train on device

mindspore/lite/include/model.h

@@ -69,6 +69,7 @@ class MS_API Model {
 
   /// \brief Free MetaGraph in MindSpore Lite Model.
   void FreeMetaGraph();
+  ModelImpl *model_impl() {return model_impl_;}
 
  protected:
   ModelImpl *model_impl_ = nullptr;

mindspore/lite/include/train_session.h

@@ -0,0 +1,63 @@
+/**
+ * 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_INCLUDE_TRAIN_SESSION_H_
+#define MINDSPORE_LITE_INCLUDE_TRAIN_SESSION_H_
+#include <vector>
+#include <string>
+#include <unordered_map>
+// #include "include/lite_session.h"
+#include "src/lite_session.h"
+
+namespace mindspore {
+namespace lite {
+class Model;
+}
+namespace lite::tensor {
+class Tensor;
+}
+namespace session {
+
+class TrainSession : public lite::LiteSession {
+ public:
+  TrainSession();
+  ~TrainSession() = default;
+
+  int RunGraph(const session::KernelCallBack &before = nullptr,
+               const session::KernelCallBack &after = nullptr) override;
+
+  int CompileGraph(lite::Model *model) override;
+  virtual void ReplaceOps();
+  virtual void* ExportToBuf(void* buf, size_t* len) const;
+
+  std::unordered_map<std::string, std::vector<mindspore::tensor::MSTensor *>> GetOutputs() const;
+  std::vector<tensor::MSTensor *> GetOutputsByName(const std::string &node_name) const;
+
+  virtual void train();
+  bool is_train() { return train_mode_ == true; }
+  virtual void eval();
+  bool is_eval() { return train_mode_ == false; }
+
+ protected:
+  bool  train_mode_ = false;
+  lite::Model* model_ = nullptr;
+  std::unordered_map<std::string, std::vector<mindspore::tensor::MSTensor *>> ext_output_map_;
+
+
+  // private:
+};
+}  // namespace session
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_INCLUDE_TRAIN_SESSION_H_

mindspore/lite/nnacl/activation_grad.c → mindspore/lite/nnacl/fp32_grad/activation_grad.c

@@ -13,9 +13,14 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include "nnacl/activation_grad.h"
 
-int ReluGrad(float *src0, float *src1, int length, float *dst) {
+#include <math.h>
+#include "nnacl/op_base.h"
+#include "nnacl/fp32/arithmetic.h"
+#include "nnacl/fp32_grad/activation_grad.h"
+#include "nnacl/errorcode.h"
+
+inline int ReluGrad(float *src0, float *src1, int length, float *dst) {
   for (int i = 0; i < length; ++i) {
     dst[i] = src1[i] > 0 ? 1.0f : 0.0f;
   }

mindspore/lite/nnacl/fp32_grad/batch_norm.c

@@ -13,11 +13,11 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include <string.h>
 #include <math.h>
+#include <string.h>
 #include "nnacl/fp32_grad/batch_norm.h"
 
-static void sumSpatialBatch(const float *in, int size, int ch, float *out) {
+void sumSpatialBatch(const float *in, int size, int ch, float *out) {
   memset(out, 0, ch * sizeof(float));
   for (int i = 0; i < size; i++) {
     const float *ptr = in + i * ch;
@@ -32,49 +32,53 @@ void scaleBias(const float *scales, int batch, int n, int size, float *output) {
     for (int c = 0; c < n; c++) output[i * n + c] *= scales[c];
 }
 
-void normalize(const float *x, const float *mean, const float *variance, float eps, int batch, int filters, int spatial,
+void normalize(const float *x, const float *mean, const float *invar, int batch, int filters, int spatial,
                float *out) {
   int b, f, i;
   for (b = 0; b < batch; ++b) {
     for (i = 0; i < spatial; ++i) {
       for (f = 0; f < filters; ++f) {
         int index = b * filters * spatial + i * filters + f;
-        out[index] = (x[index] - mean[f]) / (sqrt(variance[f]) + eps);
+        out[index] = (x[index] - mean[f]) * invar[f];
       }
     }
   }
 }
 
-void backwardScale(const float *x_norm, const float *delta, int batch, int n, int size, float *scale_updates) {
+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;
   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;
-        scale_updates[f] += delta[index] * x_norm[index];
+        float x_norm = (x[index] - mean[f]) * invar[f];
+        scale_updates[f] += delta[index] * x_norm;
       }
     }
   }
 }
 
-void meanVar(const float *in, int batch, int spatial, int ch, float *mean, float *var) {
+void meanVar(const float *in, int batch, int spatial, int ch, float eps, float *mean, float *invar) {
   float N = batch * spatial;
   sumSpatialBatch(in, N, ch, mean);
-  for (int f = 0; f < ch; ++f) mean[f] /= N;
-  memset(var, 0, ch * sizeof(float));
-  for (int i = 0; i < N; i++) {
-    for (int f = 0; f < ch; f++) {
-      float x = in[i * ch + f];
-      var[f] += (x - mean[f]) * (x - mean[f]);
+  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));
   }
-  for (int f = 0; f < ch; f++) var[f] /= N;
 }
 
-void meanDelta(float *yt, int size, int ch, float eps, float *variance, float *mean_delta) {
+void meanDelta(float *yt, int size, int ch,  float *invar, float *mean_delta) {
   sumSpatialBatch(yt, size, ch, mean_delta);
-  for (int i = 0; i < ch; i++) mean_delta[i] *= -1.f / sqrt((variance[i] + eps));
+  for (int i = 0; i < ch; i++) mean_delta[i] *= -invar[i];
 }
 
 void meanAdd(const float *x, const float *mean, const float *variance_delta, int batch, int filters, int spatial,
@@ -93,8 +97,8 @@ void meanAdd(const float *x, const float *mean, const float *variance_delta, int
   }
 }
 
-void varianceDelta(const float *x, const float *delta, const float *mean, const float *variance, int batch, int filters,
-                   int spatial, float eps, float *variance_delta) {
+void varianceDelta(const float *x, const float *delta, const float *mean, const float *invar, int batch, int filters,
+                   int spatial, float *variance_delta) {
   int i, k;
   memset(variance_delta, 0, filters * sizeof(float));
   for (k = 0; k < batch * spatial; k++) {
@@ -103,16 +107,16 @@ void varianceDelta(const float *x, const float *delta, const float *mean, const
       variance_delta[i] += delta[index] * (x[index] - mean[i]);
     }
   }
-  for (i = 0; i < filters; i++) variance_delta[i] *= -.5 * pow(variance[i] + eps, (-3.f / 2.f));
+  for (i = 0; i < filters; i++) variance_delta[i] *= -.5 * 1.0f/(invar[i]*invar[i]*invar[i]);
 }
 
-void NormalizeDelta(const float *x, const float *mean, const float *variance, const float *mean_delta,
-                    const float *variance_delta, int batch, int filters, int spatial, float eps, float *delta) {
+void NormalizeDelta(const float *x, const float *mean, const float *invar, const float *mean_delta,
+                    const float *variance_delta, int batch, int filters, int spatial, float *delta) {
   int f, k;
   for (k = 0; k < batch * spatial; k++) {
     for (f = 0; f < filters; f++) {
       int index = k * filters + f;
-      delta[index] = delta[index] * 1. / (sqrt(variance[f] + eps)) +
+      delta[index] = delta[index] * invar[f] +
                      variance_delta[f] * 2. * (x[index] - mean[f]) / (spatial * batch) +
                      mean_delta[f] / (spatial * batch);
     }

mindspore/lite/nnacl/fp32_grad/batch_norm.h

@@ -17,28 +17,33 @@
 #ifndef MINDSPORE_LITE_NNACL_FP32_BATCH_NORM_H_
 #define MINDSPORE_LITE_NNACL_FP32_BATCH_NORM_H_
 
-typedef struct bnParameter {
-  int batch;
-  int channels;
-  int spatial;
-  float eps;
-} bnParameter;
+#include "nnacl/op_base.h"
+
+typedef struct BNGradParameter {
+  OpParameter op_parameter_;
+  float epsilon_;
+  float momentum_;
+} BNGradParameter;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
+
+
+void sumSpatialBatch(const float *in, int size, int ch, float *out);
 void scaleBias(const float *scales, int batch, int n, int size, float *output);
-void normalize(const float *x, const float *mean, const float *variance, float eps, int batch, int filters, int spatial,
+void normalize(const float *x, const float *mean, const float *invar, int batch, int filters, int spatial,
                float *out);
-void backwardScale(const float *x_norm, const float *delta, int batch, int n, int size, float *scale_updates);
-void meanVar(const float *in, int batch, int size, int ch, float *mean, float *var);
-void meanDelta(float *yt, int size, int ch, float eps, float *variance, float *mean_delta);
-void varianceDelta(const float *x, const float *delta, const float *mean, const float *variance, int batch, int ch,
-                   int spatial, float eps, float *variance_delta);
+void backwardScale(const float *x, const float *mean, const float *invar, const float *delta, int batch,
+                   int n, int size, float *scale_updates);
+void meanVar(const float *in, int batch, int size, int ch, float eps, float *mean, float *invar);
+void meanDelta(float *yt, int size, int ch, float *invar, float *mean_delta);
+void varianceDelta(const float *x, const float *delta, const float *mean, const float *invar, int batch, int ch,
+                   int spatial, float *variance_delta);
 void meanAdd(const float *x, const float *mean, const float *variance_delta, int batch, int filters, int spatial,
              float *mean_add, float *mean_delta);
-void NormalizeDelta(const float *x, const float *mean, const float *variance, const float *mean_delta,
-                    const float *variance_delta, int batch, int filters, int spatial, float eps, float *delta);
+void NormalizeDelta(const float *x, const float *mean, const float *invar, const float *mean_delta,
+                    const float *variance_delta, int batch, int filters, int spatial, float *delta);
 #ifdef __cplusplus
 }
 #endif

mindspore/lite/nnacl/fp32_grad/pack_ext.c

@@ -125,9 +125,9 @@ 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) {
-  const int pad_left = /*conv_param->pad_l_*/ conv_param->pad_w_;
+  const int pad_left = /*conv_param->pad_l_*/ conv_param->pad_l_;
   // const int pad_right =  /*conv_param->pad_r_*/conv_param->pad_w_;
-  const int pad_up = /*conv_param->pad_u_*/ conv_param->pad_h_;
+  const int pad_up = /*conv_param->pad_u_*/ conv_param->pad_u_;
   // const int pad_down =   /*conv_param->pad_d/*/conv_param->pad_h_;
 
   const int stride_h = conv_param->stride_h_;

mindspore/lite/nnacl/fp32_grad/pooling_grad.c

@@ -13,7 +13,8 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-#include <cstdint>
+#include <stdint.h>
+#include <float.h>
 #include "nnacl/fp32_grad/pooling_grad.h"
 
 void AvgPoolingGrad(const float *input_ptr, float *output_ptr, PoolingParameter *pooling_param) {
@@ -31,33 +32,37 @@ void AvgPoolingGrad(const float *input_ptr, float *output_ptr, PoolingParameter
   int output_batch = pooling_param->output_batch_;
 
   const float *inPtr = NULL;
-  for (int i = 0; i < output_h * output_w * channel * output_batch; i++) output_ptr[i] = 0.0;
+  // for (int i = 0; i < output_h * output_w * channel * output_batch; i++) output_ptr[i] = 0.0;
+  for (int i = 0; i < in_h * in_w * channel * output_batch; i++) output_ptr[i] = 0.0;
 
   float kk = (float)(win_h * win_w);
 
   for (uint16_t ib = 0; ib < output_batch; ib++) {
     float *out;
-    out = &output_ptr[(ib * output_h * output_w)];
-    inPtr = (float *)(&input_ptr[(ib * in_h * in_w)]);
+    // out = &output_ptr[(ib * output_h * output_w)];
+    out = &output_ptr[(ib * in_h * in_w * channel)];
+    // inPtr = (float *)(&input_ptr[(ib * in_h * in_w)]);
+    inPtr = (float *)(&input_ptr[(ib * output_h * output_w * channel)]);
     if (1) {  // in->layout() == Tensor::nhwc)
       // iterate over yt
-      for (uint16_t yh = 0; yh < in_h; yh++) {
-        for (uint16_t yw = 0; yw < in_w; yw++) {
+      for (uint16_t yh = 0; yh < output_h; yh++) {
+        for (uint16_t yw = 0; yw < output_w; yw++) {
           for (uint16_t ic = 0; ic < channel; ic++) {
-            int idx = (yw + yh * in_w) * channel + ic;  // (ic*in_h*in_w) + (in_w*yh) + yw;
+            int idx = (yw + yh * output_w) * channel + ic;  // (ic*in_h*in_w) + (in_w*yh) + yw;
             float delta = inPtr[idx] / kk;
             for (int32_t kh = 0; kh < win_h; kh++) {
               int xh = yh * stride_h + kh - pad_h;
-              if ((xh < 0) || (xh >= output_h)) {
+              if ((xh < 0) || (xh >= in_h)) {
                 continue;
               }
               for (int32_t kw = 0; kw < win_w; kw++) {
                 int xw = yw * stride_w + kw - pad_w;
-                if ((xw < 0) || (xw >= output_w)) {
+                if ((xw < 0) || (xw >= in_w)) {
                   continue;
                 }
-                // out[(ic*output_h*output_w) + (xh*output_w) + xw] += delta;
-                out[(xw + output_w * xh) * channel + ic] += delta;
+
+                // out[(xw + output_w * xh) * channel + ic] += delta;
+                out[(xw + in_w * xh) * channel + ic] += delta;
               }
             }
           }
@@ -66,21 +71,22 @@ void AvgPoolingGrad(const float *input_ptr, float *output_ptr, PoolingParameter
     } else {  // nchw
       for (uint16_t ic = 0; ic < channel; ic++) {
         // iterate over yt
-        for (uint16_t yh = 0; yh < in_h; yh++) {
-          for (uint16_t yw = 0; yw < in_w; yw++) {
-            int idx = (ic * in_h * in_w) + (in_w * yh) + yw;
+        for (uint16_t yh = 0; yh < output_h; yh++) {
+          for (uint16_t yw = 0; yw < output_w; yw++) {
+            int idx = (ic * output_h * output_w) + (output_w * yh) + yw;
             float delta = inPtr[idx] / kk;
             for (int32_t kh = 0; kh < win_h; kh++) {
               int xh = yh * stride_h + kh - pad_h;
-              if ((xh < 0) || (xh >= output_h)) {
+              if ((xh < 0) || (xh >= in_h)) {
                 continue;
               }
               for (int32_t kw = 0; kw < win_w; kw++) {
                 int xw = yw * stride_w + kw - pad_w;
-                if ((xw < 0) || (xw >= output_w)) {
+                if ((xw < 0) || (xw >= in_w)) {
                   continue;
                 }
-                out[(ic * output_h * output_w) + (xh * output_w) + xw] += delta;
+                // out[(ic * output_h * output_w) + (xh * output_w) + xw] += delta;
+                out[(ic * in_h * in_w) + (xh * in_w) + xw] += delta;
               }
             }
           }
@@ -90,7 +96,14 @@ void AvgPoolingGrad(const float *input_ptr, float *output_ptr, PoolingParameter
   }
 }
 
-void MaxPoolingGrad(const float *dy, const int *indices, float *output_ptr, PoolingParameter *pooling_param) {
+void MaxPoolingGrad(const float *input_ptr, const float *dx_ptr, const float *dy_ptr, float *output_ptr,
+                    PoolingParameter *pooling_param) {
+  int stride_w = pooling_param->stride_w_;
+  int stride_h = pooling_param->stride_h_;
+  int pad_w = pooling_param->pad_l_;
+  int pad_h = pooling_param->pad_u_;
+  int win_w = pooling_param->window_w_;
+  int win_h = pooling_param->window_h_;
   int channel = pooling_param->input_channel_;
   int in_w = pooling_param->input_w_;
   int in_h = pooling_param->input_h_;
@@ -98,38 +111,73 @@ void MaxPoolingGrad(const float *dy, const int *indices, float *output_ptr, Pool
   int output_h = pooling_param->output_h_;
   int output_batch = pooling_param->output_batch_;
 
-  int out_img_size =
-    output_h * output_w;  // Emir -- in original code this varible is calculated according to input size ??
-  int ind_img_size = in_h * in_w;
-  // const int w_pad = (output_w + pad_w + pad_w);
+  const float *inPtr;
+  const float *dyPtr;
+
+  for (int i = 0; i < in_h * in_w * channel * output_batch; i++) output_ptr[i] = 0.0;
+
+  for (uint16_t ib = 0; ib < output_batch; ib++) {
+    float *out;
+    out = &output_ptr[(ib * in_h * in_w * channel)];
+    inPtr = (const float *)(&input_ptr[(ib * in_h * in_w * channel)]);
+    dyPtr = (const float *)(&dy_ptr[(ib * output_h * output_w * channel)]);
 
-  for (int i = 0; i < output_h * output_w * channel * output_batch; i++) output_ptr[i] = 0.0;
+    if (1) {  // nhwc
+      for (uint16_t yh = 0; yh < output_h; yh++) {
+        for (uint16_t yw = 0; yw < output_w; yw++) {
+          for (uint16_t ic = 0; ic < channel; ic++) {
+            int idx = (yw + yh * output_w) * channel + ic;
 
-  const float *yt = (const float *)(dy);
-  const int *pos = (const int *)(indices);
-  float *out = NULL;
+            float delta = dyPtr[idx];
+            float max_val = -FLT_MAX;
+            int max_idx = 0;
+            for (int32_t kh = 0; kh < win_h; kh++) {
+              int xh = yh * stride_h + kh - pad_h;
+              if ((xh < 0) || (xh >= in_h)) {
+                continue;
+              }
+              for (int32_t kw = 0; kw < win_w; kw++) {
+                int xw = yw * stride_w + kw - pad_w;
+                if ((xw < 0) || (xw >= in_w)) {
+                  continue;
+                }
 
-  if (1) {  // grads->layout() == Tensor::nhwc)
-    for (int ib = 0; ib < output_batch; ib++) {
-      out = &(output_ptr[ib * output_w * output_w * channel]);
-      for (int ix = 0; ix < ind_img_size; ix++) {
-        for (int cix = 0; cix < channel; cix++) {
-          int idx = (*pos) * channel + cix;
-          out[idx] += *yt;
-          pos++;
-          yt++;
+                if (inPtr[(xw + in_w * xh) * channel + ic] > max_val) {
+                  max_val = inPtr[(xw + in_w * xh) * channel + ic];
+                  max_idx = (xw + in_w * xh) * channel + ic;
+                }
+              }
+            }
+            out[max_idx] += delta;
+          }
         }
       }
-    }
-  } else {
-    for (int ib = 0; ib < output_batch; ib++) {
-      out = &output_ptr[(ib * out_img_size)];
-      for (int cix = 0; cix < channel; cix++) {
-        for (int ix = 0; ix < ind_img_size; ix++) {
-          int idx = cix * output_h * output_w + *pos;  // cord_y*output_w + cord_x;
-          out[idx] += *yt;
-          pos++;
-          yt++;
+    } else {  // nchw
+      for (uint16_t yh = 0; yh < output_h; yh++) {
+        for (uint16_t yw = 0; yw < output_w; yw++) {
+          for (uint16_t ic = 0; ic < channel; ic++) {
+            int idx = (ic * output_h * output_w) + (output_w * yh) + yw;
+            float delta = dyPtr[idx];
+            float max_val = -FLT_MAX;
+            int max_idx = 0;
+            for (int32_t kh = 0; kh < win_h; kh++) {
+              int xh = yh * stride_h + kh - pad_h;
+              if ((xh < 0) || (xh >= in_h)) {
+                continue;
+              }
+              for (int32_t kw = 0; kw < win_w; kw++) {
+                int xw = yw * stride_w + kw - pad_w;
+                if ((xw < 0) || (xw >= in_w)) {
+                  continue;
+                }
+                if (inPtr[(ic * in_h * in_w) + (xh * in_w) + xw] > max_val) {
+                  max_val = inPtr[(ic * in_h * in_w) + (xh * in_w) + xw];
+                  max_idx = (ic * in_h * in_w) + (xh * in_w) + xw;
+                }
+              }
+            }
+            out[max_idx] += delta;
+          }
         }
       }
     }

mindspore/lite/nnacl/fp32_grad/pooling_grad.h

@@ -23,7 +23,9 @@
 extern "C" {
 #endif
 void AvgPoolingGrad(const float *input_ptr, float *output_ptr, PoolingParameter *pooling_param);
-void MaxPoolingGrad(const float *dy, const int *indices_ptr, float *output_ptr, PoolingParameter *pooling_param);
+// void MaxPoolingGrad(const float *dy, const int *indices_ptr, float *output_ptr, PoolingParameter *pooling_param);
+void MaxPoolingGrad(const float *input_ptr, const float *dx_ptr, const float *dy_ptr, float *output_ptr,
+                    PoolingParameter *pooling_param);
 #ifdef __cplusplus
 }
 #endif

mindspore/lite/nnacl/fp32_grad/reduce_grad.c

@@ -13,10 +13,10 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  */
-
+#include <string.h>
 #include "nnacl/fp32_grad/reduce_grad.h"
 
-static inline bool NextIndex(const int num_dims, const int *dims, int *current) {
+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;
@@ -45,10 +45,10 @@ static inline size_t GetOutputOffset(const int num_dims, const int *dims, const
   size_t offset = 0;
   for (int idx = 0; idx < num_dims; ++idx) {
     // if we need to skip this axis
-    bool is_axis = false;
+    int is_axis = 0;
     for (int axis_idx = 0; axis_idx < num_axis; ++axis_idx) {
       if (idx == axes[axis_idx]) {
-        is_axis = true;
+        is_axis = 1;
         break;
       }
     }
@@ -101,10 +101,10 @@ float ReduceMeanAll(const float *src, int size) {
 
 void ReduceSumByAxes(const float *input, const int *input_dims, float *output, const int *output_dims, int num_dims) {
   int num_outputs = 1;
-  int same_shape = true;
+  int same_shape = 1;
   for (int idx = 0; idx < num_dims; ++idx) {
     num_outputs *= output_dims[idx];
-    if (output_dims[idx] != input_dims[idx]) same_shape = false;
+    if (output_dims[idx] != input_dims[idx]) same_shape = 0;
   }
   if (same_shape) {
     memcpy(output, input, num_outputs * sizeof(float));

mindspore/lite/nnacl/fp32_grad/reduce_grad.h

@@ -17,8 +17,7 @@
 #ifndef MINDSPORE_LITE_NNACL_FP32_REDUCE_GRAD_H_
 #define MINDSPORE_LITE_NNACL_FP32_REDUCE_GRAD_H_
 
-#include <cstddef.h>
-#include <algorithm.h>
+#include <stddef.h>
 
 #ifdef __cplusplus
 extern "C" {

mindspore/lite/nnacl/fp32_grad/softmax_grad.h

@@ -20,7 +20,7 @@
 #include "nnacl/op_base.h"
 
 typedef struct SoftmaxCrossEntropyParameter {
-  OpParameter op_parameter;
+  OpParameter op_parameter_;
   int32_t batch_size_;
   unsigned int number_of_classes_;
   int n_dim_;

mindspore/lite/schema/model.fbs

@@ -178,8 +178,8 @@ union PrimitiveType {
     Conv2DGradFilter,
     Conv2DGradInput,
     PoolingGrad,
-    BNGradInput,
-    OptMomentum,
+    BNGrad,
+    ApplyMomentum,
     BiasGrad,
     SoftmaxCrossEntropy,
     AddGrad,
@@ -190,6 +190,7 @@ union PrimitiveType {
     ActivationGrad,
     PriorBox,
     SpaceToBatchND,
+    Depend,
     Return,
     MakeTuple,
     ToFormat,

mindspore/lite/schema/ops.fbs

@@ -149,7 +149,8 @@ table Activation {
     alpha: float = 0.2;
 }
 table ActivationGrad {
-    type: ActivationGradType = 0;
+    type: ActivationType = 0;
+    alpha: float = 0.2;
 }
 
 
@@ -230,6 +231,9 @@ table SoftmaxCrossEntropy {
     axis: [int];
 }
 
+table make_tuple {
+}
+
 
 table PoolingGrad {
     format: Format = 0;
@@ -390,10 +394,11 @@ table DeConv2D {
     hasBias: bool = false;
     activationType: ActivationType = 0;
 }
-table BNGradInput {
+table BNGrad {
     eps : float;
-    channels: int;
+    momentum: float;
 }
+
 table Scale {
     axis: int;
 }
@@ -841,7 +846,10 @@ table SquaredDifference {
 table TupleGetItem {
 }
 
-table OptMomentum {
+table ApplyMomentum {
+    gradientScale: float;
+    useLocking: bool;
+    useNesterov: bool;
 }
 
 
@@ -884,6 +892,10 @@ table ToFormat {
     dstT: int;
 }
 
+
+table Depend {
+}
+
 table Return {
 }
 

mindspore/lite/src/CMakeLists.txt

@@ -27,7 +27,7 @@ set(LITE_SRC
     )
 
 if (SUPPORT_GPU)
-  set(LITE_SRC
+set(LITE_SRC
       ${LITE_SRC}
           ${CMAKE_CURRENT_SOURCE_DIR}/runtime/kernel/opencl/subgraph_opencl_kernel.cc
           ${CMAKE_CURRENT_SOURCE_DIR}/runtime/kernel/opencl/utils.cc
@@ -36,6 +36,24 @@ if (SUPPORT_GPU)
           ${CMAKE_CURRENT_SOURCE_DIR}/runtime/opencl/opencl_runtime.cc
           ${CMAKE_CURRENT_SOURCE_DIR}/runtime/opencl/opencl_wrapper.cc
       )
+endif()
+
+
+if (SUPPORT_TRAIN)
+    set(ANF_SRC
+             ${ANF_SRC}
+
+            )
+    set(PASS_SRC)
+    set(LITE_SRC
+            ${LITE_SRC}
+            ${ANF_SRC}
+        #   ${CMAKE_CURRENT_SOURCE_DIR}/train/ops/train_ops.cc
+            ${CMAKE_CURRENT_SOURCE_DIR}/train/train_populate_parameter.cc
+            ${CMAKE_CURRENT_SOURCE_DIR}/train/train_session.cc
+            ${CMAKE_CURRENT_SOURCE_DIR}/lite_session.cc
+            )
+            
 endif ()
 
 file(GLOB_RECURSE C_OPS_SRC ${CMAKE_CURRENT_SOURCE_DIR}/ops/*.cc)

mindspore/lite/src/common/file_utils.cc

@@ -110,6 +110,7 @@ int CompareOutputData(float *output_data, float *correct_data, int data_size) {
     }
   }
   error /= data_size;
+
   if (error > 0.0001) {
     printf("has accuracy error!\n");
     printf("%f\n", error);
@@ -118,12 +119,14 @@ int CompareOutputData(float *output_data, float *correct_data, int data_size) {
   return 0;
 }
 
-void CompareOutput(float *output_data, std::string file_path) {
+int  CompareOutput(float *output_data, std::string file_path) {
   size_t output_size;
   auto ground_truth = reinterpret_cast<float *>(mindspore::lite::ReadFile(file_path.c_str(), &output_size));
   size_t output_num = output_size / sizeof(float);
   printf("output num : %zu\n", output_num);
-  CompareOutputData(output_data, ground_truth, output_num);
+  int res = CompareOutputData(output_data, ground_truth, output_num);
+  delete [] ground_truth;
+  return res;
 }
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/common/file_utils.h

@@ -47,7 +47,7 @@ void WriteToTxt(const std::string& file_path, void *data, size_t element_size) {
 int WriteToBin(const std::string& file_path, void *data, size_t size);
 
 int CompareOutputData(float *output_data, float *correct_data, int data_size);
-void CompareOutput(float *output_data, std::string file_path);
+int  CompareOutput(float *output_data, std::string file_path);
 
 std::string GetAndroidPackageName();
 std::string GetAndroidPackagePath();

mindspore/lite/src/common/file_utils_ext.cc

@@ -47,7 +47,9 @@ int CompareRelativeOutput(float *output_data, std::string file_path) {
   auto ground_truth = reinterpret_cast<float *>(mindspore::lite::ReadFile(file_path.c_str(), &output_size));
   size_t output_num = output_size / sizeof(float);
   std::cout << "output num : " << output_num << "\n";
-  return CompareOutputRelativeData(output_data, ground_truth, output_num);
+  int res = CompareOutputRelativeData(output_data, ground_truth, output_num);
+  delete [] ground_truth;
+  return res;
 }
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/executor.cc

@@ -39,6 +39,10 @@ int Executor::Run(std::vector<tensor::Tensor *> &in_tensors, std::vector<tensor:
     }
   }
   kernel::LiteKernelUtil::InitTensorRefCount(kernels);
+  for (auto out_tensor : out_tensors) {  // increase RefCount of output tensors, such that Run will not free them
+    out_tensor->SetRefCount(out_tensor->RefCount() + 1);
+  }
+
   for (auto *kernel : kernels) {
     MS_ASSERT(nullptr != kernel);
 
@@ -48,6 +52,8 @@ int Executor::Run(std::vector<tensor::Tensor *> &in_tensors, std::vector<tensor:
         MS_LOG(ERROR) << "run kernel before_callback failed, name: " << kernel->name();
       }
     }
+    // JBDEBUG
+    // std::cout << "executing kernel " << kernel->name() << "\n";
     auto ret = kernel->Run();
     if (0 != ret) {
       MS_LOG(ERROR) << "run kernel failed, name: " << kernel->name();

mindspore/lite/src/lite_kernel.h

@@ -27,7 +27,6 @@
 #include "src/ir/tensor.h"
 #include "include/errorcode.h"
 
-
 // using mindspore::kernel::AddressPtr;
 namespace mindspore::kernel {
 using mindspore::lite::RET_ERROR;

mindspore/lite/src/model.cc

@@ -112,11 +112,11 @@ int ModelImpl::BuildOps() {
 
 Model *Model::Import(const char *model_buf, size_t size) {
   auto model = new Model();
+  model->model_impl_ = ModelImpl::Import(model_buf, size);
   if (model_buf == nullptr) {
     MS_LOG(ERROR) << "model buf is null";
     return nullptr;
   }
-  model->model_impl_ = ModelImpl::Import(model_buf, size);
   if (model->model_impl_ == nullptr) {
     MS_LOG(ERROR) << "model impl is null";
     return nullptr;

mindspore/lite/src/ops/activation_grad.cc

@@ -20,11 +20,11 @@ namespace mindspore {
 namespace lite {
 #ifdef PRIMITIVE_WRITEABLE
 int ActivationGrad::GetType() const { return this->primitive_->value.AsActivationGrad()->type; }
-
+float ActivationGrad::GetAlpha() const { return this->primitive_->value.AsActivationGrad()->alpha; }
 void ActivationGrad::SetType(int type) {
-  this->primitive_->value.AsActivationGrad()->type = (schema::ActivationGradType)type;
+  this->primitive_->value.AsActivationGrad()->type = (schema::ActivationType)type;
 }
-
+void ActivationGrad::SetAlpha(float alpha) { this->primitive_->value.AsActivationGrad()->alpha = alpha; }
 #else
 int ActivationGrad::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
   MS_ASSERT(nullptr != primitive);
@@ -40,7 +40,7 @@ int ActivationGrad::UnPackToFlatBuilder(const schema::Primitive *primitive, flat
   return RET_OK;
 }
 int ActivationGrad::GetType() const { return this->primitive_->value_as_ActivationGrad()->type(); }
-
+float ActivationGrad::GetAlpha() const { return this->primitive_->value_as_ActivationGrad()->alpha(); }
 #endif
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/ops/activation_grad.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_ACTIVATION_GRAD_H_
-#define LITE_MINDSPORE_LITE_C_OPS_ACTIVATION_GRAD_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_ACTIVATION_GRAD_H_
+#define MINDSPORE_LITE_SRC_OPS_ACTIVATION_GRAD_H_
 
 #include <vector>
 #include <set>
@@ -32,13 +32,15 @@ class ActivationGrad : public PrimitiveC {
   ActivationGrad() = default;
   explicit ActivationGrad(schema::PrimitiveT *primitive) : PrimitiveC(primitive) {}
   void SetType(int type);
+  void SetAlpha(float alpha);
 #else
   ActivationGrad() = default;
 
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
 #endif
   int GetType() const;
+  float GetAlpha() const;
 };
 }  // namespace lite
 }  // namespace mindspore
-#endif  // LITE_MINDSPORE_LITE_C_OPS_ACTIVATION_GRAD_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_ACTIVATION_GRAD_H_

mindspore/lite/src/ops/apply_momentum.cc

@@ -0,0 +1,64 @@
+/**
+ * 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/apply_momentum.h"
+namespace mindspore {
+namespace lite {
+
+
+#ifdef PRIMITIVE_WRITEABLE
+
+#else
+int ApplyMomentum::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
+  MS_ASSERT(nullptr != primitive);
+  MS_ASSERT(nullptr != fbb);
+  auto attr = primitive->value_as_ApplyMomentum();
+  if (attr == nullptr) {
+    MS_LOG(ERROR) << "value_as_ApplyMomentum return nullptr";
+    return RET_ERROR;
+  }
+  auto val_offset = schema::CreateApplyMomentum(*fbb);
+  auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_ActivationGrad, val_offset.o);
+  fbb->Finish(prim_offset);
+  return RET_OK;
+}
+#endif
+
+int ApplyMomentum::InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) {
+  if (5 != inputs.size()) {
+    MS_LOG(ERROR) << "ApplyMomentum should have at 5 input tensors";
+    return RET_ERROR;
+  }
+  // if (outputs.empty()) {
+  //  MS_LOG(ERROR) << "ApplyMomentumCPUKernel error input output size!";
+  //  return RET_ERROR;
+  // }
+
+  if (inputs[0]->ElementsNum() != inputs[1]->ElementsNum() || inputs[0]->ElementsNum() != inputs[3]->ElementsNum() ||
+      inputs[2]->ElementsNum() != 1 || inputs[4]->ElementsNum() != 1) {
+    MS_LOG(ERROR) << "error input data size!";
+    return RET_ERROR;
+  }
+  if (!outputs.empty()) {
+    auto *out = outputs.front();
+    MS_ASSERT(out != nullptr);
+    out->set_data_type(inputs[0]->data_type());
+    out->SetFormat(inputs[0]->GetFormat());
+  }
+
+  return RET_OK;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/ops/apply_momentum.h

@@ -0,0 +1,44 @@
+/**
+ * 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_APPLY_MOMENTUM_H_
+#define MINDSPORE_LITE_SRC_OPS_APPLY_MOMENTUM_H_
+
+#include <vector>
+#include <set>
+#include <cmath>
+#include "ir/dtype/type_id.h"
+#include "src/ops/primitive_c.h"
+
+namespace mindspore {
+namespace lite {
+class ApplyMomentum : public PrimitiveC {
+ public:
+#ifdef PRIMITIVE_WRITEABLE
+  MS_DECLARE_PARENT(ApplyMomentum, PrimitiveC);
+  ApplyMomentum() = default;
+  explicit ApplyMomentum(schema::PrimitiveT *primitive) : PrimitiveC(primitive) {}
+#else
+  ApplyMomentum() = default;
+
+  int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
+#endif
+  int InferShape(std::vector<lite::tensor::Tensor *> inputs_, std::vector<lite::tensor::Tensor *> outputs_) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_OPS_APPLY_MOMENTUM_H_

mindspore/lite/src/ops/arithmetic_grad.cc

@@ -0,0 +1,108 @@
+/**
+ * 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/arithmetic_grad.h"
+#include "include/errorcode.h"
+#include "utils/log_adapter.h"
+#include "src/ir/tensor.h"
+
+namespace mindspore {
+namespace lite {
+int ArithmeticGrad::InferShape(std::vector<lite::tensor::Tensor *> inputs_,
+                               std::vector<lite::tensor::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 dy = inputs_[0];
+  auto x1 = inputs_[1];
+  auto x2 = 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);
+
+  auto inShape0 = x1->shape();
+  auto inShape1 = x2->shape();
+  auto outShape = dy->shape();
+
+  if ((Type() == schema::PrimitiveType_AddGrad) || (Type() == schema::PrimitiveType_SubGrad)) {
+    ndim_ = outShape.size();
+    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];
+    }
+  } else {
+    // if (inShape0.size() < inShape1.size())
+    if (dx1->ElementsNum() < dx2->ElementsNum()) {
+      ndim_ = inShape1.size();
+      auto fillDimNum = inShape1.size() - inShape0.size();  // This will not work for batch!
+      int j = 0;
+      for (unsigned int i = 0; i < inShape1.size(); i++) {
+        if (i < fillDimNum) {
+          x2_shape_[i] = 1;
+        } else {
+          x2_shape_[i] = inShape0[j++];
+        }
+        x1_shape_[i] = inShape1[i];
+        dy_shape_[i] = outShape[i];
+      }
+    } else if (dx2->ElementsNum() < dx1->ElementsNum()) {  // if (inShape0.size() > inShape1.size())
+      ndim_ = inShape0.size();
+      broadcasting_ = true;
+      ndim_ = inShape0.size();
+      int j = 0;
+      auto fillDimNum = inShape0.size() - inShape1.size();
+      for (unsigned int i = 0; i < inShape0.size(); i++) {
+        if (i < fillDimNum) {
+          x2_shape_[i] = 1;
+        } else {
+          x2_shape_[i] = inShape1[j++];
+        }
+        x1_shape_[i] = inShape0[i];
+        dy_shape_[i] = outShape[i];
+      }
+    } else {
+      broadcasting_ = false;
+      for (unsigned int i = 0; i < inShape0.size(); i++) {
+        x2_shape_[i] = inShape1[i];
+        x1_shape_[i] = inShape0[i];
+        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

mindspore/lite/src/ops/arithmetic_grad.h

@@ -0,0 +1,58 @@
+/**
+ * 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_ARITHMETIC_GRAD_H_
+#define MINDSPORE_LITE_SRC_OPS_ARITHMETIC_GRAD_H_
+
+#include <vector>
+#include <set>
+#include <cmath>
+#include "ir/dtype/type_id.h"
+#include "src/ops/primitive_c.h"
+
+namespace mindspore {
+namespace lite {
+class ArithmeticGrad : public PrimitiveC {
+ public:
+#ifdef PRIMITIVE_WRITEABLE
+  MS_DECLARE_PARENT(ArithmeticGrad, PrimitiveC);
+  ArithmeticGrad() = default;
+  explicit ArithmeticGrad(schema::PrimitiveT *primitive) : PrimitiveC(primitive) {}
+#else
+  //  explicit Arithmetic(schema::Primitive *primitive) : PrimitiveC(primitive) {}
+  ArithmeticGrad() = default;
+  int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override {
+    return RET_ERROR;
+  }
+#endif
+  int InferShape(std::vector<lite::tensor::Tensor *> inputs_, std::vector<lite::tensor::Tensor *> outputs_) override;
+  bool Broadcasting() { return this->broadcasting_; }
+  int NDims() { return this->ndim_; }
+  std::vector<int> dyShape() { return this->dy_shape_; }
+  std::vector<int> x1Shape() { return this->x1_shape_; }
+  std::vector<int> x2Shape() { return this->x2_shape_; }
+
+ protected:
+  bool broadcasting_ = false;
+  int ndim_;
+  std::vector<int> dy_shape_;
+  std::vector<int> x1_shape_;
+  std::vector<int> x2_shape_;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_OPS_ARITHMETIC_GRAD_H_

mindspore/lite/src/ops/bias_grad.cc

@@ -48,6 +48,32 @@ std::vector<int> BiasGrad::GetAxis() const {
   return std::vector<int>(fb_vector->begin(), fb_vector->end());
 }
 
+int BiasGrad::InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) {
+  if (1 != inputs.size()) {
+    MS_LOG(ERROR) << "BiasGrad should have one input";
+    return RET_ERROR;
+  }
+  if (1 != outputs.size()) {
+    MS_LOG(ERROR) << "BiasGrad should have one output";
+    return RET_ERROR;
+  }
+  auto *in0 = inputs.front();
+  auto *out = outputs.front();
+  MS_ASSERT(in0 != nullptr);
+  MS_ASSERT(out != nullptr);
+  auto inshape = in0->shape();
+  int ndim = inshape.size();
+  for (int i = 0; i < ndim - 1; i++) {
+    inshape[i] = 1;
+  }
+  out->set_shape(inshape);
+  out->set_data_type(in0->data_type());
+  out->SetFormat(in0->GetFormat());
+
+  return RET_OK;
+}
+
+
 #endif
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/ops/bias_grad.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_BIAS_GRAD_H_
-#define LITE_MINDSPORE_LITE_C_OPS_BIAS_GRAD_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_BIAS_GRAD_H_
+#define MINDSPORE_LITE_SRC_OPS_BIAS_GRAD_H_
 
 #include <vector>
 #include <set>
@@ -38,10 +38,11 @@ class BiasGrad : public PrimitiveC {
   BiasGrad() = default;
 
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
+  int InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) override;
 #endif
   std::vector<int> GetAxis() const;
 };
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_BIAS_GRAD_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_BIAS_GRAD_H_

mindspore/lite/src/ops/bn_grad_input.cc → mindspore/lite/src/ops/bn_grad.cc

@@ -14,33 +14,33 @@
  * limitations under the License.
  */
 
-#include "src/ops/bn_grad_input.h"
+#include "src/ops/bn_grad.h"
 
 namespace mindspore {
 namespace lite {
 #ifdef PRIMITIVE_WRITEABLE
-float BNGradInput::GetEps() const { return this->primitive_->value.AsBNGradInput()->eps; }
-int BNGradInput::GetChannels() const { return this->primitive_->value.AsBNGradInput()->channels; }
+float BNGrad::GetEps() const { return this->primitive_->value.AsBNGrad()->eps; }
+float BNGrad::GetMomentum() const { return this->primitive_->value.AsBNGrad()->momentum; }
 
-void BNGradInput::SetEps(float eps) { this->primitive_->value.AsBNGradInput()->eps = eps; }
-void BNGradInput::SetChannels(int channels) { this->primitive_->value.AsBNGradInput()->channels = channels; }
+void BNGrad::SetEps(float eps) { this->primitive_->value.AsBNGrad()->eps = eps; }
+void BNGrad::SetMomentum(float momentum) { this->primitive_->value.AsBNGrad()->momentum = momentum; }
 
 #else
-int BNGradInput::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
+int BNGrad::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) {
   MS_ASSERT(nullptr != primitive);
   MS_ASSERT(nullptr != fbb);
-  auto attr = primitive->value_as_BNGradInput();
+  auto attr = primitive->value_as_BNGrad();
   if (attr == nullptr) {
     MS_LOG(ERROR) << "value_as_BNGradInput return nullptr";
     return RET_ERROR;
   }
-  auto val_offset = schema::CreateBNGradInput(*fbb, attr->eps(), attr->channels());
-  auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_BNGradInput, val_offset.o);
+  auto val_offset = schema::CreateBNGrad(*fbb, attr->eps(), attr->momentum());
+  auto prim_offset = schema::CreatePrimitive(*fbb, schema::PrimitiveType_BNGrad, val_offset.o);
   fbb->Finish(prim_offset);
   return RET_OK;
 }
-float BNGradInput::GetEps() const { return this->primitive_->value_as_BNGradInput()->eps(); }
-int BNGradInput::GetChannels() const { return this->primitive_->value_as_BNGradInput()->channels(); }
+float BNGrad::GetEps() const { return this->primitive_->value_as_BNGrad()->eps(); }
+float BNGrad::GetMomentum() const { return this->primitive_->value_as_BNGrad()->momentum(); }
 
 #endif
 }  // namespace lite

mindspore/lite/src/ops/bn_grad_input.h → mindspore/lite/src/ops/bn_grad.h

@@ -25,21 +25,20 @@
 
 namespace mindspore {
 namespace lite {
-class BNGradInput : public PrimitiveC {
+class BNGrad : public PrimitiveC {
  public:
 #ifdef PRIMITIVE_WRITEABLE
-  MS_DECLARE_PARENT(BNGradInput, PrimitiveC);
-  BNGradInput() = default;
-  explicit BNGradInput(schema::PrimitiveT *primitive) : PrimitiveC(primitive) {}
+  MS_DECLARE_PARENT(BNGrad, PrimitiveC);
+  BNGrad() = default;
+  explicit BNGrad(schema::PrimitiveT *primitive) : PrimitiveC(primitive) {}
   void SetEps(float eps);
-  void SetChannels(int channels);
+  void SetMomentum(float momentum);
 #else
-  BNGradInput() = default;
-
+  BNGrad() = default;
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
 #endif
   float GetEps() const;
-  int GetChannels() const;
+  float GetMomentum() const;
 };
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/ops/conv2d_grad_filter.cc

@@ -105,5 +105,47 @@ int Conv2DGradFilter::GetActivationType() const {
 }
 
 #endif
+
+int Conv2DGradFilter::InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) {
+  if (3 != inputs.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad Filter should have 3 inputs";
+    return RET_ERROR;
+  }
+  if (1 != outputs.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad Filter should have one output";
+    return RET_ERROR;
+  }
+
+  auto *in0 = inputs.at(0);
+  auto *in = inputs.at(2);
+  MS_ASSERT(out != nullptr);
+
+  std::vector<int> output_shape;
+  int *out_shape = reinterpret_cast<int *>(in->Data());
+  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_data_type(in0->data_type());
+  out->SetFormat(in0->GetFormat());
+
+  return RET_OK;
+}
+
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/ops/conv2d_grad_filter.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_CONV2_D_GRAD_FILTER_H_
-#define LITE_MINDSPORE_LITE_C_OPS_CONV2_D_GRAD_FILTER_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_CONV2D_GRAD_FILTER_H_
+#define MINDSPORE_LITE_SRC_OPS_CONV2D_GRAD_FILTER_H_
 
 #include <vector>
 #include <set>
@@ -53,6 +53,7 @@ class Conv2DGradFilter : public PrimitiveC {
 
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
 #endif
+  int InferShape(std::vector<lite::tensor::Tensor *> inputs_, std::vector<lite::tensor::Tensor *> outputs_) override;
   int GetFormat() const;
   int GetGroup() const;
   int GetChannelIn() const;
@@ -74,4 +75,4 @@ class Conv2DGradFilter : public PrimitiveC {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_CONV2_D_GRAD_FILTER_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_CONV2D_GRAD_FILTER_H_

mindspore/lite/src/ops/conv2d_grad_input.cc

@@ -103,5 +103,46 @@ int Conv2DGradInput::GetActivationType() const {
 }
 
 #endif
+
+int Conv2DGradInput::InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) {
+  if (3 != inputs.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad Input should have 3 inputs";
+    return RET_ERROR;
+  }
+  if (1 != outputs.size()) {
+    MS_LOG(ERROR) << "Conv2d Grad input should have one output";
+    return RET_ERROR;
+  }
+
+  auto *in0 = inputs.at(0);
+  auto *in = inputs.at(2);
+  MS_ASSERT(out != nullptr);
+
+  std::vector<int> output_shape;
+  int *out_shape = reinterpret_cast<int *>(in->Data());
+  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_data_type(in0->data_type());
+  out->SetFormat(in0->GetFormat());
+
+  return RET_OK;
+}
+
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/ops/conv2d_grad_input.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_CONV2_D_GRAD_INPUT_H_
-#define LITE_MINDSPORE_LITE_C_OPS_CONV2_D_GRAD_INPUT_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_CONV2D_GRAD_INPUT_H_
+#define MINDSPORE_LITE_SRC_OPS_CONV2D_GRAD_INPUT_H_
 
 #include <vector>
 #include <set>
@@ -53,6 +53,7 @@ class Conv2DGradInput : public PrimitiveC {
 
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
 #endif
+  int InferShape(std::vector<lite::tensor::Tensor *> inputs_, std::vector<lite::tensor::Tensor *> outputs_) override;
   int GetFormat() const;
   int GetGroup() const;
   int GetChannelIn() const;
@@ -74,4 +75,4 @@ class Conv2DGradInput : public PrimitiveC {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_CONV2_D_GRAD_INPUT_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_CONV2D_GRAD_INPUT_H_

mindspore/lite/src/ops/dedepthwise_conv2d.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_DE_DEPTHWISE_CONV2_D_H_
-#define LITE_MINDSPORE_LITE_C_OPS_DE_DEPTHWISE_CONV2_D_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_DEDEPTHWISE_CONV2D_H_
+#define MINDSPORE_LITE_SRC_OPS_DEDEPTHWISE_CONV2D_H_
 
 #include <vector>
 #include <set>
@@ -84,4 +84,4 @@ class DeDepthwiseConv2D : public PrimitiveC {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_DE_DEPTHWISE_CONV2_D_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_DEDEPTHWISE_CONV2D_H_

mindspore/lite/src/ops/depthwise_conv2d.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_DEPTHWISE_CONV2_D_H_
-#define LITE_MINDSPORE_LITE_C_OPS_DEPTHWISE_CONV2_D_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_DEPTHWISE_CONV2D_H_
+#define MINDSPORE_LITE_SRC_OPS_DEPTHWISE_CONV2D_H_
 
 #include <vector>
 #include <set>
@@ -94,4 +94,4 @@ class DepthwiseConv2D : public PrimitiveC {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_DEPTHWISE_CONV2_D_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_DEPTHWISE_CONV2D_H_

mindspore/lite/src/ops/make_tuple.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_SRC_OPS_MAKE_TUPLE_H_
-#define LITE_MINDSPORE_LITE_SRC_OPS_MAKE_TUPLE_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_MAKE_TUPLE_H_
+#define MINDSPORE_LITE_SRC_OPS_MAKE_TUPLE_H_
 #include <vector>
 #include "src/ops/primitive_c.h"
 
@@ -37,4 +37,4 @@ class MakeTuple : public PrimitiveC {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_SRC_OPS_MAKE_TUPLE_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_MAKE_TUPLE_H_

mindspore/lite/src/ops/pooling_grad.cc

@@ -86,5 +86,52 @@ int PoolingGrad::UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuf
   return RET_OK;
 }
 #endif
+
+int PoolingGrad::InferShape(std::vector<tensor::Tensor *> inputs_, std::vector<tensor::Tensor *> outputs_) {
+  MS_ASSERT(this->primitive != nullptr);
+  auto input = inputs_.at(0);
+  MS_ASSERT(input != nullptr);
+  int input_h = input->shape().at(1);
+  int input_w = input->shape().at(2);
+
+  auto window_h = GetWindowH();
+  auto window_w = GetWindowW();
+  if (GetGlobal()) {
+    window_h = input_h;
+    window_w = input_w;
+  }
+
+  pad_l_ = GetPadLeft();
+  pad_u_ = GetPadUp();
+  pad_d_ = GetPadDown();
+  pad_r_ = GetPadRight();
+  if (GetPadMode() == schema::PadMode_SAME) {
+    int output_w = std::ceil(static_cast<float>(input_w) / static_cast<float>(GetStrideW()));
+    int output_h = std::ceil(static_cast<float>(input_h) / static_cast<float>(GetStrideH()));
+    auto pad_h_all = ((output_h - 1) * GetStrideH() + (window_h - 1) + 1 - input_h);
+    auto pad_w_all = ((output_w - 1) * GetStrideW() + (window_w - 1) + 1 - input_w);
+    if (pad_h_all < 0) {
+      pad_u_ = pad_d_ = 0;
+    } else {
+      pad_u_ = pad_h_all / 2;
+      pad_d_ = pad_h_all - pad_u_;
+    }
+    if (pad_w_all < 0) {
+      pad_l_ = pad_r_ = 0;
+    } else {
+      pad_l_ = pad_w_all / 2;
+      pad_r_ = pad_w_all - pad_l_;
+    }
+  }
+  auto grad_output = outputs_.at(0);
+  // todo: fmk type
+  auto output_shape = input->shape();
+  grad_output->set_shape(output_shape);
+  grad_output->set_data_type(input->data_type());
+  // todo: temp fix
+  grad_output->SetFormat(input->GetFormat());
+  return RET_OK;
+}
+
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/ops/pooling_grad.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_POOLING_GRAD_H_
-#define LITE_MINDSPORE_LITE_C_OPS_POOLING_GRAD_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_POOLING_GRAD_H_
+#define MINDSPORE_LITE_SRC_OPS_POOLING_GRAD_H_
 
 #include <vector>
 #include <set>
@@ -49,6 +49,7 @@ class PoolingGrad : public PrimitiveC {
 
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
 #endif
+  int InferShape(std::vector<lite::tensor::Tensor *> inputs_, std::vector<lite::tensor::Tensor *> outputs_) override;
   int GetFormat() const;
   int GetPoolingMode() const;
   bool GetGlobal() const;
@@ -62,8 +63,14 @@ class PoolingGrad : public PrimitiveC {
   int GetPadLeft() const;
   int GetPadRight() const;
   int GetRoundMode() const;
+
+ protected:
+  int pad_u_ = 0;
+  int pad_d_ = 0;
+  int pad_l_ = 0;
+  int pad_r_ = 0;
 };
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_POOLING_GRAD_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_POOLING_GRAD_H_

mindspore/lite/src/ops/power_grad.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_POWER_GRAD_H_
-#define LITE_MINDSPORE_LITE_C_OPS_POWER_GRAD_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_POWER_GRAD_H_
+#define MINDSPORE_LITE_SRC_OPS_POWER_GRAD_H_
 
 #include <vector>
 #include <set>
@@ -46,4 +46,4 @@ class PowerGrad : public PrimitiveC {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_POWER_GRAD_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_POWER_GRAD_H_

mindspore/lite/src/ops/primitive_c.cc

@@ -125,6 +125,21 @@
 #ifdef PRIMITIVE_WRITEABLE
 #include "tools/converter/quantizer/quantize_util.h"
 #endif
+
+#ifdef SUPPORT_TRAIN
+#include "src/ops/activation_grad.h"
+#include "src/ops/apply_momentum.h"
+#include "src/ops/bias_grad.h"
+#include "src/ops/pooling_grad.h"
+#include "src/ops/conv2d_grad_filter.h"
+#include "src/ops/conv2d_grad_input.h"
+#include "src/ops/power_grad.h"
+#include "src/ops/softmax_cross_entropy.h"
+#include "src/ops/bn_grad.h"
+#include "src/ops/arithmetic_grad.h"
+#endif
+
+
 namespace mindspore {
 namespace lite {
 #ifdef PRIMITIVE_WRITEABLE
@@ -353,6 +368,22 @@ std::shared_ptr<PrimitiveC> PrimitiveC::UnPackFromPrimitive(const Primitive &pri
     return NewPrimitiveC<TupleGetItem>(prim, inputs, quantType);
   } else if (op_type == "Softmax") {
     return NewPrimitiveC<SoftMax>(prim, inputs, quantType);
+#ifdef SUPPORT_TRAIN0
+  } else if ((op_type == "ReluGrad" || op_type == "Relu6Grad" || op_type == "SigmoidGrad")) {
+    return NewPrimitiveC<ActivationGrad>(prim, inputs, quantType);
+  } else if ((op_type == "MaxPoolGrad") || (op_type == "MeanPoolGrad")) {
+    return NewPrimitiveC<PoolingGrad>(prim, inputs, quantType);
+  } else if (op_type == "Conv2DBackpropFilter") {
+    return NewPrimitiveC<Conv2DGradFilter>(prim, inputs, quantType);
+  } else if (op_type == "Conv2DBackpropInput") {
+    return NewPrimitiveC<Conv2DGradInput>(prim, inputs, quantType);
+  } else if (op_type == "BiasAddGrad") {
+    return NewPrimitiveC<BiasGrad>(prim, inputs, quantType);
+  } else if (op_type == "ApplyMomentum") {
+    return NewPrimitiveC<ApplyMomentum>(prim, inputs, quantType);
+  } else if (op_type == "BatchNormGrad") {
+    return NewPrimitiveC<BNGrad>(prim, inputs, quantType);
+#endif
   } else {
     MS_LOG(ERROR) << "Unsupported primitive type in UnPackFromPrimitive : " << op_type;
     return nullptr;
@@ -565,6 +596,32 @@ PrimitiveC *PrimitiveC::UnPackFromSchemaPrimitiveT(mindspore::schema::PrimitiveT
       return new SparseToDense(primitive);
     case schema::PrimitiveType_DetectionPostProcess:
       return new DetectionPostProcess(primitive);
+
+#ifdef SUPPORT_TRAIN
+    case schema::PrimitiveType_ActivationGrad:
+      return new ActivationGrad(primitive);
+    case schema::PrimitiveType_PoolingGrad:
+      return new PoolingGrad(primitive);
+    case schema::PrimitiveType_Conv2DGradFilter:
+      return new Conv2DGradFilter(primitive);
+    case schema::PrimitiveType_Conv2DGradInput:
+      return new Conv2DGradInput(primitive);
+    case schema::PrimitiveType_BiasGrad:
+      return new BiasGrad(primitive);
+    case schema::PrimitiveType_ApplyMomentum:
+      return new ApplyMomentum(primitive);
+    case schema::PrimitiveType_BNGrad:
+      return new BNGrad(primitive);
+    case schema::PrimitiveType_AddGrad:
+      return new ArithmeticGrad(primitive);
+    case schema::PrimitiveType_SubGrad:
+      return new ArithmeticGrad(primitive);
+    case schema::PrimitiveType_MulGrad:
+      return new ArithmeticGrad(primitive);
+    case schema::PrimitiveType_DivGrad:
+      return new ArithmeticGrad(primitive);
+#endif
+
     default:
       MS_LOG(ERROR) << "Unsupported primitive type in UnPackFromSchemaPrimitiveT : "
                     << schema::EnumNamePrimitiveType(op_type);
@@ -779,6 +836,31 @@ PrimitiveC *PrimitiveC::UnPackFromSchemaPrimitive(const schema::Primitive *primi
       return NewPrimitiveC<SparseToDense>(primitive);
     case schema::PrimitiveType_DetectionPostProcess:
       return NewPrimitiveC<DetectionPostProcess>(primitive);
+
+#ifdef SUPPORT_TRAIN
+    case schema::PrimitiveType_ActivationGrad:
+      return NewPrimitiveC<ActivationGrad>(primitive);
+    case schema::PrimitiveType_PoolingGrad:
+      return NewPrimitiveC<PoolingGrad>(primitive);
+    case schema::PrimitiveType_Conv2DGradFilter:
+      return NewPrimitiveC<Conv2DGradFilter>(primitive);
+    case schema::PrimitiveType_Conv2DGradInput:
+      return NewPrimitiveC<Conv2DGradInput>(primitive);
+    case schema::PrimitiveType_BiasGrad:
+      return NewPrimitiveC<BiasGrad>(primitive);
+    case schema::PrimitiveType_ApplyMomentum:
+      return NewPrimitiveC<ApplyMomentum>(primitive);
+    case schema::PrimitiveType_BNGrad:
+      return NewPrimitiveC<BNGrad>(primitive);
+    case schema::PrimitiveType_AddGrad:
+      return NewPrimitiveC<ArithmeticGrad>(primitive);
+    case schema::PrimitiveType_SubGrad:
+      return NewPrimitiveC<ArithmeticGrad>(primitive);
+    case schema::PrimitiveType_MulGrad:
+      return NewPrimitiveC<ArithmeticGrad>(primitive);
+    case schema::PrimitiveType_DivGrad:
+     return NewPrimitiveC<ArithmeticGrad>(primitive);
+#endif
     default:
       MS_LOG(ERROR) << "Unsupported primitive type in UnPackFromSchemaPrimitive : "
                     << schema::EnumNamePrimitiveType(op_type);

mindspore/lite/src/ops/reduce.cc

@@ -115,7 +115,7 @@ constexpr size_t kInputSize = 1;
 constexpr size_t kOutputSize = 1;
 }  // namespace
 int Reduce::InferShape(std::vector<tensor::Tensor *> inputs_, std::vector<tensor::Tensor *> outputs_) {
-  if (inputs_.size() != kInputSize || outputs_.size() != kOutputSize) {
+  if (inputs_.size() < kInputSize || outputs_.size() != kOutputSize) {
     return RET_ERROR;
   }
   auto input = inputs_.front();

mindspore/lite/src/ops/reshape.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_RESHAPE_H_
-#define LITE_MINDSPORE_LITE_C_OPS_RESHAPE_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_RESHAPE_H_
+#define MINDSPORE_LITE_SRC_OPS_RESHAPE_H_
 
 #include <vector>
 #include <set>
@@ -50,4 +50,4 @@ class Reshape : public PrimitiveC {
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_RESHAPE_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_RESHAPE_H_

mindspore/lite/src/ops/softmax_cross_entropy.cc

@@ -51,5 +51,31 @@ int SoftmaxCrossEntropy::UnPackToFlatBuilder(const schema::Primitive *primitive,
   return RET_OK;
 }
 #endif
+
+int SoftmaxCrossEntropy::InferShape(std::vector<tensor::Tensor *> inputs, std::vector<tensor::Tensor *> outputs) {
+  if (1 > outputs.size()) {
+    MS_LOG(ERROR) << "SoftmaxCrossEntropy should have at least one output";
+    return RET_ERROR;
+  }
+  auto *in0 = inputs.front();
+  MS_ASSERT(in0 != nullptr);
+  auto *out = outputs.front();
+  MS_ASSERT(out != nullptr);
+
+  std::vector<int> outshape;
+  outshape.push_back(1);
+  out->set_shape(outshape);
+  out->set_data_type(in0->data_type());
+
+  if (1 < outputs.size()) {
+    auto *grads = outputs.at(1);
+    MS_ASSERT(grads != nullptr);
+    grads->set_shape(in0->shape());
+    grads->set_data_type(in0->data_type());
+    grads->SetFormat(in0->GetFormat());
+  }
+  return RET_OK;
+}
+
 }  // namespace lite
 }  // namespace mindspore

mindspore/lite/src/ops/softmax_cross_entropy.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef LITE_MINDSPORE_LITE_C_OPS_SOFTMAX_CROSS_ENTROPY_H_
-#define LITE_MINDSPORE_LITE_C_OPS_SOFTMAX_CROSS_ENTROPY_H_
+#ifndef MINDSPORE_LITE_SRC_OPS_SOFTMAX_CROSS_ENTROPY_H_
+#define MINDSPORE_LITE_SRC_OPS_SOFTMAX_CROSS_ENTROPY_H_
 
 #include <vector>
 #include <set>
@@ -39,9 +39,11 @@ class SoftmaxCrossEntropy : public PrimitiveC {
 
   int UnPackToFlatBuilder(const schema::Primitive *primitive, flatbuffers::FlatBufferBuilder *fbb) override;
 #endif
+  int InferShape(std::vector<lite::tensor::Tensor *> inputs_, std::vector<lite::tensor::Tensor *> outputs_) override;
+
   std::vector<int> GetAxis() const;
 };
 }  // namespace lite
 }  // namespace mindspore
 
-#endif  // LITE_MINDSPORE_LITE_C_OPS_SOFTMAX_CROSS_ENTROPY_H_
+#endif  // MINDSPORE_LITE_SRC_OPS_SOFTMAX_CROSS_ENTROPY_H_

mindspore/lite/src/populate_parameter.cc

@@ -1678,6 +1678,13 @@ PopulateParameterFunc PopulateParameterRegistry::GetParameterFunc(int type) {
   return populate_parameter_funcs_[schema::PrimitiveType(type)];
 }
 
+int PopulateParameterRegistry::AddPopulateParameterFunc(const schema::PrimitiveType &type, PopulateParameterFunc func) {
+  if ((type <  schema::PrimitiveType_MIN)|| (type > schema::PrimitiveType_MAX))
+    return -1;
+  populate_parameter_funcs_[type]  = func;
+  return 0;
+}
+
 OpParameter *PopulateParameter(const mindspore::lite::PrimitiveC *primitive) {
   if (primitive == nullptr) {
     MS_LOG(ERROR) << "Primitive is nullptr when populating parameter for op.";

mindspore/lite/src/populate_parameter.h

@@ -30,12 +30,16 @@ class PopulateParameterRegistry {
   ~PopulateParameterRegistry() = default;
 
   static PopulateParameterRegistry *GetInstance();
+  int AddPopulateParameterFunc(const schema::PrimitiveType &type, PopulateParameterFunc func);
   PopulateParameterFunc GetParameterFunc(int type);
 
  protected:
   PopulateParameterFunc populate_parameter_funcs_[schema::PrimitiveType_MAX + 1];
 };
 
+OpParameter *PopulateActivationParameter(const lite::PrimitiveC *primitive);
+OpParameter *PopulateArithmetic(const lite::PrimitiveC *primitive);
+
 OpParameter *PopulateParameter(const mindspore::lite::PrimitiveC *primitive);
 }  // namespace mindspore::kernel
 #endif  // MINDSPORE_LITE_SRC_POPULATE_PARAMETER_H_

mindspore/lite/src/runtime/kernel/arm/base/reduce_base.cc

@@ -37,8 +37,8 @@ constexpr size_t kOutputNum = 1;
 }  // namespace
 
 int ReduceBaseCPUKernel::CheckInputsOutputs() {
-  if (in_tensors_.size() != kInputNum) {
-    MS_LOG(ERROR) << "Reduce inputs size should be " << kInputNum << " but got " << in_tensors_.size();
+  if (in_tensors_.size() < kInputNum) {
+    MS_LOG(ERROR) << "Reduce inputs size should be at least " << kInputNum << " but got " << in_tensors_.size();
     return RET_ERROR;
   }
   if (out_tensors_.size() != kOutputNum) {
@@ -99,7 +99,15 @@ int ReduceBaseCPUKernel::Init() {
   if (reduce_param == nullptr) {
     return RET_NULL_PTR;
   }
-  num_axes_ = reduce_param->num_axes_;
+  if (in_tensors_.size() > 1) {
+    auto axes_ptr = in_tensors_.at(1);
+    num_axes_ = axes_ptr->ElementsNum();
+    memcpy(axes_, axes_ptr->Data(), axes_ptr->Size());
+  } else {
+    num_axes_ = reduce_param->num_axes_;
+    memcpy(axes_, reduce_param->axes_, sizeof(reduce_param->axes_));
+  }
+
   mode_ = reduce_param->mode_;
   memcpy(axes_, reduce_param->axes_, sizeof(reduce_param->axes_));
   reduce_to_end_ = reduce_param->reduce_to_end_;

mindspore/lite/src/runtime/kernel/arm/fp32_grad/activation_grad.cc

@@ -15,6 +15,7 @@
  */
 
 #include "src/runtime/kernel/arm/fp32_grad/activation_grad.h"
+#include "nnacl/fp32_grad/activation_grad.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
 #include "src/runtime/runtime_api.h"
@@ -24,41 +25,38 @@ using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
-using mindspore::schema::ActivationGradType_HSWISH;
-using mindspore::schema::ActivationGradType_LEAKY_RELU;
-using mindspore::schema::ActivationGradType_RELU;
-using mindspore::schema::ActivationGradType_RELU6;
+using mindspore::schema::ActivationType_HSWISH;
+using mindspore::schema::ActivationType_LEAKY_RELU;
+using mindspore::schema::ActivationType_RELU;
+using mindspore::schema::ActivationType_RELU6;
 using mindspore::schema::PrimitiveType_ActivationGrad;
 
 namespace mindspore::kernel {
-int ActivationGradCPUKernel::Init() {
-  outputs_[0]->set_shape(inputs_[0]->shape());
-  return RET_OK;
-}
+int ActivationGradCPUKernel::Init() { return RET_OK; }
 
 int ActivationGradCPUKernel::ReSize() { return RET_OK; }
 
 int ActivationGradCPUKernel::DoActivation(int task_id) {
-  auto yt_addr = reinterpret_cast<float *>(inputs_.at(0)->Data());
-  auto input_addr = reinterpret_cast<float *>(inputs_.at(1)->Data());
-  auto output_addr = reinterpret_cast<float *>(outputs_.at(0)->Data());
-  auto length = inputs_.at(0)->ElementsNum();
+  auto yt_addr = reinterpret_cast<float *>(in_tensors_.at(0)->Data());
+  auto input_addr = reinterpret_cast<float *>(in_tensors_.at(1)->Data());
+  auto output_addr = reinterpret_cast<float *>(out_tensors_.at(0)->Data());
+  int length = in_tensors_.at(0)->ElementsNum();
 
   auto error_code = RET_OK;
 
-  if (type_ == schema::ActivationGradType_RELU) {
+  if (param_act_grad_->type_ == schema::ActivationType_RELU) {
     error_code = ReluGrad(yt_addr, input_addr, length, output_addr);
-  } else if (type_ == schema::ActivationGradType_RELU6) {
+  } else if (param_act_grad_->type_ == schema::ActivationType_RELU6) {
     error_code = Relu6Grad(yt_addr, input_addr, length, output_addr);
-  } else if (type_ == schema::ActivationGradType_LEAKY_RELU) {
-    error_code = LReluGrad(yt_addr, input_addr, length, output_addr, alpha_);
-  } else if (type_ == schema::ActivationGradType_SIGMOID) {
+  } else if (param_act_grad_->type_ == schema::ActivationType_LEAKY_RELU) {
+    error_code = LReluGrad(yt_addr, input_addr, length, output_addr, param_act_grad_->alpha_);
+  } else if (param_act_grad_->type_ == schema::ActivationType_SIGMOID) {
     error_code = SigmoidGrad(yt_addr, input_addr, length, output_addr);
-  } else if (type_ == schema::ActivationGradType_TANH) {
+  } else if (param_act_grad_->type_ == schema::ActivationType_TANH) {
     error_code = TanhGrad(yt_addr, input_addr, length, output_addr);
-  } else if (type_ == schema::ActivationGradType_HSWISH) {
+  } else if (param_act_grad_->type_ == schema::ActivationType_HSWISH) {
     error_code = HSwishGrad(yt_addr, input_addr, length, output_addr);
-  } else if (type_ == schema::ActivationGradType_HSIGMOID) {
+  } else if (param_act_grad_->type_ == schema::ActivationType_HSIGMOID) {
     error_code = HSigmoidGrad(yt_addr, input_addr, length, output_addr);
   } else {
     MS_LOG(ERROR) << "Activation type error";
@@ -81,6 +79,12 @@ int ActivationGradRun(void *cdata, int task_id) {
 }
 
 int ActivationGradCPUKernel::Run() {
+  auto ret = Prepare();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare failed.";
+    return ret;
+  }
+
   int error_code = ParallelLaunch(THREAD_POOL_DEFAULT, ActivationGradRun, this, thread_count_);
   if (error_code != RET_OK) {
     MS_LOG(ERROR) << "Activation function error error_code[" << error_code << "]";

mindspore/lite/src/runtime/kernel/arm/fp32_grad/activation_grad.h

@@ -20,8 +20,7 @@
 #include <vector>
 #include "src/lite_kernel.h"
 #include "ir/anf.h"
-
-#include "nnacl/activation_grad.h"
+#include "nnacl/fp32/activation.h"
 
 namespace mindspore::kernel {
 class ActivationGradCPUKernel : public LiteKernel {
@@ -30,9 +29,7 @@ class ActivationGradCPUKernel : public LiteKernel {
                                    const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
                                    const mindspore::lite::PrimitiveC *primitive)
       : LiteKernel(param, inputs, outputs, ctx, primitive) {
-    ActivationGradParameter *param_act_grad = reinterpret_cast<ActivationGradParameter *>(param);
-    type_ = param_act_grad->type_;
-    alpha_ = param_act_grad->alpha_;
+    param_act_grad_ = reinterpret_cast<ActivationParameter *>(param);
   }
   ~ActivationGradCPUKernel() override = default;
 
@@ -43,9 +40,9 @@ class ActivationGradCPUKernel : public LiteKernel {
 
  private:
   int thread_count_;
-  int type_;
-  float alpha_;
+  ActivationParameter *param_act_grad_;
 };
+
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_ACTIVATION_GRAD_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/apply_momentum.cc

@@ -0,0 +1,105 @@
+
+/**
+ * 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 "src/runtime/kernel/arm/fp32_grad/apply_momentum.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "include/errorcode.h"
+#include "src/runtime/kernel/arm/fp32/nchw2nhwc.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_ApplyMomentum;
+
+namespace mindspore::kernel {
+
+int ApplyMomentumCPUKernel::ReSize() { return RET_OK; }
+
+int ApplyMomentumCPUKernel::Run() {
+  auto prepare_ret = Prepare();
+  if (prepare_ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare fail!ret: " << prepare_ret;
+    return prepare_ret;
+  }
+
+  auto weight = reinterpret_cast<float *>(in_tensors_[0]->Data());
+  auto accumulate = reinterpret_cast<float *>(in_tensors_[1]->Data());
+  float learning_rate = reinterpret_cast<float *>(in_tensors_[2]->Data())[0];
+  auto gradient = reinterpret_cast<float *>(in_tensors_[3]->Data());
+  float moment = reinterpret_cast<float *>(in_tensors_[4]->Data())[0];
+  size_t elem_num = in_tensors_[0]->ElementsNum();
+
+  // align format
+  if (in_tensors_[3]->shape().size() == 4 &&
+      in_tensors_[3]->GetFormat() == schema::Format_NCHW &&
+      in_tensors_[0]->GetFormat() == schema::Format_KHWC) {
+    PackNCHWToNHWCFp32(gradient, workspace, in_tensors_[0]->Batch(), in_tensors_[0]->Height() * in_tensors_[0]->Width(),
+                       in_tensors_[0]->Channel());
+  } else {
+    memcpy(workspace, gradient, in_tensors_[3]->ElementsNum() * sizeof(float));
+  }
+
+  for (size_t i = 0; i < elem_num; ++i) {
+    accumulate[i] = accumulate[i] * moment + workspace[i];  // * (1.0 - moment);
+    weight[i] -= accumulate[i] * learning_rate;
+  }
+  return RET_OK;
+}
+
+int ApplyMomentumCPUKernel::Init() {
+  // Only for test with uninitialized Data
+  size_t elem_num = in_tensors_[0]->ElementsNum();
+  auto accumulate = reinterpret_cast<float *>(in_tensors_[1]->Data());
+  for (int i =0; i < elem_num; i++) accumulate[i] = 0.0;
+
+  workspace = new float[elem_num];
+  return 0;
+}
+#if 0
+OpParameter *PopulateApplyMomentumParameter(const lite::Primitive *primitive) {
+  OpParameter *param = new (std::nothrow) OpParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "new Param for OptMomentum failed.";
+    return nullptr;
+  }
+  param->type_ = primitive->Type();
+  return param;
+}
+#endif
+
+kernel::LiteKernel *CpuApplyMomentumFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                    const std::vector<lite::tensor::Tensor *> &outputs,
+                                                    OpParameter *opParameter, const lite::Context *ctx,
+                                                    const kernel::KernelKey &desc, const lite::PrimitiveC *primitive) {
+  MS_ASSERT(desc.type == schema::PrimitiveType_ApplyMomentum);
+  auto *kernel = new (std::nothrow) ApplyMomentumCPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  MS_ASSERT(kernel != nullptr);
+
+  auto ret = kernel->Init();
+  if (0 != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_ApplyMomentum, CpuApplyMomentumFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32_grad/opt_momentum.h → mindspore/lite/src/runtime/kernel/arm/fp32_grad/apply_momentum.h

@@ -14,28 +14,32 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_OPT_MOMENTUM_H_
-#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_OPT_MOMENTUM_H_
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_APPLY_MOMENTUM_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_APPLY_MOMENTUM_H_
 
 #include <vector>
 #include "src/lite_kernel.h"
 #include "ir/anf.h"
 
 namespace mindspore::kernel {
-class OptMomentumCPUKernel : public LiteKernel {
+class ApplyMomentumCPUKernel : public LiteKernel {
  public:
-  explicit OptMomentumCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+  explicit ApplyMomentumCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
                                 const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
                                 const mindspore::lite::PrimitiveC *primitive)
       : LiteKernel(parameter, inputs, outputs, ctx, primitive) {}
-  ~OptMomentumCPUKernel() override {}
+  ~ApplyMomentumCPUKernel() override {delete [] workspace;}
 
   int Init() override;
   int ReSize() override;
   int Run() override;
 
  private:
+    float *workspace;
 };
+
+// OpParameter *PopulateApplyMomentumParameter(const lite::Primitive *primitive);
+
 }  // namespace mindspore::kernel
 
-#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_OPT_MOMENTUM_H_
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_APPLY_MOMENTUM_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.cc

@@ -14,11 +14,11 @@
  * limitations under the License.
  */
 
+#include "src/runtime/kernel/arm/fp32_grad/arithmetic_grad.h"
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
 #include "nnacl/fp32_grad/reduce_grad.h"
 #include "nnacl/fp32_grad/arithmetic_grad.h"
-#include "src/runtime/kernel/arm/fp32_grad/arithmetic_grad.h"
 #include "include/errorcode.h"
 
 using mindspore::kernel::KERNEL_ARCH::kCPU;
@@ -33,108 +33,41 @@ constexpr int kArithGradOpOutputNum = 2;
 }  // namespace
 
 int ArithmeticGradCPUKernel::Init() {
-  auto ret = InferShape();
-  return ret;
-}
-
-int ArithmeticGradCPUKernel::InferShape() {
-  if (inputs_.size() != kArithGradOpInputNum) {
-    MS_LOG(ERROR) << "The number of input must be " << kArithGradOpInputNum;
-    return RET_ERROR;
-  }
-  if (outputs_.size() != kArithGradOpOutputNum) {
-    MS_LOG(ERROR) << "The number of output must be " << kArithGradOpOutputNum;
-    return RET_ERROR;
-  }
-  auto dy = inputs_[0];
-  auto x1 = inputs_[1];
-  auto x2 = inputs_[2];
-  auto dx1 = outputs_[0];
-  auto dx2 = outputs_[1];
+  auto dx1 = out_tensors_[0];
+  auto dx2 = out_tensors_[1];
 
-  MS_ASSERT(dy != nullptr);
-  MS_ASSERT(x1 != nullptr);
-  MS_ASSERT(x2 != nullptr);
   MS_ASSERT(dx1 != nullptr);
   MS_ASSERT(dx2 != nullptr);
 
-  auto inShape0 = x1->shape();
-  auto inShape1 = x2->shape();
-  auto outShape = dy->shape();
-
-  if ((type() == PrimitiveType_AddGrad) || (type() == PrimitiveType_SubGrad)) {
-    arithmeticParameter_->ndim_ = outShape.size();
-    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++) {
-      arithmeticParameter_->in_shape0_[i] = (i < fillDimNum0) ? 1 : inShape0[j0++];
-      arithmeticParameter_->in_shape1_[i] = (i < fillDimNum1) ? 1 : inShape1[j1++];
-      arithmeticParameter_->out_shape_[i] = outShape[i];
-    }
-  } else {
+  if ((Type() == PrimitiveType_MulGrad) || (Type() == PrimitiveType_DivGrad)) {
     // if (inShape0.size() < inShape1.size())
     if (dx1->ElementsNum() < dx2->ElementsNum()) {
-      arithmeticParameter_->ndim_ = inShape1.size();
-      if (type() == PrimitiveType_MulGrad)
+      if (Type() == PrimitiveType_MulGrad)
         arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradMul2L;
-      else if (type() == PrimitiveType_DivGrad)
+      else if (Type() == PrimitiveType_DivGrad)
         arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradDiv2L;
 
-      auto fillDimNum = inShape1.size() - inShape0.size();  // This will not work for batch!
-      int j = 0;
-      for (unsigned int i = 0; i < inShape1.size(); i++) {
-        if (i < fillDimNum) {
-          arithmeticParameter_->in_shape1_[i] = 1;
-        } else {
-          arithmeticParameter_->in_shape1_[i] = inShape0[j++];
-        }
-        arithmeticParameter_->in_shape0_[i] = inShape1[i];
-        arithmeticParameter_->out_shape_[i] = outShape[i];
-      }
     } else if (dx2->ElementsNum() < dx1->ElementsNum()) {  // if (inShape0.size() > inShape1.size())
-      arithmeticParameter_->ndim_ = inShape0.size();
-      if (type() == PrimitiveType_MulGrad)
+      if (Type() == PrimitiveType_MulGrad)
         arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradMul1L;
-      else if (type() == PrimitiveType_DivGrad)
+      else if (Type() == PrimitiveType_DivGrad)
         arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradDiv1L;
-      arithmeticParameter_->broadcasting_ = true;
-      arithmeticParameter_->ndim_ = inShape0.size();
-      int j = 0;
-      auto fillDimNum = inShape0.size() - inShape1.size();
-      for (unsigned int i = 0; i < inShape0.size(); i++) {
-        if (i < fillDimNum) {
-          arithmeticParameter_->in_shape1_[i] = 1;
-        } else {
-          arithmeticParameter_->in_shape1_[i] = inShape1[j++];
-        }
-        arithmeticParameter_->in_shape0_[i] = inShape0[i];
-        arithmeticParameter_->out_shape_[i] = outShape[i];
-      }
-    } else {
-      arithmeticParameter_->broadcasting_ = false;
-      for (unsigned int i = 0; i < inShape0.size(); i++) {
-        arithmeticParameter_->in_shape1_[i] = inShape1[i];
-        arithmeticParameter_->in_shape0_[i] = inShape0[i];
-        arithmeticParameter_->out_shape_[i] = outShape[i];
-      }
     }
 
-    tile_data0 = new (std::nothrow) float[inputs_.at(0)->ElementsNum()];
+    tile_data0 = new (std::nothrow) float[in_tensors_.at(0)->ElementsNum()];
     if (tile_data0 == nullptr) {
       MS_LOG(ERROR) << "new data0 fail!";
       return RET_ERROR;
     }
-    tile_data1 = new (std::nothrow) float[inputs_.at(0)->ElementsNum()];
+    tile_data1 = new (std::nothrow) float[in_tensors_.at(0)->ElementsNum()];
     if (tile_data1 == nullptr) {
       MS_LOG(ERROR) << "new data1 fail!";
       delete tile_data0;
       return RET_ERROR;
     }
 
-    if (type() == PrimitiveType_DivGrad) {
-      tile_data2 = new (std::nothrow) float[inputs_.at(0)->ElementsNum()];
+    if (Type() == PrimitiveType_DivGrad) {
+      tile_data2 = new (std::nothrow) float[in_tensors_.at(0)->ElementsNum()];
       if (tile_data2 == nullptr) {
         MS_LOG(ERROR) << "new data2 fail!";
         delete tile_data0;
@@ -144,10 +77,6 @@ int ArithmeticGradCPUKernel::InferShape() {
     }
   }
 
-  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;
 }
 
@@ -187,16 +116,16 @@ void ArithmeticGradCPUKernel::ArithmeticGradSub(float *dy, int dy_size, float *d
 
 void ArithmeticGradCPUKernel::ArithmeticGradMul(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
                                                 int dx2_size) {
-  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
-  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+  auto x1_data = reinterpret_cast<float *>(in_tensors_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(in_tensors_[2]->Data());
   ElementMul(dy, x1_data, dx2, dy_size);
   ElementMul(dy, x2_data, dx1, dy_size);
 }
 
 void ArithmeticGradCPUKernel::ArithmeticGradMul1L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
                                                   int dx2_size) {
-  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
-  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+  auto x1_data = reinterpret_cast<float *>(in_tensors_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(in_tensors_[2]->Data());
   ElementMul(dy, x1_data, tile_data0, dy_size);
   ReduceSumByAxes(tile_data0, arithmeticParameter_->in_shape0_, dx2, arithmeticParameter_->in_shape1_,
                   arithmeticParameter_->ndim_);
@@ -206,8 +135,8 @@ void ArithmeticGradCPUKernel::ArithmeticGradMul1L(float *dy, int dy_size, float
 
 void ArithmeticGradCPUKernel::ArithmeticGradMul2L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
                                                   int dx2_size) {
-  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
-  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+  auto x1_data = reinterpret_cast<float *>(in_tensors_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(in_tensors_[2]->Data());
   ElementMul(dy, x2_data, tile_data0, dy_size);
   ReduceSumByAxes(tile_data0, arithmeticParameter_->in_shape0_, dx1, arithmeticParameter_->in_shape1_,
                   arithmeticParameter_->ndim_);
@@ -217,16 +146,16 @@ void ArithmeticGradCPUKernel::ArithmeticGradMul2L(float *dy, int dy_size, float
 
 void ArithmeticGradCPUKernel::ArithmeticGradDiv(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
                                                 int dx2_size) {
-  auto x1 = reinterpret_cast<float *>(inputs_[1]->Data());
-  auto x2 = reinterpret_cast<float *>(inputs_[2]->Data());
+  auto x1 = reinterpret_cast<float *>(in_tensors_[1]->Data());
+  auto x2 = reinterpret_cast<float *>(in_tensors_[2]->Data());
   ElementDiv(dy, x2, dx1, dy_size);
   ElementMulAndDivNegSquare(dy, x1, x2, dx2, dy_size);
 }
 
 void ArithmeticGradCPUKernel::ArithmeticGradDiv1L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
                                                   int dx2_size) {
-  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
-  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+  auto x1_data = reinterpret_cast<float *>(in_tensors_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(in_tensors_[2]->Data());
 
   ElementMul(x2_data, x2_data, dx2, dx2_size);
   ElementMul(x1_data, dy, dx1, dy_size);  // use dx1 buffer
@@ -243,8 +172,8 @@ void ArithmeticGradCPUKernel::ArithmeticGradDiv1L(float *dy, int dy_size, float
 
 void ArithmeticGradCPUKernel::ArithmeticGradDiv2L(float *dy, int dy_size, float *dx1, int dx1_size, float *dx2,
                                                   int dx2_size) {
-  auto x1_data = reinterpret_cast<float *>(inputs_[1]->Data());
-  auto x2_data = reinterpret_cast<float *>(inputs_[2]->Data());
+  auto x1_data = reinterpret_cast<float *>(in_tensors_[1]->Data());
+  auto x2_data = reinterpret_cast<float *>(in_tensors_[2]->Data());
 
   // dx1 = dy/x2
   ElementDiv(dy, x2_data, tile_data0, dy_size);  // first multiply into temp
@@ -259,13 +188,13 @@ void ArithmeticGradCPUKernel::ArithmeticGradDiv2L(float *dy, int dy_size, float
 int ArithmeticGradCPUKernel::ReSize() { return RET_OK; }
 
 int ArithmeticGradCPUKernel::Run() {
-  auto dy = reinterpret_cast<float *>(inputs_[0]->Data());
-  auto dx1 = reinterpret_cast<float *>(outputs_[0]->Data());
-  auto dx2 = reinterpret_cast<float *>(outputs_[1]->Data());
+  auto dy = reinterpret_cast<float *>(in_tensors_[0]->Data());
+  auto dx1 = reinterpret_cast<float *>(out_tensors_[0]->Data());
+  auto dx2 = reinterpret_cast<float *>(out_tensors_[1]->Data());
 
-  size_t dy_size = inputs_.at(0)->ElementsNum();
-  size_t dx1_size = outputs_.at(0)->ElementsNum();
-  size_t dx2_size = outputs_[1]->ElementsNum();
+  size_t dy_size = in_tensors_.at(0)->ElementsNum();
+  size_t dx1_size = out_tensors_.at(0)->ElementsNum();
+  size_t dx2_size = out_tensors_[1]->ElementsNum();
   (this->*arithmetic_grad_)(dy, dy_size, dx1, dx1_size, dx2, dx2_size);
   return RET_OK;
 }

mindspore/lite/src/runtime/kernel/arm/fp32_grad/arithmetic_grad.h

@@ -40,7 +40,7 @@ class ArithmeticGradCPUKernel : public LiteKernel {
                                    const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
                                    const mindspore::lite::PrimitiveC *primitive)
       : LiteKernel(parameter, inputs, outputs, ctx, primitive), tile_data0(NULL), tile_data1(NULL), tile_data2(NULL) {
-    switch (type()) {
+    switch (Type()) {
       case PrimitiveType_MulGrad:
         arithmetic_grad_ = &ArithmeticGradCPUKernel::ArithmeticGradMul;  // this will be adjusted in InferShape
         break;

mindspore/lite/src/runtime/kernel/arm/fp32_grad/bias_grad.cc

@@ -27,33 +27,9 @@ using mindspore::lite::RET_OK;
 using mindspore::schema::PrimitiveType_BiasGrad;
 
 namespace mindspore::kernel {
-int BiasGradCPUKernel::InferShape() {
-  if (1 != this->inputs_.size()) {
-    MS_LOG(ERROR) << "BiasGrad should have one input";
-    return RET_ERROR;
-  }
-  if (1 != this->outputs_.size()) {
-    MS_LOG(ERROR) << "BiasGrad should have one output";
-    return RET_ERROR;
-  }
-  auto *in0 = inputs_.front();
-  auto *out = outputs_.front();
-  MS_ASSERT(in0 != nullptr);
-  MS_ASSERT(out != nullptr);
-  auto inshape = in0->shape();
-  int ndim = inshape.size();
-  for (int i = 0; i < ndim - 1; i++) {
-    inshape[i] = 1;
-  }
-  out->set_shape(inshape);
-  out->set_data_type(in0->data_type());
-  return RET_OK;
-}
 
 int BiasGradCPUKernel::Init() {
-  MS_ASSERT(InferShape() == RET_OK);
-
-  auto dims = inputs_[0]->shape();
+  auto dims = in_tensors_[0]->shape();
   bias_param->ndim_ = dims.size();
   for (unsigned int i = 0; i < bias_param->ndim_; i++) {
     bias_param->in_shape0_[i] = dims[i];
@@ -75,8 +51,8 @@ int BiasGradCPUKernel::Run() {
     MS_LOG(ERROR) << "Prepare failed.";
     return RET_ERROR;
   }
-  auto in = reinterpret_cast<float *>(inputs_.at(0)->Data());
-  auto out = reinterpret_cast<float *>(outputs_.at(0)->Data());
+  auto in = reinterpret_cast<float *>(in_tensors_.at(0)->Data());
+  auto out = reinterpret_cast<float *>(out_tensors_.at(0)->Data());
 
   size_t nhw_size = 1;
   size_t channels = bias_param->in_shape0_[bias_param->ndim_ - 1];  // C in NHWC

mindspore/lite/src/runtime/kernel/arm/fp32_grad/bias_grad.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_BIAS_GRAD_H_
-#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_BIAS_GRAD_H_
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_BIAS_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_BIAS_GRAD_H_
 
 #include <vector>
 #include "src/lite_kernel.h"
@@ -35,7 +35,6 @@ class BiasGradCPUKernel : public LiteKernel {
   ~BiasGradCPUKernel() override = default;
 
   int Init() override;
-  int InferShape();
   int ReSize() override;
   int Run() override;
 
@@ -44,4 +43,4 @@ class BiasGradCPUKernel : public LiteKernel {
 };
 }  // namespace mindspore::kernel
 
-#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_BIAS_GRAD_H_
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_BIAS_GRAD_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/bn_grad.cc

@@ -14,11 +14,11 @@
  * limitations under the License.
  */
 
+#include "src/runtime/kernel/arm/fp32_grad/bn_grad.h"
 #include <algorithm>
 #include <vector>
 #include "schema/model_generated.h"
 #include "src/kernel_registry.h"
-#include "src/runtime/kernel/arm/fp32_grad/bn_grad.h"
 #include "nnacl/fp32_grad/batch_norm.h"
 #include "include/errorcode.h"
 
@@ -27,79 +27,103 @@ using mindspore::lite::KernelRegistrar;
 using mindspore::lite::RET_ERROR;
 using mindspore::lite::RET_OK;
 // using mindspore::lite::REG_OP;
-using mindspore::schema::PrimitiveType_BNGradInput;
+using mindspore::schema::PrimitiveType_BNGrad;
 
+/*
+{dy}
+{x }
+{scale }
+{save_mean }
+{save_inv_variance }
+*/
 namespace mindspore::kernel {
-int BNGradInputCPUKernel::Init() {
-  auto bn_param = reinterpret_cast<bnParameter *>(opParameter);
-  workspace_size = 5 * bn_param->channels;
-  workspace = new  (std::nothrow) float[workspace_size];
-  if (workspace == nullptr) {
-    MS_LOG(ERROR) << "new workspace fail!";
-    return RET_ERROR;
-  }
 
-  if (2 != this->inputs_.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad should has 2 inputs";
-    return RET_ERROR;
+#if 0
+OpParameter *PopulateBNGradParameter(const lite::Primitive *primitive) {
+  BNGradParameter *param = new (std::nothrow) BNGradParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "new Param for conv grad filter failed.";
+    return nullptr;
   }
-  if (1 != this->outputs_.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad should has one output";
+  param->op_parameter_.type_ = primitive->Type();
+
+  auto bngrad_primitive = primitive->Value()->value_as_BNGrad();
+  param->epsilon_ = bngrad_primitive->eps();
+  param->momentum_ = bngrad_primitive->momentum();
+  return reinterpret_cast<OpParameter *>(param);
+}
+#endif
+int BNGradCPUKernel::Init() {
+  auto *input_x = in_tensors_.at(1);
+  int channels = input_x->shape().at(kNHWC_C);
+  workspace_size = 5 * channels;
+  workspace = new (std::nothrow) float[workspace_size];
+  if (workspace == nullptr) {
+    MS_LOG(ERROR) << "new workspace fail!";
     return RET_ERROR;
   }
-  auto *input_tensor = inputs_.at(0);
-  auto *out_tensor = outputs_.at(0);
-  auto in_shape = input_tensor->shape();
-  out_tensor->set_shape(in_shape);
-  out_tensor->set_data_type(input_tensor->data_type());
   return RET_OK;
 }
 
-int BNGradInputCPUKernel::ReSize() { return RET_OK; }
+int BNGradCPUKernel::ReSize() { return RET_OK; }
 
-int BNGradInputCPUKernel::Run() {
-  auto *input_x = inputs_.at(0);
-  auto *input_yt = inputs_.at(1);
-  auto *input_scale = inputs_.at(2);
-  auto *output_grad = outputs_.at(0);
-  auto bn_param = reinterpret_cast<bnParameter *>(opParameter);
-  int batch = bn_param->batch;
-  int channels = bn_param->channels;
-  int spatial = bn_param->spatial;
-  float eps = bn_param->eps;
+int BNGradCPUKernel::Run() {
+  // std::cout << "run succ" << std::endl;
+  auto prepare_ret = Prepare();
+  if (prepare_ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare fail!ret: " << prepare_ret;
+    return prepare_ret;
+  }
+  auto bn_param = reinterpret_cast<BNGradParameter *>(op_parameter_);
+  auto *input_yt = in_tensors_.at(0);
+  auto *input_x = in_tensors_.at(1);
+  auto *input_scale = in_tensors_.at(2);
+  auto *output_dx = out_tensors_.at(0);
+  auto *output_scale = out_tensors_.at(1);
+  auto *output_bias = out_tensors_.at(2);
+  // Tensor *bias = input[5];
+  int batch = input_x->Batch();
+  int channels = input_x->Channel();
+  int spatial = input_x->Height() * input_x->Width();
+  float eps = bn_param->epsilon_;
   std::fill(workspace, workspace + workspace_size, 0.f);
-
   float *mean = workspace;
-  float *variance = mean + channels;
-  float *mean_delta = variance + channels;
+  float *invar = mean + channels;
+  float *mean_delta = invar + channels;
   float *variance_delta = mean_delta + channels;
   float *mean_add_delta = variance_delta + channels;
 
   float *x = reinterpret_cast<float *>(input_x->Data());
   float *yt = reinterpret_cast<float *>(input_yt->Data());
   float *scale = reinterpret_cast<float *>(input_scale->Data());
-  float *out = reinterpret_cast<float *>(output_grad->Data());
+  float *dx = reinterpret_cast<float *>(output_dx->Data());
+  float *dscale = reinterpret_cast<float *>(output_scale->Data());
+  float *dbias = reinterpret_cast<float *>(output_bias->Data());
 
-  std::copy(yt, yt + batch * channels * spatial, out);
-  meanVar(x, batch, spatial, channels, mean, variance);
-  scaleBias(scale, batch, channels, spatial, out);
-  meanDelta(out, spatial, channels, eps, variance, mean_delta);
-  varianceDelta(x, out, mean, variance, batch, channels, spatial, eps, variance_delta);
+  std::copy(yt, yt + batch * channels * spatial, dx);
+  meanVar(x, batch, spatial, channels, eps, mean, invar);
+  scaleBias(scale, batch, channels, spatial, dx);
+  meanDelta(dx, spatial, channels, invar, mean_delta);
+  varianceDelta(x, dx, mean, invar, batch, channels, spatial, variance_delta);
   meanAdd(x, mean, variance_delta, batch, channels, spatial, mean_add_delta, mean_delta);
-  NormalizeDelta(x, mean, variance, mean_delta, variance_delta, batch, channels, eps, spatial, out);
+  NormalizeDelta(x, mean, invar, mean_delta, variance_delta, batch, channels, spatial, dx);
+  // dbias
+  sumSpatialBatch(yt, batch * spatial, channels, dbias);
+  // dscale
+  backwardScale(x, mean, invar, yt, batch, channels, spatial, dscale);
   return RET_OK;
 }
 
-kernel::LiteKernel *CpuBNGradInputFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                                    const std::vector<lite::tensor::Tensor *> &outputs,
-                                                    OpParameter *opParameter, const lite::Context *ctx,
-                                                    const kernel::KernelKey &desc,
-                                                    const mindspore::lite::PrimitiveC *primitive) {
+kernel::LiteKernel *CpuBNGradFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                               const std::vector<lite::tensor::Tensor *> &outputs,
+                                               OpParameter *opParameter, const lite::Context *ctx,
+                                               const kernel::KernelKey &desc,
+                                               const mindspore::lite::PrimitiveC *primitive) {
   MS_ASSERT(opParameter != nullptr);
-  MS_ASSERT(desc.type == schema::PrimitiveType_BNGradInput);
-  auto *kernel = new (std::nothrow) BNGradInputCPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  MS_ASSERT(desc.type == schema::PrimitiveType_BNGrad);
+  auto *kernel = new (std::nothrow) BNGradCPUKernel(opParameter, inputs, outputs, ctx, primitive);
   if (kernel == nullptr) {
-    MS_LOG(ERROR) << "new BNGradInputCPUKernel fail!";
+    MS_LOG(ERROR) << "new BNGradCPUKernel fail!";
     return nullptr;
   }
   auto ret = kernel->Init();
@@ -112,5 +136,5 @@ kernel::LiteKernel *CpuBNGradInputFp32KernelCreator(const std::vector<lite::tens
   return kernel;
 }
 
-REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_BNGradInput, CpuBNGradInputFp32KernelCreator)
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_BNGrad, CpuBNGradFp32KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32_grad/bn_grad.h

@@ -14,21 +14,25 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_BNGRAD_INPUT_H_
-#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_BNGRAD_INPUT_H_
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_BN_GRAD_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_BN_GRAD_H_
 
 #include <vector>
 #include "src/lite_kernel.h"
 #include "ir/anf.h"
 
+
 namespace mindspore::kernel {
-class BNGradInputCPUKernel : public LiteKernel {
+
+
+
+class BNGradCPUKernel : public LiteKernel {
  public:
-  explicit BNGradInputCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+  explicit BNGradCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
                                 const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
                                 const mindspore::lite::PrimitiveC *primitive)
       : LiteKernel(parameter, inputs, outputs, ctx, primitive) {}
-  ~BNGradInputCPUKernel() override { delete workspace; }
+  ~BNGradCPUKernel() override { delete workspace; }
 
   int Init() override;
   int ReSize() override;
@@ -38,5 +42,8 @@ class BNGradInputCPUKernel : public LiteKernel {
   float *workspace;
   int workspace_size;
 };
+
+// OpParameter *PopulateBNGradParameter(const lite::Primitive *primitive);
+
 }  // namespace mindspore::kernel
-#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_BNGRAD_INPUT_H_
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_BN_GRAD_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.cc

@@ -0,0 +1,121 @@
+/**
+ * Copyright 2019 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/runtime/kernel/arm/fp32_grad/convolution.h"
+#include "nnacl/fp32_grad/pack_ext.h"
+#include "nnacl/fp32_grad/gemm.h"
+#include "include/errorcode.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+namespace mindspore::kernel {
+int ConvolutionTrainCPUKernel::Init() {
+  auto conv_param_ = reinterpret_cast<ConvParameter *>(op_parameter_);
+  auto *input_x = in_tensors_.at(kInputIndex);
+  auto *input_weight = in_tensors_.at(kWeightIndex);
+  auto *out_y = out_tensors_.at(kOutputIndex);
+
+  conv_param_->output_batch_ = out_y->shape().at(kNHWC_N);
+  conv_param_->input_batch_ = input_x->shape().at(kNHWC_N);
+  conv_param_->input_h_ = input_x->shape().at(kNHWC_H);
+  conv_param_->input_w_ = input_x->shape().at(kNHWC_W);
+  conv_param_->output_h_ = out_y->shape().at(kNHWC_H);
+  conv_param_->output_w_ = out_y->shape().at(kNHWC_W);
+  conv_param_->input_channel_ = input_x->shape().at(kNHWC_C);
+  conv_param_->output_channel_ = input_weight->shape().at(kNHWC_N);
+  conv_param_->kernel_h_ = input_weight->shape().at(kNHWC_H);
+  conv_param_->kernel_w_ = input_weight->shape().at(kNHWC_W);
+
+  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_;
+
+  workspace = new float[ws_size];
+  return RET_OK;
+}
+
+int ConvolutionTrainCPUKernel::ReSize() { return RET_OK; }
+
+int ConvolutionTrainCPUKernel::Run() {
+  auto prepare_ret = Prepare();
+  if (prepare_ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare fail!ret: " << prepare_ret;
+    return prepare_ret;
+  }
+  auto conv_param_ = reinterpret_cast<ConvParameter *>(op_parameter_);
+  auto *input_x = in_tensors_.at(kInputIndex);
+  auto *input_w = in_tensors_.at(kWeightIndex);
+  auto *out_y = out_tensors_.at(kOutputIndex);
+
+  auto x_addr = reinterpret_cast<float *>(input_x->Data());
+  auto y_addr = reinterpret_cast<float *>(out_y->Data());
+  auto w_addr = reinterpret_cast<float *>(input_w->Data());
+
+  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;
+
+  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);
+    }
+  }
+
+  // std::cout << "run succ" << std::endl;
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuConvTrainFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                  const std::vector<lite::tensor::Tensor *> &outputs,
+                                                  OpParameter *opParameter, const lite::Context *ctx,
+                                                  const kernel::KernelKey &desc, const lite::PrimitiveC *primitive) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Conv2D);
+
+  auto *kernel = new (std::nothrow) ConvolutionTrainCPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  MS_ASSERT(kernel != nullptr);
+
+  auto ret = kernel->Init();
+  if (RET_OK != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.h

@@ -0,0 +1,47 @@
+/**
+ * Copyright 2019 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_CONVOLUTION_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_CONVOLUTION_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+namespace mindspore::kernel {
+class ConvolutionTrainCPUKernel : public LiteKernel {
+ public:
+  explicit ConvolutionTrainCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                                          const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
+                                          const lite::PrimitiveC *primitive)
+      : LiteKernel(parameter, inputs, outputs, ctx, primitive) {}
+  ~ConvolutionTrainCPUKernel() override { delete [] workspace; }
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  float *workspace;
+};
+
+kernel::LiteKernel *CpuConvTrainFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                  const std::vector<lite::tensor::Tensor *> &outputs,
+                                                  OpParameter *opParameter, const lite::Context *ctx,
+                                                  const kernel::KernelKey &desc, const lite::PrimitiveC *primitive);
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_CONVOLUTION_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.cc

@@ -33,30 +33,24 @@ int ConvolutionGradFilterCPUKernel::Init() {
   // x is in input 1
   // dw is output 0
 
-  if (2 != this->inputs_.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad should has 2 inputs";
-    return RET_ERROR;
-  }
-  if (1 != this->outputs_.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad should has one output";
-    return RET_ERROR;
-  }
-
-  auto *input_tensor = inputs_.at(1);
-  MS_ASSERT(input_tensor != nullptr);
-  auto *dy = inputs_.at(0);
-  MS_ASSERT(dy != nullptr);
-  auto *weight_tensor = outputs_.at(0);
+  auto *x_tensor = in_tensors_.at(1);
+  MS_ASSERT(x_tensor != nullptr);
+  auto *dy_tensor = in_tensors_.at(0);
+  MS_ASSERT(dy_tensor != nullptr);
+  auto *weight_tensor = out_tensors_.at(0);
   MS_ASSERT(weight_tensor != nullptr);
 
-  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
-  conv_param->output_batch_ = this->inputs_.at(0)->shape().at(kNHWC_N);
-  conv_param->input_batch_ = this->inputs_.at(1)->shape().at(kNHWC_N);
-  conv_param->input_h_ = this->inputs_.at(1)->shape().at(kNHWC_H);
-  conv_param->input_w_ = this->inputs_.at(1)->shape().at(kNHWC_W);
-  // assume OutCh|kh|kw|In
-  conv_param->input_channel_ = this->inputs_.at(1)->shape().at(kNHWC_C);
-  conv_param->output_channel_ = this->outputs_.at(0)->shape().at(kNHWC_N);
+  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter_);
+  conv_param->output_batch_ = dy_tensor->shape().at(kNHWC_N);
+  conv_param->input_batch_ = x_tensor->shape().at(kNHWC_N);
+  conv_param->input_h_ = x_tensor->shape().at(kNHWC_H);
+  conv_param->input_w_ = x_tensor->shape().at(kNHWC_W);
+  // assume OutCh|kh|kw|InCh
+  conv_param->input_channel_ = x_tensor->shape().at(kNHWC_C);
+  conv_param->output_channel_ = dy_tensor->shape().at(kNHWC_C);
+  // TBD
+  conv_param->output_h_ = dy_tensor->shape()[kNHWC_H];
+  conv_param->output_w_ = dy_tensor->shape()[kNHWC_W];
 
   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_;
@@ -67,34 +61,21 @@ int ConvolutionGradFilterCPUKernel::Init() {
     return RET_ERROR;
   }
 
-  int output_w = 0;
-  int output_h = 0;
-  output_h = dy->shape()[kNHWC_H];
-  output_w = dy->shape()[kNHWC_W];
-
-  std::vector<int> out_shape(4);
-  out_shape.at(0) = conv_param->output_channel_;
-  out_shape.at(1) = conv_param->kernel_h_;
-  out_shape.at(2) = conv_param->kernel_w_;
-  out_shape.at(3) = conv_param->input_channel_ / conv_param->group_;
-
-  // weight is output
-  weight_tensor->set_shape(out_shape);
-  weight_tensor->set_data_type(input_tensor->data_type());
-
-  conv_param->output_h_ = output_h;
-  conv_param->output_w_ = output_w;
-
   return RET_OK;
 }
 
-int ConvolutionGradFilterCPUKernel::ReSize() { return 0; }
+int ConvolutionGradFilterCPUKernel::ReSize() { return RET_OK; }
 
 int ConvolutionGradFilterCPUKernel::Run() {
-  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
-  auto *input_dy = inputs_.at(0);
-  auto *input_x = inputs_.at(1);
-  auto *out_dw = outputs_.at(0);
+  auto prepare_ret = Prepare();
+  if (prepare_ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare fail!ret: " << prepare_ret;
+    return prepare_ret;
+  }
+  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter_);
+  auto *input_dy = in_tensors_.at(0);
+  auto *input_x = in_tensors_.at(1);
+  auto *out_dw = out_tensors_.at(0);
 
   auto x_addr = reinterpret_cast<float *>(input_x->Data());
   auto dy_addr = reinterpret_cast<float *>(input_dy->Data());
@@ -135,7 +116,48 @@ int ConvolutionGradFilterCPUKernel::Run() {
   // std::cout << "run succ" << std::endl;
   return RET_OK;
 }
+#if 0
+OpParameter *PopulateConvolutionGradFilterParameter(const lite::Primitive *primitive) {
+  ConvParameter *param = new (std::nothrow) ConvParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "new Param for conv grad filter failed.";
+    return nullptr;
+  }
+  param->op_parameter_.type_ = primitive->Type();
+
+  auto convg_primitive = primitive->Value()->value_as_Conv2DGradFilter();
+  param->kernel_h_ = convg_primitive->kernelH();
+  param->kernel_w_ = convg_primitive->kernelW();
+  param->stride_h_ = convg_primitive->strideH();
+  param->stride_w_ = convg_primitive->strideW();
+  param->dilation_h_ = convg_primitive->dilateH();
+  param->dilation_w_ = convg_primitive->dilateW();
+  param->pad_h_ = convg_primitive->padUp();
+  param->pad_w_ = convg_primitive->padLeft();
+  param->pad_u_ = convg_primitive->padUp();
+  param->pad_d_ = convg_primitive->padDown();
+  param->pad_l_ = convg_primitive->padLeft();
+  param->pad_r_ = convg_primitive->padRight();
+  param->group_ = convg_primitive->group();
+  auto act_type = convg_primitive->activationType();
+  switch (act_type) {
+    case schema::ActivationType_RELU:
+      param->is_relu_ = true;
+      param->is_relu6_ = false;
+      break;
+    case schema::ActivationType_RELU6:
+      param->is_relu_ = false;
+      param->is_relu6_ = true;
+      break;
+    default:
+      param->is_relu_ = false;
+      param->is_relu6_ = false;
+      break;
+  }
 
+  return reinterpret_cast<OpParameter *>(param);
+}
+#endif
 kernel::LiteKernel *CpuConvGradFilterFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
                                                        const std::vector<lite::tensor::Tensor *> &outputs,
                                                        OpParameter *opParameter, const lite::Context *ctx,

mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.h

@@ -1,4 +1,4 @@
-/**
+ /**
  * Copyright 2019 Huawei Technologies Co., Ltd
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -28,15 +28,17 @@ class ConvolutionGradFilterCPUKernel : public LiteKernel {
                                           const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
                                           const mindspore::lite::PrimitiveC *primitive)
       : LiteKernel(parameter, inputs, outputs, ctx, primitive) {}
-  ~ConvolutionGradFilterCPUKernel() override { delete workspace; }
+  ~ConvolutionGradFilterCPUKernel() override { delete [] workspace; }
 
   int Init() override;
   int ReSize() override;
   int Run() override;
 
  private:
-  float *workspace;
+  float *workspace = nullptr;
 };
+
+
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_CONVOLUTION_GRAD_FILTER_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.cc

@@ -29,23 +29,14 @@ using mindspore::schema::PrimitiveType_Conv2DGradInput;
 
 namespace mindspore::kernel {
 int ConvolutionGradInputCPUKernel::Init() {
-  if (2 != this->inputs_.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad should has 2 inputs";
-    return RET_ERROR;
-  }
-  if (1 != this->outputs_.size()) {
-    MS_LOG(ERROR) << "Conv2d Grad should has one output";
-    return RET_ERROR;
-  }
-
-  auto *dy_tensor = inputs_.at(kInputIndex);
+  auto *dy_tensor = in_tensors_.at(kInputIndex);
   MS_ASSERT(dy_tensor != nullptr);
-  auto *weight_tensor = inputs_.at(kWeightIndex);
+  auto *weight_tensor = in_tensors_.at(kWeightIndex);
   MS_ASSERT(weight_tensor != nullptr);
-  auto *dx_tensor = outputs_.at(kOutputIndex);
+  auto *dx_tensor = out_tensors_.at(kOutputIndex);
   MS_ASSERT(dx_tensor != nullptr);
 
-  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
+  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter_);
   conv_param->output_batch_ = dx_tensor->shape()[(kNHWC_N)];
   conv_param->input_batch_ = dy_tensor->shape()[(kNHWC_N)];
 
@@ -74,10 +65,16 @@ int ConvolutionGradInputCPUKernel::Init() {
 int ConvolutionGradInputCPUKernel::ReSize() { return 0; }
 
 int ConvolutionGradInputCPUKernel::Run() {
-  auto conv_param = reinterpret_cast<ConvParameter *>(opParameter);
-  auto *input_dy = inputs_.at(0);
-  auto *input_w = inputs_.at(1);
-  auto *out_dx = outputs_.at(0);
+  auto prepare_ret = Prepare();
+  if (prepare_ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare fail!ret: " << prepare_ret;
+    return prepare_ret;
+  }
+
+  auto conv_param = reinterpret_cast<ConvParameter *>(op_parameter_);
+  auto *input_dy = in_tensors_.at(0);
+  auto *input_w = in_tensors_.at(1);
+  auto *out_dx = out_tensors_.at(0);
 
   auto dy_addr = reinterpret_cast<float *>(input_dy->Data());
   auto w_addr = reinterpret_cast<float *>(input_w->Data());
@@ -116,6 +113,49 @@ int ConvolutionGradInputCPUKernel::Run() {
   return 0;
 }
 
+#if 0
+OpParameter *PopulateConvolutionGradInputParameter(const lite::Primitive *primitive) {
+  ConvParameter *param = new (std::nothrow) ConvParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "new Param for conv grad input failed.";
+    return nullptr;
+  }
+  param->op_parameter_.type_ = primitive->Type();
+
+  auto convg_primitive = primitive->Value()->value_as_Conv2DGradInput();
+  param->kernel_h_ = convg_primitive->kernelH();
+  param->kernel_w_ = convg_primitive->kernelW();
+  param->stride_h_ = convg_primitive->strideH();
+  param->stride_w_ = convg_primitive->strideW();
+  param->dilation_h_ = convg_primitive->dilateH();
+  param->dilation_w_ = convg_primitive->dilateW();
+  param->pad_h_ = convg_primitive->padUp();
+  param->pad_w_ = convg_primitive->padLeft();
+  param->pad_u_ = convg_primitive->padUp();
+  param->pad_d_ = convg_primitive->padDown();
+  param->pad_l_ = convg_primitive->padLeft();
+  param->pad_r_ = convg_primitive->padRight();
+  param->group_ = convg_primitive->group();
+  auto act_type = convg_primitive->activationType();
+  switch (act_type) {
+    case schema::ActivationType_RELU:
+      param->is_relu_ = true;
+      param->is_relu6_ = false;
+      break;
+    case schema::ActivationType_RELU6:
+      param->is_relu_ = false;
+      param->is_relu6_ = true;
+      break;
+    default:
+      param->is_relu_ = false;
+      param->is_relu6_ = false;
+      break;
+  }
+
+  return reinterpret_cast<OpParameter *>(param);
+}
+#endif
+
 kernel::LiteKernel *CpuConvGradInputFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
                                                       const std::vector<lite::tensor::Tensor *> &outputs,
                                                       OpParameter *opParameter, const lite::Context *ctx,

mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.h

@@ -28,7 +28,7 @@ class ConvolutionGradInputCPUKernel : public LiteKernel {
                                          const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
                                          const mindspore::lite::PrimitiveC *primitive)
       : LiteKernel(parameter, inputs, outputs, ctx, primitive) {}
-  ~ConvolutionGradInputCPUKernel() override { delete workspace; }
+  ~ConvolutionGradInputCPUKernel() override { delete [] workspace; }
 
   int Init() override;
   int ReSize() override;
@@ -37,6 +37,9 @@ class ConvolutionGradInputCPUKernel : public LiteKernel {
  private:
   float *workspace;
 };
+
+// OpParameter *PopulateConvolutionGradInputParameter(const lite::Primitive *primitive);
+
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_CONVOLUTION_GRAD_INPUT_H

mindspore/lite/src/runtime/kernel/arm/fp32_grad/depend.cc

@@ -0,0 +1,73 @@
+/**
+ * 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 <vector>
+#include "src/runtime/kernel/arm/fp32_grad/depend.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "include/errorcode.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_Depend;
+
+namespace mindspore::kernel {
+
+int DependCPUKernel::Init() {
+  return RET_OK;
+}
+
+int DependCPUKernel::ReSize() { return 0; }
+
+int DependCPUKernel::Run() {
+#if 0
+  auto ret = Prepare();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare failed.";
+    return RET_ERROR;
+  }
+  auto in = reinterpret_cast<float *>(in_tensors_.at(0)->Data());
+  auto out = reinterpret_cast<float *>(out_tensors_.at(0)->Data());
+
+  memcpy(out, in, in_tensors_.at(0)->Size());
+#endif
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuDependFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                 const std::vector<lite::tensor::Tensor *> &outputs,
+                                                 OpParameter *opParameter, const lite::Context *ctx,
+                                                 const kernel::KernelKey &desc, const lite::PrimitiveC *primitive) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_Depend);
+  auto *kernel =
+    new (std::nothrow) DependCPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  MS_ASSERT(kernel != nullptr);
+
+  auto ret = kernel->Init();
+  if (RET_OK != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeBool, PrimitiveType_Depend, CpuDependFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32_grad/depend.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_RUNTIME_KERNEL_ARM_FP32_GRAD_DEPEND_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_DEPEND_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+#include "nnacl/fp32/arithmetic.h"
+
+namespace mindspore::kernel {
+class DependCPUKernel : public LiteKernel {
+ public:
+  explicit DependCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                             const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
+                             const lite::PrimitiveC *primitive)
+      : LiteKernel(parameter, inputs, outputs, ctx, primitive) {
+    param = parameter;
+  }
+  ~DependCPUKernel() override = default;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  OpParameter *param;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_DEPEND_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/make_tuple.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_RUNTIME_KERNEL_ARM_FP32_GRAD_MAKE_TUPLE_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_MAKE_TUPLE_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+#include "src/runtime/kernel/arm/nnacl/fp32/arithmetic.h"
+
+namespace mindspore::kernel {
+class MakeTupleCPUKernel : public LiteKernel {
+ public:
+  explicit MakeTupleCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                             const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
+                             const lite::Primitive *primitive)
+      : LiteKernel(parameter, inputs, outputs, ctx, primitive) {
+    param = parameter;
+  }
+  ~MakeTupleCPUKernel() override = default;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  OpParameter *param;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_MAKE_TUPLE_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/opt_momentum.cc

@@ -1,87 +0,0 @@
-
-/**
- * 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 "schema/model_generated.h"
-#include "src/kernel_registry.h"
-#include "src/runtime/kernel/arm/fp32_grad/opt_momentum.h"
-#include "include/errorcode.h"
-
-using mindspore::kernel::KERNEL_ARCH::kCPU;
-using mindspore::lite::KernelRegistrar;
-using mindspore::lite::RET_ERROR;
-using mindspore::lite::RET_OK;
-using mindspore::schema::PrimitiveType_OptMomentum;
-
-namespace mindspore::kernel {
-
-int OptMomentumCPUKernel::ReSize() { return 0; }
-
-int OptMomentumCPUKernel::Run() {
-  auto prepare_ret = Prepare();
-  if (prepare_ret != RET_OK) {
-    MS_LOG(ERROR) << "Prepare fail!ret: " << prepare_ret;
-    return prepare_ret;
-  }
-  if (inputs_.size() != 5 || !outputs_.empty()) {
-    MS_LOG(ERROR) << "OptMomentumCPUKernel error input output size!";
-    return RET_ERROR;
-  }
-
-  if (inputs_[0]->ElementsNum() != inputs_[1]->ElementsNum() ||
-      inputs_[0]->ElementsNum() != inputs_[3]->ElementsNum()) {
-    MS_LOG(ERROR) << "error input data size!";
-    return RET_ERROR;
-  }
-  auto weight = reinterpret_cast<float *>(inputs_[0]->Data());
-  auto accumulate = reinterpret_cast<float *>(inputs_[1]->Data());
-  float learning_rate = reinterpret_cast<float *>(inputs_[2]->Data())[0];
-  auto gradient = reinterpret_cast<float *>(inputs_[3]->Data());
-  float moment = reinterpret_cast<float *>(inputs_[4]->Data())[0];
-  size_t elem_num = inputs_[0]->ElementsNum();
-  for (size_t i = 0; i < elem_num; ++i) {
-    accumulate[i] = accumulate[i] * moment + gradient[i];
-    weight[i] -= accumulate[i] * learning_rate;
-  }
-  return RET_OK;
-}
-
-int OptMomentumCPUKernel::Init() { return 0; }
-
-kernel::LiteKernel *CpuOptMomentumFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
-                                                    const std::vector<lite::tensor::Tensor *> &outputs,
-                                                    OpParameter *opParameter, const lite::Context *ctx,
-                                                    const kernel::KernelKey &desc,
-                                                    const mindspore::lite::PrimitiveC *primitive) {
-  MS_ASSERT(desc.type == schema::PrimitiveType_OptMomentum);
-  auto *kernel = new (std::nothrow) OptMomentumCPUKernel(opParameter, inputs, outputs, ctx, primitive);
-  if (kernel == nullptr) {
-    MS_LOG(ERROR) << "new OptMomentumCPUKernel fail!";
-    return nullptr;
-  }
-
-  auto ret = kernel->Init();
-  if (0 != ret) {
-    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
-                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
-    delete kernel;
-    return nullptr;
-  }
-  return kernel;
-}
-
-REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_OptMomentum, CpuOptMomentumFp32KernelCreator)
-}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32_grad/pooling_grad.cc

@@ -20,6 +20,7 @@
 #include "nnacl/fp32/pooling.h"
 #include "nnacl/fp32_grad/pooling_grad.h"
 #include "include/errorcode.h"
+// #include "src/train/ops/train_ops.h"
 
 using mindspore::kernel::KERNEL_ARCH::kCPU;
 using mindspore::lite::KernelRegistrar;
@@ -29,9 +30,15 @@ using mindspore::schema::PrimitiveType_PoolingGrad;
 
 namespace mindspore::kernel {
 int PoolingGradCPUKernel::Init() {
-  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *>(opParameter);
+  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *>(op_parameter_);
 
-  auto in_shape = inputs_.at(0)->shape();
+  auto in_shape = in_tensors_.at(0)->shape();
+  auto out_shape = in_tensors_.at(1)->shape();
+
+  if (pool_param->pool_mode_ == PoolMode_AvgPool) {
+    in_shape = in_tensors_.at(1)->shape();
+    out_shape = in_tensors_.at(0)->shape();
+  }
   int input_h = in_shape.at(1);
   int input_w = in_shape.at(2);
 
@@ -40,25 +47,39 @@ int PoolingGradCPUKernel::Init() {
     pool_param->window_h_ = input_h;
   }
 
+  pool_param->input_h_ = in_shape[kNHWC_H];
+  pool_param->input_w_ = in_shape[kNHWC_W];
+  pool_param->input_batch_ = in_shape[kNHWC_N];
+  pool_param->input_channel_ = in_shape[kNHWC_C];
+
   // Emir -- here I assume we get the outputshape in the output tensor
-  auto *out_tensor = outputs_.front();
-  auto out_shape = out_tensor->shape();
+  // auto *out_tensor = out_tensors_.front();
+  // auto out_shape = in_tensors_.at(1)->shape();
+
+  pool_param->output_h_ = out_shape[kNHWC_H];
+  pool_param->output_w_ = out_shape[kNHWC_W];
+  pool_param->output_batch_ = out_shape[kNHWC_N];
+  pool_param->output_channel_ = out_shape[kNHWC_C];
 
-  out_tensor->set_shape(out_shape);
-  out_tensor->set_data_type(inputs_.at(0)->data_type());
   return RET_OK;
 }
 
 int PoolingGradCPUKernel::ReSize() { return RET_OK; }
 
 int PoolingGradCPUKernel::Run() {
-  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *>(opParameter);
-  auto input_ptr = reinterpret_cast<float *>(inputs_.at(0)->Data());
-  auto output_ptr = reinterpret_cast<float *>(outputs_.at(0)->Data());
+  auto prepare_ret = Prepare();
+  if (prepare_ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare fail!ret: " << prepare_ret;
+    return prepare_ret;
+  }
+  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *>(op_parameter_);
+  auto input_ptr = reinterpret_cast<float *>(in_tensors_.at(0)->Data());
+  auto output_ptr = reinterpret_cast<float *>(out_tensors_.at(0)->Data());
 
   if (pool_param->pool_mode_ == PoolMode_MaxPool) {
-    auto ind = reinterpret_cast<int *>(inputs_.at(1)->Data());
-    MaxPoolingGrad(input_ptr, ind, output_ptr, pool_param);
+    auto dx_ptr = reinterpret_cast<float *>(in_tensors_.at(1)->Data());
+    auto dy_ptr = reinterpret_cast<float *>(in_tensors_.at(2)->Data());
+    MaxPoolingGrad(input_ptr, dx_ptr, dy_ptr, output_ptr, pool_param);
   } else {
     AvgPoolingGrad(input_ptr, output_ptr, pool_param);
   }

mindspore/lite/src/runtime/kernel/arm/fp32_grad/pooling_grad.h

@@ -43,6 +43,7 @@ class PoolingGradCPUKernel : public LiteKernel {
  private:
   uint8_t data_shape_{0};
 };
+
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_POOLING_GRAD_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/power_grad.cc

@@ -31,10 +31,10 @@ int PowerGradCPUKernel::Init() { return RET_OK; }
 int PowerGradCPUKernel::ReSize() { return RET_OK; }
 
 int PowerGradCPUKernel::Run() {
-  auto dy_addr = reinterpret_cast<float *>(inputs_.at(0)->Data());
-  auto x_addr = reinterpret_cast<float *>(inputs_.at(1)->Data());
-  auto dx_addr = reinterpret_cast<float *>(outputs_.at(0)->Data());
-  auto size = inputs_.at(0)->ElementsNum();
+  auto dy_addr = reinterpret_cast<float *>(in_tensors_.at(0)->Data());
+  auto x_addr = reinterpret_cast<float *>(in_tensors_.at(1)->Data());
+  auto dx_addr = reinterpret_cast<float *>(out_tensors_.at(0)->Data());
+  auto size = in_tensors_.at(0)->ElementsNum();
 
   float exp = power_ - 1;
   Power(x_addr, &exp, dx_addr, size, scale_, shift_, true);
@@ -47,6 +47,7 @@ int PowerGradCPUKernel::Run() {
   return RET_OK;
 }
 
+
 kernel::LiteKernel *CpuPowerGradFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
                                                   const std::vector<lite::tensor::Tensor *> &outputs,
                                                   OpParameter *opParameter, const lite::Context *ctx,

mindspore/lite/src/runtime/kernel/arm/fp32_grad/power_grad.h

@@ -45,6 +45,7 @@ class PowerGradCPUKernel : public LiteKernel {
   float scale_;
   float shift_;
 };
+
 }  // namespace mindspore::kernel
 
 #endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_POWER_GRAD_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/sparse_softmax_cross_entropy_with_logits.cc

@@ -14,6 +14,7 @@
  * limitations under the License.
  */
 
+#include <math.h>
 #include "src/kernel_registry.h"
 #include "nnacl/softmax_parameter.h"
 #include "nnacl/fp32/softmax.h"
@@ -46,9 +47,10 @@ void SparseSoftmaxCrossEntropyWithLogitsCPUKernel::ForwardPostExecute(const int
   output[0] = total_loss / param->batch_size_;
 }
 
-void SparseSoftmaxCrossEntropyWithLogitsCPUKernel::GradPostExecute(const int *labels, const float *losses,
+void SparseSoftmaxCrossEntropyWithLogitsCPUKernel::GradPostExecute(const int *labels, const float *losses, float *grads,
                                                                    float *output) const {
   size_t row_start = 0;
+  float total_loss = 0;
   for (int i = 0; i < param->batch_size_; ++i) {
     if (labels[i] < 0) {
       MS_LOG(EXCEPTION) << "label value must >= 0";
@@ -56,78 +58,88 @@ void SparseSoftmaxCrossEntropyWithLogitsCPUKernel::GradPostExecute(const int *la
     size_t label = labels[i];
     if (label > param->number_of_classes_) {
       MS_LOG(EXCEPTION) << "error label input!";
-    }
-    for (size_t j = 0; j < param->number_of_classes_; ++j) {
-      size_t index = row_start + j;
-      if (j == label) {
-        output[index] = (losses[index] - 1) / param->batch_size_;
-      } else {
-        output[index] = losses[index] / param->batch_size_;
+    } else {
+      total_loss -= logf(losses[i * param->number_of_classes_ + label]);
+      for (size_t j = 0; j < param->number_of_classes_; ++j) {
+        size_t index = row_start + j;
+        if (j == label) {
+          grads[index] = (losses[index] - 1) / param->batch_size_;
+        } else {
+          grads[index] = losses[index] / param->batch_size_;
+        }
       }
     }
     row_start += param->number_of_classes_;
   }
+  output[0] = total_loss / param->batch_size_;
 }
 
 int SparseSoftmaxCrossEntropyWithLogitsCPUKernel::Run() {
-  auto ins = reinterpret_cast<float *>(inputs_.at(0)->Data());
-  auto labels = reinterpret_cast<int *>(inputs_.at(1)->Data());
-  auto out = reinterpret_cast<float *>(outputs_.at(1)->Data());
+  auto ret = Prepare();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare failed.";
+    return ret;
+  }
+
+  auto ins = reinterpret_cast<float *>(in_tensors_.at(0)->Data());
+  auto labels = reinterpret_cast<int *>(in_tensors_.at(1)->Data());
+  float *out = reinterpret_cast<float *>(out_tensors_.at(0)->Data());
   float *grads = NULL;
-  if (is_train()) {  // outputs_.size() > 1)
-    grads = reinterpret_cast<float *>(outputs_.at(0)->Data());
+  if (is_train() && out_tensors_.size() > 1) {
+    grads = reinterpret_cast<float *>(out_tensors_.at(1)->Data());
   }
-  size_t data_size = inputs_.at(0)->ElementsNum();
+  size_t data_size = in_tensors_.at(0)->ElementsNum();
   float *losses = new (std::nothrow) float[data_size];
   if (losses == nullptr) {
     MS_LOG(ERROR) << "losses is null";
-    return nullptr;
+    return RET_ERROR;
   }
 
-  std::fill(losses, losses + data_size, 0);
-
   MS_ASSERT(out != nullptr);
   MS_ASSERT(labels != nullptr);
   MS_ASSERT(ins != nullptr);
-
-  SoftmaxParameter sm_params;
-  sm_params.n_dim_ = param->n_dim_;
-  sm_params.element_size_ = data_size;
-  sm_params.axis_ = 0;
-  for (int i = 0; i < 4; i++)  // softmax has only 4 params in shape
-    sm_params.input_shape_[i] = param->input_shape_[i];
-  float sum_data[sm_params.input_shape_[sm_params.axis_]] = {0};
-  std::fill(sum_data, sum_data + sm_params.input_shape_[sm_params.axis_], 0);
-  Softmax(ins, losses, sum_data, &sm_params);
-
+  std::fill(losses_, losses_ + data_size, 0);
+  std::fill(sum_data_, sum_data_ + sm_params_.input_shape_[0], 0);
+  Softmax(ins, losses_, sum_data_, &sm_params_);
   if (is_train()) {
-    GradPostExecute(labels, losses, grads);
-  } else {
-    ForwardPostExecute(labels, losses, out);
+    GradPostExecute(labels, losses_, grads, out);
+  } else if (out != nullptr) {
+    ForwardPostExecute(labels, losses_, out);
   }
   return RET_OK;
 }
 
 int SparseSoftmaxCrossEntropyWithLogitsCPUKernel::Init() {
-  if (context_->infer_shape_interrupt_ && !context_->running_) {
-    SetNeedReInit();
-    return RET_OK;
-  }
-  auto dims = inputs_[0]->shape();
+  // if (context_ && context_->infer_shape_interrupt_ && !context_->running_) {
+  // set_need_reinit();
+  //  return RET_OK;
+  // }
+  auto dims = in_tensors_[0]->shape();
   param->n_dim_ = 2;
   param->number_of_classes_ = dims[1];
   param->batch_size_ = dims[0];
   for (unsigned int i = 0; i < dims.size(); i++) param->input_shape_[i] = dims[i];
-  if (2 != this->inputs_.size()) {
+  if (2 != this->in_tensors_.size()) {
     MS_LOG(ERROR) << "softmax entropy loss should have two inputs";
     return RET_ERROR;
   }
-  auto *in0 = inputs_.front();
+  auto *in0 = in_tensors_.front();
   if (in0 == nullptr) {
     MS_LOG(ERROR) << "softmax etropy loss in0 have no data";
     return RET_ERROR;
   }
 
+  size_t data_size = in_tensors_.at(0)->ElementsNum();
+  losses_ = new (std::nothrow) float[data_size];
+  sum_data_ = new (std::nothrow) float[dims[0]];
+  MS_ASSERT(losses_ != nullptr);
+  MS_ASSERT(sum_data_ != nullptr);
+
+  sm_params_.n_dim_ = 2;
+  sm_params_.element_size_ = data_size;
+  sm_params_.axis_ = 1;
+  for (int i = 0; i < dims.size(); i++) sm_params_.input_shape_[i] = dims[i];
+
   return RET_OK;
 }
 

mindspore/lite/src/runtime/kernel/arm/fp32_grad/sparse_softmax_cross_entropy_with_logits.h

@@ -14,31 +14,32 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS_H_
-#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS_H_
+#ifndef MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS_H_
 
 #include <vector>
-#include "src/lite_kernel.h"
+#include "src/train/loss_kernel.h"
 #include "ir/anf.h"
 #include "nnacl/fp32_grad/softmax_grad.h"
 #include "nnacl/fp32/arithmetic.h"
+#include "nnacl/softmax_parameter.h"
 
 namespace mindspore::kernel {
 
-class SparseSoftmaxCrossEntropyWithLogitsCPUKernel : public LiteKernel {
+class SparseSoftmaxCrossEntropyWithLogitsCPUKernel : public LossKernel {
  public:
   explicit SparseSoftmaxCrossEntropyWithLogitsCPUKernel(OpParameter *parameter,
                                                         const std::vector<lite::tensor::Tensor *> &inputs,
                                                         const std::vector<lite::tensor::Tensor *> &outputs,
                                                         const lite::Context *ctx,
                                                         const mindspore::lite::PrimitiveC *primitive)
-      : LiteKernel(parameter, inputs, outputs, ctx, primitive) {
+      : LossKernel(parameter, inputs, outputs, ctx, primitive) {
     param = reinterpret_cast<SoftmaxCrossEntropyParameter *>(parameter);
   }
-  ~SparseSoftmaxCrossEntropyWithLogitsCPUKernel() override = default;
+  ~SparseSoftmaxCrossEntropyWithLogitsCPUKernel() override { delete[] losses_; delete[] sum_data_; }
 
   void ForwardPostExecute(const int *labels, const float *losses, float *output) const;
-  void GradPostExecute(const int *labels, const float *losses, float *output) const;
+  void GradPostExecute(const int *labels, const float *losses, float* grads, float *output) const;
 
   int Init() override;
   int ReSize() override;
@@ -46,7 +47,11 @@ class SparseSoftmaxCrossEntropyWithLogitsCPUKernel : public LiteKernel {
 
  private:
   SoftmaxCrossEntropyParameter *param;
+  SoftmaxParameter sm_params_;
+  float *losses_ = nullptr;
+  float *sum_data_ = nullptr;
 };
+
 }  // namespace mindspore::kernel
 
-#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS_H_
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_SPARSE_SOFTMAX_CROSS_ENTROPY_WITH_LOGITS_H_

mindspore/lite/src/runtime/kernel/arm/fp32_grad/tuple_getitem.cc

@@ -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.
+ */
+
+#include <vector>
+#include "src/runtime/kernel/arm/fp32_grad/tuple_getitem.h"
+#include "schema/model_generated.h"
+#include "src/kernel_registry.h"
+#include "include/errorcode.h"
+
+using mindspore::kernel::KERNEL_ARCH::kCPU;
+using mindspore::lite::KernelRegistrar;
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+using mindspore::schema::PrimitiveType_TupleGetItem;
+
+namespace mindspore::kernel {
+
+int TupleGetItemCPUKernel::Init() {
+  return RET_OK;
+}
+
+int TupleGetItemCPUKernel::ReSize() { return 0; }
+
+int TupleGetItemCPUKernel::Run() {
+  auto ret = Prepare();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Prepare failed.";
+    return RET_ERROR;
+  }
+  auto in = reinterpret_cast<float *>(in_tensors_.at(0)->Data());
+  auto out = reinterpret_cast<float *>(out_tensors_.at(0)->Data());
+
+  memcpy(out, in, in_tensors_.at(0)->Size());
+
+  return RET_OK;
+}
+
+kernel::LiteKernel *CpuTupleGetItemFp32KernelCreator(const std::vector<lite::tensor::Tensor *> &inputs,
+                                                 const std::vector<lite::tensor::Tensor *> &outputs,
+                                                 OpParameter *opParameter, const lite::Context *ctx,
+                                                 const kernel::KernelKey &desc, const lite::PrimitiveC *primitive) {
+  MS_ASSERT(opParameter != nullptr);
+  MS_ASSERT(desc.type == schema::PrimitiveType_TupleGetItem);
+  auto *kernel =
+    new (std::nothrow) TupleGetItemCPUKernel(opParameter, inputs, outputs, ctx, primitive);
+  MS_ASSERT(kernel != nullptr);
+
+  auto ret = kernel->Init();
+  if (RET_OK != ret) {
+    MS_LOG(ERROR) << "Init kernel failed, name: " << opParameter->name_ << ", type: "
+                  << schema::EnumNamePrimitiveType(static_cast<schema::PrimitiveType>(opParameter->type_));
+    delete kernel;
+    return nullptr;
+  }
+  return kernel;
+}
+
+REG_KERNEL(kCPU, kNumberTypeFloat32, PrimitiveType_TupleGetItem, CpuTupleGetItemFp32KernelCreator)
+}  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp32_grad/tuple_getitem.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_RUNTIME_KERNEL_ARM_FP32_GRAD_TUPLE_GETITEM_H_
+#define MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_TUPLE_GETITEM_H_
+
+#include <vector>
+#include "src/lite_kernel.h"
+#include "ir/anf.h"
+
+#include "nnacl/fp32/arithmetic.h"
+
+namespace mindspore::kernel {
+class TupleGetItemCPUKernel : public LiteKernel {
+ public:
+  explicit TupleGetItemCPUKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                             const std::vector<lite::tensor::Tensor *> &outputs, const lite::Context *ctx,
+                             const lite::PrimitiveC *primitive)
+      : LiteKernel(parameter, inputs, outputs, ctx, primitive) {
+    param = parameter;
+  }
+  ~TupleGetItemCPUKernel() override = default;
+
+  int Init() override;
+  int ReSize() override;
+  int Run() override;
+
+ private:
+  OpParameter *param;
+};
+}  // namespace mindspore::kernel
+
+#endif  // MINDSPORE_LITE_SRC_RUNTIME_KERNEL_ARM_FP32_GRAD_TUPLE_GETITEM_H_

mindspore/lite/src/scheduler.cc

@@ -94,8 +94,10 @@ int Scheduler::InferShape(const lite::Model *model, std::vector<tensor::Tensor *
       inputs.emplace_back(tensors->at(size_t(inIndexes->GetAs<uint32_t>(j))));
     }
     auto outIndexes = cNode->outputIndex();
-    for (size_t j = 0; j < outIndexes->size(); j++) {
-      outputs.emplace_back(tensors->at(size_t(outIndexes->GetAs<uint32_t>(j))));
+    if (outIndexes != nullptr) {
+      for (size_t j = 0; j < outIndexes->size(); j++) {
+        outputs.emplace_back(tensors->at(size_t(outIndexes->GetAs<uint32_t>(j))));
+      }
     }
     auto *primitive = model->GetOp(cNode->name()->str());
     if (primitive == nullptr) {

mindspore/lite/src/train/loss_kernel.h

@@ -0,0 +1,34 @@
+/**
+ * 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_TRAIN_LOSS_KERNEL_H_
+#define MINDSPORE_LITE_SRC_TRAIN_LOSS_KERNEL_H_
+#include <vector>
+#include "src/lite_kernel.h"
+namespace mindspore::kernel {
+
+class LossKernel : public LiteKernel {
+ public:
+  LossKernel()  = default;
+  explicit LossKernel(OpParameter *parameter, const std::vector<lite::tensor::Tensor *> &inputs,
+                      const std::vector<lite::tensor::Tensor *> &outputs,
+                      const lite::Context *ctx,
+                      const lite::PrimitiveC *primitive)
+      : LiteKernel(parameter, inputs, outputs, ctx, primitive) {}
+  ~LossKernel() = default;
+};
+
+}  // namespace mindspore::kernel
+#endif  // MINDSPORE_LITE_SRC_TRAIN_LOSS_KERNEL_H_

mindspore/lite/src/train/train_populate_parameter.cc

@@ -0,0 +1,250 @@
+/**
+ * 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 "src/populate_parameter.h"
+#include "src/train/train_populate_parameter.h"
+#include "src/ops/pooling_grad.h"
+#include "nnacl/pooling_parameter.h"
+#include "src/ops/softmax_cross_entropy.h"
+#include "nnacl/fp32_grad/softmax_grad.h"
+#include "src/ops/activation_grad.h"
+#include "nnacl/fp32/activation.h"
+#include "src/ops/conv2d_grad_filter.h"
+#include "src/ops/conv2d_grad_input.h"
+#include "nnacl/conv_parameter.h"
+#include "src/ops/power_grad.h"
+#include "nnacl/power_parameter.h"
+
+namespace mindspore::kernel {
+
+OpParameter *DefaultPopulateParameter(const mindspore::lite::PrimitiveC *primitive) {
+  if (primitive == nullptr) {
+    MS_LOG(ERROR) << "Primitive is nullptr when populating parameter for op.";
+    return nullptr;
+  }
+
+  OpParameter *param = new (std::nothrow) OpParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "new Param for primitive failed.";
+    return nullptr;
+  }
+
+  param->type_ = primitive->Type();
+  return param;
+}
+
+OpParameter *PopulateSoftmaxCrossEntropyParameter(const mindspore::lite::PrimitiveC *primitive) {
+  if (primitive == nullptr) {
+    MS_LOG(ERROR) << "Primitive is nullptr when populating parameter for op.";
+    return nullptr;
+  }
+  SoftmaxCrossEntropyParameter *sce_param = new (std::nothrow) SoftmaxCrossEntropyParameter();
+  if (sce_param == nullptr) {
+    MS_LOG(ERROR) << "new SoftmaxCrossEntropyParameter failed.";
+    return nullptr;
+  }
+  sce_param->op_parameter_.type_ = primitive->Type();
+  return reinterpret_cast<OpParameter *>(sce_param);
+}
+
+OpParameter *PopulatePoolingGradParameter(const mindspore::lite::PrimitiveC *primitive) {
+  if (primitive == nullptr) {
+    MS_LOG(ERROR) << "Primitive is nullptr when populating parameter for op.";
+    return nullptr;
+  }
+  PoolingParameter *pooling_param = new (std::nothrow) PoolingParameter();
+  if (pooling_param == nullptr) {
+    MS_LOG(ERROR) << "new PoolingParameter failed.";
+    return nullptr;
+  }
+  pooling_param->op_parameter_.type_ = primitive->Type();
+  auto pooling_primitive =
+    reinterpret_cast<mindspore::lite::PoolingGrad *>(const_cast<mindspore::lite::PrimitiveC *>(primitive));
+
+  pooling_param->global_ = pooling_primitive->GetGlobal();
+  pooling_param->window_w_ = pooling_primitive->GetWindowW();
+  pooling_param->window_h_ = pooling_primitive->GetWindowH();
+
+  pooling_param->pad_u_ = pooling_primitive->GetPadUp();
+  pooling_param->pad_d_ = pooling_primitive->GetPadDown();
+  pooling_param->pad_l_ = pooling_primitive->GetPadLeft();
+  pooling_param->pad_r_ = pooling_primitive->GetPadRight();
+  pooling_param->stride_w_ = pooling_primitive->GetStrideW();
+  pooling_param->stride_h_ = pooling_primitive->GetStrideH();
+
+  pooling_param->pool_mode_ = PoolMode_No;
+  pooling_param->round_mode_ = RoundMode_No;
+
+  switch (pooling_primitive->GetPoolingMode()) {
+    case schema::PoolMode_MAX_POOLING:
+      pooling_param->pool_mode_ = PoolMode_MaxPool;
+      break;
+    case schema::PoolMode_MEAN_POOLING:
+      pooling_param->pool_mode_ = PoolMode_AvgPool;
+      break;
+    default:
+      break;
+  }
+
+  switch (pooling_primitive->GetRoundMode()) {
+    case schema::RoundMode_FLOOR:
+      pooling_param->round_mode_ = RoundMode_Floor;
+      break;
+    case schema::RoundMode_CEIL:
+      pooling_param->round_mode_ = RoundMode_Ceil;
+      break;
+    default:
+      break;
+  }
+  return reinterpret_cast<OpParameter *>(pooling_param);
+}
+
+OpParameter *PopulateActivationGradParameter(const mindspore::lite::PrimitiveC *primitive) {
+  if (primitive == nullptr) {
+    MS_LOG(ERROR) << "Primitive is nullptr when populating parameter for op.";
+    return nullptr;
+  }
+
+  ActivationParameter *act_param = new (std::nothrow) ActivationParameter();
+  if (act_param == nullptr) {
+    MS_LOG(ERROR) << "new ActivationParameter failed.";
+    return nullptr;
+  }
+  act_param->op_parameter_.type_ = primitive->Type();
+  auto activation =
+    reinterpret_cast<mindspore::lite::ActivationGrad *>(const_cast<mindspore::lite::PrimitiveC *>(primitive));
+  act_param->type_ = static_cast<int>(activation->GetType());
+  act_param->alpha_ = activation->GetAlpha();
+  return reinterpret_cast<OpParameter *>(act_param);
+}
+
+OpParameter *PopulateConvolutionGradFilterParameter(const mindspore::lite::PrimitiveC *primitive) {
+  if (primitive == nullptr) {
+    MS_LOG(ERROR) << "Primitive is nullptr when populating parameter for op.";
+    return nullptr;
+  }
+
+  ConvParameter *param = new (std::nothrow) ConvParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "new Param for conv grad filter failed.";
+    return nullptr;
+  }
+  param->op_parameter_.type_ = primitive->Type();
+
+  auto convg_primitive =
+    reinterpret_cast<mindspore::lite::Conv2DGradFilter *>(const_cast<mindspore::lite::PrimitiveC *>(primitive));
+  param->kernel_h_   = convg_primitive->GetKernelH();
+  param->kernel_w_   = convg_primitive->GetKernelW();
+  param->stride_h_   = convg_primitive->GetStrideH();
+  param->stride_w_   = convg_primitive->GetStrideW();
+  param->dilation_h_ = convg_primitive->GetDilateH();
+  param->dilation_w_ = convg_primitive->GetDilateW();
+  param->pad_u_ = convg_primitive->GetPadUp();
+  param->pad_d_ = convg_primitive->GetPadDown();
+  param->pad_l_ = convg_primitive->GetPadLeft();
+  param->pad_r_ = convg_primitive->GetPadRight();
+  param->group_ = convg_primitive->GetGroup();
+  param->act_type_ = ActType_No;
+  switch (convg_primitive->GetActivationType()) {
+    case schema::ActivationType_RELU:
+      param->act_type_ = ActType_Relu;
+      break;
+    case schema::ActivationType_RELU6:
+      param->act_type_ = ActType_Relu6;
+      break;
+    default:
+      break;
+  }
+
+  return reinterpret_cast<OpParameter *>(param);
+}
+
+OpParameter *PopulateConvolutionGradInputParameter(const mindspore::lite::PrimitiveC *primitive) {
+  if (primitive == nullptr) {
+    MS_LOG(ERROR) << "Primitive is nullptr when populating parameter for op.";
+    return nullptr;
+  }
+
+  ConvParameter *param = new (std::nothrow) ConvParameter();
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "new Param for conv grad filter failed.";
+    return nullptr;
+  }
+  param->op_parameter_.type_ = primitive->Type();
+
+  auto convg_primitive =
+    reinterpret_cast<mindspore::lite::Conv2DGradInput *>(const_cast<mindspore::lite::PrimitiveC *>(primitive));
+  param->kernel_h_   = convg_primitive->GetKernelH();
+  param->kernel_w_   = convg_primitive->GetKernelW();
+  param->stride_h_   = convg_primitive->GetStrideH();
+  param->stride_w_   = convg_primitive->GetStrideW();
+  param->dilation_h_ = convg_primitive->GetDilateH();
+  param->dilation_w_ = convg_primitive->GetDilateW();
+  param->pad_u_ = convg_primitive->GetPadUp();
+  param->pad_d_ = convg_primitive->GetPadDown();
+  param->pad_l_ = convg_primitive->GetPadLeft();
+  param->pad_r_ = convg_primitive->GetPadRight();
+  param->group_ = convg_primitive->GetGroup();
+  param->act_type_ = ActType_No;
+  switch (convg_primitive->GetActivationType()) {
+    case schema::ActivationType_RELU:
+      param->act_type_ = ActType_Relu;
+      break;
+    case schema::ActivationType_RELU6:
+      param->act_type_ = ActType_Relu6;
+      break;
+    default:
+      break;
+  }
+
+  return reinterpret_cast<OpParameter *>(param);
+}
+
+OpParameter *PopulatePowerGradParameter(const mindspore::lite::PrimitiveC *primitive) {
+  if (primitive == nullptr) {
+    MS_LOG(ERROR) << "Primitive is nullptr when populating parameter for op.";
+    return nullptr;
+  }
+
+  PowerParameter *power_param = new (std::nothrow) PowerParameter();
+  if (power_param == nullptr) {
+    MS_LOG(ERROR) << "new PowerParameter failed.";
+    return nullptr;
+  }
+  power_param->op_parameter_.type_ = primitive->Type();
+  auto power = reinterpret_cast<mindspore::lite::PowerGrad *>(const_cast<mindspore::lite::PrimitiveC *>(primitive));
+  power_param->power_ = power->GetPower();
+  power_param->scale_ = power->GetScale();
+  power_param->shift_ = power->GetShift();
+  return reinterpret_cast<OpParameter *>(power_param);
+}
+
+void PopulateTrainParameters() {
+  auto ppr = PopulateParameterRegistry::GetInstance();
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_ApplyMomentum, DefaultPopulateParameter);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_BiasGrad, PopulateArithmetic);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_SoftmaxCrossEntropy, PopulateSoftmaxCrossEntropyParameter);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_ActivationGrad, PopulateActivationGradParameter);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_TupleGetItem, DefaultPopulateParameter);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_Depend, DefaultPopulateParameter);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_BNGrad, DefaultPopulateParameter);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_Conv2DGradFilter, PopulateConvolutionGradFilterParameter);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_Conv2DGradInput, PopulateConvolutionGradInputParameter);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_PoolingGrad, PopulatePoolingGradParameter);
+  ppr->AddPopulateParameterFunc(schema::PrimitiveType_PowerGrad, PopulatePowerGradParameter);
+}
+
+}  // namespace mindspore::kernel

mindspore/lite/src/train/train_populate_parameter.h

@@ -0,0 +1,28 @@
+/**
+ * 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_TRAIN_TRAIN_POPULATE_PARAMETER_H_
+#define MINDSPORE_LITE_SRC_TRAIN_TRAIN_POPULATE_PARAMETER_H_
+
+#include "src/ops/primitive_c.h"
+
+namespace mindspore::kernel {
+
+  void PopulateTrainParameters();
+
+
+}  // namespace mindspore::kernel
+#endif  // MINDSPORE_LITE_SRC_TRAIN_TRAIN_POPULATE_PARAMETER_H_

mindspore/lite/src/train/train_session.cc

@@ -0,0 +1,136 @@
+/**
+ * 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 "include/train_session.h"
+#include <algorithm>
+#include "utils/log_adapter.h"
+#include "include/context.h"
+#include "src/common/utils.h"
+#include "mindspore/lite/src/ir/tensor.h"
+#include "src/train/loss_kernel.h"
+#include "src/train/train_populate_parameter.h"
+#include "src/runtime/runtime_api.h"
+#include "src/executor.h"
+#include "src/kernel_registry.h"
+#include "src/runtime/kernel/arm/fp32_grad/convolution.h"
+
+namespace mindspore::session {
+
+TrainSession::TrainSession() { kernel::PopulateTrainParameters(); }
+
+void TrainSession::ReplaceOps() {
+  mindspore::lite::KernelRegistrar tmp(mindspore::kernel::KERNEL_ARCH::kCPU, kNumberTypeFloat32,
+                                       mindspore::schema::PrimitiveType_Conv2D,
+                                       mindspore::kernel::CpuConvTrainFp32KernelCreator);
+}
+
+int TrainSession::CompileGraph(lite::Model *model) {
+  model_ = model;
+  ReplaceOps();
+  return LiteSession::CompileGraph(model);
+}
+
+void* TrainSession::ExportToBuf(void* buf, size_t *len) const {
+//  auto train_model_impl = (dynamic_cast<lite::train::TrainModelImpl*>(model_->model_impl()));
+//  return train_model_impl->ExportToBuf(buf, len);
+  return nullptr;
+}
+
+
+int TrainSession::RunGraph(const session::KernelCallBack &before, const session::KernelCallBack &after) {
+  auto ms_output_tensors = GetOutputs();
+  this->outputs_.clear();
+  for (auto ms_tensors : ms_output_tensors)
+    for (auto ms_tensor : ms_tensors.second)
+    this->outputs_.push_back((dynamic_cast<lite::tensor::LiteTensor*>(ms_tensor))->tensor());
+  if (train_mode_)
+    return LiteSession::RunGraph(before, after);
+
+  // object is expected to run only inference part of graph
+  // prepare a lit of kernels till the loss function -- temporary solution
+  std::vector<kernel::LiteKernel *> infference_kernels;
+  for (auto kernel : this->kernels_) {
+    if (dynamic_cast<const kernel::LossKernel*>(kernel) != nullptr)
+      break;
+    infference_kernels.push_back(kernel);
+  }
+
+  MS_EXCEPTION_IF_NULL(this->context_);
+  // TODO(Emir)
+  // SetMaxWokerNum(context_->thread_num_);
+  // context_->running_ = true;
+  lite::Executor executor;
+  if (before == nullptr && after == nullptr) {
+    return executor.Run(this->inputs_, this->outputs_, infference_kernels, this->context_->allocator.get());
+  } else {
+    return executor.Run(this->inputs_, this->outputs_, infference_kernels, this->context_->allocator.get(),
+           before, after);
+  }
+}
+
+void TrainSession::train() {
+  for (auto *kernel : kernels_) {
+    MS_ASSERT(nullptr != kernel);
+    kernel->train();
+  }
+  train_mode_ = true;
+  ext_output_map_.clear();
+  for (auto kernel : this->kernels_) {
+    if (dynamic_cast<const kernel::LossKernel*>(kernel) != nullptr) {
+      auto *ms_tensor = new lite::tensor::LiteTensor(kernel->out_tensors().at(0));
+      ext_output_map_[kernel->name()].emplace_back(ms_tensor);
+    }
+  }
+}
+
+void TrainSession::eval() {
+  for (auto *kernel : kernels_) {
+    MS_ASSERT(nullptr != kernel);
+    kernel->eval();
+  }
+  train_mode_ = false;
+  kernel::LiteKernel* last_kernel = nullptr;
+  // We should get in_kernels and then get all last kernels
+  ext_output_map_ = output_node_map_;
+  for (auto kernel : this->kernels_) {
+    if ((dynamic_cast<const kernel::LossKernel*>(kernel) != nullptr) &&
+        (last_kernel != nullptr)) {
+      auto *ms_tensor = new lite::tensor::LiteTensor(last_kernel->out_tensors().at(0));
+      ext_output_map_[last_kernel->name()].emplace_back(ms_tensor);
+    }
+    last_kernel = kernel;
+  }
+}
+
+std::unordered_map<std::string, std::vector<mindspore::tensor::MSTensor *>> TrainSession::GetOutputs() const {
+  return ext_output_map_;
+}
+std::vector<tensor::MSTensor *> TrainSession::GetOutputsByName(const std::string &name) const {
+  auto ret_vect = LiteSession::GetOutputsByNodeName(name);  // TODO(emir):  GetOutputsByTensorName?
+  if (ret_vect.size() > 0)
+    return ret_vect;
+  auto ret = ext_output_map_.find(name);
+  if (ret == ext_output_map_.end()) {
+    MS_LOG(WARNING) << "Node  " << name << " is not an output node";
+    std::vector<mindspore::tensor::MSTensor *> empty_ret;
+    return empty_ret;
+  }
+  return ret->second;
+}
+
+
+
+}  // namespace mindspore::session

mindspore/lite/test/CMakeLists.txt

@@ -259,6 +259,10 @@ endif()
 if (SUPPORT_TRAIN)
     set(TEST_LITE_SRC
             ${TEST_LITE_SRC}
+           # ${LITE_DIR}/src/train/ops/train_ops.cc
+            ${LITE_DIR}/src/train/train_populate_parameter.cc
+            ${LITE_DIR}/src/train/train_session.cc
+            ${LITE_DIR}/src/lite_session.cc
             #            ${SRC_DIR}/common/trans.cc
             #            ${SRC_DIR}/common/lite/trans_extends.cc
             #            ${SRC_DIR}/kernel/kernel_build_info.cc

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/activation_grad_fp32_tests.cc

@@ -25,9 +25,10 @@
 #include "mindspore/lite/src/ir/tensor.h"
 #include "mindspore/lite/src/lite_kernel.h"
 #include "mindspore/lite/src/runtime/kernel/arm/fp32_grad/activation_grad.h"
+#include "nnacl/fp32_grad/activation_grad.h"
 
 namespace mindspore {
-class TestActGradFp32 :  public mindspore::CommonTest {
+class TestActGradFp32 : public mindspore::CommonTest {
  public:
   TestActGradFp32() {}
 };
@@ -41,13 +42,14 @@ 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));
+  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));
-
+  EXPECT_EQ(input_size, output_data_size * sizeof(float));
   auto output_data = new float[output_data_size];
   // warm up loop
   for (int i = 0; i < 3; i++) {
-    ReluGrad(yt_data, input_data, 50, output_data);
+    ReluGrad(yt_data, input_data, output_data_size, output_data);
   }
 
   int loop_count = 100;
@@ -72,9 +74,9 @@ TEST_F(TestActGradFp32, ReluGradFp32) {
 
   EXPECT_EQ(res, 0);
 
-  delete input_data;
+  delete[] input_data;
   delete[] output_data;
-  delete yt_data;
+  delete[] yt_data;
 
   MS_LOG(INFO) << "ReluGradFp32 passed";
 }
@@ -118,9 +120,9 @@ TEST_F(TestActGradFp32, Relu6GradFp32) {
 
   EXPECT_EQ(res, 0);
 
-  delete input_data;
+  delete[] input_data;
   delete[] output_data;
-  delete yt_data;
+  delete[] yt_data;
 
   MS_LOG(INFO) << "Relu6GradFp32 passed";
 }
@@ -164,9 +166,9 @@ TEST_F(TestActGradFp32, LReluGradFp32) {
 
   EXPECT_EQ(res, 0);
 
-  delete input_data;
+  delete[] input_data;
   delete[] output_data;
-  delete yt_data;
+  delete[] yt_data;
 
   MS_LOG(INFO) << "LReluGradFp32 passed";
 }
@@ -211,9 +213,9 @@ TEST_F(TestActGradFp32, SigmoidGradFp32) {
   EXPECT_EQ(res, 0);
   // lite::CompareOutput(output_data, output_path);
 
-  delete input_data;
+  delete[] input_data;
   delete[] output_data;
-  delete yt_data;
+  delete[] yt_data;
 
   MS_LOG(INFO) << "SigmoidGradFp32 passed";
 }
@@ -257,9 +259,9 @@ TEST_F(TestActGradFp32, tanhGradFp32) {
 
   EXPECT_EQ(res, 0);
 
-  delete input_data;
+  delete[] input_data;
   delete[] output_data;
-  delete yt_data;
+  delete[] yt_data;
   MS_LOG(INFO) << "TanhGradFp32 passed";
 }
 
@@ -267,24 +269,25 @@ TEST_F(TestActGradFp32, hswishGradFp32) {
   // runtime part
   printf("Calculating runtime cost...\n");
   uint64_t time_avg = 0;
-  size_t output_data_size = 50;
+  const size_t output_data_size = 10;
 
   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));
+  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));
-
+  EXPECT_EQ(input_size, output_data_size * sizeof(float));
   auto output_data = new float[output_data_size];
   // warm up loop
   for (int i = 0; i < 3; i++) {
-    HSwishGrad(yt_data, input_data, 50, output_data);
+    HSwishGrad(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++) {
-    HSwishGrad(yt_data, input_data, 50, output_data);
+    HSwishGrad(yt_data, input_data, output_data_size, output_data);
   }
   auto time_end = mindspore::lite::GetTimeUs();
   auto cost = time_end - time_start;
@@ -292,7 +295,7 @@ TEST_F(TestActGradFp32, hswishGradFp32) {
   printf("single thread running time : %f ms\n", time_avg / 1000.0f);
 
   printf("==================output data=================\n");
-  for (int i = 0; i < 20; i++) {
+  for (int i = 0; i < std::min(output_data_size, 20UL); i++) {
     std::cout << output_data[i] << " ,";
   }
   std::cout << std::endl;
@@ -302,9 +305,9 @@ TEST_F(TestActGradFp32, hswishGradFp32) {
 
   EXPECT_EQ(res, 0);
 
-  delete input_data;
+  delete[] input_data;
   delete[] output_data;
-  delete yt_data;
+  delete[] yt_data;
   MS_LOG(INFO) << "hswishGradFp32 passed";
 }
 

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/arithmetic_grad_fp32_tests.cc

@@ -106,9 +106,14 @@ TEST_F(TestArithmeticGradFp32, TestAddGradFp32) {
 
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_1_dx2_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
-
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // delete all_tensors;
+  // delete param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestAddGradFp32 passed";
 }
 
@@ -137,9 +142,14 @@ TEST_F(TestArithmeticGradFp32, TestAddGrad2Fp32) {
 
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_1_dx2_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
-
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // for (int i = 0; i < 5; i++) delete all_tensors[i]; //TODO tensor data is unique pointer
+  // delete param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestAddGrad2Fp32 passed";
 }
 
@@ -169,8 +179,14 @@ TEST_F(TestArithmeticGradFp32, TestAddGrad3Fp32) {
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_8_dx1_5_4_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
 
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // for (int i = 0; i < 5; i++) delete all_tensors[i];
+  // delete param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestAddGrad3Fp32 passed";
 }
 
@@ -200,8 +216,14 @@ TEST_F(TestArithmeticGradFp32, TestSubGradFp32) {
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_2_dx2_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
 
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // for (int i = 0; i < 5; i++) delete all_tensors[i];
+  // delete param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestSubGradFp32 passed";
 }
 
@@ -231,8 +253,12 @@ TEST_F(TestArithmeticGradFp32, TestSubGrad2Fp32) {
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_3_dx2_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
 
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  delete kernel_obj;
   MS_LOG(INFO) << "TestSubGrad2Fp32 passed";
 }
 
@@ -271,9 +297,13 @@ TEST_F(TestArithmeticGradFp32, TestMulGradFp32) {
 
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_4_dx2_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
-
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  delete kernel_obj;
+  // delete param;
   MS_LOG(INFO) << "TestMulGradFp32 passed";
 }
 
@@ -302,9 +332,14 @@ TEST_F(TestArithmeticGradFp32, TestMulGrad2Fp32) {
 
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_4_dx2_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
-
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // for (int i = 0; i < 5; i++) delete all_tensors[i];
+  // delete param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestMulGrad2Fp32 passed";
 }
 
@@ -333,9 +368,14 @@ TEST_F(TestArithmeticGradFp32, TestMulGrad3Fp32) {
 
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_9_dx2_5_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
-
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // for (int i = 0; i < 5; i++) delete all_tensors[i];
+  // delete param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestMulGrad3Fp32 passed";
 }
 
@@ -364,9 +404,14 @@ TEST_F(TestArithmeticGradFp32, TestMulGrad4Fp32) {
 
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_9_dx2_5_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
-
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // for (int i = 0; i < 5; i++) delete all_tensors[i];
+  // delete param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestMulGrad4Fp32 passed";
 }
 
@@ -395,9 +440,14 @@ TEST_F(TestArithmeticGradFp32, TestDivGradFp32) {
 
   std::string dx2_path = "./test_data/operators/arithmetic_fp32_5_dx2_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, dx2_path));
-
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // for (int i = 0; i < 5; i++) delete all_tensors[i];
+  delete kernel_obj;
+  // delete param;
   MS_LOG(INFO) << "TestDivGradFp32 passed";
 }
 
@@ -427,8 +477,14 @@ TEST_F(TestArithmeticGradFp32, TestDivGrad2Fp32) {
   std::string output_path = "./test_data/operators/arithmetic_fp32_6_dx1_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, output_path));
 
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // for (int i = 0; i < 5; i++) delete all_tensors[i];
+  // delete param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestDivGrad2Fp32 passed";
 }
 
@@ -457,9 +513,14 @@ TEST_F(TestArithmeticGradFp32, TestDivGrad3Fp32) {
 
   std::string output_path = "./test_data/operators/arithmetic_fp32_10_dx2_5_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, output_path));
-
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  // for (int i = 0; i < 5; i++) delete all_tensors[i];
+  // delete param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestDivGrad3Fp32 passed";
 }
 
@@ -488,9 +549,12 @@ TEST_F(TestArithmeticGradFp32, Test3DDivGrad2Fp32) {
 
   std::string output_path = "./test_data/operators/arithmetic_fp32_7_dx2_1_1_6.bin";
   EXPECT_EQ(0, lite::CompareRelativeOutput(output_ptr, output_path));
-
-  for (int i = 0; i < 5; i++) delete all_tensors[i];
-  delete param;
+  for (auto tensor : all_tensors) {
+    delete[] reinterpret_cast<float *>(tensor->Data());
+    tensor->SetData(nullptr);
+    delete tensor;
+  }
+  delete kernel_obj;
   MS_LOG(INFO) << "TestDivGrad2Fp32 passed";
 }
 

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bias_grad_fp32_tests.cc

@@ -18,8 +18,8 @@
 #include "utils/log_adapter.h"
 #include "common/common_test.h"
 #include "src/common/file_utils.h"
-#include "mindspore/lite/src/runtime/kernel/arm/fp32_grad/bias_grad.h"
-#include "mindspore/lite/src/kernel_registry.h"
+#include "src/runtime/kernel/arm/fp32_grad/bias_grad.h"
+#include "src/kernel_registry.h"
 
 namespace mindspore {
 
@@ -40,9 +40,8 @@ TEST_F(TestBiasGradFp32, BiasGradFp32) {
   dy_tensor.SetData(input_data);
 
   std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor};
-
   auto output_data = new float[7];
-  std::vector<int> dim_dw({7});
+  std::vector<int> dim_dw = {7};
   lite::tensor::Tensor dw_tensor(TypeId::kNumberTypeFloat32, dim_dw);
   dw_tensor.SetData(output_data);
   std::vector<lite::tensor::Tensor *> outputs = {&dw_tensor};
@@ -62,9 +61,12 @@ TEST_F(TestBiasGradFp32, BiasGradFp32) {
   std::string output_path = "./test_data/operators/biasgradfp32_1_db_7.bin";
   lite::CompareOutput(output_data, output_path);
 
-  // delete input_data;
-  // delete[] output_data;
-  delete bias_param;
+  delete [] input_data;
+  delete[] output_data;
+  // delete bias_param;
+  dy_tensor.SetData(nullptr);
+  dw_tensor.SetData(nullptr);
+  delete kernel_obj;
   MS_LOG(INFO) << "BiasGradFp32 passed";
 }
 

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/bn_grad_fp32_test.cc

@@ -0,0 +1,111 @@
+/**
+ * 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 <iostream>
+#include <memory>
+#include "utils/log_adapter.h"
+#include "common/common_test.h"
+#include "src/common/file_utils.h"
+#include "src/common/file_utils_ext.h"
+#include "src/runtime/kernel/arm/fp32_grad/bn_grad.h"
+#include "nnacl/fp32_grad/batch_norm.h"
+#include "src/kernel_registry.h"
+#
+
+namespace mindspore {
+
+class TestBNGradFp32 : public mindspore::CommonTest {
+ public:
+  TestBNGradFp32() {}
+  lite::tensor::Tensor *CreateInTensor(std::string file_name, std::vector<int> dim);
+};
+
+lite::tensor::Tensor *TestBNGradFp32::CreateInTensor(std::string file_name, std::vector<int> dim) {
+  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::Tensor(TypeId::kNumberTypeFloat32, dim);
+  tensor->SetData(input_data);
+  EXPECT_EQ(input_size, tensor->Size());
+  return tensor;
+}
+
+TEST_F(TestBNGradFp32, BNGradFp32) {
+  // prepare stage
+  auto bn_param = new BNGradParameter();
+  bn_param->epsilon_ = 0.00001;
+  bn_param->momentum_ = 0.1;
+  const int batch = 2;
+  const int channels = 3;
+  const int height = 4;
+  const int width = 5;
+
+  auto dy_tensor = CreateInTensor("./test_data/bngrad/dy_2_4_5_3.bin", {batch, height, width, channels});
+  auto x_tensor = CreateInTensor("./test_data/bngrad/input_x_2_4_5_3.bin", {batch, height, width, channels});
+  auto scale_tensor = CreateInTensor("./test_data/bngrad/scale_3.bin", {1, 1, 1, channels});
+  auto mean_tensor = CreateInTensor("./test_data/bngrad/save_mean_3.bin", {1, 1, 1, channels});
+  auto var_tensor = CreateInTensor("././test_data/bngrad/save_var_3.bin", {1, 1, 1, channels});
+  // prepare output tensors
+  lite::tensor::Tensor dx_tensor(TypeId::kNumberTypeFloat32, {batch, height, width, channels});
+  dx_tensor.MallocData();
+  lite::tensor::Tensor dscale_tensor(TypeId::kNumberTypeFloat32, {1, 1, 1, channels});
+  dscale_tensor.MallocData();
+  lite::tensor::Tensor dbias_tensor(TypeId::kNumberTypeFloat32, {1, 1, 1, channels});
+  dbias_tensor.MallocData();
+
+  std::vector<lite::tensor::Tensor *> inputs = {dy_tensor, x_tensor, scale_tensor, mean_tensor, var_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&dx_tensor, &dscale_tensor, &dbias_tensor};
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_BNGrad};
+
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(bn_param), NULL, desc, nullptr);
+
+  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";
+  std::cout << "==========dx==========\n";
+  auto dx = reinterpret_cast<float *>(outputs[0]->Data());
+  for (int i = 0; i < 7; i++) std::cout << dx[i] << " ";
+  std::cout << "\n=======dscale=======\n";
+  auto dscale = reinterpret_cast<float *>(outputs[1]->Data());
+  for (int i = 0; i < channels; i++) std::cout << dscale[i] << " ";
+  std::cout << "\n";
+  int res = mindspore::lite::CompareRelativeOutput(dscale, "./test_data/bngrad/output_dscale_3.bin");
+  EXPECT_EQ(res, 0);
+  std::cout << "==========dbias==========\n";
+  auto dbias = reinterpret_cast<float *>(outputs[2]->Data());
+  for (int i = 0; i < 3; i++) std::cout << dbias[i] << " ";
+  std::cout << "\n";
+  res = mindspore::lite::CompareRelativeOutput(dscale, "./test_data/bngrad/output_dscale_3.bin");
+  EXPECT_EQ(res, 0);
+  for (auto v : inputs) {
+    delete[] reinterpret_cast<float *>(v->Data());
+    v->SetData(nullptr);
+    // delete v;
+  }
+  delete kernel_obj;
+  MS_LOG(INFO) << "BNGradFp32 passed";
+}
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/convolution_grad_fp32_tests.cc

@@ -21,6 +21,7 @@
 #include "common/common_test.h"
 #include "src/common/file_utils.h"
 #include "src/common/file_utils_ext.h"
+#include "mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution.h"
 #include "mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_filter.h"
 #include "mindspore/lite/src/runtime/kernel/arm/fp32_grad/convolution_grad_input.h"
 #include "mindspore/lite/nnacl/conv_parameter.h"
@@ -130,11 +131,14 @@ TEST_F(TestConvolutionGradFp32, ConvFp32FilterGrad) {
 
   EXPECT_EQ(res, 0);
 
-  // delete input_data;
-  // delete dy_data;
-  // delete [] dw_data;
+  delete [] input_data;
+  delete [] dy_data;
+  delete [] dw_data;
   delete kernel;
-  delete conv_param;
+  // delete conv_param;
+  dw_tensor.SetData(nullptr);
+  x_tensor.SetData(nullptr);
+  dy_tensor.SetData(nullptr);
   MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
 }
 
@@ -193,9 +197,15 @@ TEST_F(TestConvolutionGradFp32, ConvFp32InputGrad) {
   std::string output_path = "./test_data/conv/convfp32_dx_1_28_28_3.bin";
   auto res = lite::CompareRelativeOutput(dx_data, output_path);
   EXPECT_EQ(res, 0);
-
+  delete [] dx_data;
+  delete [] w_data;
+  delete [] dy_data;
+  w_tensor.SetData(nullptr);
+  dy_tensor.SetData(nullptr);
+  dx_tensor.SetData(nullptr);
   delete kernel;
-  delete conv_param;
+  // delete conv_param;
+
   MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
 }
 
@@ -254,11 +264,14 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupFilterGrad) {
   auto res = lite::CompareRelativeOutput(dw_data, output_path);
   EXPECT_EQ(res, 0);
 
-  // delete input_data;
-  // delete dy_data;
-  // delete [] dw_data;
+  delete [] input_data;
+  delete [] dy_data;
+  delete [] dw_data;
+  dw_tensor.SetData(nullptr);
+  x_tensor.SetData(nullptr);
+  dy_tensor.SetData(nullptr);
   delete kernel;
-  delete conv_param;
+  // delete conv_param;
   MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
 }
 
@@ -317,9 +330,15 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupInputGrad) {
   std::string output_path = "./test_data/conv/convfp32_dx_g3_1_28_28_3.bin";
   auto res = lite::CompareRelativeOutput(dx_data, output_path);
   EXPECT_EQ(res, 0);
+  delete [] dx_data;
+  delete [] w_data;
+  delete [] dy_data;
+  dx_tensor.SetData(nullptr);
+  w_tensor.SetData(nullptr);
+  dy_tensor.SetData(nullptr);
 
   delete kernel;
-  delete conv_param;
+  // delete conv_param;
   MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
 }
 
@@ -378,11 +397,14 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationFilterGrad) {
   std::string output_path = "./test_data/conv/convfp32_dw_g3_d2_18_3_3_3.bin";
   auto res = lite::CompareRelativeOutput(dw_data, output_path);
   EXPECT_EQ(res, 0);
-  // delete input_data;
-  // delete dy_data;
-  // delete [] dw_data;
+  delete [] input_data;
+  delete [] dy_data;
+  delete [] dw_data;
+  dw_tensor.SetData(nullptr);
+  dy_tensor.SetData(nullptr);
+  x_tensor.SetData(nullptr);
   delete kernel;
-  delete conv_param;
+  // delete conv_param;
   MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
 }
 
@@ -441,80 +463,93 @@ TEST_F(TestConvolutionGradFp32, ConvFp32GroupDilationInputGrad) {
   std::string output_path = "./test_data/conv/convfp32_dx_g3_d2_1_28_28_3.bin";
   auto res = lite::CompareRelativeOutput(dx_data, output_path);
   EXPECT_EQ(res, 0);
-
+  delete [] dx_data;
+  delete [] w_data;
+  delete [] dy_data;
+  dx_tensor.SetData(nullptr);
+  dy_tensor.SetData(nullptr);
+  w_tensor.SetData(nullptr);
   delete kernel;
-  delete conv_param;
+  // delete conv_param;
   MS_LOG(INFO) << "TestConvolutionGradFp32 Filter Grad passed";
 }
 
-// TEST_F(TestConvolutionGradFp32, ConvGroupDilation) {
-//   // prepare stage
-//   auto conv_param = new ConvParameter();
-//   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));
-//   std::vector<int> dim_x({1, 28, 28, 3});
-//   tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
-//   x_tensor.SetData(x_data);
-
-//   size_t w_size;
-//   std::string w_path = "./test_data/conv/convfp32_w_g3_d2_18_3_3_3.bin";
-//   auto w_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(w_path.c_str(), &w_size));
-//   std::vector<int> dim_w({18, 3, 3, 1});
-//   tensor::Tensor w_tensor(TypeId::kNumberTypeFloat32, dim_w);
-//   w_tensor.SetData(w_data);
-
-//   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];
-//   std::vector<int> dim_y({1, 26, 26, 18});
-//   tensor::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
-//   y_tensor.SetData(y_data);
-
-//   std::vector<tensor::Tensor *> inputs = {&x_tensor, &w_tensor};
-//   std::vector<tensor::Tensor *> outputs = {&y_tensor};
-//   // runtime part
-
-//   printf("Calculating runtime cost...\n");
-//   uint64_t time_avg = 0;
-
-//   lite::Context context;
-//   ;
-//   context.deviceCtx.type = lite::DT_CPU;
-//   context.threadNum = 1;
-
-//   kernel::KernelKey desc = {kernel::kCPU, kNumberTypeFloat32, schema::PrimitiveType_Conv2D};
-//   auto creator = lite::KernelRegistry::GetInstance()->GetKernelCreator(desc);
-//   auto kernel = creator(inputs, outputs, (OpParameter *)conv_param, &context, desc);
-
-//   kernel->train();
-//   EXPECT_EQ(kernel->is_train(), 1);
-
-//   // warm up loop
-//   for (int i = 0; i < 3; i++) {
-//     kernel->Run();
-//   }
-
-//   int loop_count = 100;
-//   auto time_start = mindspore::lite::GetTimeUs();
-//   for (int i = 0; i < loop_count; i++) {
-//     kernel->Run();
-//   }
-//   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);
-
-//   std::string output_path = "./test_data/conv/convfp32_y_g3_d2_1_26_26_18.bin";
-//   auto res = lite::CompareRelativeOutput(y_data, output_path);
-//   EXPECT_EQ(res, 0);
-
-//   delete kernel;
-//   delete conv_param;
-
-//   MS_LOG(INFO) << "TestConvolutionFp32 Filter Grad passed";
-// }
+TEST_F(TestConvolutionGradFp32, ConvGroupDilation) {
+  // prepare stage
+  auto conv_param = new ConvParameter();
+  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));
+  std::vector<int> dim_x({1, 28, 28, 3});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(x_data);
+
+  size_t w_size;
+  std::string w_path = "./test_data/conv/convfp32_w_g3_d2_18_3_3_3.bin";
+  auto w_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(w_path.c_str(), &w_size));
+  std::vector<int> dim_w({18, 3, 3, 1});
+  lite::tensor::Tensor w_tensor(TypeId::kNumberTypeFloat32, dim_w);
+  w_tensor.SetData(w_data);
+
+  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];
+  std::vector<int> dim_y({1, 26, 26, 18});
+  lite::tensor::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  y_tensor.SetData(y_data);
+
+  std::vector<lite::tensor::Tensor *> inputs = {&x_tensor, &w_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&y_tensor};
+  // runtime part
+
+  printf("Calculating runtime cost...\n");
+  uint64_t time_avg = 0;
+
+  lite::Context context;
+  context.device_ctx_.type = lite::DT_CPU;
+  context.thread_num_ = 1;
+
+
+  auto *kernel = new mindspore::kernel::ConvolutionTrainCPUKernel(reinterpret_cast<OpParameter *>(conv_param),
+   inputs, outputs, &context, 0);
+  kernel->Init();
+  // kernel::KernelKey desc = {kernel::kCPU, kNumberTypeFloat32, schema::PrimitiveType_Conv2D};
+  // auto creator = lite::KernelRegistry::GetInstance()->GetKernelCreator(desc);
+  // auto kernel = creator(inputs, outputs, (OpParameter *)conv_param, &context, desc);
+
+  kernel->train();
+  EXPECT_EQ(kernel->is_train(), 1);
+
+  // warm up loop
+  for (int i = 0; i < 3; i++) {
+    kernel->Run();
+  }
+
+  int loop_count = 100;
+  auto time_start = mindspore::lite::GetTimeUs();
+  for (int i = 0; i < loop_count; i++) {
+    kernel->Run();
+  }
+  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);
+
+  std::string output_path = "./test_data/conv/convfp32_y_g3_d2_1_26_26_18.bin";
+  auto res = lite::CompareRelativeOutput(y_data, output_path);
+  EXPECT_EQ(res, 0);
+
+  delete [] y_data;
+  delete [] x_data;
+  delete [] w_data;
+  x_tensor.SetData(nullptr);
+  y_tensor.SetData(nullptr);
+  w_tensor.SetData(nullptr);
+  delete kernel;
+
+  MS_LOG(INFO) << "TestConvolutionFp32 Filter Grad passed";
+}
 
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/network_test.cc

@@ -0,0 +1,564 @@
+/**
+ * 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 <dirent.h>
+#include <climits>
+#include <cmath>
+#include <iostream>
+#include <fstream>
+#include <memory>
+#include <string>
+#include <functional>
+
+#include "mindspore/lite/schema/inner/model_generated.h"
+#include "mindspore/lite/include/model.h"
+#include "common/common_test.h"
+#include "include/train_session.h"
+// #include "include/lite_session.h"
+#include "include/context.h"
+#include "include/errorcode.h"
+#include "utils/log_adapter.h"
+#include "src/common/file_utils.h"
+#include "src/common/file_utils_ext.h"
+
+namespace mindspore {
+class NetworkTest : public mindspore::CommonTest {
+ public:
+  NetworkTest() {}
+};
+
+
+//             INPUT(0)
+//                V
+//        +-------------+
+//        |     ReLU    |
+//        +-------------+
+//  +---output(1) V
+//  |             V   V weights(2) <----+
+//  |     +-------------+               |
+//  |     |    MatMul   |               |
+//  |     +-------------+               |
+//  |   output(3) V                     |
+//  |             V   V weights(4)<-+   |
+//  |     +-------------+           |   |
+//  |     |    Bias     |           |   |
+//  |     +-------------+           |   |
+//  |   output(5) V                 |   |
+//  |             V   V LABELS(6)   |   |
+//  |     +-------------+           |   |
+//  |     | CrossEntropy|           |   |
+//  |     +-------------+           |   |
+//  |  +-dy(7) V  V------------------------->Loss (14)
+//  |  |       V                    |   |
+//  |  |  +-------------+           |   |
+//  |  |  |  BiasGrad   |           |   |
+//  |  |  +-------------+           |   |
+//  |  |          V db(8)           |   |
+//  |  |          +--------Update---+   |
+//  |  +-------+                        |
+//  +------V   V                        |
+//        +-------------+               |
+//        |  MatMul     |               |
+//        +-------------+               |
+//               V dw(9)                |
+//               +-----------Update-----+
+
+
+TEST_F(NetworkTest, tuning_layer) {
+  const int BATCH_SIZE = 32;
+  const int NUM_CLASSES = 10;
+  const int FEATURE_SIZE = 1000;
+  auto meta_graph = std::make_shared<schema::MetaGraphT>();
+  meta_graph->name = "graph";
+  // define nodes
+  {
+    auto node = std::make_unique<schema::CNodeT>();
+    node->inputIndex = {0};
+    node->outputIndex = {1};
+    node->primitive = std::make_unique<schema::PrimitiveT>();
+    node->primitive->value.type = schema::PrimitiveType_Activation;
+    auto primitive = new schema::ActivationT;
+    primitive->type = schema::ActivationType_RELU;
+    node->primitive->value.value = primitive;
+    node->name = "ReLU";
+    meta_graph->nodes.emplace_back(std::move(node));
+  }
+  {
+    auto node = std::make_unique<schema::CNodeT>();
+    node->inputIndex = {1, 2};
+    node->outputIndex = {3};
+    node->primitive = std::make_unique<schema::PrimitiveT>();
+    node->primitive->value.type = schema::PrimitiveType_MatMul;
+    auto primitive = new schema::MatMulT;
+    primitive->transposeA = false;
+    primitive->transposeB = true;
+    node->primitive->value.value = primitive;
+    node->name = "MatMul1";
+    meta_graph->nodes.emplace_back(std::move(node));
+  }
+  {
+    auto node = std::make_unique<schema::CNodeT>();
+    node->inputIndex = {3, 4};
+    node->outputIndex = {5};
+    node->primitive = std::make_unique<schema::PrimitiveT>();
+    node->primitive->value.type = schema::PrimitiveType_BiasAdd;
+    auto primitive = new schema::BiasAddT;
+    primitive->axis.push_back(0);
+    node->primitive->value.value = primitive;
+    node->name = "BiasAdd";
+    meta_graph->nodes.emplace_back(std::move(node));
+  }
+  {
+    auto node = std::make_unique<schema::CNodeT>();
+    node->inputIndex = {5, 6};
+    node->outputIndex = {14, 7};
+    node->primitive = std::make_unique<schema::PrimitiveT>();
+    node->primitive->value.type = schema::PrimitiveType_SoftmaxCrossEntropy;
+    auto primitive = new schema::SoftmaxCrossEntropyT;
+    primitive->axis.push_back(0);
+    node->primitive->value.value = primitive;
+    node->name = "SoftmaxCrossEntropy";
+    meta_graph->nodes.emplace_back(std::move(node));
+  }
+  {
+    auto node = std::make_unique<schema::CNodeT>();
+    node->inputIndex = {7};
+    node->outputIndex = {8};
+    node->primitive = std::make_unique<schema::PrimitiveT>();
+    node->primitive->value.type = schema::PrimitiveType_BiasGrad;
+    auto primitive = new schema::BiasGradT;
+    primitive->axis.push_back(0);
+    node->primitive->value.value = primitive;
+    node->name = "BiasGrad";
+    meta_graph->nodes.emplace_back(std::move(node));
+  }
+  {
+    auto node = std::make_unique<schema::CNodeT>();
+    node->inputIndex = {7, 1};
+    node->outputIndex = {9};
+    node->primitive = std::make_unique<schema::PrimitiveT>();
+    node->primitive->value.type = schema::PrimitiveType_MatMul;
+    auto primitive = new schema::MatMulT;
+    primitive->transposeA = true;
+    primitive->transposeB = false;
+    node->primitive->value.value = primitive;
+    node->name = "MatMul2";
+    meta_graph->nodes.emplace_back(std::move(node));
+  }
+  {
+    auto node = std::make_unique<schema::CNodeT>();
+    node->inputIndex = {2, 10, 11, 9, 12};
+    node->outputIndex = {};
+    node->primitive = std::make_unique<schema::PrimitiveT>();
+    node->primitive->value.type = schema::PrimitiveType_ApplyMomentum;
+    auto primitive = new schema::ApplyMomentumT;
+    node->primitive->value.value = primitive;
+    node->name = "Momentum";
+    meta_graph->nodes.emplace_back(std::move(node));
+  }
+  {
+    auto node = std::make_unique<schema::CNodeT>();
+    node->inputIndex = {4, 13, 11, 8, 12};
+    node->outputIndex = {};
+    node->primitive = std::make_unique<schema::PrimitiveT>();
+    node->primitive->value.type = schema::PrimitiveType_ApplyMomentum;
+    auto primitive = new schema::ApplyMomentumT;
+    node->primitive->value.value = primitive;
+    node->name = "Momentum";
+    meta_graph->nodes.emplace_back(std::move(node));
+  }
+  meta_graph->inputIndex = {6, 0};  // XXX TODO why is it reverse?
+  meta_graph->outputIndex = {5, 14};
+  const int NUM_OF_OUTPUTS = 2;
+
+  auto input0 = std::make_unique<schema::TensorT>();
+  input0->nodeType = schema::NodeType::NodeType_ValueNode;
+  input0->format = schema::Format_NHWC;
+  input0->dataType = TypeId::kNumberTypeFloat32;
+  input0->dims = {BATCH_SIZE, FEATURE_SIZE};
+  input0->offset = -1;
+  meta_graph->allTensors.emplace_back(std::move(input0));
+  // tensor 1 - relu
+  auto relu_out = std::make_unique<schema::TensorT>();
+  relu_out->nodeType = schema::NodeType::NodeType_Parameter;
+  relu_out->format = schema::Format_NHWC;
+  relu_out->dataType = TypeId::kNumberTypeFloat32;
+  relu_out->dims = {BATCH_SIZE, FEATURE_SIZE};
+  relu_out->offset = -1;
+  meta_graph->allTensors.emplace_back(std::move(relu_out));
+  // tensor 2 - matmul weights
+  auto weight = std::make_unique<schema::TensorT>();
+  weight->nodeType = schema::NodeType::NodeType_ValueNode;
+  weight->format = schema::Format_KHWC;
+  weight->dataType = TypeId::kNumberTypeFloat32;
+  weight->dims = {NUM_CLASSES, FEATURE_SIZE};
+  size_t weight_size;
+  char *buf;
+  std::string weight_path = "./test_data/train/train_weight_10_1000.bin";
+  ReadFile(weight_path.c_str(), &weight_size, &buf);
+  ASSERT_NE(nullptr, buf);
+  weight->data.resize(weight_size);
+  std::copy(buf, buf + weight_size, weight->data.data());
+  meta_graph->allTensors.emplace_back(std::move(weight));
+  // tensor 3 - matmul
+  auto input3 = std::make_unique<schema::TensorT>();
+  input3->nodeType = schema::NodeType::NodeType_Parameter;
+  input3->format = schema::Format_NHWC;
+  input3->dataType = TypeId::kNumberTypeFloat32;
+  input3->dims = {BATCH_SIZE, NUM_CLASSES};
+  input3->offset = -1;
+  meta_graph->allTensors.emplace_back(std::move(input3));
+  // tensor 4 - fc bias
+  auto bias = std::make_unique<schema::TensorT>();
+  bias->nodeType = schema::NodeType::NodeType_ValueNode;
+  bias->format = schema::Format_NHWC;
+  bias->dataType = TypeId::kNumberTypeFloat32;
+  bias->dims = {NUM_CLASSES};
+  bias->offset = -1;
+  std::string bias_path = "./test_data/train/train_bias_10.bin";
+  size_t bias_size;
+  ReadFile(bias_path.c_str(), &bias_size, &buf);
+  ASSERT_NE(nullptr, buf);
+  bias->data.resize(bias_size);
+  std::copy(buf, buf + bias_size, bias->data.data());
+  meta_graph->allTensors.emplace_back(std::move(bias));
+
+  // tensor 5 - bias_add
+  auto input5 = std::make_unique<schema::TensorT>();
+  input5->nodeType = schema::NodeType::NodeType_Parameter;
+  input5->format = schema::Format_NHWC;
+  input5->dataType = TypeId::kNumberTypeFloat32;
+  input5->dims = {BATCH_SIZE, NUM_CLASSES};
+  input5->offset = -1;
+  meta_graph->allTensors.emplace_back(std::move(input5));
+  // tensor 6 - Label
+  {
+    auto label = std::make_unique<schema::TensorT>();
+    label->nodeType = schema::NodeType::NodeType_ValueNode;
+    label->format = schema::Format_NHWC;
+    label->dataType = TypeId::kNumberTypeInt32;
+    label->dims = {BATCH_SIZE};
+    label->offset = -1;
+    label->data.resize(BATCH_SIZE * NUM_CLASSES * sizeof(float));
+    int *data = reinterpret_cast<int *>(label->data.data());
+    for (int i = 0; i < BATCH_SIZE; i++)
+      for (int j = 0; j < NUM_CLASSES; j++) *(data + i * NUM_CLASSES + j) = j;
+    meta_graph->allTensors.emplace_back(std::move(label));
+  }
+  // tensor 7 - Softmaxentropy
+  auto input7 = std::make_unique<schema::TensorT>();
+  input7->nodeType = schema::NodeType::NodeType_Parameter;
+  input7->format = schema::Format_NHWC;
+  input7->dataType = TypeId::kNumberTypeFloat32;
+  input7->dims = {BATCH_SIZE, NUM_CLASSES};
+  input7->offset = -1;
+  meta_graph->allTensors.emplace_back(std::move(input7));
+  // tensor 8 - biasGrad
+  auto input8 = std::make_unique<schema::TensorT>();
+  input8->nodeType = schema::NodeType::NodeType_Parameter;
+  input8->format = schema::Format_NHWC;
+  input8->dataType = TypeId::kNumberTypeFloat32;
+  input8->dims = {NUM_CLASSES};
+  input8->offset = -1;
+  meta_graph->allTensors.emplace_back(std::move(input8));
+  // tensor 9 - matmul2
+  auto input9 = std::make_unique<schema::TensorT>();
+  input9->nodeType = schema::NodeType::NodeType_Parameter;
+  input9->format = schema::Format_NHWC;
+  input9->dataType = TypeId::kNumberTypeFloat32;
+  input9->dims = {NUM_CLASSES, FEATURE_SIZE};
+  input9->offset = -1;
+  meta_graph->allTensors.emplace_back(std::move(input9));
+  // tensor 10 weights accumulate
+  auto input10 = std::make_unique<schema::TensorT>();
+  input10->nodeType = schema::NodeType::NodeType_ValueNode;
+  input10->format = schema::Format_NHWC;
+  input10->dataType = TypeId::kNumberTypeFloat32;
+  input10->dims = {NUM_CLASSES, FEATURE_SIZE};
+  input10->offset = -1;
+  size_t input10_size = NUM_CLASSES * FEATURE_SIZE * sizeof(float);
+  input10->data.resize(input10_size);
+  std::fill(input10->data.data(), input10->data.data() + input10_size, 0.f);
+  meta_graph->allTensors.emplace_back(std::move(input10));
+  // tensor 11 - lr
+  {
+    auto lr = std::make_unique<schema::TensorT>();
+    lr->nodeType = schema::NodeType::NodeType_ValueNode;
+    lr->format = schema::Format_NHWC;
+    lr->dataType = TypeId::kNumberTypeFloat32;
+    lr->dims = {1};
+    lr->offset = -1;
+    lr->data.resize(sizeof(float));
+    float *data = reinterpret_cast<float *>(lr->data.data());
+    *data = 0.01f;
+    meta_graph->allTensors.emplace_back(std::move(lr));
+  }
+  // tensor 12  - momentum
+  {
+    auto input12 = std::make_unique<schema::TensorT>();
+    input12->nodeType = schema::NodeType::NodeType_ValueNode;
+    input12->format = schema::Format_NHWC;
+    input12->dataType = TypeId::kNumberTypeFloat32;
+    input12->dims = {1};
+    input12->offset = -1;
+    input12->data.resize(sizeof(float));
+    float *data = reinterpret_cast<float *>(input12->data.data());
+    *data = 0.f;
+    meta_graph->allTensors.emplace_back(std::move(input12));
+  }
+  // tensor 13 - bias accumulate
+  auto input13 = std::make_unique<schema::TensorT>();
+  input13->nodeType = schema::NodeType::NodeType_ValueNode;
+  input13->format = schema::Format_NHWC;
+  input13->dataType = TypeId::kNumberTypeFloat32;
+  input13->dims = {NUM_CLASSES};
+  input13->offset = -1;
+  size_t input13_size = NUM_CLASSES * sizeof(float);
+  input13->data.resize(input13_size);
+  std::fill(input13->data.data(), input13->data.data() + input13_size, 0.f);
+  meta_graph->allTensors.emplace_back(std::move(input13));
+
+  // tensor 14 - loss
+  {
+    auto loss14 = std::make_unique<schema::TensorT>();
+    loss14->nodeType = schema::NodeType::NodeType_ValueNode;
+    loss14->format = schema::Format_NHWC;
+    loss14->dataType = TypeId::kNumberTypeFloat32;
+    loss14->dims = {1};
+    loss14->offset = -1;
+    loss14->data.resize(sizeof(float));
+    float *data = reinterpret_cast<float *>(loss14->data.data());
+    *data = 0.0f;
+    meta_graph->allTensors.emplace_back(std::move(loss14));
+  }
+
+  //================================================================
+  buf = nullptr;
+
+  flatbuffers::FlatBufferBuilder builder(1024);
+  auto offset = schema::MetaGraph::Pack(builder, meta_graph.get());
+  builder.Finish(offset);
+  size_t size = builder.GetSize();
+  const char *content = reinterpret_cast<char *>(builder.GetBufferPointer());
+  std::cout << "build fb size= " << size << "\n";
+
+#if 0  // EXPORT_FILE
+  std::string path = std::string("hcdemo_train.fb");
+  std::ofstream ofs(path);
+  ASSERT_EQ(true, ofs.good());
+  ASSERT_EQ(true, ofs.is_open());
+
+  ofs.seekp(0, std::ios::beg);
+  ofs.write(content, size);
+  ofs.close();
+#endif
+
+  auto model = lite::Model::Import(content, size);
+  ASSERT_NE(nullptr, model);
+  meta_graph.reset();
+  content = nullptr;
+  auto context = new lite::Context;
+  context->device_ctx_.type = lite::DT_CPU;
+  context->cpu_bind_mode_ = lite::NO_BIND;
+  context->thread_num_ = 1;
+  auto session = new session::TrainSession();
+  ASSERT_NE(nullptr, session);
+  session->Init(context);
+  auto ret = session->CompileGraph(model);
+  ASSERT_EQ(lite::RET_OK, ret);
+  session->train();
+
+  auto inputs = session->GetInputs();
+  ASSERT_EQ(inputs.size(), 2);
+  auto inTensor = inputs.at(0);
+  ASSERT_NE(nullptr, inTensor);
+  auto data = inTensor->MutableData();
+  //===================================================
+  size_t input_size;
+  std::string input_path = "./test_data/train/train_input_32_1000.bin";
+  ReadFile(input_path.c_str(), &input_size, &buf);
+  ASSERT_NE(nullptr, buf);
+  auto input_data = reinterpret_cast<float *>(buf);
+  ASSERT_NE(nullptr, input_data);
+  //===================================================
+  ASSERT_EQ(input_size, inTensor->Size());
+  memcpy(data, input_data, input_size);
+
+  auto labelTensor = inputs.at(1);
+  ASSERT_NE(nullptr, labelTensor);
+  ASSERT_EQ(BATCH_SIZE, labelTensor->ElementsNum());
+  auto labels = reinterpret_cast<int *>(labelTensor->MutableData());
+  for (int i = 0; i < BATCH_SIZE; i++) labels[i] = (i * 97) % NUM_CLASSES;
+
+  ret = session->RunGraph();
+  ASSERT_EQ(lite::RET_OK, ret);
+  auto outputs = session->GetOutputsByName("BiasAdd");
+  ASSERT_EQ(outputs.size(), 1);
+  auto outTensor = (outputs.at(0));
+  ASSERT_NE(nullptr, outTensor);
+  ASSERT_EQ(TypeId::kNumberTypeFloat32, outTensor->data_type());
+  auto *outData = reinterpret_cast<float *>(outTensor->MutableData());
+  ASSERT_NE(nullptr, outData);
+  std::cout << "========================dW=====================" << std::endl;
+  for (int i = 0; i < 20; i++) {
+    std::cout << outData[i] << ", ";
+  }
+  std::cout << std::endl;
+  ret = session->RunGraph();
+  outputs = session->GetOutputsByName("BiasAdd");
+  ASSERT_EQ(outputs.size(), 1);
+  outTensor = (outputs.at(0));
+  ASSERT_NE(nullptr, outTensor);
+  // ASSERT_EQ(28 * 28 * 32, outTensor->ElementsNum());
+  ASSERT_EQ(TypeId::kNumberTypeFloat32, outTensor->data_type());
+  outData = reinterpret_cast<float *>(outTensor->MutableData());
+  ASSERT_NE(nullptr, outData);
+  std::cout << "========================dW=====================" << std::endl;
+  for (int i = 0; i < 20; i++) {
+    std::cout << outData[i] << ", ";
+  }
+//===================================================
+#if 0
+  size_t output_size;
+  std::string output_path = "./convfp32_out_1_28_28_32.bin";
+  buf = mindspore::lite::ReadFile(output_path.c_str(), &output_size);
+  ASSERT_NE(nullptr, buf);
+  auto output_data = reinterpret_cast<float *>(buf);
+  ASSERT_NE(nullptr, output_data);
+  //===================================================
+  ASSERT_EQ(output_size, runOutput->Size());
+  for (size_t i = 0; i < runOutput->ElementsNum(); i++) {
+    ASSERT_EQ(output_data[i], outData[i]);
+  }
+#endif
+  MS_LOG(INFO) << "Passed";
+}
+
+int32_t fileIterator(mindspore::session::TrainSession *session, const std::string &path,
+ std::function<int32_t(mindspore::session::TrainSession *session,
+  const std::string &)> cb) {
+  int32_t res = 0;
+  if (auto dir = opendir(path.c_str())) {
+    while (auto f = readdir(dir)) {
+      if (!f->d_name || f->d_name[0] == '.') continue;
+      if (f->d_type == DT_DIR) fileIterator(session, path + f->d_name + "/", cb);
+
+      if (f->d_type == DT_REG)
+       res |= cb(session, path + f->d_name);
+    }
+    closedir(dir);
+  }
+  return res;
+}
+#if 0
+void replaceExt(const std::string &src, std::string *dst) {
+  dst = &std::move(src.substr(0, src.find_last_of('.')) + ".emb");
+}
+#endif
+int32_t runEffNet(mindspore::session::TrainSession *session, const std::string &in, const std::string &out) {
+  // setup input
+  auto inputs = session->GetInputs();
+  // ASSERT_EQ(inputs.size(), 1);
+  auto inTensor = inputs.at(0);
+  // ASSERT_NE(nullptr, inTensor);
+  float *data = reinterpret_cast<float *>(inTensor->MutableData());
+
+  size_t input_size;
+  float *in_buf =  reinterpret_cast<float *>(lite::ReadFile(in.c_str(), &input_size));
+  // ASSERT_NE(nullptr, data);
+  auto input_data = reinterpret_cast<float *>(in_buf);
+  // ASSERT_EQ(input_size, inTensor->Size());
+  std::copy(input_data, input_data + inTensor->ElementsNum(), data);
+
+  // execute network
+  session->RunGraph();
+
+  // compare outputs
+  auto outputs = session->GetOutputs();
+  auto output = ((outputs.begin())->second);
+  float *output_data = reinterpret_cast<float *>(output.at(0)->MutableData());
+
+  return mindspore::lite::CompareRelativeOutput(output_data, out.c_str());
+}
+
+TEST_F(NetworkTest, efficient_net) {
+  const int NUM_OF_INPUTS = 1;
+  char *buf = nullptr;
+  size_t net_size = 0;
+  std::string net = "./test_data/nets/efficientnet_b0_f.ms";
+  ReadFile(net.c_str(), &net_size, &buf);
+  auto model = lite::Model::Import(buf, net_size);
+  auto context = new lite::Context;
+  context->device_ctx_.type = lite::DT_CPU;
+  context->cpu_bind_mode_ = lite::NO_BIND;
+  context->thread_num_ = 1;
+
+
+  auto session = new mindspore::session::TrainSession();
+  ASSERT_NE(session, nullptr);
+  auto ret = session->Init(context);
+  ASSERT_EQ(lite::RET_OK, ret);
+  ret = session->CompileGraph(model);
+  ASSERT_EQ(lite::RET_OK, ret);
+  session->eval();
+
+#if 0
+  std::string path = "/opt/share/MiniBinEmbDataset/";
+  auto res = fileIterator(session, path, [](mindspore::session::TrainSession *session, const std::string &in) {
+    int32_t res = 0;
+    if (in.find(".bin") != std::string::npos) {
+      std::string out;
+      replaceExt(in, out);
+      res = runEffNet(session, in, out);
+      std::cout << "input file: " << in << (res ?  " Fail" : " Pass") << std::endl;
+    }
+    return res;
+  });
+#else
+  std::string in = "./test_data/nets/effNet_input_x_1_3_224_224.bin";
+  std::string out = "./test_data/nets/effNet_output_y_1_1000.bin";
+  auto res = runEffNet(session, in, out);
+#endif
+  // auto inputs = session->GetInputs();
+  // ASSERT_EQ(inputs.size(), NUM_OF_INPUTS);
+  // auto inTensor = inputs.at(0);
+  // ASSERT_NE(nullptr, inTensor);
+  // float *data = reinterpret_cast<float *>(inTensor->MutableData());
+
+  // // fill input
+  // std::string input_path = "./test_data/nets/effNet_input_x_1_3_224_224.bin";
+  // // std::string input_path = "/opt/share/MiniBinEmbDataset/2_pet/n02099601_3111.bin";
+  // size_t input_size;
+  // char *in_buf = nullptr;
+  // ReadFile(input_path.c_str(), &input_size, &in_buf);
+  // ASSERT_NE(nullptr, data);
+  // auto input_data = reinterpret_cast<float *>(in_buf);
+  // ASSERT_EQ(input_size, inTensor->Size());
+  // std::copy(input_data, input_data+inTensor->ElementsNum(), data);
+
+  // // execute network
+  // ret = session->RunGraph();
+
+  // // compare outputs
+  // std::string output_path = "./test_data/nets/effNet_output_y_1_1000.bin";
+  // // std::string output_path = "/opt/share/MiniBinEmbDataset/2_pet/n02099601_3111.emb";
+  // auto outputs = session->GetOutputs();
+  // auto output = ((outputs.begin())->second);
+  // float* output_data = reinterpret_cast<float *>(output.at(0)->MutableData());
+  // int res = lite::CompareRelativeOutput(output_data, output_path);
+  ASSERT_EQ(res, 0);
+}
+
+}  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/pooling_grad_fp32_tests.cc

@@ -22,6 +22,7 @@
 #include "mindspore/lite/src/kernel_registry.h"
 #include "src/common/utils.h"
 #include "src/common/file_utils.h"
+#include "src/common/file_utils_ext.h"
 #include "src/runtime/kernel/arm/fp32_grad/pooling_grad.h"
 #include "nnacl/fp32_grad/pooling_grad.h"
 
@@ -60,6 +61,7 @@ TEST_F(TestPoolingGradFp32, AvgPoolingGradFp32) {
   auto pooling_param = new PoolingParameter();
   InitPoolingParamFP32(pooling_param);
   pooling_param->output_channel_ = 3;
+  pooling_param->pool_mode_ = PoolMode_AvgPool;
 
   // runtime part
   printf("Calculating runtime cost...\n");
@@ -95,7 +97,7 @@ TEST_F(TestPoolingGradFp32, AvgPoolingGradFp32) {
   std::string output_path = "./test_data/pooling/avgpoolgradfp32_1_dx_1_28_28_3.bin";
   lite::CompareOutput(output_data, output_path);
 
-  delete input_data;
+  delete[] input_data;
   delete[] output_data;
   delete pooling_param;
   MS_LOG(INFO) << "TestAvgPoolingGradFp32 passed";
@@ -122,10 +124,10 @@ TEST_F(TestPoolingGradFp32, AvgPoolingKernelGradFp32) {
   dy_tensor.SetData(input_data);
 
   std::string input1_path = "./test_data/pooling/avgpoolgradfp32_1_x_1_28_28_3.bin";
-  input_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input1_path.c_str(), &input_size));
+  auto input1_data = reinterpret_cast<float *>(mindspore::lite::ReadFile(input1_path.c_str(), &input_size));
   std::vector<int> dim_x({1, 28, 28, 3});
   lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
-  x_tensor.SetData(input_data);
+  x_tensor.SetData(input1_data);
 
   std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &x_tensor};
 
@@ -150,12 +152,205 @@ TEST_F(TestPoolingGradFp32, AvgPoolingKernelGradFp32) {
   std::string output_path = "./test_data/pooling/avgpoolgradfp32_1_dx_1_28_28_3.bin";
   lite::CompareOutput(output_data, output_path);
 
-  // delete input_data;
-  // delete[] output_data;
-  delete pooling_param;
+  delete[] input_data;
+  delete[] input1_data;
+  delete[] output_data;
+  dx_tensor.SetData(nullptr);
+  x_tensor.SetData(nullptr);
+  dy_tensor.SetData(nullptr);
+  // delete pooling_param;
+  delete kernel_obj;
   MS_LOG(INFO) << "TestAvgPoolingGradFp32 passed";
 }
 
+TEST_F(TestPoolingGradFp32, AvgPoolingBatchGradFp32) {
+  // prepare stage
+  auto pooling_param = new PoolingParameter();
+  InitPoolingParamFP32(pooling_param);
+
+  pooling_param->output_channel_ = 3;
+  pooling_param->input_batch_ = 3;
+  pooling_param->output_batch_ = 3;
+
+  // runtime part
+  printf("Calculating runtime cost...\n");
+  // uint64_t time_avg = 0;
+  size_t output_data_size =
+    pooling_param->output_batch_ * pooling_param->output_channel_ * pooling_param->input_h_ * pooling_param->input_w_;
+
+  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));
+  std::vector<int> dim_dy({1, 28, 28, 3});
+  lite::tensor::Tensor dy_tensor(TypeId::kNumberTypeFloat32, dim_dy);
+  dy_tensor.SetData(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));
+  std::vector<int> dim_x({1, 28, 28, 3});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(input1_data);
+
+  std::vector<lite::tensor::Tensor *> inputs = {&dy_tensor, &x_tensor};
+
+  auto output_data = new float[output_data_size];
+  std::vector<int> dim_dx({1, 28, 28, 3});
+  lite::tensor::Tensor dx_tensor(TypeId::kNumberTypeFloat32, dim_dx);
+  dx_tensor.SetData(output_data);
+  std::vector<lite::tensor::Tensor *> outputs = {&dx_tensor};
+
+  kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
+
+  auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);
+  auto kernel_obj = creator(inputs, outputs, reinterpret_cast<OpParameter *>(pooling_param), NULL, desc, 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/pooling/avgpoolgradfp32_1_dx_3_28_28_3.bin";
+  lite::CompareOutput(output_data, output_path);
+
+  delete[] input_data;
+  delete[] input1_data;
+  delete[] output_data;
+  dx_tensor.SetData(nullptr);
+  x_tensor.SetData(nullptr);
+  dy_tensor.SetData(nullptr);
+  // delete pooling_param;
+  delete kernel_obj;
+  MS_LOG(INFO) << "TestAvgPoolingGradBatchFp32 passed";
+}
+
+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 = new PoolingParameter();
+  InitPoolingParamFP32(pool);
+  pool->output_channel_ = 3;
+  pool->pool_mode_ = PoolMode_AvgPool;
+  pool->input_batch_ = 3;
+  pool->output_batch_ = 3;
+  pool->output_h_ = 14;
+  pool->output_w_ = 14;
+  pool->stride_h_ = 2;
+  pool->stride_w_ = 2;
+
+  size_t input_size;
+  size_t y_data_size = pool->output_batch_ * pool->output_channel_ * pool->input_h_ * pool->input_w_;
+
+  auto x_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/avgpoolgradfp32_s2_x_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_x({pool->output_batch_, pool->input_h_, pool->input_w_, pool->input_channel_});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(x_data);
+
+  auto yt_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/avgpoolgradfp32_s2_dy_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_y({pool->output_batch_, pool->output_h_, pool->output_w_, pool->output_channel_});
+  lite::tensor::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  yt_tensor.SetData(yt_data);
+
+  auto out_data = new float[y_data_size];
+  lite::tensor::Tensor out_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  out_tensor.SetData(out_data);
+
+  std::vector<lite::tensor::Tensor *> inputs = {&yt_tensor, &x_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&out_tensor};
+  // ----------------------------------------
+  kernel::KernelKey pool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
+  auto pool_creator = lite::KernelRegistry::GetInstance()->GetCreator(pool_desc);
+  auto kernel = pool_creator(inputs, outputs, reinterpret_cast<OpParameter *>(pool), NULL, pool_desc, nullptr);
+
+  kernel->Init();
+
+  auto time_start = mindspore::lite::GetTimeUs();
+  kernel->Run();
+  auto time_end = mindspore::lite::GetTimeUs();
+  printf("single thread running time : %ld ms\n", time_end - time_start);
+
+  std::string output_path = "./test_data/pooling/avgpoolgradfp32_s2_dx_3_28_28_3.bin";
+  auto res = lite::CompareRelativeOutput(out_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete[] x_data;
+  delete[] yt_data;
+  // delete[] out_data;
+  // delete conv_param;
+  x_tensor.SetData(nullptr);
+  yt_tensor.SetData(nullptr);
+  out_tensor.SetData(nullptr);
+  delete kernel;
+  MS_LOG(INFO) << "AvgPoolGradStride2Fp32 Filter Grad passed";
+}
+
+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 = new PoolingParameter();
+  InitPoolingParamFP32(pool);
+  pool->output_channel_ = 3;
+  pool->pool_mode_ = PoolMode_AvgPool;
+  pool->input_batch_ = 3;
+  pool->output_batch_ = 3;
+  pool->output_h_ = 10;
+  pool->output_w_ = 10;
+  pool->stride_h_ = 3;
+  pool->stride_w_ = 3;
+
+  size_t input_size;
+  size_t y_data_size = pool->output_batch_ * pool->output_channel_ * pool->input_h_ * pool->input_w_;
+
+  auto x_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/avgpoolgradfp32_s3_x_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_x({pool->output_batch_, pool->input_h_, pool->input_w_, pool->input_channel_});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(x_data);
+
+  auto yt_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/avgpoolgradfp32_s3_dy_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_y({pool->output_batch_, pool->output_h_, pool->output_w_, pool->output_channel_});
+  lite::tensor::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  yt_tensor.SetData(yt_data);
+
+  auto out_data = new float[y_data_size];
+  lite::tensor::Tensor out_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  out_tensor.SetData(out_data);
+
+  std::vector<lite::tensor::Tensor *> inputs = {&yt_tensor, &x_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&out_tensor};
+  // ----------------------------------------
+  kernel::KernelKey pool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
+  auto pool_creator = lite::KernelRegistry::GetInstance()->GetCreator(pool_desc);
+  auto kernel = pool_creator(inputs, outputs, reinterpret_cast<OpParameter *>(pool), NULL, pool_desc, nullptr);
+
+  kernel->Init();
+
+  auto time_start = mindspore::lite::GetTimeUs();
+  kernel->Run();
+  auto time_end = mindspore::lite::GetTimeUs();
+  printf("single thread running time : %ld ms\n", time_end - time_start);
+
+  std::string output_path = "./test_data/pooling/avgpoolgradfp32_s3_dx_3_28_28_3.bin";
+  auto res = lite::CompareRelativeOutput(out_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete[] x_data;
+  delete[] yt_data;
+  // delete[] out_data;
+  // delete conv_param;
+  x_tensor.SetData(nullptr);
+  yt_tensor.SetData(nullptr);
+  out_tensor.SetData(nullptr);
+  delete kernel;
+  MS_LOG(INFO) << "AvgPoolGradStride3Fp32 Filter Grad passed";
+}
+
 TEST_F(TestPoolingGradFp32, MaxPoolingGradFp32) {
   // prepare stage
   auto pooling_param = new PoolingParameter();
@@ -169,26 +364,25 @@ TEST_F(TestPoolingGradFp32, MaxPoolingGradFp32) {
     pooling_param->output_batch_ * pooling_param->output_channel_ * pooling_param->output_h_ * pooling_param->output_w_;
 
   size_t input_size;
-  std::string i_path = "./test_data/pooling/maxpoolgradfp32_1_i_1_28_28_3.bin";
-  auto ill_data = reinterpret_cast<int64_t *>(mindspore::lite::ReadFile(i_path.c_str(), &input_size));
-  auto i_data = new int[output_data_size];
-  for (uint32_t i = 0; i < output_data_size; i++) {
-    i_data[i] = static_cast<int>(ill_data[i]);
-  }
+  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));
 
   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));
 
+  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));
+
   auto output_data = new float[output_data_size];
   // warm up loop
   for (int i = 0; i < 3; i++) {
-    MaxPoolingGrad(dy_data, i_data, output_data, pooling_param);
+    MaxPoolingGrad(in_data, dx_data, dy_data, output_data, pooling_param);
   }
 
   int loop_count = 100;
   auto time_start = mindspore::lite::GetTimeUs();
   for (int i = 0; i < loop_count; i++) {
-    MaxPoolingGrad(dy_data, i_data, output_data, pooling_param);
+    MaxPoolingGrad(in_data, dx_data, dy_data, output_data, pooling_param);
   }
   auto time_end = mindspore::lite::GetTimeUs();
   auto cost = time_end - time_start;
@@ -200,11 +394,13 @@ TEST_F(TestPoolingGradFp32, MaxPoolingGradFp32) {
     std::cout << output_data[i] << " ,";
   }
   std::cout << std::endl;
-  std::string output_path = "./test_data/pooling/maxpoolgradfp32_1_dx_1_28_28_3.bin";
+  std::string output_path = "./test_data/pooling/maxpoolgradfp32_1_xgrad_1_28_28_3.bin";
   lite::CompareOutput(output_data, output_path);
 
-  // delete input_data;
+  delete[] in_data;
   delete pooling_param;
+  delete[] dy_data;
+  delete[] dx_data;
   delete[] output_data;
   MS_LOG(INFO) << "TestMaxPoolingGradFp32 passed";
 }
@@ -326,4 +522,216 @@ TEST_F(TestPoolingGradFp32, MaxPoolingKernelGradFp32) {
   MS_LOG(INFO) << "TestMaxPoolingKernelGradFp32 passed";
 }
 #endif  // if 0 before MaxPoolingKernelGradFp32
+
+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 = new PoolingParameter();
+  InitPoolingParamFP32(maxpool);
+  maxpool->output_channel_ = 3;
+  maxpool->pool_mode_ = PoolMode_MaxPool;
+  maxpool->input_batch_ = 3;
+  maxpool->output_batch_ = 3;
+
+  size_t input_size;
+  size_t y_data_size = maxpool->output_batch_ * maxpool->output_channel_ * maxpool->input_h_ * maxpool->input_w_;
+
+  auto x_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_1_x_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_x({3, 28, 28, 3});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(x_data);
+
+  auto y_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_1_dx_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_y({3, 28, 28, 3});
+  lite::tensor::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  y_tensor.SetData(y_data);
+
+  auto yt_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_1_dy_3_28_28_3.bin", &input_size));
+  lite::tensor::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  yt_tensor.SetData(yt_data);
+
+  auto out_data = new float[y_data_size];
+  lite::tensor::Tensor out_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  out_tensor.SetData(out_data);
+
+  std::vector<lite::tensor::Tensor *> maxpool_inputs = {&x_tensor, &y_tensor, &yt_tensor};
+  std::vector<lite::tensor::Tensor *> maxpool_outputs = {&out_tensor};
+  // ----------------------------------------
+  kernel::KernelKey maxpool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
+  auto maxpool_creator = lite::KernelRegistry::GetInstance()->GetCreator(maxpool_desc);
+  auto kernel = maxpool_creator(maxpool_inputs, maxpool_outputs, reinterpret_cast<OpParameter *>(maxpool), NULL,
+                                maxpool_desc, nullptr);
+
+  kernel->Init();
+
+  auto time_start = mindspore::lite::GetTimeUs();
+  kernel->Run();
+  auto time_end = mindspore::lite::GetTimeUs();
+  printf("single thread running time : %ld ms\n", time_end - time_start);
+
+  std::string output_path = "./test_data/pooling/maxpoolgradfp32_1_xgrad_3_28_28_3.bin";
+  auto res = lite::CompareRelativeOutput(out_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete[] x_data;
+  delete[] y_data;
+  delete[] yt_data;
+  // delete[] out_data;
+  // delete conv_param;
+  x_tensor.SetData(nullptr);
+  y_tensor.SetData(nullptr);
+  yt_tensor.SetData(nullptr);
+  out_tensor.SetData(nullptr);
+  delete kernel;
+  MS_LOG(INFO) << "MaxPoolGradBatchFp32 Filter Grad passed";
+}
+
+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 = new PoolingParameter();
+  InitPoolingParamFP32(maxpool);
+  maxpool->output_channel_ = 3;
+  maxpool->input_channel_ = 3;
+  maxpool->pool_mode_ = PoolMode_MaxPool;
+  maxpool->input_batch_ = 3;
+  maxpool->output_batch_ = 3;
+  maxpool->output_h_ = 14;
+  maxpool->output_w_ = 14;
+  maxpool->stride_h_ = 2;
+  maxpool->stride_w_ = 2;
+
+  size_t input_size;
+  size_t y_data_size = maxpool->output_batch_ * maxpool->output_channel_ * maxpool->input_h_ * maxpool->input_w_;
+
+  auto x_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s2_x_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_x({maxpool->output_batch_, maxpool->input_h_, maxpool->input_w_, maxpool->input_channel_});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(x_data);
+
+  auto y_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s2_dx_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_y({maxpool->output_batch_, maxpool->output_h_, maxpool->output_w_, maxpool->output_channel_});
+  lite::tensor::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  y_tensor.SetData(y_data);
+
+  auto yt_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s2_dy_3_28_28_3.bin", &input_size));
+  lite::tensor::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  yt_tensor.SetData(yt_data);
+
+  auto out_data = new float[y_data_size];
+  lite::tensor::Tensor out_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  out_tensor.SetData(out_data);
+
+  std::vector<lite::tensor::Tensor *> maxpool_inputs = {&x_tensor, &y_tensor, &yt_tensor};
+  std::vector<lite::tensor::Tensor *> maxpool_outputs = {&out_tensor};
+  // ----------------------------------------
+  kernel::KernelKey maxpool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
+  auto maxpool_creator = lite::KernelRegistry::GetInstance()->GetCreator(maxpool_desc);
+  auto kernel = maxpool_creator(maxpool_inputs, maxpool_outputs, reinterpret_cast<OpParameter *>(maxpool), NULL,
+                                maxpool_desc, nullptr);
+
+  kernel->Init();
+
+  auto time_start = mindspore::lite::GetTimeUs();
+  kernel->Run();
+  auto time_end = mindspore::lite::GetTimeUs();
+  printf("single thread running time : %ld ms\n", time_end - time_start);
+
+  std::string output_path = "./test_data/pooling/maxpoolgradfp32_s2_xgrad_3_28_28_3.bin";
+  auto res = lite::CompareRelativeOutput(out_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete[] x_data;
+  delete[] y_data;
+  delete[] yt_data;
+  // delete[] out_data;
+  // delete conv_param;
+  x_tensor.SetData(nullptr);
+  y_tensor.SetData(nullptr);
+  yt_tensor.SetData(nullptr);
+  out_tensor.SetData(nullptr);
+  delete kernel;
+  MS_LOG(INFO) << "MaxPoolGradStride2Fp32 Filter Grad passed";
+}
+
+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 = new PoolingParameter();
+  InitPoolingParamFP32(maxpool);
+  maxpool->output_channel_ = 3;
+  maxpool->input_channel_ = 3;
+  maxpool->pool_mode_ = PoolMode_MaxPool;
+  maxpool->input_batch_ = 3;
+  maxpool->output_batch_ = 3;
+  maxpool->output_h_ = 10;
+  maxpool->output_w_ = 10;
+  maxpool->stride_h_ = 3;
+  maxpool->stride_w_ = 3;
+
+  size_t input_size;
+  size_t y_data_size = maxpool->output_batch_ * maxpool->output_channel_ * maxpool->input_h_ * maxpool->input_w_;
+
+  auto x_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s3_x_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_x({maxpool->output_batch_, maxpool->input_h_, maxpool->input_w_, maxpool->input_channel_});
+  lite::tensor::Tensor x_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  x_tensor.SetData(x_data);
+
+  auto y_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s3_dx_3_28_28_3.bin", &input_size));
+  std::vector<int> dim_y({maxpool->output_batch_, maxpool->output_h_, maxpool->output_w_, maxpool->output_channel_});
+  lite::tensor::Tensor y_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  y_tensor.SetData(y_data);
+
+  auto yt_data = reinterpret_cast<float *>(
+    mindspore::lite::ReadFile("./test_data/pooling/maxpoolgradfp32_s3_dy_3_28_28_3.bin", &input_size));
+  lite::tensor::Tensor yt_tensor(TypeId::kNumberTypeFloat32, dim_y);
+  yt_tensor.SetData(yt_data);
+
+  auto out_data = new float[y_data_size];
+  lite::tensor::Tensor out_tensor(TypeId::kNumberTypeFloat32, dim_x);
+  out_tensor.SetData(out_data);
+
+  std::vector<lite::tensor::Tensor *> maxpool_inputs = {&x_tensor, &y_tensor, &yt_tensor};
+  std::vector<lite::tensor::Tensor *> maxpool_outputs = {&out_tensor};
+  // ----------------------------------------
+  kernel::KernelKey maxpool_desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_PoolingGrad};
+  auto maxpool_creator = lite::KernelRegistry::GetInstance()->GetCreator(maxpool_desc);
+  auto kernel = maxpool_creator(maxpool_inputs, maxpool_outputs, reinterpret_cast<OpParameter *>(maxpool), NULL,
+                                maxpool_desc, nullptr);
+
+  kernel->Init();
+
+  auto time_start = mindspore::lite::GetTimeUs();
+  kernel->Run();
+  auto time_end = mindspore::lite::GetTimeUs();
+  printf("single thread running time : %ld ms\n", time_end - time_start);
+
+  std::string output_path = "./test_data/pooling/maxpoolgradfp32_s3_xgrad_3_28_28_3.bin";
+  auto res = lite::CompareRelativeOutput(out_data, output_path);
+
+  EXPECT_EQ(res, 0);
+
+  delete[] x_data;
+  delete[] y_data;
+  delete[] yt_data;
+  // delete[] out_data;
+  // delete conv_param;
+  x_tensor.SetData(nullptr);
+  y_tensor.SetData(nullptr);
+  yt_tensor.SetData(nullptr);
+  out_tensor.SetData(nullptr);
+  delete kernel;
+  MS_LOG(INFO) << "MaxPoolGradStride3Fp32 Filter Grad passed";
+}
+
 }  // namespace mindspore

mindspore/lite/test/ut/src/runtime/kernel/arm/fp32_grad/softmax_crossentropy_fp32_tests.cc

@@ -40,7 +40,7 @@ TEST_F(TestSoftmaxCrossEntropyFp32, SoftmaxCrossEntropyFp32) {
   y_tensor.SetData(input_data);
 
   std::string label_path = "./test_data/operators/sce_fp32_1_l_6.bin";
-  auto ll_labels = reinterpret_cast<int64 *>(mindspore::lite::ReadFile(label_path.c_str(), &input_size));
+  auto ll_labels = reinterpret_cast<int64_t *>(mindspore::lite::ReadFile(label_path.c_str(), &input_size));
   auto labels = new int[6];
   for (int i = 0; i < 6; i++) labels[i] = static_cast<int>(ll_labels[i]);
 
@@ -57,7 +57,7 @@ TEST_F(TestSoftmaxCrossEntropyFp32, SoftmaxCrossEntropyFp32) {
   auto grad = new float[24];
   lite::tensor::Tensor grad_tensor(TypeId::kNumberTypeFloat32, dim_y);
   grad_tensor.SetData(grad);
-  std::vector<lite::tensor::Tensor *> outputs = {&grad_tensor, &loss_tensor};
+  std::vector<lite::tensor::Tensor *> outputs = {&loss_tensor, &grad_tensor};
 
   kernel::KernelKey desc = {kernel::kCPU, TypeId::kNumberTypeFloat32, schema::PrimitiveType_SoftmaxCrossEntropy};
   auto creator = lite::KernelRegistry::GetInstance()->GetCreator(desc);

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/input_x_2_4_5_3.bin

@@ -0,0 +1,2 @@
+V_‚?�K¿ŒÏ§¿øÓà¿…�î>J?/=Å"m?LÒu¿œÂj@!U$?f=ˆ?e¥[?W·Ú¾òÎ ½m©?æ	e¾O™·?}4¿¬žˆ?˜B<–ÚK¿íÕÀše›¾¶ÆÚ½Ýø¿p~|>½?ƒ¿Š7ì½ç§ø½E	:?JÍ¿̬>—	? Ä~?Ï«¾óN1¿?>HV¿ú|ʾ¨œØ=€{‰¿IðU?x›¿©v¾°W>úÉ[¾$˜î? ú•¿]›¿4Bu¿û
 ¾ç­4@Ç×
¿¦+?Ÿ…z>ušB?Åä™=|e >MÚê>¢Ÿ>í¶ß?}Þ0?¹ö§¾©¿ëœ>�û†¾	ù
@<ç€?Âv�?våZ?zäÅ¿@±í¾è.ο•8B?ðîo½
Œ¿ªâÔ¾q"m¿n¯‰?”k>=ì"ê>:©¿³»@ÇÉ<>+R¿
+b±¿6.Œ?“i?¶›v?`j6¿R~]?çJU¼6„s¿GöŒ?M·—¿% ž?Äh”>£È¿½ÇÍ	¿¯ÚG¼½?„¦>³Ó“¾«'6?Æ÷@¿�¥ð¿2¿ƒ/V¾K5è¿TÆ>X]?„[Ý?v_Ø¿¥ü¤¾”j?pý˜?€\l?ã.l=°äÀb©?

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/output_dscale_3.bin

@@ -0,0 +1 @@
+���B���B���B

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_mean_3.bin

@@ -0,0 +1 @@
+å:c2;@e<

mindspore/lite/test/ut/src/runtime/kernel/arm/test_data/bngrad/save_var_3.bin

@@ -0,0 +1 @@
+=}…?ÂM€?Z|?