primitive type is error :" << this->primitive_->value.type;
+    return RET_ERROR;
+  }
+  if (this->primitive_->value.value == nullptr) {
+    auto attr = new (std::nothrow) schema::DropoutT();
+    if (attr == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT value failed";
+      return RET_ERROR;
+    }
+    if (prim.GetAttr("keep_prob") != nullptr) {
+      attr->ratio = GetValue<float>(prim.GetAttr("keep_prob"));
+    }
+    this->primitive_->value.value = attr;
+    if (this->primitive_->value.value == nullptr) {
+      MS_LOG(ERROR) << "primitive value is nullptr";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
 #else
 int Dropout::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
   MS_ASSERT(nullptr != primitive);
@@ -46,5 +77,29 @@ float Dropout::GetRatio() const { return this->primitive_->value_as_Dropout()->r
 PrimitiveC *DropoutCreator(const schema::Primitive *primitive) { return PrimitiveC::NewPrimitiveC<Dropout>(primitive); }
 Registry DropoutRegistry(schema::PrimitiveType_Dropout, DropoutCreator);
 #endif
+int Dropout::InferShape(std::vector<Tensor *> inputs_, std::vector<Tensor *> outputs_) {
+  MS_ASSERT(this->primitive_ != nullptr);
+  auto input = inputs_.front();
+  MS_ASSERT(input != nullptr);
+  auto output0 = outputs_.front();
+  MS_ASSERT(output0 != nullptr);
+  if (!GetInferFlag()) {
+    return RET_OK;
+  }
+  output0->set_shape(input->shape());
+  output0->set_data_type(input->data_type());
+  output0->SetFormat(input->GetFormat());
+
+  if (outputs_.size() > 1) {
+    auto output1 = outputs_[1];
+    MS_ASSERT(output1 != nullptr);
+    output1->set_shape(input->shape());
+    output1->set_data_type(input->data_type());
+    output1->SetFormat(input->GetFormat());
+  }
+
+  return RET_OK;
+}
+
 }  // namespace lite
 }  // namespace mindspore
diff --git a/mindspore/lite/src/ops/dropout.h b/mindspore/lite/src/ops/dropout.h
index 5bb645f765..21310974b6 100644
--- a/mindspore/lite/src/ops/dropout.h
+++ b/mindspore/lite/src/ops/dropout.h
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_DROPOUT_H_
-#define LITE_MINDSPORE_LITE_C_OPS_DROPOUT_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_DROPOUT_H_
+#define MINDSPORE_LITE_SRC_OPS_DROPOUT_H_
 
 #include <vector>
 #include <set>
@@ -32,13 +32,16 @@ class Dropout : public PrimitiveC {
   MS_DECLARE_PARENT(Dropout, PrimitiveC);
   explicit Dropout(schema::PrimitiveT *primitive) : PrimitiveC(primitive) {}
   void SetRatio(float ratio);
+  int UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) override;
 
 #else
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
 #endif
   float GetRatio() const;
+  int InferShape(std::vector<lite::Tensor *> inputs_, std::vector<lite::Tensor *> outputs_) override;
 };
+
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_DROPOUT_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_DROPOUT_H_
diff --git a/mindspore/lite/src/ops/dropout_grad.cc b/mindspore/lite/src/ops/dropout_grad.cc
new file mode 100644
index 0000000000..99b348f005
--- /dev/null
+++ b/mindspore/lite/src/ops/dropout_grad.cc
@@ -0,0 +1,100 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "src/ops/dropout_grad.h"
+
+#ifndef PRIMITIVE_WRITEABLE
+#include "src/ops/ops_register.h"
+#endif
+
+namespace mindspore {
+namespace lite {
+#ifdef PRIMITIVE_WRITEABLE
+float DropoutGrad::GetRatio() const { return this->primitive_->value.AsDropout()->ratio; }
+
+void DropoutGrad::SetRatio(float ratio) { this->primitive_->value.AsDropout()->ratio = ratio; }
+
+int DropoutGrad::UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) {
+  if (this->primitive_ == nullptr) {
+    this->primitive_ = new (std::nothrow) schema::PrimitiveT;
+    if (this->primitive_ == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT failed";
+      return RET_ERROR;
+    }
+    this->primitive_->value.type = schema::PrimitiveType_DropoutGrad;
+  }
+  if (this->primitive_->value.type != schema::PrimitiveType_DropoutGrad) {
+    MS_LOG(ERROR) << "Primitive type is error :" << this->primitive_->value.type;
+    return RET_ERROR;
+  }
+  if (this->primitive_->value.value == nullptr) {
+    auto attr = new (std::nothrow) schema::DropoutGradT();
+    if (attr == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT value failed";
+      return RET_ERROR;
+    }
+    if (prim.GetAttr("keep_prob") != nullptr) {
+      attr->ratio = GetValue<float>(prim.GetAttr("keep_prob"));
+    }
+    this->primitive_->value.value = attr;
+    if (this->primitive_->value.value == nullptr) {
+      MS_LOG(ERROR) << "primitive value is nullptr";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+#else
+int DropoutGrad::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
+  MS_ASSERT(nullptr != primitive);
+  MS_ASSERT(nullptr != fbb);
+  auto attr = primitive->value_as_DropoutGrad();
+  if (attr == nullptr) {
+    MS_LOG(ERROR) << "value_as_DropoutGrad return nullptr";
+    return RET_ERROR;
+  }
+  auto val_offset = schema::CreateDropoutGrad(*fbb, attr->ratio());
+  auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_DropoutGrad, val_offset.o);
+  fbb->Finish(prim_offset);
+  return RET_OK;
+}
+float DropoutGrad::GetRatio() const { return this->primitive_->value_as_DropoutGrad()->ratio(); }
+
+PrimitiveC *DropoutGradCreator(const schema::Primitive *primitive) {
+  return PrimitiveC::NewPrimitiveC<DropoutGrad>(primitive);
+}
+Registry DropoutGradRegistry(schema::PrimitiveType_DropoutGrad, DropoutGradCreator);
+
+#endif
+int DropoutGrad::InferShape(std::vector<Tensor *> inputs_, std::vector<Tensor *> outputs_) {
+  MS_ASSERT(this->primitive_ != nullptr);
+  MS_ASSERT(inputs_.size() == 2);
+  auto input = inputs_.front();
+  MS_ASSERT(input != nullptr);
+  auto output = outputs_.front();
+  MS_ASSERT(output != nullptr);
+  if (!GetInferFlag()) {
+    return RET_OK;
+  }
+  output->set_shape(input->shape());
+  output->set_data_type(input->data_type());
+  output->SetFormat(input->GetFormat());
+
+  return RET_OK;
+}
+
+}  // namespace lite
+}  // namespace mindspore
diff --git a/mindspore/lite/src/ops/dropout_grad.h b/mindspore/lite/src/ops/dropout_grad.h
new file mode 100644
index 0000000000..c0d0d11c29
--- /dev/null
+++ b/mindspore/lite/src/ops/dropout_grad.h
@@ -0,0 +1,47 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_OPS_DROPOUT_GRAD_H_
+#define MINDSPORE_LITE_SRC_OPS_DROPOUT_GRAD_H_
+
+#include <vector>
+#include <set>
+#include <cmath>
+#include "src/ops/primitive_c.h"
+
+namespace mindspore {
+namespace lite {
+class DropoutGrad : public PrimitiveC {
+ public:
+#ifdef PRIMITIVE_WRITEABLE
+  MS_DECLARE_PARENT(DropoutGrad, PrimitiveC);
+  DropoutGrad() = default;
+  explicit DropoutGrad(schema::PrimitiveT *primitive) : PrimitiveC(primitive) {}
+  void SetRatio(float ratio);
+  int UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) override;
+
+#else
+  DropoutGrad() = default;
+
+  int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
+#endif
+  float GetRatio() const;
+  int InferShape(std::vector<lite::Tensor *> inputs_, std::vector<lite::Tensor *> outputs_) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_OPS_DROPOUT_GRAD_H_
diff --git a/mindspore/lite/src/ops/group_conv2d_grad_input.cc b/mindspore/lite/src/ops/group_conv2d_grad_input.cc
index 1ce21b2506..673d89a8cd 100644
--- a/mindspore/lite/src/ops/group_conv2d_grad_input.cc
+++ b/mindspore/lite/src/ops/group_conv2d_grad_input.cc
@@ -39,6 +39,9 @@ int GroupConv2DGradInput::GetPadRight() const { return this->primitive_->value.A
 int GroupConv2DGradInput::GetDilateW() const { return this->primitive_->value.AsGroupConv2DGradInput()->dilateW; }
 int GroupConv2DGradInput::GetDilateH() const { return this->primitive_->value.AsGroupConv2DGradInput()->dilateH; }
 bool GroupConv2DGradInput::GetHasBias() const { return this->primitive_->value.AsGroupConv2DGradInput()->hasBias; }
+std::vector<int> GroupConv2DGradInput::GetInputShape() const {
+  return this->primitive_->value.AsGroupConv2DGradInput()->input_shape;
+}
 int GroupConv2DGradInput::GetActivationType() const {
   return this->primitive_->value.AsGroupConv2DGradInput()->activationType;
 }
@@ -99,10 +102,16 @@ int GroupConv2DGradInput::UnPackToFlatBuilder(const schema::Primitive *primitive
     MS_LOG(ERROR) << "value_as_GroupConv2DGradInput return nullptr";
     return RET_ERROR;
   }
-  auto val_offset = schema::CreateGroupConv2DGradInput(
+  std::vector<int32_t> input_shape;
+  if (attr->input_shape() != nullptr) {
+    for (int i = 0; i < static_cast<int>(attr->input_shape()->size()); i++) {
+      input_shape.push_back(attr->input_shape()->data()[i]);
+    }
+  }
+  auto val_offset = schema::CreateGroupConv2DGradInputDirect(
     *fbb, attr->format(), attr->group(), attr->channelIn(), attr->channelOut(), attr->kernelW(), attr->kernelH(),
     attr->strideW(), attr->strideH(), attr->padMode(), attr->padUp(), attr->padDown(), attr->padLeft(),
-    attr->padRight(), attr->dilateW(), attr->dilateH(), attr->hasBias(), attr->activationType());
+    attr->padRight(), attr->dilateW(), attr->dilateH(), attr->hasBias(), &input_shape, attr->activationType());
   auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_GroupConv2DGradInput, val_offset.o);
   fbb->Finish(prim_offset);
   return RET_OK;
@@ -127,51 +136,38 @@ int GroupConv2DGradInput::GetPadRight() const { return this->primitive_->value_a
 int GroupConv2DGradInput::GetDilateW() const { return this->primitive_->value_as_GroupConv2DGradInput()->dilateW(); }
 int GroupConv2DGradInput::GetDilateH() const { return this->primitive_->value_as_GroupConv2DGradInput()->dilateH(); }
 bool GroupConv2DGradInput::GetHasBias() const { return this->primitive_->value_as_GroupConv2DGradInput()->hasBias(); }
+std::vector<int> GroupConv2DGradInput::GetInputShape() const {
+  auto fb_vector = this->primitive_->value_as_GroupConv2DGradInput()->input_shape();
+  return std::vector<int>(fb_vector->begin(), fb_vector->end());
+}
 int GroupConv2DGradInput::GetActivationType() const {
   return this->primitive_->value_as_GroupConv2DGradInput()->activationType();
 }
-
 PrimitiveC *GroupConv2DGradInputCreator(const schema::Primitive *primitive) {
   return PrimitiveC::NewPrimitiveC<GroupConv2DGradInput>(primitive);
 }
 Registry GroupConv2DGradInputRegistry(schema::PrimitiveType_GroupConv2DGradInput, GroupConv2DGradInputCreator);
+
 #endif
 
 int GroupConv2DGradInput::InferShape(std::vector<Tensor *> inputs, std::vector<Tensor *> outputs) {
-  if (3 != inputs.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad Input should have 3 inputs";
+  if (2 != inputs.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad input should have 2 inputs";
     return RET_ERROR;
   }
   if (1 != outputs.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad input should have one output";
+    MS_LOG(ERROR) << "Conv2d Grad output should have one output";
     return RET_ERROR;
   }
 
   auto *in0 = inputs.at(0);
-  auto *in = inputs.at(2);
-  MS_ASSERT(in0 != nullptr);
-  MS_ASSERT(in != nullptr);
 
-  std::vector<int> output_shape;
-  int *out_shape = reinterpret_cast<int *>(in->MutableData());
-  int new_size = in->ElementsNum();
-  if (in0->GetFormat() == in->GetFormat()) {
-    for (int i = 0; i < new_size; i++) output_shape.push_back(out_shape[i]);
-  } else {
-    if ((in0->GetFormat() == schema::Format_NHWC) && (in->GetFormat() == schema::Format_NCHW)) {
-      output_shape.push_back(out_shape[0]);
-      output_shape.push_back(out_shape[2]);
-      output_shape.push_back(out_shape[3]);
-      output_shape.push_back(out_shape[1]);
-    } else {
-      MS_LOG(ERROR) << "Shape covnert is not supported";
-      return RET_ERROR;
-    }
-  }
+  MS_ASSERT(in0 != nullptr);
 
   auto *out = outputs.at(0);
   MS_ASSERT(out != nullptr);
-  out->set_shape(output_shape);
+  out->set_shape(GetInputShape());
+
   out->set_data_type(in0->data_type());
   out->SetFormat(in0->GetFormat());
 
diff --git a/mindspore/lite/src/ops/group_conv2d_grad_input.h b/mindspore/lite/src/ops/group_conv2d_grad_input.h
index a005ced74d..a53af0f316 100644
--- a/mindspore/lite/src/ops/group_conv2d_grad_input.h
+++ b/mindspore/lite/src/ops/group_conv2d_grad_input.h
@@ -70,6 +70,7 @@ class GroupConv2DGradInput : public PrimitiveC {
   int GetDilateW() const;
   int GetDilateH() const;
   bool GetHasBias() const;
+  std::vector<int> GetInputShape() const;
   int GetActivationType() const;
 };
 }  // namespace lite
diff --git a/mindspore/lite/src/ops/maximum.h b/mindspore/lite/src/ops/maximum.h
index 9b85e202c3..052088ebab 100644
--- a/mindspore/lite/src/ops/maximum.h
+++ b/mindspore/lite/src/ops/maximum.h
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_MAXIMUM_H_
-#define LITE_MINDSPORE_LITE_C_OPS_MAXIMUM_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_MAXIMUM_H_
+#define MINDSPORE_LITE_SRC_OPS_MAXIMUM_H_
 
 #include <vector>
 #include <set>
@@ -41,4 +41,4 @@ class Maximum : public Arithmetic {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_MAXIMUM_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_MAXIMUM_H_
diff --git a/mindspore/lite/src/ops/maximum_grad.cc b/mindspore/lite/src/ops/maximum_grad.cc
new file mode 100644
index 0000000000..9644957f55
--- /dev/null
+++ b/mindspore/lite/src/ops/maximum_grad.cc
@@ -0,0 +1,124 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "include/errorcode.h"
+#include "src/ops/maximum_grad.h"
+#include "src/common/log_adapter.h"
+#ifdef PRIMITIVE_WRITEABLE
+#include <float.h>
+#include "tools/converter/quantizer/quantize_util.h"
+#endif
+
+#ifndef PRIMITIVE_WRITEABLE
+#include "src/ops/ops_register.h"
+#endif
+
+namespace mindspore {
+namespace lite {
+#ifdef PRIMITIVE_WRITEABLE
+int MaximumGrad::UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) {
+  if (this->primitive_ == nullptr) {
+    this->primitive_ = new (std::nothrow) schema::PrimitiveT;
+    if (this->primitive_ == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT failed";
+      return RET_ERROR;
+    }
+    this->primitive_->value.type = schema::PrimitiveType_MaximumGrad;
+  }
+  if (this->primitive_->value.type != schema::PrimitiveType_MaximumGrad) {
+    MS_LOG(ERROR) << "Primitive type is error :" << this->primitive_->value.type;
+    return RET_ERROR;
+  }
+  if (this->primitive_->value.value == nullptr) {
+    auto attr = new (std::nothrow) schema::MaximumGradT();
+    if (attr == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT value failed";
+      return RET_ERROR;
+    }
+    this->primitive_->value.value = attr;
+    if (this->primitive_->value.value == nullptr) {
+      MS_LOG(ERROR) << "primitive value is nullptr";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+#else
+int MaximumGrad::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
+  MS_ASSERT(nullptr != primitive);
+  MS_ASSERT(nullptr != fbb);
+  auto val_offset = schema::CreateMaximumGrad(*fbb);
+  auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_MaximumGrad, val_offset.o);
+  fbb->Finish(prim_offset);
+  return RET_OK;
+}
+PrimitiveC *MaximumGradCreator(const schema::Primitive *primitive) {
+  return PrimitiveC::NewPrimitiveC<MaximumGrad>(primitive);
+}
+Registry MaximumGradRegistry(schema::PrimitiveType_MaximumGrad, MaximumGradCreator);
+
+#endif
+int MaximumGrad::InferShape(std::vector<Tensor *> inputs_, std::vector<Tensor *> outputs_) {
+  if (inputs_.size() != 3) {
+    MS_LOG(ERROR) << "The number of input must be 3";
+    return RET_ERROR;
+  }
+  if (outputs_.size() != 2) {
+    MS_LOG(ERROR) << "The number of output must be 2";
+    return RET_ERROR;
+  }
+
+  auto x1 = inputs_[0];
+  auto x2 = inputs_[1];
+  auto dy = inputs_[2];
+  auto dx1 = outputs_[0];
+  auto dx2 = outputs_[1];
+
+  MS_ASSERT(dy != nullptr);
+  MS_ASSERT(x1 != nullptr);
+  MS_ASSERT(x2 != nullptr);
+  MS_ASSERT(dx1 != nullptr);
+  MS_ASSERT(dx2 != nullptr);
+  if (!GetInferFlag()) {
+    return RET_OK;
+  }
+
+  auto inShape0 = x1->shape();
+  auto inShape1 = x2->shape();
+  auto outShape = dy->shape();
+
+  ndim_ = outShape.size();
+  x1_shape_.resize(ndim_);
+  x2_shape_.resize(ndim_);
+  dy_shape_.resize(ndim_);
+  auto fillDimNum0 = outShape.size() - inShape0.size();
+  auto fillDimNum1 = outShape.size() - inShape1.size();
+  int j0 = 0;
+  int j1 = 0;
+  for (unsigned int i = 0; i < outShape.size(); i++) {
+    x1_shape_[i] = (i < fillDimNum0) ? 1 : inShape0[j0++];
+    x2_shape_[i] = (i < fillDimNum1) ? 1 : inShape1[j1++];
+    dy_shape_[i] = outShape[i];
+  }
+
+  dx1->set_shape(x1->shape());
+  dx2->set_shape(x2->shape());
+  dx1->set_data_type(dy->data_type());
+  dx2->set_data_type(dy->data_type());
+  return RET_OK;
+}
+}  // namespace lite
+}  // namespace mindspore
diff --git a/mindspore/lite/src/ops/maximum_grad.h b/mindspore/lite/src/ops/maximum_grad.h
new file mode 100644
index 0000000000..10e73b485a
--- /dev/null
+++ b/mindspore/lite/src/ops/maximum_grad.h
@@ -0,0 +1,46 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_OPS_MAXIMUM_GRAD_H_
+#define MINDSPORE_LITE_SRC_OPS_MAXIMUM_GRAD_H_
+
+#include <vector>
+#include <set>
+#include <cmath>
+
+#include "src/ops/arithmetic_grad.h"
+#include "src/ops/primitive_c.h"
+
+namespace mindspore {
+namespace lite {
+class MaximumGrad : public ArithmeticGrad {
+ public:
+#ifdef PRIMITIVE_WRITEABLE
+  MS_DECLARE_PARENT(MaximumGrad, ArithmeticGrad);
+  MaximumGrad() = default;
+  explicit MaximumGrad(schema::PrimitiveT *primitive) : ArithmeticGrad(primitive) {}
+  int UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) override;
+#else
+  MaximumGrad() = default;
+
+  int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
+#endif
+  int InferShape(std::vector<lite::Tensor *> inputs_, std::vector<lite::Tensor *> outputs_) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_OPS_MAXIMUM_GRAD_H_
diff --git a/mindspore/lite/src/ops/minimum.cc b/mindspore/lite/src/ops/minimum.cc
index a7126c9502..5881976ad3 100644
--- a/mindspore/lite/src/ops/minimum.cc
+++ b/mindspore/lite/src/ops/minimum.cc
@@ -23,6 +23,33 @@
 namespace mindspore {
 namespace lite {
 #ifdef PRIMITIVE_WRITEABLE
+int Minimum::UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) {
+  if (this->primitive_ == nullptr) {
+    this->primitive_ = new (std::nothrow) schema::PrimitiveT;
+    if (this->primitive_ == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT failed";
+      return RET_ERROR;
+    }
+    this->primitive_->value.type = schema::PrimitiveType_Minimum;
+  }
+  if (this->primitive_->value.type != schema::PrimitiveType_Minimum) {
+    MS_LOG(ERROR) << "Primitive type is error :" << this->primitive_->value.type;
+    return RET_ERROR;
+  }
+  if (this->primitive_->value.value == nullptr) {
+    auto attr = new (std::nothrow) schema::MinimumT();
+    if (attr == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT value failed";
+      return RET_ERROR;
+    }
+    this->primitive_->value.value = attr;
+    if (this->primitive_->value.value == nullptr) {
+      MS_LOG(ERROR) << "primitive value is nullptr";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
 #else
 int Minimum::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
   MS_ASSERT(nullptr != primitive);
diff --git a/mindspore/lite/src/ops/minimum.h b/mindspore/lite/src/ops/minimum.h
index 72f3ae26c2..de69645c70 100644
--- a/mindspore/lite/src/ops/minimum.h
+++ b/mindspore/lite/src/ops/minimum.h
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_MINIMUM_H_
-#define LITE_MINDSPORE_LITE_C_OPS_MINIMUM_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_MINIMUM_H_
+#define MINDSPORE_LITE_SRC_OPS_MINIMUM_H_
 
 #include <vector>
 #include <set>
@@ -32,6 +32,7 @@ class Minimum : public Arithmetic {
 #ifdef PRIMITIVE_WRITEABLE
   MS_DECLARE_PARENT(Arithmetic, Arithmetic);
   explicit Minimum(schema::PrimitiveT *primitive) : Arithmetic(primitive) {}
+  int UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) override;
 #else
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
 #endif
@@ -39,4 +40,4 @@ class Minimum : public Arithmetic {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_MINIMUM_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_MINIMUM_H_
diff --git a/mindspore/lite/src/ops/minimum_grad.cc b/mindspore/lite/src/ops/minimum_grad.cc
new file mode 100644
index 0000000000..f4a03b7d95
--- /dev/null
+++ b/mindspore/lite/src/ops/minimum_grad.cc
@@ -0,0 +1,76 @@
+/**
+ * Copyright 2019-2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "include/errorcode.h"
+#include "src/ops/minimum_grad.h"
+#include "src/common/log_adapter.h"
+#ifdef PRIMITIVE_WRITEABLE
+#include <float.h>
+#include "tools/converter/quantizer/quantize_util.h"
+#endif
+
+#ifndef PRIMITIVE_WRITEABLE
+#include "src/ops/ops_register.h"
+#endif
+
+namespace mindspore {
+namespace lite {
+#ifdef PRIMITIVE_WRITEABLE
+int MinimumGrad::UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) {
+  if (this->primitive_ == nullptr) {
+    this->primitive_ = new (std::nothrow) schema::PrimitiveT;
+    if (this->primitive_ == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT failed";
+      return RET_ERROR;
+    }
+    this->primitive_->value.type = schema::PrimitiveType_MinimumGrad;
+  }
+  if (this->primitive_->value.type != schema::PrimitiveType_MinimumGrad) {
+    MS_LOG(ERROR) << "Primitive type is error :" << this->primitive_->value.type;
+    return RET_ERROR;
+  }
+  if (this->primitive_->value.value == nullptr) {
+    auto attr = new (std::nothrow) schema::MinimumGradT();
+    if (attr == nullptr) {
+      MS_LOG(ERROR) << "new primitiveT value failed";
+      return RET_ERROR;
+    }
+    this->primitive_->value.value = attr;
+    if (this->primitive_->value.value == nullptr) {
+      MS_LOG(ERROR) << "primitive value is nullptr";
+      return RET_ERROR;
+    }
+  }
+  return RET_OK;
+}
+
+#else
+PrimitiveC *MinimumGradCreator(const schema::Primitive *primitive) {
+  return PrimitiveC::NewPrimitiveC<MinimumGrad>(primitive);
+}
+Registry MinimumGradRegistry(schema::PrimitiveType_MinimumGrad, MinimumGradCreator);
+
+int MinimumGrad::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
+  MS_ASSERT(nullptr != primitive);
+  MS_ASSERT(nullptr != fbb);
+  auto val_offset = schema::CreateMinimumGrad(*fbb);
+  auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_MinimumGrad, val_offset.o);
+  fbb->Finish(prim_offset);
+  return RET_OK;
+}
+#endif
+}  // namespace lite
+}  // namespace mindspore
diff --git a/mindspore/lite/src/ops/minimum_grad.h b/mindspore/lite/src/ops/minimum_grad.h
new file mode 100644
index 0000000000..2d46776419
--- /dev/null
+++ b/mindspore/lite/src/ops/minimum_grad.h
@@ -0,0 +1,45 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_SRC_OPS_MINIMUM_GRAD_H_
+#define MINDSPORE_LITE_SRC_OPS_MINIMUM_GRAD_H_
+
+#include <vector>
+#include <set>
+#include <cmath>
+
+#include "src/ops/arithmetic_grad.h"
+#include "src/ops/primitive_c.h"
+
+namespace mindspore {
+namespace lite {
+class MinimumGrad : public ArithmeticGrad {
+ public:
+#ifdef PRIMITIVE_WRITEABLE
+  MS_DECLARE_PARENT(MinimumGrad, ArithmeticGrad);
+  MinimumGrad() = default;
+  explicit MinimumGrad(schema::PrimitiveT *primitive) : ArithmeticGrad(primitive) {}
+  int UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) override;
+#else
+  MinimumGrad() = default;
+
+  int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
+#endif
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_OPS_MINIMUM_GRAD_H_
diff --git a/mindspore/lite/src/ops/primitive_c.cc b/mindspore/lite/src/ops/primitive_c.cc
index 71f3235e48..5cef1e50d0 100644
--- a/mindspore/lite/src/ops/primitive_c.cc
+++ b/mindspore/lite/src/ops/primitive_c.cc
@@ -18,6 +18,7 @@
 #ifdef PRIMITIVE_WRITEABLE
 #include <memory>
 #include <map>
+
 #include "tools/converter/quantizer/quantize_util.h"
 #include "src/ops/space_to_batch.h"
 #include "src/ops/space_to_batch_nd.h"
@@ -167,12 +168,14 @@
 #include "src/ops/sgd.h"
 #include "src/ops/adam.h"
 #include "src/ops/assign.h"
+#include "src/ops/dropout_grad.h"
+#include "src/ops/maximum_grad.h"
+#include "src/ops/minimum_grad.h"
 #include "src/ops/control_depend.h"
 #include "src/ops/assign_add.h"
 #include "src/ops/binary_cross_entropy.h"
 #include "src/ops/binary_cross_entropy_grad.h"
 #endif
-
 #endif
 namespace mindspore {
 namespace lite {
@@ -506,10 +509,12 @@ std::shared_ptr<PrimitiveC> PrimitiveC::Create(const Primitive &prim, const std:
     return NewPrimitiveC<Maximum>(prim, inputs, quantType);
   } else if (op_type == "Split") {
     return NewPrimitiveC<Split>(prim, inputs, quantType);
-  } else if (op_type == "While") {
-    return NewPrimitiveC<While>(prim, inputs, quantType);
   } else if (op_type == "OneHot") {
     return NewPrimitiveC<OneHot>(prim, inputs, quantType);
+  } else if (op_type == "Dropout") {
+    return NewPrimitiveC<Dropout>(prim, inputs, quantType);
+  } else if (op_type == "While") {
+    return NewPrimitiveC<While>(prim, inputs, quantType);
   } else if (op_type == "GatherV2") {
     return NewPrimitiveC<Gather>(prim, inputs, quantType);
   } else if (op_type == "OnesLike") {
@@ -537,7 +542,7 @@ std::shared_ptr<PrimitiveC> PrimitiveC::Create(const Primitive &prim, const std:
   } else if ((op_type == "ReluGrad" || op_type == "ReLU6Grad" || op_type == "SigmoidGrad" ||
               op_type == "HSigmoidGrad" || op_type == "HSwishGrad")) {
     return NewPrimitiveC<ActivationGrad>(prim, inputs, quantType);
-  } else if ((op_type == "MaxPoolGrad") || (op_type == "MeanPoolGrad")) {
+  } else if ((op_type == "MaxPoolGrad") || (op_type == "MeanPoolGrad") || (op_type == "AvgPoolGradGpu")) {
     return NewPrimitiveC<PoolingGrad>(prim, inputs, quantType);
   } else if (op_type == "Conv2DBackpropFilter") {
     return NewPrimitiveC<Conv2DGradFilter>(prim, inputs, quantType);
@@ -559,6 +564,12 @@ std::shared_ptr<PrimitiveC> PrimitiveC::Create(const Primitive &prim, const std:
     return NewPrimitiveC<Adam>(prim, inputs, quantType);
   } else if (op_type == "Assign") {
     return NewPrimitiveC<Assign>(prim, inputs, quantType);
+  } else if (op_type == "DropoutGrad") {
+    return NewPrimitiveC<DropoutGrad>(prim, inputs, quantType);
+  } else if (op_type == "MaximumGrad") {
+    return NewPrimitiveC<MaximumGrad>(prim, inputs, quantType);
+  } else if (op_type == "MinimumGrad") {
+    return NewPrimitiveC<MinimumGrad>(prim, inputs, quantType);
   } else if (op_type == "AssignAdd") {
     return NewPrimitiveC<AssignAdd>(prim, inputs, quantType);
   } else if (op_type == "BinaryCrossEntropy") {
@@ -884,7 +895,12 @@ PrimitiveC *PrimitiveC::Create(mindspore::schema::PrimitiveT *primitive) {
       return new BinaryCrossEntropyGrad(primitive);
     case schema::PrimitiveType_BinaryCrossEntropy:
       return new BinaryCrossEntropy(primitive);
-
+    case schema::PrimitiveType_DropoutGrad:
+      return new DropoutGrad(primitive);
+    case schema::PrimitiveType_MaximumGrad:
+      return new MaximumGrad(primitive);
+    case schema::PrimitiveType_MinimumGrad:
+      return new MinimumGrad(primitive);
 #endif
     default:
       MS_LOG(ERROR) << "Unsupported primitive type in Create : " << schema::EnumNamePrimitiveType(op_type);
@@ -892,6 +908,7 @@ PrimitiveC *PrimitiveC::Create(mindspore::schema::PrimitiveT *primitive) {
   }
   return nullptr;
 }
+
 #else
 void PrimitiveC::SetQuantType(schema::QuantType quant_type) { this->quant_type_ = quant_type; }
 schema::QuantType PrimitiveC::GetQuantType() const { return quant_type_; }
diff --git a/mindspore/lite/src/ops/squeeze.cc b/mindspore/lite/src/ops/squeeze.cc
index 3e3d8589a7..eaa1c654dd 100644
--- a/mindspore/lite/src/ops/squeeze.cc
+++ b/mindspore/lite/src/ops/squeeze.cc
@@ -50,8 +50,7 @@ int Squeeze::UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &in
       MS_LOG(INFO) << "Squeeze's attr xis is set to default";
       attr->axis = {0};
     } else {
-      int axis = GetValue<int>(prim.GetAttr("axis"));
-      attr->axis = {axis};
+      attr->axis = GetValue<std::vector<int>>(prim.GetAttr("axis"));
     }
     this->primitive_->value.value = attr;
   }
diff --git a/mindspore/lite/src/ops/sub.h b/mindspore/lite/src/ops/sub.h
index 559c1df443..d431851ee3 100644
--- a/mindspore/lite/src/ops/sub.h
+++ b/mindspore/lite/src/ops/sub.h
@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_SUB_H_
-#define LITE_MINDSPORE_LITE_C_OPS_SUB_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_SUB_H_
+#define MINDSPORE_LITE_SRC_OPS_SUB_H_
 
 #include <vector>
 #include <set>
@@ -34,7 +34,6 @@ class Sub : public Arithmetic {
   explicit Sub(schema::PrimitiveT *primitive) : Arithmetic(primitive) {}
   void SetActivationType(int activation_type);
   int UnPackAttr(const Primitive &prim, const std::vector<AnfNodePtr> &inputs) override;
-
 #else
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
 #endif
@@ -43,4 +42,4 @@ class Sub : public Arithmetic {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_SUB_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_SUB_H_
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.cc b/mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.cc
index fe3a955165..e55ab6efb2 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.cc
@@ -39,14 +39,6 @@ void FusedBatchnormCPUKernel::FreeScaleAndOffset() {
     free(offset_);
     offset_ = nullptr;
   }
-  if (save_mean_ != nullptr) {
-    free(save_mean_);
-    save_mean_ = nullptr;
-  }
-  if (save_variance_ != nullptr) {
-    free(save_variance_);
-    save_variance_ = nullptr;
-  }
 }
 
 int FusedBatchnormCPUKernel::InitConstTensor() {
@@ -59,11 +51,8 @@ int FusedBatchnormCPUKernel::InitConstTensor() {
   offset_ = malloc(offset->Size());
   mean_ = malloc(mean->Size());
   variance_ = malloc(variance->Size());
-  save_mean_ = malloc(mean->Size());
-  save_variance_ = malloc(variance->Size());
 
-  if (scale_ == nullptr || offset_ == nullptr || mean_ == nullptr || variance_ == nullptr || save_mean_ == nullptr ||
-      save_variance_ == nullptr) {
+  if (scale_ == nullptr || offset_ == nullptr || mean_ == nullptr || variance_ == nullptr) {
     FreeMeanAndVariance();
     FreeScaleAndOffset();
     MS_LOG(ERROR) << "Memory allocation failed";
@@ -73,61 +62,64 @@ int FusedBatchnormCPUKernel::InitConstTensor() {
   memcpy(offset_, offset->MutableData(), offset->Size());
   memcpy(mean_, mean->MutableData(), mean->Size());
   memcpy(variance_, variance->MutableData(), variance->Size());
-  memset(save_mean_, 0, mean->Size());
-  memset(save_variance_, 0, variance->Size());
-  if (out_tensors_.size() > 4) {
-    for (size_t i = 1; i < out_tensors_.size(); i++) {
-      auto *data = static_cast<float *>(out_tensors_[i]->MutableData());
-      std::fill(data, data + out_tensors_[i]->ElementsNum(), 0.f);
-    }
-  }
 
   return RET_OK;
 }
 
 int FusedBatchnormCPUKernel::Run() {
   auto param = reinterpret_cast<BatchNormParameter *>(op_parameter_);
-  if (is_train() && in_tensors_.size() >= 5) {
+  if (IsTrain() && IsTrainable() && in_tensors_.size() >= 5) {
     float *in = static_cast<float *>(in_tensors_[0]->MutableData());
     float *scale = static_cast<float *>(in_tensors_[1]->MutableData());
-    float *bias = static_cast<float *>(in_tensors_[2]->MutableData());
-    float *mean = static_cast<float *>(in_tensors_[3]->MutableData());
-    float *var = static_cast<float *>(in_tensors_[4]->MutableData());
-    std::fill(mean, mean + in_tensors_[3]->ElementsNum(), 0.f);
-    std::fill(var, var + in_tensors_[4]->ElementsNum(), 0.f);
-    FusedBatchNormFp32MeanVar(in, mean, var, param, static_cast<float *>(save_mean_),
-                              static_cast<float *>(save_variance_));
-    memcpy(out_tensors_[3]->MutableData(), save_mean_, out_tensors_[3]->Size());
-    memcpy(out_tensors_[4]->MutableData(), save_variance_, out_tensors_[3]->Size());
-    memcpy(mean_, mean, in_tensors_[3]->Size());
-    memcpy(variance_, var, in_tensors_[4]->Size());
+    float *offset = static_cast<float *>(in_tensors_[2]->MutableData());
+    float *current_mean = static_cast<float *>(mean_);
+    float *current_var = static_cast<float *>(variance_);
+    float *save_mean = static_cast<float *>(in_tensors_[3]->MutableData());
+    float *save_variance = static_cast<float *>(in_tensors_[4]->MutableData());
+
+    std::fill(current_mean, current_mean + in_tensors_[3]->ElementsNum(), 0.f);
+    std::fill(current_var, current_var + in_tensors_[4]->ElementsNum(), 0.f);
+    FusedBatchNormFp32MeanVar(in, current_mean, current_var, param, static_cast<float *>(save_mean),
+                              static_cast<float *>(save_variance));
+
+    memcpy(out_tensors_[1]->MutableData(), scale, out_tensors_[1]->Size());
+    memcpy(out_tensors_[2]->MutableData(), offset, out_tensors_[2]->Size());
+    memcpy(out_tensors_[3]->MutableData(), current_mean, out_tensors_[3]->Size());
+    memcpy(out_tensors_[4]->MutableData(), current_var, out_tensors_[4]->Size());
+
+    // Copy to local variables
     memcpy(scale_, scale, in_tensors_[1]->Size());
-    memcpy(offset_, bias, in_tensors_[2]->Size());
+    memcpy(offset_, offset, in_tensors_[2]->Size());
+
+    // save for next iteration
+    memcpy(in_tensors_[3]->MutableData(), save_mean, in_tensors_[3]->Size());
+    memcpy(in_tensors_[4]->MutableData(), save_variance, in_tensors_[4]->Size());
+
     trained_ = true;  // trained at least once
   }
   auto ret = ParallelLaunch(this->context_->thread_pool_, BatchNormRun, this, op_parameter_->thread_num_);
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "BatchnormRun error error_code[" << ret << "]";
   }
+
   return ret;
 }
 
-void FusedBatchnormCPUKernel::eval() {
-  LiteKernel::eval();
+int FusedBatchnormCPUKernel::Eval() {
+  LiteKernel::Eval();
   if (trained_) {
-    float *run_mean = static_cast<float *>(in_tensors_[3]->MutableData());
-    float *run_var = static_cast<float *>(in_tensors_[4]->MutableData());
+    float *save_mean = static_cast<float *>(in_tensors_[3]->MutableData());
+    float *save_var = static_cast<float *>(in_tensors_[4]->MutableData());
     float *scale = static_cast<float *>(in_tensors_[1]->MutableData());
     float *bias = static_cast<float *>(in_tensors_[2]->MutableData());
-    // Copy to input tensors for Model export
-    memcpy(run_mean, save_mean_, in_tensors_[3]->Size());
-    memcpy(run_var, save_variance_, in_tensors_[4]->Size());
+
     // Copy to local variables
-    memcpy(mean_, run_mean, in_tensors_[3]->Size());
-    memcpy(variance_, run_var, in_tensors_[4]->Size());
     memcpy(scale_, scale, in_tensors_[1]->Size());
     memcpy(offset_, bias, in_tensors_[2]->Size());
+    memcpy(mean_, save_mean, in_tensors_[3]->Size());
+    memcpy(variance_, save_var, in_tensors_[4]->Size());
   }
+  return RET_OK;
 }
 
 int FusedBatchnormCPUKernel::DoExecute(int task_id) {
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.h b/mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.h
index cea5a532d0..0265549c9f 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.h
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32/fused_batchnorm_fp32.h
@@ -29,7 +29,7 @@ class FusedBatchnormCPUKernel : public BatchnormCPUKernel {
       : BatchnormCPUKernel(parameter, inputs, outputs, ctx, primitive) {}
   ~FusedBatchnormCPUKernel() { FreeScaleAndOffset(); }
 
-  void eval() override;
+  int Eval() override;
   int ReSize() override;
   int Run() override;
   int InitConstTensor() override;
@@ -39,8 +39,6 @@ class FusedBatchnormCPUKernel : public BatchnormCPUKernel {
   void FreeScaleAndOffset();
   void *scale_ = nullptr;
   void *offset_ = nullptr;
-  void *save_mean_ = nullptr;
-  void *save_variance_ = nullptr;
   bool trained_ = false;
 };
 }  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/matmul_fp32.cc b/mindspore/lite/src/runtime/kernel/arm/fp32/matmul_fp32.cc
index c0d37b7dfd..d0c14e244b 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/matmul_fp32.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32/matmul_fp32.cc
@@ -140,10 +140,12 @@ int MatmulCPUKernel::InitBias() {
                     : (c_shape[c_shape.size() - 1]);
   params_->col_8_ = UP_ROUND(params_->col_, 8);
   auto col_tmp = is_vector_a_ ? params_->col_ : params_->col_8_;
-  bias_ptr_ = reinterpret_cast<float *>(malloc(col_tmp * sizeof(float)));
   if (bias_ptr_ == nullptr) {
-    FreeTmpBuffer();
-    return RET_MEMORY_FAILED;
+    bias_ptr_ = reinterpret_cast<float *>(malloc(col_tmp * sizeof(float)));
+    if (bias_ptr_ == nullptr) {
+      FreeTmpBuffer();
+      return RET_MEMORY_FAILED;
+    }
   }
   memset(bias_ptr_, 0, col_tmp * sizeof(float));
   if (in_tensors_.size() == 3) {
@@ -154,6 +156,8 @@ int MatmulCPUKernel::InitBias() {
 
 int MatmulCPUKernel::ReSize() {
   if (!params_->b_const_) {
+    free(bias_ptr_);
+    bias_ptr_ = nullptr;
     auto ret = InitBias();
     if (ret != RET_OK) {
       MS_LOG(ERROR) << "Matmul fp32 init bias failed";
@@ -277,7 +281,7 @@ int MatmulCPUKernel::Run() {
   auto b_src = reinterpret_cast<float *>(in_tensors_[1]->data_c());
   auto c_src = reinterpret_cast<float *>(out_tensors_[0]->data_c());
 
-  if (!params_->a_const_ || is_train()) {
+  if (!params_->a_const_ || IsTrain()) {
     if (a_pack_ptr_ != nullptr) {
       params_->a_const_ ? free(a_pack_ptr_) : context_->allocator->Free(a_pack_ptr_);
       a_pack_ptr_ = nullptr;
@@ -294,7 +298,7 @@ int MatmulCPUKernel::Run() {
       a_ptr_ = a_pack_ptr_;
     }
   }
-  if (!params_->b_const_ || is_train()) {
+  if (!params_->b_const_ || IsTrain()) {
     if (b_pack_ptr_ != nullptr) {
       params_->b_const_ ? free(b_pack_ptr_) : context_->allocator->Free(b_pack_ptr_);
       b_pack_ptr_ = nullptr;
@@ -311,7 +315,9 @@ int MatmulCPUKernel::Run() {
       b_ptr_ = b_pack_ptr_;
     }
   }
-
+  if (IsTrain()) {
+    InitBias();
+  }
   for (int i = 0; i < params_->batch; ++i) {
     if (is_vector_a_) {
       cur_a_ptr_ = a_ptr_ + i * params_->deep_;
@@ -329,26 +335,54 @@ int MatmulCPUKernel::Run() {
       return RET_ERROR;
     }
   }
-  if (!params_->a_const_ || is_train()) {
-    context_->allocator->Free(a_pack_ptr_);
+  if (!params_->a_const_ || IsTrain()) {
+    params_->a_const_ ? free(a_pack_ptr_) : context_->allocator->Free(a_pack_ptr_);
     a_pack_ptr_ = nullptr;
   }
-  if (!params_->b_const_ || is_train()) {
-    context_->allocator->Free(b_pack_ptr_);
+  if (!params_->b_const_ || IsTrain()) {
+    params_->b_const_ ? free(b_pack_ptr_) : context_->allocator->Free(b_pack_ptr_);
     b_pack_ptr_ = nullptr;
   }
   return RET_OK;
 }
 
-void MatmulCPUKernel::eval() {
+int MatmulCPUKernel::Eval() {
   // Copy weights after training
-  LiteKernel::eval();
+  auto a_src = reinterpret_cast<float *>(in_tensors_[0]->data_c());
+  auto b_src = reinterpret_cast<float *>(in_tensors_[1]->data_c());
+  LiteKernel::Eval();
   if (params_->a_const_) {
-    InitMatrixA(reinterpret_cast<float *>(in_tensors_[0]->MutableData()), a_pack_ptr_);
+    if (a_pack_ptr_ == nullptr) {
+      auto ret = MallocMatrixABuffer();
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Matmul fp32 malloc matrix a buffer failed";
+        return RET_ERROR;
+      }
+    }
+    if (is_vector_a_) {
+      a_ptr_ = a_src;
+    } else {
+      InitMatrixA(a_src, a_pack_ptr_);
+      a_ptr_ = a_pack_ptr_;
+    }
   }
   if (params_->b_const_) {
-    InitMatrixB(reinterpret_cast<float *>(in_tensors_[1]->MutableData()), b_pack_ptr_);
+    if (b_pack_ptr_ == nullptr) {
+      auto ret = MallocMatrixBBuffer();
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << "Matmul fp32 malloc matrix b buffer failed";
+        return RET_ERROR;
+      }
+    }
+    if (is_vector_a_ && params_->b_transpose_) {
+      b_ptr_ = b_src;
+    } else {
+      InitMatrixB(b_src, b_pack_ptr_);
+      b_ptr_ = b_pack_ptr_;
+    }
   }
+  InitBias();
+  return RET_OK;
 }
 
 kernel::LiteKernel *CpuMatmulFp32KernelCreator(const std::vector<lite::Tensor *> &inputs,
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/matmul_fp32.h b/mindspore/lite/src/runtime/kernel/arm/fp32/matmul_fp32.h
index 68ca588761..f48e53d608 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/matmul_fp32.h
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32/matmul_fp32.h
@@ -34,7 +34,7 @@ class MatmulCPUKernel : public MatmulBaseCPUKernel {
   int ReSize() override;
   int Run() override;
   int RunImpl(int task_id);
-  void eval() override;
+  int Eval() override;
 
  private:
   int MallocMatrixABuffer();
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32/one_hot_fp32.cc b/mindspore/lite/src/runtime/kernel/arm/fp32/one_hot_fp32.cc
index 619e33d03b..4b452297cd 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32/one_hot_fp32.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32/one_hot_fp32.cc
@@ -214,5 +214,5 @@ kernel::LiteKernel *CpuOneHotFp32KernelCreator(const std::vector<lite::Tensor *>
   return kernel;
 }
 
-REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_OneHot, CpuOneHotFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeInt32, PrimitiveType_OneHot, CpuOneHotFp32KernelCreator)
 }  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/adam.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/adam.cc
index 81cfa992d0..65572eebf0 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/adam.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/adam.cc
@@ -45,24 +45,19 @@ int AdamCPUKernel::Execute(int task_id) {
   auto eps = reinterpret_cast<float *>(in_tensors_[8]->MutableData())[0];
   auto gradient = reinterpret_cast<float *>(in_tensors_[9]->MutableData());
   size_t elem_num = in_tensors_[0]->ElementsNum();
+  auto update_lr = learning_rate * std::sqrt(1 - beta2_power) / (1 - beta1_power);
 
   if (adam_param_->use_nesterov_) {  // Nadam
     for (size_t i = 0; i < elem_num; ++i) {
-      m[i] = (m[i] * beta1) + (gradient[i] * (1.f - beta1));
-      v[i] = (v[i] * beta2) + (gradient[i] * gradient[i] * (1.f - beta2));
-      auto g_hat = gradient[i] / (1 - beta1_power);
-      auto m_hat = m[i] / (1 - beta1_power);
-      auto v_hat = v[i] / (1 - beta2_power);
-      auto m_tag = (1.f - beta1) * g_hat + beta1 * m_hat;
-      weight[i] -= learning_rate * m_tag / (sqrtf(v_hat) + eps);
+      m[i] += (gradient[i] - m[i]) * (1 - beta1);
+      v[i] += (gradient[i] * gradient[i] - v[i]) * (1 - beta2);
+      weight[i] -= update_lr * (m[i] * beta1 + (1 - beta1) * gradient[i]) / (std::sqrt(v[i]) + eps);
     }
   } else {
     for (size_t i = 0; i < elem_num; ++i) {
-      m[i] = (m[i] * beta1) + (gradient[i] * (1.f - beta1));
-      v[i] = (v[i] * beta2) + (gradient[i] * gradient[i] * (1.f - beta2));
-      auto m_hat = m[i] / (1 - beta1_power);
-      auto v_hat = v[i] / (1 - beta2_power);
-      weight[i] -= learning_rate * m_hat / (sqrtf(v_hat) + eps);
+      m[i] += (gradient[i] - m[i]) * (1 - beta1);
+      v[i] += (gradient[i] * gradient[i] - v[i]) * (1 - beta2);
+      weight[i] -= update_lr * m[i] / (std::sqrt(v[i]) + eps);
     }
   }
   return RET_OK;
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.cc
index e847c3cf64..833f022522 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.cc
@@ -177,6 +177,28 @@ void ArithmeticGradCPUKernel::ArithmeticGradDiv2L(float *dy, int dy_size, float
   ElementDivNegSquare(tile_data2, x2_data, dx2, dy_size);
 }
 
+void ArithmeticGradCPUKernel::ArithmeticGradMaximum(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                    int dx2_size) {
+  // For some reason, input order is x0, x1, dy
+  auto x1 = reinterpret_cast<float *>(in_tensors_[0]->MutableData());
+  auto x2 = reinterpret_cast<float *>(in_tensors_[1]->MutableData());
+  dy = reinterpret_cast<float *>(in_tensors_[2]->MutableData());
+
+  MaximumByAxes(x1, x2, dy, arithmeticParameter_->in_shape0_, arithmeticParameter_->in_shape1_,
+                arithmeticParameter_->out_shape_, dx1, dx2, arithmeticParameter_->ndim_);
+}
+
+void ArithmeticGradCPUKernel::ArithmeticGradMinimum(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
+                                                    int dx2_size) {
+  // For some reason, input order is x0, x1, dy
+  auto x1 = reinterpret_cast<float *>(in_tensors_[0]->MutableData());
+  auto x2 = reinterpret_cast<float *>(in_tensors_[1]->MutableData());
+  dy = reinterpret_cast<float *>(in_tensors_[2]->MutableData());
+
+  MinimumByAxes(x1, x2, dy, arithmeticParameter_->out_shape_, arithmeticParameter_->in_shape0_,
+                arithmeticParameter_->in_shape1_, dx1, dx2, arithmeticParameter_->ndim_);
+}
+
 int ArithmeticGradCPUKernel::ReSize() { return RET_OK; }
 
 int ArithmeticGradCPUKernel::Execute(int task_id) {
@@ -240,4 +262,6 @@ REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_MulGrad, CpuArithmeticGradFp3
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_AddGrad, CpuArithmeticGradFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_SubGrad, CpuArithmeticGradFp32KernelCreator)
 REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_DivGrad, CpuArithmeticGradFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_MaximumGrad, CpuArithmeticGradFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_MinimumGrad, CpuArithmeticGradFp32KernelCreator)
 }  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.h b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.h
index 7f480daf4c..c11d01f079 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.h
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.h
@@ -24,6 +24,8 @@
 
 using mindspore::schema::PrimitiveType_AddGrad;
 using mindspore::schema::PrimitiveType_DivGrad;
+using mindspore::schema::PrimitiveType_MaximumGrad;
+using mindspore::schema::PrimitiveType_MinimumGrad;
 using mindspore::schema::PrimitiveType_MulGrad;
 using mindspore::schema::PrimitiveType_SubGrad;
 
@@ -52,6 +54,12 @@ class ArithmeticGradCPUKernel : public LiteKernel {
       case PrimitiveType_DivGrad:
         arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradDiv;  // this will be adjusted in InferShape
         break;
+      case PrimitiveType_MaximumGrad:
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradMaximum;
+        break;
+      case PrimitiveType_MinimumGrad:
+        arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradMinimum;
+        break;
       default:
         MS_LOG(ERROR) << "Error Operator type " << parameter->type_;
         break;
@@ -79,6 +87,8 @@ class ArithmeticGradCPUKernel : public LiteKernel {
   void ArithmeticGradDiv(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
   void ArithmeticGradDiv1L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
   void ArithmeticGradDiv2L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  void ArithmeticGradMaximum(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
+  void ArithmeticGradMinimum(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2, int dx2_size);
   ArithmeticParameter *arithmeticParameter_;
   ArithmeticGradOperation arithmetic_grad_;
   float *tile_data0;
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/bn_grad.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/bn_grad.cc
index c821003cec..12c8002b70 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/bn_grad.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/bn_grad.cc
@@ -15,6 +15,7 @@
  */
 
 #include "src/runtime/kernel/arm/fp32_grad/bn_grad.h"
+#include <math.h>
 #include <algorithm>
 #include <vector>
 #include "schema/model_generated.h"
@@ -34,7 +35,7 @@ namespace mindspore::kernel {
 int BNGradCPUKernel::Init() {
   auto *input_x = in_tensors_.at(1);
   int channels = input_x->shape().at(kNHWC_C);
-  SetWorkspaceSize(4 * channels * sizeof(float));
+  SetWorkspaceSize(2 * channels * sizeof(float));
   return RET_OK;
 }
 
@@ -45,19 +46,23 @@ int BNGradCPUKernel::Execute(int task_id) {
   auto *input_yt = in_tensors_.at(0);
   auto *input_x = in_tensors_.at(1);
   auto *input_scale = in_tensors_.at(2);
+  auto *input_mean = in_tensors_.at(3);
+  auto *input_var = in_tensors_.at(4);
+
+  float *save_mean = reinterpret_cast<float *>(input_mean->MutableData());
+  float *save_var = reinterpret_cast<float *>(input_var->MutableData());
+
   auto *output_dx = out_tensors_.at(0);
   auto *output_scale = out_tensors_.at(1);
   auto *output_bias = out_tensors_.at(2);
-  int batch = input_x->Batch();
-  int channels = input_x->Channel();
-  int spatial = input_x->Height() * input_x->Width();
+  size_t batch = input_x->Batch();
+  size_t channels = input_x->Channel();
+  size_t spatial = input_x->Height() * input_x->Width();
   float eps = bn_param->epsilon_;
 
   float *workspace = static_cast<float *>(GetWorkspace());
   std::fill(workspace, workspace + GetWorkspaceSize() / sizeof(*workspace), 0.f);
-  float *mean = workspace;
-  float *invar = mean + channels;
-  float *dxhat_sum = invar + channels;
+  float *dxhat_sum = workspace;
   float *dxhathat_sum = dxhat_sum + channels;
 
   float *x = reinterpret_cast<float *>(input_x->MutableData());
@@ -67,11 +72,14 @@ int BNGradCPUKernel::Execute(int task_id) {
   float *dscale = reinterpret_cast<float *>(output_scale->MutableData());
   float *dbias = reinterpret_cast<float *>(output_bias->MutableData());
 
-  backwardX(x, yt, scale, batch * spatial, channels, eps, mean, invar, dxhat_sum, dxhathat_sum, dx);
+  var2Invar(save_var, input_var->ElementsNum(), eps);
+  // dx
+  backwardX(x, yt, scale, batch * spatial, channels, save_mean, save_var, dxhat_sum, dxhathat_sum, dx);
   // dbias
   sumSpatialBatch(yt, batch * spatial, channels, dbias);
   // dscale
-  backwardScale(x, mean, invar, yt, batch, channels, spatial, dscale);
+  backwardScale(x, save_mean, save_var, yt, batch, channels, spatial, dscale);
+
   return RET_OK;
 }
 
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.cc
index e3f27962ba..af2f29b4ef 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.cc
@@ -19,6 +19,7 @@
 #include "nnacl/fp32_grad/gemm.h"
 #include "include/errorcode.h"
 #include "src/runtime/runtime_api.h"
+#include "nnacl/pack.h"
 
 using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::RET_ERROR;
@@ -26,8 +27,8 @@ using mindspore::lite::RET_OK;
 
 namespace mindspore::kernel {
 int ConvolutionTrainCPUKernel::Init() {
-  if (2 != in_tensors_.size()) {
-    MS_LOG(ERROR) << "Convolution should have two inputs";
+  if (2 > in_tensors_.size()) {
+    MS_LOG(ERROR) << "Convolution should have at least two inputs";
     return RET_ERROR;
   }
   if (1 != out_tensors_.size()) {
@@ -51,11 +52,11 @@ int ConvolutionTrainCPUKernel::Init() {
   conv_param_->kernel_w_ = input_weight->shape().at(kNHWC_W);
 
   conv_param_->group_ = (conv_param_->group_ == 0) ? conv_param_->input_channel_ : conv_param_->group_;
-
-  int ws_size = conv_param_->output_h_ * conv_param_->output_w_ * conv_param_->kernel_h_ * conv_param_->kernel_w_ *
-                conv_param_->input_channel_ / conv_param_->group_;
-
-  SetWorkspaceSize(ws_size * sizeof(float));
+  const int n = conv_param_->output_channel_ * conv_param_->group_;
+  const int k = conv_param_->kernel_h_ * conv_param_->kernel_w_ * conv_param_->input_channel_ / conv_param_->group_;
+  ws_size = chunk * k;
+  int mat_alloc = MatSizeTotal(chunk, n, k, 0);
+  SetWorkspaceSize((ws_size + mat_alloc) * sizeof(float));
   return RET_OK;
 }
 
@@ -71,36 +72,35 @@ int ConvolutionTrainCPUKernel::Execute(int task_id) {
   auto y_addr = reinterpret_cast<float *>(out_y->MutableData());
   auto w_addr = reinterpret_cast<float *>(input_w->MutableData());
 
-  int i, j;
-  int nweights = input_w->ElementsNum();
-  int in_ch = conv_param_->input_channel_;
-  int in_h = conv_param_->input_h_;
-  int in_w = conv_param_->input_w_;
-  int k_h = conv_param_->kernel_h_;
-  int k_w = conv_param_->kernel_w_;
-  int batch = conv_param_->output_batch_;
-  int out_ch = conv_param_->output_channel_;  // out_y->shape()[3];
-  int groups = conv_param_->group_;
-  int out_h = conv_param_->output_h_;
-  int out_w = conv_param_->output_w_;
-  int m = out_h * out_w;
-  int n = out_ch / groups;
-  int k = k_h * k_w * in_ch / groups;
+  const int nweights = input_w->ElementsNum();
+  const int in_ch = conv_param_->input_channel_;
+  const int in_h = conv_param_->input_h_;
+  const int in_w = conv_param_->input_w_;
+  const int k_h = conv_param_->kernel_h_;
+  const int k_w = conv_param_->kernel_w_;
+  const int batch = conv_param_->output_batch_;
+  const int out_ch = conv_param_->output_channel_;  // out_y->shape()[3];
+  const int groups = conv_param_->group_;
+  const int out_h = conv_param_->output_h_;
+  const int out_w = conv_param_->output_w_;
+  const int m = out_h * out_w;
+  const int n = out_ch / groups;
+  const int k = k_h * k_w * in_ch / groups;
   float *workspace = static_cast<float *>(GetWorkspace());
-
-  memset(y_addr, 0, out_y->Size());
-
-  for (i = 0; i < batch; ++i) {
-    for (j = 0; j < groups; ++j) {
-      float *mat_a = workspace;
-      float *mat_b = w_addr + j * nweights / groups;
-      float *mat_c = y_addr + (i * groups) * n * m + j * (out_ch / groups);
-      float *im = x_addr + (i * groups) * (in_ch / groups) * in_h * in_w + j * (in_ch / groups);
-      im2col_hwc(im, mat_a, conv_param_);
-      gemm(0, 1, m, n, k, 1, mat_a, k, mat_b, k, 1, mat_c, out_ch);
+  float *mat_workspace = workspace + ws_size;
+  for (int i = 0; i < batch; ++i) {
+    for (int j = 0; j < groups; ++j) {
+      for (int ci = 0; ci < m; ci += chunk) {
+        int real_chunk = MSMIN(m - ci, chunk);
+        float *mat_a = workspace;
+        const float *mat_b = w_addr + j * nweights / groups;
+        float *mat_c = y_addr + (i * groups) * n * m + j * (out_ch / groups) + ci * out_ch;
+        float *im = x_addr + (i * groups) * (in_ch / groups) * in_h * in_w + j * (in_ch / groups);
+        RollingIm2ColPackUnitFp32(im, conv_param_, mat_a, real_chunk, ci);
+        GemmMatmul(0, 1, real_chunk, n, k, 1, mat_a, k, mat_b, k, 0, mat_c, out_ch, mat_workspace);
+      }
     }
   }
-
   return RET_OK;
 }
 
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.h b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.h
index dd92d28183..dd212e7f87 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.h
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.h
@@ -33,6 +33,14 @@ class ConvolutionTrainCPUKernel : public LiteKernel {
   int ReSize() override;
   int Run() override;
   int Execute(int task_id);
+
+ private:
+  int ws_size = 0;
+#ifdef ENABLE_ARM32
+  const int chunk = C4NUM;
+#else
+  const int chunk = C12NUM;
+#endif
 };
 
 kernel::LiteKernel *CpuConvTrainFp32KernelCreator(const std::vector<lite::Tensor *> &inputs,
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.cc
index c72306b009..5c73bbfe63 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.cc
@@ -51,10 +51,12 @@ int ConvolutionGradFilterCPUKernel::Init() {
   conv_param->output_h_ = dy_tensor->shape()[kNHWC_H];
   conv_param->output_w_ = dy_tensor->shape()[kNHWC_W];
 
-  size_t ws_size = conv_param->output_h_ * conv_param->output_w_ * conv_param->kernel_h_ * conv_param->kernel_w_ *
-                   conv_param->input_channel_ / conv_param->group_;
+  ws_size = chunk * conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_ / conv_param->group_;
 
-  SetWorkspaceSize(ws_size * sizeof(float));
+  int n = conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_ / conv_param->group_;
+  int k = conv_param->output_channel_ / conv_param->group_;
+  size_t mat_alloc = MatSizeTotal(k, n, chunk, n);
+  SetWorkspaceSize((ws_size + mat_alloc) * sizeof(float));
   return RET_OK;
 }
 
@@ -88,19 +90,21 @@ int ConvolutionGradFilterCPUKernel::Execute(int task_id) {
   int k = out_ch / groups;
 
   float *workspace = reinterpret_cast<float *>(GetWorkspace());
-
+  float *mat_workspace = workspace + ws_size;
   // zero out pointer
   memset(dw_addr, 0, out_dw->Size());
-
   for (i = 0; i < batch; ++i) {
     for (j = 0; j < groups; ++j) {
-      float *mat_a = dy_addr + (i * groups) * m * k + j * (out_ch / groups);
-      float *mat_b = workspace;
-      float *mat_c = dw_addr + j * nweights / groups;
-      float *im = x_addr + (i * in_ch * in_h * in_w) + j * (in_ch / groups);
-
-      im2row_hwc(im, mat_b, conv_param, false);
-      gemm(1, 1, k, n, m, 1, mat_a, out_ch, mat_b, m, 1, mat_c, n);
+      for (int ci = 0; ci < m; ci += chunk) {
+        int real_chunk = MSMIN(m - ci, chunk);
+        float *mat_a = dy_addr + (i * groups) * m * k + j * (out_ch / groups) + ci * out_ch;
+        float *mat_b = workspace;
+        float *mat_c = dw_addr + j * nweights / groups;
+        float *im = x_addr + (i * in_ch * in_h * in_w) + j * (in_ch / groups);
+        memset(mat_b, 0, n * real_chunk * sizeof(float));
+        RollingIm2ColPackUnitFp32(im, conv_param, mat_b, real_chunk, ci);
+        GemmMatmul(1, 0, k, n, real_chunk, 1, mat_a, out_ch, mat_b, n, 1, mat_c, n, mat_workspace);
+      }
     }
   }
   return RET_OK;
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.h b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.h
index a8eaefdafc..763abc7612 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.h
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.h
@@ -34,6 +34,14 @@ class ConvolutionGradFilterCPUKernel : public LiteKernel {
   int ReSize() override;
   int Run() override;
   int Execute(int task_id);
+
+ private:
+  size_t ws_size = 0;
+#ifdef ENABLE_ARM32
+  const int chunk = C4NUM;
+#else
+  const int chunk = C12NUM;
+#endif
 };
 
 }  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.cc
index e7bbd1ce5b..72082b5066 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.cc
@@ -51,11 +51,14 @@ int ConvolutionGradInputCPUKernel::Init() {
 
   conv_param->output_h_ = dy_tensor->shape()[kNHWC_H];
   conv_param->output_w_ = dy_tensor->shape()[kNHWC_W];
+  ws_size = chunk * conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_ / conv_param->group_;
 
-  size_t ws_size = conv_param->output_h_ * conv_param->output_w_ * conv_param->kernel_h_ * conv_param->kernel_w_ *
-                   conv_param->input_channel_ / conv_param->group_;
+  int n = conv_param->kernel_w_ * conv_param->kernel_h_ * conv_param->input_channel_ / conv_param->group_;
+  int k = conv_param->output_channel_ / conv_param->group_;
 
-  SetWorkspaceSize(ws_size * sizeof(float));
+  size_t mat_alloc = MatSizeTotal(chunk, n, k, 0);
+
+  SetWorkspaceSize((ws_size + mat_alloc) * sizeof(float));
   return RET_OK;
 }
 
@@ -88,16 +91,30 @@ int ConvolutionGradInputCPUKernel::Execute(int task_id) {
   int n = k_w * k_h * in_ch / groups;
   int k = out_ch / groups;
   float *workspace = reinterpret_cast<float *>(GetWorkspace());
-
+  float *mat_workspace = workspace + ws_size;
   memset(dx_addr, 0, sizeof(float) * batch * in_ch * in_h * in_w);
-
   for (i = 0; i < batch; ++i) {
     for (j = 0; j < groups; ++j) {
-      float *mat_a = dy_addr + (i * groups) * m * k + j * (out_ch / groups);
-      float *mat_b = w_addr + j * nweights / groups;
-      float *mat_c = workspace;
-      gemm(0, 0, m, n, k, 1, mat_a, out_ch, mat_b, n, 0, mat_c, n);
-      col2im_hwc(mat_c, dx_addr + (i * groups) * (in_ch / groups) * in_h * in_w + j * (in_ch / groups), conv_param);
+      GemmCb gcb;
+      for (int ci = 0; ci < m; ci += chunk) {
+        float *mat_b;
+        if (ci == 0) {
+          mat_b = w_addr + j * nweights / groups;
+          gcb.ca = 0;
+          gcb.cb = 0;
+          gcb.bias = nullptr;
+          gcb.atype = ActType_No;
+        } else {
+          mat_b = gcb.mat_b;
+          gcb.cb = 1;
+        }
+        int real_chunk = MSMIN(m - ci, chunk);
+        float *mat_a = dy_addr + (i * groups) * m * k + j * (out_ch / groups) + ci * out_ch;
+        float *mat_c = workspace;
+        GemmMatmulPlus(0, 0, real_chunk, n, k, 1, mat_a, out_ch, mat_b, n, 0, mat_c, n, mat_workspace, &gcb);
+        rolling_col2im_hwc(mat_c, dx_addr + (i * groups) * (in_ch / groups) * in_h * in_w + j * (in_ch / groups),
+                           conv_param, real_chunk, ci);
+      }
     }
   }
 
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.h b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.h
index 6bea61b59c..d4b226dd9b 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.h
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.h
@@ -33,6 +33,14 @@ class ConvolutionGradInputCPUKernel : public LiteKernel {
   int ReSize() override;
   int Run() override;
   int Execute(int task_id);
+
+ private:
+  size_t ws_size = 0;
+#ifdef ENABLE_ARM32
+  const int chunk = C4NUM;
+#else
+  const int chunk = C12NUM;
+#endif
 };
 }  // namespace mindspore::kernel
 
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_filter.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_filter.cc
index 0133ffb0f4..34e425d10c 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_filter.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_filter.cc
@@ -50,10 +50,14 @@ int DeConvolutionGradFilterCPUKernel::Init() {
   conv_param->output_h_ = dy_tensor->shape()[kNHWC_H];
   conv_param->output_w_ = dy_tensor->shape()[kNHWC_W];
 
-  int ws_size = conv_param->input_h_ * conv_param->input_w_ * conv_param->kernel_h_ * conv_param->kernel_w_ *
-                conv_param->output_channel_ / conv_param->group_;
+  ws_size = chunk * conv_param->input_w_ * conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->output_channel_ /
+            conv_param->group_;
 
-  SetWorkspaceSize(ws_size * sizeof(float));
+  int m = conv_param->input_channel_ / conv_param->group_;
+  int n = conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->output_channel_ / conv_param->group_;
+  size_t mat_alloc = MatSizeTotal(n, m, chunk * conv_param->input_w_, conv_param->input_channel_);
+
+  SetWorkspaceSize((ws_size + mat_alloc) * sizeof(float));
 
   return RET_OK;
 }
@@ -82,21 +86,25 @@ int DeConvolutionGradFilterCPUKernel::Execute(int task_id) {
   int out_h = conv_param->output_h_;
   int out_w = conv_param->output_w_;
 
-  int m = in_ch / groups;
-  int n = k_h * k_w * out_ch / groups;
-  int k = in_h * in_w;
+  const int m = in_ch / groups;
+  const int n = k_h * k_w * out_ch / groups;
 
   float *workspace = reinterpret_cast<float *>(GetWorkspace());
+  float *mat_workspace = workspace + ws_size;
   // zero out pointer
   memset(dw_addr, 0, out_dw->Size());
   for (i = 0; i < batch; ++i) {
     for (j = 0; j < groups; ++j) {
-      float *mat_a = x_addr + (i * (in_ch * in_h * in_w) + j * (in_ch / groups));
-      float *mat_b = workspace;
-      float *mat_c = dw_addr + j * m;
-      float *im = dy_addr + (i * (out_h * out_w * out_ch) + j * (out_ch / groups));
-      im2row_hwc(im, mat_b, conv_param, true);
-      gemm(0, 0, n, m, k, 1, mat_b, k, mat_a, in_ch, 1, mat_c, in_ch);
+      for (int ci = 0; ci < in_h; ci += chunk) {
+        int real_chunk = MSMIN(in_h - ci, chunk);
+        float *mat_a = x_addr + (i * (in_ch * in_h * in_w) + j * (in_ch / groups)) + ci * in_w * in_ch;
+        float *mat_b = workspace;
+        float *mat_c = dw_addr + j * m;
+        float *im = dy_addr + (i * (out_h * out_w * out_ch) + j * (out_ch / groups));
+        rolling_im2row_hwc(im, mat_b, conv_param, real_chunk, ci);
+        GemmMatmul(0, 0, n, m, real_chunk * in_w, 1, mat_b, real_chunk * in_w, mat_a, in_ch, 1, mat_c, in_ch,
+                   mat_workspace);
+      }
     }
   }
   return RET_OK;
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_filter.h b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_filter.h
index 0737cb1009..a95b4e484a 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_filter.h
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_filter.h
@@ -33,6 +33,10 @@ class DeConvolutionGradFilterCPUKernel : public LiteKernel {
   int ReSize() override;
   int Run() override;
   int Execute(int task_id);
+
+ private:
+  size_t ws_size = 0;
+  const int chunk = 1;
 };
 
 }  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout.cc
new file mode 100644
index 0000000000..7fa2eafa8b
--- /dev/null
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout.cc
@@ -0,0 +1,131 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <random>
+#include <algorithm>
+#include "src/runtime/kernel/arm/fp32_grad/dropout.h"
+#include "schema/model_generated.h"
+#include "src/runtime/runtime_api.h"
+#include "src/kernel_registry.h"
+#include "include/errorcode.h"
+#include "nnacl/fp32_grad/dropout_parameter.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_NULL_PTR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_Dropout;
+
+namespace mindspore::kernel {
+
+int DropoutCPUKernel::Init() {
+  auto param = reinterpret_cast<DropoutParameter *>(op_parameter_);
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Dropout op_parameter_ nullptr";
+    return RET_NULL_PTR;
+  }
+
+  if ((param->ratio_ > 1.0f) || (param->ratio_ < 0.0f)) {
+    MS_LOG(ERROR) << "unsupported ratio value - Dropout ratio should be between zero to one";
+    return RET_ERROR;
+  }
+
+  if (param->ratio_ >= 1.0f) {
+    scale_ = 1.0f;
+  } else {
+    scale_ = 1. / (1. - param->ratio_);
+  }
+  if (!InferShapeDone()) {
+    return RET_OK;
+  }
+  return ReSize();
+}
+
+int DropoutCPUKernel::ReSize() { return RET_OK; }
+
+int DropoutCPUKernel::Execute(int task_id) {
+  auto input_ptr = reinterpret_cast<float *>(in_tensors_.at(kInputIndex)->MutableData());
+  auto output_ptr = reinterpret_cast<float *>(out_tensors_.at(kOutputIndex)->MutableData());
+  auto mask = reinterpret_cast<float *>(out_tensors_.at(1)->MutableData());
+  auto length = in_tensors_.at(kInputIndex)->ElementsNum();
+  auto param = reinterpret_cast<DropoutParameter *>(op_parameter_);
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Dropout op_parameter_ nullptr";
+    return RET_NULL_PTR;
+  }
+  if (IsEval()) {
+    std::copy(input_ptr, input_ptr + length, output_ptr);
+  } else {
+    std::default_random_engine generator;
+    std::bernoulli_distribution distribution(param->ratio_);
+
+    for (int i = 0; i < length; i++) {
+      mask[i] = distribution(generator);
+      output_ptr[i] = input_ptr[i] * mask[i] * scale_;
+    }
+  }
+  return RET_OK;
+}
+
+int RunDropout(void *cdata, int task_id) {
+  auto dropout = reinterpret_cast<DropoutCPUKernel *>(cdata);
+  auto error_code = dropout->Execute(task_id);
+  if (error_code != RET_OK) {
+    MS_LOG(ERROR) << "Dropout Run error task_id[" << task_id << "] error_code[" << error_code << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int DropoutCPUKernel::Run() {
+  int error_code = ParallelLaunch(this->context_->thread_pool_, RunDropout, this, 1);
+  if (error_code != RET_OK) {
+    MS_LOG(ERROR) << "Dropout function error error_code[" << error_code << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuDropoutFp32KernelCreator(const std::vector<lite::Tensor *> &inputs,
+                                                const std::vector<lite::Tensor *> &outputs, OpParameter *opParameter,
+                                                const lite::InnerContext *ctx, const kernel::KernelKey &desc,
+                                                const mindspore::lite::PrimitiveC *primitive) {
+  if (opParameter == nullptr) {
+    MS_LOG(ERROR) << "Dropout opParameter nullptr.";
+    return nullptr;
+  }
+  if (desc.type != schema::PrimitiveType_Dropout) {
+    MS_LOG(ERROR) << "Dropout desc type should be " << schema::PrimitiveType_Dropout << " got " << desc.type;
+    return nullptr;
+  }
+  auto *kernel = new (std::nothrow) DropoutCPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "Dropout new kernel failed.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_Dropout, CpuDropoutFp32KernelCreator)
+}  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout.h b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout.h
new file mode 100644
index 0000000000..2b4093a73b
--- /dev/null
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout.h
@@ -0,0 +1,43 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_DROPOUT_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_DROPOUT_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+
+namespace mindspore::kernel {
+class DropoutCPUKernel : public LiteKernel {
+ public:
+  DropoutCPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs,
+                   const std::vector<lite::Tensor *> &outputs, const lite::InnerContext *ctx,
+                   const mindspore::lite::PrimitiveC *primitive)
+      : LiteKernel(parameter, inputs, outputs, ctx, primitive) {}
+
+  ~DropoutCPUKernel() override = default;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+  int Execute(int task_id);
+
+ private:
+  float scale_;
+};
+
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_DROPOUT_H_
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout_grad.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout_grad.cc
new file mode 100644
index 0000000000..bb62ba40f8
--- /dev/null
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout_grad.cc
@@ -0,0 +1,118 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <random>
+#include "src/runtime/kernel/arm/fp32_grad/dropout_grad.h"
+#include "nnacl/fp32_grad/dropout_grad.h"
+#include "schema/model_generated.h"
+#include "src/runtime/runtime_api.h"
+#include "src/kernel_registry.h"
+#include "include/errorcode.h"
+#include "nnacl/fp32_grad/dropout_parameter.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_NULL_PTR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_DropoutGrad;
+
+namespace mindspore::kernel {
+
+int DropoutGradCPUKernel::Init() {
+  auto param = reinterpret_cast<DropoutParameter *>(op_parameter_);
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Dropout op_parameter_ nullptr";
+    return RET_NULL_PTR;
+  }
+
+  if ((param->ratio_ > 1.0f) || (param->ratio_ < 0.0f)) {
+    MS_LOG(ERROR) << "unsupported ratio value - Dropout ratio should be between zero to one";
+    return RET_ERROR;
+  }
+
+  if (param->ratio_ >= 1.0f) {
+    scale_ = 1.0f;
+  } else {
+    scale_ = 1. / (1. - param->ratio_);
+  }
+  if (!InferShapeDone()) {
+    return RET_OK;
+  }
+  return ReSize();
+}
+
+int DropoutGradCPUKernel::ReSize() { return RET_OK; }
+
+int DropoutGradCPUKernel::Execute(int task_id) {
+  auto yt_ptr = reinterpret_cast<float *>(in_tensors_.at(kInputIndex)->MutableData());
+  auto mask_ptr = reinterpret_cast<float *>(in_tensors_.at(1)->MutableData());
+  auto output_ptr = reinterpret_cast<float *>(out_tensors_.at(kOutputIndex)->MutableData());
+  auto length = in_tensors_.at(kInputIndex)->ElementsNum();
+  DropoutGrad(yt_ptr, mask_ptr, output_ptr, length, scale_);
+
+  return RET_OK;
+}
+
+int RunDropoutGrad(void *cdata, int task_id) {
+  auto dropout = reinterpret_cast<DropoutGradCPUKernel *>(cdata);
+  auto error_code = dropout->Execute(task_id);
+  if (error_code != RET_OK) {
+    MS_LOG(ERROR) << "Dropout Grad Run error task_id[" << task_id << "] error_code[" << error_code << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+int DropoutGradCPUKernel::Run() {
+  int error_code = ParallelLaunch(this->context_->thread_pool_, RunDropoutGrad, this, 1);
+  if (error_code != RET_OK) {
+    MS_LOG(ERROR) << "Dropout Grad function error error_code[" << error_code << "]";
+    return RET_ERROR;
+  }
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuDropoutGradFp32KernelCreator(const std::vector<lite::Tensor *> &inputs,
+                                                    const std::vector<lite::Tensor *> &outputs,
+                                                    OpParameter *opParameter, const lite::InnerContext *ctx,
+                                                    const kernel::KernelKey &desc,
+                                                    const mindspore::lite::PrimitiveC *primitive) {
+  if (opParameter == nullptr) {
+    MS_LOG(ERROR) << "DropoutGrad opParameter nullptr.";
+    return nullptr;
+  }
+  if (desc.type != schema::PrimitiveType_DropoutGrad) {
+    MS_LOG(ERROR) << "DropoutGrad desc type should be " << schema::PrimitiveType_DropoutGrad << " got " << desc.type;
+    return nullptr;
+  }
+  auto *kernel = new (std::nothrow) DropoutGradCPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  if (kernel == nullptr) {
+    MS_LOG(ERROR) << "DropoutGrad new kernel failed.";
+    return nullptr;
+  }
+  auto ret = kernel->Init();
+  if (ret != RET_OK) {
+    delete kernel;
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_DropoutGrad, CpuDropoutGradFp32KernelCreator)
+}  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout_grad.h b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout_grad.h
new file mode 100644
index 0000000000..1740656dde
--- /dev/null
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/dropout_grad.h
@@ -0,0 +1,43 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_DROPOUT_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_DROPOUT_GRAD_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+
+namespace mindspore::kernel {
+class DropoutGradCPUKernel : public LiteKernel {
+ public:
+  DropoutGradCPUKernel(OpParameter *parameter, const std::vector<lite::Tensor *> &inputs,
+                       const std::vector<lite::Tensor *> &outputs, const lite::InnerContext *ctx,
+                       const mindspore::lite::PrimitiveC *primitive)
+      : LiteKernel(parameter, inputs, outputs, ctx, primitive) {}
+
+  ~DropoutGradCPUKernel() override = default;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+  int Execute(int task_id);
+
+ private:
+  float scale_;
+};
+
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_DROPOUT_GRAD_H_
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/softmax_cross_entropy_with_logits.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/softmax_cross_entropy_with_logits.cc
index 6a8d7c55a9..762349084f 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/softmax_cross_entropy_with_logits.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/softmax_cross_entropy_with_logits.cc
@@ -41,8 +41,7 @@ void SoftmaxCrossEntropyWithLogitsCPUKernel::ForwardPostExecute(const float *lab
         float logit =
           -logf(logits[i * param_->number_of_classes_ + j] <= 0.0 ? eps : logits[i * param_->number_of_classes_ + j]);
         grads[i * param_->number_of_classes_ + j] =
-          (logits[i * param_->number_of_classes_ + j] - labels[i * param_->number_of_classes_ + j]) /
-          param_->batch_size_;
+          (logits[i * param_->number_of_classes_ + j] - labels[i * param_->number_of_classes_ + j]);
         total_loss += labels[i * param_->number_of_classes_ + j] * logit;
       }
     }
@@ -63,7 +62,7 @@ int SoftmaxCrossEntropyWithLogitsCPUKernel::Execute(int task_id) {
   auto labels = reinterpret_cast<float *>(in_tensors_.at(1)->MutableData());
   float *out = reinterpret_cast<float *>(out_tensors_.at(0)->MutableData());
   float *grads = NULL;
-  if (is_train() && out_tensors_.size() > 1) {
+  if (IsTrain() && out_tensors_.size() > 1) {
     grads = reinterpret_cast<float *>(out_tensors_.at(1)->MutableData());
   }
   size_t data_size = in_tensors_.at(0)->ElementsNum();
diff --git a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/sparse_softmax_cross_entropy_with_logits.cc b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/sparse_softmax_cross_entropy_with_logits.cc
index 01135de869..1d3fb6feb2 100644
--- a/mindspore/lite/src/runtime/kernel/arm/fp32_grad/sparse_softmax_cross_entropy_with_logits.cc
+++ b/mindspore/lite/src/runtime/kernel/arm/fp32_grad/sparse_softmax_cross_entropy_with_logits.cc
@@ -86,7 +86,7 @@ int SparseSoftmaxCrossEntropyWithLogitsCPUKernel::Execute(int task_id) {
   auto labels = reinterpret_cast<int *>(in_tensors_.at(1)->data_c());
   float *out = reinterpret_cast<float *>(out_tensors_.at(0)->data_c());
   float *grads = NULL;
-  if (is_train() && out_tensors_.size() > 1) {
+  if (IsTrain() && out_tensors_.size() > 1) {
     grads = reinterpret_cast<float *>(out_tensors_.at(1)->MutableData());
   }
   size_t data_size = in_tensors_.at(0)->ElementsNum();
@@ -99,7 +99,7 @@ int SparseSoftmaxCrossEntropyWithLogitsCPUKernel::Execute(int task_id) {
   std::fill(losses_, losses_ + data_size, 0.f);
   std::fill(sum_data_, sum_data_ + sm_params_.input_shape_[0], 0.f);
   Softmax(ins, losses_, sum_data_, &sm_params_);
-  if (is_train()) {
+  if (IsTrain()) {
     GradPostExecute(labels, losses_, grads, out);
   } else {
     ForwardPostExecute(labels, losses_, out);
diff --git a/mindspore/lite/src/train/train_populate_parameter.cc b/mindspore/lite/src/train/train_populate_parameter.cc
index c490b34a25..3d5f4c028e 100644
--- a/mindspore/lite/src/train/train_populate_parameter.cc
+++ b/mindspore/lite/src/train/train_populate_parameter.cc
@@ -36,6 +36,10 @@
 #include "src/ops/bn_grad.h"
 #include "nnacl/fp32_grad/batch_norm.h"
 #include "src/ops/adam.h"
+#include "nnacl/fp32_grad/dropout_parameter.h"
+#include "src/ops/dropout.h"
+#include "src/ops/dropout_grad.h"
+#include "src/ops/arithmetic.h"
 #include "src/ops/oneslike.h"
 #include "src/ops/binary_cross_entropy.h"
 #include "src/ops/binary_cross_entropy_grad.h"
@@ -399,10 +403,66 @@ OpParameter *PopulateBNGradParameter(const mindspore::lite::PrimitiveC *primitiv
   bnGrad_param->op_parameter_.type_ = primitive->Type();
   auto bngrad = reinterpret_cast<mindspore::lite::BNGrad *>(const_cast<mindspore::lite::PrimitiveC *>(primitive));
   bnGrad_param->epsilon_ = bngrad->GetEps();
-  bnGrad_param->momentum_ = 0.1;
+  bnGrad_param->momentum_ = bngrad->GetMomentum();
   return reinterpret_cast<OpParameter *>(bnGrad_param);
 }
 
+OpParameter *PopulateDropoutParameter(const mindspore::lite::PrimitiveC *primitive) {
+  DropoutParameter *dropout_parameter = reinterpret_cast<DropoutParameter *>(malloc(sizeof(DropoutParameter)));
+  if (dropout_parameter == nullptr) {
+    MS_LOG(ERROR) << "malloc Dropout Parameter failed.";
+    return nullptr;
+  }
+  memset(dropout_parameter, 0, sizeof(DropoutParameter));
+  dropout_parameter->op_parameter_.type_ = primitive->Type();
+  auto param = reinterpret_cast<mindspore::lite::Dropout *>(const_cast<mindspore::lite::PrimitiveC *>(primitive));
+  dropout_parameter->ratio_ = param->GetRatio();
+  if (dropout_parameter->ratio_ < 0.f || dropout_parameter->ratio_ > 1.f) {
+    MS_LOG(ERROR) << "Dropout ratio must be between 0 to 1, got " << dropout_parameter->ratio_;
+    free(dropout_parameter);
+    return nullptr;
+  }
+  return reinterpret_cast<OpParameter *>(dropout_parameter);
+}
+
+OpParameter *PopulateDropoutGradParameter(const mindspore::lite::PrimitiveC *primitive) {
+  DropoutParameter *dropoutGrad_parameter = reinterpret_cast<DropoutParameter *>(malloc(sizeof(DropoutParameter)));
+  if (dropoutGrad_parameter == nullptr) {
+    MS_LOG(ERROR) << "malloc Dropout Grad Parameter failed.";
+    return nullptr;
+  }
+  memset(dropoutGrad_parameter, 0, sizeof(DropoutParameter));
+  dropoutGrad_parameter->op_parameter_.type_ = primitive->Type();
+  auto param = reinterpret_cast<mindspore::lite::DropoutGrad *>(const_cast<mindspore::lite::PrimitiveC *>(primitive));
+  dropoutGrad_parameter->ratio_ = param->GetRatio();
+  if (dropoutGrad_parameter->ratio_ < 0.f || dropoutGrad_parameter->ratio_ > 1.f) {
+    MS_LOG(ERROR) << "Dropout Grad ratio must be between 0 to 1, got " << dropoutGrad_parameter->ratio_;
+    free(dropoutGrad_parameter);
+    return nullptr;
+  }
+  return reinterpret_cast<OpParameter *>(dropoutGrad_parameter);
+}
+
+OpParameter *PopulateArithmeticGradParameter(const mindspore::lite::PrimitiveC *primitive) {
+  ArithmeticParameter *arithmetic_param = reinterpret_cast<ArithmeticParameter *>(malloc(sizeof(ArithmeticParameter)));
+  if (arithmetic_param == nullptr) {
+    MS_LOG(ERROR) << "malloc ArithmeticParameter failed.";
+    return nullptr;
+  }
+  memset(arithmetic_param, 0, sizeof(ArithmeticParameter));
+  arithmetic_param->op_parameter_.type_ = primitive->Type();
+  arithmetic_param->broadcasting_ = ((lite::Arithmetic *)primitive)->Broadcasting();
+  arithmetic_param->ndim_ = ((lite::Arithmetic *)primitive)->NDims();
+
+  auto tmp_shape = ((lite::Arithmetic *)primitive)->InShape0();
+  memcpy(arithmetic_param->in_shape0_, static_cast<void *>(tmp_shape.data()), tmp_shape.size() * sizeof(int));
+  tmp_shape = ((lite::Arithmetic *)primitive)->InShape1();
+  memcpy(arithmetic_param->in_shape1_, static_cast<void *>(tmp_shape.data()), tmp_shape.size() * sizeof(int));
+  tmp_shape = ((lite::Arithmetic *)primitive)->OutputShape();
+  memcpy(arithmetic_param->out_shape_, static_cast<void *>(tmp_shape.data()), tmp_shape.size() * sizeof(int));
+  return reinterpret_cast<OpParameter *>(arithmetic_param);
+}
+
 void PopulateTrainParameters() {
   lite::Registry ApplyMomentumParameterRegistry(schema::PrimitiveType_ApplyMomentum, PopulateApplyMomentumParameter);
   lite::Registry BiasGradParameterRegistry(schema::PrimitiveType_BiasGrad, PopulateBiasGradParameter);
@@ -430,6 +490,10 @@ void PopulateTrainParameters() {
   lite::Registry OnesLikeParameterRegistry(schema::PrimitiveType_OnesLike, DefaultPopulateParameter);
   lite::Registry UnsortedSegmentSumParameterRegistry(schema::PrimitiveType_UnsortedSegmentSum,
                                                      DefaultPopulateParameter);
+  lite::Registry DropoutParameterRegistry(schema::PrimitiveType_Dropout, PopulateDropoutParameter);
+  lite::Registry DropGradParameterRegistry(schema::PrimitiveType_DropoutGrad, PopulateDropoutGradParameter);
+  lite::Registry MaximumGradParameterRegistry(schema::PrimitiveType_MaximumGrad, PopulateArithmeticGradParameter);
+  lite::Registry MinimumGradParameterRegistry(schema::PrimitiveType_MinimumGrad, PopulateArithmeticGradParameter);
 }
 
 }  // namespace mindspore::kernel
diff --git a/mindspore/lite/src/train/train_session.cc b/mindspore/lite/src/train/train_session.cc
index 34b427139f..1257089069 100644
--- a/mindspore/lite/src/train/train_session.cc
+++ b/mindspore/lite/src/train/train_session.cc
@@ -15,9 +15,12 @@
  */
 
 #include "src/train/train_session.h"
+#include <sys/stat.h>
 #include <algorithm>
 #include <utility>
 #include <vector>
+#include <iostream>
+#include <fstream>
 #include "include/errorcode.h"
 #include "include/train_model.h"
 #include "src/common/utils.h"
@@ -98,6 +101,21 @@ int TrainSession::CompileTrainGraph(mindspore::lite::TrainModel *model) {
   for (auto inTensor : inputs_) inTensor->MutableData();
   RestoreOps(restore);
   AllocWorkSpace();
+  MarkOptimizedKernels();
+  CompileTrainKernels();
+  if (train_mode_) {
+    auto ret1 = Train();
+    if (ret1 != RET_OK) {
+      MS_LOG(ERROR) << "faild to initialize network in train mode";
+      return RET_ERROR;
+    }
+  } else {
+    auto ret1 = Eval();
+    if (ret1 != RET_OK) {
+      MS_LOG(ERROR) << "faild to initialize network in eval mode";
+      return RET_ERROR;
+    }
+  }
   return ret;
 }
 
@@ -110,34 +128,67 @@ void *TrainSession::ExportToBuf(char *buf, size_t *len) const { return model_->E
 
 int TrainSession::RunGraph(const KernelCallBack &before, const KernelCallBack &after) {
   this->outputs_.clear();
-  for (auto ms_tensors : output_node_map_)
-    for (auto ms_tensor : ms_tensors.second) this->outputs_.push_back((static_cast<lite::Tensor *>(ms_tensor)));
-  if (train_mode_) return lite::LiteSession::RunGraph(before, after);
+
+  // build out tensor
+  for (auto ms_tensors : output_node_map_) {
+    for (auto ms_tensor : ms_tensors.second) {
+      this->outputs_.push_back((static_cast<lite::Tensor *>(ms_tensor)));
+    }
+  }
 
   if (this->context_ == nullptr) {
     MS_LOG(ERROR) << "context is null";
     return lite::RET_NULL_PTR;
   }
+  auto run_kernel = (train_mode_) ? train_kernels_ : inference_kernels_;
   lite::Executor executor;
   if (before == nullptr && after == nullptr) {
-    return executor.Run(this->inputs_, this->outputs_, inference_kernels_, this->context_->allocator.get());
+    return executor.Run(this->inputs_, this->outputs_, run_kernel, this->context_->allocator.get());
   } else {
-    return executor.Run(this->inputs_, this->outputs_, inference_kernels_, this->context_->allocator.get(), before,
-                        after);
+    return executor.Run(this->inputs_, this->outputs_, run_kernel, this->context_->allocator.get(), before, after);
   }
 }
 
-void TrainSession::Train() {
+int TrainSession::SaveToFile(const std::string &filename) const {
+  size_t fb_size = 0;
+  auto *buf = reinterpret_cast<char *>(ExportToBuf(nullptr, &fb_size));
+  if (buf == NULL) {
+    MS_LOG(ERROR) << "Could not Export Trained model";
+    return lite::RET_NULL_PTR;
+  }
+  std::ofstream ofs(filename);
+  if ((true != ofs.good()) || (true != ofs.is_open())) {
+    MS_LOG(ERROR) << "Could not open file \"" << filename << "\" for writing";
+    free(buf);
+    return RET_ERROR;
+  }
+
+  ofs.seekp(0, std::ios::beg);
+  ofs.write(buf, fb_size);
+  ofs.close();
+  free(buf);
+  return chmod(filename.c_str(), S_IRUSR);
+}
+
+int TrainSession::Train() {
   for (auto ori_kernel : kernels_) {
     MS_ASSERT(nullptr != ori_kernel);
     if (ori_kernel->subgraph_type() == kernel::kNotSubGraph) {
-      ori_kernel->train();
+      auto ret = ori_kernel->Train();
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << ori_kernel->name() << " failed to set train mode";
+        return RET_ERROR;
+      }
     } else {
       auto sub_graph = reinterpret_cast<kernel::SubGraphKernel *>(ori_kernel);
       MS_ASSERT(nullptr != sub_graph);
       for (auto kernel : sub_graph->nodes()) {
         MS_ASSERT(nullptr != kernel);
-        kernel->train();
+        auto ret = kernel->Train();
+        if (ret != RET_OK) {
+          MS_LOG(ERROR) << kernel->name() << " failed to set train mode";
+          return RET_ERROR;
+        }
       }
     }
   }
@@ -157,6 +208,7 @@ void TrainSession::Train() {
       }
     }
   }
+  return RET_OK;
 }
 
 void TrainSession::UpdateOutputMapByLossKernel(const kernel::LiteKernel *kernel) {
@@ -190,17 +242,25 @@ void TrainSession::UpdateOutputMapByInKernel(const kernel::LiteKernel *kernel) {
   }
 }
 
-void TrainSession::Eval() {
+int TrainSession::Eval() {
   for (auto ori_kernel : kernels_) {
     MS_ASSERT(nullptr != ori_kernel);
     if (ori_kernel->subgraph_type() == kernel::kNotSubGraph) {
-      ori_kernel->eval();
+      auto ret = ori_kernel->Eval();
+      if (ret != RET_OK) {
+        MS_LOG(ERROR) << ori_kernel->name() << " failed to set eval mode";
+        return RET_ERROR;
+      }
     } else {
       auto sub_graph = reinterpret_cast<kernel::SubGraphKernel *>(ori_kernel);
       MS_ASSERT(nullptr != sub_graph);
       for (auto kernel : sub_graph->nodes()) {
         MS_ASSERT(nullptr != kernel);
-        kernel->eval();
+        auto ret = kernel->Eval();
+        if (ret != RET_OK) {
+          MS_LOG(ERROR) << kernel->name() << " failed to set eval mode";
+          return RET_ERROR;
+        }
       }
     }
   }
@@ -221,6 +281,7 @@ void TrainSession::Eval() {
   if (inference_kernels_.size() == 0) {
     BuildInferenceKernelsMap();
   }
+  return RET_OK;
 }
 
 void TrainSession::BuildInferenceKernelsRecursive(kernel::LiteKernel *kernel, std::vector<kernel::LiteKernel *> *v) {
@@ -234,24 +295,25 @@ void TrainSession::BuildInferenceKernelsRecursive(kernel::LiteKernel *kernel, st
 
 void TrainSession::BuildInferenceKernelsMap() {
   std::vector<kernel::LiteKernel *> req_kernels;
-  for (auto ori_kernel : kernels_) {
-    if (ori_kernel->subgraph_type() == kernel::kNotSubGraph) {
-      if (IsLossKernel(ori_kernel)) {  // For each loss in the system add backward tree
-        for (auto in_node : ori_kernel->in_kernels()) {
+  for (auto kernel : this->kernels_) {
+    if (kernel->subgraph_type() == kernel::kNotSubGraph) {
+      if (IsLossKernel(kernel)) {  // For each loss in the system add backward tree
+        for (auto in_node : kernel->in_kernels()) {
           BuildInferenceKernelsRecursive(in_node, &req_kernels);
         }
       }
     } else {
-      auto sub_graph = reinterpret_cast<kernel::SubGraphKernel *>(ori_kernel);
-      for (auto kernel : sub_graph->nodes()) {
-        if (IsLossKernel(kernel)) {  // For each loss in the system add backward tree
-          for (auto in_node : kernel->in_kernels()) {
+      auto sub_graph = reinterpret_cast<kernel::SubGraphKernel *>(kernel);
+      for (auto sb_kernel : sub_graph->nodes()) {
+        if (IsLossKernel(sb_kernel)) {  // For each loss in the system add backward tree
+          for (auto in_node : sb_kernel->in_kernels()) {
             BuildInferenceKernelsRecursive(in_node, &req_kernels);
           }
         }
       }
     }
   }
+
   inference_kernels_.clear();
   for (auto ori_kernel : kernels_) {
     if (ori_kernel->subgraph_type() == kernel::kNotSubGraph) {
@@ -272,10 +334,71 @@ void TrainSession::BuildInferenceKernelsMap() {
   }
 }
 
-bool TrainSession::IsLossKernel(const kernel::LiteKernel *kernel) {
+void TrainSession::CompileTrainKernels() {
+  train_kernels_.clear();
+  for (auto ori_kernel : kernels_) {
+    if (ori_kernel->subgraph_type() == kernel::kNotSubGraph) {
+      train_kernels_.push_back(ori_kernel);
+    } else {
+      auto sub_graph = reinterpret_cast<kernel::SubGraphKernel *>(ori_kernel);
+      for (auto kernel : sub_graph->nodes()) {
+        train_kernels_.push_back(kernel);
+      }
+    }
+  }
+}
+
+void TrainSession::MarkOptimizedKernels() {
+  std::vector<lite::Tensor *> ot;
+  for (auto kernel : this->kernels_) {
+    if (kernel->subgraph_type() == kernel::kNotSubGraph) {
+      if (IsOptimizer(kernel)) {
+        std::copy(kernel->in_tensors().begin(), kernel->in_tensors().end(), std::back_inserter(ot));
+      }
+    } else {
+      auto sub_graph = reinterpret_cast<kernel::SubGraphKernel *>(kernel);
+      for (auto sb_kernel : sub_graph->nodes()) {
+        if (IsOptimizer(sb_kernel)) {
+          std::copy(sb_kernel->in_tensors().begin(), sb_kernel->in_tensors().end(), std::back_inserter(ot));
+        }
+      }
+    }
+  }
+  for (auto kernel : this->kernels_) {
+    if (kernel->subgraph_type() == kernel::kNotSubGraph) {
+      if (!IsOptimizer(kernel)) {
+        for (auto it : kernel->in_tensors()) {
+          if (std::find(ot.begin(), ot.end(), it) != ot.end()) {
+            kernel->SetTrainable(true);
+            break;
+          }
+        }
+      }
+    } else {
+      auto sub_graph = reinterpret_cast<kernel::SubGraphKernel *>(kernel);
+      for (auto sb_kernel : sub_graph->nodes()) {
+        if (!IsOptimizer(sb_kernel)) {
+          for (auto it : sb_kernel->in_tensors()) {
+            if (std::find(ot.begin(), ot.end(), it) != ot.end()) {
+              sb_kernel->SetTrainable(true);
+              break;
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
+bool TrainSession::IsLossKernel(const kernel::LiteKernel *kernel) const {
   return (kernel->Type() == schema::PrimitiveType_SoftmaxCrossEntropy);
 }
 
+bool TrainSession::IsOptimizer(kernel::LiteKernel *kernel) const {
+  return ((kernel->Type() == schema::PrimitiveType_Adam) || (kernel->Type() == schema::PrimitiveType_Sgd) ||
+          (kernel->Type() == schema::PrimitiveType_ApplyMomentum));
+}
+
 }  // namespace lite
 
 session::TrainSession *session::TrainSession::CreateSession(lite::Context *context) {
diff --git a/mindspore/lite/src/train/train_session.h b/mindspore/lite/src/train/train_session.h
index 6e676a3957..e11a8d92d3 100644
--- a/mindspore/lite/src/train/train_session.h
+++ b/mindspore/lite/src/train/train_session.h
@@ -55,9 +55,10 @@ class TrainSession : virtual public session::TrainSession, virtual public lite::
   int CompileTrainGraph(lite::TrainModel *model) override;
 
   void *ExportToBuf(char *buf, size_t *len) const override;
+  int SaveToFile(const std::string &filename) const override;
 
-  void Train() override;
-  void Eval() override;
+  int Train() override;
+  int Eval() override;
 
   void BindThread(bool if_bind) override { return lite::LiteSession::BindThread(if_bind); }
   std::vector<tensor::MSTensor *> GetInputs() const override { return lite::LiteSession::GetInputs(); }
@@ -84,16 +85,19 @@ class TrainSession : virtual public session::TrainSession, virtual public lite::
 
  protected:
   void AllocWorkSpace();
-  bool IsLossKernel(const kernel::LiteKernel *kernel);
+  bool IsLossKernel(const kernel::LiteKernel *kernel) const;
+  bool IsOptimizer(kernel::LiteKernel *kernel) const;
+  virtual void MarkOptimizedKernels();
   virtual std::vector<CreatorOp> ReplaceOps();
   virtual void RestoreOps(const std::vector<CreatorOp> &restore);
   virtual void BuildInferenceKernelsMap();
   virtual void BuildInferenceKernelsRecursive(kernel::LiteKernel *ker, std::vector<kernel::LiteKernel *> *req_kernels);
-
+  virtual void CompileTrainKernels();
   TrainModel *model_ = nullptr;
   std::unordered_map<std::string, std::vector<mindspore::tensor::MSTensor *>> orig_output_map_;
   std::unordered_map<std::string, mindspore::tensor::MSTensor *> orig_output_tensor_map_;
   std::vector<kernel::LiteKernel *> inference_kernels_;
+  std::vector<kernel::LiteKernel *> train_kernels_;
 };
 }  // namespace lite
 }  // namespace mindspore
diff --git a/mindspore/lite/test/models_ms_train.cfg b/mindspore/lite/test/models_ms_train.cfg
new file mode 100644
index 0000000000..60192579f4
--- /dev/null
+++ b/mindspore/lite/test/models_ms_train.cfg
@@ -0,0 +1,8 @@
+mini_alexnet
+mobilenetv1
+mobilenetv2
+mobilenetv3
+lenet
+effnet
+effnet_tune
+resnet
diff --git a/mindspore/lite/test/run_net_train.sh b/mindspore/lite/test/run_net_train.sh
new file mode 100755
index 0000000000..0a616457d2
--- /dev/null
+++ b/mindspore/lite/test/run_net_train.sh
@@ -0,0 +1,394 @@
+#!/bin/bash
+
+# Run Export on x86 platform and create output test files:
+function Run_Export(){
+    cd $models_path || exit 1
+    if [[ -z "${CLOUD_MODEL_ZOO}" ]]; then
+        echo "CLOUD_MODEL_ZOO is not defined - exiting export models"
+        exit 1
+    fi    
+    # Export mindspore train models:
+    while read line; do
+        model_name=${line}
+        if [[ $model_name == \#* ]]; then
+          continue
+        fi
+        echo ${model_name}'_train_export.py' >> "${export_log_file}"
+        echo 'exporting' ${model_name}
+        echo 'docker run --user $(id -u):$(id -g) --env CLOUD_MODEL_ZOO=${CLOUD_MODEL_ZOO} -w $PWD --runtime=nvidia -v /home/$USER:/home/$USER -v /opt/share:/opt/share --privileged=true mindspore_dev:5 python '${models_path}'/'${model_name}'_train_export.py' >>  "${export_log_file}"
+        docker run --user $(id -u):$(id -g) --env CLOUD_MODEL_ZOO=${CLOUD_MODEL_ZOO} -w $PWD --runtime=nvidia -v /home/$USER:/home/$USER -v /opt/share:/opt/share --privileged=true mindspore_dev:5 python ${models_path}'/'${model_name}_train_export.py
+        if [ $? = 0 ]; then
+            export_result='export mindspore '${model_name}'_train_export pass';echo ${export_result} >> ${export_result_file}
+        else
+            export_result='export mindspore '${model_name}'_train_export failed';echo ${export_result} >> ${export_result_file}
+        fi
+    done < ${models_mindspore_train_config}
+}
+
+# Run converter on x86 platform:
+function Run_Converter() {
+    # Unzip x86 runtime and convertor
+    cd ${x86_path} || exit 1
+    tar -zxf mindspore-lite-${version}-runtime-x86-${process_unit_x86}-train.tar.gz || exit 1
+
+    tar -zxf mindspore-lite-${version}-converter-ubuntu-train.tar.gz || exit 1
+    cd ${x86_path}/mindspore-lite-${version}-converter-ubuntu-train || exit 1
+    cp converter/converter_lite ./ || exit 1
+    
+
+    # Convert the models
+    cd ${x86_path}/mindspore-lite-${version}-converter-ubuntu-train || exit 1
+
+    rm -rf ${ms_models_path}
+    mkdir -p ${ms_models_path}
+
+    # Convert mindspore train models:
+    while read line; do
+        model_name=${line}
+        if [[ $model_name == \#* ]]; then
+          continue
+        fi
+        echo ${model_name}'_train' >> "${run_converter_log_file}"
+        echo './converter_lite  --fmk=MINDIR --modelFile='${models_path}'/'${model_name}'_train.mindir --outputFile='${ms_models_path}'/'${model_name}'_train  --trainModel=true' >> "${run_converter_log_file}"
+        LD_LIBRARY_PATH=./lib/:./third_party/protobuf/lib:./third_party/flatbuffers/lib:./third_party/glog/lib \
+         ./converter_lite  --fmk=MINDIR --modelFile=${models_path}/${model_name}_train.mindir \
+         --outputFile=${ms_models_path}/${model_name}'_train' \
+         --trainModel=true
+        if [ $? = 0 ]; then
+            converter_result='converter mindspore '${model_name}'_train pass';echo ${converter_result} >> ${run_converter_result_file}
+        else
+            converter_result='converter mindspore '${model_name}'_train failed';echo ${converter_result} >> ${run_converter_result_file}
+        fi
+    done < ${models_mindspore_train_config}
+}
+
+# Run on x86 platform:
+function Run_x86() {
+    # Run mindspore converted train models:
+    while read line; do
+        model_name=${line}
+        if [[ $model_name == \#* ]]; then
+          continue
+        fi
+        
+        echo ${model_name}'_train' >> "${run_x86_log_file}"
+        echo 'cd  '${x86_path}'/mindspore-lite-'${version}'-runtime-x86-'${process_unit_x86}-train >> "${run_x86_log_file}"
+        cd ${x86_path}/mindspore-lite-${version}-runtime-x86-${process_unit_x86}-train || return 1
+        echo 'LD_LIBRARY_PATH='${LD_LIBRARY_PATH}':./lib:./third_party/libjpeg-turbo/lib:./third_party/opencv/lib;./net_train/net_train --modelFile='${ms_models_path}'/'${model_name}'_train.ms --inDataFile='${input_path}'/'${model_name}'_input1.bin,'${train_io_path}'/'${model_name}'_input2.bin --expectedDataFile='${train_io_path}'/'${model_name}'_outputs.bin --exportFile='${ms_models_path}'/'${model_name}'_train_exported.ms' >> "${run_x86_log_file}"
+        echo '-------------------------------------------------------------------------------' >> "${run_x86_log_file}"
+        LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:./lib:./third_party/libjpeg-turbo/lib:./third_party/opencv/lib \
+        ${run_valgrind}./net_train/net_train \
+        --modelFile=${ms_models_path}/${model_name}_train.ms \
+        --inDataFile=${train_io_path}/${model_name}_input1.bin,${train_io_path}/${model_name}_input2.bin \
+        --expectedDataFile=${train_io_path}/${model_name}_outputs.bin \
+        --exportFile=${ms_models_path}/${model_name}_train_exported.ms >> "${run_x86_log_file}"
+        if [ $? = 0 ]; then
+            run_result='x86: '${model_name}'_train pass'; echo ${run_result} >> ${run_net_train_result_file}
+        else
+            run_result='x86: '${model_name}'_train failed'; echo ${run_result} >> ${run_net_train_result_file}
+        fi
+    done < ${models_mindspore_train_config}
+}
+
+# Run on arm platform: 
+# Gets a parameter - arm64/arm32
+function Run_arm() {
+    if [ "$1" == arm64 ]; then
+        arm_path=${arm64_path}
+        process_unit=${process_unit_arm64}
+        version_arm=${version_arm64}
+        run_arm_log_file=${run_arm64_log_file} 
+        adb_cmd_run_file=${adb_cmd_arm64_run_file} 
+        adb_push_log_file=${adb_push_arm64_log_file} 
+        adb_cmd_file=${adb_cmd_arm64_file}
+    elif [ "$1" == arm32 ]; then
+        arm_path=${arm32_path}
+        process_unit=${process_unit_arm32}
+        version_arm=${version_arm32}
+        run_arm_log_file=${run_arm32_log_file} 
+        adb_cmd_run_file=${adb_cmd_arm32_run_file} 
+        adb_push_log_file=${adb_push_arm32_log_file} 
+        adb_cmd_file=${adb_cmd_arm32_file}
+    else
+        echo 'type ' $1 'is not supported'
+        exit 1
+    fi   
+    arm_type=$1
+
+    # Unzip
+    cd ${arm_path} || exit 1
+    tar -zxf mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train.tar.gz || exit 1
+
+    # If build with minddata, copy the minddata related libs
+    cd ${net_train_test_path} || exit 1
+    if [ -f ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/lib/libminddata-lite.so ]; then
+        cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/third_party/libjpeg-turbo/lib/libjpeg.so ${net_train_test_path}/libjpeg.so || exit 1
+        cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/third_party/libjpeg-turbo/lib/libturbojpeg.so ${net_train_test_path}/libturbojpeg.so || exit 1
+        cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/third_party/opencv/lib/libopencv_core.so ${net_train_test_path}/libopencv_core.so || exit 1
+        cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/third_party/opencv/lib/libopencv_imgcodecs.so ${net_train_test_path}/libopencv_imgcodecs.so || exit 1
+        cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/third_party/opencv/lib/libopencv_imgproc.so ${net_train_test_path}/libopencv_imgproc.so || exit 1
+        cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/lib/libminddata-lite.so ${net_train_test_path}/libminddata-lite.so || exit 1
+    fi
+
+    cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/lib/libmindspore-lite.so ${net_train_test_path}/libmindspore-lite.so || exit 1
+    if [ "$1" == arm64 ]; then    
+        cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/lib/libmindspore-lite-fp16.so ${net_train_test_path}/libmindspore-lite-fp16.so || exit 1
+        cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/lib/libmindspore-lite-optimize.so ${net_train_test_path}/libmindspore-lite-optimize.so || exit 1
+    fi    
+    cp -a ${arm_path}/mindspore-lite-${version_arm}-runtime-${arm_type}-${process_unit}-train/net_train/net_train ${net_train_test_path}/net_train || exit 1
+
+    # adb push all needed files to the phone
+    adb -s ${device_id} push ${net_train_test_path} /data/local/tmp/ > ${adb_push_log_file}
+
+    # run adb ,run session ,check the result:
+    echo 'cd  /data/local/tmp/net_train_test' > ${adb_cmd_file}
+    echo 'chmod 777 net_train' >> ${adb_cmd_file}
+
+    adb -s ${device_id} shell < ${adb_cmd_file}
+    
+    # Run mindir converted train models:
+    while read line; do
+        model_name=${line}
+        if [[ $model_name == \#* ]]; then
+          continue
+        fi
+
+        # run net_train test without clib data
+        echo ${model_name}'_train' >> "${run_arm_log_file}"
+        adb -s ${device_id} push ${train_io_path}/${model_name}_input*.bin ${train_io_path}/${model_name}_outputs.bin  /data/local/tmp/net_train_test >> ${adb_push_log_file}
+        echo 'cd /data/local/tmp/net_train_test' > ${adb_cmd_run_file}
+        if [ "$1" == arm64 ]; then
+            echo 'export LD_LIBRARY_PATH=/data/local/tmp/net_train_test;./net_train --modelFile='${model_name}'_train.ms --inDataFile=/data/local/tmp/net_train_test/'${model_name}'_input1.bin,/data/local/tmp/net_train_test/'${model_name}'_input2.bin --expectedDataFile=/data/local/tmp/net_train_test/'${model_name}'_outputs.bin --exportFile='${model_name}'_train_exported.ms' >> "${run_arm_log_file}"
+            echo 'export LD_LIBRARY_PATH=/data/local/tmp/net_train_test;./net_train --modelFile='${model_name}'_train.ms --inDataFile=/data/local/tmp/net_train_test/'${model_name}'_input1.bin,/data/local/tmp/net_train_test/'${model_name}'_input2.bin --expectedDataFile=/data/local/tmp/net_train_test/'${model_name}'_outputs.bin --exportFile='${model_name}'_train_exported.ms' >> "${adb_cmd_run_file}"
+        elif [ "$1" == arm32 ]; then 
+            echo 'export LD_LIBRARY_PATH=/data/local/tmp/:/data/local/tmp/net_train_test;./net_train --modelFile='${model_name}'_train.ms --inDataFile=/data/local/tmp/net_train_test/'${model_name}'_input1.bin,/data/local/tmp/net_train_test/'${model_name}'_input2.bin --expectedDataFile=/data/local/tmp/net_train_test/'${model_name}'_outputs.bin --exportFile='${model_name}'_train_exported.ms' >> "${run_arm_log_file}"
+            echo 'export LD_LIBRARY_PATH=/data/local/tmp/:/data/local/tmp/net_train_test;./net_train --modelFile='${model_name}'_train.ms --inDataFile=/data/local/tmp/net_train_test/'${model_name}'_input1.bin,/data/local/tmp/net_train_test/'${model_name}'_input2.bin --expectedDataFile=/data/local/tmp/net_train_test/'${model_name}'_outputs.bin --exportFile='${model_name}'_train_exported.ms' >> "${adb_cmd_run_file}"
+        fi    
+
+        adb -s ${device_id} shell < ${adb_cmd_run_file} >> ${run_arm_log_file}
+        # TODO: change to arm_type
+        if [ $? = 0 ]; then
+            run_result=$1': '${model_name}'_train pass'; echo ${run_result} >> ${run_net_train_result_file}
+        else
+            run_result=$1': '${model_name}'_train failed'; echo ${run_result} >> ${run_net_train_result_file}
+        fi
+    done < ${models_mindspore_train_config}
+}
+
+# Print start msg before run testcase
+function MS_PRINT_TESTCASE_START_MSG() {
+    echo ""
+    echo -e "-----------------------------------------------------------------------------------------------------------------------------------"
+    echo -e "env                  Testcase                                                                                             Result   "
+    echo -e "---                  --------                                                                                             ------   "
+}
+
+# Print start msg after run testcase
+function MS_PRINT_TESTCASE_END_MSG() {
+    echo -e "-----------------------------------------------------------------------------------------------------------------------------------"
+}
+
+function Print_Result() {
+    MS_PRINT_TESTCASE_END_MSG
+    while read line; do
+        arr=("${line}")
+        printf "%-15s %-20s %-90s %-7s\n" ${arr[0]} ${arr[1]} ${arr[2]} ${arr[3]}
+    done < $1
+    MS_PRINT_TESTCASE_END_MSG
+}
+
+basepath=$(pwd)
+echo ${basepath}
+
+# Example:run_net_train.sh -r /home/emir/Work/TestingEnv/release -m /home/emir/Work/TestingEnv/train_models -i /home/emir/Work/TestingEnv/train_io -d "8KE5T19620002408"
+# For running on arm64, use -t to set platform tools path (for using adb commands)
+while getopts "r:m:d:i:e:v" opt; do
+    case ${opt} in
+        r)
+	    release_path=${OPTARG}
+            echo "release_path is ${OPTARG}"
+            ;;
+        m)
+	    models_path=${OPTARG}
+            echo "models_path is ${OPTARG}"
+            ;;
+        i)
+	    train_io_path=${OPTARG}
+            echo "train_io_path is ${OPTARG}"
+            ;;
+        d)
+	    device_id=${OPTARG}
+            echo "device_id is ${OPTARG}"
+            ;;
+        e)
+	    enable_export=${OPTARG}
+            echo "enable_export = ${OPTARG}"
+            ;;          
+        v)
+	    run_valgrind="valgrind "
+            echo "Run x86 with valgrind"
+            ;;            
+        ?)
+            echo "unknown para"
+            exit 1;;
+    esac
+done
+
+arm64_path=${release_path}/android_aarch64
+file=$(ls ${arm64_path}/*runtime-arm64*train.tar.gz)
+file_name="${file##*/}"
+IFS="-" read -r -a file_name_array <<< "$file_name"
+version_arm64=${file_name_array[2]}
+process_unit_arm64=${file_name_array[5]}
+
+arm32_path=${release_path}/android_aarch32
+file=$(ls ${arm32_path}/*runtime-arm32*train.tar.gz)
+file_name="${file##*/}"
+IFS="-" read -r -a file_name_array <<< "$file_name"
+version_arm32=${file_name_array[2]}
+process_unit_arm32=${file_name_array[5]}
+
+x86_path=${release_path}/ubuntu_x86
+file=$(ls ${x86_path}/*runtime-x86*train.tar.gz)
+file_name="${file##*/}"
+IFS="-" read -r -a file_name_array <<< "$file_name"
+version=${file_name_array[2]}
+process_unit_x86=${file_name_array[5]}
+
+# Set models config filepath
+models_mindspore_train_config=${basepath}/models_ms_train.cfg
+
+ms_models_path=${models_path}/ms_models
+
+logs_path=${models_path}/logs
+rm -rf ${logs_path}
+mkdir -p ${logs_path}
+
+# Export model if enabled
+if [[ $enable_export == 1 ]]; then
+    echo "Start Exporting models ..."
+    # Write export result to temp file
+    export_log_file=${logs_path}/export_log.txt
+    echo ' ' > ${export_log_file}
+
+    export_result_file=${logs_path}/export_result.txt
+    echo ' ' > ${export_result_file}
+    # Run export
+    Run_Export
+    Print_Result ${export_result_file}
+
+fi    
+
+# Write converter result to temp file
+run_converter_log_file=${logs_path}/run_converter_log.txt
+echo ' ' > ${run_converter_log_file}
+
+run_converter_result_file=${logs_path}/run_converter_result.txt
+echo ' ' > ${run_converter_result_file}
+
+START=$(date +%s.%N)
+
+# Run converter
+echo "start run converter ..."
+Run_Converter
+Run_converter_PID=$!
+sleep 1
+
+wait ${Run_converter_PID}
+Run_converter_status=$?
+
+# Check converter result and return value
+if [[ ${Run_converter_status} = 0 ]];then
+    echo "Run converter success"
+    Print_Result ${run_converter_result_file}
+else
+    echo "Run converter failed"
+    cat ${run_converter_log_file}
+    Print_Result ${run_converter_result_file}
+    exit 1
+fi
+
+
+# Write net_train result to temp file
+run_net_train_result_file=${logs_path}/run_net_train_result.txt
+echo ' ' > ${run_net_train_result_file}
+
+# Create log files
+run_x86_log_file=${logs_path}/run_x86_log.txt
+echo 'run x86 logs: ' > ${run_x86_log_file}
+
+run_arm64_log_file=${logs_path}/run_arm64_log.txt
+echo 'run arm64 logs: ' > ${run_arm64_log_file}
+adb_push_arm64_log_file=${logs_path}/adb_push_arm64_log.txt
+adb_cmd_arm64_file=${logs_path}/adb_arm64_cmd.txt
+adb_cmd_arm64_run_file=${logs_path}/adb_arm64_cmd_run.txt
+
+run_arm32_log_file=${logs_path}/run_arm32_log.txt
+echo 'run arm32 logs: ' > ${run_arm64_log_file}
+adb_push_arm32_log_file=${logs_path}/adb_push_arm32_log.txt
+adb_cmd_arm32_file=${logs_path}/adb_arm32_cmd.txt
+adb_cmd_arm32_run_file=${logs_path}/adb_arm32_cmd_run.txt
+
+# Copy the MindSpore models:
+echo "Push files to net_train_test folder and run net_train"
+net_train_test_path=${models_path}/net_train_test
+rm -rf ${net_train_test_path}
+mkdir -p ${net_train_test_path}
+cp -a ${ms_models_path}/*.ms ${net_train_test_path} || exit 1
+
+# Run on x86
+echo "start Run x86 ..."
+Run_x86 &
+Run_x86_PID=$!
+sleep 1
+
+# wait ${Run_x86_PID}
+cat ${run_net_train_result_file}
+wait ${Run_x86_PID}
+Run_x86_status=$?
+# exit 0
+
+# Run on arm64
+echo "start Run arm64 ..."
+Run_arm arm64 
+Run_arm64_status=$?
+sleep 3
+
+# Run on arm32
+echo "start Run arm32 ..."
+Run_arm arm32
+Run_arm32_status=$?
+sleep 1
+
+END=$(date +%s.%N)
+DIFF=$(echo "$END - $START" | bc)
+
+function Print_Benchmark_Result() {
+    MS_PRINT_TESTCASE_START_MSG
+    while read line; do
+        arr=("${line}")
+        printf "%-20s %-100s %-7s\n" ${arr[0]} ${arr[1]} ${arr[2]}
+    done < ${run_net_train_result_file}
+    MS_PRINT_TESTCASE_END_MSG
+}
+
+# Check net_train result and return value
+if [[ ${Run_x86_status} != 0 ]];then
+    echo "Run_x86 failed"
+    cat ${run_x86_log_file}
+fi
+
+if [[ ${Run_arm64_status} != 0 ]];then
+    echo "Run_arm64 failed"
+    cat ${run_arm64_log_file}
+fi
+
+if [[ ${Run_arm32_status} != 0 ]];then
+    echo "Run_arm32 failed"
+    cat ${run_arm32_log_file}
+fi
+
+echo "Test ended - Results:"
+Print_Benchmark_Result
+echo "Test run Time:" $DIFF
+exit 0
diff --git a/mindspore/lite/test/run_train_ut.sh b/mindspore/lite/test/run_train_ut.sh
index 669373b950..6bcccf4d17 100755
--- a/mindspore/lite/test/run_train_ut.sh
+++ b/mindspore/lite/test/run_train_ut.sh
@@ -1,5 +1,5 @@
 #!/bin/bash
 cd ./ut/src/runtime/kernel/arm || exit 1
 ../../../../../../build/test/lite-test --gtest_filter=NetworkTest.efficient_net
-../../../../../../build/test/lite-test --gtest_filter=NetworkTest.tuning_layer
-../../../../../../build/test/lite-test --gtest_filter=NetworkTest.lenetnet
\ No newline at end of file
+# ../../../../../../build/test/lite-test --gtest_filter=NetworkTest.tuning_layer
+# ../../../../../../build/test/lite-test --gtest_filter=NetworkTest.lenetnet
\ No newline at end of file
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/activation_grad_fp32_tests.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/activation_grad_fp32_tests.cc
index f5863fffc7..2f52e3ff18 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/activation_grad_fp32_tests.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/activation_grad_fp32_tests.cc
@@ -42,11 +42,17 @@ TEST_F(TestActGradFp32, ReluGradFp32) {
   size_t input_size;
   std::string input_path = "./test_data/activationGrad/relu_y_50.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   EXPECT_EQ(input_size, output_data_size * sizeof(float));
+
   std::string yt_path = "./test_data/activationGrad/relu_yt_50.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
   EXPECT_EQ(input_size, output_data_size * sizeof(float));
+
   auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
+
   // warm up loop
   for (int i = 0; i < 3; i++) {
     ReluGrad(yt_data, input_data, output_data_size, output_data);
@@ -90,10 +96,15 @@ TEST_F(TestActGradFp32, Relu6GradFp32) {
   size_t input_size;
   std::string input_path = "./test_data/activationGrad/relu6_y_50.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
+
   std::string yt_path = "./test_data/activationGrad/relu6_yt_50.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
 
   auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
+
   // warm up loop
   for (int i = 0; i < 3; i++) {
     Relu6Grad(yt_data, input_data, 50, output_data);
@@ -136,10 +147,15 @@ TEST_F(TestActGradFp32, LReluGradFp32) {
   size_t input_size;
   std::string input_path = "./test_data/activationGrad/lrelu_y_50.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
+
   std::string yt_path = "./test_data/activationGrad/lrelu_yt_50.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
 
   auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
+
   // warm up loop
   for (int i = 0; i < 3; i++) {
     LReluGrad(yt_data, input_data, 50, output_data, 0.1);
@@ -182,10 +198,15 @@ TEST_F(TestActGradFp32, SigmoidGradFp32) {
   size_t input_size;
   std::string input_path = "./test_data/activationGrad/sigmoid_y_50.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
+
   std::string yt_path = "./test_data/activationGrad/sigmoid_yt_50.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
 
   auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
+
   // warm up loop
   for (int i = 0; i < 3; i++) {
     SigmoidGrad(yt_data, input_data, 50, output_data);
@@ -229,10 +250,15 @@ TEST_F(TestActGradFp32, tanhGradFp32) {
   size_t input_size;
   std::string input_path = "./test_data/activationGrad/tanh_y_50.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
+
   std::string yt_path = "./test_data/activationGrad/tanh_yt_50.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
 
   auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
+
   // warm up loop
   for (int i = 0; i < 3; i++) {
     TanhGrad(yt_data, input_data, 50, output_data);
@@ -274,11 +300,17 @@ TEST_F(TestActGradFp32, hswishGradFp32) {
   size_t input_size;
   std::string input_path = "./test_data/activationGrad/hswish_x_50.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   EXPECT_EQ(input_size, output_data_size * sizeof(float));
+
   std::string yt_path = "./test_data/activationGrad/hswish_yt_50.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
   EXPECT_EQ(input_size, output_data_size * sizeof(float));
+
   auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
+
   // warm up loop
   for (int i = 0; i < 3; i++) {
     HSwishGrad(yt_data, input_data, static_cast<int>(output_data_size), output_data);
@@ -311,4 +343,58 @@ TEST_F(TestActGradFp32, hswishGradFp32) {
   delete[] yt_data;
   MS_LOG(INFO) << "hswishGradFp32 passed";
 }
+
+TEST_F(TestActGradFp32, hsigmoidGradFp32) {
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+  const size_t output_data_size = 10;
+
+  size_t input_size;
+  std::string input_path = "./test_data/activationGrad/hsig_x_50.bin";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
+  EXPECT_EQ(input_size, output_data_size * sizeof(float));
+
+  std::string yt_path = "./test_data/activationGrad/hsig_yt_50.bin";
+  auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
+  EXPECT_EQ(input_size, output_data_size * sizeof(float));
+
+  auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
+
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    HSigmoidGrad(yt_data, input_data, static_cast<int>(output_data_size), output_data);
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    HSigmoidGrad(yt_data, input_data, output_data_size, output_data);
+  }
+  auto time_end = mindspore::lite::GetTimeUs();
+  auto cost = time_end - time_start;
+  time_avg = cost / loop_count;
+  printf("single thread running time : %f ms\n", time_avg / 1000.0f);
+
+  printf("==================output data=================\n");
+  size_t min = (output_data_size < 20UL) ? output_data_size : 20UL;
+  for (size_t i = 0; i < min; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string output_path = "./test_data/activationGrad/hsig_out_50.bin";
+  int res = CompareRelativeOutput(output_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete[] input_data;
+  delete[] output_data;
+  delete[] yt_data;
+  MS_LOG(INFO) << "hsigmoidGradFp32 passed";
+}
+
 }  // namespace mindspore
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/arithmetic_grad_fp32_tests.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/arithmetic_grad_fp32_tests.cc
index 1b98f1db55..788fb7971c 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/arithmetic_grad_fp32_tests.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/arithmetic_grad_fp32_tests.cc
@@ -24,9 +24,9 @@
 #include "src/kernel_registry.h"
 #include "src/ops/arithmetic_grad.h"
 
-#ifdef PRIMITIVE_WRITEABLE
 namespace mindspore {
 
+#ifdef PRIMITIVE_WRITEABLE
 ArithmeticParameter *PopulateArithmeticParameter(mindspore::schema::PrimitiveType type,
                                                  std::vector<lite::Tensor *> inputs,
                                                  std::vector<lite::Tensor *> outputs) {
@@ -37,6 +37,12 @@ ArithmeticParameter *PopulateArithmeticParameter(mindspore::schema::PrimitiveTyp
   }
   arithmetic_param->op_parameter_.type_ = type;
   schema::PrimitiveT *prim = new schema::PrimitiveT;
+  if (prim == nullptr) {
+    free(arithmetic_param);
+    MS_LOG(ERROR) << "new PrimitiveT failed.";
+    return nullptr;
+  }
+
   prim->value.type = type;
   auto agrad = mindspore::lite::ArithmeticGrad(prim);
   agrad.InferShape(inputs, outputs);
@@ -55,6 +61,7 @@ class TestArithmeticGradFp32 : public mindspore::CommonTest {
 
 std::vector<lite::Tensor *> GenerateTensorsForTest(const char *test, int test_id) {
   size_t input_size;
+  std::vector<lite::Tensor *> ret_vector;
   std::vector<int> large_dim({4, 6});
   std::vector<int> small_dim({6});
   int large_size = (4 * 6);
@@ -80,36 +87,127 @@ std::vector<lite::Tensor *> GenerateTensorsForTest(const char *test, int test_id
   }
 
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(test, &input_size));
+  if (dy_data == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    return ret_vector;
+  }
   lite::Tensor *dy_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, large_dim);
+  if (dy_tensor == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    delete[] dy_data;
+    return ret_vector;
+  }
   dy_tensor->set_data(dy_data);
 
   auto x1_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dx1_file, &input_size));
+  if (x1_data == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    delete[] dy_data;
+    delete dy_tensor;
+    return ret_vector;
+  }
   lite::Tensor *x1_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, large_dim);
+  if (x1_tensor == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    delete[] dy_data;
+    delete dy_tensor;
+    delete[] x1_data;
+    return ret_vector;
+  }
   x1_tensor->set_data(x1_data);
 
   auto x2_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dx2_file, &input_size));
+  if (x2_data == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    delete[] dy_data;
+    delete dy_tensor;
+    delete[] x1_data;
+    delete x1_tensor;
+    return ret_vector;
+  }
   lite::Tensor *x2_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, small_dim);
+  if (x2_tensor == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    delete[] dy_data;
+    delete dy_tensor;
+    delete[] x1_data;
+    delete x1_tensor;
+    delete[] x2_data;
+    return ret_vector;
+  }
   x2_tensor->set_data(x2_data);
 
   auto dx1_data = new float[large_size];
+  if (dx1_data == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    delete[] dy_data;
+    delete dy_tensor;
+    delete[] x1_data;
+    delete x1_tensor;
+    delete[] x2_data;
+    delete x2_tensor;
+    return ret_vector;
+  }
   lite::Tensor *dx1_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, large_dim);
+  if (dx1_tensor == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    delete[] dy_data;
+    delete dy_tensor;
+    delete[] x1_data;
+    delete x1_tensor;
+    delete[] x2_data;
+    delete x2_tensor;
+    delete[] dx1_data;
+    return ret_vector;
+  }
   dx1_tensor->set_data(dx1_data);
 
   auto dx2_data = new float[small_size];
+  if (dx2_data == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    delete[] dy_data;
+    delete dy_tensor;
+    delete[] x1_data;
+    delete x1_tensor;
+    delete[] x2_data;
+    delete x2_tensor;
+    delete[] dx1_data;
+    delete dx1_tensor;
+    return ret_vector;
+  }
   lite::Tensor *dx2_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, small_dim);
+  if (dx2_tensor == nullptr) {
+    MS_LOG(ERROR) << "new operator failed";
+    delete[] dy_data;
+    delete dy_tensor;
+    delete[] x1_data;
+    delete x1_tensor;
+    delete[] x2_data;
+    delete x2_tensor;
+    delete[] dx1_data;
+    delete dx1_tensor;
+    delete[] dx2_data;
+    return ret_vector;
+  }
   dx2_tensor->set_data(dx2_data);
 
-  std::vector<lite::Tensor *> ret_vector = {dy_tensor, x1_tensor, x2_tensor, dx1_tensor, dx2_tensor};
+  ret_vector.push_back(dy_tensor);
+  ret_vector.push_back(x1_tensor);
+  ret_vector.push_back(x2_tensor);
+  ret_vector.push_back(dx1_tensor);
+  ret_vector.push_back(dx2_tensor);
+
   return ret_vector;
 }
 
 TEST_F(TestArithmeticGradFp32, TestAddGradFp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_1_dy_4_6.bin", 1);
-
+  ASSERT_NE(all_tensors.size(), 0);
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
   std::vector<lite::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_AddGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -117,7 +215,9 @@ TEST_F(TestArithmeticGradFp32, TestAddGradFp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_AddGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[1]->MutableData());
@@ -144,10 +244,12 @@ TEST_F(TestArithmeticGradFp32, TestAddGradFp32) {
 TEST_F(TestArithmeticGradFp32, TestAddGrad2Fp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_1_dy_4_6.bin", 1);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
   std::vector<lite::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_AddGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -155,7 +257,9 @@ TEST_F(TestArithmeticGradFp32, TestAddGrad2Fp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_AddGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[0]->MutableData());
@@ -184,10 +288,12 @@ TEST_F(TestArithmeticGradFp32, TestAddGrad2Fp32) {
 TEST_F(TestArithmeticGradFp32, TestAddGrad3Fp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_8_dy_5_4_6.bin", 8);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
   std::vector<lite::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_AddGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -195,7 +301,9 @@ TEST_F(TestArithmeticGradFp32, TestAddGrad3Fp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_AddGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[0]->MutableData());
@@ -225,10 +333,12 @@ TEST_F(TestArithmeticGradFp32, TestAddGrad3Fp32) {
 TEST_F(TestArithmeticGradFp32, TestSubGradFp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_2_dy_4_6.bin", 2);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
   std::vector<lite::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_SubGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -236,7 +346,9 @@ TEST_F(TestArithmeticGradFp32, TestSubGradFp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_SubGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[1]->MutableData());
@@ -266,10 +378,12 @@ TEST_F(TestArithmeticGradFp32, TestSubGradFp32) {
 TEST_F(TestArithmeticGradFp32, TestSubGrad2Fp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_3_dy_4_6.bin", 3);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
   std::vector<lite::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_SubGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -277,7 +391,9 @@ TEST_F(TestArithmeticGradFp32, TestSubGrad2Fp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_SubGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[0]->MutableData());
@@ -305,10 +421,12 @@ TEST_F(TestArithmeticGradFp32, TestSubGrad2Fp32) {
 TEST_F(TestArithmeticGradFp32, TestMulGradFp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_4_dy_4_6.bin", 4);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
   std::vector<lite::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_MulGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -316,8 +434,9 @@ TEST_F(TestArithmeticGradFp32, TestMulGradFp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_MulGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
-
+  ASSERT_NE(kernel_obj, nullptr);
   int loop_count = 1000;
   auto time_start = mindspore::lite::GetTimeUs();
   for (int i = 0; i < loop_count; i++) {
@@ -354,10 +473,12 @@ TEST_F(TestArithmeticGradFp32, TestMulGradFp32) {
 TEST_F(TestArithmeticGradFp32, TestMulGrad2Fp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_4_dy_4_6.bin", 4);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
   std::vector<lite::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_MulGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -365,7 +486,9 @@ TEST_F(TestArithmeticGradFp32, TestMulGrad2Fp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_MulGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[0]->MutableData());
@@ -394,10 +517,12 @@ TEST_F(TestArithmeticGradFp32, TestMulGrad2Fp32) {
 TEST_F(TestArithmeticGradFp32, TestMulGrad3Fp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_9_dy_5_4_6.bin", 9);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
   std::vector<lite::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_MulGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -405,7 +530,9 @@ TEST_F(TestArithmeticGradFp32, TestMulGrad3Fp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_MulGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[1]->MutableData());
@@ -434,10 +561,12 @@ TEST_F(TestArithmeticGradFp32, TestMulGrad3Fp32) {
 TEST_F(TestArithmeticGradFp32, TestMulGrad4Fp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_9_dy_5_4_6.bin", 9);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
   std::vector<lite::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_MulGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -445,7 +574,9 @@ TEST_F(TestArithmeticGradFp32, TestMulGrad4Fp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_MulGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[0]->MutableData());
@@ -474,10 +605,12 @@ TEST_F(TestArithmeticGradFp32, TestMulGrad4Fp32) {
 TEST_F(TestArithmeticGradFp32, TestDivGradFp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_5_dy_4_6.bin", 5);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
   std::vector<lite::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_DivGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -485,7 +618,9 @@ TEST_F(TestArithmeticGradFp32, TestDivGradFp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DivGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[1]->MutableData());
@@ -514,10 +649,12 @@ TEST_F(TestArithmeticGradFp32, TestDivGradFp32) {
 TEST_F(TestArithmeticGradFp32, TestDivGrad2Fp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_6_dy_4_6.bin", 6);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[2], all_tensors[1]};
   std::vector<lite::Tensor *> outputs = {all_tensors[4], all_tensors[3]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_DivGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -525,7 +662,9 @@ TEST_F(TestArithmeticGradFp32, TestDivGrad2Fp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DivGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[0]->MutableData());
@@ -555,10 +694,12 @@ TEST_F(TestArithmeticGradFp32, TestDivGrad2Fp32) {
 TEST_F(TestArithmeticGradFp32, TestDivGrad3Fp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_10_dy_5_4_6.bin", 10);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
   std::vector<lite::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_DivGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -566,7 +707,9 @@ TEST_F(TestArithmeticGradFp32, TestDivGrad3Fp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DivGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[1]->MutableData());
@@ -595,10 +738,12 @@ TEST_F(TestArithmeticGradFp32, TestDivGrad3Fp32) {
 TEST_F(TestArithmeticGradFp32, Test3DDivGrad2Fp32) {
   std::vector<lite::Tensor *> all_tensors =
     GenerateTensorsForTest("./test_data/operators/arithmetic_fp32_7_dy_4_5_6.bin", 7);
+  ASSERT_NE(all_tensors.size(), 0);
 
   std::vector<lite::Tensor *> inputs = {all_tensors[0], all_tensors[1], all_tensors[2]};
   std::vector<lite::Tensor *> outputs = {all_tensors[3], all_tensors[4]};
   auto param = PopulateArithmeticParameter(schema::PrimitiveType_DivGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
 
   lite::InnerContext ctx;
   ctx.thread_num_ = 1;
@@ -606,7 +751,9 @@ TEST_F(TestArithmeticGradFp32, Test3DDivGrad2Fp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DivGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   float *output_ptr = reinterpret_cast<float *>(outputs[1]->MutableData());
@@ -630,6 +777,91 @@ TEST_F(TestArithmeticGradFp32, Test3DDivGrad2Fp32) {
   MS_LOG(INFO) << "TestDivGrad2Fp32 passed";
 }
 
-}  // namespace mindspore
+TEST_F(TestArithmeticGradFp32, TestMaximumGradBroadcastFp32) {
+  std::vector<int> large_dim({4, 6});
+  std::vector<int> small_dim({6});
+
+  large_dim = std::vector<int>({1, 2, 3});
+  small_dim = std::vector<int>({1, 3});
+  int large_size = (2 * 3);
+  int small_size = 3;
+  size_t input_size;
+  char *dx1_file = const_cast<char *>("./test_data/operators/x1_maximum.bin");
+  char *dx2_file = const_cast<char *>("./test_data/operators/x2_maximum.bin");
+
+  std::string yt_path = "./test_data/operators/yt_maximum.bin";
+  auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(dy_data, nullptr);
+  EXPECT_EQ(input_size, large_size * sizeof(float));
+  lite::Tensor *dy_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, large_dim);
+  ASSERT_NE(dy_tensor, nullptr);
+  dy_tensor->set_data(dy_data);
+
+  auto x1_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dx1_file, &input_size));
+  ASSERT_NE(x1_data, nullptr);
+  lite::Tensor *x1_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, small_dim);
+  ASSERT_NE(x1_tensor, nullptr);
+  x1_tensor->set_data(x1_data);
+
+  auto x2_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dx2_file, &input_size));
+  ASSERT_NE(x2_data, nullptr);
+  lite::Tensor *x2_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, large_dim);
+  ASSERT_NE(x2_tensor, nullptr);
+  x2_tensor->set_data(x2_data);
+
+  auto dx1_data = new float[small_size];
+  ASSERT_NE(dx1_data, nullptr);
+  lite::Tensor *dx1_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, small_dim);
+  ASSERT_NE(dx1_tensor, nullptr);
+  dx1_tensor->set_data(dx1_data);
 
+  auto dx2_data = new float[large_size];
+  ASSERT_NE(dx2_data, nullptr);
+  lite::Tensor *dx2_tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, large_dim);
+  ASSERT_NE(dx2_tensor, nullptr);
+  dx2_tensor->set_data(dx2_data);
+
+  std::vector<lite::Tensor *> inputs = {x1_tensor, x2_tensor, dy_tensor};
+  std::vector<lite::Tensor *> outputs = {dx1_tensor, dx2_tensor};
+
+  auto param = PopulateArithmeticParameter(schema::PrimitiveType_MaximumGrad, inputs, outputs);
+  ASSERT_NE(param, nullptr);
+
+  lite::InnerContext ctx;
+  ctx.thread_num_ = 1;
+  ASSERT_EQ(lite::RET_OK, ctx.Init());
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_MaximumGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
+  kernel_obj->Run();
+
+  float *output_ptr = reinterpret_cast<float *>(outputs[1]->MutableData());
+  printf("==================output data=================\n");
+  for (int i = 0; i < 6; i++) {
+    std::cout << output_ptr[i] << " ,";
+  }
+  std::cout << std::endl;
+
+  std::string dx1_path = "./test_data/operators/x1_grad_maximum.bin";
+  EXPECT_EQ(0, CompareRelativeOutput(reinterpret_cast<float *>(outputs[0]->MutableData()), dx1_path));
+
+  std::string output_path = "./test_data/operators/x2_grad_maximum.bin";
+  EXPECT_EQ(0, CompareRelativeOutput(output_ptr, output_path));
+  for (auto tensor : inputs) {
+    delete[] reinterpret_cast<float *>(tensor->MutableData());
+    tensor->set_data(nullptr);
+    delete tensor;
+  }
+  for (auto tensor : outputs) {
+    delete[] reinterpret_cast<float *>(tensor->MutableData());
+    tensor->set_data(nullptr);
+    delete tensor;
+  }
+  delete kernel_obj;
+  MS_LOG(INFO) << "TestMaximumGradBroadcastFp32 passed";
+}
 #endif
+}  // namespace mindspore
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bias_grad_fp32_tests.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bias_grad_fp32_tests.cc
index 03659e39ee..b0d94b9d2c 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bias_grad_fp32_tests.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bias_grad_fp32_tests.cc
@@ -31,15 +31,20 @@ class TestBiasGradFp32 : public mindspore::CommonTest {
 TEST_F(TestBiasGradFp32, BiasGradFp32) {
   // prepare stage
   ArithmeticParameter *bias_param = static_cast<ArithmeticParameter *>(malloc(sizeof(ArithmeticParameter)));
+  ASSERT_NE(bias_param, nullptr);
+
   size_t input_size;
   std::string input_path = "./test_data/operators/biasgradfp32_1_dy_10_28_28_7.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
+
   std::vector<int> dim_dy({10, 28, 28, 7});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(input_data);
 
   std::vector<lite::Tensor *> inputs = {&dy_tensor};
   auto output_data = new float[7];
+  ASSERT_NE(output_data, nullptr);
   std::vector<int> dim_dw = {7};
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(output_data);
@@ -51,8 +56,9 @@ TEST_F(TestBiasGradFp32, BiasGradFp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_BiasGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(bias_param), &ctx, desc, nullptr);
-
+  ASSERT_NE(kernel_obj, nullptr);
   kernel_obj->Run();
 
   printf("==================output data=================\n");
@@ -61,7 +67,57 @@ TEST_F(TestBiasGradFp32, BiasGradFp32) {
   }
   std::cout << std::endl;
   std::string output_path = "./test_data/operators/biasgradfp32_1_db_7.bin";
-  CompareOutput(output_data, 7, output_path);
+  auto res = CompareRelativeOutput(output_data, output_path);
+  EXPECT_EQ(res, 0);
+
+  delete[] input_data;
+  delete[] output_data;
+  // delete bias_param;
+  dy_tensor.set_data(nullptr);
+  dw_tensor.set_data(nullptr);
+  delete kernel_obj;
+  MS_LOG(INFO) << "BiasGradFp32 passed";
+}
+
+TEST_F(TestBiasGradFp32, BiasGrad2DFp32) {
+  // prepare stage
+  ArithmeticParameter *bias_param = static_cast<ArithmeticParameter *>(malloc(sizeof(ArithmeticParameter)));
+  ASSERT_NE(bias_param, nullptr);
+
+  size_t input_size;
+  std::string input_path = "./test_data/operators/fc_yt.f32";
+  auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  std::vector<int> dim_dy({2, 20});
+  lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.set_data(input_data);
+
+  std::vector<lite::Tensor *> inputs = {&dy_tensor};
+  auto output_data = new float[20];
+  ASSERT_NE(output_data, nullptr);
+  std::vector<int> dim_dw = {20};
+  lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
+  dw_tensor.set_data(output_data);
+  std::vector<lite::Tensor *> outputs = {&dw_tensor};
+
+  lite::InnerContext ctx;
+  ctx.thread_num_ = 1;
+  ASSERT_EQ(lite::RET_OK, ctx.Init());
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_BiasGrad};
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(bias_param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
+  kernel_obj->Run();
+
+  printf("==================output data=================\n");
+  for (int i = 0; i < 20; i++) {
+    std::cout << output_data[i] << " ,";
+  }
+  std::cout << std::endl;
+  std::string output_path = "./test_data/operators/fc_b_grad.f32";
+  auto res = CompareRelativeOutput(output_data, output_path);
+  EXPECT_EQ(res, 0);
 
   delete[] input_data;
   delete[] output_data;
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bn_grad_fp32_test.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bn_grad_fp32_test.cc
index 913d685725..bd0554c04d 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bn_grad_fp32_test.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bn_grad_fp32_test.cc
@@ -35,6 +35,10 @@ lite::Tensor *TestBNGradFp32::CreateInTensor(std::string file_name, std::vector<
   size_t input_size = 0;
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(file_name.c_str(), &input_size));
   auto tensor = new lite::Tensor(TypeId::kNumberTypeFloat32, dim);
+  if (tensor == nullptr) {
+    MS_LOG(ERROR) << "new tensor failed";
+    return nullptr;
+  }
   tensor->set_data(input_data);
   EXPECT_EQ(input_size, tensor->Size());
   return tensor;
@@ -43,7 +47,9 @@ lite::Tensor *TestBNGradFp32::CreateInTensor(std::string file_name, std::vector<
 TEST_F(TestBNGradFp32, BNGradFp32) {
   // prepare stage
   auto bn_param = static_cast<BNGradParameter *>(malloc(sizeof(BNGradParameter)));
-  bn_param->epsilon_ = 0.00001;
+  ASSERT_NE(bn_param, nullptr);
+
+  bn_param->epsilon_ = 1e-2;
   bn_param->momentum_ = 0.1;
   const int batch = 2;
   const int channels = 3;
@@ -51,10 +57,16 @@ TEST_F(TestBNGradFp32, BNGradFp32) {
   const int width = 5;
 
   auto dy_tensor = CreateInTensor("./test_data/bngrad/dy_2_4_5_3.bin", {batch, height, width, channels});
+  ASSERT_NE(dy_tensor, nullptr);
   auto x_tensor = CreateInTensor("./test_data/bngrad/input_x_2_4_5_3.bin", {batch, height, width, channels});
+  ASSERT_NE(x_tensor, nullptr);
   auto scale_tensor = CreateInTensor("./test_data/bngrad/scale_3.bin", {1, 1, 1, channels});
+  ASSERT_NE(scale_tensor, nullptr);
   auto mean_tensor = CreateInTensor("./test_data/bngrad/save_mean_3.bin", {1, 1, 1, channels});
+  ASSERT_NE(mean_tensor, nullptr);
   auto var_tensor = CreateInTensor("././test_data/bngrad/save_var_3.bin", {1, 1, 1, channels});
+  ASSERT_NE(var_tensor, nullptr);
+
   // prepare output tensors
   lite::Tensor dx_tensor(TypeId::kNumberTypeFloat32, {batch, height, width, channels});
   ASSERT_EQ(dx_tensor.MallocData(), 0);
@@ -72,27 +84,18 @@ TEST_F(TestBNGradFp32, BNGradFp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_BNGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(bn_param), &ctx, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel_obj->GetWorkspaceSize());
 
-  for (int i = 0; i < 3; i++) {
-    kernel_obj->Run();
-  }
-
-  int loop_count = 100;
-  auto time_start = mindspore::lite::GetTimeUs();
-  for (int i = 0; i < loop_count; i++) {
-    kernel_obj->Run();
-  }
-  auto time_end = mindspore::lite::GetTimeUs();
-  auto cost = time_end - time_start;
-  auto time_avg = cost / loop_count;
-  std::cout << "single thread running time : " << time_avg << "us\n";
+  kernel_obj->Run();
   std::cout << "==========dx==========\n";
   auto dx = reinterpret_cast<float *>(outputs[0]->MutableData());
   for (int i = 0; i < 7; i++) std::cout << dx[i] << " ";
   std::cout << "\n";
   auto res = CompareRelativeOutput(dx, "./test_data/bngrad/output_dx_2_4_5_3.bin");
+  EXPECT_EQ(res, 0);
   std::cout << "\n=======dscale=======\n";
   auto dscale = reinterpret_cast<float *>(outputs[1]->MutableData());
   for (int i = 0; i < channels; i++) std::cout << dscale[i] << " ";
@@ -104,7 +107,6 @@ TEST_F(TestBNGradFp32, BNGradFp32) {
   for (int i = 0; i < 3; i++) std::cout << dbias[i] << " ";
   std::cout << "\n";
   res = CompareRelativeOutput(dbias, "./test_data/bngrad/output_dbias_3.bin");
-  EXPECT_EQ(res, 0);
   for (auto v : inputs) {
     delete[] reinterpret_cast<float *>(v->MutableData());
     v->set_data(nullptr);
@@ -117,8 +119,10 @@ TEST_F(TestBNGradFp32, BNGradFp32) {
 
 TEST_F(TestBNGradFp32, BNTtrainFp32) {
   auto bn_param = static_cast<BatchNormParameter *>(malloc(sizeof(BatchNormParameter)));
-  bn_param->epsilon_ = 0.00001;
-  bn_param->momentum_ = 0.;
+  ASSERT_NE(bn_param, nullptr);
+
+  bn_param->epsilon_ = 1e-2;
+  bn_param->momentum_ = 0.1;
   const int batch = 2;
   const int channels = 3;
   const int height = 4;
@@ -173,27 +177,34 @@ TEST_F(TestBNGradFp32, BNTtrainFp32) {
   ASSERT_EQ(lite::RET_OK, context.Init());
 
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(bn_param), &context, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel_obj->GetWorkspaceSize());
 
   float *save_mean = reinterpret_cast<float *>(save_mean_tensor.MutableData());
   float *save_var = reinterpret_cast<float *>(save_var_tensor.MutableData());
-  std::fill(save_mean, save_mean + channels, 0.f);
-  std::fill(save_var, save_var + channels, 0.f);
+  for (int i = 0; i < channels; i++) {
+    save_var[i] = 1.f;
+    save_mean[i] = 0.f;
+  }
+  float *curr_mean = reinterpret_cast<float *>(mean_tensor.MutableData());
+  float *curr_var = reinterpret_cast<float *>(var_tensor.MutableData());
 
-  kernel_obj->train();
+  kernel_obj->Train();
+  kernel_obj->SetTrainable(true);
   kernel_obj->Run();
 
   std::cout << "================save_mean==============================\n";
-  for (int i = 0; i < channels; i++) std::cout << save_mean[i] << " ";
+  for (int i = 0; i < channels; i++) std::cout << curr_mean[i] << " ";
   std::cout << "\n";
   std::cout << "===============save_var==============================\n";
-  for (int i = 0; i < channels; i++) std::cout << save_var[i] << " ";
+  for (int i = 0; i < channels; i++) std::cout << curr_var[i] << " ";
   std::cout << "\n";
   delete[] reinterpret_cast<float *>(x_tensor->MutableData());
-  auto res = CompareRelativeOutput(save_mean, "./test_data/bngrad/running_mean_3.bin");
+  auto res = CompareRelativeOutput(curr_mean, "./test_data/bngrad/running_mean_3.bin");
   EXPECT_EQ(res, 0);
-  res = CompareRelativeOutput(save_var, "./test_data/bngrad/running_var_3.bin");
+  res = CompareRelativeOutput(curr_var, "./test_data/bngrad/running_var_3.bin");
   EXPECT_EQ(res, 0);
 
   x_tensor->set_data(nullptr);
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/convolution_grad_fp32_tests.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/convolution_grad_fp32_tests.cc
index 080cb1d17c..0625adb70f 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/convolution_grad_fp32_tests.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/convolution_grad_fp32_tests.cc
@@ -77,11 +77,13 @@ void InitConvParamGroup3Dilation2FP32(ConvParameter *conv_param) {
 TEST_F(TestConvolutionGradFp32, ConvFp32FilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
-  InitConvParamGroup1FP32(conv_param);
+  ASSERT_NE(conv_param, nullptr);
 
+  InitConvParamGroup1FP32(conv_param);
   size_t dy_size;
   std::string dy_path = "./test_data/conv/convfp32_dy_1_28_28_32.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  ASSERT_NE(dy_data, nullptr);
   std::vector<int> dim_dy({1, 28, 28, 32});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(dy_data);
@@ -95,11 +97,13 @@ TEST_F(TestConvolutionGradFp32, ConvFp32FilterGrad) {
   size_t input_size;
   std::string input_path = "./test_data/conv/convfp32_x_1_28_28_3.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_x({1, 28, 28, 3});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({32, 3, 3, 3});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -112,7 +116,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32FilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
   // warm up loop
   for (int i = 0; i < 3; i++) {
@@ -149,8 +155,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32FilterGrad) {
 TEST_F(TestConvolutionGradFp32, ConvFp32InputGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
-  InitConvParamGroup1FP32(conv_param);
+  ASSERT_NE(conv_param, nullptr);
 
+  InitConvParamGroup1FP32(conv_param);
   size_t dy_size;
   std::string dy_path = "./test_data/conv/convfp32_dy_1_28_28_32.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
@@ -168,6 +175,7 @@ TEST_F(TestConvolutionGradFp32, ConvFp32InputGrad) {
   size_t output_data_size =
     conv_param->input_batch_ * conv_param->input_h_ * conv_param->input_w_ * conv_param->input_channel_;
   auto dx_data = new float[output_data_size];
+  ASSERT_NE(dx_data, nullptr);
   std::vector<int> dim_dx({1, 28, 28, 3});
   lite::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
   dx_tensor.set_data(dx_data);
@@ -185,7 +193,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32InputGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradInput};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
@@ -222,8 +232,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32InputGrad) {
 TEST_F(TestConvolutionGradFp32, ConvFp32GroupFilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
-  InitConvParamGroup3FP32(conv_param);
+  ASSERT_NE(conv_param, nullptr);
 
+  InitConvParamGroup3FP32(conv_param);
   size_t dy_size;
   std::string dy_path = "./test_data/conv/convfp32_dy_g3_1_28_28_18.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
@@ -245,6 +256,7 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupFilterGrad) {
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({18, 3, 3, 1});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -257,7 +269,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupFilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
   // warm up loop
   for (int i = 0; i < 3; i++) {
@@ -293,8 +307,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupFilterGrad) {
 TEST_F(TestConvolutionGradFp32, ConvFp32GroupInputGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
-  InitConvParamGroup3FP32(conv_param);
+  ASSERT_NE(conv_param, nullptr);
 
+  InitConvParamGroup3FP32(conv_param);
   size_t dy_size;
   std::string dy_path = "./test_data/conv/convfp32_dy_g3_1_28_28_18.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
@@ -312,6 +327,7 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupInputGrad) {
   size_t output_data_size =
     conv_param->input_batch_ * conv_param->input_h_ * conv_param->input_w_ * conv_param->input_channel_;
   auto dx_data = new float[output_data_size];
+  ASSERT_NE(dx_data, nullptr);
   std::vector<int> dim_dx({1, 28, 28, 3});
   lite::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
   dx_tensor.set_data(dx_data);
@@ -329,7 +345,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupInputGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradInput};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
   // warm up loop
   for (int i = 0; i < 3; i++) {
@@ -365,9 +383,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupInputGrad) {
 TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationFilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  ASSERT_NE(conv_param, nullptr);
 
   InitConvParamGroup3Dilation2FP32(conv_param);
-
   size_t dy_size;
   std::string dy_path = "./test_data/conv/convfp32_dy_g3_d2_1_26_26_18.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
@@ -389,6 +407,7 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationFilterGrad) {
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({18, 3, 3, 1});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -401,7 +420,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationFilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
@@ -437,8 +458,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationFilterGrad) {
 TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationInputGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
-  InitConvParamGroup3Dilation2FP32(conv_param);
+  ASSERT_NE(conv_param, nullptr);
 
+  InitConvParamGroup3Dilation2FP32(conv_param);
   size_t dy_size;
   std::string dy_path = "./test_data/conv/convfp32_dy_g3_d2_1_26_26_18.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
@@ -456,6 +478,7 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationInputGrad) {
   size_t output_data_size =
     conv_param->input_batch_ * conv_param->input_h_ * conv_param->input_w_ * conv_param->input_channel_;
   auto dx_data = new float[output_data_size];
+  ASSERT_NE(dx_data, nullptr);
   std::vector<int> dim_dx({1, 28, 28, 3});
   lite::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
   dx_tensor.set_data(dx_data);
@@ -473,7 +496,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationInputGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradInput};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   int loop_count = 100;
@@ -504,8 +529,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationInputGrad) {
 TEST_F(TestConvolutionGradFp32, ConvGroupDilation) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
-  InitConvParamGroup3Dilation2FP32(conv_param);
+  ASSERT_NE(conv_param, nullptr);
 
+  InitConvParamGroup3Dilation2FP32(conv_param);
   size_t x_size;
   std::string x_path = "./test_data/conv/convfp32_x_g3_d2_1_28_28_3.bin";
   auto x_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(x_path.c_str(), &x_size));
@@ -523,6 +549,7 @@ TEST_F(TestConvolutionGradFp32, ConvGroupDilation) {
   size_t output_data_size =
     conv_param->output_batch_ * conv_param->output_h_ * conv_param->output_w_ * conv_param->output_channel_;
   auto y_data = new float[output_data_size];
+  ASSERT_NE(y_data, nullptr);
   std::vector<int> dim_y({1, 26, 26, 18});
   lite::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
   y_tensor.set_data(y_data);
@@ -540,11 +567,12 @@ TEST_F(TestConvolutionGradFp32, ConvGroupDilation) {
 
   auto *kernel = new mindspore::kernel::ConvolutionTrainCPUKernel(reinterpret_cast<OpParameter *>(conv_param), inputs,
                                                                   outputs, &context, 0);
+  ASSERT_NE(kernel, nullptr);
   kernel->Init();
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
-  kernel->train();
-  EXPECT_EQ(kernel->is_train(), 1);
+  kernel->Train();
+  EXPECT_EQ(kernel->IsTrain(), 1);
 
   // warm up loop
   for (int i = 0; i < 3; i++) {
@@ -580,6 +608,8 @@ TEST_F(TestConvolutionGradFp32, ConvGroupDilation) {
 TEST_F(TestConvolutionGradFp32, ConvFp32Dilation2Group2Stride2FilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  ASSERT_NE(conv_param, nullptr);
+
   conv_param->input_batch_ = 2;
   conv_param->input_h_ = 32;
   conv_param->input_w_ = 32;
@@ -624,11 +654,13 @@ TEST_F(TestConvolutionGradFp32, ConvFp32Dilation2Group2Stride2FilterGrad) {
   size_t input_size;
   std::string input_path = "./test_data/conv/convfp32_input0_d2_g2_s2_2_4_32_32.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_x({2, 32, 32, 4});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({12, 3, 3, 2});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -641,7 +673,9 @@ TEST_F(TestConvolutionGradFp32, ConvFp32Dilation2Group2Stride2FilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
@@ -679,6 +713,8 @@ TEST_F(TestConvolutionGradFp32, ConvFp32Dilation2Group2Stride2FilterGrad) {
 TEST_F(TestConvolutionGradFp32, ConvGroup2Dilation2Stride2) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  ASSERT_NE(conv_param, nullptr);
+
   conv_param->input_batch_ = 2;
   conv_param->input_h_ = 32;
   conv_param->input_w_ = 32;
@@ -710,6 +746,7 @@ TEST_F(TestConvolutionGradFp32, ConvGroup2Dilation2Stride2) {
   size_t dy_size;
   std::string dy_path = "./test_data/conv/convfp32_dy_d2_g2_s2_2_12_15_15.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  ASSERT_NE(dy_data, nullptr);
   std::vector<int> dim_dy({2, 15, 15, 12});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(dy_data);
@@ -717,6 +754,7 @@ TEST_F(TestConvolutionGradFp32, ConvGroup2Dilation2Stride2) {
   size_t w_size;
   std::string w_path = "./test_data/conv/convfp32_w_d2_g2_s2_12_2_3_3.bin";
   auto w_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(w_path.c_str(), &w_size));
+  ASSERT_NE(w_data, nullptr);
   std::vector<int> dim_w({12, 3, 3, 2});
   lite::Tensor w_tensor(TypeId::kNumberTypeFloat32, dim_w);
   w_tensor.set_data(w_data);
@@ -724,6 +762,7 @@ TEST_F(TestConvolutionGradFp32, ConvGroup2Dilation2Stride2) {
   size_t output_data_size =
     conv_param->input_batch_ * conv_param->input_h_ * conv_param->input_w_ * conv_param->input_channel_;
   auto dx_data = new float[output_data_size];
+  ASSERT_NE(dx_data, nullptr);
   std::vector<int> dim_dx({2, 32, 32, 4});
   lite::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
   dx_tensor.set_data(dx_data);
@@ -741,7 +780,9 @@ TEST_F(TestConvolutionGradFp32, ConvGroup2Dilation2Stride2) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_Conv2DGradInput};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_fp32_tests.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_fp32_tests.cc
index 571ffa81f8..ecdee356e7 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_fp32_tests.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/deconvolution_grad_fp32_tests.cc
@@ -32,6 +32,8 @@ class TestDeConvolutionGradFp32 : public mindspore::CommonTest {
 TEST_F(TestDeConvolutionGradFp32, DeConvFp32FilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  ASSERT_NE(conv_param, nullptr);
+
   conv_param->input_batch_ = 2;
   conv_param->input_h_ = 32;
   conv_param->input_w_ = 32;
@@ -63,24 +65,24 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32FilterGrad) {
   size_t dy_size;
   std::string dy_path = "./test_data/deconv/deconvfp32_dy_2_9_63_63.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  ASSERT_NE(dy_data, nullptr);
   std::vector<int> dim_dy({2, 63, 63, 9});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(dy_data);
 
-  // runtime part
-  printf("Calculating runtime cost...\n");
-  uint64_t time_avg = 0;
   size_t output_data_size =
     conv_param->output_channel_ * conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_;
 
   size_t input_size;
   std::string input_path = "./test_data/deconv/deconvfp32_input0_2_3_32_32.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_x({2, 32, 32, 3});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({3, 3, 3, 9});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -93,7 +95,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32FilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DeConv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
@@ -101,6 +105,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32FilterGrad) {
     kernel->Run();
   }
 
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
   int loop_count = 100;
   auto time_start = mindspore::lite::GetTimeUs();
   for (int i = 0; i < loop_count; i++) {
@@ -131,6 +138,8 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32FilterGrad) {
 TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2FilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  ASSERT_NE(conv_param, nullptr);
+
   conv_param->input_batch_ = 2;
   conv_param->input_h_ = 32;
   conv_param->input_w_ = 32;
@@ -162,24 +171,24 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2FilterGrad) {
   size_t dy_size;
   std::string dy_path = "./test_data/deconv/deconvfp32_dy_d2_2_9_65_65.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  ASSERT_NE(dy_data, nullptr);
   std::vector<int> dim_dy({2, 65, 65, 9});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(dy_data);
 
-  // runtime part
-  printf("Calculating runtime cost...\n");
-  uint64_t time_avg = 0;
   size_t output_data_size =
     conv_param->output_channel_ * conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_;
 
   size_t input_size;
   std::string input_path = "./test_data/deconv/deconvfp32_input0_d2_2_3_32_32.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_x({2, 32, 32, 3});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({9, 3, 3, 3});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -192,7 +201,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2FilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DeConv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
@@ -200,6 +211,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2FilterGrad) {
     kernel->Run();
   }
 
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
   int loop_count = 100;
   auto time_start = mindspore::lite::GetTimeUs();
   for (int i = 0; i < loop_count; i++) {
@@ -230,6 +244,8 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2FilterGrad) {
 TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3FilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  ASSERT_NE(conv_param, nullptr);
+
   conv_param->input_batch_ = 2;
   conv_param->input_h_ = 32;
   conv_param->input_w_ = 32;
@@ -261,6 +277,7 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3FilterGrad) {
   size_t dy_size;
   std::string dy_path = "./test_data/deconv/deconvfp32_dy_d2_g3_2_9_65_65.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  ASSERT_NE(dy_data, nullptr);
   std::vector<int> dim_dy({2, 65, 65, 9});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(dy_data);
@@ -274,11 +291,13 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3FilterGrad) {
   size_t input_size;
   std::string input_path = "./test_data/deconv/deconvfp32_input0_d2_g3_2_3_32_32.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_x({2, 32, 32, 3});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({3, 3, 3, 3});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -291,7 +310,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3FilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DeConv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
@@ -329,6 +350,8 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3FilterGrad) {
 TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3Stride1FilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  ASSERT_NE(conv_param, nullptr);
+
   conv_param->input_batch_ = 2;
   conv_param->input_h_ = 32;
   conv_param->input_w_ = 32;
@@ -360,24 +383,24 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3Stride1FilterGrad) {
   size_t dy_size;
   std::string dy_path = "./test_data/deconv/deconvfp32_dy_d2_g3_s1_2_9_34_34.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  ASSERT_NE(dy_data, nullptr);
   std::vector<int> dim_dy({2, 34, 34, 9});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(dy_data);
 
-  // runtime part
-  printf("Calculating runtime cost...\n");
-  uint64_t time_avg = 0;
   size_t output_data_size =
     conv_param->output_channel_ * conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_;
 
   size_t input_size;
   std::string input_path = "./test_data/deconv/deconvfp32_input0_d2_g3_s1_2_3_32_32.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_x({2, 32, 32, 3});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({3, 3, 3, 3});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -390,7 +413,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3Stride1FilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DeConv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
@@ -398,6 +423,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3Stride1FilterGrad) {
     kernel->Run();
   }
 
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
   int loop_count = 100;
   auto time_start = mindspore::lite::GetTimeUs();
   for (int i = 0; i < loop_count; i++) {
@@ -428,6 +456,8 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group3Stride1FilterGrad) {
 TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group2Stride2FilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  ASSERT_NE(conv_param, nullptr);
+
   conv_param->input_batch_ = 2;
   conv_param->input_h_ = 32;
   conv_param->input_w_ = 32;
@@ -459,24 +489,24 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group2Stride2FilterGrad) {
   size_t dy_size;
   std::string dy_path = "./test_data/deconv/deconvfp32_dy_d2_g2_s2_2_12_65_65.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  ASSERT_NE(dy_data, nullptr);
   std::vector<int> dim_dy({2, 65, 65, 12});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(dy_data);
 
-  // runtime part
-  printf("Calculating runtime cost...\n");
-  uint64_t time_avg = 0;
   size_t output_data_size =
     conv_param->output_channel_ * conv_param->kernel_h_ * conv_param->kernel_w_ * conv_param->input_channel_;
 
   size_t input_size;
   std::string input_path = "./test_data/deconv/deconvfp32_input0_d2_g2_s2_2_4_32_32.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_x({2, 32, 32, 4});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({6, 3, 3, 4});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -489,7 +519,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group2Stride2FilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DeConv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
@@ -497,6 +529,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group2Stride2FilterGrad) {
     kernel->Run();
   }
 
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
   int loop_count = 100;
   auto time_start = mindspore::lite::GetTimeUs();
   for (int i = 0; i < loop_count; i++) {
@@ -527,6 +562,8 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group2Stride2FilterGrad) {
 TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group12Stride2FilterGrad) {
   // prepare stage
   auto conv_param = static_cast<ConvParameter *>(malloc(sizeof(ConvParameter)));
+  ASSERT_NE(conv_param, nullptr);
+
   conv_param->input_batch_ = 2;
   conv_param->input_h_ = 32;
   conv_param->input_w_ = 32;
@@ -558,6 +595,7 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group12Stride2FilterGrad) {
   size_t dy_size;
   std::string dy_path = "./test_data/deconv/deconvfp32_dy_d2_g12_s2_2_12_65_65.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &dy_size));
+  ASSERT_NE(dy_data, nullptr);
   std::vector<int> dim_dy({2, 65, 65, 12});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(dy_data);
@@ -571,11 +609,13 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group12Stride2FilterGrad) {
   size_t input_size;
   std::string input_path = "./test_data/deconv/deconvfp32_input0_d2_g12_s2_2_12_32_32.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_x({2, 32, 32, 12});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input_data);
 
   auto dw_data = new float[output_data_size];
+  ASSERT_NE(dw_data, nullptr);
   std::vector<int> dim_dw({1, 3, 3, 12});
   lite::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.set_data(dw_data);
@@ -588,7 +628,9 @@ TEST_F(TestDeConvolutionGradFp32, DeConvFp32Dilation2Group12Stride2FilterGrad) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_DeConv2DGradFilter};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel = creator(inputs, outputs, reinterpret_cast<OpParameter *>(conv_param), &context, desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel->GetWorkspaceSize());
 
   // warm up loop
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/network_test.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/network_test.cc
index 0e1275cbdf..6f1aa1f51e 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/network_test.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/network_test.cc
@@ -90,6 +90,7 @@ TEST_F(NetworkTest, tuning_layer) {
     node->primitive = std::make_unique<schema::PrimitiveT>();
     node->primitive->value.type = schema::PrimitiveType_Activation;
     auto primitive = new schema::ActivationT;
+    ASSERT_NE(primitive, nullptr);
     primitive->type = schema::ActivationType_RELU;
     node->primitive->value.value = primitive;
     node->name = "ReLU";
@@ -102,6 +103,7 @@ TEST_F(NetworkTest, tuning_layer) {
     node->primitive = std::make_unique<schema::PrimitiveT>();
     node->primitive->value.type = schema::PrimitiveType_MatMul;
     auto primitive = new schema::MatMulT;
+    ASSERT_NE(primitive, nullptr);
     primitive->transposeA = false;
     primitive->transposeB = true;
     node->primitive->value.value = primitive;
@@ -115,6 +117,7 @@ TEST_F(NetworkTest, tuning_layer) {
     node->primitive = std::make_unique<schema::PrimitiveT>();
     node->primitive->value.type = schema::PrimitiveType_BiasAdd;
     auto primitive = new schema::BiasAddT;
+    ASSERT_NE(primitive, nullptr);
     primitive->axis.push_back(0);
     node->primitive->value.value = primitive;
     node->name = "BiasAdd";
@@ -127,6 +130,7 @@ TEST_F(NetworkTest, tuning_layer) {
     node->primitive = std::make_unique<schema::PrimitiveT>();
     node->primitive->value.type = schema::PrimitiveType_SoftmaxCrossEntropy;
     auto primitive = new schema::SoftmaxCrossEntropyT;
+    ASSERT_NE(primitive, nullptr);
     primitive->axis.push_back(0);
     node->primitive->value.value = primitive;
     node->name = "SoftmaxCrossEntropy";
@@ -139,6 +143,7 @@ TEST_F(NetworkTest, tuning_layer) {
     node->primitive = std::make_unique<schema::PrimitiveT>();
     node->primitive->value.type = schema::PrimitiveType_BiasGrad;
     auto primitive = new schema::BiasGradT;
+    ASSERT_NE(primitive, nullptr);
     primitive->axis.push_back(0);
     node->primitive->value.value = primitive;
     node->name = "BiasGrad";
@@ -151,6 +156,7 @@ TEST_F(NetworkTest, tuning_layer) {
     node->primitive = std::make_unique<schema::PrimitiveT>();
     node->primitive->value.type = schema::PrimitiveType_MatMul;
     auto primitive = new schema::MatMulT;
+    ASSERT_NE(primitive, nullptr);
     primitive->transposeA = true;
     primitive->transposeB = false;
     node->primitive->value.value = primitive;
@@ -164,6 +170,7 @@ TEST_F(NetworkTest, tuning_layer) {
     node->primitive = std::make_unique<schema::PrimitiveT>();
     node->primitive->value.type = schema::PrimitiveType_ApplyMomentum;
     auto primitive = new schema::ApplyMomentumT;
+    ASSERT_NE(primitive, nullptr);
     node->primitive->value.value = primitive;
     node->name = "Momentum";
     meta_graph->nodes.emplace_back(std::move(node));
@@ -175,6 +182,7 @@ TEST_F(NetworkTest, tuning_layer) {
     node->primitive = std::make_unique<schema::PrimitiveT>();
     node->primitive->value.type = schema::PrimitiveType_ApplyMomentum;
     auto primitive = new schema::ApplyMomentumT;
+    ASSERT_NE(primitive, nullptr);
     node->primitive->value.value = primitive;
     node->name = "Momentum";
     meta_graph->nodes.emplace_back(std::move(node));
@@ -450,9 +458,6 @@ TEST_F(NetworkTest, tuning_layer) {
   std::cout << std::endl;
   error = RelativeOutputError(outData, output_path);
   EXPECT_LT(error, 2e-3);
-
-  delete session;
-  MS_LOG(INFO) << "TuningLayer passed";
 }
 
 int32_t fileIterator(mindspore::session::TrainSession *session, const std::string &path,
@@ -516,6 +521,7 @@ TEST_F(NetworkTest, efficient_net) {
   auto model = lite::TrainModel::Import(buf, net_size);
   delete[] buf;
   auto context = new lite::Context;
+  ASSERT_NE(context, nullptr);
   context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = lite::NO_BIND;
   context->thread_num_ = 1;
 
@@ -533,48 +539,6 @@ TEST_F(NetworkTest, efficient_net) {
   ASSERT_EQ(res, 0);
 }
 
-TEST_F(NetworkTest, lenetnet) {
-  char *buf = nullptr;
-  size_t net_size = 0;
-  std::string net = "./test_data/nets/lenet_train.ms";
-  ReadFile(net.c_str(), &net_size, &buf);
-  auto model = lite::TrainModel::Import(buf, net_size);
-  delete[] buf;
-  auto context = new lite::Context;
-  context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = lite::NO_BIND;
-  context->thread_num_ = 1;
-
-  // check registration
-  mindspore::lite::KernelRegistry *reg = mindspore::lite::KernelRegistry::GetInstance();
-  mindspore::kernel::KernelKey desc1 = {mindspore::kernel::KERNEL_ARCH::kCPU, kNumberTypeFloat32,
-                                        mindspore::schema::PrimitiveType_Conv2D};
-  mindspore::kernel::KernelKey desc2 = {mindspore::kernel::KERNEL_ARCH::kCPU, kNumberTypeFloat32,
-                                        mindspore::schema::PrimitiveType_DepthwiseConv2D};
-  auto regb1 = reg->GetCreator(desc1);
-  auto regb2 = reg->GetCreator(desc2);
-  ASSERT_EQ(regb1 == mindspore::kernel::CpuConvTrainFp32KernelCreator, false);
-
-  auto session = session::TrainSession::CreateSession(context);
-  ASSERT_NE(session, nullptr);
-  auto ret = session->CompileTrainGraph(model);
-  ASSERT_EQ(lite::RET_OK, ret);
-
-  auto rega1 = reg->GetCreator(desc1);
-  auto rega2 = reg->GetCreator(desc2);
-  ASSERT_EQ(regb1, rega1);
-  ASSERT_EQ(regb2, rega2);
-  ASSERT_EQ(rega1 == mindspore::kernel::CpuConvTrainFp32KernelCreator, false);
-  // end of check registration
-
-  session->Eval();
-  std::string in = "./test_data/nets/x_lenet.bin";
-  std::string out = "./test_data/nets/y_lenet.bin";
-  auto res = runNet(session, in, out, "24");
-  delete session;
-  delete context;
-  ASSERT_EQ(res, 0);
-}
-
 TEST_F(NetworkTest, retina_net) {
   char *buf = nullptr;
   size_t net_size = 0;
@@ -585,6 +549,7 @@ TEST_F(NetworkTest, retina_net) {
   auto model = lite::Model::Import(buf, net_size);
   delete[] buf;
   auto context = new lite::Context;
+  ASSERT_NE(context, nullptr);
   context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = lite::NO_BIND;
   context->thread_num_ = 1;
 
@@ -592,7 +557,7 @@ TEST_F(NetworkTest, retina_net) {
   auto session = session::LiteSession::CreateSession(context);
   ASSERT_NE(session, nullptr);
   auto ret = session->CompileGraph(model);
-  ASSERT_EQ(lite::RET_OK, ret);
+  EXPECT_EQ(lite::RET_OK, ret);
   // session->Eval();
 
   std::string in = "./test_data/nets/test1.hwc_normalized_f32";
@@ -619,8 +584,9 @@ TEST_F(NetworkTest, retina_net) {
     final_res |= res;
   }
 
-  ASSERT_EQ(final_res, 0);
+  EXPECT_EQ(final_res, 0);
 
+  delete model;
   delete session;
   delete context;
 }
@@ -635,6 +601,7 @@ TEST_F(NetworkTest, mobileface_net) {
   auto model = lite::Model::Import(buf, net_size);
   delete[] buf;
   auto context = new lite::Context;
+  ASSERT_NE(context, nullptr);
   context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = lite::NO_BIND;
   context->thread_num_ = 1;
 
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/pooling_grad_fp32_tests.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/pooling_grad_fp32_tests.cc
index ae7d061f47..212c5d0c81 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/pooling_grad_fp32_tests.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/pooling_grad_fp32_tests.cc
@@ -60,6 +60,8 @@ void InitPoolingParamFP32(PoolingParameter *pooling_param) {
 TEST_F(TestPoolingGradFp32, AvgPoolingGradFp32) {
   // prepare stage
   auto pooling_param = static_cast<PoolingParameter *>(malloc(sizeof(PoolingParameter)));
+  ASSERT_NE(pooling_param, nullptr);
+
   InitPoolingParamFP32(pooling_param);
   pooling_param->output_channel_ = 3;
   pooling_param->pool_mode_ = PoolMode_AvgPool;
@@ -73,8 +75,10 @@ TEST_F(TestPoolingGradFp32, AvgPoolingGradFp32) {
   size_t input_size;
   std::string input_path = "./test_data/pooling/avgpoolgradfp32_1_dy_1_28_28_3.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
 
   auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
   // warm up loop
   for (int i = 0; i < 3; i++) {
     AvgPoolingGrad(input_data, output_data, pooling_param, 1);
@@ -108,6 +112,8 @@ TEST_F(TestPoolingGradFp32, AvgPoolingGradFp32) {
 TEST_F(TestPoolingGradFp32, AvgPoolingKernelGradFp32) {
   // prepare stage
   auto pooling_param = static_cast<PoolingParameter *>(malloc(sizeof(PoolingParameter)));
+  ASSERT_NE(pooling_param, nullptr);
+
   InitPoolingParamFP32(pooling_param);
   pooling_param->output_channel_ = 3;
   pooling_param->pool_mode_ = PoolMode_AvgPool;
@@ -121,12 +127,14 @@ TEST_F(TestPoolingGradFp32, AvgPoolingKernelGradFp32) {
   size_t input_size;
   std::string input_path = "./test_data/pooling/avgpoolgradfp32_1_dy_1_28_28_3.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_dy({1, 28, 28, 3});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(input_data);
 
   std::string input1_path = "./test_data/pooling/avgpoolgradfp32_1_x_1_28_28_3.bin";
   auto input1_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input1_path.c_str(), &input_size));
+  ASSERT_NE(input1_data, nullptr);
   std::vector<int> dim_x({1, 28, 28, 3});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input1_data);
@@ -134,6 +142,7 @@ TEST_F(TestPoolingGradFp32, AvgPoolingKernelGradFp32) {
   std::vector<lite::Tensor *> inputs = {&dy_tensor, &x_tensor};
 
   auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
   std::vector<int> dim_dx({1, 28, 28, 3});
   lite::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
   dx_tensor.set_data(output_data);
@@ -145,7 +154,9 @@ TEST_F(TestPoolingGradFp32, AvgPoolingKernelGradFp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(pooling_param), &context, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
 
   kernel_obj->Run();
 
@@ -172,8 +183,9 @@ TEST_F(TestPoolingGradFp32, AvgPoolingKernelGradFp32) {
 TEST_F(TestPoolingGradFp32, AvgPoolingBatchGradFp32) {
   // prepare stage
   auto pooling_param = static_cast<PoolingParameter *>(malloc(sizeof(PoolingParameter)));
-  InitPoolingParamFP32(pooling_param);
+  ASSERT_NE(pooling_param, nullptr);
 
+  InitPoolingParamFP32(pooling_param);
   pooling_param->output_channel_ = 3;
   pooling_param->input_batch_ = 3;
   pooling_param->output_batch_ = 3;
@@ -185,12 +197,14 @@ TEST_F(TestPoolingGradFp32, AvgPoolingBatchGradFp32) {
   size_t input_size;
   std::string input_path = "./test_data/pooling/avgpoolgradfp32_1_dy_3_28_28_3.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_dy({3, 28, 28, 3});
   lite::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
   dy_tensor.set_data(input_data);
 
   std::string input1_path = "./test_data/pooling/avgpoolgradfp32_1_x_3_28_28_3.bin";
   auto input1_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input1_path.c_str(), &input_size));
+  ASSERT_NE(input1_data, nullptr);
   std::vector<int> dim_x({3, 28, 28, 3});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(input1_data);
@@ -209,7 +223,9 @@ TEST_F(TestPoolingGradFp32, AvgPoolingBatchGradFp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(pooling_param), &context, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
 
   kernel_obj->Run();
 
@@ -236,6 +252,8 @@ TEST_F(TestPoolingGradFp32, AvgPoolGradStride2Fp32) {
   // prepare stage
   // input size will be equal to the original size of x, output size will be the output size as in forward
   auto pool = static_cast<PoolingParameter *>(malloc(sizeof(PoolingParameter)));
+  ASSERT_NE(pool, nullptr);
+
   InitPoolingParamFP32(pool);
   pool->output_channel_ = 3;
   pool->pool_mode_ = PoolMode_AvgPool;
@@ -250,12 +268,14 @@ TEST_F(TestPoolingGradFp32, AvgPoolGradStride2Fp32) {
 
   auto x_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/avgpoolgradfp32_s2_x_3_28_28_3.bin", &input_size));
+  ASSERT_NE(x_data, nullptr);
   std::vector<int> dim_x({pool->output_batch_, pool->input_h_, pool->input_w_, pool->input_channel_});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(x_data);
 
   auto yt_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/avgpoolgradfp32_s2_dy_3_28_28_3.bin", &input_size));
+  ASSERT_NE(yt_data, nullptr);
   std::vector<int> dim_y({pool->output_batch_, pool->output_h_, pool->output_w_, pool->output_channel_});
   lite::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
   yt_tensor.set_data(yt_data);
@@ -271,7 +291,9 @@ TEST_F(TestPoolingGradFp32, AvgPoolGradStride2Fp32) {
 
   kernel::KernelKey pool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
   auto pool_creator = lite::KernelRegistry::GetInstance()->GetCreator(pool_desc);
+  ASSERT_NE(pool_creator, nullptr);
   auto kernel = pool_creator(inputs, outputs, reinterpret_cast<OpParameter *>(pool), &context, pool_desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
 
   kernel->Init();
 
@@ -295,6 +317,8 @@ TEST_F(TestPoolingGradFp32, AvgPoolGradStride3Fp32) {
   // prepare stage
   // input size will be equal to the original size of x, output size will be the output size as in forward
   auto pool = static_cast<PoolingParameter *>(malloc(sizeof(PoolingParameter)));
+  ASSERT_NE(pool, nullptr);
+
   InitPoolingParamFP32(pool);
   pool->output_channel_ = 3;
   pool->pool_mode_ = PoolMode_AvgPool;
@@ -309,12 +333,14 @@ TEST_F(TestPoolingGradFp32, AvgPoolGradStride3Fp32) {
 
   auto x_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/avgpoolgradfp32_s3_x_3_28_28_3.bin", &input_size));
+  ASSERT_NE(x_data, nullptr);
   std::vector<int> dim_x({pool->output_batch_, pool->input_h_, pool->input_w_, pool->input_channel_});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(x_data);
 
   auto yt_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/avgpoolgradfp32_s3_dy_3_28_28_3.bin", &input_size));
+  ASSERT_NE(yt_data, nullptr);
   std::vector<int> dim_y({pool->output_batch_, pool->output_h_, pool->output_w_, pool->output_channel_});
   lite::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
   yt_tensor.set_data(yt_data);
@@ -332,7 +358,9 @@ TEST_F(TestPoolingGradFp32, AvgPoolGradStride3Fp32) {
 
   kernel::KernelKey pool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
   auto pool_creator = lite::KernelRegistry::GetInstance()->GetCreator(pool_desc);
+  ASSERT_NE(pool_creator, nullptr);
   auto kernel = pool_creator(inputs, outputs, reinterpret_cast<OpParameter *>(pool), &context, pool_desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
 
   kernel->Init();
 
@@ -356,6 +384,8 @@ TEST_F(TestPoolingGradFp32, AvgPoolGradStride3Fp32) {
 TEST_F(TestPoolingGradFp32, MaxPoolingGradFp32) {
   // prepare stage
   auto pooling_param = static_cast<PoolingParameter *>(malloc(sizeof(PoolingParameter)));
+  ASSERT_NE(pooling_param, nullptr);
+
   InitPoolingParamFP32(pooling_param);
   pooling_param->output_channel_ = 3;
   pooling_param->pool_mode_ = PoolMode_MaxPool;
@@ -368,14 +398,18 @@ TEST_F(TestPoolingGradFp32, MaxPoolingGradFp32) {
   size_t input_size;
   std::string i_path = "./test_data/pooling/maxpoolgradfp32_1_x_1_28_28_3.bin";
   auto in_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(i_path.c_str(), &input_size));
+  ASSERT_NE(in_data, nullptr);
 
   std::string dy_path = "./test_data/pooling/maxpoolgradfp32_1_dy_1_28_28_3.bin";
   auto dy_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dy_path.c_str(), &input_size));
+  ASSERT_NE(dy_data, nullptr);
 
   std::string dx_path = "./test_data/pooling/maxpoolgradfp32_1_dx_1_28_28_3.bin";
   auto dx_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(dx_path.c_str(), &input_size));
+  ASSERT_NE(dx_data, nullptr);
 
   auto output_data = new float[output_data_size];
+  ASSERT_NE(output_data, nullptr);
   // warm up loop
   for (int i = 0; i < 3; i++) {
     MaxPoolingGrad(in_data, dx_data, dy_data, output_data, pooling_param, 1);
@@ -412,6 +446,8 @@ TEST_F(TestPoolingGradFp32, MaxPoolGradBatchFp32) {
   // prepare stage
   // input size will be equal to the original size of x, output size will be the output size as in forward
   auto maxpool = static_cast<PoolingParameter *>(malloc(sizeof(PoolingParameter)));
+  ASSERT_NE(maxpool, nullptr);
+
   InitPoolingParamFP32(maxpool);
   maxpool->output_channel_ = 3;
   maxpool->pool_mode_ = PoolMode_MaxPool;
@@ -422,18 +458,21 @@ TEST_F(TestPoolingGradFp32, MaxPoolGradBatchFp32) {
 
   auto x_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_1_x_3_28_28_3.bin", &input_size));
+  ASSERT_NE(x_data, nullptr);
   std::vector<int> dim_x({3, 28, 28, 3});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(x_data);
 
   auto y_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_1_dx_3_28_28_3.bin", &input_size));
+  ASSERT_NE(y_data, nullptr);
   std::vector<int> dim_y({3, 28, 28, 3});
   lite::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
   y_tensor.set_data(y_data);
 
   auto yt_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_1_dy_3_28_28_3.bin", &input_size));
+  ASSERT_NE(yt_data, nullptr);
   lite::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
   yt_tensor.set_data(yt_data);
 
@@ -449,8 +488,10 @@ TEST_F(TestPoolingGradFp32, MaxPoolGradBatchFp32) {
 
   kernel::KernelKey maxpool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
   auto maxpool_creator = lite::KernelRegistry::GetInstance()->GetCreator(maxpool_desc);
+  ASSERT_NE(maxpool_creator, nullptr);
   auto kernel = maxpool_creator(maxpool_inputs, maxpool_outputs, reinterpret_cast<OpParameter *>(maxpool), &context,
                                 maxpool_desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
 
   kernel->Init();
 
@@ -477,6 +518,8 @@ TEST_F(TestPoolingGradFp32, MaxPoolGradStride2Fp32) {
   // prepare stage
   // input size will be equal to the original size of x, output size will be the output size as in forward
   auto maxpool = static_cast<PoolingParameter *>(malloc(sizeof(PoolingParameter)));
+  ASSERT_NE(maxpool, nullptr);
+
   InitPoolingParamFP32(maxpool);
   maxpool->output_channel_ = 3;
   maxpool->input_channel_ = 3;
@@ -492,18 +535,21 @@ TEST_F(TestPoolingGradFp32, MaxPoolGradStride2Fp32) {
 
   auto x_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s2_x_3_28_28_3.bin", &input_size));
+  ASSERT_NE(x_data, nullptr);
   std::vector<int> dim_x({maxpool->output_batch_, maxpool->input_h_, maxpool->input_w_, maxpool->input_channel_});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(x_data);
 
   auto y_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s2_dx_3_28_28_3.bin", &input_size));
+  ASSERT_NE(y_data, nullptr);
   std::vector<int> dim_y({maxpool->output_batch_, maxpool->output_h_, maxpool->output_w_, maxpool->output_channel_});
   lite::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
   y_tensor.set_data(y_data);
 
   auto yt_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s2_dy_3_28_28_3.bin", &input_size));
+  ASSERT_NE(yt_data, nullptr);
   lite::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
   yt_tensor.set_data(yt_data);
 
@@ -520,8 +566,10 @@ TEST_F(TestPoolingGradFp32, MaxPoolGradStride2Fp32) {
 
   kernel::KernelKey maxpool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
   auto maxpool_creator = lite::KernelRegistry::GetInstance()->GetCreator(maxpool_desc);
+  ASSERT_NE(maxpool_creator, nullptr);
   auto kernel = maxpool_creator(maxpool_inputs, maxpool_outputs, reinterpret_cast<OpParameter *>(maxpool), &context,
                                 maxpool_desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
 
   kernel->Init();
 
@@ -548,6 +596,8 @@ TEST_F(TestPoolingGradFp32, MaxPoolGradStride3Fp32) {
   // prepare stage
   // input size will be equal to the original size of x, output size will be the output size as in forward
   auto maxpool = static_cast<PoolingParameter *>(malloc(sizeof(PoolingParameter)));
+  ASSERT_NE(maxpool, nullptr);
+
   InitPoolingParamFP32(maxpool);
   maxpool->output_channel_ = 3;
   maxpool->input_channel_ = 3;
@@ -563,18 +613,21 @@ TEST_F(TestPoolingGradFp32, MaxPoolGradStride3Fp32) {
 
   auto x_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s3_x_3_28_28_3.bin", &input_size));
+  ASSERT_NE(x_data, nullptr);
   std::vector<int> dim_x({maxpool->output_batch_, maxpool->input_h_, maxpool->input_w_, maxpool->input_channel_});
   lite::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
   x_tensor.set_data(x_data);
 
   auto y_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s3_dx_3_28_28_3.bin", &input_size));
+  ASSERT_NE(y_data, nullptr);
   std::vector<int> dim_y({maxpool->output_batch_, maxpool->output_h_, maxpool->output_w_, maxpool->output_channel_});
   lite::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
   y_tensor.set_data(y_data);
 
   auto yt_data = reinterpret_cast<float *>(
     mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s3_dy_3_28_28_3.bin", &input_size));
+  ASSERT_NE(yt_data, nullptr);
   lite::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
   yt_tensor.set_data(yt_data);
 
@@ -591,11 +644,12 @@ TEST_F(TestPoolingGradFp32, MaxPoolGradStride3Fp32) {
 
   kernel::KernelKey maxpool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
   auto maxpool_creator = lite::KernelRegistry::GetInstance()->GetCreator(maxpool_desc);
+  ASSERT_NE(maxpool_creator, nullptr);
   auto kernel = maxpool_creator(maxpool_inputs, maxpool_outputs, reinterpret_cast<OpParameter *>(maxpool), &context,
                                 maxpool_desc, nullptr);
+  ASSERT_NE(kernel, nullptr);
 
   kernel->Init();
-
   kernel->Run();
 
   std::string output_path = "./test_data/pooling/maxpoolgradfp32_s3_xgrad_3_28_28_3.bin";
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/softmax_crossentropy_fp32_tests.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/softmax_crossentropy_fp32_tests.cc
index 9988320431..91de1005d4 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/softmax_crossentropy_fp32_tests.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/softmax_crossentropy_fp32_tests.cc
@@ -31,17 +31,21 @@ class TestSoftmaxCrossEntropyFp32 : public mindspore::CommonTest {
 TEST_F(TestSoftmaxCrossEntropyFp32, SoftmaxCrossEntropyFp32) {
   // prepare stage
   auto sce_param = reinterpret_cast<SoftmaxCrossEntropyParameter *>(malloc(sizeof(SoftmaxCrossEntropyParameter)));
+  ASSERT_NE(sce_param, nullptr);
   size_t input_size;
 
   std::string input_path = "./test_data/operators/sce_fp32_1_y_6_4.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
   std::vector<int> dim_y({6, 4});
   lite::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
   y_tensor.set_data(input_data);
 
   std::string label_path = "./test_data/operators/sce_fp32_1_l_6.bin";
   auto ll_labels = reinterpret_cast<int64_t *>(mindspore::lite::ReadFile(label_path.c_str(), &input_size));
+  ASSERT_NE(ll_labels, nullptr);
   auto labels = new float[6 * 4];
+  ASSERT_NE(labels, nullptr);
   std::fill(labels, labels + 6 * 4, 0.f);
   for (int i = 0; i < 6; i++) labels[i * 4 + ll_labels[i]] = 1.0;
 
@@ -52,10 +56,12 @@ TEST_F(TestSoftmaxCrossEntropyFp32, SoftmaxCrossEntropyFp32) {
   std::vector<lite::Tensor *> inputs = {&y_tensor, &l_tensor};
 
   auto loss = new float[1];
+  ASSERT_NE(loss, nullptr);
   std::vector<int> dim_dw({1});
   lite::Tensor loss_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   loss_tensor.set_data(loss);
   auto grad = new float[24];
+  ASSERT_NE(grad, nullptr);
   lite::Tensor grad_tensor(TypeId::kNumberTypeFloat32, dim_y);
   grad_tensor.set_data(grad);
   std::vector<lite::Tensor *> outputs = {&loss_tensor, &grad_tensor};
@@ -66,7 +72,9 @@ TEST_F(TestSoftmaxCrossEntropyFp32, SoftmaxCrossEntropyFp32) {
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_SoftmaxCrossEntropy};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  ASSERT_NE(creator, nullptr);
   auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(sce_param), &context, desc, nullptr);
+  ASSERT_NE(kernel_obj, nullptr);
   mindspore::kernel::LiteKernel::AllocWorkspace(kernel_obj->GetWorkspaceSize());
   kernel_obj->Run();
 
@@ -78,16 +86,20 @@ TEST_F(TestSoftmaxCrossEntropyFp32, SoftmaxCrossEntropyFp32) {
   std::string output_path = "./test_data/operators/sce_fp32_1_loss_1.bin";
   CompareOutput(loss, 1, output_path);
 
-  ((mindspore::kernel::SparseSoftmaxCrossEntropyWithLogitsCPUKernel *)kernel_obj)->train();
+  ((mindspore::kernel::SparseSoftmaxCrossEntropyWithLogitsCPUKernel *)kernel_obj)->Train();
   kernel_obj->Run();
-
+  // normalize by batch size the result
+  for (int i = 0; i < 24; i++) {
+    grad[i] /= 6;
+  }
   printf("==================output data=================\n");
   for (int i = 0; i < 12; i++) {
     std::cout << grad[i] << " ,";
   }
   std::cout << std::endl;
   std::string grad_path = "./test_data/operators/sce_fp32_1_dy_6_4.bin";
-  CompareOutput(grad, 24, grad_path);
+  auto res = CompareRelativeOutput(grad, grad_path);
+  EXPECT_EQ(res, 0);
 
   delete[] ll_labels;
   delete[] labels;
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/softmax_grad_fp32_tests.cc b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/softmax_grad_fp32_tests.cc
index 6c61cd1e7f..3300a6974a 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/softmax_grad_fp32_tests.cc
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/softmax_grad_fp32_tests.cc
@@ -55,6 +55,7 @@ void InitSoftMaxParam(SoftmaxParameter *softmax_param, int axis, int n, int c, i
 
 TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis0) {
   auto softmax_param = new SoftmaxParameter();
+  ASSERT_NE(softmax_param, nullptr);
   // set parameters
   InitSoftMaxParam(softmax_param, 0);
 
@@ -64,21 +65,23 @@ TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis0) {
     inner_size *= softmax_param->input_shape_[i];
   }
   float *sum_data = new (std::nothrow) float[inner_size];
+  ASSERT_NE(sum_data, nullptr);
   float *sum_mul = new (std::nothrow) float[inner_size * softmax_param->input_shape_[softmax_param->axis_]];
+  ASSERT_NE(sum_mul, nullptr);
   std::vector<int> shape = {1, 9, 11, 12};
   size_t input_size;
   std::string input_path = "./test_data/softmax/softmaxgrad_yinput.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
-
+  ASSERT_NE(input_data, nullptr);
   std::string yt_path = "./test_data/softmax/softmaxgrad_yt_input.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
-
+  ASSERT_NE(yt_data, nullptr);
   // runtime part
   printf("Calculating runtime cost...\n");
   uint64_t time_avg = 0;
 
   auto out_data = new float[softmax_param->element_size_];
-
+  ASSERT_NE(out_data, nullptr);
   // warm up loop
   for (int i = 0; i < 3; i++) {
     SoftmaxGrad(input_data, yt_data, out_data, sum_data, sum_mul, softmax_param);
@@ -112,6 +115,7 @@ TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis0) {
 
 TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis1) {
   auto softmax_param = new SoftmaxParameter();
+  ASSERT_NE(softmax_param, nullptr);
   // set parameters
   InitSoftMaxParam(softmax_param, 1);
 
@@ -121,21 +125,26 @@ TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis1) {
     inner_size *= softmax_param->input_shape_[i];
   }
   float *sum_data = new (std::nothrow) float[inner_size];
+  ASSERT_NE(sum_data, nullptr);
   float *sum_mul = new (std::nothrow) float[inner_size * softmax_param->input_shape_[softmax_param->axis_]];
+  ASSERT_NE(sum_mul, nullptr);
 
   std::vector<int> shape = {1, 9, 11, 12};
   size_t input_size;
   std::string input_path = "./test_data/softmax/softmaxgrad_1_yinput.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
 
   std::string yt_path = "./test_data/softmax/softmaxgrad_1_yt_input.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
 
   // runtime part
   printf("Calculating runtime cost...\n");
   uint64_t time_avg = 0;
 
   auto out_data = new float[softmax_param->element_size_];
+  ASSERT_NE(out_data, nullptr);
 
   // warm up loop
   for (int i = 0; i < 3; i++) {
@@ -171,6 +180,7 @@ TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis1) {
 
 TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis2) {
   auto softmax_param = new SoftmaxParameter();
+  ASSERT_NE(softmax_param, nullptr);
   // set parameters
   InitSoftMaxParam(softmax_param, 2);
 
@@ -180,21 +190,26 @@ TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis2) {
     inner_size *= softmax_param->input_shape_[i];
   }
   float *sum_data = new (std::nothrow) float[inner_size];
+  ASSERT_NE(sum_data, nullptr);
   float *sum_mul = new (std::nothrow) float[inner_size * softmax_param->input_shape_[softmax_param->axis_]];
+  ASSERT_NE(sum_mul, nullptr);
 
   std::vector<int> shape = {1, 9, 11, 12};
   size_t input_size;
   std::string input_path = "./test_data/softmax/softmaxgrad_2_yinput.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
 
   std::string yt_path = "./test_data/softmax/softmaxgrad_2_yt_input.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
 
   // runtime part
   printf("Calculating runtime cost...\n");
   uint64_t time_avg = 0;
 
   auto out_data = new float[softmax_param->element_size_];
+  ASSERT_NE(out_data, nullptr);
 
   // warm up loop
   for (int i = 0; i < 3; i++) {
@@ -230,6 +245,7 @@ TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis2) {
 
 TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis3) {
   auto softmax_param = new SoftmaxParameter();
+  ASSERT_NE(softmax_param, nullptr);
   // set parameters
   InitSoftMaxParam(softmax_param, 3);
 
@@ -239,21 +255,25 @@ TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis3) {
     inner_size *= softmax_param->input_shape_[i];
   }
   float *sum_data = new (std::nothrow) float[inner_size];
+  ASSERT_NE(sum_data, nullptr);
   float *sum_mul = new (std::nothrow) float[inner_size * softmax_param->input_shape_[softmax_param->axis_]];
+  ASSERT_NE(sum_mul, nullptr);
 
   std::vector<int> shape = {1, 9, 11, 12};
   size_t input_size;
   std::string input_path = "./test_data/softmax/softmaxgrad_3_yinput.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
-
+  ASSERT_NE(input_data, nullptr);
   std::string yt_path = "./test_data/softmax/softmaxgrad_3_yt_input.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
+  ASSERT_NE(yt_data, nullptr);
 
   // runtime part
   printf("Calculating runtime cost...\n");
   uint64_t time_avg = 0;
 
   auto out_data = new float[softmax_param->element_size_];
+  ASSERT_NE(out_data, nullptr);
 
   // warm up loop
   for (int i = 0; i < 3; i++) {
@@ -289,6 +309,8 @@ TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxis3) {
 
 TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxisMinus1) {
   auto softmax_param = new SoftmaxParameter();
+  ASSERT_NE(softmax_param, nullptr);
+
   // set parameters
   InitSoftMaxParam(softmax_param, -1);
 
@@ -298,21 +320,25 @@ TEST_F(TestSoftmaxGradFp32, SoftmaxGradAxisMinus1) {
     inner_size *= softmax_param->input_shape_[i];
   }
   float *sum_data = new (std::nothrow) float[inner_size];
+  ASSERT_NE(sum_data, nullptr);
   float *sum_mul = new (std::nothrow) float[inner_size * softmax_param->input_shape_[softmax_param->axis_]];
+  ASSERT_NE(sum_mul, nullptr);
 
   std::vector<int> shape = {1, 9, 11, 12};
   size_t input_size;
   std::string input_path = "./test_data/softmax/softmaxgrad_-1_yinput.bin";
   auto input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input_path.c_str(), &input_size));
+  ASSERT_NE(input_data, nullptr);
 
   std::string yt_path = "./test_data/softmax/softmaxgrad_-1_yt_input.bin";
   auto yt_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(yt_path.c_str(), &input_size));
-
+  ASSERT_NE(yt_data, nullptr);
   // runtime part
   printf("Calculating runtime cost...\n");
   uint64_t time_avg = 0;
 
   auto out_data = new float[softmax_param->element_size_];
+  ASSERT_NE(out_data, nullptr);
 
   // warm up loop
   for (int i = 0; i < 3; i++) {
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsig_out_50.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsig_out_50.bin
new file mode 100644
index 0000000000..2e65e6d409
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsig_out_50.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsig_x_50.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsig_x_50.bin
new file mode 100644
index 0000000000..c01986eae4
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsig_x_50.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsig_yt_50.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsig_yt_50.bin
new file mode 100644
index 0000000000..7502f7121b
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hsig_yt_50.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_out_50.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_out_50.bin
index 6ff3dd84c9..71785be6c3 100644
Binary files a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_out_50.bin and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_out_50.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_x_50.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_x_50.bin
index b9341ce212..eaa2049f01 100644
Binary files a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_x_50.bin and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_x_50.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_yt_50.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_yt_50.bin
index b25a5c7787..6e0ffb821d 100644
Binary files a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_yt_50.bin and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/activationGrad/hswish_yt_50.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/dy_2_4_5_3.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/dy_2_4_5_3.bin
index 2ccfc68d73..e02252dd6b 100644
Binary files a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/dy_2_4_5_3.bin and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/dy_2_4_5_3.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/input_x_2_4_5_3.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/input_x_2_4_5_3.bin
index a194c59c0a..9754647732 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/input_x_2_4_5_3.bin
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/input_x_2_4_5_3.bin
@@ -1,2 +1,2 @@
-V_�?�K��ϧ���࿅��>J?/=�"m?L�u���j@!U$?f=�?e�[?W�ھ�� �m�?�	e�O��?}4����?�B<��K�����e����ڽ���p~|>�?���7����E	:?JͿ̬>�	? �~?ϫ��N1�?>HV��|ʾ���=�{��I�U?x���v��W>��[�$��?����]��4Bu��
���4@��
��+?��z>u�B?��=|e >M��>��>���?}�0?��������>����	�
@<�?�v�?v�Z?z�ſ@����.ο�8B?��o�
����Ծq"m�n��?�k>=�"�>:����@��<>+R�
-b��6.�?�i?��v?`j6�R~]?�JU�6�s�G��?M���% �?�h�>�ȿ���	���G�½?��>�ӓ��'6?��@����2��/V�K5�T�>X]?�[�?v_ؿ�����j?p��?�\l?�.l=���b�?
\ No newline at end of file
+�CE�€��&s���ľ�pѿ ��fe'���?�
+��
ϫ?u�P>�(�?��/>��¿[�Q���R>�Y�;��o�v��1�?rUt?jb?�WM>������,�;4%�\4�|�
��_k�!Ȉ��T���ھ
f�h^�>��&?�%V��}�';��#�f>ND�m�\>�X�>�1�?��[>c�@‹B@3	��H�?H��K�6?B��{{�� �z?*>0�?�1*��j�?	�q?/!"�I5�>HA?Pc�<��O@P?+ڳ�v�>���?������(@n�?��T�=�j��?�f:?MJC>��H��D�뼵=���?����Ɔ�!J���\�U�?
U��F����l����>��%�����?���D�޿c�K>�?YQd��i0>�y�?�_�?b�*��� ?f<��d�Ͼ6���?"��/����.����q>�23?��5>iU�>J^(?�8�=����~\��K>e��?��=KK~=
\ No newline at end of file
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dbias_3.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dbias_3.bin
index ac2915f01f..6279d91483 100644
Binary files a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dbias_3.bin and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dbias_3.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dscale_3.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dscale_3.bin
index cf8cfa4b05..4d31c84f93 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dscale_3.bin
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dscale_3.bin
@@ -1 +1 @@
-���B���B���B
\ No newline at end of file
+r��B���B]��B
\ No newline at end of file
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dx_2_4_5_3.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dx_2_4_5_3.bin
index daf908d964..110652f810 100644
Binary files a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dx_2_4_5_3.bin and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dx_2_4_5_3.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/running_mean_3.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/running_mean_3.bin
index 7cf94f5cd0..4cc4654b7e 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/running_mean_3.bin
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/running_mean_3.bin
@@ -1 +1 @@
-�-�<���<���=
\ No newline at end of file
+8�<"�>���
\ No newline at end of file
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/running_var_3.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/running_var_3.bin
index c124f424e8..10a5cce277 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/running_var_3.bin
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/running_var_3.bin
@@ -1 +1 @@
-c�?�	�?;�[?
\ No newline at end of file
+���?`�?��?
\ No newline at end of file
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_mean_3.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_mean_3.bin
index a3cd501da7..3351ce87e0 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_mean_3.bin
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_mean_3.bin
@@ -1 +1 @@
-�:c2;@e<
\ No newline at end of file
+�:>������)�
\ No newline at end of file
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_var_3.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_var_3.bin
index aa2b0f6ede..74e5b4d0f5 100644
--- a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_var_3.bin
+++ b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_var_3.bin
@@ -1 +1 @@
-=}�?�M�?Z|?
\ No newline at end of file
+vr�?90�?��?
\ No newline at end of file
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/fc_b_grad.f32 b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/fc_b_grad.f32
new file mode 100644
index 0000000000..098b242555
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/fc_b_grad.f32 differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/fc_yt.f32 b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/fc_yt.f32
new file mode 100644
index 0000000000..58779f9fc0
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/fc_yt.f32 differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x1_grad_maximum.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x1_grad_maximum.bin
new file mode 100644
index 0000000000..0f7481b380
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x1_grad_maximum.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x1_maximum.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x1_maximum.bin
new file mode 100644
index 0000000000..b0ec219f21
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x1_maximum.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x2_grad_maximum.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x2_grad_maximum.bin
new file mode 100644
index 0000000000..b2caa77eb7
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x2_grad_maximum.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x2_maximum.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x2_maximum.bin
new file mode 100644
index 0000000000..8337142839
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/x2_maximum.bin differ
diff --git a/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/yt_maximum.bin b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/yt_maximum.bin
new file mode 100644
index 0000000000..c064a92962
Binary files /dev/null and b/mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/operators/yt_maximum.bin differ
diff --git a/mindspore/lite/tools/anf_exporter/anf_exporter.cc b/mindspore/lite/tools/anf_exporter/anf_exporter.cc
index d28eabbc09..61799f64b5 100644
--- a/mindspore/lite/tools/anf_exporter/anf_exporter.cc
+++ b/mindspore/lite/tools/anf_exporter/anf_exporter.cc
@@ -297,6 +297,7 @@ int AnfExporter::ConvertInputCNode(const std::shared_ptr<AnfNode> input_anode, s
 #endif
   } else {
     auto inputs = input_cnode->inputs();
+
     if (inputs.size() != 3) {
       MS_LOG(ERROR) << "TupleGetItem should have 3 inputs, got " << inputs.size();
       return RET_ERROR;
@@ -440,43 +441,8 @@ int AnfExporter::ConvertInputValueNode(std::shared_ptr<AnfNode> input_anode,
     output_cnode->inputIndex.emplace_back(meta_graphT->allTensors.size());
     meta_graphT->allTensors.emplace_back(std::move(paramTensor));
   } else if (value->isa<mindspore::ValueSequeue>()) {
-#ifndef SUPPORT_TRAIN
     MS_LOG(DEBUG) << "Value type is ValueSequence.";
     return RET_OK;
-#else
-    auto valueAbstract = valueNode->abstract();
-    auto abstractSequnce = utils::cast<abstract::AbstractSequeuePtr>(valueAbstract);
-    if (abstractSequnce->isa<abstract::AbstractTuple>()) {
-      auto abstractTuple = utils::cast<abstract::AbstractTuplePtr>(valueAbstract);
-      auto x_shape_data = abstractTuple->elements();
-      std::vector<int32_t> shape;
-      for (std::size_t i = 0; i < abstractTuple->size(); ++i) {
-        auto value_track = x_shape_data[i]->GetValueTrack();
-        if (value_track == nullptr) {
-          ReturnCode::GetSingleReturnCode()->UpdateReturnCode(RET_NULL_PTR);
-          return RET_NULL_PTR;
-        }
-        if (value_track->isa<Int32Imm>()) {
-          shape.push_back((GetValue<int>(value_track)));
-        } else {
-          MS_LOG(ERROR) << "Value type is ValueSequence is not integer, it is " << value_track->ToString() << ".";
-        }
-      }
-      if (shape.size()) {
-        auto typePtr = abstractTuple->elements()[0]->GetTypeTrack();  // abstractTuple->GetTypeTrack();
-        paramTensor->dataType = typePtr->type_id();
-        paramTensor->dims = {static_cast<int32_t>(shape.size())};
-        paramTensor->nodeType = schema::NodeType_ValueNode;
-        paramTensor->data.resize(shape.size() * sizeof(int));
-        memcpy(paramTensor->data.data(), shape.data(), shape.size() * sizeof(int));
-        node_id_map_[valueNode->fullname_with_scope()] = meta_graphT->allTensors.size();
-        output_cnode->inputIndex.emplace_back(meta_graphT->allTensors.size());
-        meta_graphT->allTensors.emplace_back(std::move(paramTensor));
-      }
-    } else {
-      MS_LOG(ERROR) << "Value type is ValueSequence not supported - " << valueAbstract->type_name() << ".";
-    }
-#endif
   } else if (value->isa<mindspore::BoolImm>()) {
     auto valueAbstract = valueNode->abstract();
     auto abstractScalar = utils::cast<abstract::AbstractScalarPtr>(valueAbstract);
diff --git a/mindspore/lite/tools/common/node_util.cc b/mindspore/lite/tools/common/node_util.cc
index dcc082ef0f..273f164f43 100644
--- a/mindspore/lite/tools/common/node_util.cc
+++ b/mindspore/lite/tools/common/node_util.cc
@@ -52,15 +52,10 @@ static const std::vector<schema::PrimitiveType> nhwcOpList = {
   schema::PrimitiveType_SpaceToDepth,
   schema::PrimitiveType_DepthToSpace};
 
-static const std::vector<schema::PrimitiveType> nhwcOpDualInputList = {
-#ifdef SUPPORT_TRAIN
-  schema::PrimitiveType_Conv2DGradFilter, schema::PrimitiveType_BNGrad
-#endif
-};
-
 static const std::vector<schema::PrimitiveType> nhwcOpAllInputList = {
 #ifdef SUPPORT_TRAIN
-  schema::PrimitiveType_PoolingGrad, schema::PrimitiveType_ActivationGrad
+  schema::PrimitiveType_PoolingGrad, schema::PrimitiveType_ActivationGrad, schema::PrimitiveType_Conv2DGradFilter,
+  schema::PrimitiveType_BNGrad
 #endif
 };
 
@@ -156,8 +151,6 @@ std::vector<schema::PrimitiveType> Getfp32FullOpList() { return fp32FullOpList;
 
 std::vector<schema::PrimitiveType> GetNhwcOpList() { return nhwcOpList; }
 
-std::vector<schema::PrimitiveType> GetNhwcDualInputOpList() { return nhwcOpDualInputList; }
-
 std::vector<schema::PrimitiveType> GetNhwcAllInputOpList() { return nhwcOpAllInputList; }
 
 std::vector<schema::PrimitiveType> GetUint8NhwcOpList() { return int8NeedNhwcOpList; }
diff --git a/mindspore/lite/tools/common/node_util.h b/mindspore/lite/tools/common/node_util.h
index 51d85ab1a5..d78ea34104 100644
--- a/mindspore/lite/tools/common/node_util.h
+++ b/mindspore/lite/tools/common/node_util.h
@@ -48,8 +48,6 @@ std::vector<schema::PrimitiveType> GetInsertOpList();
 
 std::vector<schema::PrimitiveType> GetNhwcOpList();
 
-std::vector<schema::PrimitiveType> GetNhwcDualInputOpList();
-
 std::vector<schema::PrimitiveType> GetNhwcAllInputOpList();
 
 std::vector<schema::PrimitiveType> Getfp32FullOpList();
diff --git a/mindspore/lite/tools/converter/converter_flags.cc b/mindspore/lite/tools/converter/converter_flags.cc
index ab7a44d54a..cd3d1cc839 100644
--- a/mindspore/lite/tools/converter/converter_flags.cc
+++ b/mindspore/lite/tools/converter/converter_flags.cc
@@ -154,6 +154,25 @@ int Flags::Init(int argc, const char **argv) {
     std::cerr << "INPUT ILLEGAL: trainModel must be true|false ";
     return RET_INPUT_PARAM_INVALID;
   }
+
+  if (this->trainModel == true) {
+    if (this->fmk != FmkType_MS) {
+      std::cerr << "INPUT ILLEGAL: train model convertor supporting only MINDIR format";
+      return RET_INPUT_PARAM_INVALID;
+    }
+    if ((this->inputDataType != TypeId::kNumberTypeFloat32) && (this->inputDataType != TypeId::kTypeUnknown)) {
+      std::cerr << "INPUT ILLEGAL: train model convertor supporting only FP32 input tensors";
+      return RET_INPUT_PARAM_INVALID;
+    }
+    if ((this->outputDataType != TypeId::kNumberTypeFloat32) && (this->outputDataType != TypeId::kTypeUnknown)) {
+      std::cerr << "INPUT ILLEGAL: train model convertor supporting only FP32 output tensors";
+      return RET_INPUT_PARAM_INVALID;
+    }
+    if (this->quantType != QuantType_QUANT_NONE) {
+      std::cerr << "INPUT ILLEGAL: train model convertor is not supporting quantization";
+      return RET_INPUT_PARAM_INVALID;
+    }
+  }
   return RET_OK;
 }
 }  // namespace converter
diff --git a/mindspore/lite/tools/converter/graphdef_transform.cc b/mindspore/lite/tools/converter/graphdef_transform.cc
index 6f682e2b89..5a0597e04d 100644
--- a/mindspore/lite/tools/converter/graphdef_transform.cc
+++ b/mindspore/lite/tools/converter/graphdef_transform.cc
@@ -49,7 +49,9 @@ int GraphDefTransform::Transform(const converter::Flags &ctx) {
   {
     Optimizer unusedOpRemoveOptimizer;
     unusedOpRemoveOptimizer.AddPass(new UnusedNodeRemovePass());
-    unusedOpRemoveOptimizer.AddPass(new DropoutNodeRemovePass());
+    if (ctx.trainModel == false) {
+      unusedOpRemoveOptimizer.AddPass(new DropoutNodeRemovePass());
+    }
     unusedOpRemoveOptimizer.AddPass(new IsolatedNodeRemovePass());
     status = unusedOpRemoveOptimizer.Run(graphDefT);
     if (status != RET_OK && status != RET_NO_CHANGE) {
diff --git a/mindspore/lite/tools/converter/legacy_optimizer/graph/dropout_node_remove_pass.cc b/mindspore/lite/tools/converter/legacy_optimizer/graph/dropout_node_remove_pass.cc
index 6b6ec109b7..bec2c0328a 100644
--- a/mindspore/lite/tools/converter/legacy_optimizer/graph/dropout_node_remove_pass.cc
+++ b/mindspore/lite/tools/converter/legacy_optimizer/graph/dropout_node_remove_pass.cc
@@ -54,8 +54,8 @@ STATUS IsolateDropoutNode(schema::MetaGraphT *graphT, size_t nodeIdx) {
     }
     auto postNodeIdxes = GetOutputNodeIdx(*graphT, nodeIdx, 1);
     if (postNodeIdxes.size() != 0) {
-      MS_LOG(ERROR) << "Unsupported Dropout: " << node->name.c_str() << " with mask output.";
-      return RET_ERROR;
+      MS_LOG(WARNING) << "Unsupported Dropout: " << node->name.c_str() << " with mask output.";
+      return RET_OK;
     }
   }
   auto inDataTensorIdx = inputTensorIdxes.front();
diff --git a/mindspore/lite/tools/converter/legacy_optimizer/graph/format_trans_pass.cc b/mindspore/lite/tools/converter/legacy_optimizer/graph/format_trans_pass.cc
index aaf9324392..6699974d5b 100644
--- a/mindspore/lite/tools/converter/legacy_optimizer/graph/format_trans_pass.cc
+++ b/mindspore/lite/tools/converter/legacy_optimizer/graph/format_trans_pass.cc
@@ -148,6 +148,7 @@ STATUS FormatTransPass::DoNodeInoutFormatTrans(schema::MetaGraphT *graph) {
 #ifdef SUPPORT_TRAIN
     if (IsContain(GetNhwcAllInputOpList(), GetCNodeTType(**iter))) {
       int idx_num = node->inputIndex.size();
+      if (GetCNodeTType(**iter) == schema::PrimitiveType_BNGrad) idx_num = 2;
       for (int i = 0; i < idx_num; i++) {
         iter = InsertFormatTransNode(graph, iter, kBefore, i, beforeNodeType, &status);
         if (status != RET_OK) {
@@ -155,14 +156,6 @@ STATUS FormatTransPass::DoNodeInoutFormatTrans(schema::MetaGraphT *graph) {
           return RET_ERROR;
         }
       }
-    } else if (IsContain(GetNhwcDualInputOpList(), GetCNodeTType(**iter))) {
-      for (int i = 0; i < 2; i++) {
-        iter = InsertFormatTransNode(graph, iter, kBefore, i, beforeNodeType, &status);
-        if (status != RET_OK) {
-          MS_LOG(ERROR) << "InsertNchw2NhwcNode before " << nodeName << "failed";
-          return RET_ERROR;
-        }
-      }
     } else {
       int idx = 0;
       if (GetCNodeTType(**iter) == schema::PrimitiveType_ApplyMomentum) idx = 3;
diff --git a/mindspore/lite/tools/net_train/CMakeLists.txt b/mindspore/lite/tools/net_train/CMakeLists.txt
new file mode 100644
index 0000000000..566b9bf705
--- /dev/null
+++ b/mindspore/lite/tools/net_train/CMakeLists.txt
@@ -0,0 +1,38 @@
+# add shared link library
+set(COMMON_SRC
+        ${CMAKE_CURRENT_SOURCE_DIR}/../common/flag_parser.cc
+        ${CMAKE_CURRENT_SOURCE_DIR}/../../src/common/file_utils.cc
+        ${CMAKE_CURRENT_SOURCE_DIR}/../../src/common/utils.cc
+        )
+
+add_executable(net_train
+        ${CMAKE_CURRENT_SOURCE_DIR}/main.cc
+        ${CMAKE_CURRENT_SOURCE_DIR}/net_train.cc
+        ${COMMON_SRC})
+if (WIN32)
+    add_dependencies(net_train fbs_src mindspore-lite_static)
+else ()
+    add_dependencies(net_train fbs_src)
+endif ()
+
+if (PLATFORM_ARM32 OR PLATFORM_ARM64)
+    target_link_libraries(net_train mindspore-lite)
+else()
+    if (WIN32)
+        target_link_libraries(net_train mindspore-lite_static pthread cpu_kernel_mid nnacl_mid)
+    else ()
+        target_link_libraries(net_train mindspore-lite pthread)
+    endif ()
+endif()
+if (PLATFORM_ARM32 OR PLATFORM_ARM64)
+    install(TARGETS net_train
+                RUNTIME DESTINATION ${MAIN_DIR}-${COMPONENT_NAME}/net_train COMPONENT ${COMPONENT_NAME})
+else()
+    if (WIN32)
+        install(TARGETS net_train
+            RUNTIME DESTINATION ${MAIN_DIR}-${WIN_RUN_X86_NAME}/net_train COMPONENT ${WIN_RUN_X86_NAME})
+    else ()
+        install(TARGETS net_train
+            RUNTIME DESTINATION ${MAIN_DIR}-${RUN_X86_COMPONENT_NAME}/net_train COMPONENT ${RUN_X86_COMPONENT_NAME})
+    endif ()
+endif()
diff --git a/mindspore/lite/tools/net_train/main.cc b/mindspore/lite/tools/net_train/main.cc
new file mode 100644
index 0000000000..cfde9d319f
--- /dev/null
+++ b/mindspore/lite/tools/net_train/main.cc
@@ -0,0 +1,27 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <malloc.h>
+#include "tools/net_train/net_train.h"
+#include "include/version.h"
+
+int main(int argc, const char **argv) {
+  MS_LOG(INFO) << mindspore::lite::Version();
+  int res = mindspore::lite::RunNetTrain(argc, argv);
+  struct mallinfo info = mallinfo();
+  std::cout << "total allocation: " << info.arena << "\n";
+  return res;
+}
diff --git a/mindspore/lite/tools/net_train/net_train.cc b/mindspore/lite/tools/net_train/net_train.cc
new file mode 100644
index 0000000000..c4e53eb039
--- /dev/null
+++ b/mindspore/lite/tools/net_train/net_train.cc
@@ -0,0 +1,780 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "tools/net_train/net_train.h"
+#define __STDC_FORMAT_MACROS
+#include <cinttypes>
+#undef __STDC_FORMAT_MACROS
+#include <algorithm>
+#include <utility>
+#include "src/common/common.h"
+#include "include/ms_tensor.h"
+#include "include/context.h"
+#include "src/runtime/runtime_api.h"
+#include "include/version.h"
+
+namespace mindspore {
+namespace lite {
+static const char *DELIM_COLON = ":";
+static const char *DELIM_COMMA = ",";
+static const char *DELIM_SLASH = "/";
+
+void SaveFile(std::string path, void *buf, size_t size) {
+  std::ofstream ofs(path);
+  assert(true == ofs.good());
+  assert(true == ofs.is_open());
+
+  ofs.seekp(0, std::ios::beg);
+  ofs.write((const char *)buf, size);
+  ofs.close();
+}
+
+int NetTrain::GenerateRandomData(size_t size, void *data) {
+  MS_ASSERT(data != nullptr);
+  char *casted_data = static_cast<char *>(data);
+  for (size_t i = 0; i < size; i++) {
+    casted_data[i] = static_cast<char>(i);
+  }
+  return RET_OK;
+}
+
+int NetTrain::GenerateInputData() {
+  for (auto tensor : ms_inputs_) {
+    MS_ASSERT(tensor != nullptr);
+    auto input_data = tensor->MutableData();
+    if (input_data == nullptr) {
+      MS_LOG(ERROR) << "MallocData for inTensor failed";
+      return RET_ERROR;
+    }
+    auto tensor_byte_size = tensor->Size();
+    auto status = GenerateRandomData(tensor_byte_size, input_data);
+    if (status != 0) {
+      std::cerr << "GenerateRandomData for inTensor failed: " << status << std::endl;
+      MS_LOG(ERROR) << "GenerateRandomData for inTensor failed:" << status;
+      return status;
+    }
+  }
+  return RET_OK;
+}
+
+int NetTrain::LoadInput() {
+  if (flags_->in_data_file_.empty()) {
+    auto status = GenerateInputData();
+    if (status != 0) {
+      std::cerr << "Generate input data error " << status << std::endl;
+      MS_LOG(ERROR) << "Generate input data error " << status;
+      return status;
+    }
+  } else {
+    auto status = ReadInputFile();
+    if (status != 0) {
+      std::cerr << "ReadInputFile error, " << status << std::endl;
+      MS_LOG(ERROR) << "ReadInputFile error, " << status;
+      return status;
+    }
+  }
+  return RET_OK;
+}
+
+int NetTrain::ReadInputFile() {
+  if (ms_inputs_.empty()) {
+    return RET_OK;
+  }
+
+  if (this->flags_->in_data_type_ == kImage) {
+    MS_LOG(ERROR) << "Not supported image input";
+    return RET_ERROR;
+  } else {
+    for (size_t i = 0; i < flags_->input_data_list_.size(); i++) {
+      auto cur_tensor = ms_inputs_.at(i);
+      MS_ASSERT(cur_tensor != nullptr);
+      size_t size;
+      char *bin_buf = ReadFile(flags_->input_data_list_[i].c_str(), &size);
+      if (bin_buf == nullptr) {
+        MS_LOG(ERROR) << "ReadFile return nullptr";
+        return RET_ERROR;
+      }
+      auto tensor_data_size = cur_tensor->Size();
+      if (size != tensor_data_size) {
+        std::cerr << "Input binary file size error, required: " << tensor_data_size << ", in fact: " << size
+                  << std::endl;
+        MS_LOG(ERROR) << "Input binary file size error, required: " << tensor_data_size << ", in fact: " << size;
+        delete bin_buf;
+        return RET_ERROR;
+      }
+      auto input_data = cur_tensor->MutableData();
+      memcpy(input_data, bin_buf, tensor_data_size);
+      delete[](bin_buf);
+    }
+  }
+  return RET_OK;
+}
+
+// calibData is FP32
+int NetTrain::ReadCalibData() {
+  const char *calib_data_path = flags_->data_file_.c_str();
+  // read calib data
+  std::ifstream in_file(calib_data_path);
+  if (!in_file.good()) {
+    std::cerr << "file: " << calib_data_path << " is not exist" << std::endl;
+    MS_LOG(ERROR) << "file: " << calib_data_path << " is not exist";
+    return RET_ERROR;
+  }
+
+  if (!in_file.is_open()) {
+    std::cerr << "file: " << calib_data_path << " open failed" << std::endl;
+    MS_LOG(ERROR) << "file: " << calib_data_path << " open failed";
+    in_file.close();
+    return RET_ERROR;
+  }
+
+  std::string line;
+
+  MS_LOG(INFO) << "Start reading calibData file";
+  std::string tensor_name;
+  while (!in_file.eof()) {
+    getline(in_file, line);
+    std::stringstream string_line1(line);
+    size_t dim = 0;
+    string_line1 >> tensor_name >> dim;
+    std::vector<size_t> dims;
+    size_t shape_size = 1;
+    for (size_t i = 0; i < dim; i++) {
+      size_t tmp_dim;
+      string_line1 >> tmp_dim;
+      dims.push_back(tmp_dim);
+      shape_size *= tmp_dim;
+    }
+
+    getline(in_file, line);
+    std::stringstream string_line2(line);
+    std::vector<float> tensor_data;
+    for (size_t i = 0; i < shape_size; i++) {
+      float tmp_data;
+      string_line2 >> tmp_data;
+      tensor_data.push_back(tmp_data);
+    }
+    auto *check_tensor = new CheckTensor(dims, tensor_data);
+    this->data_.insert(std::make_pair(tensor_name, check_tensor));
+  }
+  in_file.close();
+  MS_LOG(INFO) << "Finish reading calibData file";
+  return RET_OK;
+}
+
+int NetTrain::CompareOutput() {
+  std::cout << "================ Comparing Output data ================" << std::endl;
+  float total_bias = 0;
+  int total_size = 0;
+  bool has_error = false;
+
+  for (const auto &calib_tensor : data_) {
+    std::string node_or_tensor_name = calib_tensor.first;
+    auto tensors = session_->GetOutputsByNodeName(node_or_tensor_name);
+    mindspore::tensor::MSTensor *tensor = nullptr;
+    if (tensors.empty() || tensors.size() != 1) {
+      MS_LOG(INFO) << "Cannot find output node: " << node_or_tensor_name
+                   << " or node has more than one output tensor, switch to GetOutputByTensorName";
+      tensor = session_->GetOutputByTensorName(node_or_tensor_name);
+      if (tensor == nullptr) {
+        MS_LOG(ERROR) << "Cannot find output tensor " << node_or_tensor_name << ", get model output failed";
+        return RET_ERROR;
+      }
+    } else {
+      tensor = tensors.front();
+    }
+    MS_ASSERT(tensor->MutableData() != nullptr);
+    auto outputs = tensor->MutableData();
+    float bias = CompareData<float>(node_or_tensor_name, tensor->shape(), reinterpret_cast<float *>(outputs));
+
+    if (bias >= 0) {
+      total_bias += bias;
+      total_size++;
+    } else {
+      has_error = true;
+      break;
+    }
+  }
+
+  if (!has_error) {
+    float mean_bias;
+    if (total_size != 0) {
+      mean_bias = total_bias / total_size * 100;
+    } else {
+      mean_bias = 0;
+    }
+
+    std::cout << "Mean bias of all nodes/tensors: " << mean_bias << "%" << std::endl;
+    std::cout << "=======================================================" << std::endl << std::endl;
+
+    if (mean_bias > this->flags_->accuracy_threshold_) {
+      MS_LOG(ERROR) << "Mean bias of all nodes/tensors is too big: " << mean_bias << "%";
+      std::cerr << "Mean bias of all nodes/tensors is too big: " << mean_bias << "%" << std::endl;
+      return RET_ERROR;
+    } else {
+      return RET_OK;
+    }
+  } else {
+    MS_LOG(ERROR) << "Error in CompareData";
+    std::cerr << "Error in CompareData" << std::endl;
+    std::cout << "=======================================================" << std::endl << std::endl;
+    return RET_ERROR;
+  }
+}
+
+int NetTrain::MarkPerformance() {
+  MS_LOG(INFO) << "Running train loops...";
+  std::cout << "Running train loops..." << std::endl;
+  uint64_t time_min = 1000000;
+  uint64_t time_max = 0;
+  uint64_t time_avg = 0;
+
+  for (int i = 0; i < flags_->epochs_; i++) {
+    session_->BindThread(true);
+    auto start = GetTimeUs();
+    auto status =
+      flags_->time_profiling_ ? session_->RunGraph(before_call_back_, after_call_back_) : session_->RunGraph();
+    if (status != 0) {
+      MS_LOG(ERROR) << "Inference error " << status;
+      std::cerr << "Inference error " << status;
+      return status;
+    }
+
+    auto end = GetTimeUs();
+    auto time = end - start;
+    time_min = std::min(time_min, time);
+    time_max = std::max(time_max, time);
+    time_avg += time;
+    session_->BindThread(false);
+  }
+
+  if (flags_->time_profiling_) {
+    const std::vector<std::string> per_op_name = {"opName", "avg(ms)", "percent", "calledTimes", "opTotalTime"};
+    const std::vector<std::string> per_op_type = {"opType", "avg(ms)", "percent", "calledTimes", "opTotalTime"};
+    PrintResult(per_op_name, op_times_by_name_);
+    PrintResult(per_op_type, op_times_by_type_);
+  }
+
+  if (flags_->epochs_ > 0) {
+    time_avg /= flags_->epochs_;
+    MS_LOG(INFO) << "Model = " << flags_->model_file_.substr(flags_->model_file_.find_last_of(DELIM_SLASH) + 1).c_str()
+                 << ", NumThreads = " << flags_->num_threads_ << ", MinRunTime = " << time_min / 1000.0f
+                 << ", MaxRuntime = " << time_max / 1000.0f << ", AvgRunTime = " << time_avg / 1000.0f;
+    printf("Model = %s, NumThreads = %d, MinRunTime = %f ms, MaxRuntime = %f ms, AvgRunTime = %f ms\n",
+           flags_->model_file_.substr(flags_->model_file_.find_last_of(DELIM_SLASH) + 1).c_str(), flags_->num_threads_,
+           time_min / 1000.0f, time_max / 1000.0f, time_avg / 1000.0f);
+  }
+  return RET_OK;
+}
+
+int NetTrain::MarkAccuracy() {
+  MS_LOG(INFO) << "MarkAccuracy";
+  std::cout << "MarkAccuracy" << std::endl;
+  for (auto &msInput : ms_inputs_) {
+    switch (msInput->data_type()) {
+      case TypeId::kNumberTypeFloat:
+        PrintInputData<float>(msInput);
+        break;
+      case TypeId::kNumberTypeFloat32:
+        PrintInputData<float>(msInput);
+        break;
+      case TypeId::kNumberTypeInt32:
+        PrintInputData<int>(msInput);
+        break;
+      default:
+        MS_LOG(ERROR) << "Datatype " << msInput->data_type() << " is not supported.";
+        return RET_ERROR;
+    }
+  }
+  session_->Eval();
+
+  auto status = session_->RunGraph();
+  if (status != RET_OK) {
+    MS_LOG(ERROR) << "Inference error " << status;
+    std::cerr << "Inference error " << status << std::endl;
+    return status;
+  }
+
+  status = ReadCalibData();
+  if (status != RET_OK) {
+    MS_LOG(ERROR) << "Read calib data error " << status;
+    std::cerr << "Read calib data error " << status << std::endl;
+    return status;
+  }
+
+  status = CompareOutput();
+  if (status != RET_OK) {
+    MS_LOG(ERROR) << "Compare output error " << status;
+    std::cerr << "Compare output error " << status << std::endl;
+    return status;
+  }
+  return RET_OK;
+}
+
+int NetTrain::RunExportedNet() {
+  auto start_prepare_time = GetTimeUs();
+  // Load graph
+  std::string model_name = flags_->export_file_.substr(flags_->export_file_.find_last_of(DELIM_SLASH) + 1);
+
+  MS_LOG(INFO) << "start reading exported model file";
+  std::cout << "start reading exported model file" << std::endl;
+  size_t size = 0;
+  char *graph_buf = ReadFile(flags_->export_file_.c_str(), &size);
+  if (graph_buf == nullptr) {
+    MS_LOG(ERROR) << "Read exported model file failed while running " << model_name.c_str();
+    std::cerr << "Read exported model file failed while running " << model_name.c_str() << std::endl;
+    return RET_ERROR;
+  }
+  auto model = lite::TrainModel::Import(graph_buf, size);
+  delete[](graph_buf);
+  if (model == nullptr) {
+    MS_LOG(ERROR) << "Import exported model file failed while running " << model_name.c_str();
+    std::cerr << "Import exported model file failed while running " << model_name.c_str() << std::endl;
+    return RET_ERROR;
+  }
+  auto context = std::make_shared<Context>();
+  if (context == nullptr) {
+    MS_LOG(ERROR) << "New context failed while running " << model_name.c_str();
+    std::cerr << "New context failed while running " << model_name.c_str() << std::endl;
+    return RET_ERROR;
+  }
+
+  if (flags_->cpu_bind_mode_ == 2) {
+    context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = MID_CPU;
+  } else if (flags_->cpu_bind_mode_ == 1) {
+    context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = HIGHER_CPU;
+  } else {
+    context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = NO_BIND;
+  }
+
+  context->thread_num_ = flags_->num_threads_;
+  // context->enable_float16_ = flags_->enable_fp16_;
+  session_ = session::TrainSession::CreateSession(context.get());
+  if (session_ == nullptr) {
+    MS_LOG(ERROR) << "CreateSession failed while running ", model_name.c_str();
+    std::cout << "CreateSession failed while running ", model_name.c_str();
+    return RET_ERROR;
+  }
+  auto ret = session_->CompileTrainGraph(model);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "CompileGraph failed while running ", model_name.c_str();
+    std::cout << "CompileGraph failed while running ", model_name.c_str();
+    return ret;
+  }
+
+  ms_inputs_ = session_->GetInputs();
+  auto end_prepare_time = GetTimeUs();
+  MS_LOG(INFO) << "Exported model PrepareTime = " << (end_prepare_time - start_prepare_time) / 1000 << " ms";
+  std::cout << "Exported model PrepareTime = " << (end_prepare_time - start_prepare_time) / 1000 << " ms" << std::endl;
+
+  // Load input
+  MS_LOG(INFO) << "start generate input data";
+  auto status = LoadInput();
+  if (status != 0) {
+    MS_LOG(ERROR) << "Generate input data error";
+    return status;
+  }
+
+  status = session_->RunGraph();
+  if (status != 0) {
+    MS_LOG(ERROR) << "Inference error " << status;
+    std::cerr << "Inference error " << status << std::endl;
+    return status;
+  }
+
+  if (!flags_->data_file_.empty()) {
+    MS_LOG(INFO) << "Check accuracy for exported model";
+    std::cout << "Check accuracy for exported model " << std::endl;
+    status = MarkAccuracy();
+    for (auto &data : data_) {
+      data.second->shape.clear();
+      data.second->data.clear();
+      delete data.second;
+    }
+    data_.clear();
+    if (status != 0) {
+      MS_LOG(ERROR) << "Run MarkAccuracy on exported model error: " << status;
+      std::cout << "Run MarkAccuracy on exported model error: " << status << std::endl;
+      return status;
+    }
+  }
+  return RET_OK;
+}
+
+int NetTrain::RunNetTrain() {
+  auto start_prepare_time = GetTimeUs();
+  // Load graph
+  std::string model_name = flags_->model_file_.substr(flags_->model_file_.find_last_of(DELIM_SLASH) + 1);
+
+  MS_LOG(INFO) << "start reading model file";
+  std::cout << "start reading model file" << std::endl;
+  size_t size = 0;
+  char *graph_buf = ReadFile(flags_->model_file_.c_str(), &size);
+  if (graph_buf == nullptr) {
+    MS_LOG(ERROR) << "Read model file failed while running " << model_name.c_str();
+    std::cerr << "Read model file failed while running " << model_name.c_str() << std::endl;
+    return RET_ERROR;
+  }
+  auto model = lite::TrainModel::Import(graph_buf, size);
+  delete[](graph_buf);
+  if (model == nullptr) {
+    MS_LOG(ERROR) << "Import model file failed while running " << model_name.c_str();
+    std::cerr << "Import model file failed while running " << model_name.c_str() << std::endl;
+    return RET_ERROR;
+  }
+  auto context = std::make_shared<Context>();
+  if (context == nullptr) {
+    MS_LOG(ERROR) << "New context failed while running " << model_name.c_str();
+    std::cerr << "New context failed while running " << model_name.c_str() << std::endl;
+    return RET_ERROR;
+  }
+
+  if (flags_->cpu_bind_mode_ == 2) {
+    context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = MID_CPU;
+  } else if (flags_->cpu_bind_mode_ == 1) {
+    context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = HIGHER_CPU;
+  } else {
+    context->device_list_[0].device_info_.cpu_device_info_.cpu_bind_mode_ = NO_BIND;
+  }
+  context->thread_num_ = flags_->num_threads_;
+  // context->enable_float16_ = flags_->enable_fp16_;
+  session_ = session::TrainSession::CreateSession(context.get());
+  if (session_ == nullptr) {
+    MS_LOG(ERROR) << "CreateSession failed while running ", model_name.c_str();
+    std::cout << "CreateSession failed while running ", model_name.c_str();
+    return RET_ERROR;
+  }
+  auto ret = session_->CompileTrainGraph(model);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "CompileGraph failed while running ", model_name.c_str();
+    std::cout << "CompileGraph failed while running ", model_name.c_str();
+    return ret;
+  }
+
+  session_->Train();
+
+  ms_inputs_ = session_->GetInputs();
+  auto end_prepare_time = GetTimeUs();
+  MS_LOG(INFO) << "PrepareTime = " << (end_prepare_time - start_prepare_time) / 1000 << " ms";
+  std::cout << "PrepareTime = " << (end_prepare_time - start_prepare_time) / 1000 << " ms" << std::endl;
+
+  // Load input
+  MS_LOG(INFO) << "start generate input data";
+  auto status = LoadInput();
+  if (status != 0) {
+    MS_LOG(ERROR) << "Generate input data error";
+    return status;
+  }
+  if (flags_->epochs_ > 0) {
+    status = MarkPerformance();
+    if (status != 0) {
+      MS_LOG(ERROR) << "Run MarkPerformance error: " << status;
+      std::cout << "Run MarkPerformance error: " << status << std::endl;
+      return status;
+    }
+  }
+  if (!flags_->data_file_.empty()) {
+    status = MarkAccuracy();
+    for (auto &data : data_) {
+      data.second->shape.clear();
+      data.second->data.clear();
+      delete data.second;
+    }
+    data_.clear();
+    if (status != 0) {
+      MS_LOG(ERROR) << "Run MarkAccuracy error: " << status;
+      std::cout << "Run MarkAccuracy error: " << status << std::endl;
+      return status;
+    }
+  }
+  if (!flags_->export_file_.empty()) {
+    size_t tsize = 0;
+    auto buf = session_->ExportToBuf(nullptr, &tsize);
+    if (buf == nullptr) {
+      MS_LOG(ERROR) << "Run ExportToBuf error";
+      std::cout << "Run ExportToBuf error";
+      return RET_ERROR;
+    }
+    SaveFile(flags_->export_file_, buf, size);
+
+    status = RunExportedNet();
+    if (status != 0) {
+      MS_LOG(ERROR) << "Run Exported model error: " << status;
+      std::cout << "Run Exported model error: " << status << std::endl;
+      return status;
+    }
+  }
+  return RET_OK;
+}
+
+void NetTrainFlags::InitInputDataList() {
+  char *saveptr1;
+  char *input_list = new char[this->in_data_file_.length() + 1];
+  snprintf(input_list, this->in_data_file_.length() + 1, "%s", this->in_data_file_.c_str());
+  char *cur_input;
+  const char *split_c = ",";
+  cur_input = strtok_r(input_list, split_c, &saveptr1);
+  while (cur_input != nullptr) {
+    input_data_list_.emplace_back(cur_input);
+    cur_input = strtok_r(nullptr, split_c, &saveptr1);
+  }
+  delete[] input_list;
+}
+
+void NetTrainFlags::InitResizeDimsList() {
+  std::string content;
+  content = this->resize_dims_in_;
+  std::vector<int64_t> shape;
+  auto shape_strs = StringSplit(content, std::string(DELIM_COLON));
+  for (const auto &shape_str : shape_strs) {
+    shape.clear();
+    auto dim_strs = StringSplit(shape_str, std::string(DELIM_COMMA));
+    std::cout << "Resize Dims: ";
+    for (const auto &dim_str : dim_strs) {
+      std::cout << dim_str << " ";
+      shape.emplace_back(static_cast<int64_t>(std::stoi(dim_str)));
+    }
+    std::cout << std::endl;
+    this->resize_dims_.emplace_back(shape);
+  }
+}
+
+int NetTrain::InitCallbackParameter() {
+  // before callback
+  before_call_back_ = [&](const std::vector<mindspore::tensor::MSTensor *> &before_inputs,
+                          const std::vector<mindspore::tensor::MSTensor *> &before_outputs,
+                          const mindspore::CallBackParam &callParam) {
+    if (before_inputs.empty()) {
+      MS_LOG(INFO) << "The num of beforeInputs is empty";
+    }
+    if (before_outputs.empty()) {
+      MS_LOG(INFO) << "The num of beforeOutputs is empty";
+    }
+    if (op_times_by_type_.find(callParam.node_type) == op_times_by_type_.end()) {
+      op_times_by_type_.insert(std::make_pair(callParam.node_type, std::make_pair(0, 0.0f)));
+    }
+    if (op_times_by_name_.find(callParam.node_name) == op_times_by_name_.end()) {
+      op_times_by_name_.insert(std::make_pair(callParam.node_name, std::make_pair(0, 0.0f)));
+    }
+
+    op_call_times_total_++;
+    op_begin_ = GetTimeUs();
+    return true;
+  };
+
+  // after callback
+  after_call_back_ = [&](const std::vector<mindspore::tensor::MSTensor *> &after_inputs,
+                         const std::vector<mindspore::tensor::MSTensor *> &after_outputs,
+                         const mindspore::CallBackParam &call_param) {
+    uint64_t opEnd = GetTimeUs();
+
+    if (after_inputs.empty()) {
+      MS_LOG(INFO) << "The num of after inputs is empty";
+    }
+    if (after_outputs.empty()) {
+      MS_LOG(INFO) << "The num of after outputs is empty";
+    }
+
+    float cost = static_cast<float>(opEnd - op_begin_) / 1000.0f;
+    op_cost_total_ += cost;
+    op_times_by_type_[call_param.node_type].first++;
+    op_times_by_type_[call_param.node_type].second += cost;
+    op_times_by_name_[call_param.node_name].first++;
+    op_times_by_name_[call_param.node_name].second += cost;
+    return true;
+  };
+
+  return RET_OK;
+}
+
+int NetTrain::Init() {
+  if (this->flags_ == nullptr) {
+    return 1;
+  }
+  MS_LOG(INFO) << "ModelPath = " << this->flags_->model_file_;
+  MS_LOG(INFO) << "InDataPath = " << this->flags_->in_data_file_;
+  MS_LOG(INFO) << "InDataType = " << this->flags_->in_data_type_in_;
+  MS_LOG(INFO) << "Epochs = " << this->flags_->epochs_;
+  MS_LOG(INFO) << "AccuracyThreshold = " << this->flags_->accuracy_threshold_;
+  MS_LOG(INFO) << "WarmUpLoopCount = " << this->flags_->warm_up_loop_count_;
+  MS_LOG(INFO) << "NumThreads = " << this->flags_->num_threads_;
+  MS_LOG(INFO) << "expectedDataFile = " << this->flags_->data_file_;
+  MS_LOG(INFO) << "exportDataFile = " << this->flags_->export_file_;
+
+  if (this->flags_->epochs_ < 0) {
+    MS_LOG(ERROR) << "epochs:" << this->flags_->epochs_ << " must be equal/greater than 0";
+    std::cerr << "epochs:" << this->flags_->epochs_ << " must be equal/greater than 0" << std::endl;
+    return RET_ERROR;
+  }
+
+  if (this->flags_->num_threads_ < 1) {
+    MS_LOG(ERROR) << "numThreads:" << this->flags_->num_threads_ << " must be greater than 0";
+    std::cerr << "numThreads:" << this->flags_->num_threads_ << " must be greater than 0" << std::endl;
+    return RET_ERROR;
+  }
+
+  this->flags_->in_data_type_ = this->flags_->in_data_type_in_ == "img" ? kImage : kBinary;
+
+  if (flags_->in_data_file_.empty() && !flags_->data_file_.empty()) {
+    MS_LOG(ERROR) << "expectedDataFile not supported in case that inDataFile is not provided";
+    std::cerr << "expectedDataFile is not supported in case that inDataFile is not provided" << std::endl;
+    return RET_ERROR;
+  }
+
+  if (flags_->in_data_file_.empty() && !flags_->export_file_.empty()) {
+    MS_LOG(ERROR) << "exportDataFile not supported in case that inDataFile is not provided";
+    std::cerr << "exportDataFile is not supported in case that inDataFile is not provided" << std::endl;
+    return RET_ERROR;
+  }
+
+  if (flags_->model_file_.empty()) {
+    MS_LOG(ERROR) << "modelPath is required";
+    std::cerr << "modelPath is required" << std::endl;
+    return 1;
+  }
+  flags_->InitInputDataList();
+  flags_->InitResizeDimsList();
+  if (!flags_->resize_dims_.empty() && flags_->resize_dims_.size() != flags_->input_data_list_.size()) {
+    MS_LOG(ERROR) << "Size of input resizeDims should be equal to size of input inDataPath";
+    std::cerr << "Size of input resizeDims should be equal to size of input inDataPath" << std::endl;
+    return RET_ERROR;
+  }
+
+  if (flags_->time_profiling_) {
+    auto status = InitCallbackParameter();
+    if (status != RET_OK) {
+      MS_LOG(ERROR) << "Init callback Parameter failed.";
+      std::cerr << "Init callback Parameter failed." << std::endl;
+      return RET_ERROR;
+    }
+  }
+
+  return RET_OK;
+}
+
+int NetTrain::PrintResult(const std::vector<std::string> &title,
+                          const std::map<std::string, std::pair<int, float>> &result) {
+  std::vector<size_t> columnLenMax(5);
+  std::vector<std::vector<std::string>> rows;
+
+  for (auto &iter : result) {
+    char stringBuf[5][100] = {};
+    std::vector<std::string> columns;
+    size_t len;
+
+    len = iter.first.size();
+    if (len > columnLenMax.at(0)) {
+      columnLenMax.at(0) = len + 4;
+    }
+    columns.push_back(iter.first);
+
+    len = snprintf(stringBuf[1], sizeof(stringBuf[1]), "%f", iter.second.second / flags_->epochs_);
+    if (len > columnLenMax.at(1)) {
+      columnLenMax.at(1) = len + 4;
+    }
+    columns.emplace_back(stringBuf[1]);
+
+    len = snprintf(stringBuf[2], sizeof(stringBuf[2]), "%f", iter.second.second / op_cost_total_);
+    if (len > columnLenMax.at(2)) {
+      columnLenMax.at(2) = len + 4;
+    }
+    columns.emplace_back(stringBuf[2]);
+
+    len = snprintf(stringBuf[3], sizeof(stringBuf[3]), "%d", iter.second.first);
+    if (len > columnLenMax.at(3)) {
+      columnLenMax.at(3) = len + 4;
+    }
+    columns.emplace_back(stringBuf[3]);
+
+    len = snprintf(stringBuf[4], sizeof(stringBuf[4]), "%f", iter.second.second);
+    if (len > columnLenMax.at(4)) {
+      columnLenMax.at(4) = len + 4;
+    }
+    columns.emplace_back(stringBuf[4]);
+
+    rows.push_back(columns);
+  }
+
+  printf("-------------------------------------------------------------------------\n");
+  for (int i = 0; i < 5; i++) {
+    auto printBuf = title[i];
+    if (printBuf.size() > columnLenMax.at(i)) {
+      columnLenMax.at(i) = printBuf.size();
+    }
+    printBuf.resize(columnLenMax.at(i), ' ');
+    printf("%s\t", printBuf.c_str());
+  }
+  printf("\n");
+  for (size_t i = 0; i < rows.size(); i++) {
+    for (int j = 0; j < 5; j++) {
+      auto printBuf = rows[i][j];
+      printBuf.resize(columnLenMax.at(j), ' ');
+      printf("%s\t", printBuf.c_str());
+    }
+    printf("\n");
+  }
+  return RET_OK;
+}
+
+NetTrain::~NetTrain() {
+  for (auto iter : this->data_) {
+    delete (iter.second);
+  }
+  this->data_.clear();
+  delete (session_);
+}
+
+int RunNetTrain(int argc, const char **argv) {
+  NetTrainFlags flags;
+  Option<std::string> err = flags.ParseFlags(argc, argv);
+
+  if (err.IsSome()) {
+    std::cerr << err.Get() << std::endl;
+    std::cerr << flags.Usage() << std::endl;
+    return RET_ERROR;
+  }
+
+  if (flags.help) {
+    std::cerr << flags.Usage() << std::endl;
+    return RET_OK;
+  }
+
+  NetTrain net_trainer(&flags);
+  auto status = net_trainer.Init();
+  if (status != 0) {
+    MS_LOG(ERROR) << "NetTrain init Error : " << status;
+    std::cerr << "NetTrain init Error : " << status << std::endl;
+    return RET_ERROR;
+  }
+
+  status = net_trainer.RunNetTrain();
+  if (status != 0) {
+    MS_LOG(ERROR) << "Run NetTrain "
+                  << flags.model_file_.substr(flags.model_file_.find_last_of(DELIM_SLASH) + 1).c_str()
+                  << " Failed : " << status;
+    std::cerr << "Run NetTrain " << flags.model_file_.substr(flags.model_file_.find_last_of(DELIM_SLASH) + 1).c_str()
+              << " Failed : " << status << std::endl;
+    return RET_ERROR;
+  }
+
+  MS_LOG(INFO) << "Run NetTrain " << flags.model_file_.substr(flags.model_file_.find_last_of(DELIM_SLASH) + 1).c_str()
+               << " Success.";
+  std::cout << "Run NetTrain " << flags.model_file_.substr(flags.model_file_.find_last_of(DELIM_SLASH) + 1).c_str()
+            << " Success." << std::endl;
+  return RET_OK;
+}
+}  // namespace lite
+}  // namespace mindspore
diff --git a/mindspore/lite/tools/net_train/net_train.h b/mindspore/lite/tools/net_train/net_train.h
new file mode 100644
index 0000000000..df600a06ac
--- /dev/null
+++ b/mindspore/lite/tools/net_train/net_train.h
@@ -0,0 +1,243 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_TOOLS_NET_TRAIN_NET_TRAIN_H_
+#define MINDSPORE_LITE_TOOLS_NET_TRAIN_NET_TRAIN_H_
+
+#include <getopt.h>
+#include <signal.h>
+#include <unordered_map>
+#include <fstream>
+#include <iostream>
+#include <map>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <memory>
+#include <cfloat>
+#include <utility>
+#include "include/train_model.h"
+#include "tools/common/flag_parser.h"
+#include "src/common/file_utils.h"
+#include "src/common/utils.h"
+#include "include/train_session.h"
+
+namespace mindspore::lite {
+enum MS_API DataType { kImage = 0, kBinary = 1 };
+
+constexpr float relativeTolerance = 1e-5;
+constexpr float absoluteTolerance = 1e-8;
+
+struct MS_API CheckTensor {
+  CheckTensor(const std::vector<size_t> &shape, const std::vector<float> &data) {
+    this->shape = shape;
+    this->data = data;
+  }
+  std::vector<size_t> shape;
+  std::vector<float> data;
+};
+
+class MS_API NetTrainFlags : public virtual FlagParser {
+ public:
+  NetTrainFlags() {
+    // common
+    AddFlag(&NetTrainFlags::model_file_, "modelFile", "Input model file", "");
+    AddFlag(&NetTrainFlags::in_data_file_, "inDataFile", "Input data file, if not set, use random input", "");
+    // MarkPerformance
+    AddFlag(&NetTrainFlags::warm_up_loop_count_, "warmUpLoopCount", "Run warm up loop", 0);
+    AddFlag(&NetTrainFlags::time_profiling_, "timeProfiling", "Run time profiling", false);
+    AddFlag(&NetTrainFlags::epochs_, "epochs", "Number of training epochs to run", 1);
+    // MarkAccuracy
+    AddFlag(&NetTrainFlags::data_file_, "expectedDataFile", "Expected results data file path", "");
+    AddFlag(&NetTrainFlags::export_file_, "exportFile", "MS File to export trained model into", "");
+    AddFlag(&NetTrainFlags::accuracy_threshold_, "accuracyThreshold", "Threshold of accuracy", 0.5);
+  }
+
+  ~NetTrainFlags() override = default;
+
+  void InitInputDataList();
+
+  void InitResizeDimsList();
+
+ public:
+  // common
+  std::string model_file_;
+  std::string in_data_file_;
+  std::vector<std::string> input_data_list_;
+  DataType in_data_type_;
+  std::string in_data_type_in_ = "bin";
+  int cpu_bind_mode_ = 0;
+  // MarkPerformance
+  int num_threads_ = 1;
+  int warm_up_loop_count_ = 0;
+  bool time_profiling_;
+  int epochs_ = 1;
+  // MarkAccuracy
+  std::string data_file_;
+  std::string data_type_ = "FLOAT";
+  float accuracy_threshold_;
+  // Resize
+  std::string export_file_ = "";
+  std::string resize_dims_in_ = "";
+  std::vector<std::vector<int64_t>> resize_dims_;
+};
+
+class MS_API NetTrain {
+ public:
+  explicit NetTrain(NetTrainFlags *flags) : flags_(flags) {}
+
+  virtual ~NetTrain();
+
+  int Init();
+  int RunNetTrain();
+  int RunExportedNet();
+
+ private:
+  // call GenerateInputData or ReadInputFile to init inputTensors
+  int LoadInput();
+
+  // call GenerateRandomData to fill inputTensors
+  int GenerateInputData();
+
+  int GenerateRandomData(size_t size, void *data);
+
+  int ReadInputFile();
+
+  int ReadCalibData();
+
+  int CompareOutput();
+
+  int InitCallbackParameter();
+
+  int PrintResult(const std::vector<std::string> &title, const std::map<std::string, std::pair<int, float>> &result);
+
+  template <typename T>
+  void PrintInputData(tensor::MSTensor *input) {
+    MS_ASSERT(input != nullptr);
+    static int i = 0;
+    auto inData = reinterpret_cast<T *>(input->MutableData());
+    std::cout << "InData" << i++ << ": ";
+    for (size_t j = 0; j < 20; j++) {
+      std::cout << inData[j] << " ";
+    }
+    std::cout << std::endl;
+  }
+
+  // tensorData need to be converter first
+  template <typename T>
+  float CompareData(const std::string &nodeName, std::vector<int> msShape, T *msTensorData) {
+    auto iter = this->data_.find(nodeName);
+    if (iter != this->data_.end()) {
+      std::vector<size_t> castedMSShape;
+      size_t shapeSize = 1;
+      for (int64_t dim : msShape) {
+        castedMSShape.push_back(size_t(dim));
+        shapeSize *= dim;
+      }
+
+      CheckTensor *calibTensor = iter->second;
+      if (calibTensor->shape != castedMSShape) {
+        std::ostringstream oss;
+        oss << "Shape of mslite output(";
+        for (auto dim : castedMSShape) {
+          oss << dim << ",";
+        }
+        oss << ") and shape source model output(";
+        for (auto dim : calibTensor->shape) {
+          oss << dim << ",";
+        }
+        oss << ") are different";
+        std::cerr << oss.str() << std::endl;
+        MS_LOG(ERROR) << oss.str().c_str();
+        return RET_ERROR;
+      }
+      size_t errorCount = 0;
+      float meanError = 0;
+      std::cout << "Data of node " << nodeName << " : ";
+      for (size_t j = 0; j < shapeSize; j++) {
+        if (j < 50) {
+          std::cout << static_cast<float>(msTensorData[j]) << " ";
+        }
+
+        if (std::isnan(msTensorData[j]) || std::isinf(msTensorData[j])) {
+          std::cerr << "Output tensor has nan or inf data, compare fail" << std::endl;
+          MS_LOG(ERROR) << "Output tensor has nan or inf data, compare fail";
+          return RET_ERROR;
+        }
+
+        auto tolerance = absoluteTolerance + relativeTolerance * fabs(calibTensor->data.at(j));
+        auto absoluteError = std::fabs(msTensorData[j] - calibTensor->data.at(j));
+        if (absoluteError > tolerance) {
+          if (fabs(calibTensor->data.at(j)) == 0) {
+            if (absoluteError > 1e-5) {
+              meanError += absoluteError;
+              errorCount++;
+            } else {
+              continue;
+            }
+          } else {
+            // just assume that atol = rtol
+            if (absoluteError > 1e-5) {
+              meanError += absoluteError / (fabs(calibTensor->data.at(j)) + FLT_MIN);
+              errorCount++;
+            }
+          }
+        }
+      }
+      std::cout << std::endl;
+      if (meanError > 0.0f) {
+        meanError /= errorCount;
+      }
+
+      if (meanError <= 0.0000001) {
+        std::cout << "Mean bias of node/tensor " << nodeName << " : 0%" << std::endl;
+      } else {
+        std::cout << "Mean bias of node/tensor " << nodeName << " : " << meanError * 100 << "%" << std::endl;
+      }
+      return meanError;
+    } else {
+      MS_LOG(INFO) << "%s is not in Source Model output", nodeName.c_str();
+      return RET_ERROR;
+    }
+  }
+
+  int MarkPerformance();
+
+  int MarkAccuracy();
+
+ private:
+  NetTrainFlags *flags_;
+  session::TrainSession *session_;
+  std::vector<mindspore::tensor::MSTensor *> ms_inputs_;
+  std::unordered_map<std::string, std::vector<mindspore::tensor::MSTensor *>> ms_outputs_;
+  std::unordered_map<std::string, CheckTensor *> data_;
+  std::unordered_map<std::string, TypeId> data_type_map_{{"FLOAT", TypeId::kNumberTypeFloat},
+                                                         {"INT32", TypeId::kNumberTypeInt32}};
+
+  // callback parameters
+  uint64_t op_begin_ = 0;
+  int op_call_times_total_ = 0;
+  float op_cost_total_ = 0.0f;
+  std::map<std::string, std::pair<int, float>> op_times_by_type_;
+  std::map<std::string, std::pair<int, float>> op_times_by_name_;
+
+  mindspore::KernelCallBack before_call_back_;
+  mindspore::KernelCallBack after_call_back_;
+};
+
+int MS_API RunNetTrain(int argc, const char **argv);
+}  // namespace mindspore::lite
+#endif  // MINDSPORE_LITE_TOOLS_NET_TRAIN_NET_TRAIN_H_