lite expression

include/api/cfg.h

@@ -1,5 +1,5 @@
 /**
- * Copyright 2021 Huawei Technologies Co., Ltd
+ * Copyright 2022 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.
@@ -25,13 +25,13 @@
 #include "include/api/types.h"
 
 namespace mindspore {
-
+constexpr int iter_th = 1000;
 class MixPrecisionCfg {
  public:
   MixPrecisionCfg() {
     this->dynamic_loss_scale_ = false;
     this->loss_scale_ = 128.0f;
-    this->num_of_not_nan_iter_th_ = 1000;
+    this->num_of_not_nan_iter_th_ = iter_th;
   }
 
   ~MixPrecisionCfg() = default;
@@ -53,6 +53,5 @@ class TrainCfg {
   MixPrecisionCfg mix_precision_cfg_; /**< Mix precision configuration */
   bool accumulate_gradients_ = false;
 };
-
 }  // namespace mindspore
 #endif  // MINDSPORE_INCLUDE_API_CFG_H

include/api/graph.h

@@ -24,23 +24,38 @@
 #include "include/api/types.h"
 
 namespace mindspore {
+class NetData;
+class Net;
+
 class MS_API Graph {
  public:
   class GraphData;
+  enum Type : uint32_t {
+    kExpressionGraph = 0,  ///< graph as expression - can auto grad
+    kExecutableGraph = 1,  ///< graph is loaded as is
+    kUnknownTypeGraph = 0xffffffff
+  };
   Graph();
   explicit Graph(const std::shared_ptr<GraphData> &graph_data);
   explicit Graph(std::shared_ptr<GraphData> &&graph_data);
   explicit Graph(std::nullptr_t);
   ~Graph();
+  explicit Graph(Type executable);
+  explicit Graph(Net *net);
 
   enum ModelType ModelType() const;
   bool operator==(std::nullptr_t) const;
   bool operator!=(std::nullptr_t) const;
+  bool IsExecutable() { return graph_type_ == kExecutableGraph; }
 
  private:
   friend class GraphCell;
   friend class ModelImpl;
+  friend class NetImpl;
+  friend class Model;
   std::shared_ptr<GraphData> graph_data_;
+  std::shared_ptr<NetData> net_data_;
+  Type graph_type_ = kExecutableGraph;
 };
 }  // namespace mindspore
 #endif  // MINDSPORE_INCLUDE_API_GRAPH_H

include/api/model.h

@@ -33,6 +33,9 @@
 namespace mindspore {
 class ModelImpl;
 class Metrics;
+class Net;
+class Node;
+class Expr;
 
 namespace dataset {
 class Dataset;
@@ -109,6 +112,17 @@ class MS_API Model {
   Status Build(GraphCell graph, const std::shared_ptr<Context> &model_context = nullptr,
                const std::shared_ptr<TrainCfg> &train_cfg = nullptr);
 
+  /// \brief Build train model
+  ///
+  /// \param[in] graph A forward network
+  /// \param[in] optimizer An optimizer node
+  /// \param[in] inputs Inputs expression for the trained network (ex: input, label )
+  /// \param[in] model_contex A context used to store options during execution.
+  /// \param[in] train_cfg A config used by training
+  /// \return Status
+
+  Status Build(GraphCell graph, Node *optimizer, std::vector<Expr *> inputs,
+               const std::shared_ptr<Context> &model_context, const std::shared_ptr<TrainCfg> &train_cfg);
   /// \brief Builds a Transfer Learning model where the backbone weights are fixed and the head weights are trainable
   ///
   /// \param[in] backbone The static, non-learnable part of the graph

include/api/net.h

@@ -0,0 +1,141 @@
+/**
+ * Copyright 2022 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_INCLUDE_API_NET_H
+#define MINDSPORE_INCLUDE_API_NET_H
+
+#include <memory>
+#include <vector>
+#include <unordered_set>
+#include <string>
+#include "include/api/types.h"
+#include "include/api/data_type.h"
+#include "include/api/cfg.h"
+
+namespace mindspore {
+/// \brief Register node or sub network
+#define REG(_name) Register(_name, #_name)
+
+class Expr;
+class NodeImpl;
+class NetImpl;
+class NodeSet;
+class Graph;
+class NetData;
+
+class NetBase {
+ public:
+  NetBase() = default;
+  virtual std::vector<Expr *> operator()(const std::vector<Expr *> &inputs) = 0;
+  virtual uint32_t type() = 0;
+};
+
+class Node : public NetBase {
+ public:
+  Node();
+  virtual ~Node();
+  /// \brief Create output expression from node
+
+  /// \param[in] name Name of input (like "labels" etc.)
+  ///
+  /// \return Expression
+  Expr *Create(std::string name);
+  /// \brief Run node on inputs. This operator is used in Net::construct()
+  ///
+  /// \param[in] inputs Inputs expression for the node.
+  /// \return Output node expression vector
+  std::vector<Expr *> operator()(const std::vector<Expr *> &inputs) override;
+  uint32_t type() final;
+
+ private:
+  friend NodeImpl;
+  std::shared_ptr<NodeImpl> impl_ = nullptr;
+};
+
+class Net : public NetBase, public std::enable_shared_from_this<Net> {
+ public:
+  Net();
+  virtual ~Net();
+  explicit Net(std::string name);
+  explicit Net(const Graph &g);
+  /// \brief Define the relation between network inputs and outputs
+  ///
+  /// \param[in] inputs expression vector
+  ///
+  /// \return expression vector
+
+  virtual std::vector<Expr *> construct(const std::vector<Expr *> &inputs);
+  /// \brief Addition operation
+  ///
+  /// \param[in] inputs Two elements to add
+  ///
+  /// \return expression vector (single element)
+
+  /// \brief Execution operator. Connect inputs to outputs via user defined construct
+  ///
+  /// \return expression vector
+
+  std::vector<Expr *> operator()(const std::vector<Expr *> &inputs);
+  void Register(Net *net, std::string &&name);
+  void Register(Node *node, std::string &&name);
+  /// \brief Find the trainable params for the trained network
+  ///
+  /// \return NodeSet for all trainable nodes
+  std::shared_ptr<NodeSet> trainable_params();
+  virtual void Add(NetBase *element);
+  /// \brief Input shape
+  ///
+  /// \param[in] idx input index
+  ///
+  /// \return Specific input shape vector
+  const std::vector<int> InputShape(int idx);
+  /// \brief Output shape
+  ///
+  /// \param[in] idx Output index
+  ///
+  /// \return Specific output shape vector
+  const std::vector<int> OutputShape(int idx);
+  uint32_t type() final;
+
+ private:
+  friend NetImpl;
+  friend NetData;
+  std::shared_ptr<NetImpl> impl_;
+};
+
+class SoftMaxCrossEntropyCfg {
+ public:
+  std::string reduction = "mean"; /**<  Specifies reduction mode. The optional values are "none", "mean", "sum" */
+};
+
+class AdamConfig {
+ public:
+  float learning_rate_ = 1e-3;
+  float beta1_ = 0.9;
+  float beta2_ = 0.999;
+  float eps_ = 1e-08;
+  bool use_nesterov_ = false;
+};
+
+namespace NN {
+Net *NetWithLoss(Net *net, Node *loss);
+Graph *GraphWithLoss(Graph *g, Node *loss);
+Node *Adam(std::shared_ptr<NodeSet> learn, const AdamConfig &cfg);
+Node *SoftmaxCrossEntropy(const SoftMaxCrossEntropyCfg &cfg);
+std::unique_ptr<Node> Input(std::vector<int> dims, DataType data_type = DataType::kNumberTypeFloat32, int fmt = NHWC);
+};  // namespace NN
+}  // namespace mindspore
+#endif  // MINDSPORE_INCLUDE_API_NET_H

mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/infer/conv2d_grad_filter_infer.c

@@ -14,6 +14,7 @@
  * limitations under the License.
  */
 
+#include <stdio.h>
 #include "nnacl/infer/conv2d_grad_filter_infer.h"
 #include "nnacl/infer/infer_register.h"
 
@@ -26,28 +27,33 @@ int Conv2dGradFilterInferShape(const TensorC *const *inputs, size_t inputs_size,
   if (inputs_size < 3 || outputs_size != 1) {
     return NNACL_ERR;
   }
-  if (inputs[0]->format_ != Format_NHWC || inputs[1]->format_ != Format_NHWC) {
+  if (inputs[FIRST_INPUT]->format_ != Format_NHWC || inputs[SECOND_INPUT]->format_ != Format_NHWC) {
     return NNACL_FORMAT_ERROR;
   }
-  SetDataTypeFormat(outputs[0], inputs[0]);
+  SetDataTypeFormat(outputs[FIRST_INPUT], inputs[FIRST_INPUT]);
 
-  if (inputs[2]->shape_size_ < 1 || inputs[2]->data_ == NULL) {
+  if (inputs[THIRD_INPUT]->shape_size_ < DIMENSION_1D || inputs[THIRD_INPUT]->data_ == NULL) {
     return NNACL_ERR;
   }
-  if (inputs[2]->shape_[0] < 0) {
+  if (inputs[THIRD_INPUT]->shape_[kNCHW_N] < 0) {
     return NNACL_ERR;
   }
-  size_t filter_shape_size = (size_t)(inputs[2]->shape_[0]);
-  if (filter_shape_size != 4) {
+  size_t filter_shape_size = (size_t)(inputs[THIRD_INPUT]->shape_[kNCHW_N]);
+  if (filter_shape_size != DIMENSION_4D) {
     return NNACL_ERR;
   }
 
   int filter_shape[MAX_SHAPE_SIZE];
-  const int nchw2nhwc[4] = {0, 2, 3, 1};
-  for (size_t i = 0; i < filter_shape_size; i++) {
-    filter_shape[i] = *((int *)(inputs[2]->data_) + nchw2nhwc[i]);
+  if (inputs[THIRD_INPUT]->format_ == Format_NCHW || inputs[THIRD_INPUT]->format_ == Format_KCHW) {
+    const int nchw2nhwc[] = {kNCHW_N, kNCHW_H, kNCHW_W, kNCHW_C};
+    for (size_t i = 0; i < filter_shape_size; i++) {
+      filter_shape[i] = *((int *)(inputs[THIRD_INPUT]->data_) + nchw2nhwc[i]);
+    }
+  } else if (inputs[THIRD_INPUT]->format_ == Format_NHWC || inputs[THIRD_INPUT]->format_ == Format_KHWC) {
+    memcpy(filter_shape, inputs[THIRD_INPUT]->data_, filter_shape_size * sizeof(int));
+  } else {
+    return NNACL_ERR;
   }
-
   SetShapeArray(outputs[0], filter_shape, filter_shape_size);
   return NNACL_OK;
 }

mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/infer/conv2d_grad_input_infer.c

@@ -37,20 +37,26 @@ int Conv2dGradInputInferShape(const TensorC *const *inputs, size_t inputs_size,
   }
   SetDataTypeFormat(out, in0);
 
-  if (inputs[2]->shape_size_ < 1 || inputs[2]->data_ == NULL) {
+  if (inputs[THIRD_INPUT]->shape_size_ < 1 || inputs[THIRD_INPUT]->data_ == NULL) {
     return NNACL_ERR;
   }
   size_t data_size = (size_t)inputs[2]->shape_[0];
   if (data_size != 4) {
     return NNACL_ERR;
   }
+
   int shape[MAX_SHAPE_SIZE];
-  const int nchw2nhwc[4] = {0, 2, 3, 1};
-  for (size_t i = 0; i < data_size; i++) {
-    shape[i] = *((int *)(inputs[2]->data_) + nchw2nhwc[i]);
+  if (inputs[THIRD_INPUT]->format_ == Format_NCHW || inputs[THIRD_INPUT]->format_ == Format_KCHW) {
+    const int nchw2nhwc[4] = {kNCHW_N, kNCHW_H, kNCHW_W, kNCHW_C};
+    for (size_t i = 0; i < data_size; i++) {
+      shape[i] = *((int *)(inputs[THIRD_INPUT]->data_) + nchw2nhwc[i]);
+    }
+  } else if (inputs[THIRD_INPUT]->format_ == Format_NHWC || inputs[THIRD_INPUT]->format_ == Format_KHWC) {
+    memcpy(shape, inputs[THIRD_INPUT]->data_, data_size * sizeof(int));
+  } else {
+    return NNACL_ERR;
   }
   SetShapeArray(out, shape, data_size);
-
   return NNACL_OK;
 }
 

mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/infer/conv2d_infer.c

@@ -157,7 +157,7 @@ int Conv2dInferShape(const TensorC *const *inputs, size_t inputs_size, TensorC *
   param->output_h_ = out_shape[DIMENSION_1D];
   param->output_w_ = out_shape[DIMENSION_2D];
   param->output_channel_ = out_shape[DIMENSION_3D];
-
+  param->out_format_ = out_tensor->format_;
   return NNACL_OK;
 }
 

mindspore/ccsrc/plugin/device/cpu/kernel/nnacl/op_base.h

@@ -25,6 +25,7 @@
 
 #include "nnacl/intrinsics/ms_simd_instructions.h"
 
+#define C0NUM 0
 #define C1NUM 1
 #define C2NUM 2
 #define C3NUM 3
@@ -33,6 +34,8 @@
 #define C6NUM 6
 #define C7NUM 7
 #define C8NUM 8
+#define C9NUM 9
+#define C10NUM 10
 #define C12NUM 12
 #define C13NUM 13
 #define C16NUM 16

mindspore/core/load_mindir/load_model.cc

@@ -235,7 +235,7 @@ FuncGraphPtr MindIRLoader::LoadMindIR(const std::string &file_name) {
   _fullpath(abs_path_buff, file_name.c_str(), PATH_MAX);
 #else
   if (!realpath(file_name.c_str(), abs_path_buff)) {
-    MS_LOG(ERROR) << "Load MindIR get absolute path failed";
+    MS_LOG(ERROR) << "Load MindIR get absolute path of " << file_name << " failed";
   }
 #endif
   // Read graph

mindspore/lite/src/CMakeLists.txt

@@ -275,9 +275,34 @@ if(MSLITE_MINDDATA_IMPLEMENT STREQUAL "full")
           ${CMAKE_CURRENT_SOURCE_DIR}/train/train_loop.cc
           )
 endif()
+
+
+file(GLOB CXX_API_EXPRESSION
+  ${CMAKE_CURRENT_SOURCE_DIR}/cxx_api/expression/*.cc
+  )
+
+file(GLOB EXPRESSION_OPS
+  ${CMAKE_CURRENT_SOURCE_DIR}/expression/ops/*.cc
+  )
+
+set(EXPRESSION_SRC
+    ${CXX_API_EXPRESSION}
+    ${CMAKE_CURRENT_SOURCE_DIR}/expression/export.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/expression/expr.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/expression/import.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/expression/net.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/expression/node.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/expression/ops.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/expression/ops_utils.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/expression/param.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/expression/sequential.cc
+    ${EXPRESSION_OPS}
+  )
+
 set(TRAIN_SRC
         ${API_TRAIN_SRC}
         ${TRAIN_SRC_WITH_MD}
+        ${EXPRESSION_SRC}
         ${CMAKE_CURRENT_SOURCE_DIR}/common/quant_utils.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/train/train_populate_parameter.cc
         ${CMAKE_CURRENT_SOURCE_DIR}/train/train_session.cc

mindspore/lite/src/cxx_api/expression/net.cc

@@ -0,0 +1,145 @@
+/**
+ * Copyright 2022 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/api/net.h"
+#include "include/api/status.h"
+#include "src/cxx_api/expression/node_impl.h"
+#include "src/cxx_api/expression/net_impl.h"
+#include "src/expression/ops.h"
+#include "src/expression/cfg.h"
+
+namespace mindspore {
+uint32_t Node::type() { return kNodeType; }
+
+std::vector<Expr *> Node::operator()(const std::vector<Expr *> &inputs) {
+  auto in = Expr::convert(inputs);
+  if (impl_ == nullptr) {
+    MS_LOG(ERROR) << "empty implementation";
+    return {};
+  }
+  if (impl_->node() == nullptr) {
+    MS_LOG(ERROR) << "expression node is not attached";
+    return {};
+  }
+  auto out = impl_->node()->construct(in);
+  return Expr::convert(out);
+}
+
+Expr *Node::Create(std::string name) {
+  auto expr = impl_->node()->create(name);
+  return reinterpret_cast<Expr *>(expr);
+}
+
+Node::Node() {
+  auto impl = std::make_shared<NodeImpl>();
+  impl_ = impl;
+  impl_->set_pnode(this);
+}
+
+Node::~Node() {
+  impl_->set_pnode(nullptr);
+  auto node = impl_->node();
+  if (node != nullptr) {
+    impl_->set_node(nullptr);
+    delete node;
+  }
+}
+
+Net::Net(std::string name) {
+  auto impl = std::make_shared<NetImpl>();
+  if (impl == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate network implementation";
+    return;
+  }
+  impl_ = impl;
+  impl_->set_pnet(std::shared_ptr<Net>(this));
+  auto netl = new (std::nothrow) lite::Net(name);
+  if (netl == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate network lite";
+    return;
+  }
+  netl->set_impl(impl);
+  impl_->set_net(netl);
+}
+
+Net::Net() : Net("") {}
+
+Net::Net(const Graph &g) {
+  auto net = NetImpl::GetNet(g);
+  impl_ = net->impl_;
+}
+
+void Net::Add(NetBase *element) { MS_LOG(WARNING) << "Only sequential can add element"; }
+
+uint32_t Net::type() { return kNetType; }
+
+std::vector<Expr *> Net::construct(const std::vector<Expr *> &inputs) {
+  auto in = Expr::convert(inputs);
+  auto out = impl_->net()->construct(in);
+  return Expr::convert(out);
+}
+
+std::vector<Expr *> Net::operator()(const std::vector<Expr *> &inputs) {
+  auto in = Expr::convert(inputs);
+  auto x = construct(inputs);
+  impl_->net()->input_ = in;
+  auto out = Expr::convert(x);
+  impl_->net()->output_ = out;
+  impl_->net()->real_output_ = out;
+  return x;
+}
+void Net::Register(Net *net, std::string &&name) {
+  if (net != nullptr) {
+    auto net_lite = net->impl_->net();
+    impl_->net()->Register(net_lite, std::move(name));
+  }
+}
+
+void Net::Register(Node *node, std::string &&name) {
+  if (node != nullptr) {
+    auto impl = NodeImpl::GetImpl(node);
+    if (impl == nullptr) {
+      MS_LOG(ERROR) << "missing implementation";
+      return;
+    }
+    auto node_lite = impl->node();
+    impl_->net()->Register(node_lite, std::move(name));
+  }
+}
+
+std::shared_ptr<NodeSet> Net::trainable_params() {
+  auto node_set = std::make_shared<NodeSet>();
+  if (node_set == nullptr) {
+    MS_LOG(ERROR) << "new NodeSet failed.";
+    return nullptr;
+  }
+  node_set->set_ = impl_->net()->trainable_params();
+  return node_set;
+}
+
+const std::vector<int> Net::InputShape(int idx) { return impl_->InputShape(idx); }
+const std::vector<int> Net::OutputShape(int idx) { return impl_->OutputShape(idx); }
+
+Net::~Net() {
+  if (impl_ != nullptr) {
+    if ((impl_->pnet() == nullptr) || (impl_->pnet() == this)) {
+      impl_->set_pnet(nullptr);
+      impl_->set_net(nullptr);
+      impl_.reset();
+    }
+  }
+}
+}  // namespace mindspore

mindspore/lite/src/cxx_api/expression/net_impl.cc

@@ -0,0 +1,222 @@
+/**
+ * Copyright 2022 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/cxx_api/expression/net_impl.h"
+#include <vector>
+#include <utility>
+#include "include/api/serialization.h"
+#include "src/expression/import.h"
+#include "src/expression/ops.h"
+#include "src/cxx_api/model/model_impl.h"
+
+namespace {
+constexpr size_t kFlatbuffersBuilderInitSize = 1024;
+};
+
+namespace mindspore {
+Sequential::Sequential() {}
+
+lite::Node *Sequential::GetNode(NetBase *element) {
+  lite::Node *lite_node = nullptr;
+  switch (element->type()) {
+    case kNodeType: {
+      Node *node = reinterpret_cast<Node *>(element);
+      auto impl = NodeImpl::GetImpl(node);
+      if (impl == nullptr) {
+        MS_LOG(ERROR) << "cannot find node implement";
+        return nullptr;
+      }
+      lite_node = impl->node();
+      break;
+    }
+    case kNetType: {
+      auto net = reinterpret_cast<Net *>(element);
+      auto impl = NetImpl::GetImpl(net);
+      if (impl == nullptr) {
+        MS_LOG(ERROR) << "cannot find node implement";
+        return nullptr;
+      }
+      lite_node = impl->net();
+      break;
+    }
+  }
+  return lite_node;
+}
+
+void Sequential::Add(NetBase *element) {
+  lite::Node *node = GetNode(element);
+  auto impl = NetImpl::GetImpl(this);
+  if (impl == nullptr) {
+    MS_LOG(ERROR) << "No implementation";
+    return;
+  }
+  impl->net()->Add(node);
+}
+
+NetWithLoss::NetWithLoss(Net *net, Node *loss) : net_(net), loss_fn_(loss) {
+  REG(net_);
+  Register(loss_fn_, "_loss_fn");
+}
+
+std::vector<Expr *> NetWithLoss::construct(const std::vector<Expr *> &inputs) {
+  if (inputs.size() != C2NUM) {
+    MS_LOG(ERROR) << "need 2 inputs for loss";
+    return {};
+  }
+  auto input = inputs[FIRST_INPUT];
+  auto label = inputs[SECOND_INPUT];
+  auto x = (*net_)({input});
+  x = (*loss_fn_)({x[FIRST_INPUT], label});
+  return x;
+}
+
+NetImpl::~NetImpl() {}
+
+NetImpl::NetImpl(std::shared_ptr<Net> p) { pnet_ = p; }
+
+NetImpl::NetImpl(Graph *g) { pnet_ = g->net_data_->net(); }
+
+std::vector<lite::EXPR *> MS_API NetImpl::construct(const std::vector<lite::EXPR *> &inputs) {
+  auto in = Expr::convert(inputs);
+  auto out = pnet_->construct(in);
+  return Expr::convert(out);
+}
+
+Net *MS_API NetImpl::Connect(std::shared_ptr<Net> net, lite::Net *lnet) {
+  auto impl = GetImpl(net.get());
+  if (impl == nullptr) {
+    MS_LOG(ERROR) << "missing implementation";
+    return nullptr;
+  }
+  impl->set_pnet(net);
+  lnet->set_impl(impl);
+  impl->set_net(lnet);
+  return net.get();
+}
+
+Status NetImpl::Import(const char *model_buf, Graph *graph) {
+  auto mg = schema::GetMetaGraph(model_buf);
+  auto net = new (std::nothrow) Net();
+  if (net == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate network";
+    return kLiteMemoryFailed;
+  }
+  lite::Import import;
+  auto lite_net = import.ImportMs(mg);
+  if (lite_net == nullptr) {
+    MS_LOG(ERROR) << "failed to import net";
+    return kLiteMemoryFailed;
+  }
+  lite_net->SetRealOutput();
+  Connect(net->shared_from_this(), lite_net);
+  *graph = Graph(net);
+  return kSuccess;
+}
+
+Status NetImpl::TrainNet(Node *optimizer, const std::vector<Expr *> &inputs) {
+  auto impl = NodeImpl::GetImpl(optimizer);
+  if (impl == nullptr) {
+    MS_LOG(ERROR) << "missing implementation ";
+    return kLiteNullptr;
+  }
+  auto opt = impl->node();
+  auto in = Expr::convert(inputs);
+  auto ret_net = net()->TrainNet(opt, in);
+  if (ret_net == nullptr) {
+    MS_LOG(ERROR) << "failed to train network";
+    return kLiteNullptr;
+  }
+  return kSuccess;
+}
+
+std::unique_ptr<Graph> NetImpl::MakeMs() {
+  auto mgraph = std::make_unique<Graph>(Graph::Type::kExecutableGraph);
+  if (mgraph == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate graph";
+    return nullptr;
+  }
+  auto trained_graph = net()->MakeMs();
+  if (trained_graph == nullptr) {
+    MS_LOG(ERROR) << "cannot create flat buffer";
+    return nullptr;
+  }
+  flatbuffers::FlatBufferBuilder builder(kFlatbuffersBuilderInitSize);
+  auto offset = schema::MetaGraph::Pack(builder, trained_graph.get());
+  builder.Finish(offset);
+  schema::FinishMetaGraphBuffer(builder, offset);
+  auto buffer = builder.GetBufferPointer();
+  size_t size = builder.GetSize();
+  auto status = Serialization::Load(buffer, size, mindspore::kMindIR, mgraph.get());
+  if (status != kSuccess) {
+    MS_LOG(ERROR) << "failed to load flatbuffer to graph";
+    return nullptr;
+  }
+  return mgraph;
+}
+
+const std::vector<int> NetImpl::InputShape(int idx) { return net_->InputShape(idx); }
+
+const std::vector<int> NetImpl::OutputShape(int idx) { return net_->OutputShape(idx); }
+
+void NetImpl::ReplaceNet(Graph *g, std::shared_ptr<Net> n) { g->net_data_->net().swap(n); }
+
+ExpressionLoader expression_registrator = CreateExpressionLoader(NetImpl::Import);
+
+namespace NN {
+Net *Sequential() {
+  auto net = new (std::nothrow) mindspore::Sequential();
+  if (net == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate ";
+    return nullptr;
+  }
+  auto netl = lite::NN::Sequential();
+  return NetImpl::Connect(net->shared_from_this(), netl);
+}
+
+Net *NetWithLoss(Net *net, Node *loss) {
+  auto loss_net = new (std::nothrow) mindspore::NetWithLoss(net, loss);
+  if (net == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate loss net";
+    return nullptr;
+  }
+  return loss_net;
+}
+
+Graph *GraphWithLoss(Graph *graph, Node *loss) {
+  auto net = NetImpl::GetNet(*graph);
+  if (net == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate network";
+    return nullptr;
+  }
+  auto loss_net = NetWithLoss(net.get(), loss);
+  if (loss_net == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate network";
+    return nullptr;
+  }
+  NetImpl::ReplaceNet(graph, loss_net->shared_from_this());
+  return graph;
+}
+
+Net *NetWithLoss(Graph *g, Node *loss) {
+  auto net = new (std::nothrow) Net(*g);
+  if (net == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate net";
+    return nullptr;
+  }
+  return NetWithLoss(net, loss);
+}
+}  // namespace NN
+}  // namespace mindspore

mindspore/lite/src/cxx_api/expression/net_impl.h

@@ -0,0 +1,95 @@
+/**
+ * Copyright 2022 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_CXX_API_EXPRESSION_NET_IMPL_H_
+#define MINDSPORE_LITE_SRC_CXX_API_EXPRESSION_NET_IMPL_H_
+
+#include <algorithm>
+#include <set>
+#include <memory>
+#include <vector>
+#include <utility>
+#include "include/api/cfg.h"
+#include "include/api/data_type.h"
+#include "include/api/graph.h"
+#include "include/api/status.h"
+#include "include/api/net.h"
+#include "src/cxx_api/expression/node_impl.h"
+#include "src/cxx_api/graph/net_data.h"
+#include "src/expression/net.h"
+#include "src/expression/ops.h"
+
+namespace mindspore {
+constexpr uint32_t kNodeType = 1;
+constexpr uint32_t kNetType = 2;
+class Sequential : public Net {
+ public:
+  Sequential();
+  void Add(NetBase *n) override;
+
+ private:
+  std::vector<NetBase *> ops_;
+  lite::Node *GetNode(NetBase *element);
+};
+
+class NetWithLoss : public Net {
+ public:
+  NetWithLoss(Net *net, Node *loss);
+  std::vector<Expr *> construct(const std::vector<Expr *> &inputs) override;
+
+ private:
+  Net *net_{nullptr};
+  Node *loss_fn_{nullptr};
+};
+
+class MS_API NetImpl {
+ public:
+  virtual ~NetImpl();
+  explicit NetImpl(std::shared_ptr<Net> p);
+  explicit NetImpl(Graph *g);
+  NetImpl() = default;
+  void set_net(lite::Net *net) {
+    if (net_ != nullptr) {
+      net_->set_impl(nullptr);
+      delete net_;
+    }
+    net_ = net;
+  }
+  void erase_net() { net_ = nullptr; }
+  void set_pnet(std::shared_ptr<Net> net) { pnet_ = net; }
+  Net *pnet() { return pnet_.get(); }
+  lite::Net *net() { return net_; }
+
+  std::vector<lite::EXPR *> construct(const std::vector<lite::EXPR *> &inputs);
+  static std::shared_ptr<mindspore::NetImpl> &GetImpl(Net *net) { return net->impl_; }
+  static Net *Connect(std::shared_ptr<Net> net, lite::Net *lnet);
+  static std::shared_ptr<Net> &GetNet(const Graph &g) { return g.net_data_->net(); }
+  static void SetNet(Graph *g, std::shared_ptr<Net> n) { g->net_data_->set_net(n); }
+  static void ReplaceNet(Graph *g, std::shared_ptr<Net> n);
+  static Status Import(const char *model_buf, Graph *graph);
+  Status TrainNet(Node *optimizer, const std::vector<Expr *> &inputs);
+  const std::vector<int> InputShape(int idx);
+  const std::vector<int> OutputShape(int idx);
+  std::unique_ptr<Graph> MakeMs();
+  void Release() { pnet_.reset(); }
+
+ private:
+  std::shared_ptr<Net> pnet_;
+  lite::Net *net_ = nullptr;
+};
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_CXX_API_EXPRESSION_NET_IMPL_H_

mindspore/lite/src/cxx_api/expression/node_impl.cc

@@ -0,0 +1,50 @@
+/**
+ * Copyright 2022 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/cxx_api/expression/node_impl.h"
+#include <vector>
+#include "include/api/net.h"
+#include "src/expression/ops.h"
+
+namespace mindspore {
+Node *NodeImpl::Connect(lite::Node *lnode) {
+  auto node = std::make_unique<Node>();
+  if (node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate node";
+    return nullptr;
+  }
+  if (lnode == nullptr) {
+    MS_LOG(ERROR) << "lite node is null";
+    return nullptr;
+  }
+  auto pnode = node.release();
+  auto impl = GetImpl(pnode);
+  if (impl == nullptr) {
+    MS_LOG(ERROR) << "missing implementation";
+    return nullptr;
+  }
+  impl->set_node(lnode);
+  lnode->set_impl(impl);
+  return pnode;
+}
+namespace NN {
+std::unique_ptr<Node> Input(std::vector<int> dims, DataType data_type, int fmt) {
+  auto type = static_cast<TypeId>(data_type);
+  auto lite_node = lite::NN::Input(dims, type, fmt);
+  return std::unique_ptr<Node>(NodeImpl::Connect(lite_node));
+}
+}  // namespace NN
+}  // namespace mindspore

mindspore/lite/src/cxx_api/expression/node_impl.h

@@ -0,0 +1,71 @@
+/**
+ * Copyright 2022 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_CXX_API_EXPRESSION_NODE_IMPL_H_
+#define MINDSPORE_LITE_SRC_CXX_API_EXPRESSION_NODE_IMPL_H_
+
+#include <algorithm>
+#include <set>
+#include <memory>
+#include <vector>
+#include "include/api/net.h"
+#include "include/api/cfg.h"
+#include "include/api/data_type.h"
+#include "src/expression/node.h"
+#include "src/expression/expr.h"
+
+namespace mindspore {
+using lite::EXPR;
+class NodeSet {
+ public:
+  std::set<lite::Node *> set_;
+};
+
+class Expr : public EXPR {
+ public:
+  static std::vector<EXPR *> convert(const std::vector<Expr *> &input) {
+    std::vector<EXPR *> vec(input.size());
+    (void)std::transform(input.begin(), input.end(), vec.begin(), [](Expr *e) { return reinterpret_cast<EXPR *>(e); });
+    return vec;
+  }
+  static std::vector<Expr *> convert(const std::vector<EXPR *> &input) {
+    std::vector<Expr *> vec(input.size());
+    (void)std::transform(input.begin(), input.end(), vec.begin(), [](EXPR *e) { return reinterpret_cast<Expr *>(e); });
+    return vec;
+  }
+};
+
+class MS_API NodeImpl {
+ public:
+  std::vector<Expr *> operator()(const std::vector<Expr *> &inputs) {
+    auto in = Expr::convert(inputs);
+    auto out = (*node_)(in);
+    return Expr::convert(out);
+  }
+  lite::Node *node() { return node_; }
+  void set_node(lite::Node *node) { node_ = node; }
+  void set_pnode(Node *node) { pnode_ = node; }
+  Node *pnode() { return pnode_; }
+  static Node *Connect(lite::Node *lnode);
+  static std::shared_ptr<NodeImpl> &GetImpl(Node *node) { return node->impl_; }
+
+ private:
+  Node *pnode_{nullptr};
+  lite::Node *node_{nullptr};
+};
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_CXX_API_EXPRESSION_NODE_IMPL_H_

mindspore/lite/src/cxx_api/graph/graph.cc

@@ -16,7 +16,9 @@
 
 #include "include/api/graph.h"
 #include "include/api/cell.h"
+#include "include/api/net.h"
 #include "src/cxx_api/graph/graph_data.h"
+#include "src/cxx_api/graph/net_data.h"
 
 namespace mindspore {
 Graph::Graph() : graph_data_(nullptr) {}
@@ -25,8 +27,15 @@ Graph::Graph(const std::shared_ptr<GraphData> &graph_data) : graph_data_(graph_d
 
 Graph::Graph(std::shared_ptr<GraphData> &&graph_data) : graph_data_(graph_data) {}
 
+Graph::Graph(Graph::Type type) : graph_type_(type) {}
+
 Graph::~Graph() {}
 
+Graph::Graph(Net *net) : graph_type_(kExpressionGraph) {
+  auto shared = std::make_shared<NetData>(net->shared_from_this());
+  net_data_ = shared;
+}
+
 Graph::Graph(std::nullptr_t) : graph_data_(nullptr) {}
 
 bool Graph::operator==(std::nullptr_t) const { return graph_data_ == nullptr; }

mindspore/lite/src/cxx_api/graph/net_data.cc

@@ -0,0 +1,21 @@
+/**
+ * Copyright 2022 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/cxx_api/graph/net_data.h"
+#include "src/cxx_api/expression/net_impl.h"
+
+namespace mindspore {
+NetData::~NetData() { net_->impl_->Release(); }
+}  // namespace mindspore

mindspore/lite/src/cxx_api/graph/net_data.h

@@ -0,0 +1,35 @@
+/**
+ * Copyright 2022 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_CXX_API_GRAPH_NET_DATA_H_
+#define MINDSPORE_LITE_SRC_CXX_API_GRAPH_NET_DATA_H_
+
+#include <memory>
+#include "include/api/net.h"
+
+namespace mindspore {
+class NetData {
+ public:
+  explicit NetData(const std::shared_ptr<Net> &net) : net_(net) {}
+  virtual ~NetData();
+  void set_net(std::shared_ptr<Net> net) { net_ = net; }
+  std::shared_ptr<Net> &net() { return net_; }
+
+ private:
+  std::shared_ptr<Net> net_;
+};
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_CXX_API_GRAPH_NET_DATA_H_

mindspore/lite/src/cxx_api/model/model.cc

@@ -22,11 +22,15 @@
 #ifdef ENABLE_LITE_ACL
 #include "acl/acl_base.h"
 #endif
+#include "flatbuffers/flatbuffers.h"
 #include "include/api/callback/callback.h"
 #include "include/api/context.h"
 #include "include/api/dual_abi_helper.h"
 #include "include/api/types.h"
+#include "include/api/serialization.h"
+#include "include/api/graph.h"
 #include "src/common/log_adapter.h"
+#include "src/cxx_api/expression/net_impl.h"
 #include "src/cxx_api/callback/callback_adapter.h"
 #include "src/cxx_api/callback/callback_impl.h"
 #include "src/cxx_api/model/model_impl.h"

mindspore/lite/src/cxx_api/model/model_impl.cc

@@ -56,6 +56,14 @@ CreateTrainSessionProto *CreateTrainSessionCallbackHolder(CreateTrainSessionProt
   return proto_;
 }
 
+ExpressionLoader CreateExpressionLoader(ExpressionLoader loader) {
+  static ExpressionLoader loader_ = nullptr;
+  if (loader != nullptr) {
+    loader_ = loader;
+  }
+  return loader_;
+}
+
 Status ModelImpl::Build(const void *model_data, size_t data_size, ModelType model_type,
                         const std::shared_ptr<Context> &ms_context) {
   if (model_data == nullptr) {

mindspore/lite/src/cxx_api/model/model_impl.h

@@ -49,6 +49,9 @@ typedef std::shared_ptr<lite::LiteSession>(CreateTrainSessionProto)(std::shared_
                                                                     lite::InnerContext *context);
 CreateTrainSessionProto *CreateTrainSessionCallbackHolder(CreateTrainSessionProto *proto = nullptr);
 
+using ExpressionLoader = std::function<Status(const char *, Graph *)>;
+ExpressionLoader CreateExpressionLoader(ExpressionLoader loader = nullptr);
+
 namespace session {
 class Metrics;
 class TrainLoopCallBack;
@@ -90,6 +93,7 @@ class ModelImpl {
 
   static bool CheckModelSupport(const std::string &device_type, ModelType model_type);
   bool IsTrainModel();
+  std::unique_ptr<Graph> BuildTrain(Node *optimizer, std::vector<Expr *> inputs);
   Status SetupVirtualBatch(int virtual_batch_multiplier, float lr, float momentum);
   Status SetLearningRate(float learning_rate);
   float GetLearningRate();

mindspore/lite/src/cxx_api/serialization.cc

@@ -20,6 +20,7 @@
 #include "include/api/graph.h"
 #include "include/api/types.h"
 #include "include/model.h"
+#include "src/cxx_api/expression/net_impl.h"
 #include "src/cxx_api/graph/graph_data.h"
 #include "src/cxx_api/model/model_impl.h"
 #include "src/cxx_api/converters.h"
@@ -27,6 +28,8 @@
 #include "src/lite_session.h"
 
 namespace mindspore {
+std::function<int(void *)> ExpressionCallback;
+
 Key::Key(const char *dec_key, size_t key_len) {
   len = 0;
   if (key_len >= max_key_len) {
@@ -115,19 +118,31 @@ Status Serialization::Load(const std::vector<char> &file, ModelType model_type,
     MS_LOG(ERROR) << "Read model file failed";
     return kLiteNullptr;
   }
-  auto model =
-    std::shared_ptr<lite::Model>(lite::ImportFromBuffer(static_cast<const char *>(model_buf), model_size, true));
-  if (model == nullptr) {
-    MS_LOG(ERROR) << "New model failed.";
-    return kLiteNullptr;
+  if (graph->IsExecutable()) {
+    auto model =
+      std::shared_ptr<lite::Model>(lite::ImportFromBuffer(static_cast<const char *>(model_buf), model_size, true));
+    if (model == nullptr) {
+      MS_LOG(ERROR) << "New model failed.";
+      return kLiteNullptr;
+    }
+    auto graph_data = std::shared_ptr<Graph::GraphData>(new (std::nothrow) Graph::GraphData(model));
+    if (graph_data == nullptr) {
+      MS_LOG(ERROR) << "New graph data failed.";
+      return kLiteMemoryFailed;
+    }
+    *graph = Graph(graph_data);
+    return kSuccess;
+  } else {
+    auto loader = CreateExpressionLoader();
+    if (loader == nullptr) {
+      MS_LOG(ERROR) << "Unsupported Feature.";
+      delete[] model_buf;
+      return kLiteError;
+    }
+    loader(model_buf, graph);
+    delete[] model_buf;
+    return kSuccess;
   }
-  auto graph_data = std::shared_ptr<Graph::GraphData>(new (std::nothrow) Graph::GraphData(model));
-  if (graph_data == nullptr) {
-    MS_LOG(ERROR) << "New graph data failed.";
-    return kLiteMemoryFailed;
-  }
-  *graph = Graph(graph_data);
-  return kSuccess;
 }
 
 Status Serialization::Load(const std::vector<std::vector<char>> &files, ModelType model_type,

mindspore/lite/src/cxx_api/train/model.cc

@@ -17,6 +17,7 @@
 #include "include/api/model.h"
 #include "include/api/types.h"
 #include "include/api/context.h"
+#include "include/api/net.h"
 #include "include/api/callback/callback.h"
 #include "include/api/dual_abi_helper.h"
 #include "src/cxx_api/model/model_impl.h"
@@ -106,11 +107,39 @@ Status Model::Evaluate(std::shared_ptr<dataset::Dataset> ds, std::vector<TrainCa
   return (ret == mindspore::lite::RET_OK) ? kSuccess : kLiteError;
 }
 
+Status Model::Build(GraphCell lossGraphCell, Node *optimizer, std::vector<Expr *> inputs,
+                    const std::shared_ptr<Context> &model_context, const std::shared_ptr<TrainCfg> &train_cfg) {
+  std::stringstream err_msg;
+  if (impl_ == nullptr) {
+    impl_ = std::make_shared<ModelImpl>();
+    if (impl_ == nullptr) {
+      MS_LOG(ERROR) << "Model implement is null.";
+      return kLiteFileError;
+    }
+  }
+  auto lossGraph = lossGraphCell.GetGraph();
+  if (lossGraph == nullptr) {
+    err_msg << "Invalid null graph";
+    MS_LOG(ERROR) << err_msg.str();
+    return Status(kLiteNullptr, err_msg.str());
+  }
+  impl_->SetContext(model_context);
+  impl_->SetConfig(train_cfg);
+  impl_->SetGraph(lossGraph);
+  auto graph = impl_->BuildTrain(optimizer, inputs);
+  auto status = Build(GraphCell(*graph), model_context, train_cfg);
+  if (status != mindspore::kSuccess) {
+    MS_LOG(ERROR) << "Error " << status << " during model build";
+    return status;
+  }
+  return kSuccess;  // status
+}
+
 Status Model::BuildTransferLearning(GraphCell backbone, GraphCell head, const std::shared_ptr<Context> &context,
                                     const std::shared_ptr<TrainCfg> &train_cfg) {
   std::stringstream err_msg;
   if (impl_ == nullptr) {
-    impl_ = std::shared_ptr<ModelImpl>(new (std::nothrow) ModelImpl());
+    impl_ = std::make_shared<ModelImpl>();
     if (impl_ == nullptr) {
       MS_LOG(ERROR) << "Model implement is null.";
       return kLiteFileError;
@@ -137,5 +166,4 @@ Status Model::BuildTransferLearning(GraphCell backbone, GraphCell head, const st
   }
   return kSuccess;
 }
-
 }  // namespace mindspore

mindspore/lite/src/cxx_api/train/model_impl.cc

@@ -18,15 +18,18 @@
 #include <memory>
 #include <unordered_map>
 #include <algorithm>
+#include "src/cxx_api/expression/node_impl.h"
 #include "include/api/types.h"
 #include "include/api/context.h"
 #include "include/api/dual_abi_helper.h"
+#include "include/api/serialization.h"
 #include "include/lite_session.h"
 #include "include/context.h"
 #include "include/api/callback/callback.h"
 #include "include/api/metrics/metrics.h"
 #include "src/lite_model.h"
 #include "src/runtime/inner_allocator.h"
+#include "src/cxx_api/expression/net_impl.h"
 #include "src/cxx_api/converters.h"
 #include "src/cxx_api/graph/graph_data.h"
 #include "src/cxx_api/tensor/tensor_impl.h"
@@ -79,6 +82,32 @@ Status ModelImpl::BuildTransferLearning(const std::shared_ptr<Graph> &backbone,
   return kLiteError;
 }
 
+std::unique_ptr<Graph> ModelImpl::BuildTrain(Node *optimizer, std::vector<Expr *> inputs) {
+  auto opt_impl = NodeImpl::GetImpl(optimizer);
+  if (opt_impl == nullptr) {
+    MS_LOG(ERROR) << "missing optimizer node implementation";
+    return nullptr;
+  }
+  auto opt = opt_impl->node();
+  auto in = Expr::convert(inputs);
+  auto net_impl = NetImpl::GetImpl(graph_->net_data_->net().get());
+  if (net_impl == nullptr) {
+    MS_LOG(ERROR) << "missing net implementation";
+    return nullptr;
+  }
+  auto trained_net = net_impl->net()->TrainNet(opt, in);
+  if (trained_net == nullptr) {
+    MS_LOG(ERROR) << "failed to train network";
+    return nullptr;
+  }
+  auto mgraph = net_impl->MakeMs();
+  if (mgraph == nullptr) {
+    MS_LOG(ERROR) << "failed to create graph";
+    return nullptr;
+  }
+  return mgraph;
+}
+
 Status ModelImpl::PrepareMetrics(Model *model, std::vector<session::Metrics *> *out_ms,
                                  std::vector<session::Metrics *> *adapter_ms) {
   if (out_ms == nullptr || adapter_ms == nullptr) {

mindspore/lite/src/expression/cfg.h

@@ -0,0 +1,68 @@
+/**
+ * Copyright 2022 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_EXPRESSION_CFG_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_CFG_H_
+
+#include <vector>
+#include <string>
+
+namespace mindspore {
+namespace lite {
+class ConvConfig {
+ public:
+  ConvConfig() = default;
+  int in_channel_ = 3;                          /**< The channel number of the input of the Conv2d layer */
+  int out_channel_ = 3;                         /**< The channel number of the output tensor of the Conv2d layer */
+  std::vector<int64_t> kernel_size_ = {3, 3};   /**< Specifies the height and width of the 2D convolution kernel. */
+  std::vector<int64_t> stride_ = {1, 1};        /**< The movement stride of the 2D convolution kernel */
+  std::vector<int64_t> padding_ = {0, 0, 0, 0}; /**< The top, bottom, left, and right padding input */
+  std::vector<int64_t> dilation_ = {1, 1};      /**< diletion height and width*/
+  int group_ = 1;                               // < Splits filter into groups, `in_channels` and `out_channels` must be
+                   // divisible by `group`. If the group is equal to `in_channels` and `out_channels`,
+                   // this 2D convolution layer also can be called 2D depthwise convolution layer */
+  bool has_bias = false; /** < Whether the Conv2d layer has a bias parameter */
+  std::string weight_init_ =
+    "normal"; /**< Initialization method of weight parameter ("normal","uniform", "ones", "zeros") */
+  std::string pad_mode_ = "same"; /**<  Specifies padding mode. The optional values are "same", "valid", "pad" */
+
+ private:
+  std::string bias_init_ = "zeros";
+  std::string data_format;
+};
+
+class DenseConfig {
+ public:
+  int in_channels_;       /**< The number of channels in the input space */
+  int out_channels_;      /**< The number of channels in the output space */
+  bool has_bias_ = false; /** Specifies whether the layer uses a bias vector **/
+ private:
+  std::string weight_init_ = "normal";
+  std::string bias_init_ = "zeros";
+  std::string activation_ = "none";
+};
+
+class PoolingConfig {
+ public:
+  PoolingConfig() = default;
+  std::vector<int64_t> kernel_size_ = {1, 1}; /**< Specifies the height and width of the 2D kernel. */
+  std::vector<int64_t> stride_ = {1, 1};      /**< The movement stride of the 2D kernel */
+  std::string pad_mode_ = "same";             /**<  Specifies padding mode. The optional values are "same", "valid" */
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_CFG_H_

mindspore/lite/src/expression/export.cc

@@ -0,0 +1,76 @@
+/**
+ * Copyright 2022 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 <utility>
+#include "src/expression/export.h"
+#include "src/expression/ops.h"
+#include "src/common/utils.h"
+#include "nnacl/conv_parameter.h"
+#include "include/errorcode.h"
+
+namespace mindspore {
+namespace lite {
+constexpr static int kFmkVal = 3;
+
+int ExportSession::Init(const std::string model_name, std::string version) {
+  meta_graph_ = new (std::nothrow) mindspore::schema::MetaGraphT();
+  if (meta_graph_ == nullptr) {
+    MS_LOG(ERROR) << "cannot allocate meta_graph";
+    return RET_ERROR;
+  }
+  meta_graph_->fmkType = kFmkVal;
+  meta_graph_->name = model_name;
+  meta_graph_->version = version;
+  return RET_OK;
+}
+
+bool ExportSession::IsToDependOnly(EXPR *expr) {
+  auto itr = outmap_.find(expr);
+  if (itr != outmap_.end() && !itr->second.empty()) {
+    for (auto expr : itr->second) {
+      auto node = expr->node();
+      if (node->primitive() != schema::PrimitiveType_Depend) return false;
+    }
+    return true;
+  }
+  return false;
+}
+
+int ExportSession::SetInputOutput(const std::vector<EXPR *> &inputs, const std::vector<EXPR *> &outputs) {
+  for (auto &in : inputs) {
+    auto id = GetOutput(in);
+    meta_graph_->inputIndex.push_back(id);
+  }
+  for (auto &out : outputs) {
+    auto id = GetOutput(out);
+    meta_graph_->outputIndex.push_back(id);
+  }
+  auto sub_graph = std::make_unique<mindspore::schema::SubGraphT>();
+  if (sub_graph == nullptr) {
+    MS_LOG(ERROR) << "cannot allocate SubGraphT";
+    return RET_ERROR;
+  }
+  auto model_name = meta_graph_->name;
+  sub_graph->name = model_name + "_subgraph";
+  sub_graph->inputIndices = meta_graph_->inputIndex;
+  sub_graph->outputIndices = meta_graph_->outputIndex;
+  for (size_t i = 0; i < meta_graph_->nodes.size(); i++) sub_graph->nodeIndices.push_back(i);
+  for (size_t i = 0; i < meta_graph_->allTensors.size(); i++) sub_graph->tensorIndices.push_back(i);
+  meta_graph_->subGraph.emplace_back(std::move(sub_graph));
+  return RET_OK;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/export.h

@@ -0,0 +1,52 @@
+/**
+ * Copyright 2022 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_EXPRESSION_EXPORT_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_EXPORT_H_
+
+#include <vector>
+#include <memory>
+#include <unordered_map>
+#include <string>
+#include <list>
+#include <map>
+#include <iostream>
+#include "src/expression/expr.h"
+
+namespace mindspore {
+namespace schema {
+struct MetaGraphT;
+}
+namespace lite {
+class ExportSession {
+ public:
+  explicit ExportSession(std::map<EXPR *, std::list<EXPR *>> &outmap) : outmap_(outmap) {}
+  int Init(const std::string model_name, std::string version);
+  void UpdateOutput(EXPR *expr, int id) { output_tensors_[expr] = id; }
+  int GetOutput(EXPR *expr) { return output_tensors_.at(expr); }
+  schema::MetaGraphT *&meta_graph() { return meta_graph_; }
+  int SetInputOutput(const std::vector<EXPR *> &inputs, const std::vector<EXPR *> &outputs);
+  bool IsToDependOnly(EXPR *expr);
+
+ private:
+  schema::MetaGraphT *meta_graph_{nullptr};
+  std::unordered_map<EXPR *, int> output_tensors_;  // output tensors per EXPR
+  std::map<EXPR *, std::list<EXPR *>> &outmap_;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_EXPORT_H_

mindspore/lite/src/expression/expr.cc

@@ -0,0 +1,98 @@
+/**
+ * Copyright 2022 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 <string>
+#include <iostream>
+#include "src/expression/expr.h"
+#include "src/expression/node.h"
+
+namespace mindspore {
+namespace lite {
+std::string EXPR::name() { return node_->name(); }
+void EXPR::Travers(std::function<bool(EXPR *e, EXPR *itr)> cb) {
+  if (!visited) {
+    visited = true;
+    for (auto &itr : params_) {
+      if (cb(this, itr)) {
+        itr->Travers(cb);
+      }
+    }
+  }
+}
+
+void EXPR::Replace(EXPR **old, EXPR **n, std::vector<Node *> *to_delete) {
+  if (!visited) {
+    visited = true;
+    for (auto &itr : params_)
+      if (itr == *old) {
+        to_delete->push_back(itr->node());
+        itr = *n;
+      }
+    for (auto &itr : params_) itr->Replace(old, n, to_delete);
+  }
+}
+
+void EXPR::Replace(const std::vector<EXPR *> &vec, std::vector<EXPR *> *old, std::vector<EXPR *> *n) {
+  std::vector<Node *> to_delete;
+  for (auto &e : vec) {
+    for (std::size_t i = 0; i < old->size(); i++) {
+      e->Replace(&old->at(i), &n->at(i), &to_delete);
+    }
+  }
+  for (auto &itr : to_delete) delete itr;
+  for (auto e : vec) e->Clear();
+}
+
+void EXPR::Clear() {
+  EXPR *item = this;
+  if (visited == false) return;
+  visited = false;
+  while (item->params_.size() == 1) {
+    item = item->params_.front();
+    if (item->visited == false) return;
+    item->visited = false;
+  }
+  for (auto &itr : item->params_) itr->Clear();
+}
+
+void EXPR::Clear(std::vector<EXPR *> vec) {
+  for (auto e : vec) e->Clear();
+}
+
+void EXPR::CreateOutputMap(std::vector<EXPR *> vec, std::map<EXPR *, std::list<EXPR *>> *outmap) {
+  for (auto e : vec) {
+    e->Travers([&](EXPR *e, EXPR *itr) {
+      (*outmap)[itr].push_back(e);
+      return true;
+    });
+  }
+  Clear(vec);
+}
+
+void EXPR::PrintDot(std::vector<EXPR *> vec) {
+  std::cout << "digraph \"expr\" { " << std::endl;
+  for (auto e : vec) {
+    e->Travers([](EXPR *e, EXPR *itr) {
+      std::cout << "\"" << itr->node_->name() << "\"->"
+                << "\"" << e->node_->name() << "\"" << std::endl;
+      return true;
+    });
+  }
+  std::cout << "}" << std::endl;
+  Clear(vec);
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/expr.h

@@ -0,0 +1,70 @@
+/**
+ * Copyright 2022 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_EXPRESSION_EXPR_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_EXPR_H_
+
+#include <vector>
+#include <list>
+#include <memory>
+#include <map>
+#include <functional>
+#include <string>
+#include "include/ms_tensor.h"
+#include "include/api/format.h"
+
+namespace mindspore {
+namespace lite {
+class Node;
+
+class EXPR {
+ public:
+  explicit EXPR(Node *node) : node_(node) { SetSize(1); }
+  static void PrintDot(std::vector<EXPR *> vec);
+  static void Replace(const std::vector<EXPR *> &vec, std::vector<EXPR *> *old, std::vector<EXPR *> *n);
+  static void CreateOutputMap(std::vector<EXPR *> vec, std::map<EXPR *, std::list<EXPR *>> *outmap);
+  static void Clear(std::vector<EXPR *> vec);
+  void Travers(std::function<bool(EXPR *e, EXPR *itr)> cb);
+  std::string name();
+  EXPR *GetInput(int idx) { return params_.at(idx); }
+  void set_node(Node *node) { node_ = node; }
+  Node *node() { return node_; }
+  bool visited = false;
+  void set_params(std::vector<EXPR *> params) { params_ = params; }
+  void set_params(int idx, EXPR *expr) { params_[idx] = expr; }
+  void add_params(EXPR *e) { params_.push_back(e); }
+  std::vector<EXPR *> &params() { return params_; }
+  EXPR *params(int i) { return params_[i]; }
+  void SetSize(int n) { params_.resize(n); }
+  void SetDims(std::vector<int> dims) { dims_ = dims; }
+  std::vector<int> &dims() { return dims_; }
+  void set_format(int fmt) { format_ = fmt; }
+  int format() { return format_; }
+  void set_data_type(TypeId data_type) { data_type_ = data_type; }
+  TypeId data_type() { return data_type_; }
+
+ private:
+  void Replace(EXPR **old, EXPR **n, std::vector<Node *> *to_delete);
+  std::vector<EXPR *> params_;
+  Node *node_{nullptr};
+  void Clear();
+  std::vector<int> dims_;
+  int format_ = NHWC;
+  TypeId data_type_ = kNumberTypeFloat32;
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_EXPR_H_

mindspore/lite/src/expression/import.cc

@@ -0,0 +1,180 @@
+/**
+ * Copyright 2022 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/expression/import.h"
+#include "ops/populate/populate_register.h"
+#include "src/expression/ops.h"
+#include "src/expression/ops/activation.h"
+#include "src/expression/ops/batchnorm.h"
+#include "src/expression/ops/biasadd.h"
+#include "src/expression/ops/conv.h"
+#include "src/expression/ops/dense.h"
+#include "src/expression/ops/pooling.h"
+#include "src/expression/ops/reshape.h"
+#include "src/expression/ops/transpose.h"
+
+namespace mindspore {
+namespace lite {
+std::unordered_map<mindspore::schema::PrimitiveType, import_func> ImportReg::import_map_;
+
+import_func ImportReg::GetImportFunc(mindspore::schema::PrimitiveType type) {
+  auto f = import_map_.find(type);
+  if (f == import_map_.end()) {
+    return nullptr;
+  }
+  return f->second;
+}
+
+OpParameter *Import::GetAttr(const schema::Primitive *prim) {
+  auto parameter_gen = PopulateRegistry::GetInstance()->GetParameterCreator(prim->value_type(), SCHEMA_CUR);
+  if (parameter_gen == nullptr) {
+    MS_LOG(ERROR) << "PopulateParameter return nullptr, type: " << schema::EnumNamePrimitiveType(prim->value_type());
+    return nullptr;
+  }
+  auto parameter = parameter_gen(prim);
+  if (parameter == nullptr) {
+    MS_LOG(ERROR) << "parameter is nullptr.";
+    return nullptr;
+  }
+  return parameter;
+}
+
+std::unique_ptr<Node> Import::CreateNode(const schema::CNode *cnode) {
+  auto param = GetAttr(cnode->primitive());
+  auto type = cnode->primitive()->value_type();
+  auto fn = ImportReg::GetImportFunc(type);
+  if (fn == nullptr) {
+    MS_LOG(ERROR) << "Cannot find importer for " << schema::EnumNamePrimitiveType(type);
+    return nullptr;
+  }
+  auto node = fn();
+  if (node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate node" << cnode->name()->str();
+    return nullptr;
+  }
+  node->SetOpParam(param);
+  node->set_name(cnode->name()->str());
+  node->set_primitive(type);
+  return std::unique_ptr<Node>(node);
+}
+
+Net *Import::ImportMs(std::string file_name) {
+  std::ifstream infile;
+  infile.open(file_name, std::ios::binary | std::ios::in);
+  if (!infile.good()) {
+    MS_LOG(ERROR) << "cannot read " << file_name << std::endl;
+    return nullptr;
+  }
+  infile.seekg(0, std::ios::end);
+  int length = infile.tellg();
+  infile.seekg(0, std::ios::beg);
+  auto data_ptr = std::make_unique<int8_t[]>(length);
+  auto *data = data_ptr.get();
+  infile.read(reinterpret_cast<char *>(data), length);
+  infile.close();
+  flatbuffers::Verifier verifier = flatbuffers::Verifier(reinterpret_cast<const uint8_t *>(data), length);
+  bool res = schema::VerifyMetaGraphBuffer(verifier);
+  if (res != true) {
+    MS_LOG(ERROR) << "fault file: " << file_name << "(" << length << ")\n";
+    return nullptr;
+  } else {
+    MS_LOG(INFO) << "verify pass file: " << file_name << "(" << length << ")\n";
+  }
+  buffer_ = data_ptr.get();
+  auto metaGraph = schema::GetMetaGraph(data_ptr.release());
+  return ImportMs(metaGraph);
+}
+
+Net *Import::ImportMs(const schema::MetaGraph *metaGraph) {
+  if (metaGraph == nullptr) {
+    MS_LOG(ERROR) << "null input";
+    return nullptr;
+  }
+  std::string NetName = "Network";
+  if (metaGraph->name() != nullptr) NetName = metaGraph->name()->str();
+  auto net = std::make_unique<Net>(NetName);
+  std::unordered_map<int, EXPR *> outputs;
+  // save inputs
+  for (size_t i = 0; i < metaGraph->inputIndex()->size(); i++) {
+    auto tensor_id = metaGraph->inputIndex()->Get(i);
+    const schema::Tensor *tensor = metaGraph->allTensors()->Get(tensor_id);
+    auto input = new (std::nothrow) InputM(tensor);
+    if (input == nullptr) {
+      MS_LOG(ERROR) << "Cannot allocate input";
+      return nullptr;
+    }
+    auto e = input->expr();
+    outputs[tensor_id] = e;
+    net->PushInput(e);
+  }
+  for (size_t i = 0; i < metaGraph->nodes()->size(); i++) {
+    auto Cnode = metaGraph->nodes()->Get(i);
+    std::vector<EXPR *> param_tensors;
+    for (size_t j = 0; j < Cnode->inputIndex()->size(); j++) {
+      int tensor_id = Cnode->inputIndex()->Get(j);
+      const schema::Tensor *tensor = metaGraph->allTensors()->Get(tensor_id);
+      auto iter = outputs.find(tensor_id);
+      if (iter == outputs.end()) {
+        // create value node if not exist
+        if (tensor->nodeType() != NodeType::NodeType_CNode) {
+          auto valnode = new (std::nothrow) InputM(tensor);
+          if (valnode == nullptr) {
+            MS_LOG(ERROR) << "Cannot allocate valnode";
+            return nullptr;
+          }
+          outputs[tensor_id] = valnode->expr();
+          param_tensors.push_back(valnode->expr());
+          net->PushOp(valnode);
+        } else {
+          MS_LOG(ERROR) << "did not found input tensor " << tensor_id;
+          return nullptr;
+        }
+      } else {
+        param_tensors.push_back(iter->second);
+      }
+    }
+    // create expression from node //
+    auto node = CreateNode(Cnode);
+    if (node != nullptr) {
+      node->SetOutputs(Cnode->outputIndex()->size());
+      std::vector<EXPR *> e = (*node)(param_tensors);
+      for (size_t j = 0; j < Cnode->outputIndex()->size(); j++) {
+        int tensor_id = Cnode->outputIndex()->Get(j);
+        outputs[tensor_id] = e.at(j);
+      }
+    } else {
+      MS_LOG(ERROR) << "failed to create node " << Cnode->name();
+      return nullptr;
+    }
+    auto node_ptr = node.release();
+    net->PushOp(node_ptr);
+    node_ptr->SetLearn();
+  }
+  for (size_t i = 0; i < metaGraph->outputIndex()->size(); i++) {
+    auto tensor_id = metaGraph->outputIndex()->Get(i);
+    auto iter = outputs.find(tensor_id);
+    if (iter == outputs.end()) {
+      MS_LOG(ERROR) << "could not find source for tensor " << tensor_id;
+      return nullptr;
+    } else {
+      net->PushOutput(iter->second);
+    }
+  }
+  return net.release();
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/import.h

@@ -0,0 +1,61 @@
+/**
+ * Copyright 2022 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_EXPRESSION_IMPORT_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_IMPORT_H_
+
+#include <string>
+#include <unordered_map>
+#include <memory>
+#include "nnacl/op_base.h"
+#include "src/expression/net.h"
+
+namespace mindspore {
+namespace lite {
+using import_func = std::function<Node *()>;
+
+template <typename T>
+Node *ReturnNode() {
+  return new (std::nothrow) T();
+}
+
+class ImportReg {
+ public:
+  explicit ImportReg(mindspore::schema::PrimitiveType type, import_func func) { import_map_[type] = func; }
+  static import_func GetImportFunc(mindspore::schema::PrimitiveType type);
+
+ private:
+  static std::unordered_map<mindspore::schema::PrimitiveType, import_func> import_map_;
+};
+
+class Import {
+ private:
+  int8_t *buffer_ = nullptr;
+  OpParameter *GetAttr(const schema::Primitive *prim);
+  std::unique_ptr<Node> CreateNode(const schema::CNode *cnode);
+
+ public:
+  Net *ImportMs(const schema::MetaGraph *meta_graph);
+  Net *ImportMs(std::string file_name);
+  ~Import() {
+    delete[] buffer_;
+    buffer_ = nullptr;
+  }
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  //  MINDSPORE_LITE_SRC_EXPRESSION_IMPORT_H_

mindspore/lite/src/expression/net.cc

@@ -0,0 +1,268 @@
+/**
+ * Copyright 2022 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/expression/net.h"
+#include <vector>
+#include "src/cxx_api/expression/net_impl.h"
+#include "src/expression/ops.h"
+#include "src/expression/export.h"
+#include "src/expression/ops/addn.h"
+#include "src/expression/ops/arithmetic.h"
+#include "src/common/storage.h"
+#include "tools/common/meta_graph_serializer.h"
+namespace mindspore {
+namespace lite {
+void Net::update_name(std::string name) {
+  if (!this->name().empty())
+    Node::update_name(name);
+  else
+    set_name(name);
+  for (auto &itr : ops_) {
+    itr->update_name(name);
+  }
+}
+
+std::vector<EXPR *> Net::operator()(const std::initializer_list<EXPR *> &&inputs) {
+  std::vector<EXPR *> vec = inputs;
+  std::vector<EXPR *> x;
+  if (impl_ == nullptr) {
+    x = construct(inputs);
+  } else {
+    x = impl_->construct(vec);
+  }
+  return x;
+}
+
+std::vector<EXPR *> Net::operator()(const std::vector<EXPR *> &inputs) {
+  std::vector<EXPR *> x;
+  if (impl_ == nullptr) {
+    x = construct(inputs);
+  } else {
+    x = impl_->construct(inputs);
+  }
+  input_ = inputs;
+  output_ = x;
+  real_output_ = x;
+  return x;
+}
+
+std::vector<EXPR *> Net::construct(const std::vector<EXPR *> &inputs) {
+  if (!output_.empty()) {
+    if (input_.size() != inputs.size()) {
+      MS_LOG(ERROR) << "input size mismatch, should be " << input_.size() << " got " << inputs.size();
+      return {};
+    }
+    auto in_ptr = inputs;
+    EXPR::Replace(output_, &input_, &in_ptr);
+  } else {
+    MS_LOG(ERROR) << "no network construction function";
+  }
+  return output_;
+}
+
+void Net::TopoSortUtil(Node *node, std::stack<Node *> *stack) {
+  visited_.insert(node);
+  for (size_t i = 0; i < node->OutputsNum(); i++) {
+    auto expr = node->expr(i);
+    auto itr = outmap_.find(expr);
+    if (itr != outmap_.end()) {
+      for (auto &e : itr->second)
+        if (visited_.find(e->node()) == visited_.end()) {
+          TopoSortUtil(e->node(), stack);
+        }
+    }
+  }
+  stack->push(node);
+}
+
+std::vector<Node *> Net::Sort() {
+  std::stack<Node *> stack;
+  outmap_.clear();
+  EXPR::CreateOutputMap(output_, &outmap_);
+  for (auto &itr : outmap_) {
+    EXPR *e = itr.first;
+    if (visited_.find(e->node()) == visited_.end()) {
+      TopoSortUtil(e->node(), &stack);
+    }
+  }
+  std::vector<Node *> res;
+  while (stack.empty() == false) {
+    res.push_back(stack.top());
+    stack.pop();
+  }
+  visited_.clear();
+  return res;
+}
+
+std::unique_ptr<schema::MetaGraphT> Net::MakeMs() {
+  auto nodes = Sort();
+  auto s = new (std::nothrow) ExportSession(outmap_);
+  if (s == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate export session";
+    return nullptr;
+  }
+  session_.reset(s);
+  session_->Init(name(), Version());
+  for (auto node : nodes) {
+    auto res = node->MakeEntry(session_.get());
+    if (res != RET_OK) {
+      MS_LOG(ERROR) << "failed in MakeEntry: " << node->name();
+      return nullptr;
+    }
+  }
+  session_->SetInputOutput(input_, real_output_);
+  auto res = session_->meta_graph();
+  return std::unique_ptr<schema::MetaGraphT>(res);
+}
+
+std::unique_ptr<schema::MetaGraphT> Net::MakeMs(const std::string file_name) {
+  auto graph = MakeMs();
+  Save(*graph, file_name);
+  return graph;
+}
+
+std::set<Node *> Net::trainable_params() {
+  std::set<Node *> res;
+  for (auto &node : ops_) {
+    res.merge(node->trainable_params());
+  }
+  return res;
+}
+
+int Net::BuildGrad(Node *optimizer) {
+  std::set<Node *> learn = optimizer->trainable_params();
+  auto NetOrder = Sort();
+  optimizer_.reset(optimizer);
+  optimizer->AddNetOutput(&output_);
+  std::map<std::pair<EXPR *, EXPR *>, EXPR *> backprop;
+  for (auto itr = NetOrder.rbegin(); itr != NetOrder.rend(); itr++) {
+    Node *node = *itr;
+    EXPR *yt = nullptr;
+    if (node->primitive() == schema::PrimitiveType_NONE) continue;
+    if (outmap_.find(node->expr()) == outmap_.end() || outmap_[node->expr()].size() == 0) {
+      yt = node->expr();
+    } else {
+      std::vector<EXPR *> add_params;
+      for (auto &output : outmap_[node->expr()]) {
+        auto link = std::make_pair(node->expr(), output);
+        auto grad = backprop[link];
+        add_params.push_back(grad);
+      }
+      if (add_params.size() == 1) {
+        yt = add_params.front();
+      } else {
+        auto addn = new (std::nothrow) AddN(0);
+        if (addn == nullptr) {
+          MS_LOG(ERROR) << "Cannot allocate add operator";
+          return RET_ERROR;
+        }
+        PushOp(addn);
+        addn->update_name(name());
+        yt = (*addn)(add_params).front();
+      }
+    }
+    auto inGrads = node->Grad(yt);
+    for (size_t i = 0; i < node->inputs().size(); i++) {
+      EXPR *inGrad{nullptr};
+      if (i < inGrads.size()) {
+        inGrad = inGrads[i];
+      } else {
+        inGrad = nullptr;
+      }
+      auto input = node->input(i);
+      if (learn.find(input->node()) != learn.end()) {
+        auto opt = optimizer->Clone(inGrad, input);
+        if (opt.size() == 0) {
+          MS_LOG(ERROR) << "failed to create optimizer";
+          return RET_ERROR;
+        }
+        if (inGrad == nullptr) {
+          MS_LOG(ERROR) << "illegal null value for grad";
+          return RET_ERROR;
+        }
+        if (opt.size() == 0) {
+          MS_LOG(ERROR) << "optimizer for " << input->node()->name() << " failure";
+          return RET_ERROR;
+        }
+        auto opt_op = opt.at(0)->node();
+        PushOp(opt_op);
+        opt_op->update_name(node->name());
+        output_.push_back(opt.at(0));
+      }
+      auto link = std::make_pair(input, node->expr());
+      backprop[link] = inGrad;
+    }
+  }
+  return RET_OK;
+}
+
+std::vector<EXPR *> Net::add(const std::vector<EXPR *> &input) {
+  auto _add = NN::Add();
+  _add->set_name(name() + "/" + _add->name());
+  ops_.push_back(_add);
+  return (*_add)(input);
+}
+
+Net *Net::TrainNet(Node *optimizer, Node *loss_fn, const std::vector<EXPR *> &inputs) {
+  auto net = new (std::nothrow) NetWithLoss(this, loss_fn);
+  if (net == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate loss network";
+    return nullptr;
+  }
+  return net->TrainNet(optimizer, inputs);
+}
+
+Net *Net::TrainNet(Node *optimizer, const std::vector<EXPR *> &inputs) {
+  auto x = (*this)(inputs);
+  auto res = BuildGrad(optimizer);
+  if (res != RET_OK) {
+    MS_LOG(ERROR) << "Build gradient network failed";
+    return nullptr;
+  }
+  real_output_ = x;
+  return this;
+}
+
+int Net::Save(const schema::MetaGraphT &graph, std::string file_name) { return Storage::Save(graph, file_name); }
+
+const std::vector<int> Net::OutputShape(int idx) {
+  if (static_cast<size_t>(idx) >= real_output_.size()) {
+    MS_LOG(ERROR) << "index (" << idx << ") exceed output size (" << real_output_.size() << ")";
+    return {};
+  }
+  return real_output_.at(idx)->dims();
+}
+
+const std::vector<int> Net::InputShape(int idx) {
+  if (static_cast<size_t>(idx) >= input_.size()) {
+    MS_LOG(ERROR) << "index (" << idx << ") exceed input size (" << input_.size() << ")";
+    return {};
+  }
+  return input_.at(idx)->dims();
+}
+
+Net::~Net() {
+  if (impl_ != nullptr) {
+    impl_->erase_net();
+    auto pnet = impl_->pnet();
+    if (pnet != nullptr) {
+      impl_->set_pnet(nullptr);
+    }
+  }
+  impl_ = nullptr;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/net.h

@@ -0,0 +1,114 @@
+/**
+ * Copyright 2022 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_EXPRESSION_NET_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_NET_H_
+#include <stack>
+#include <memory>
+#include <set>
+#include <map>
+#include <utility>
+#include <string>
+#include <unordered_set>
+#include <list>
+#include <vector>
+#include "src/expression/node.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+class Net;
+class NetImpl;
+namespace lite {
+#define REG(_name) Register(_name, #_name)
+
+class ExportSession;
+
+class Net : public Node {
+ public:
+  Net() = default;
+  virtual ~Net();
+  explicit Net(std::string name) : Node(name) {}
+  std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs) override;
+  std::vector<EXPR *> operator()(const std::vector<EXPR *> &inputs) override;
+  std::vector<EXPR *> operator()(const std::initializer_list<EXPR *> &&inputs) override;
+  void update_name(std::string name) override;
+  Net *TrainNet(Node *optimizer, Node *loss_fn, const std::vector<EXPR *> &inputs);
+  Net *TrainNet(Node *optimizer, const std::vector<EXPR *> &inputs);
+  void PrintDot() { EXPR::PrintDot(output_); }
+
+  void PushOutput(EXPR *e) { output_.push_back(e); }
+  void PushInput(EXPR *e) { input_.push_back(e); }
+  void SetRealOutput() { real_output_ = output_; }
+  std::set<Node *> trainable_params() override;
+  std::vector<Node *> Sort();
+  int BuildGrad(Node *optimizer);
+  int BuildGrad(Node *optimizer, std::set<Node *> learnable);
+  std::unique_ptr<schema::MetaGraphT> MakeMs();
+  std::unique_ptr<schema::MetaGraphT> MakeMs(std::string file_name);
+  schema::MetaGraph *meta_graph() { return meta_graph_; }
+  int Save(const schema::MetaGraphT &graph, const std::string filename);
+  void set_impl(std::shared_ptr<mindspore::NetImpl> impl) { impl_ = impl; }
+  const std::vector<int> InputShape(int idx);
+  const std::vector<int> OutputShape(int idx);
+
+ protected:
+  std::vector<EXPR *> add(const std::vector<EXPR *> &input);
+  void Register(Node *node, std::string &&name) {
+    if (node != nullptr) {
+      PushOp(node);
+      node->update_name(name);
+    }
+  }
+
+ private:
+  friend mindspore::Net;
+  std::unordered_set<Node *> visited_;
+  std::map<EXPR *, std::list<EXPR *>> outmap_;  // outputs per expression
+  std::map<EXPR *, std::list<EXPR *>> inmap_;   // inputs per expression
+  std::vector<EXPR *> output_;                  // network output expression
+  std::vector<EXPR *> real_output_;             // network output for export
+  std::vector<EXPR *> input_;                   // network input expression
+  schema::MetaGraph *meta_graph_;               // imported meta_graph
+  std::unique_ptr<ExportSession> session_;      // export session
+  std::unique_ptr<Node> optimizer_;
+  void TopoSortUtil(Node *v, std::stack<Node *> *stack);
+  void CreateOutputMap(std::vector<EXPR *> vec, std::map<Node *, std::list<Node *>> *outmap);
+  std::shared_ptr<mindspore::NetImpl> impl_;
+};
+
+class NetWithLoss : public Net {
+ public:
+  NetWithLoss(Net *net, Node *loss) : net_(net), loss_fn_(loss) {
+    REG(net_);
+    REG(loss_fn_);
+    loss_fn_->set_name("_loss_fn");
+  }
+  std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs) {
+    auto input = inputs[0];
+    auto label = inputs[1];
+    auto x = (*net_)({input});
+    x = (*loss_fn_)({x[0], label});
+    return {x.front()};
+  }
+
+ private:
+  Net *net_{nullptr};
+  Node *loss_fn_{nullptr};
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_NET_H_

mindspore/lite/src/expression/node.cc

@@ -0,0 +1,271 @@
+/**
+ * Copyright 2022 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 <algorithm>
+#include <utility>
+#include <functional>
+#include "src/expression/node.h"
+#include "src/expression/ops.h"
+#include "src/expression/export.h"
+#include "src/runtime/infer_manager.h"
+#include "src/common/utils.h"
+#include "src/cxx_api/expression/net_impl.h"
+
+namespace mindspore {
+namespace lite {
+int Node::name_id;
+
+std::vector<EXPR *> Node::construct(const std::vector<EXPR *> &inputs) {
+  if (inputs.size() >= expr()->params().size()) {
+    expr()->set_params(inputs);
+  } else {
+    for (std::size_t i = 0; i < inputs.size(); i++) {
+      expr()->set_params(i, inputs[i]);
+    }
+  }
+  auto ret = InferShape();
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "error infershape for node " << name();
+    return {};
+  }
+  std::vector<EXPR *> res(expr_.size());
+  (void)std::transform(expr_.begin(), expr_.end(), res.begin(), [](const EXPR &e) { return const_cast<EXPR *>(&e); });
+  return res;
+}
+
+std::vector<EXPR *> Node::Grad(EXPR *expr) {
+  MS_LOG(ERROR) << name() << " (" << schema::EnumNamePrimitiveType(primitive()) << ") does not have grad defined";
+  return {};
+}
+
+int Node::CreateTensorFromExpr(const std::vector<EXPR *> &expr, std::vector<Tensor *> *tensors, bool is_input) {
+  MS_ASSERT(tensors != nullptr);
+  int ret = RET_OK;
+  for (auto e : expr) {
+    // Tensor -> TensorC
+    if (is_input && e->node()->primitive() == schema::PrimitiveType_Depend) {
+      continue;
+    }
+    auto type = (e->node()->primitive() != schema::PrimitiveType_NONE) ? Category::VAR : Category::CONST_TENSOR;
+    auto t = std::make_unique<Tensor>(e->data_type(), e->dims(), (mindspore::Format)e->format(), type);
+    if (t == nullptr) {
+      ret = RET_NULL_PTR;
+      break;
+    }
+    // copy data if any
+    if (type == Category::CONST_TENSOR) {
+      void *dst = t->MutableData();
+      if (dst == nullptr) {
+        ret = RET_NULL_PTR;
+        break;
+      }
+      if (e->node()->data() && (e->node()->data()->data().size() > 0)) {
+        uint8_t *src = e->node()->data()->data().data();
+        memcpy(dst, src, t->Size());
+      }
+    }
+    tensors->push_back(t.release());
+  }
+  return ret;
+}
+
+void Node::FreeAllTensors(std::vector<Tensor *> *tensors) {
+  MS_ASSERT(tensors != nullptr);
+  for (auto &t : *tensors) {
+    delete t;
+  }
+  tensors->clear();
+}
+
+int Node::InferShape() {
+  auto ret = RET_OK;
+  std::vector<Tensor *> in_tensors;
+  std::vector<Tensor *> out_tensors;
+  // build in \ out tensors
+  ret = CreateTensorFromExpr(expr()->params(), &in_tensors, true);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Failed in create in tensors";
+    FreeAllTensors(&in_tensors);
+    return RET_ERROR;
+  }
+  std::vector<EXPR *> expr(expr_.size());
+  (void)std::transform(expr_.begin(), expr_.end(), expr.begin(), [](const EXPR &e) { return const_cast<EXPR *>(&e); });
+  ret = CreateTensorFromExpr(expr, &out_tensors);
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "Failed in create out tensors";
+    FreeAllTensors(&in_tensors);
+    FreeAllTensors(&out_tensors);
+    return RET_ERROR;
+  }
+  // Do infer Shape
+  ret = KernelInferShape(in_tensors, out_tensors, OpParam());
+  if (ret != RET_OK) {
+    MS_LOG(ERROR) << "failed in infer shape for " << name();
+    FreeAllTensors(&in_tensors);
+    FreeAllTensors(&out_tensors);
+    return RET_ERROR;
+  }
+  // copy infer shape into expr
+  for (uint32_t i = 0; i < expr_.size(); i++) {
+    auto e = &expr_.at(i);
+    auto o = out_tensors.at(i);
+    e->set_format((o->format()));
+    e->set_data_type(o->data_type());
+    e->SetDims(o->shape());
+  }
+  // cleanup
+  FreeAllTensors(&in_tensors);
+  FreeAllTensors(&out_tensors);
+
+  return ret;
+}
+
+EXPR *Node::CreateWeights(std::vector<int> dims, TypeId data_type, int format, Param::Mode mode, std::string name) {
+  auto weights = new (std::nothrow) InputM(dims);
+  if (weights == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate weights";
+    return nullptr;
+  }
+  weights->set_name(this->name() + "/" + name);
+  int size = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<int>());
+  weights->data()->SetSize(size);
+  weights->data()->Fill(mode);
+  PushOp(weights);
+  return weights->expr();
+}
+
+Node *Node::CreateConstTensor(int index, std::vector<int> dims, TypeId data_type, int format, std::string name,
+                              const void *data) {
+  auto tensor = NN::Input(dims, data_type, format);
+  int elem_size = DataTypeSize(data_type);
+  tensor->set_name(this->name() + "/" + name);
+  int size = std::accumulate(dims.begin(), dims.end(), 1, std::multiplies<int>()) * elem_size;
+  tensor->data()->SetSize(size);
+  tensor->data()->Copy(reinterpret_cast<const uint8_t *>(data), size);
+  expr()->set_params(index, tensor->expr());
+  PushOp(tensor);
+  return tensor;
+}
+
+int Node::MakeEntry(ExportSession *session) {
+  std::vector<uint32_t> input_idx;
+  std::vector<uint32_t> output_idx;
+  std::vector<uint8_t> empty;
+  if (primitive() == schema::PrimitiveType_Depend) return RET_OK;
+  // create node input
+  size_t inputs = InputsNum();
+  for (size_t i = 0; i < inputs; i++) {
+    EXPR *ex = expr()->GetInput(i);
+    if (ex->node()->primitive() == schema::PrimitiveType_Depend) continue;
+    uint32_t id = session->GetOutput(ex);
+    input_idx.push_back(id);
+  }
+  size_t outputs = OutputsNum();
+  size_t last_id = session->meta_graph()->allTensors.size();
+  int type = (primitive() == schema::PrimitiveType_NONE) ? NodeType_ValueNode : NodeType_CNode;
+  auto data = (type == NodeType_ValueNode) ? this->data()->data() : empty;
+  if (data.empty()) type = NodeType_CNode;  // input is Cnode !!?
+  int idx = 0;
+  for (size_t i = 0; i < outputs; i++) {
+    if (session->IsToDependOnly(expr(i))) continue;
+    output_idx.push_back(last_id + idx);
+    session->UpdateOutput(expr(i), last_id + idx);
+    auto odims = dims(i);
+    auto data_type = expr(i)->data_type();
+    auto format = expr(i)->format();
+    std::string footer = (i > 0) ? ("-" + std::to_string(i)) : "";
+    auto otensor = CreateTensor(name() + footer, type, data_type, odims, format, data);
+    std::cout << "tensor -" << last_id + idx << ": " << name() + footer << std::endl;
+    idx++;
+    session->meta_graph()->allTensors.emplace_back(std::move(otensor));
+  }
+  if (primitive() != schema::PrimitiveType_NONE) {
+    if (output_idx.size() == 0) {
+      return RET_OK;
+    }
+    auto cnode = CreateCNode(input_idx, output_idx);
+
+    auto ret = UnPopulate(cnode);
+    if (ret != RET_OK) {
+      MS_LOG(ERROR) << "failed to populate cnode";
+      return RET_ERROR;
+    }
+    session->meta_graph()->nodes.emplace_back(std::move(cnode));
+  }
+
+  return RET_OK;
+}
+
+std::unique_ptr<schema::CNodeT> Node::CreateCNode(std::vector<uint32_t> inputIndex, std::vector<uint32_t> outputIndex) {
+  auto cnode = std::make_unique<schema::CNodeT>();
+  cnode->primitive = std::make_unique<schema::PrimitiveT>();
+  cnode->primitive->value.type = primitive();
+  cnode->name = name();
+  cnode->inputIndex = inputIndex;
+  cnode->outputIndex = outputIndex;
+  return cnode;
+}
+
+int Node::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  MS_LOG(ERROR) << "Node " << schema::EnumNamePrimitiveType(primitive()) << " cannot be exported";
+  return RET_ERROR;
+}
+
+std::unique_ptr<mindspore::schema::TensorT> Node::CreateTensor(std::string name, int type, int data_type,
+                                                               const std::vector<int32_t> dims, int format,
+                                                               const std::vector<uint8_t> &data) {
+  auto tensorT = std::make_unique<mindspore::schema::TensorT>();
+  tensorT->nodeType = type;
+  tensorT->dims = dims;
+  tensorT->format = static_cast<schema::Format>(format);
+  tensorT->name = name;
+  tensorT->refCount = 0;
+  tensorT->offset = 0;
+  tensorT->dataType = data_type;
+  tensorT->data = data;
+  tensorT->enableHuffmanCode = false;
+  if (tensorT->nodeType == mindspore::lite::NodeType_ValueNode) {
+    tensorT->data = data;
+  }
+  return tensorT;
+}
+
+int Node::SetOutputs(int num) {
+  EXPR e(this);
+  e.SetSize(0);
+  for (auto i = expr_.size(); i < static_cast<size_t>(num); i++) {
+    expr_.emplace_back(e);
+  }
+  return RET_OK;
+}
+
+Node::~Node() {
+  for (auto &op : ops_) {
+    delete op;
+  }
+  ops_.clear();
+  if (impl_ != nullptr) {
+    impl_->set_node(nullptr);
+    auto pnode = impl_->pnode();
+    if (pnode != nullptr) {
+      impl_->set_pnode(nullptr);
+      delete pnode;
+    }
+  }
+  impl_ = nullptr;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/node.h

@@ -0,0 +1,156 @@
+/**
+ * Copyright 2022 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_EXPRESSION_NODE_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_NODE_H_
+
+#include <stdlib.h>
+#include <vector>
+#include <string>
+#include <iostream>
+#include <memory>
+#include <set>
+#include "src/expression/export.h"
+#include "inner/model_generated.h"
+#include "src/expression/param.h"
+#include "src/expression/expr.h"
+#include "src/tensor.h"
+#include "nnacl/op_base.h"
+
+namespace mindspore {
+class NodeImpl;
+namespace schema {
+struct TensorT;
+struct CNodeT;
+}  // namespace schema
+
+namespace lite {
+class Node {
+ public:
+  const std::string kGradName = "Gradients";
+  explicit Node(const std::string name) : opParam_(nullptr), name_(name) { expr_.emplace_back(this); }
+  virtual ~Node();
+  Node() : Node("") {}
+  explicit Node(Node *node) : Node(*node) {}
+  EXPR *create(std::string name) {
+    name_ = name;
+    return &expr_[0];
+  }
+  virtual std::vector<EXPR *> operator()(const std::vector<EXPR *> &inputs) {
+    auto x = construct(inputs);
+    return x;
+  }
+  virtual std::vector<EXPR *> operator()(const std::initializer_list<EXPR *> &&inputs) {
+    std::vector<EXPR *> vec = inputs;
+    auto x = construct(vec);
+    return x;
+  }
+  virtual std::vector<EXPR *> operator()(const std::initializer_list<EXPR *> &inputs) {
+    std::vector<EXPR *> vec = inputs;
+    auto x = construct(vec);
+    return x;
+  }
+  void set_primitive(schema::PrimitiveType primitive) {
+    primitive_ = primitive;
+    if (OpParam() != nullptr) opParam_->type_ = primitive_;
+  }
+  schema::PrimitiveType primitive() { return primitive_; }
+  virtual std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs);
+  std::string name() { return name_; }
+  void set_name(std::string name) { name_ = name; }
+  virtual void update_name(std::string name) { set_name(name + "/" + name_); }
+  size_t Load(std::string file_name, size_t offset = 0) { return offset; }
+  OpParameter *OpParam() const { return opParam_.get(); }
+  virtual void Add(Node *node) {}
+  virtual std::vector<EXPR *> Clone(EXPR *grad, EXPR *weight) { return {}; }
+  void SetOpParam(std::shared_ptr<OpParameter> opParam) { opParam_ = opParam; }
+  void SetOpParam(void *opParam) { opParam_.reset(reinterpret_cast<OpParameter *>(opParam), free); }
+  static std::string UniqueName(const std::string &name) { return name + "-" + std::to_string(name_id++); }
+  static std::string UniqueName(std::string &&name) { return name + "-" + std::to_string(name_id++); }
+  template <typename T>
+  int CloneOpParam(std::shared_ptr<OpParameter> opParam) {
+    auto t = reinterpret_cast<T *>(opParam.get());
+    auto obj = new (std::nothrow) T(*t);  // copy content
+    if (obj == nullptr) {
+      MS_LOG(ERROR) << "Cannot allocate obj";
+      return RET_ERROR;
+    }
+    opParam_.reset(reinterpret_cast<OpParameter *>(obj));
+    return RET_OK;
+  }
+  template <typename T>
+  int CloneOpParam(OpParameter *opParam) {
+    auto t = reinterpret_cast<T *>(opParam);
+    auto obj = new (std::nothrow) T(*t);  // copy content
+    if (obj == nullptr) {
+      MS_LOG(ERROR) << "Cannot allocate obj";
+      return RET_ERROR;
+    }
+    opParam_.reset(reinterpret_cast<OpParameter *>(obj));
+    return RET_OK;
+  }
+  virtual Param *weight() { return nullptr; }
+  EXPR *expr(int i) { return &expr_[i]; }
+  EXPR *expr() { return expr(0); }
+  std::vector<EXPR *> inputs() { return expr()[0].params(); }
+  size_t InputsNum() { return expr()[0].params().size(); }
+  size_t OutputsNum() { return expr_.size(); }
+  EXPR *input(int idx) { return expr()[0].params().at(idx); }
+  EXPR *output(int idx) { return expr(idx); }
+  EXPR *CreateWeights(std::vector<int> dims, TypeId data_type, int format, Param::Mode mode, std::string name);
+  Node *CreateConstTensor(int index, std::vector<int> dims, TypeId data_type, int format, std::string name,
+                          const void *data);
+  virtual std::vector<EXPR *> Grad(EXPR *expr);
+  virtual Param *data() { return nullptr; }
+  bool IsLearn(Node *node) { return learnable_.find(node) != learnable_.end(); }
+  virtual void SetLearn() {}
+  virtual std::set<Node *> trainable_params() { return learnable_; }
+  std::vector<int> &dims() { return expr()->dims(); }
+  std::vector<int> &dims(int i) { return expr(i)->dims(); }
+  // export
+  int MakeEntry(ExportSession *session);
+  void PushOp(Node *n) { ops_.push_back(n); }
+  virtual void AddNetOutput(std::vector<EXPR *> *output) {}
+  int SetOutputs(int num);
+  std::shared_ptr<OpParameter> opParam_;
+  void set_impl(std::shared_ptr<NodeImpl> impl) { impl_ = impl; }
+
+ protected:
+  std::vector<EXPR> expr_;   // hold outputs
+  std::vector<Node *> ops_;  // all nodes or subnets
+  int InferShape();
+  void AddLearn(Node *node) { learnable_.insert(node); }
+  void AssignLearn(std::set<Node *> &&learn) { learnable_ = learn; }
+
+  std::unique_ptr<schema::CNodeT> CreateCNode(std::vector<uint32_t> inputIndex, std::vector<uint32_t> outputIndex);
+  virtual int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode);
+  std::unique_ptr<schema::TensorT> CreateTensor(std::string name, int type, int data_type,
+                                                const std::vector<int32_t> dims, int format,
+                                                const std::vector<uint8_t> &data);
+
+ private:
+  int CreateTensorFromExpr(const std::vector<EXPR *> &expr, std::vector<Tensor *> *tensors, bool is_input = false);
+  void FreeAllTensors(std::vector<Tensor *> *tensors);
+  static int name_id;
+  std::set<Node *> learnable_;  // set of nodes with learnable parameters
+  std::string name_;
+  schema::PrimitiveType primitive_;
+  std::shared_ptr<NodeImpl> impl_;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_NODE_H_

mindspore/lite/src/expression/ops.cc

@@ -0,0 +1,66 @@
+
+/**
+ * Copyright 2022 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 <numeric>
+#include <algorithm>
+#include "src/expression/ops.h"
+#include "src/expression/ops_utils.h"
+#include "src/expression/param.h"
+#include "include/api/cfg.h"
+#include "src/expression/sequential.h"
+
+namespace mindspore {
+namespace lite {
+void InputM::SetUp(const std::vector<int> &dims, TypeId data_type, int fmt) {
+  expr()->SetSize(C0NUM);
+  expr()->SetDims(dims);
+  expr()->set_data_type(data_type);
+  expr()->set_format(fmt);
+  set_primitive(schema::PrimitiveType_NONE);
+}
+
+InputM::InputM(const std::vector<int> &dims, TypeId data_type, int fmt) : Node() { SetUp(dims, data_type, fmt); }
+
+InputM::InputM(const schema::Tensor *tensor) : Node() {
+  std::vector<int> dims(tensor->dims()->size());
+  (void)std::transform(tensor->dims()->begin(), tensor->dims()->end(), dims.begin(), [](int32_t x) { return x; });
+  SetUp(dims, static_cast<TypeId>(tensor->dataType()), tensor->format());
+  if (tensor->name()) set_name(tensor->name()->str());
+  if (tensor->data() != nullptr) data_.Copy(tensor->data()->data(), tensor->data()->size());
+}
+
+namespace NN {
+Node *Input(const std::vector<int> &dims, TypeId data_type, int fmt) {
+  auto i = new (std::nothrow) InputM(dims, data_type, fmt);
+  if (i == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate input expression ";
+    return nullptr;
+  }
+  return i;
+}
+
+Net *Sequential() {
+  auto s = new (std::nothrow) mindspore::lite::Sequential();
+  if (s == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate sequential expression";
+    return nullptr;
+  }
+  return s;
+}
+};  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops.h

@@ -0,0 +1,69 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_H_
+
+#include <vector>
+#include <string>
+#include <set>
+#include "include/api/net.h"
+#include "src/expression/cfg.h"
+#include "src/expression/net.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class InputM : public Node {
+ public:
+  explicit InputM(const schema::Tensor *tensor);
+  explicit InputM(const std::vector<int> &dims, TypeId data_type = kNumberTypeFloat32, int fmt = NHWC);
+  Param *data() override { return &data_; }
+
+ private:
+  void SetUp(const std::vector<int> &dims, TypeId data_type, int fmt);
+  Param data_;
+};
+namespace NN {
+Node *Conv2D(const ConvConfig &cfg);
+Node *Relu();
+Node *Dense(const DenseConfig &cfg);
+Node *Flatten();
+Node *Input(const std::vector<int> &dims, TypeId data_type = kNumberTypeFloat32, int fmt = NHWC);
+Node *Add();
+Node *Sub();
+Node *Div();
+Node *Mul();
+Node *Neg();
+Node *SoftmaxCrossEntropy();
+Net *Sequential();
+Node *Adam(std::set<Node *> &&learn, const AdamConfig &cfg);
+
+Node *Softmax(int axis = -1);
+Node *BatchNorm2D(int outp, float momentum = 0.1, float epsilon = 1e-5f);
+Node *Sigmoid();
+Node *DropOut(float ration = 0.5);
+Node *ReLU6();
+Node *Reshape(const std::vector<int> &shape);
+Node *ReduceMean(bool keep_dims, const std::vector<int> &dims);
+Node *ReduceSum(bool keep_dims, const std::vector<int> &dims);
+Node *Tile(const std::vector<int> &multiples);
+Node *MaxPool2D(const PoolingConfig &cfg);
+Node *AvgPool2D(const PoolingConfig &cfg);
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_H_

mindspore/lite/src/expression/ops/activation.cc

@@ -0,0 +1,133 @@
+/**
+ * Copyright 2022 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/expression/ops/activation.h"
+#include "nnacl/fp32/activation_fp32.h"
+#include "src/expression/import.h"
+#include "src/expression/ops.h"
+#include "src/cxx_api/expression/node_impl.h"
+
+namespace mindspore {
+namespace lite {
+ActM::ActM(schema::ActivationType type) : Node() {
+  auto op_param = malloc(sizeof(ActivationParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate ActivationParameter";
+    return;
+  }
+  SetOpParam(op_param);
+  set_primitive(schema::PrimitiveType_Activation);
+  ActivationParameter *act_param = reinterpret_cast<ActivationParameter *>(opParam_.get());
+  act_param->type_ = type;
+  act_param->alpha_ = 0.f;
+  act_param->min_val_ = 0.f;
+  act_param->max_val_ = 0.f;
+}
+
+std::vector<EXPR *> ActM::Grad(EXPR *yt) {
+  auto actGrad = new (std::nothrow) ActGradM(this);
+  if (actGrad == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate activation grad node";
+    return {};
+  }
+  PushOp(actGrad);
+  auto param = reinterpret_cast<ActivationParameter *>(actGrad->OpParam());
+  EXPR *ag = nullptr;
+  actGrad->expr()->SetSize(C2NUM);
+  if ((param->type_ == schema::ActivationType_SIGMOID) || (param->type_ == schema::ActivationType_TANH)) {
+    ag = (*actGrad)({output(0), yt}).front();
+  } else if ((param->type_ == schema::ActivationType_HSWISH) || (param->type_ == schema::ActivationType_HSIGMOID) ||
+             (param->type_ == schema::ActivationType_RELU6)) {
+    ag = (*actGrad)({yt, input(0)}).front();
+  } else if (param->type_ == schema::ActivationType_GELU) {
+    actGrad->expr()->SetSize(C3NUM);
+    ag = (*actGrad)({yt, input(0), output(0)}).front();
+  } else {
+    ag = (*actGrad)({yt, output(0)}).front();
+  }
+  std::vector<EXPR *> res = {ag};
+  return res;
+}
+
+int ActM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto act_param = reinterpret_cast<const ActivationParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::ActivationT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate activation primitive";
+    return RET_ERROR;
+  }
+  prim->activation_type = static_cast<decltype(prim->activation_type)>(act_param->type_);
+  prim->alpha = act_param->alpha_;
+  prim->min_val = act_param->min_val_;
+  prim->max_val = act_param->max_val_;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg reg(schema::PrimitiveType_Activation, ReturnNode<ActM>);
+
+ActGradM::ActGradM(Node *node) {
+  CloneOpParam<ActivationParameter>(node->OpParam());
+  set_primitive(schema::PrimitiveType_ActivationGrad);
+  set_name(node->name() + "/" + kGradName + "/actGrad");
+}
+
+std::vector<EXPR *> ActGradM::Grad(EXPR *yt) { return {}; }
+
+int ActGradM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto act_param = reinterpret_cast<const ActivationParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::ActivationGradT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate activation grad primitive";
+    return RET_ERROR;
+  }
+  prim->activation_type = static_cast<decltype(prim->activation_type)>(act_param->type_);
+  prim->alpha = act_param->alpha_;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg regGrad(schema::PrimitiveType_ActivationGrad, ReturnNode<ActGradM>);
+namespace NN {
+Node *ReLU6() {
+  auto r = new (std::nothrow) ActM(schema::ActivationType_RELU6);
+  if (r == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate relu6";
+    return nullptr;
+  }
+  r->set_name(Node::UniqueName("ReLU6"));
+  return r;
+}
+Node *Sigmoid() {
+  auto s = new (std::nothrow) ActM(schema::ActivationType_SIGMOID);
+  if (s == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate sigmoid";
+    return nullptr;
+  }
+  s->set_name(Node::UniqueName("Sigmoid"));
+  return s;
+}
+Node *Relu() {
+  auto r = new (std::nothrow) ActM(schema::ActivationType_RELU);
+  if (r == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate relu";
+    return nullptr;
+  }
+  r->set_name(r->UniqueName("Relu"));
+  return r;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/activation.h

@@ -0,0 +1,44 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_ACTIVATION_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_ACTIVATION_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/net.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class ActM : public Node {
+ public:
+  ActM() = default;
+  explicit ActM(schema::ActivationType type);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+class ActGradM : public Node {
+ public:
+  ActGradM() : Node() {}         // for Import
+  explicit ActGradM(Node *act);  // for Grad
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_ACTIVATION_H_

mindspore/lite/src/expression/ops/adam.cc

@@ -0,0 +1,142 @@
+/**
+ * Copyright 2022 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/expression/ops/adam.h"
+#include <memory>
+#include <set>
+#include <utility>
+#include "src/expression/ops.h"
+#include "src/expression/ops/assign.h"
+#include "src/expression/ops/arithmetic.h"
+#include "nnacl/fp32_grad/optimizer.h"
+#include "include/api/net.h"
+#include "src/cxx_api/expression/node_impl.h"
+
+namespace mindspore {
+namespace NN {
+Node *Adam(std::shared_ptr<NodeSet> learn, const AdamConfig &cfg) {
+  auto lite_node = lite::NN::Adam(std::move(learn->set_), cfg);
+  return NodeImpl::Connect(lite_node);
+}
+}  // namespace NN
+
+namespace lite {
+std::vector<EXPR *> AdamM::Clone(EXPR *grad, EXPR *weight) {
+  auto adam = new (std::nothrow) AdamM();
+  if (adam == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate adam";
+    return {};
+  }
+  adam->set_name("optimizer-Adam");
+  adam->CloneOpParam<AdamParameter>(OpParam());
+  adam->update_name(weight->node()->name());
+  adam->set_primitive(primitive());
+  adam->expr()->SetSize(C10NUM);
+  // setup weight and momentum
+  adam->expr()->set_params(C0NUM, weight);
+  auto dims = grad->dims();
+  auto m = adam->CreateWeights(dims, kNumberTypeFloat32, KHWC, Param::ZEROS, "m");
+  adam->expr()->set_params(C1NUM, m);
+  auto v = adam->CreateWeights(dims, kNumberTypeFloat32, KHWC, Param::ZEROS, "v");
+  adam->expr()->set_params(C2NUM, v);
+  // copy parameters
+  for (int i = C3NUM; i < C9NUM; i++) {
+    adam->expr()->set_params(i, this->input(i));
+  }
+  adam->expr()->set_params(C9NUM, grad);
+  return (*adam)(adam->inputs());
+}
+
+AdamM::AdamM(std::set<Node *> &&learn, const AdamConfig &cfg) {
+  auto op_param = reinterpret_cast<AdamParameter *>(malloc(sizeof(AdamParameter)));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate ActivationParameter";
+    return;
+  }
+  AssignLearn(std::move(learn));
+  memset(op_param, 0, sizeof(AdamParameter));
+  op_param->use_nesterov_ = cfg.use_nesterov_;
+  SetOpParam(op_param);
+  set_primitive(schema::PrimitiveType_Adam);
+  set_name("optimizer-Adam");
+  // Adam Network
+  expr()->SetSize(C10NUM);
+  auto assign1 = new (std::nothrow) AssignM(0);
+  if (assign1 == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate assign";
+    return;
+  }
+  PushOp(assign1);
+  auto assign2 = new (std::nothrow) AssignM(0);
+  if (assign2 == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate assign";
+    return;
+  }
+  PushOp(assign2);
+  auto mul1 = NN::Mul();
+  if (mul1 == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate mul";
+    return;
+  }
+  PushOp(mul1);
+  auto mul2 = NN::Mul();
+  if (mul2 == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate mul";
+    return;
+  }
+  PushOp(mul2);
+  auto tmp = 1.0f;
+  mul1->CreateConstTensor(C0NUM, {1}, kNumberTypeFloat32, KHWC, "beta1-power", &tmp);
+  mul1->CreateConstTensor(C1NUM, {1}, kNumberTypeFloat32, KHWC, "beta1-data", &cfg.beta1_);
+  auto o1 = (*mul1)({});
+  assign1_ = (*assign1)({mul1->input(0), o1.front()}).front();
+  mul2->CreateConstTensor(C0NUM, {1}, kNumberTypeFloat32, KHWC, "beta2-power", &tmp);
+  mul2->CreateConstTensor(C1NUM, {1}, kNumberTypeFloat32, KHWC, "beta2-data", &cfg.beta2_);
+  auto o2 = (*mul2)({});
+  assign2_ = (*assign2)({mul2->input(0), o2.front()}).front();
+  expr()->set_params(C3NUM, o1.front());
+  expr()->set_params(C4NUM, o2.front());
+  CreateConstTensor(C5NUM, {1}, kNumberTypeFloat32, KHWC, "learning-rate", &cfg.learning_rate_);
+  CreateConstTensor(C6NUM, {1}, kNumberTypeFloat32, KHWC, "beta1", &cfg.beta1_);
+  CreateConstTensor(C7NUM, {1}, kNumberTypeFloat32, KHWC, "beta2", &cfg.beta2_);
+  CreateConstTensor(C8NUM, {1}, kNumberTypeFloat32, KHWC, "epsilon", &cfg.eps_);
+}
+
+int AdamM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto param = reinterpret_cast<const AdamParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::AdamT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate " << cnode->name;
+    return RET_ERROR;
+  }
+  prim->use_nesterov = param->use_nesterov_;
+  prim->use_locking = false;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+namespace NN {
+Node *Adam(std::set<Node *> &&learn, const AdamConfig &cfg) {
+  auto a = new (std::nothrow) AdamM(std::move(learn), cfg);
+  if (a == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate adam";
+    return nullptr;
+  }
+  return a;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/adam.h

@@ -0,0 +1,48 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_ADAM_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_ADAM_H_
+
+#include <vector>
+#include <set>
+#include <memory>
+#include "include/api/net.h"
+#include "src/expression/net.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class AdamM : public Node {
+ public:
+  AdamM() = default;
+  AdamM(std::set<Node *> &&learn, const AdamConfig &cfg);
+  std::vector<EXPR *> Clone(EXPR *grad, EXPR *weight) override;
+  void AddNetOutput(std::vector<EXPR *> *output) override {
+    output->push_back(assign1_);
+    output->push_back(assign2_);
+  }
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+
+ private:
+  float weight_decay_;
+  float loss_scale_;
+  EXPR *assign1_{nullptr};
+  EXPR *assign2_{nullptr};
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_ADAM_H_

mindspore/lite/src/expression/ops/addn.cc

@@ -0,0 +1,42 @@
+/**
+ * Copyright 2022 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/expression/ops/addn.h"
+
+namespace mindspore {
+namespace lite {
+AddN::AddN(int dummy) {
+  auto op_param = calloc(1, sizeof(OpParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate parameter ";
+    return;
+  }
+  set_name(UniqueName("addN"));
+  SetOpParam(op_param);
+  set_primitive(schema::PrimitiveType_AddN);
+}
+
+int AddN::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::AddNT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/addn.h

@@ -0,0 +1,34 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_ADDN_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_ADDN_H_
+
+#include <memory>
+#include "src/expression/node.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class AddN : public Node {
+ public:
+  explicit AddN(int dummy);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_ADDN_H_

mindspore/lite/src/expression/ops/arithmetic.cc

@@ -0,0 +1,223 @@
+/**
+ * Copyright 2022 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/expression/ops/arithmetic.h"
+#include <memory>
+#include "src/expression/ops/reduce.h"
+#include "src/expression/ops/reshape.h"
+#include "src/expression/ops_utils.h"
+#include "src/expression/ops/arithmetic_self.h"
+#include "src/expression/ops.h"
+#include "nnacl/arithmetic.h"
+#include "src/expression/import.h"
+#include "src/cxx_api/expression/node_impl.h"
+
+namespace mindspore {
+namespace lite {
+// Common Arithmetic Functionality
+ArithmeticM::ArithmeticM(schema::PrimitiveType type) : Node() {
+  auto op_param = malloc(sizeof(ArithmeticParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate ActivationParameter";
+    return;
+  }
+  SetOpParam(op_param);
+  expr()->SetSize(C2NUM);
+  set_primitive(type);
+}
+
+std::vector<EXPR *> ArithmeticM::binop_grad_common(EXPR *x, EXPR *y, EXPR *dx, EXPR *dy) {
+  auto shape_of_x = x->dims();
+  auto shape_of_y = y->dims();
+  auto reduce_dx = dx;
+  auto reduce_dy = dy;
+  auto rx = (BroadcastGradientArgs(shape_of_x, shape_of_y))();
+  if (rx[0].size()) {
+    auto reduce_sum = NN::ReduceSum(false, rx[0]);
+    PushOp(reduce_sum);
+    reduce_dx = (*reduce_sum)({reduce_dx}).front();
+    auto reshape = NN::Reshape(shape_of_x);
+    PushOp(reshape);
+    reduce_dx = (*reshape)({reduce_dx}).front();
+  }
+  if (rx[1].size()) {
+    auto reduce_sum = NN::ReduceSum(false, rx[1]);
+    PushOp(reduce_sum);
+    reduce_dy = (*reduce_sum)({reduce_dy}).front();
+    auto reshape = NN::Reshape(shape_of_y);
+    PushOp(reshape);
+    reduce_dy = (*reshape)({reduce_dy}).front();
+  }
+  std::vector<EXPR *> out = {reduce_dx, reduce_dy};
+  return out;
+}
+
+// Add Op
+AddM::AddM(int dummy) : ArithmeticM(schema::PrimitiveType_AddFusion) { set_name(UniqueName("Add")); }
+
+std::vector<EXPR *> AddM::Grad(EXPR *yt) { return binop_grad_common(input(0), input(1), yt, yt); }
+
+int AddM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::AddFusionT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate prim";
+    return RET_ERROR;
+  }
+  prim->activation_type = schema::ActivationType_NO_ACTIVATION;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg AddReg(schema::PrimitiveType_AddFusion, ReturnNode<AddM>);
+
+// Div op
+DivM::DivM(int dummy) : ArithmeticM(schema::PrimitiveType_RealDiv) { set_name(UniqueName("RealDiv")); }
+std::vector<EXPR *> DivM::Grad(EXPR *yt) {
+  auto x = input(0);
+  auto y = input(1);
+  auto o = output(0);
+  auto div_op = NN::Div();
+  if (div_op == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate div_op";
+    return {};
+  }
+  PushOp(div_op);
+  auto neg_op = NN::Neg();
+  if (neg_op == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate neg_op";
+    return {};
+  }
+  PushOp(neg_op);
+  auto mul_op = NN::Mul();
+  if (mul_op == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate mul_op";
+    return {};
+  }
+  PushOp(mul_op);
+  auto bc_x = (*div_op)({yt, y}).front();
+  auto bc_y = (*neg_op)((*mul_op)({bc_x, o})).front();
+  return binop_grad_common(x, y, bc_x, bc_y);
+}
+int DivM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::RealDivT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+static ImportReg DivReg(schema::PrimitiveType_DivFusion, ReturnNode<DivM>);
+
+// Mul op
+MulM::MulM(int dummy) : ArithmeticM(schema::PrimitiveType_MulFusion) { set_name(UniqueName("Mul")); }
+
+std::vector<EXPR *> MulM::Grad(EXPR *yt) {
+  auto mul_dx = NN::Mul();
+  if (mul_dx == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate mul dx";
+    return {};
+  }
+  PushOp(mul_dx);
+  auto mul_dy = NN::Mul();
+  if (mul_dy == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate mul_dy";
+    return {};
+  }
+  PushOp(mul_dy);
+  auto x = input(0);
+  auto y = input(1);
+  auto bc_dx = (*mul_dx)({y, yt}).front();
+  auto bc_dy = (*mul_dy)({x, yt}).front();
+  return binop_grad_common(x, y, bc_dx, bc_dy);
+}
+
+int MulM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::MulFusionT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate prim";
+    return RET_ERROR;
+  }
+  prim->activation_type = schema::ActivationType_NO_ACTIVATION;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+static ImportReg MulReg(schema::PrimitiveType_MulFusion, ReturnNode<MulM>);
+
+// Sub op
+SubM::SubM(int dummy) : ArithmeticM(schema::PrimitiveType_SubFusion) { set_name(UniqueName("Sub")); }
+
+std::vector<EXPR *> SubM::Grad(EXPR *yt) {
+  auto x = input(0);
+  auto y = input(1);
+  auto neg = NN::Neg();
+  if (neg == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate neg";
+    return {};
+  }
+  PushOp(neg);
+  auto neg_grad = (*neg)({yt}).front();
+  return binop_grad_common(x, y, yt, neg_grad);
+}
+int SubM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::SubFusionT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate prim";
+    return RET_ERROR;
+  }
+  prim->activation_type = schema::ActivationType_NO_ACTIVATION;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+static ImportReg SubReg(schema::PrimitiveType_SubFusion, ReturnNode<SubM>);
+
+namespace NN {
+Node *Add() {
+  auto a = new (std::nothrow) AddM(0);
+  if (a == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate a";
+    return nullptr;
+  }
+  return a;
+}
+Node *Sub() {
+  auto a = new (std::nothrow) SubM(0);
+  if (a == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate a";
+    return nullptr;
+  }
+  return a;
+}
+
+Node *Mul() {
+  auto a = new (std::nothrow) MulM(0);
+  if (a == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate a";
+    return nullptr;
+  }
+  return a;
+}
+Node *Div() {
+  auto a = new (std::nothrow) DivM(0);
+  if (a == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate a";
+    return nullptr;
+  }
+  return a;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/arithmetic.h

@@ -0,0 +1,75 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_ARITHMETIC_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_ARITHMETIC_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/net.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class ArithmeticM : public Node {
+ public:
+  ArithmeticM() = default;
+  explicit ArithmeticM(schema::PrimitiveType type);
+
+ protected:
+  std::vector<EXPR *> binop_grad_common(EXPR *x, EXPR *y, EXPR *dx, EXPR *dy);
+};
+
+class AddM : public ArithmeticM {
+ public:
+  AddM() = default;
+  explicit AddM(int dummy);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+class DivM : public ArithmeticM {
+ public:
+  DivM() = default;
+  explicit DivM(int dummy);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+class MulM : public ArithmeticM {
+ public:
+  MulM() = default;
+  explicit MulM(int dummy);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+class SubM : public ArithmeticM {
+ public:
+  SubM() = default;
+  explicit SubM(int dummy);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+namespace NN {
+Node *Add();
+Node *Sub();
+Node *Mul();
+Node *Div();
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_ARITHMETIC_H_

mindspore/lite/src/expression/ops/arithmetic_self.cc

@@ -0,0 +1,72 @@
+/**
+ * Copyright 2022 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/expression/ops/arithmetic_self.h"
+#include <memory>
+#include "src/expression/ops_utils.h"
+#include "src/expression/ops.h"
+#include "nnacl/arithmetic_self_parameter.h"
+#include "src/expression/import.h"
+
+namespace mindspore {
+namespace lite {
+// Common Arithmetic Self Functionality
+ArithmeticSelfM::ArithmeticSelfM(schema::PrimitiveType type) : Node() {
+  auto op_param = malloc(sizeof(ArithmeticSelfParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate ArithmeticSelfParameter";
+    return;
+  }
+  memset(op_param, 0, sizeof(ArithmeticSelfParameter));
+  SetOpParam(op_param);
+  expr()->SetSize(C1NUM);
+  set_primitive(type);
+}
+
+// NEG OP
+NegM::NegM(int dummy) : ArithmeticSelfM(schema::PrimitiveType_NegGrad) { set_name(UniqueName("Neg")); }
+
+std::vector<EXPR *> NegM::Grad(EXPR *yt) {
+  auto grad_neg = new (std::nothrow) NegM(0);
+  if (grad_neg == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate neg gradient";
+    return {};
+  }
+  return (*grad_neg)({yt});
+}
+
+int NegM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::NegT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+namespace NN {
+Node *Neg() {
+  auto a = new (std::nothrow) NegM(0);
+  if (a == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate neg node";
+    return nullptr;
+  }
+  return a;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/arithmetic_self.h

@@ -0,0 +1,46 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_ARITHMETIC_SELF_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_ARITHMETIC_SELF_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/net.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class ArithmeticSelfM : public Node {
+ public:
+  explicit ArithmeticSelfM(schema::PrimitiveType type);
+
+ protected:
+};
+
+class NegM : public ArithmeticSelfM {
+ public:
+  explicit NegM(int dummy);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+namespace NN {
+Node *Neg();
+}
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_ARITHMETIC_SELF_H_

mindspore/lite/src/expression/ops/assign.cc

@@ -0,0 +1,60 @@
+/**
+ * Copyright 2022 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/expression/ops/assign.h"
+#include <memory>
+#include "nnacl/reshape_parameter.h"
+#include "src/expression/import.h"
+
+namespace mindspore {
+namespace lite {
+AssignM::AssignM(int dummy) {
+  auto op_param = calloc(1, sizeof(OpParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate ReshapeParameter";
+    return;
+  }
+  expr()->SetSize(C2NUM);
+  SetOpParam(op_param);
+  set_primitive(schema::PrimitiveType_Assign);
+  set_name(UniqueName("Assign"));
+}
+
+int AssignM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::AssignT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg reg(schema::PrimitiveType_Reshape, ReturnNode<AssignM>);
+
+namespace NN {
+Node *Assign() {
+  auto node = new (std::nothrow) AssignM(0);
+  if (node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate node";
+    return nullptr;
+  }
+  node->set_name(Node::UniqueName("Assign"));
+  return node;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/assign.h

@@ -0,0 +1,35 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_ASSIGN_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_ASSIGN_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/net.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class AssignM : public Node {
+ public:
+  AssignM() = default;
+  explicit AssignM(int dummy);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_ASSIGN_H_

mindspore/lite/src/expression/ops/batchnorm.cc

@@ -0,0 +1,135 @@
+/**
+ * Copyright 2022 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/expression/ops/batchnorm.h"
+#include <memory>
+#include "nnacl/batchnorm_parameter.h"
+#include "nnacl/fp32_grad/batch_norm.h"
+#include "src/expression/import.h"
+#include "src/expression/ops.h"
+#include "src/cxx_api/expression/node_impl.h"
+
+namespace mindspore {
+namespace lite {
+BatchNorm2dM::BatchNorm2dM(int outp, float momentum, float epsilon) {
+  constexpr int bn_inputs = 5;
+  constexpr int bn_outputs = 5;
+
+  auto op_param = calloc(1, sizeof(BatchNormParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate BatchNormParameter";
+    return;
+  }
+  expr()->SetSize(bn_inputs);
+  set_name(UniqueName("BatchNorm2D"));
+  auto bn_param = reinterpret_cast<BatchNormParameter *>(op_param);
+  bn_param->channel_ = outp;
+  bn_param->momentum_ = momentum;
+  bn_param->epsilon_ = epsilon;
+  SetOpParam(op_param);
+  set_primitive(schema::PrimitiveType_FusedBatchNorm);
+  std::vector<int> dims = {outp};
+  auto scale = Node::CreateWeights(dims, kNumberTypeFloat32, KHWC, Param::Mode::ONES, "scale");
+  expr()->set_params(C1NUM, scale);
+  auto offset = Node::CreateWeights(dims, kNumberTypeFloat32, KHWC, Param::Mode::ZEROS, "offset");
+  expr()->set_params(C2NUM, offset);
+  auto mean = Node::CreateWeights(dims, kNumberTypeFloat32, KHWC, Param::Mode::ZEROS, "mean");
+  expr()->set_params(C3NUM, mean);
+  auto var = Node::CreateWeights(dims, kNumberTypeFloat32, KHWC, Param::Mode::ONES, "var");
+  expr()->set_params(C4NUM, var);
+  SetOutputs(bn_outputs);
+  SetLearn();
+}
+
+BatchNorm2dGradM::BatchNorm2dGradM(BatchNorm2dM *bn_node) : Node() {
+  auto op_param = calloc(1, sizeof(BNGradParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate BNGradParameter";
+    return;
+  }
+  expr()->SetSize(C6NUM);
+  set_name(bn_node->name() + "/" + kGradName + "/bnGrad");
+  auto bn_grad_param = reinterpret_cast<BNGradParameter *>(op_param);
+  auto bn_param = reinterpret_cast<BatchNormParameter *>(bn_node->OpParam());
+  bn_param->is_training_ = true;
+  bn_grad_param->epsilon_ = bn_param->epsilon_;
+  bn_grad_param->is_training_ = true;
+  SetOpParam(op_param);
+  set_primitive(schema::PrimitiveType_BatchNormGrad);
+  EXPR e(this);
+  e.SetSize(0);
+  // Dgamma
+  expr_.emplace_back(e);
+  // Doffset
+  expr_.emplace_back(e);
+}
+
+std::vector<EXPR *> BatchNorm2dM::Grad(EXPR *yt) {
+  auto bn_grad_node = new (std::nothrow) BatchNorm2dGradM(this);
+  if (bn_grad_node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate batchnorm grad";
+    return {};
+  }
+  PushOp(bn_grad_node);
+  auto bn_grad = (*bn_grad_node)({yt, input(0), output(1), output(3), output(4), output(2)});
+  return bn_grad;
+}
+
+int BatchNorm2dM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto bn_param = reinterpret_cast<const BatchNormParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::FusedBatchNormT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->epsilon = bn_param->epsilon_;
+  prim->momentum = bn_param->momentum_;
+  prim->mode = (bn_param->is_training_ == false) ? 0 : 1;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+void BatchNorm2dM::SetLearn() {
+  AddLearn(input(C1NUM)->node());
+  AddLearn(input(C2NUM)->node());
+}
+
+int BatchNorm2dGradM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto param = reinterpret_cast<const BNGradParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::BatchNormGradT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->epsilon = param->epsilon_;
+  prim->is_training = param->is_training_;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg reg(schema::PrimitiveType_FusedBatchNorm, ReturnNode<BatchNorm2dM>);
+namespace NN {
+Node *BatchNorm2D(int outp, float momentum, float epsilon) {
+  auto node = new (std::nothrow) BatchNorm2dM(outp, momentum, epsilon);
+  if (node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate node";
+    return nullptr;
+  }
+  return node;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/batchnorm.h

@@ -0,0 +1,43 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_BATCHNORM_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_BATCHNORM_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/net.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class BatchNorm2dM : public Node {
+ public:
+  BatchNorm2dM() = default;
+  BatchNorm2dM(int outp, float momentum, float epsilon);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+  void SetLearn() override;
+};
+
+class BatchNorm2dGradM : public Node {
+ public:
+  explicit BatchNorm2dGradM(BatchNorm2dM *bn_node);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_BATCHNORM_H_

mindspore/lite/src/expression/ops/biasadd.cc

@@ -0,0 +1,93 @@
+/**
+ * Copyright 2022 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/expression/ops/biasadd.h"
+#include "src/expression/ops/transpose.h"
+#include "nnacl/arithmetic.h"
+#include "src/expression/import.h"
+
+namespace mindspore {
+namespace lite {
+BiasAddM::BiasAddM(Format data_format) {
+  auto op_param = calloc(1, sizeof(ArithmeticParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate ConvParameter";
+    return;
+  }
+  auto bias_param = reinterpret_cast<ArithmeticParameter *>(op_param);
+  SetOpParam(bias_param);
+  set_primitive(schema::PrimitiveType_BiasAdd);
+}
+
+std::vector<EXPR *> BiasAddM::construct(const std::vector<EXPR *> &inputs) {
+  auto x = Node::construct(inputs);
+  AddLearn(inputs.at(C1NUM)->node());
+  return x;
+}
+
+void BiasAddM::SetLearn() { AddLearn(input(C1NUM)->node()); }
+
+std::vector<EXPR *> BiasAddM::Grad(EXPR *yt) {
+  auto in = yt;
+  if (yt->format() != NHWC && yt->dims().size() == C4NUM) {
+    in = TransposeM::TransposeCHW2HWC(yt);
+    in->node()->set_name(kGradName + "/" + name() + "/" + in->node()->name());
+    PushOp(in->node());
+  }
+  auto grad_node = new (std::nothrow) BiasAddGradM(*this);
+  if (grad_node == nullptr) {
+    MS_LOG(ERROR) << "Cannon allocate Bias Grad";
+    return {};
+  }
+  PushOp(grad_node);
+  auto bias_grad = (*grad_node)({in});
+  return {in, bias_grad.front()};
+}
+int BiasAddM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::BiasAddT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "cannot allocate prim";
+    return RET_ERROR;
+  }
+  prim->format = static_cast<schema::Format>(KHWC);
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+BiasAddGradM::BiasAddGradM(const BiasAddM &bias) {
+  auto op_param = calloc(1, sizeof(OpParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate op_param";
+  }
+  SetOpParam(op_param);
+  set_primitive(schema::PrimitiveType_BiasAddGrad);
+}
+
+int BiasAddGradM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::BiasAddGradT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg reg(schema::PrimitiveType_BiasAdd, ReturnNode<BiasAddM>);
+
+namespace NN {}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/biasadd.h

@@ -0,0 +1,44 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_BIASADD_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_BIASADD_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/net.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class BiasAddM : public Node {
+ public:
+  BiasAddM() = default;
+  explicit BiasAddM(Format data_format);
+  std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs) override;
+  std::vector<EXPR *> Grad(EXPR *yt) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+  void SetLearn() override;
+};
+
+class BiasAddGradM : public Node {
+ public:
+  explicit BiasAddGradM(const BiasAddM &bias);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_BIASADD_H_

mindspore/lite/src/expression/ops/conv.cc

@@ -0,0 +1,241 @@
+/**
+ * Copyright 2022 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/expression/ops/conv.h"
+#include <memory>
+#include "src/expression/ops/biasadd.h"
+#include "src/expression/ops/depend.h"
+#include "src/expression/ops/transpose.h"
+#include "nnacl/conv_parameter.h"
+#include "inner/model_generated.h"
+#include "src/expression/import.h"
+#include "src/expression/ops.h"
+#include "src/cxx_api/expression/node_impl.h"
+
+namespace mindspore {
+namespace lite {
+ConvM::ConvM(const ConvConfig &cfg) : Node() {
+  auto op_param = calloc(1, sizeof(ConvParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate ConvParameter";
+    return;
+  }
+  SetOpParam(op_param);
+  ConvParameter *conv_param = reinterpret_cast<ConvParameter *>(OpParam());
+  conv_param->input_channel_ = cfg.in_channel_;
+  conv_param->output_channel_ = cfg.out_channel_;
+  conv_param->kernel_h_ = cfg.kernel_size_[0];
+  conv_param->kernel_w_ = cfg.kernel_size_[1];
+  conv_param->stride_h_ = cfg.stride_[0];
+  conv_param->stride_w_ = cfg.stride_[1];
+  auto pad_mode = GetMode(cfg.pad_mode_);
+  if (pad_mode == -1) {
+    MS_LOG(ERROR) << "bad pad mode";
+    return;
+  }
+  conv_param->pad_mode_ = static_cast<PadMode>(pad_mode);
+  conv_param->pad_u_ = cfg.padding_[C0NUM];
+  conv_param->pad_d_ = cfg.padding_[C1NUM];
+  conv_param->pad_l_ = cfg.padding_[C2NUM];
+  conv_param->pad_r_ = cfg.padding_[C3NUM];
+  conv_param->dilation_h_ = cfg.dilation_[C0NUM];
+  conv_param->dilation_w_ = cfg.dilation_[C1NUM];
+  conv_param->group_ = cfg.group_;
+  conv_param->out_format_ = NHWC;
+  conv_param->act_type_ = ActType_No;
+  expr()->SetSize(C2NUM);
+  set_primitive(schema::PrimitiveType_Conv2DFusion);
+  set_name(UniqueName("Conv"));
+  Param::Mode mode = Param::String2Enum(cfg.weight_init_);
+  std::vector<int> dims = {conv_param->output_channel_, conv_param->kernel_h_, conv_param->kernel_w_,
+                           conv_param->input_channel_ / conv_param->group_};
+  auto w = CreateWeights(dims, kNumberTypeFloat32, KHWC, mode, "weights");
+  expr()->set_params(C1NUM, w);
+  if (cfg.has_bias) {
+    bias_ = new (std::nothrow) BiasAddM(KHWC);
+    if (bias_ == nullptr) {
+      MS_LOG(ERROR) << "Cannot allocate bias";
+      return;
+    }
+    bias_->update_name(name());
+    std::vector<int> dim_bias = {conv_param->output_channel_};
+    wbias_ = CreateWeights(dim_bias, kNumberTypeFloat32, KHWC, Param::Mode::ZEROS, "weights");
+    AddLearn(wbias_->node());
+    PushOp(bias_);
+  }
+  SetLearn();
+}
+
+std::vector<EXPR *> ConvM::construct(const std::vector<EXPR *> &inputs) {
+  auto in = inputs;
+  auto x = in.front();
+  if (x->format() != NHWC && x->dims().size() == C4NUM) {
+    x = TransposeM::TransposeCHW2HWC(x);
+    x->node()->set_name(name() + "/" + x->node()->name());
+    PushOp(x->node());
+    in.at(0) = x;
+  }
+  auto y = Node::construct(in);
+  if (bias_ != nullptr) {
+    y = (*bias_)({y.front(), wbias_});
+  }
+  return y;
+}
+
+void ConvM::SetLearn() { AddLearn(input(C1NUM)->node()); }
+
+int ConvM::GetMode(std::string mode) {
+  const std::vector<std::string> list = {"pad", "same", "valid"};
+  auto itr = std::find(list.begin(), list.end(), mode);
+  if (itr == list.end()) {
+    MS_LOG(ERROR) << "illegal mode" << mode;
+    return -1;
+  }
+  return std::distance(list.begin(), itr);
+}
+
+std::vector<EXPR *> ConvM::Grad(EXPR *yt) {
+  // Generate Input Grad
+  EXPR *in = yt;
+  if (yt->format() != NHWC && yt->dims().size() == C4NUM) {
+    in = TransposeM::TransposeCHW2HWC(yt);
+    in->node()->set_name(kGradName + "/" + name() + "/" + in->node()->name());
+    PushOp(in->node());
+  }
+  auto inGrad = new (std::nothrow) ConvInputGradM(this);
+  if (inGrad == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate convolution input grad";
+    return {};
+  }
+  PushOp(inGrad);
+  auto ig = (*inGrad)({in, input(1), inGrad->input(2)});
+  // Execution Control Flow !
+  auto depend = NN::Depend();
+  if (depend == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate depend";
+    return {};
+  }
+  PushOp(depend);
+  depend->update_name(name());
+  auto de = (*depend)({inGrad->expr()});
+  // Generate Filter Grad
+  auto filterGrad = new (std::nothrow) ConvFilterGradM(this);
+  if (filterGrad == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate convolution filter grad";
+    return {};
+  }
+  PushOp(filterGrad);
+  filterGrad->update_name(name());
+  auto fg = (*filterGrad)({in, input(0), filterGrad->input(2), de[0]});
+  std::vector<EXPR *> res = {ig[0], fg[0]};
+  return res;
+}
+
+int ConvM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto conv_param = reinterpret_cast<const ConvParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::Conv2DFusionT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->activation_type = static_cast<schema::ActivationType>(conv_param->act_type_);
+  prim->format = static_cast<schema::Format>(conv_param->out_format_);
+  prim->stride = {conv_param->stride_h_, conv_param->stride_w_};
+  prim->kernel_size = {conv_param->kernel_h_, conv_param->kernel_w_};
+  prim->dilation = {conv_param->dilation_h_, conv_param->dilation_w_};
+  prim->out_channel = conv_param->output_channel_;
+  prim->in_channel = conv_param->input_channel_;
+  prim->group = conv_param->group_;
+  prim->pad_mode = static_cast<schema::PadMode>(conv_param->pad_mode_);
+  prim->pad_list = {conv_param->pad_u_, conv_param->pad_d_, conv_param->pad_l_, conv_param->pad_r_};
+  prim->mode = 1;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+ConvInputGradM::ConvInputGradM(ConvM *conv_node) : Node() {
+  CloneOpParam<ConvParameter>(conv_node->OpParam());
+  set_primitive(schema::PrimitiveType_Conv2DBackpropInputFusion);
+  set_name(kGradName + "/conv2DBackpropInput");
+  expr()->SetSize(C3NUM);
+  auto const x = conv_node->input(0);
+  CreateConstTensor(C2NUM, {static_cast<int32_t>(x->dims().size())}, kNumberTypeInt32, KHWC, "shape", x->dims().data());
+}
+
+int ConvInputGradM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto conv_param = reinterpret_cast<const ConvParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::Conv2DBackpropInputFusionT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->activation_type = static_cast<schema::ActivationType>(conv_param->act_type_);
+  prim->format = static_cast<schema::Format>(conv_param->out_format_);
+  prim->stride = {conv_param->stride_h_, conv_param->stride_w_};
+  prim->kernel_size = {conv_param->kernel_h_, conv_param->kernel_w_};
+  prim->dilation = {conv_param->dilation_h_, conv_param->dilation_w_};
+  prim->out_channel = conv_param->output_channel_;
+  prim->in_channel = conv_param->input_channel_;
+  prim->group = conv_param->group_;
+  prim->pad_mode = static_cast<schema::PadMode>(conv_param->pad_mode_);
+  prim->pad_list = {conv_param->pad_u_, conv_param->pad_d_, conv_param->pad_l_, conv_param->pad_r_};
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+ConvFilterGradM::ConvFilterGradM(ConvM *conv_node) : Node() {
+  CloneOpParam<ConvParameter>(conv_node->OpParam());
+  set_primitive(schema::PrimitiveType_Conv2DBackpropFilterFusion);
+  set_name(kGradName + "/conv2DBackpropFilter");
+  expr()->SetSize(C4NUM);
+  auto w = conv_node->input(1);
+  CreateConstTensor(C2NUM, {static_cast<int32_t>(w->dims().size())}, kNumberTypeInt32, KHWC, "shape", w->dims().data());
+}
+int ConvFilterGradM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto conv_param = reinterpret_cast<const ConvParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::Conv2DBackpropFilterFusionT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->activation_type = static_cast<schema::ActivationType>(conv_param->act_type_);
+  prim->format = static_cast<schema::Format>(conv_param->out_format_);
+  prim->stride = {conv_param->stride_h_, conv_param->stride_w_};
+  prim->kernel_size = {conv_param->kernel_h_, conv_param->kernel_w_};
+  prim->dilation = {conv_param->dilation_h_, conv_param->dilation_w_};
+  prim->out_channel = conv_param->output_channel_;
+  prim->in_channel = conv_param->input_channel_;
+  prim->group = conv_param->group_;
+  prim->pad_mode = static_cast<schema::PadMode>(conv_param->pad_mode_);
+  prim->pad_list = {conv_param->pad_u_, conv_param->pad_d_, conv_param->pad_l_, conv_param->pad_r_};
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg reg(schema::PrimitiveType_Conv2DFusion, ReturnNode<ConvM>);
+
+namespace NN {
+Node *Conv2D(const ConvConfig &cfg) {
+  auto c = new (std::nothrow) ConvM(cfg);
+  if (c == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate Convolution object";
+    return nullptr;
+  }
+  return c;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/conv.h

@@ -0,0 +1,58 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_CONV_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_CONV_H_
+
+#include <vector>
+#include <memory>
+#include <string>
+#include "src/expression/cfg.h"
+#include "src/expression/node.h"
+
+namespace mindspore {
+namespace lite {
+class ConvM : public Node {
+ public:
+  ConvM() = default;
+  explicit ConvM(const ConvConfig &cfg);
+  std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs) override;
+  Param *weight() override { return input(1)->node()->data(); }
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+  void SetLearn() override;
+
+ private:
+  int GetMode(std::string mode);
+  Node *bias_ = nullptr;
+  EXPR *wbias_ = nullptr;
+};
+
+class ConvInputGradM : public Node {
+ public:
+  explicit ConvInputGradM(ConvM *conv_node);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+class ConvFilterGradM : public Node {
+ public:
+  explicit ConvFilterGradM(ConvM *conv_node);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_CONV_H_

mindspore/lite/src/expression/ops/dense.cc

@@ -0,0 +1,151 @@
+/**
+ * Copyright 2022 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/expression/ops/dense.h"
+#include <memory>
+#include "include/api/cfg.h"
+#include "src/expression/ops/biasadd.h"
+#include "src/expression/ops/depend.h"
+#include "src/expression/ops.h"
+#include "nnacl/matmul_parameter.h"
+#include "src/expression/import.h"
+#include "inner/model_generated.h"
+#include "src/cxx_api/expression/node_impl.h"
+
+namespace mindspore {
+namespace lite {
+DenseM::DenseM(const DenseConfig &cfg) : Node() {
+  auto op_param = calloc(1, sizeof(MatMulParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate MatMulParameter";
+    return;
+  }
+  set_name(UniqueName("Dense"));
+  SetOpParam(op_param);
+  expr()->SetSize(C2NUM);
+  set_primitive(schema::PrimitiveType_MatMulFusion);
+  auto param = reinterpret_cast<MatMulParameter *>(opParam_.get());
+  param->row_ = cfg.out_channels_;
+  param->col_ = cfg.in_channels_;
+  param->a_transpose_ = false;
+  param->b_transpose_ = true;
+  std::vector<int> dims = {param->row_, param->col_};
+  auto w = Node::CreateWeights(dims, kNumberTypeFloat32, KHWC, Param::Mode::NORMAL, "weights");
+  expr()->set_params(C1NUM, w);
+  if (cfg.has_bias_) {
+    wbias_ = CreateWeights({cfg.out_channels_}, kNumberTypeFloat32, KHWC, Param::Mode::ZEROS, "bias_weights");
+    bias_ = new (std::nothrow) BiasAddM(KHWC);
+    if (bias_ == nullptr) {
+      MS_LOG(ERROR) << "Cannot allocate bias";
+      return;
+    }
+    bias_->update_name(name());
+    AddLearn(wbias_->node());
+    PushOp(bias_);
+  }
+  SetLearn();
+}
+
+std::vector<EXPR *> DenseM::construct(const std::vector<EXPR *> &inputs) {
+  auto x = Node::construct(inputs);
+  if (bias_ != nullptr) {
+    x = (*bias_)({x.front(), wbias_});
+  }
+  return x;
+}
+
+std::vector<EXPR *> DenseM::Grad(EXPR *yt) {
+  auto src_param = reinterpret_cast<MatMulParameter *>(opParam_.get());
+  bool ta = src_param->a_transpose_;
+  bool tb = src_param->b_transpose_;
+
+  // dx grad op
+  auto dxGrad = new (std::nothrow) DenseM();
+  if (dxGrad == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate dxGrad ";
+    return {};
+  }
+  PushOp(dxGrad);
+  dxGrad->CloneOpParam<MatMulParameter>(opParam_);
+  dxGrad->set_primitive(schema::PrimitiveType_MatMulFusion);
+  auto dxGradParam = reinterpret_cast<MatMulParameter *>(dxGrad->OpParam());
+  dxGradParam->a_transpose_ = (ta && tb);
+  dxGradParam->b_transpose_ = (ta || !tb);
+  dxGrad->set_name(name() + kGradName + "/dxGrad");
+  EXPR *dx = nullptr;
+  if (ta) {
+    dx = (*dxGrad)({input(1), yt}).front();
+  } else {
+    dx = (*dxGrad)({yt, input(1)}).front();
+  }
+  // Control execution flow
+  auto depend = NN::Depend();
+  if (depend == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate depend ";
+    return {};
+  }
+  PushOp(depend);
+  auto de = (*depend)({dxGrad->expr()}).front();
+
+  // dw grad op
+  auto dwGrad = new (std::nothrow) DenseM();
+  if (dwGrad == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate dwGrad ";
+    return {};
+  }
+  PushOp(dwGrad);
+  dwGrad->CloneOpParam<MatMulParameter>(opParam_);
+  dwGrad->set_primitive(schema::PrimitiveType_MatMulFusion);
+  auto dwGradParam = reinterpret_cast<MatMulParameter *>(dwGrad->OpParam());
+  dwGradParam->a_transpose_ = (!ta || tb);
+  dwGradParam->b_transpose_ = ta && tb;
+  dwGrad->set_name(name() + kGradName + "/dwGrad");
+  EXPR *dw = nullptr;
+  if (tb) {
+    dw = (*dwGrad)({yt, input(0), de}).front();
+  } else {
+    dw = (*dwGrad)({input(0), yt, de}).front();
+  }
+  return {dx, dw};
+}
+int DenseM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto dense_param = reinterpret_cast<const MatMulParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::MatMulFusionT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->transpose_a = dense_param->a_transpose_;
+  prim->transpose_b = dense_param->b_transpose_;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+void DenseM::SetLearn() { AddLearn(input(C1NUM)->node()); }
+
+static ImportReg reg(schema::PrimitiveType_MatMulFusion, ReturnNode<DenseM>);
+
+namespace NN {
+Node *Dense(const DenseConfig &cfg) {
+  auto l = new (std::nothrow) DenseM(cfg);
+  if (l == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate Dense object";
+  }
+  return l;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/dense.h

@@ -0,0 +1,44 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_DENSE_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_DENSE_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/node.h"
+#include "src/expression/cfg.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class DenseM : public Node {
+ public:
+  DenseM() = default;
+  explicit DenseM(const DenseConfig &cfg);
+  std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs) override;
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+  void SetLearn() override;
+
+ private:
+  Param *weight() override { return input(1)->node()->data(); }
+  Node *bias_{nullptr};
+  EXPR *wbias_{nullptr};
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_DENSE_H_

mindspore/lite/src/expression/ops/depend.cc

@@ -0,0 +1,43 @@
+/**
+ * Copyright 2022 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/expression/ops/depend.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+DependM::DependM() : Node() {
+  auto param = calloc(1, sizeof(OpParameter));
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate parameter";
+    return;
+  }
+  SetOpParam(param);
+  set_primitive(schema::PrimitiveType_Depend);
+  set_name(UniqueName("Depend"));
+}
+namespace NN {
+Node *Depend() {
+  auto d = new (std::nothrow) DependM();
+  if (d == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate depend object";
+    return nullptr;
+  }
+  return d;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/depend.h

@@ -0,0 +1,37 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_DEPEND_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_DEPEND_H_
+
+#include <vector>
+#include "src/expression/node.h"
+#include "src/expression/ops.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class DependM : public Node {
+ public:
+  DependM();
+};
+
+namespace NN {
+Node *Depend();
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_DEPEND_H_

mindspore/lite/src/expression/ops/dropout.cc

@@ -0,0 +1,91 @@
+/**
+ * Copyright 2022 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/expression/ops/dropout.h"
+#include <vector>
+#include "nnacl/fp32_grad/dropout_parameter.h"
+#include "inner/model_generated.h"
+#include "src/expression/import.h"
+#include "src/expression/ops.h"
+#include "src/cxx_api/expression/node_impl.h"
+
+namespace mindspore {
+namespace lite {
+DropOutM::DropOutM(float ratio) {
+  auto param = reinterpret_cast<DropoutParameter *>(calloc(1, sizeof(DropoutParameter)));
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate parameter";
+    return;
+  }
+  param->ratio_ = ratio;
+  SetOpParam(param);
+  set_primitive(schema::PrimitiveType_Dropout);
+  set_name(UniqueName("DropOut"));
+}
+
+std::vector<EXPR *> DropOutM::Grad(EXPR *yt) {
+  auto inGrad = new (std::nothrow) DropOutGradM(this);
+  if (inGrad == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate drop grad";
+    return {};
+  }
+  return (*inGrad)({yt, expr()});
+}
+
+int DropOutM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto param = reinterpret_cast<const DropoutParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::DropoutT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->keep_prob = param->ratio_;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+DropOutGradM::DropOutGradM(DropOutM *node) {
+  CloneOpParam<DropoutParameter>(node->OpParam());
+  set_primitive(schema::PrimitiveType_DropoutGrad);
+  set_name(kGradName + "/DropOutGrad");
+}
+
+int DropOutGradM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto param = reinterpret_cast<const DropoutParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::DropoutGradT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->keep_prob = param->ratio_;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg reg(schema::PrimitiveType_Dropout, ReturnNode<DropOutM>);
+
+namespace NN {
+Node *DropOut(float ratio) {
+  auto node = new (std::nothrow) DropOutM(ratio);
+  if (node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate dropout node";
+    return nullptr;
+  }
+  return node;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/dropout.h

@@ -0,0 +1,42 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_DROPOUT_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_DROPOUT_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/node.h"
+
+namespace mindspore {
+namespace lite {
+class DropOutM : public Node {
+ public:
+  DropOutM() = default;
+  explicit DropOutM(float ratio);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+class DropOutGradM : public Node {
+ public:
+  explicit DropOutGradM(DropOutM *node);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_DROPOUT_H_

mindspore/lite/src/expression/ops/flatten.cc

@@ -0,0 +1,71 @@
+/**
+ * Copyright 2022 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/expression/ops/flatten.h"
+#include <vector>
+#include "inner/model_generated.h"
+#include "src/expression/import.h"
+#include "src/expression/ops.h"
+
+#include "src/cxx_api/expression/node_impl.h"
+#include "nnacl/op_base.h"
+
+namespace mindspore {
+namespace lite {
+FlattenM::FlattenM(int dummy) {
+  auto param = reinterpret_cast<OpParameter *>(calloc(C1NUM, sizeof(OpParameter)));
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate parameter";
+    return;
+  }
+  SetOpParam(param);
+  set_primitive(schema::PrimitiveType_Flatten);
+  set_name(UniqueName("Flatten"));
+}
+
+std::vector<EXPR *> FlattenM::construct(const std::vector<EXPR *> &inputs) {
+  auto in = inputs;
+  auto y = Node::construct(in);
+  return y;
+}
+
+std::vector<EXPR *> FlattenM::Grad(EXPR *yt) {
+  auto shape_of_x = input(0)->dims();
+  auto reshape = NN::Reshape(shape_of_x);
+  PushOp(reshape);
+  return (*reshape)({yt});
+}
+
+int FlattenM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::DropoutT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg reg(schema::PrimitiveType_Flatten, ReturnNode<FlattenM>);
+
+namespace NN {
+Node *Flatten() {
+  auto node = new (std::nothrow) FlattenM(0);
+  return node;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/flatten.h

@@ -0,0 +1,36 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_FLATTEN_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_FLATTEN_H_
+
+#include <memory>
+#include <vector>
+#include "src/expression/node.h"
+
+namespace mindspore {
+namespace lite {
+class FlattenM : public Node {
+ public:
+  FlattenM() = default;
+  explicit FlattenM(int dummy);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+  std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_FLATTEN_H_

mindspore/lite/src/expression/ops/pooling.cc

@@ -0,0 +1,215 @@
+/**
+ * Copyright 2022 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/expression/ops/pooling.h"
+#include "src/expression/ops.h"
+#include "src/expression/import.h"
+#include "src/cxx_api/expression/node_impl.h"
+#include "src/expression/ops/transpose.h"
+
+namespace mindspore {
+namespace lite {
+PoolingM::PoolingM(const PoolingConfig &cfg) : Node() {
+  auto op_param = calloc(1, sizeof(PoolingParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate PoolingParameter";
+    return;
+  }
+  SetOpParam(op_param);
+  PoolingParameter *pool_param = reinterpret_cast<PoolingParameter *>(OpParam());
+
+  pool_param->window_h_ = cfg.kernel_size_[0];
+  pool_param->window_w_ = cfg.kernel_size_[1];
+  pool_param->stride_h_ = cfg.stride_[0];
+  pool_param->stride_w_ = cfg.stride_[1];
+  auto pad_mode = GetMode(cfg.pad_mode_);
+  if (pad_mode == -1) {
+    MS_LOG(ERROR) << "bad pad mode";
+    return;
+  }
+  pool_param->pad_mode_ = static_cast<PadMode>(pad_mode + Pad_pad);
+  pool_param->round_mode_ = RoundMode_Floor;
+  pool_param->act_type_ = ActType_No;
+}
+
+std::vector<EXPR *> PoolingM::construct(const std::vector<EXPR *> &inputs) {
+  auto in = inputs;
+  auto x = in.front();
+  if (x->format() != NHWC && x->dims().size() == C4NUM) {
+    x = TransposeM::TransposeCHW2HWC(x);
+    x->node()->set_name(name() + "/" + x->node()->name());
+    PushOp(x->node());
+    in.at(0) = x;
+  }
+  auto y = Node::construct(in);
+  return y;
+}
+
+int PoolingM::GetMode(std::string mode) {
+  const std::vector<std::string> list = {"same", "valid"};
+  auto itr = std::find(list.begin(), list.end(), mode);
+  if (itr == list.end()) {
+    MS_LOG(ERROR) << "illegal mode" << mode;
+    return -1;
+  }
+  return std::distance(list.begin(), itr);
+}
+
+void PoolingM::UpdateRoundMode(const PoolingParameter *param, schema::RoundMode *round_mode) {
+  switch (param->round_mode_) {
+    case RoundMode_Floor:
+      *round_mode = schema::RoundMode_FLOOR;
+      break;
+    case RoundMode_Ceil:
+      *round_mode = schema::RoundMode_CEIL;
+      break;
+    default:
+      *round_mode = schema::RoundMode_FLOOR;
+      break;
+  }
+}
+
+template <typename T>
+int PoolingM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto param = reinterpret_cast<const PoolingParameter *>(OpParam());
+  auto prim = new (std::nothrow) T;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->kernel_size = {param->window_h_, param->window_w_};
+  prim->strides = {param->stride_h_, param->stride_w_};
+  prim->pad = {param->pad_u_, param->pad_d_, param->pad_l_, param->pad_r_};
+  prim->pad_mode = static_cast<schema::PadMode>(param->pad_mode_);
+  UpdateRoundMode(param, &prim->round_mode);
+  prim->global = param->global_;
+  prim->activation_type = schema::ActivationType_NO_ACTIVATION;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+template <typename T>
+int PoolingM::UnPopulateGrad(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto param = reinterpret_cast<const PoolingParameter *>(OpParam());
+  auto prim = new (std::nothrow) T;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->kernel_size = {param->window_h_, param->window_w_};
+  prim->strides = {param->stride_h_, param->stride_w_};
+  prim->pad_mode = static_cast<schema::PadMode>(param->pad_mode_);
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+// Max pooling Definition
+MaxPoolM::MaxPoolM(const PoolingConfig &cfg) : PoolingM(cfg) {
+  auto param = reinterpret_cast<PoolingParameter *>(OpParam());
+  param->pool_mode_ = PoolMode_MaxPool;
+  set_primitive(schema::PrimitiveType_MaxPoolFusion);
+  set_name(UniqueName("MaxPool"));
+}
+
+int MaxPoolM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  return PoolingM::UnPopulate<schema::MaxPoolFusionT>(cnode);
+}
+
+std::vector<EXPR *> MaxPoolM::Grad(EXPR *yt) {
+  auto in = yt;
+  if (yt->format() != NHWC && yt->dims().size() == C4NUM) {
+    in = TransposeM::TransposeCHW2HWC(yt);
+    in->node()->set_name(kGradName + "/" + name() + "/" + in->node()->name());
+    PushOp(in->node());
+  }
+  auto pool_grad = new (std::nothrow) MaxPoolGradM(this);
+  PushOp(pool_grad);
+  return (*pool_grad)({input(0), output(0), in});
+}
+
+static ImportReg maxPoolReg(schema::PrimitiveType_MaxPoolFusion, ReturnNode<MaxPoolM>);
+
+// Avg pooling Definition
+AvgPoolM::AvgPoolM(const PoolingConfig &cfg) : PoolingM(cfg) {
+  auto param = reinterpret_cast<PoolingParameter *>(OpParam());
+  param->pool_mode_ = PoolMode_AvgPool;
+  set_primitive(schema::PrimitiveType_AvgPoolFusion);
+  set_name(UniqueName("AvgPool"));
+}
+
+int AvgPoolM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  return PoolingM::UnPopulate<schema::AvgPoolFusionT>(cnode);
+}
+
+std::vector<EXPR *> AvgPoolM::Grad(EXPR *yt) {
+  auto in = yt;
+  if (yt->format() != NHWC && yt->dims().size() == C4NUM) {
+    in = TransposeM::TransposeCHW2HWC(yt);
+    in->node()->set_name(kGradName + "/" + name() + "/" + in->node()->name());
+    PushOp(in->node());
+  }
+  auto pool_grad = new (std::nothrow) AvgPoolGradM(this);
+  PushOp(pool_grad);
+  return (*pool_grad)({input(0), output(0), in});
+}
+
+static ImportReg avgPoolReg(schema::PrimitiveType_AvgPoolFusion, ReturnNode<AvgPoolM>);
+
+// Max Pool Grad Definition
+MaxPoolGradM::MaxPoolGradM(MaxPoolM *node) {
+  Node();
+  CloneOpParam<PoolingParameter>(node->OpParam());
+  set_primitive(schema::PrimitiveType_MaxPoolGrad);
+  set_name(kGradName + "/" + node->name() + "/MaxPoolGrad");
+}
+
+int MaxPoolGradM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  return PoolingM::UnPopulateGrad<schema::MaxPoolGradT>(cnode);
+}
+
+// Avg Pool Grad Definition
+AvgPoolGradM::AvgPoolGradM(AvgPoolM *node) {
+  Node();
+  CloneOpParam<PoolingParameter>(node->OpParam());
+  set_primitive(schema::PrimitiveType_AvgPoolGrad);
+  set_name(kGradName + "/" + node->name() + "/AvgPoolGrad");
+}
+
+int AvgPoolGradM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  return PoolingM::UnPopulateGrad<schema::AvgPoolGradT>(cnode);
+}
+
+namespace NN {
+Node *MaxPool2D(const PoolingConfig &cfg) {
+  auto c = new (std::nothrow) MaxPoolM(cfg);
+  if (c == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate max pool object";
+    return nullptr;
+  }
+  return c;
+}
+
+Node *AvgPool2D(const PoolingConfig &cfg) {
+  auto c = new (std::nothrow) AvgPoolM(cfg);
+  if (c == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate average pool object";
+    return nullptr;
+  }
+  return c;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/pooling.h

@@ -0,0 +1,74 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_POOLING_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_POOLING_H_
+
+#include <vector>
+#include <string>
+#include <memory>
+#include "src/expression/node.h"
+#include "inner/model_generated.h"
+#include "src/expression/cfg.h"
+#include "nnacl/pooling_parameter.h"
+
+namespace mindspore {
+namespace lite {
+class PoolingM : public Node {
+ public:
+  PoolingM() = default;
+  explicit PoolingM(const PoolingConfig &cfg);
+  template <typename T>
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode);
+  template <typename T>
+  int UnPopulateGrad(const std::unique_ptr<schema::CNodeT> &cnode);
+  std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs);
+
+ private:
+  void UpdateRoundMode(const PoolingParameter *param, enum schema::RoundMode *round_mode);
+  int GetMode(std::string mode);
+};
+
+class MaxPoolM : public PoolingM {
+ public:
+  MaxPoolM() = default;
+  explicit MaxPoolM(const PoolingConfig &cfg);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+class AvgPoolM : public PoolingM {
+ public:
+  AvgPoolM() = default;
+  explicit AvgPoolM(const PoolingConfig &cfg);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+class MaxPoolGradM : public PoolingM {
+ public:
+  explicit MaxPoolGradM(MaxPoolM *node);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+class AvgPoolGradM : public PoolingM {
+ public:
+  explicit AvgPoolGradM(AvgPoolM *node);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_POOLING_H_

mindspore/lite/src/expression/ops/reduce.cc

@@ -0,0 +1,126 @@
+/**
+ * Copyright 2022 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/expression/ops/reduce.h"
+#include <functional>
+#include "src/expression/ops/tile.h"
+#include "src/expression/ops/reshape.h"
+#include "src/expression/ops/arithmetic.h"
+#include "src/expression/ops.h"
+#include "src/expression/ops_utils.h"
+#include "src/expression/import.h"
+#include "nnacl/reduce_parameter.h"
+#include "src/cxx_api/expression/node_impl.h"
+
+namespace mindspore {
+namespace lite {
+ReduceM::ReduceM(schema::ReduceMode mode, bool keep_dims, const std::vector<int> &axis) : Node() {
+  expr()->SetSize(C2NUM);
+  ReduceParameter *param = reinterpret_cast<ReduceParameter *>(calloc(1, sizeof(ReduceParameter)));
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate parameter";
+    return;
+  }
+  param->mode_ = mode;
+  param->keep_dims_ = keep_dims;
+  param->reduce_to_end_ = false;
+  param->coeff = 1.f;
+  SetOpParam(param);
+  set_name(UniqueName("Reduce"));
+  set_primitive(schema::PrimitiveType_ReduceFusion);
+  Node::CreateConstTensor(C1NUM, {static_cast<int32_t>(axis.size())}, kNumberTypeInt32, KHWC, "axis", axis.data());
+}
+
+int ReduceM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto reduce_param = reinterpret_cast<const ReduceParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::ReduceFusionT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->keep_dims = reduce_param->keep_dims_;
+  prim->mode = static_cast<schema::ReduceMode>(reduce_param->mode_);
+  prim->coeff = reduce_param->coeff;
+  prim->reduce_to_end = reduce_param->reduce_to_end_;
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+std::vector<EXPR *> ReduceM::Grad(EXPR *yt) {
+  auto shape_of_x = input(0)->dims();
+  std::vector<int> shape_of_axis;
+
+  auto data = input(1)->node()->data()->data().data();
+  int size = input(1)->dims().at(0);
+  auto int_data = reinterpret_cast<int *>(data);
+  for (int i = 0; i < size; i++) {
+    shape_of_axis.push_back(int_data[i]);
+  }
+
+  // assume no dynamic shape
+  ShapeReduce reduce_shape;
+  auto output_shape_kept_dims = ShapeReduce()(shape_of_x, shape_of_axis);
+  auto tile_scaling = VectorDiv()(shape_of_x, output_shape_kept_dims);
+  auto reshape = NN::Reshape(output_shape_kept_dims);
+  PushOp(reshape);
+  reshape->set_name(name() + "/reshape");
+  auto g = (*reshape)({yt}).front();
+  auto tile = NN::Tile(tile_scaling);
+  PushOp(tile);
+  tile->set_name(name() + "/tile");
+  auto sum_grad = (*tile)({g}).front();
+  auto reduce_param = reinterpret_cast<const ReduceParameter *>(OpParam());
+  if (reduce_param->mode_ == schema::ReduceMode_ReduceSum) {
+    return {sum_grad};
+  } else if (reduce_param->mode_ == schema::ReduceMode_ReduceMean) {
+    auto shape_of_y = output(0)->dims();
+    auto shape_x_mul = std::accumulate(shape_of_x.begin(), shape_of_x.end(), 1, std::multiplies<int>());
+    auto shape_y_mul = std::accumulate(shape_of_y.begin(), shape_of_y.end(), 1, std::multiplies<int>());
+    auto div_shape = static_cast<float>(shape_x_mul) / static_cast<float>(shape_y_mul);
+    auto div_op = NN::Div();
+    PushOp(div_op);
+    auto d = div_op->CreateConstTensor(C1NUM, {1}, kNumberTypeFloat32, KHWC, "div_shape", &div_shape);
+    auto dx = (*div_op)({sum_grad, d->expr()});
+    return dx;
+  } else {
+    return {};
+  }
+}
+
+static ImportReg reg(schema::PrimitiveType_ReduceFusion, ReturnNode<ReduceM>);
+
+namespace NN {
+Node *ReduceSum(bool keep_dims, const std::vector<int> &axis) {
+  auto node = new (std::nothrow) ReduceM(schema::ReduceMode_ReduceSum, keep_dims, axis);
+  if (node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate  reduce sum node";
+    return nullptr;
+  }
+  node->set_name(Node::UniqueName("ReduceSum"));
+  return node;
+}
+Node *ReduceMean(bool keep_dims, const std::vector<int> &axis) {
+  auto node = new (std::nothrow) ReduceM(schema::ReduceMode_ReduceMean, keep_dims, axis);
+  if (node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate reduce mean node";
+    return nullptr;
+  }
+  node->set_name(Node::UniqueName("ReduceMean"));
+  return node;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/reduce.h

@@ -0,0 +1,42 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_REDUCE_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_REDUCE_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/node.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class ReduceM : public Node {
+ public:
+  ReduceM() = default;
+  ReduceM(schema::ReduceMode mode, bool keep_dims, const std::vector<int> &axis);
+  Param *weight() override { return input(1)->node()->data(); }
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+};
+
+namespace NN {
+Node *ReduceMean(bool keep_dims, const std::vector<int> &axis);
+Node *ReduceSum(bool keep_dims, const std::vector<int> &axis);
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_REDUCE_H_

mindspore/lite/src/expression/ops/reshape.cc

@@ -0,0 +1,74 @@
+/**
+ * Copyright 2022 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/expression/ops/reshape.h"
+#include "src/expression/ops.h"
+#include "nnacl/reshape_parameter.h"
+#include "src/expression/import.h"
+
+namespace mindspore {
+namespace lite {
+ReshapeM::ReshapeM(const std::vector<int> &shape) : Node() {
+  auto op_param = calloc(1, sizeof(ReshapeParameter));
+  if (op_param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate ReshapeParameter";
+    return;
+  }
+  set_name(UniqueName("Reshape"));
+  expr()->SetSize(C2NUM);
+  SetOpParam(op_param);
+  set_primitive(schema::PrimitiveType_Reshape);
+
+  ReshapeParameter *reshape_param = reinterpret_cast<ReshapeParameter *>(opParam_.get());
+  reshape_param->shape_dim_ = shape.size();
+  for (int i = 0; i < reshape_param->shape_dim_; i++) {
+    reshape_param->shape_[i] = shape.at(i);
+  }
+  Node::CreateConstTensor(C1NUM, {static_cast<int32_t>(shape.size())}, kNumberTypeInt32, KHWC, "shape", shape.data());
+}
+
+int ReshapeM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::ReshapeT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+std::vector<EXPR *> ReshapeM::Grad(EXPR *yt) {
+  auto shape_of_x = input(0)->dims();
+  auto reshape = NN::Reshape(shape_of_x);
+  PushOp(reshape);
+  return (*reshape)({yt});
+}
+
+static ImportReg reg(schema::PrimitiveType_Reshape, ReturnNode<ReshapeM>);
+
+namespace NN {
+Node *Reshape(const std::vector<int> &shape) {
+  auto node = new (std::nothrow) ReshapeM(shape);
+  if (node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate reshape node";
+    return nullptr;
+  }
+  node->set_name(Node::UniqueName("Reshape"));
+  return node;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/reshape.h

@@ -0,0 +1,37 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_RESHAPE_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_RESHAPE_H_
+
+#include <vector>
+#include <memory>
+
+#include "src/expression/node.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class ReshapeM : public Node {
+ public:
+  ReshapeM() = default;
+  explicit ReshapeM(const std::vector<int> &shape);
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_RESHAPE_H_

mindspore/lite/src/expression/ops/softmax.cc

@@ -0,0 +1,119 @@
+/**
+ * Copyright 2022 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/expression/ops/softmax.h"
+#include "nnacl/softmax_parameter.h"
+#include "inner/model_generated.h"
+#include "src/expression/import.h"
+#include "src/expression/ops/reshape.h"
+#include "src/expression/ops/reduce.h"
+#include "src/expression/ops/arithmetic.h"
+#include "src/expression/ops/transpose.h"
+#include "src/cxx_api/expression/node_impl.h"
+#include "src/expression/ops.h"
+
+namespace mindspore {
+namespace lite {
+SoftmaxM::SoftmaxM(int axis) {
+  auto param = reinterpret_cast<SoftmaxParameter *>(calloc(1, sizeof(SoftmaxParameter)));
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate parameter";
+    return;
+  }
+  param->axis_ = axis;
+  SetOpParam(param);
+  set_primitive(schema::PrimitiveType_Softmax);
+  set_name(UniqueName("Softmax"));
+}
+
+std::vector<int> SoftmaxM::getTransposeAxis(const std::vector<int> &shape, int axis) {
+  int rank = shape.size();
+  if (axis < 0) {
+    axis += rank;
+  }
+  std::vector<int> reverse_axis(rank);
+  std::iota(reverse_axis.begin(), reverse_axis.end(), 0);
+  reverse_axis.at(axis) = rank - 1;
+  reverse_axis.at(rank - 1) = axis;
+  return reverse_axis;
+}
+
+std::vector<EXPR *> SoftmaxM::Grad(EXPR *yt) {
+  auto x = input(0);
+  auto out = output(0);
+  auto shape_of_x = x->dims();
+  auto param = reinterpret_cast<const SoftmaxParameter *>(OpParam());
+  auto reverse_axis = getTransposeAxis(shape_of_x, param->axis_);
+
+  auto transpose_out = NN::Transpose(reverse_axis);
+  transpose_out->set_name(kGradName + "/" + name() + "/" + transpose_out->name() + "/out/");
+  PushOp(transpose_out);
+  auto y_trn = (*transpose_out)({out}).front();
+
+  auto transpose_dout = NN::Transpose(reverse_axis);
+  transpose_dout->set_name(kGradName + "/" + name() + "/" + transpose_dout->name() + "/dout/");
+  PushOp(transpose_dout);
+  auto yt_trn = (*transpose_dout)({yt}).front();
+
+  auto mul0 = NN::Mul();
+  mul0->set_name(kGradName + "/" + name() + "/" + mul0->name() + "0");
+  PushOp(mul0);
+  auto tmp0 = (*mul0)({y_trn, yt_trn}).front();
+
+  auto sum_func = NN::ReduceSum(true, {-1});
+  sum_func->set_name(kGradName + "/" + name() + "/" + sum_func->name());
+  PushOp(sum_func);
+  auto tmp1 = (*sum_func)({tmp0}).front();
+
+  auto sub = NN::Sub();
+  sub->set_name(kGradName + "/" + name() + "/" + sub->name());
+  PushOp(sub);
+  auto tmp2 = (*sub)({yt_trn, tmp1}).front();
+
+  auto mul1 = NN::Mul();
+  mul1->set_name(kGradName + "/" + name() + "/" + mul1->name() + "1");
+  PushOp(mul1);
+  auto tmp3 = (*mul1)({y_trn, tmp2});
+
+  auto transpose_dx = NN::Transpose(reverse_axis);
+  transpose_dx->set_name(kGradName + "/" + name() + "/" + transpose_dx->name() + "/dx");
+  PushOp(transpose_dx);
+  auto dx = (*transpose_dx)({tmp3});
+  return dx;
+}
+
+int SoftmaxM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto param = reinterpret_cast<const SoftmaxParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::SoftmaxT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  prim->axis.push_back(param->axis_);
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg reg(schema::PrimitiveType_Softmax, ReturnNode<SoftmaxM>);
+
+namespace NN {
+Node *Softmax(int axis) {
+  auto node = new (std::nothrow) SoftmaxM(axis);
+  return node;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/softmax.h

@@ -0,0 +1,39 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_SOFTMAX_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_SOFTMAX_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/node.h"
+
+namespace mindspore {
+namespace lite {
+class SoftmaxM : public Node {
+ public:
+  SoftmaxM() = default;
+  explicit SoftmaxM(int axis);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+
+ private:
+  std::vector<int> getTransposeAxis(const std::vector<int> &shape, int axis);
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_SOFTMAX_H_

mindspore/lite/src/expression/ops/softmaxCE.cc

@@ -0,0 +1,93 @@
+/**
+ * Copyright 2022 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/expression/ops/softmaxCE.h"
+#include "include/api/net.h"
+#include "src/cxx_api/expression/node_impl.h"
+#include "src/expression/ops/reduce.h"
+namespace mindspore {
+namespace NN {
+Node *SoftmaxCrossEntropy(const SoftMaxCrossEntropyCfg &cfg) {
+  auto lite_node = lite::NN::SoftmaxCrossEntropy(cfg);
+  return NodeImpl::Connect(lite_node);
+}
+}  // namespace NN
+
+namespace lite {
+SoftmaxCrossEntropyM::SoftmaxCrossEntropyM() {
+  auto param = calloc(1, sizeof(OpParameter));
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate parameter";
+    return;
+  }
+  expr()->SetSize(C2NUM);
+  SetOpParam(param);
+  set_name("SoftmaxCrossEntropy");
+  set_primitive(schema::PrimitiveType_SoftmaxCrossEntropyWithLogits);
+  EXPR e(this);
+  e.SetSize(0);
+  expr_.emplace_back(e);
+}
+
+Node *SoftmaxCrossEntropyM::GetReductionNode(const std::string &mode, const std::vector<int> &axis) {
+  if (mode == "mean") {
+    return NN::ReduceMean(false, axis);
+  } else if (mode == "sum") {
+    return NN::ReduceSum(false, axis);
+  } else {
+    return nullptr;
+  }
+}
+
+SoftmaxCrossEntropyM::SoftmaxCrossEntropyM(const SoftMaxCrossEntropyCfg &cfg) : SoftmaxCrossEntropyM() {
+  std::vector<int> axis = {0};
+  reduce_ = GetReductionNode(cfg.reduction, axis);
+  if (reduce_ != nullptr) {
+    PushOp(reduce_);
+  }
+}
+
+std::vector<EXPR *> SoftmaxCrossEntropyM::construct(const std::vector<EXPR *> &inputs) {
+  auto y = Node::construct(inputs);
+  if (reduce_ != nullptr) {
+    y = (*reduce_)({y.front()});
+  }
+  return y;
+}
+
+std::vector<EXPR *> SoftmaxCrossEntropyM::Grad(EXPR *expr) { return {this->expr(1)}; }
+
+int SoftmaxCrossEntropyM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::SoftmaxCrossEntropyWithLogitsT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+namespace NN {
+Node *SoftmaxCrossEntropy(const SoftMaxCrossEntropyCfg &cfg) {
+  auto s = new (std::nothrow) SoftmaxCrossEntropyM(cfg);
+  if (s == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate softmax node";
+  }
+  return s;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/softmaxCE.h

@@ -0,0 +1,47 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_SOFTMAXCE_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_SOFTMAXCE_H_
+
+#include <vector>
+#include <memory>
+#include <string>
+#include "src/expression/node.h"
+#include "inner/model_generated.h"
+#include "include/api/net.h"
+
+namespace mindspore {
+namespace lite {
+class SoftmaxCrossEntropyM : public Node {
+ public:
+  SoftmaxCrossEntropyM();
+  explicit SoftmaxCrossEntropyM(const SoftMaxCrossEntropyCfg &cfg);
+  std::vector<EXPR *> Grad(EXPR *expr) override;
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+  std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs) override;
+
+ private:
+  Node *GetReductionNode(const std::string &mode, const std::vector<int> &axis);
+  Node *reduce_ = nullptr;
+};
+
+namespace NN {
+Node *SoftmaxCrossEntropy(const SoftMaxCrossEntropyCfg &cfg);
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_SOFTMAXCE_H_

mindspore/lite/src/expression/ops/tile.cc

@@ -0,0 +1,62 @@
+/**
+ * Copyright 2022 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/expression/ops/tile.h"
+#include <memory>
+#include "src/expression/ops.h"
+#include "nnacl/base/tile_base.h"
+
+namespace mindspore {
+namespace lite {
+TileM::TileM(const std::vector<int> &multiples) : Node() {
+  expr()->SetSize(C2NUM);
+  TileParameter *param = reinterpret_cast<TileParameter *>(calloc(1, sizeof(TileParameter)));
+  if (param == nullptr) {
+    MS_LOG(ERROR) << " cannot allocate ConvParameter";
+    return;
+  }
+  SetOpParam(param);
+  set_name(UniqueName("Tile"));
+  set_primitive(schema::PrimitiveType_TileFusion);
+  Node::CreateConstTensor(C1NUM, {static_cast<int32_t>(multiples.size())}, kNumberTypeInt32, KHWC, "axis",
+                          multiples.data());
+}
+
+int TileM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto tile_param = reinterpret_cast<const TileParameter *>(OpParam());
+  auto prim = new (std::nothrow) schema::TileFusionT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  for (size_t i = 0; i < tile_param->dims_size_; i++) {
+    prim->dims.push_back(tile_param->dims_[i]);
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+namespace NN {
+Node *Tile(const std::vector<int> &multiples) {
+  auto node = new (std::nothrow) TileM(multiples);
+  if (node == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate tile node";
+  }
+  return node;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/tile.h

@@ -0,0 +1,40 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_TILE_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_TILE_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/node.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class TileM : public Node {
+ public:
+  TileM() = default;
+  explicit TileM(const std::vector<int> &multiples);
+  Param *weight() override { return input(1)->node()->data(); }
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+};
+
+namespace NN {
+Node *Tile(const std::vector<int> &multiples);
+}
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_TILE_H_

mindspore/lite/src/expression/ops/transpose.cc

@@ -0,0 +1,88 @@
+/**
+ * Copyright 2022 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/expression/ops/transpose.h"
+#include <memory>
+#include "nnacl/transpose.h"
+#include "inner/model_generated.h"
+#include "src/expression/import.h"
+
+namespace mindspore {
+namespace lite {
+TransposeM::TransposeM(const std::vector<int> &vector) {
+  auto param = calloc(1, sizeof(TransposeParameter));
+  if (param == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate transpose parameter";
+    return;
+  }
+  SetOpParam(param);
+  expr()->SetSize(C2NUM);
+  set_primitive(schema::PrimitiveType_Transpose);
+  std::vector<int> dims = {static_cast<int>(vector.size())};
+  set_name(UniqueName("Transpose"));
+  CreateConstTensor(C1NUM, dims, kNumberTypeInt32, KHWC, "axis", vector.data());
+}
+
+std::vector<int> TransposeM::Invert(const std::vector<int> &vector) {
+  std::vector<int> res;
+  for (size_t i = 0; i < vector.size(); i++) {
+    int idx = static_cast<int>(i);
+    auto val = std::find_if(vector.begin(), vector.end(), [idx](int x) { return (x == idx) ? true : false; });
+    if (val == vector.end()) {
+      MS_LOG(ERROR) << "Wrong index for " << idx;
+      return {};
+    }
+    res.push_back(std::distance(vector.begin(), val));
+  }
+  return res;
+}
+
+std::vector<EXPR *> TransposeM::Grad(EXPR *yt) {
+  auto tensor = input(1)->node();
+  auto data = tensor->data();
+  auto vec = data->Extract<int>();
+  auto invert = Invert(vec);
+  auto tran = new (std::nothrow) TransposeM(invert);
+  if (tran == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate transpose grad";
+    return {};
+  }
+  tran->set_name(kGradName + "/" + name() + "/" + tran->name());
+  PushOp(tran);
+  auto grad = (*tran)({yt});
+  return grad;
+}
+
+int TransposeM::UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) {
+  auto prim = new (std::nothrow) schema::TransposeT;
+  if (prim == nullptr) {
+    MS_LOG(ERROR) << "Cannot allocate primitive";
+    return RET_ERROR;
+  }
+  cnode->primitive->value.value = prim;
+  return RET_OK;
+}
+
+static ImportReg reg(schema::PrimitiveType_Transpose, ReturnNode<TransposeM>);
+
+namespace NN {
+Node *Transpose(const std::vector<int> &permute) {
+  auto node = new (std::nothrow) TransposeM(permute);
+  return node;
+}
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops/transpose.h

@@ -0,0 +1,59 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_TRANSPOSE_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_TRANSPOSE_H_
+
+#include <vector>
+#include <memory>
+#include "src/expression/node.h"
+#include "inner/model_generated.h"
+
+namespace mindspore {
+namespace lite {
+class TransposeM : public Node {
+ public:
+  TransposeM() = default;
+  explicit TransposeM(const std::vector<int> &vector);
+  static EXPR *TransposeCHW2HWC(EXPR *in) {
+    std::vector<int> res = {0, 2, 3, 1};
+    auto trans = new (std::nothrow) TransposeM(res);
+    if (trans == nullptr) {
+      return nullptr;
+    }
+    return (*trans)({in}).front();
+  }
+  static EXPR *TransposeHWC2CHW(EXPR *in) {
+    std::vector<int> res = {0, 3, 1, 2};
+    auto trans = new (std::nothrow) TransposeM(res);
+    if (trans == nullptr) {
+      return nullptr;
+    }
+    return (*trans)({in}).front();
+  }
+  int UnPopulate(const std::unique_ptr<schema::CNodeT> &cnode) override;
+  std::vector<EXPR *> Grad(EXPR *yt) override;
+
+ private:
+  std::vector<int> Invert(const std::vector<int> &vec);
+};
+
+namespace NN {
+Node *Transpose(const std::vector<int> &permute);
+}  // namespace NN
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_TRANSPOSE_H_

mindspore/lite/src/expression/ops_utils.cc

@@ -0,0 +1,275 @@
+/**
+ * Copyright 2022 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/expression/ops_utils.h"
+#include <set>
+#include <algorithm>
+
+namespace mindspore {
+namespace lite {
+enum class State {
+  SAME,
+  X_ONE,
+  Y_ONE,
+};
+
+bool CompareShape(const std::vector<int> &x_shape, const std::vector<int> &y_shape) {
+  if (x_shape.size() != y_shape.size()) {
+    return false;
+  }
+
+  for (size_t i = 0; i < x_shape.size(); ++i) {
+    if (x_shape.at(i) != y_shape.at(i)) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+void ComputeReduceIndex(const std::vector<int> &reverse_x, const std::vector<int> &reverse_y,
+                        std::vector<int> *grad_x_reduce_idx, std::vector<int> *grad_y_reduce_idy) {
+  MS_ASSERT(grad_x_reduce_idx != nullptr);
+  MS_ASSERT(grad_y_reduce_idy != nullptr);
+  const size_t n = reverse_x.size();
+  if (reverse_y.size() < n) {
+    MS_LOG_ERROR << "The size of reverse_y is less than the size of reverse_x.";
+  }
+  for (size_t i = 0; i < n; ++i) {
+    State curr = State::SAME;
+    const int x_i = reverse_x[i];
+    const int y_i = reverse_y[i];
+    const int reduce_idx = (n - 1 - i);
+    if (x_i == y_i) {
+      curr = State::SAME;
+    } else if (x_i == 1) {
+      grad_x_reduce_idx->push_back(reduce_idx);
+      curr = State::X_ONE;
+    } else if (y_i == 1) {
+      grad_y_reduce_idy->push_back(reduce_idx);
+      curr = State::Y_ONE;
+    } else {
+      MS_LOG_ERROR << "not compatible shape input for BroadcastGradientArgs";
+    }
+    if (curr == State::SAME && x_i == 1) {
+      grad_x_reduce_idx->push_back(reduce_idx);
+      grad_y_reduce_idy->push_back(reduce_idx);
+      continue;
+    }
+  }
+
+  std::reverse(grad_x_reduce_idx->begin(), grad_x_reduce_idx->end());
+  std::reverse(grad_y_reduce_idy->begin(), grad_y_reduce_idy->end());
+}
+
+std::vector<std::vector<int>> BroadcastGradientArgs::operator()() {
+  std::vector<std::vector<int>> input_dim(kInNum);
+  input_dim[0] = dim0_;
+  input_dim[1] = dim1_;
+  auto same_shape = CompareShape(dim0_, dim1_);
+  if (same_shape) {
+    return {{}, {}};
+  }
+
+  std::vector<int> reverse_x;
+  std::vector<int> reverse_y;
+
+  (void)std::transform(dim0_.rbegin(), dim0_.rend(), std::back_inserter(reverse_x), [](const int &v) { return v; });
+  (void)std::transform(dim1_.rbegin(), dim1_.rend(), std::back_inserter(reverse_y), [](const int &v) { return v; });
+
+  if (reverse_x.size() > reverse_y.size()) {
+    reverse_y.resize(reverse_x.size(), 1);
+  } else {
+    reverse_x.resize(reverse_y.size(), 1);
+  }
+
+  std::vector<int> grad_x_reduce_idx;
+  std::vector<int> grad_y_reduce_idy;
+  ComputeReduceIndex(reverse_x, reverse_y, &grad_x_reduce_idx, &grad_y_reduce_idy);
+  return {grad_x_reduce_idx, grad_y_reduce_idy};
+}
+
+void DynamicBroadcastGradientArgs::AddElementToGradReduceIdx(std::vector<std::vector<int>> *grad_reduce_idx,
+                                                             std::vector<bool> current_is_one, bool none_is_one,
+                                                             const size_t largest_rank, size_t j) {
+  for (size_t i = 0; i < kInNum; ++i) {
+    if (current_is_one[i] && !none_is_one) {
+      (void)(*grad_reduce_idx)[i].emplace_back(largest_rank - 1 - j);
+    }
+  }
+}
+
+void DynamicBroadcastGradientArgs::UpdatePreIsOne(std::vector<bool> *prev_is_one, std::vector<bool> current_is_one) {
+  for (size_t i = 0; i < kInNum; ++i) {
+    (*prev_is_one)[i] = current_is_one[i];
+  }
+}
+
+std::vector<std::vector<int>> DynamicBroadcastGradientArgs::GetGradientIndices(
+  const std::vector<std::vector<int>> &reverse_shape, const size_t largest_rank) {
+  std::vector<std::vector<int>> grad_reduce_idx(kInNum);
+  // indices of j-th component of each input.
+  std::vector<bool> prev_is_one(kInNum);
+  std::vector<bool> current_is_one(kInNum);
+  for (size_t i = 0; i < kInNum; ++i) {
+    prev_is_one[i] = false;
+    current_is_one[i] = false;
+  }
+
+  bool set_one = false;
+  for (size_t j = 0; j < largest_rank; ++j) {
+    int output_dim = -1;
+    bool output_dim_set = false;
+    bool none_is_one = true;
+    // Find which indices are 1.
+    for (size_t i = 0; i < kInNum; ++i) {
+      if (reverse_shape[i][j] == 1) {
+        current_is_one[i] = true;
+        none_is_one = false;
+      } else {
+        current_is_one[i] = false;
+        if (!output_dim_set || reverse_shape[i][j] == static_cast<int>(output_dim)) {
+          output_dim = reverse_shape[i][j];
+          output_dim_set = true;
+        } else {
+          std::cout << "Input[0] and input[1] Cannot broadcast!";
+        }
+      }
+    }
+    // All dimensions are 1.
+    if (!output_dim_set) {
+      for (size_t i = 0; i < kInNum; ++i) {
+        (void)grad_reduce_idx[i].emplace_back(largest_rank - 1 - j);
+      }
+      continue;
+    } else if (std::equal(current_is_one.begin(), current_is_one.end(), prev_is_one.begin()) && set_one) {
+      AddElementToGradReduceIdx(&grad_reduce_idx, current_is_one, none_is_one, largest_rank, j);
+    } else {
+      AddElementToGradReduceIdx(&grad_reduce_idx, current_is_one, none_is_one, largest_rank, j);
+    }
+    set_one = true;
+    UpdatePreIsOne(&prev_is_one, current_is_one);
+  }
+  return grad_reduce_idx;
+}
+
+std::vector<std::vector<int>> DynamicBroadcastGradientArgs::CalculateOutput(const std::vector<std::vector<int>> &x) {
+  std::vector<std::vector<int>> grad_reduce_idx(kInNum);
+  bool all_equal = true;
+  size_t largest_rank = 0;
+  for (size_t i = 0; i < kInNum; ++i) {
+    if (x[i] != x[0]) {
+      all_equal = false;
+    }
+    if (x[i].size() > largest_rank) {
+      largest_rank = x[i].size();
+    }
+  }
+  if (all_equal) {
+    return grad_reduce_idx;
+  }
+
+  // Reverse input the shapes
+  std::vector<std::vector<int>> reverse_shape(kInNum);
+  for (size_t i = 0; i < kInNum; ++i) {
+    reverse_shape[i] = x[i];
+    std::reverse(reverse_shape[i].begin(), reverse_shape[i].end());
+  }
+
+  // 1-extend and align all vectors.
+  for (size_t i = 0; i < kInNum; ++i) {
+    if (reverse_shape[i].size() < largest_rank) {
+      reverse_shape[i].resize(largest_rank, 1);
+    }
+  }
+  grad_reduce_idx = GetGradientIndices(reverse_shape, largest_rank);
+  return grad_reduce_idx;
+}
+
+std::vector<std::vector<int>> DynamicBroadcastGradientArgs::SetOutputValue(
+  const std::vector<std::vector<int>> &grad_reduce_idx, const std::vector<std::vector<int>> &input_dim) {
+  std::vector<std::vector<int>> output(kInNum);
+  for (size_t index = 0; index < kInNum; ++index) {
+    auto idx_num = grad_reduce_idx[index].size();
+    for (size_t k = 0; k < idx_num; ++k) {
+      output[index].push_back(grad_reduce_idx[index][idx_num - 1 - k]);
+    }
+    if (idx_num == 0) {
+      auto input_num = input_dim[index].size();
+      for (size_t k = 0; k < input_num; ++k) {
+        output[index].push_back(k);
+      }
+    }
+  }
+  return output;
+}
+
+std::vector<std::vector<int>> DynamicBroadcastGradientArgs::operator()() {
+  std::vector<std::vector<int>> input_dim(kInNum);
+  input_dim[0] = dim0_;
+  input_dim[1] = dim1_;
+  auto grad_reduce_idx = CalculateOutput(input_dim);
+  auto output = SetOutputValue(grad_reduce_idx, input_dim);
+  return output;
+}
+
+std::vector<int> VectorDiv::operator()(const std::vector<int> &x, const std::vector<int> &d) {
+  if (d.size() != x.size()) {
+    MS_LOG(ERROR) << "x and divider must have same size";
+    return {};
+  }
+  std::vector<int> res;
+  for (size_t i = 0; i < d.size(); i++) {
+    auto x_value = x.at(i);
+    auto d_value = d.at(i);
+    if (d_value == 0) {
+      MS_LOG(ERROR) << "Divisor is zero";
+      return {};
+    }
+    if ((x_value % d_value) != 0) {
+      MS_LOG(ERROR) << "x and d and not dividable";
+    }
+    auto r = x_value / d_value;
+    res.push_back(r);
+  }
+  return res;
+}
+
+std::vector<int> ShapeReduce::operator()(const std::vector<int> &x_shape, const std::vector<int> &axis) {
+  int x_rank = x_shape.size();
+  std::set<int> axis_set;
+
+  auto min = -x_rank;
+  auto max = x_rank - 1;
+  for (auto &elem : axis) {
+    if (elem > max || elem < min) {
+      MS_LOG(ERROR) << "illegal axis value";
+      return {};
+    }
+    axis_set.insert(elem);
+  }
+  std::vector<int> res;
+  for (int i = 0; i < x_rank; i++) {
+    if (axis_set.count(i) || axis_set.count(i - x_rank)) {
+      res.push_back(1);
+    } else {
+      res.push_back(x_shape.at(i));
+    }
+  }
+  return res;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/ops_utils.h

@@ -0,0 +1,69 @@
+/**
+ * Copyright 2022 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_EXPRESSION_OPS_UTILS_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_OPS_UTILS_H_
+
+#include "include/api/cfg.h"
+#include "src/expression/net.h"
+#include "vector"
+
+namespace mindspore {
+namespace lite {
+class BroadcastGradientArgs {
+ public:
+  BroadcastGradientArgs(const std::vector<int> &dim0, const std::vector<int> &dim1) : dim0_(dim0), dim1_(dim1) {}
+  std::vector<std::vector<int>> operator()();
+
+ private:
+  static const int kInNum = 2;
+  const std::vector<int> &dim0_;
+  const std::vector<int> &dim1_;
+};
+
+class DynamicBroadcastGradientArgs {
+ public:
+  DynamicBroadcastGradientArgs(const std::vector<int> &dim0, const std::vector<int> &dim1) : dim0_(dim0), dim1_(dim1) {}
+  std::vector<std::vector<int>> operator()();
+
+ private:
+  void AddElementToGradReduceIdx(std::vector<std::vector<int>> *grad_reduce_idx, std::vector<bool> current_is_one,
+                                 bool none_is_one, const size_t largest_rank, size_t j);
+  void UpdatePreIsOne(std::vector<bool> *prev_is_one, std::vector<bool> current_is_one);
+  std::vector<std::vector<int>> GetGradientIndices(const std::vector<std::vector<int>> &reverse_shape,
+                                                   const size_t largest_rank);
+  std::vector<std::vector<int>> CalculateOutput(const std::vector<std::vector<int>> &x);
+  std::vector<std::vector<int>> SetOutputValue(const std::vector<std::vector<int>> &grad_reduce_idx,
+                                               const std::vector<std::vector<int>> &input_dim);
+  static const int kInNum = 2;
+  const std::vector<int> &dim0_;
+  const std::vector<int> &dim1_;
+};
+
+class VectorDiv {
+ public:
+  VectorDiv() {}
+  std::vector<int> operator()(const std::vector<int> &x, const std::vector<int> &d);
+};
+
+class ShapeReduce {
+ public:
+  ShapeReduce() {}
+  std::vector<int> operator()(const std::vector<int> &x_shape, const std::vector<int> &axis);
+};
+}  // namespace lite
+}  // namespace mindspore
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_OPS_UTILS_H_

mindspore/lite/src/expression/param.cc

@@ -0,0 +1,70 @@
+/**
+ * Copyright 2022 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/expression/param.h"
+#include <random>
+#include <algorithm>
+#include <string>
+#include "include/errorcode.h"
+
+using mindspore::lite::RET_ERROR;
+using mindspore::lite::RET_OK;
+
+constexpr float kZero = 0.0f;
+constexpr float kOne = 1.0f;
+
+namespace mindspore {
+namespace lite {
+int Param::Fill(Mode mode) {
+  std::default_random_engine engine{static_cast<unsigned int>(0)};
+  std::vector<float> data(size_);
+  switch (mode) {
+    case NORMAL: {
+      constexpr float scale = 0.01;
+      std::normal_distribution<float> n{0, 1};
+      std::generate_n(data.begin(), size_, [&]() { return n(engine); });
+      (void)std::transform(data.begin(), data.end(), data.begin(), [](float x) { return x * scale; });
+      break;
+    }
+    case UNIFORM: {
+      constexpr float scale = 0.07;
+      std::uniform_real_distribution<float> u{-1.0, 1.0};
+      std::generate_n(data.begin(), size_, [&]() { return u(engine) * scale; });
+      break;
+    }
+    case ZEROS:
+      std::fill_n(data.begin(), size_, kZero);
+      break;
+    case ONES:
+      std::fill_n(data.begin(), size_, kOne);
+      break;
+    case NOT_SUPPORTED:
+      return RET_ERROR;
+  }
+  Copy(data);
+  return RET_OK;
+}
+
+Param::Mode Param::String2Enum(std::string mode) {
+  (void)std::transform(mode.begin(), mode.end(), mode.begin(), ::tolower);
+  if (mode == "normal") return NORMAL;
+  if (mode == "uniform") return UNIFORM;
+  if (mode == "ones") return ONES;
+  if (mode == "zeors") return ZEROS;
+  return NOT_SUPPORTED;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/param.h

@@ -0,0 +1,60 @@
+/**
+ * Copyright 2022 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_EXPRESSION_PARAM_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_PARAM_H_
+
+#include <vector>
+#include <iostream>
+#include <fstream>
+#include <string>
+
+namespace mindspore {
+namespace lite {
+class Param {
+ public:
+  enum Mode { NORMAL, UNIFORM, ONES, ZEROS, NOT_SUPPORTED };
+  int Fill(Mode type);
+  static Mode String2Enum(std::string);
+  std::vector<uint8_t> &data() { return data_; }
+  size_t Load(std::string file_name, size_t offset = 0) { return data_.size() * sizeof(float); }
+  size_t Load(std::ifstream &s, int offset = 0) { return data_.size() * sizeof(float); }
+  void SetSize(size_t size) { size_ = size; }
+  template <typename T>
+  void Copy(const T *data, size_t size) {
+    auto cast_data = reinterpret_cast<const uint8_t *>(data);
+    data_ = decltype(data_)(cast_data, cast_data + size * sizeof(T) / sizeof(uint8_t));
+  }
+  template <typename T>
+  void Copy(const std::vector<T> data) {
+    Copy<T>(data.data(), data.size());
+  }
+
+  template <typename T>
+  std::vector<T> Extract() {
+    T *num = reinterpret_cast<T *>(data_.data());
+    std::vector<T> res(num, num + data_.size() / sizeof(T));
+    return res;
+  }
+
+ private:
+  size_t size_;
+  std::vector<uint8_t> data_;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_PARAM_H_

mindspore/lite/src/expression/sequential.cc

@@ -0,0 +1,30 @@
+/**
+ * Copyright 2022 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/expression/sequential.h"
+
+namespace mindspore {
+namespace lite {
+void Sequential::Add(Node *node) { PushOp(node); }
+
+std::vector<EXPR *> Sequential::construct(const std::vector<EXPR *> &inputs) {
+  auto x = inputs;
+  for (auto &node : ops_) {
+    x = (*node)({x.front()});
+  }
+  return x;
+}
+}  // namespace lite
+}  // namespace mindspore

mindspore/lite/src/expression/sequential.h

@@ -0,0 +1,32 @@
+/**
+ * Copyright 2022 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_EXPRESSION_SEQUENTIAL_H_
+#define MINDSPORE_LITE_SRC_EXPRESSION_SEQUENTIAL_H_
+#include <vector>
+#include "src/expression/net.h"
+
+namespace mindspore {
+namespace lite {
+class Sequential : public Net {
+ public:
+  std::vector<EXPR *> construct(const std::vector<EXPR *> &inputs) override;
+  void Add(Node *node) override;
+};
+}  // namespace lite
+}  // namespace mindspore
+
+#endif  // MINDSPORE_LITE_SRC_EXPRESSION_SEQUENTIAL_H_

mindspore/lite/src/runtime/kernel/cpu/fp32/convolution_1x1_fp32.cc

@@ -25,6 +25,7 @@ using mindspore::lite::RET_OK;
 namespace mindspore::kernel {
 Convolution1x1CPUKernel::~Convolution1x1CPUKernel() {
   FreeTmpBuffer();
+
   if (matmul_param_ != nullptr) {
     delete matmul_param_;
     matmul_param_ = nullptr;
@@ -185,7 +186,7 @@ int Convolution1x1CPUKernel::DoConv1x1Hw(int task_id) {
 
   float *thread_input_ptr = input_ptr_ + task_id * thread_stride_ * matmul_param_->deep_;
   float *thread_pack_input = pack_input_ + task_id * row_tile_ * matmul_param_->deep_;
-  float *thread_output_ptr;
+  float *thread_output_ptr = nullptr;
   if (out_tensors()[0]->format() != NC4HW4) {
     thread_output_ptr = output_ptr_ + task_id * thread_stride_ * matmul_param_->col_;
   } else {

mindspore/lite/src/runtime/kernel/cpu/fp32/convolution_depthwise_fp32.cc

@@ -127,10 +127,12 @@ int ConvolutionDepthwiseCPUKernel::MallocWeightBiasData() {
     }
   }
   CHECK_LESS_RETURN(MAX_MALLOC_SIZE, channel * sizeof(float));
-  bias_data_ = malloc(channel * sizeof(float));
   if (bias_data_ == nullptr) {
-    MS_LOG(ERROR) << "Malloc buffer failed.";
-    return RET_ERROR;
+    bias_data_ = malloc(channel * sizeof(float));
+    if (bias_data_ == nullptr) {
+      MS_LOG(ERROR) << "Malloc buffer failed.";
+      return RET_ERROR;
+    }
   }
   memset(bias_data_, 0, channel * sizeof(float));
   return RET_OK;

mindspore/lite/test/config_level0/cropped_size.cfg

@@ -1,2 +1,2 @@
 Note: This is the mindspore Lite inference framework size threshold. Offline review is required before modify this value!!!
-1034600
+1040696

mindspore/lite/test/config_level0/models_ms_train.cfg

@@ -48,5 +48,7 @@ vae
 unified_api code_example
 train_lenet code_example
 train_lenet_java code_example
+lenet expression
+mobilenetv2 expression noarm32
 # LAST
 #test_resize inputShapes 16,10,10,1:16,10,10,1 0.5

mindspore/lite/test/config_level1/cropped_size.cfg

@@ -1,2 +1,2 @@
 Note: This is the mindspore Lite inference framework size threshold. Modifying this threshold requires meeting review.
-1034600
+1040696

mindspore/lite/test/st/scripts/run_net_train.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 source ./scripts/base_functions.sh
-version=1.5.0
+version=1.6.1
 
 # Run Export on x86 platform and create output test files:
 docker_image=mindspore_build:210301
@@ -69,6 +69,9 @@ function Run_Converter() {
             model_prefix="${line_array[0]}_head"
             model_name=${line_array[0]}'_head'
         fi
+        if [[ "${expression}" == "1" ]]; then
+            model_prefix="${model_name}_fwd"
+        fi
         if [[ ${check_convert} == "1" ]]; then
             ms_file=$ms_models_path'/'$model_name'.ms'
             if [ -f "$ms_file" ]; then
@@ -80,7 +83,7 @@ function Run_Converter() {
         {
           echo ${model_name} >> "${run_converter_log_file}"
           echo './converter_lite  --fmk=MINDIR --modelFile='${models_path}'/'${model_prefix}'.mindir --outputFile='${ms_models_path}'/'${model_name}' --trainModel=true' ${WEIGHT_QUANT} >> "${run_converter_log_file}"
-          ./converter_lite --fmk=MINDIR --modelFile=${models_path}/${model_prefix}.mindir --outputFile=${ms_models_path}/${model_name} --trainModel=true ${WEIGHT_QUANT}
+          ./converter_lite --fmk=MINDIR --modelFile=${models_path}/${model_prefix}.mindir --outputFile=${ms_models_path}/${model_name} --trainModel=true ${no_opt} ${WEIGHT_QUANT}
           if [ $? = 0 ]; then
               converter_result='converter mindspore '${model_name}' pass';echo ${converter_result} >> ${run_converter_result_file}
           else
@@ -118,9 +121,12 @@ function should_run_example() {
       continue
     fi
     if [[ $model_name == "$1" ]]; then
-      if [[ ${line_array[1]} == "code_example" ]]; then
+      if [[ "${line_array[1]}" == "code_example" ]]; then
         ret=1
       fi
+      if [[ "${line_array[2]}" == "expression" ]]; then
+        ret=2
+      fi
     fi
   done < ${models_ms_train_config}
   return $ret
@@ -146,6 +152,9 @@ function parse_line() {
     enable_transfer=0
     suffix_print=""
     check_convert=0
+    do_api=false
+    no_opt="--NoFusion=false"
+    expression=0
     model_name=${line_array[0]}_train
     while [[ $i < ${#line_array[@]} ]] ; do
         case ${line_array[i]} in
@@ -189,6 +198,12 @@ function parse_line() {
             i=$(($i+1))
             inputShapes=${line_array[i]}
             ;;
+          "expression")
+            model_name="${line_array[0]}_expr"
+            do_api=true
+            no_opt="--NoFusion=true"
+            expression=1
+            ;;
           *)
             check=`echo "${line_array[i]}" | grep -E '^\-?[0-9]*\.?[0-9]+$'`
             if [ "${check}" != "" ] ; then
@@ -214,12 +229,14 @@ function Run_x86() {
           continue
         fi
         local model_prefix=${line_array[0]}
+        local bb_model_file=""
         local log_suffix="_train"
         parse_line x86
         if [[ "$?" == "1" ]]; then continue; fi
+        if [[ "${expression}" == "1" ]]; then
+            model_prefix=${model_prefix}_expr
+        fi
         local model_file="${ms_models_path}/${model_name}.ms"
-        local bb_model_file=""
-        local export_file="${ms_models_path}/${model_name}_tod"
         local inference_file="${ms_models_path}/${model_name}_infer"
         if [[ "${enable_transfer}" == "1" ]]; then
             model_file="${ms_models_path}/${model_prefix}_head.ms"
@@ -247,7 +264,8 @@ function Run_x86() {
             --exportFile=${export_file} \
             --virtualBatch=${virtual_batch} \
             --inputShapes=${inputShapes} \
-            --lossName=${loss_name} " >> ${run_x86_log_file}
+            --lossName=${loss_name}  \
+            --unifiedApi=${do_api}" >> ${run_x86_log_file}
         ${run_valgrind} ./tools/benchmark_train/benchmark_train \
             --modelFile=${model_file} \
             --bbModelFile=${bb_model_file} \
@@ -257,7 +275,8 @@ function Run_x86() {
             --exportFile=${export_file} \
             --virtualBatch=${virtual_batch} \
             --inputShapes=${inputShapes} \
-            --lossName=${loss_name} >> "${run_x86_log_file}"
+            --lossName=${loss_name} \
+            --unifiedApi=${do_api} >> "${run_x86_log_file}"
         if [ $? = 0 ]; then
             run_result='x86'${log_suffix}': '${model_name}''${suffix_print}' pass'; echo ${run_result} >> ${run_benchmark_train_result_file}
         else
@@ -327,24 +346,27 @@ function Run_arm() {
     echo "cd ${tmp_dir}" > ${adb_cmd_file}
     echo 'chmod 777 benchmark_train' >> ${adb_cmd_file}
     adb -s ${device_id} shell < ${adb_cmd_file}
-
     local fail=0
     # Run mindir converted train models:
     while read line; do
         local line_array
+        local bb_model_file=""
         LFS=" " read -r -a line_array <<< ${line}
         if [[ ${line_array[0]} == \#* || ${line_array[0]} == "" ]]; then
           continue
         fi
+        parse_line $1
+        if [[ "$?" == "1" ]]; then continue; fi
         local model_prefix=${line_array[0]}
+        if [[ ${expression} == "1" ]]; then
+          model_prefix=${model_prefix}_expr
+        fi
         local run_result=""
         local log_suffix="_train"
-        parse_line $1
         if [[ "$?" == "1" ]]; then continue; fi
         local export_file="${tmp_dir}/${model_name}_tod"
         local inference_file="${tmp_dir}/${model_name}_infer"
         local model_file="${model_name}.ms"
-        local bb_model_file=""
         if [[ "${enable_transfer}" == "1" ]]; then
             model_file="${model_prefix}_head.ms"
             bb_model_file="${model_prefix}_bb.ms"
@@ -388,7 +410,8 @@ function Run_arm() {
         --exportFile=${export_file} \
         --virtualBatch=${virtual_batch} \
         --inputShapes=${inputShapes} \
-        --lossName=${loss_name}
+        --lossName=${loss_name} \
+        --unifiedApi=${do_api}
 ENDM
 )
         echo "${adb_cmd}" >> ${run_arm_log_file}
@@ -443,11 +466,16 @@ function Run_CodeExamples() {
 
       should_run_example "unified_api"
       should_run=$?
+      local expression_flag=""
+      if [[ "$should_run" == "2" ]]; then
+        expression_flag="-x"
+        should_run=1
+      fi
       if [[ "$should_run" == "1" ]]; then
         cd ${basepath}/../../examples/unified_api || exit 1
         chmod 777 ./prepare_and_run.sh
         chmod 777 ./*/*.sh
-        ./prepare_and_run.sh -D ${datasets_path}/mnist -r ${tarball_path} -t ${target} -m ${models_path}/code_example.mindir -e 1 >> ${run_code_examples_log_file}
+        ./prepare_and_run.sh -D ${datasets_path}/mnist -r ${tarball_path} -t ${target} -m ${models_path}/code_example.mindir -e 1 ${expression_flag} >> ${run_code_examples_log_file}
         if [ "$?" != "0" ]; then
           echo "Unified API prepare_and_run.sh failed"
           exit 1
@@ -492,6 +520,10 @@ while getopts "r:c:m:d:i:e:vt:q:D:M:l:" opt; do
             models_path=${OPTARG}
             echo "models_path is ${OPTARG}"
             ;;
+         c)
+           models_ms_train_config=${OPTARG}
+           echo  "models_ms_train_config ${models_ms_train_config}"
+           ;;
         i)
             train_io_path=${OPTARG}
             echo "train_io_path is ${OPTARG}"
@@ -539,7 +571,9 @@ config_folder="config_level0"
 if [[ ${level} == "level1" ]];then
     config_folder="config_level1"
 fi
-models_ms_train_config=${basepath}/../${config_folder}/models_ms_train.cfg
+if [[ "${models_ms_train_config}" == "" ]]; then
+    models_ms_train_config=${basepath}/../${config_folder}/models_ms_train.cfg
+fi
 
 if [[ $train_io_path == "" ]]; then
   echo "train_io path is empty"

mindspore/lite/tools/benchmark_train/CMakeLists.txt

@@ -5,10 +5,23 @@ set(COMMON_SRC
         ${CMAKE_CURRENT_SOURCE_DIR}/../../src/common/utils.cc
         )
 
+
+set(TEST_SRC
+    ${CMAKE_CURRENT_SOURCE_DIR}/main.cc
+    ${CMAKE_CURRENT_SOURCE_DIR}/net_train.cc
+    )
+
+if(MSLITE_MINDDATA_IMPLEMENT STREQUAL "full")
+    set(TEST_SRC
+    ${TEST_SRC}
+    ${CMAKE_CURRENT_SOURCE_DIR}/net_runner.cc
+    )
+endif()
+
 add_executable(benchmark_train
-        ${CMAKE_CURRENT_SOURCE_DIR}/main.cc
-        ${CMAKE_CURRENT_SOURCE_DIR}/net_train.cc
+        ${TEST_SRC}
         ${COMMON_SRC})
+
 if(WIN32)
     add_dependencies(benchmark_train fbs_src mindspore-lite_static mindspore-lite-train_static)
 else()

mindspore/lite/tools/benchmark_train/net_runner.cc

@@ -0,0 +1,371 @@
+/**
+ * Copyright 2022 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "tools/benchmark_train/net_runner.h"
+#include "tools/benchmark_train/net_train.h"
+#include <getopt.h>
+#include <malloc.h>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <iostream>
+#include <fstream>
+#include <utility>
+#include <chrono>
+#include "include/api/types.h"
+#include "include/api/context.h"
+#include "include/api/serialization.h"
+#include "include/api/callback/loss_monitor.h"
+#include "include/api/metrics/accuracy.h"
+#include "include/api/callback/ckpt_saver.h"
+#include "include/api/callback/train_accuracy.h"
+#include "include/api/callback/lr_scheduler.h"
+#include "include/dataset/datasets.h"
+#include "include/dataset/vision_lite.h"
+#include "include/dataset/transforms.h"
+#include "include/api/cfg.h"
+#include "include/api/net.h"
+
+using mindspore::AccuracyMetrics;
+using mindspore::Model;
+using mindspore::TrainAccuracy;
+using mindspore::TrainCallBack;
+using mindspore::TrainCallBackData;
+using mindspore::dataset::Dataset;
+using mindspore::dataset::Mnist;
+using mindspore::dataset::SequentialSampler;
+using mindspore::dataset::TensorOperation;
+using mindspore::dataset::transforms::TypeCast;
+using mindspore::dataset::vision::Normalize;
+using mindspore::dataset::vision::Resize;
+
+constexpr int kNCHWCDim = 2;
+constexpr int kPrintTimes = 100;
+constexpr float kBetta1 = 0.9f;
+constexpr float kBetta2 = 0.999f;
+
+class Rescaler : public mindspore::TrainCallBack {
+ public:
+  explicit Rescaler(float scale) : scale_(scale) {
+    if (scale_ == 0) {
+      scale_ = 1.0;
+    }
+  }
+  ~Rescaler() override = default;
+  void StepBegin(const mindspore::TrainCallBackData &cb_data) override {
+    auto inputs = cb_data.model_->GetInputs();
+    auto *input_data = reinterpret_cast<float *>(inputs.at(0).MutableData());
+    for (int k = 0; k < inputs.at(0).ElementNum(); k++) input_data[k] /= scale_;
+  }
+
+ private:
+  float scale_ = 1.0;
+};
+
+/* This is an example of a user defined Callback to measure memory and latency of execution */
+class Measurement : public mindspore::TrainCallBack {
+ public:
+  explicit Measurement(unsigned int epochs)
+      : time_avg_(std::chrono::duration<double, std::milli>(0)), epochs_(epochs) {}
+  ~Measurement() override = default;
+  void EpochBegin(const mindspore::TrainCallBackData &cb_data) override {
+    start_time_ = std::chrono::high_resolution_clock::now();
+  }
+  mindspore::CallbackRetValue EpochEnd(const mindspore::TrainCallBackData &cb_data) override {
+    end_time_ = std::chrono::high_resolution_clock::now();
+    auto time = std::chrono::duration<double, std::milli>(end_time_ - start_time_);
+    time_avg_ += time;
+    return mindspore::kContinue;
+  }
+  void End(const mindspore::TrainCallBackData &cb_data) override {
+    if (epochs_ > 0) {
+      std::cout << "AvgRunTime: " << time_avg_.count() / epochs_ << " ms" << std::endl;
+    }
+
+    struct mallinfo info = mallinfo();
+    std::cout << "Total allocation: " << info.arena + info.hblkhd << std::endl;
+  }
+
+ private:
+  std::chrono::time_point<std::chrono::high_resolution_clock> start_time_;
+  std::chrono::time_point<std::chrono::high_resolution_clock> end_time_;
+  std::chrono::duration<double, std::milli> time_avg_;
+  unsigned int epochs_;
+};
+
+NetRunner::~NetRunner() {
+  if (model_ != nullptr) {
+    delete model_;
+  }
+  if (graph_ != nullptr) {
+    delete graph_;
+  }
+}
+
+mindspore::Status NetRunner::InitAndFigureInputs() {
+  auto context = std::make_shared<mindspore::Context>();
+  auto cpu_context = std::make_shared<mindspore::CPUDeviceInfo>();
+  cpu_context->SetEnableFP16(enable_fp16_);
+  context->MutableDeviceInfo().push_back(cpu_context);
+
+  graph_ = new (std::nothrow) mindspore::Graph(mindspore::Graph::Type::kExpressionGraph);
+  if (graph_ == nullptr) {
+    std::cout << "Cannot allocate graph" << std::endl;
+    return mindspore::kLiteMemoryFailed;
+  }
+  auto status = mindspore::Serialization::Load(ms_file_, mindspore::kMindIR, graph_);
+  if (status != mindspore::kSuccess) {
+    std::cout << "Error " << status << " during serialization of graph " << ms_file_;
+    return status;
+  }
+  auto net = std::make_unique<mindspore::Net>(*graph_);
+  auto input_shape = net->InputShape(0);
+  auto label_shape = net->OutputShape(0);
+  auto inputM = mindspore::NN::Input(input_shape);
+  auto labelM = mindspore::NN::Input(label_shape);
+  auto label = labelM->Create("label");
+  auto input = inputM->Create("input");
+
+  auto cfg = std::make_shared<mindspore::TrainCfg>();
+  if (enable_fp16_) {
+    cfg.get()->optimization_level_ = mindspore::kO2;
+  }
+
+  model_ = new (std::nothrow) mindspore::Model();
+  if (model_ == nullptr) {
+    std::cout << "model allocation failed" << std::endl;
+    return mindspore::kLiteMemoryFailed;
+  }
+  mindspore::SoftMaxCrossEntropyCfg softmax_ce_cfg;
+  softmax_ce_cfg.reduction = "none";
+  auto netWithLoss = mindspore::NN::GraphWithLoss(graph_, mindspore::NN::SoftmaxCrossEntropy(softmax_ce_cfg));
+  mindspore::AdamConfig AdamCfg;
+  AdamCfg.beta1_ = kBetta1;
+  AdamCfg.beta2_ = kBetta2;
+  AdamCfg.eps_ = 1e-8;
+  AdamCfg.learning_rate_ = 1e-2;
+  auto optimizer = mindspore::NN::Adam(net->trainable_params(), AdamCfg);
+  status = model_->Build(mindspore::GraphCell(*netWithLoss), optimizer, {input, label}, context, cfg);
+  if (status != mindspore::kSuccess) {
+    std::cout << "Error " << status << " during build of model " << ms_file_ << std::endl;
+    return status;
+  }
+  delete graph_;
+  graph_ = nullptr;
+  auto inputs = model_->GetInputs();
+  if (inputs.size() < 1) {
+    return mindspore::kLiteError;
+  }
+  auto nhwc_input_dims = inputs.at(0).Shape();
+  batch_size_ = nhwc_input_dims.at(0);
+  h_ = nhwc_input_dims.at(1);
+  w_ = nhwc_input_dims.at(kNCHWCDim);
+  return mindspore::kSuccess;
+}
+
+int NetRunner::CompareOutput(const std::vector<mindspore::MSTensor> &outputs) {
+  std::cout << "================ Comparing Forward Output data ================" << std::endl;
+  float total_bias = 0;
+  int total_size = 0;
+  bool has_error = false;
+  int i = 1;
+  for (auto &tensor : outputs) {
+    std::cout << "output is tensor " << tensor.Name() << "\n";
+    auto output = tensor.Data();
+    size_t size;
+    std::string output_file = flags_->data_file_ + std::to_string(i) + ".bin";
+    auto bin_buf = std::unique_ptr<float[]>(mindspore::lite::NetTrain::ReadFileBuf(output_file.c_str(), &size));
+    if (bin_buf == nullptr) {
+      MS_LOG(ERROR) << "ReadFile return nullptr";
+      std::cout << "ReadFile return nullptr" << std::endl;
+      return mindspore::kLiteNullptr;
+    }
+    if (size != tensor.DataSize()) {
+      MS_LOG(ERROR) << "Output buffer and output file differ by size. Tensor size: " << tensor.DataSize()
+                    << ", read size: " << size;
+      std::cout << "Output buffer and output file differ by size. Tensor size: " << tensor.DataSize()
+                << ", read size: " << size << std::endl;
+      return mindspore::kLiteError;
+    }
+    float bias = mindspore::lite::NetTrain::CompareData<float>(bin_buf.get(), tensor.ElementNum(),
+                                                               reinterpret_cast<const float *>(output.get()));
+    if (bias >= 0) {
+      total_bias += bias;
+      total_size++;
+    } else {
+      has_error = true;
+      break;
+    }
+    i++;
+  }
+
+  if (!has_error) {
+    float mean_bias;
+    if (total_size != 0) {
+      mean_bias = total_bias / total_size * kPrintTimes;
+    } else {
+      mean_bias = 0;
+    }
+
+    std::cout << "Mean bias of all nodes/tensors: " << mean_bias << "%"
+              << " threshold is:" << this->flags_->accuracy_threshold_ << std::endl;
+    std::cout << "=======================================================" << std::endl << std::endl;
+
+    if (mean_bias > this->flags_->accuracy_threshold_) {
+      MS_LOG(INFO) << "Mean bias of all nodes/tensors is too big: " << mean_bias << "%";
+      std::cout << "Mean bias of all nodes/tensors is too big: " << mean_bias << "%" << std::endl;
+      return mindspore::kLiteError;
+    } else {
+      return mindspore::kSuccess;
+    }
+  } else {
+    MS_LOG(ERROR) << "Error in CompareData";
+    std::cout << "Error in CompareData" << std::endl;
+    std::cout << "=======================================================" << std::endl << std::endl;
+    return mindspore::kSuccess;
+  }
+}
+
+void NetRunner::CheckSum(const mindspore::MSTensor &tensor, std::string node_type, int id, std::string in_out) {
+  constexpr int kPrintLen = 4;
+  int tensor_size = tensor.ElementNum();
+  const void *data = tensor.Data().get();
+  const float *fdata = reinterpret_cast<const float *>(data);
+  mindspore::DataType type = tensor.DataType();
+  std::cout << node_type << " " << in_out << id << std::endl;
+  std::cout << "tensor name: " << tensor.Name() << std::endl;
+  if ((tensor_size) == 0 || (data == nullptr)) {
+    std::cout << "Empty tensor" << std::endl;
+    return;
+  }
+  switch (type) {
+    case mindspore::DataType::kNumberTypeFloat32:
+      std::cout << "sum=" << mindspore::lite::TensorSum<float>(data, tensor_size) << std::endl;
+      std::cout << "data: ";
+      for (int i = 0; i <= kPrintLen && i < tensor_size; i++) {
+        std::cout << static_cast<float>(fdata[i]) << ", ";
+      }
+      std::cout << std::endl;
+      break;
+    case mindspore::DataType::kNumberTypeInt32:
+      std::cout << "sum=" << mindspore::lite::TensorSum<int>(data, tensor_size) << std::endl;
+      break;
+    default:
+      std::cout << "unsupported type:" << static_cast<int>(type) << std::endl;
+      break;
+  }
+}
+
+int NetRunner::InitCallbackParameter() {
+  // after callback
+  after_call_back_ = [&](const std::vector<mindspore::MSTensor> &after_inputs,
+                         const std::vector<mindspore::MSTensor> &after_outputs,
+                         const mindspore::MSCallBackParam &call_param) {
+    if (after_inputs.empty()) {
+      MS_LOG(INFO) << "The num of after inputs is empty";
+    }
+    if (after_outputs.empty()) {
+      MS_LOG(INFO) << "The num of after outputs is empty";
+    }
+    if (flags_->layer_checksum_) {
+      for (size_t i = 0; i < after_inputs.size(); i++) {
+        CheckSum(after_inputs.at(i), call_param.node_type, i, "in");
+      }
+      for (size_t i = 0; i < after_outputs.size(); i++) {
+        CheckSum(after_outputs.at(i), call_param.node_type, i, "out");
+      }
+      std::cout << std::endl;
+    }
+    return true;
+  };
+  return false;
+}
+
+int NetRunner::RunOnce() {
+  auto inputs = model_->GetInputs();
+  std::vector<mindspore::MSTensor> output;
+  auto status = LoadInput(&inputs);
+  if (status != mindspore::kSuccess) {
+    std::cout << "cannot load data";
+    return status;
+  }
+  model_->SetTrainMode(true);
+  model_->RunStep(nullptr, nullptr);
+  model_->SetTrainMode(false);
+  model_->Predict(inputs, &output, nullptr, nullptr);
+  return CompareOutput(output);
+}
+
+int NetRunner::LoadInput(std::vector<mindspore::MSTensor> *ms_inputs) {
+  auto status = ReadInputFile(ms_inputs);
+  if (status != mindspore::kSuccess) {
+    std::cout << "Read Input File error, " << status << std::endl;
+    MS_LOG(ERROR) << "Read Input File error, " << status;
+    return status;
+  }
+  return mindspore::kSuccess;
+}
+
+int NetRunner::ReadInputFile(std::vector<mindspore::MSTensor> *ms_inputs) {
+  if (ms_inputs->empty()) {
+    std::cout << "no inputs to input" << std::endl;
+    return mindspore::kLiteError;
+  }
+  for (size_t i = 0; i < ms_inputs->size(); i++) {
+    auto cur_tensor = ms_inputs->at(i);
+    if (cur_tensor == nullptr) {
+      std::cout << "empty tensor " << i << std::endl;
+      MS_LOG(ERROR) << "empty tensor " << i;
+    }
+    size_t size;
+    std::string file_name = flags_->in_data_file_ + std::to_string(i + 1) + ".bin";
+    auto bin_buf = std::unique_ptr<float[]>(mindspore::lite::NetTrain::ReadFileBuf(file_name.c_str(), &size));
+    if (bin_buf == nullptr) {
+      MS_LOG(ERROR) << "ReadFile return nullptr";
+      std::cout << "ReadFile return nullptr" << std::endl;
+      return mindspore::kLiteNullptr;
+    }
+    auto tensor_data_size = cur_tensor.DataSize();
+    if (size != tensor_data_size) {
+      std::cout << "Input binary file size error, required: " << tensor_data_size << ", in fact: " << size
+                << " ,file_name: " << file_name.c_str() << std::endl;
+      MS_LOG(ERROR) << "Input binary file size error, required: " << tensor_data_size << ", in fact: " << size
+                    << " ,file_name: " << file_name.c_str();
+      return mindspore::kLiteError;
+    }
+    auto input_data = cur_tensor.MutableData();
+    memcpy(input_data, bin_buf.get(), tensor_data_size);
+  }
+  return mindspore::kSuccess;
+}
+
+int NetRunner::Main() {
+  ms_file_ = flags_->model_file_;
+  InitCallbackParameter();
+  auto status = InitAndFigureInputs();
+  if (status != mindspore::kSuccess) {
+    std::cout << "failed to initialize network" << std::endl;
+    return status.StatusCode();
+  }
+  return RunOnce();
+}
+
+int CallBack(mindspore::lite::NetTrainFlags *flags) {
+  NetRunner nr(flags);
+  return nr.Main();
+}
+
+int init = mindspore::lite::NetTrain::SetNr(CallBack);

mindspore/lite/tools/benchmark_train/net_runner.h

@@ -0,0 +1,81 @@
+/**
+ * Copyright 2022 Huawei Technologies Co., Ltd
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef MINDSPORE_LITE_TOOLS_BENCHMARK_TRAIN_NET_RUNNER_H_
+#define MINDSPORE_LITE_TOOLS_BENCHMARK_TRAIN_NET_RUNNER_H_
+
+#include <tuple>
+#include <iomanip>
+#include <map>
+#include <vector>
+#include <memory>
+#include <string>
+#include "include/api/model.h"
+#include "include/api/graph.h"
+#include "include/api/types.h"
+#include "include/api/status.h"
+#include "include/api/metrics/accuracy.h"
+#include "include/dataset/datasets.h"
+
+using mindspore::AccuracyMetrics;
+using mindspore::dataset::Dataset;
+
+namespace mindspore::lite {
+class NetTrainFlags;
+}
+
+class NetRunner {
+ public:
+  int Main();
+  explicit NetRunner(mindspore::lite::NetTrainFlags *flags) : flags_(flags) {}
+  bool ReadArgs(int argc, int8_t *argv[]);
+  virtual ~NetRunner();
+
+ private:
+  void Usage();
+  mindspore::Status InitAndFigureInputs();
+  void CheckSum(const mindspore::MSTensor &tensor, std::string node_type, int id, std::string in_out);
+  int InitCallbackParameter();
+  int TrainLoop();
+  float CalculateAccuracy(int max_tests = 0);
+  float GetLoss() const;
+  int RunOnce();
+  int CompareOutput(const std::vector<mindspore::MSTensor> &outputs);
+  int LoadInput(std::vector<mindspore::MSTensor> *ms_inputs);
+  int ReadInputFile(std::vector<mindspore::MSTensor> *ms_inputs);
+
+  mindspore::Model *model_ = nullptr;
+  mindspore::Graph *graph_ = nullptr;
+
+  std::shared_ptr<Dataset> train_ds_;
+  std::shared_ptr<Dataset> test_ds_;
+  std::shared_ptr<AccuracyMetrics> acc_metrics_;
+
+  std::string ms_file_ = "";
+  std::string data_dir_ = "";
+  unsigned int epochs_ = 10;
+  bool verbose_ = false;
+  bool enable_fp16_ = false;
+  int virtual_batch_ = -1;
+  int save_checkpoint_ = 0;
+  int batch_size_ = 32;
+  int h_ = 32;
+  int w_ = 32;
+  mindspore::lite::NetTrainFlags *flags_{nullptr};
+  mindspore::MSKernelCallBack after_call_back_;
+};
+
+#endif  // MINDSPORE_LITE_TOOLS_BENCHMARK_TRAIN_NET_RUNNER_H_

mindspore/lite/tools/benchmark_train/net_train.cc

@@ -24,6 +24,7 @@
 #ifdef ENABLE_NEON
 #include <arm_neon.h>
 #endif
+#include "tools/benchmark_train/net_runner.h"
 #include "src/common/common.h"
 #include "include/ms_tensor.h"
 #include "include/context.h"
@@ -49,14 +50,20 @@ constexpr int kCPUBindFlag2 = 2;
 constexpr int kCPUBindFlag1 = 1;
 static const int kTHOUSAND = 1000;
 
-namespace {
-float *ReadFileBuf(const char *file, size_t *size) {
-  if (file == nullptr) {
+std::function<int(NetTrainFlags *)> NetTrain::nr_cb_ = nullptr;
+
+int NetTrain::SetNr(std::function<int(NetTrainFlags *)> param) {
+  nr_cb_ = param;
+  return 0;
+}
+
+float *NetTrain::ReadFileBuf(const std::string file, size_t *size) {
+  if (file.empty()) {
     MS_LOG(ERROR) << "file is nullptr";
     return nullptr;
   }
   MS_ASSERT(size != nullptr);
-  std::string real_path = RealPath(file);
+  std::string real_path = RealPath(file.c_str());
   std::ifstream ifs(real_path);
   if (!ifs.good()) {
     MS_LOG(ERROR) << "file: " << real_path << " is not exist";
@@ -83,7 +90,6 @@ float *ReadFileBuf(const char *file, size_t *size) {
 
   return buf.release();
 }
-}  // namespace
 
 int NetTrain::GenerateRandomData(mindspore::tensor::MSTensor *tensor) {
   auto input_data = tensor->MutableData();
@@ -832,7 +838,9 @@ int RunNetTrain(int argc, const char **argv) {
     std::cerr << flags.Usage() << std::endl;
     return RET_OK;
   }
-
+  if (flags.unified_api_) {
+    return NetTrain::RunNr(&flags);
+  }
   NetTrain net_trainer(&flags);
   auto status = net_trainer.Init();
   if (status != RET_OK) {

mindspore/lite/tools/benchmark_train/net_train.h

@@ -53,8 +53,8 @@ constexpr float relativeTolerance = 1e-5;
 constexpr float absoluteTolerance = 1e-8;
 
 template <typename T>
-float TensorSum(void *data, int size) {
-  T *typed_data = reinterpret_cast<T *>(data);
+float TensorSum(const void *data, int size) {
+  const T *typed_data = reinterpret_cast<const T *>(data);
   float sum = 0.f;
   for (int i = 0; i < size; i++) {
     sum += static_cast<float>(typed_data[i]);
@@ -85,6 +85,7 @@ class MS_API NetTrainFlags : public virtual FlagParser {
     AddFlag(&NetTrainFlags::virtual_batch_, "virtualBatch", "use virtual batch", false);
     AddFlag(&NetTrainFlags::resize_dims_in_, "inputShapes",
             "Shape of input data, the format should be NHWC. e.g. 1,32,32,32:1,1,32,32,1", "");
+    AddFlag(&NetTrainFlags::unified_api_, "unifiedApi", "do unified api test", false);
   }
 
   ~NetTrainFlags() override = default;
@@ -117,6 +118,7 @@ class MS_API NetTrainFlags : public virtual FlagParser {
   std::vector<std::vector<int>> resize_dims_;
   std::string loss_name_ = "";
   std::string inference_file_ = "";
+  bool unified_api_ = false;
 };
 
 class MS_API NetTrain {
@@ -126,49 +128,19 @@ class MS_API NetTrain {
 
   int Init();
   int RunNetTrain();
-
- private:
-  // call GenerateInputData or ReadInputFile to init inputTensors
-  int LoadInput(Vector<tensor::MSTensor *> *ms_inputs);
-  void CheckSum(mindspore::tensor::MSTensor *tensor, std::string node_type, int id, std::string in_out);
-  // call GenerateRandomData to fill inputTensors
-  int GenerateInputData(std::vector<mindspore::tensor::MSTensor *> *ms_inputs);
-
-  int GenerateRandomData(mindspore::tensor::MSTensor *tensor);
-
-  int ReadInputFile(std::vector<mindspore::tensor::MSTensor *> *ms_inputs);
-  int CreateAndRunNetwork(const std::string &filename, const std::string &bb_filename, int train_session, int epochs,
-                          bool check_accuracy = true);
-
-  std::unique_ptr<session::LiteSession> CreateAndRunNetworkForInference(const std::string &filename,
-                                                                        const Context &context);
-
-  std::unique_ptr<session::LiteSession> CreateAndRunNetworkForTrain(const std::string &filename,
-                                                                    const std::string &bb_filename,
-                                                                    const Context &context, const TrainCfg &train_cfg,
-                                                                    int epochs);
-
-  int InitCallbackParameter();
-
-  int PrintResult(const std::vector<std::string> &title, const std::map<std::string, std::pair<int, float>> &result);
-
-  template <typename T>
-  void PrintInputData(tensor::MSTensor *input) {
-    MS_ASSERT(input != nullptr);
-    static int i = 0;
-    auto inData = reinterpret_cast<T *>(input->MutableData());
-    size_t tensorSize = input->ElementsNum();
-    size_t len = (tensorSize < 20) ? tensorSize : 20;
-    std::cout << "InData" << i++ << ": ";
-    for (size_t j = 0; j < len; j++) {
-      std::cout << inData[j] << " ";
+  static float *ReadFileBuf(const std::string file, size_t *size);
+  static int SetNr(std::function<int(NetTrainFlags *)> param);
+  static int RunNr(NetTrainFlags *flags) {
+    if (nr_cb_ != nullptr) {
+      return nr_cb_(flags);
     }
-    std::cout << std::endl;
+    MS_LOG(WARNING) << "unified api was not tested";
+    std::cout << "unified api was not tested";
+    return RET_OK;
   }
-
   // tensorData need to be converter first
   template <typename T>
-  float CompareData(const float *refOutput, int size, T *msTensorData) {
+  static float CompareData(const float *refOutput, int size, const T *msTensorData) {
     size_t errorCount = 0;
     float meanError = 0;
     std::cout << "Out tensor size is: " << size << std::endl;
@@ -219,6 +191,45 @@ class MS_API NetTrain {
     return meanError;
   }
 
+ private:
+  // call GenerateInputData or ReadInputFile to init inputTensors
+  int LoadInput(Vector<tensor::MSTensor *> *ms_inputs);
+  void CheckSum(mindspore::tensor::MSTensor *tensor, std::string node_type, int id, std::string in_out);
+  // call GenerateRandomData to fill inputTensors
+  int GenerateInputData(std::vector<mindspore::tensor::MSTensor *> *ms_inputs);
+
+  int GenerateRandomData(mindspore::tensor::MSTensor *tensor);
+
+  int ReadInputFile(std::vector<mindspore::tensor::MSTensor *> *ms_inputs);
+  int CreateAndRunNetwork(const std::string &filename, const std::string &bb_filename, int train_session, int epochs,
+                          bool check_accuracy = true);
+
+  std::unique_ptr<session::LiteSession> CreateAndRunNetworkForInference(const std::string &filename,
+                                                                        const Context &context);
+
+  std::unique_ptr<session::LiteSession> CreateAndRunNetworkForTrain(const std::string &filename,
+                                                                    const std::string &bb_filename,
+                                                                    const Context &context, const TrainCfg &train_cfg,
+                                                                    int epochs);
+
+  int InitCallbackParameter();
+
+  int PrintResult(const std::vector<std::string> &title, const std::map<std::string, std::pair<int, float>> &result);
+
+  template <typename T>
+  void PrintInputData(tensor::MSTensor *input) {
+    MS_ASSERT(input != nullptr);
+    static int i = 0;
+    auto inData = reinterpret_cast<T *>(input->MutableData());
+    size_t tensorSize = input->ElementsNum();
+    size_t len = (tensorSize < 20) ? tensorSize : 20;
+    std::cout << "InData" << i++ << ": ";
+    for (size_t j = 0; j < len; j++) {
+      std::cout << inData[j] << " ";
+    }
+    std::cout << std::endl;
+  }
+
   int MarkPerformance(const std::unique_ptr<session::LiteSession> &session);
   int MarkAccuracy(const std::unique_ptr<session::LiteSession> &session, bool enforce_accuracy = true);
   int CompareOutput(const session::LiteSession &lite_session);
@@ -242,8 +253,8 @@ class MS_API NetTrain {
     }
   }
 #endif
-  NetTrainFlags *flags_;
-
+  NetTrainFlags *flags_{nullptr};
+  static std::function<int(NetTrainFlags *)> nr_cb_;
   // callback parameters
   uint64_t op_begin_ = 0;
   int op_call_times_total_ = 0;

mindspore/lite/tools/converter/converter_flags.cc

@@ -101,6 +101,7 @@ Flags::Flags() {
           "Whether to export MindIR pb. "
           "true | false",
           "false");
+  AddFlag(&Flags::noFusionStr, "NoFusion", "Avoid fusion optimization true|false", "false");
 }
 
 int Flags::InitInputOutputDataType() {
@@ -359,6 +360,18 @@ int Flags::InitPreInference() {
   return RET_OK;
 }
 
+int Flags::InitNoFusion() {
+  if (this->noFusionStr == "true") {
+    this->disableFusion = true;
+  } else if (this->noFusionStr == "false") {
+    this->disableFusion = false;
+  } else {
+    std::cerr << "INPUT ILLEGAL: NoFusion must be true|false " << std::endl;
+    return RET_INPUT_PARAM_INVALID;
+  }
+  return RET_OK;
+}
+
 int Flags::InitExportMindIR() {
   if (this->exportMindIR == "true") {
     this->export_mindir = true;
@@ -495,12 +508,16 @@ int Flags::Init(int argc, const char **argv) {
     std::cerr << "Init encrypt failed." << std::endl;
     return RET_INPUT_PARAM_INVALID;
   }
-
   ret = InitPreInference();
   if (ret != RET_OK) {
     std::cerr << "Init pre inference failed." << std::endl;
     return RET_INPUT_PARAM_INVALID;
   }
+  ret = InitNoFusion();
+  if (ret != RET_OK) {
+    std::cerr << "Init no fusion failed." << std::endl;
+    return RET_INPUT_PARAM_INVALID;
+  }
 
   ret = InitExportMindIR();
   if (ret != RET_OK) {
@@ -509,6 +526,7 @@ int Flags::Init(int argc, const char **argv) {
   }
   return RET_OK;
 }
+
 Flags::~Flags() {
   dec_key.clear();
   encKeyStr.clear();
@@ -591,7 +609,7 @@ std::string GetStrFromConfigFile(const std::string &file, const std::string &tar
   }
 
 #ifdef _WIN32
-  char *real_path = _fullpath(resolved_path.get(), file.c_str(), kPathLengthUpperLimit);
+  auto *real_path = _fullpath(resolved_path.get(), file.c_str(), kPathLengthUpperLimit);
 #else
   char *real_path = realpath(file.c_str(), resolved_path.get());
 #endif

mindspore/lite/tools/converter/converter_flags.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef MINDSPORE_LITE_TOOLS_CONVERTER_CONVERTER_FLAGS_H
-#define MINDSPORE_LITE_TOOLS_CONVERTER_CONVERTER_FLAGS_H
+#ifndef MINDSPORE_LITE_TOOLS_CONVERTER_CONVERTER_FLAGS_H_
+#define MINDSPORE_LITE_TOOLS_CONVERTER_CONVERTER_FLAGS_H_
 
 #include <string>
 #include <vector>
@@ -73,6 +73,8 @@ class Flags : public virtual mindspore::lite::FlagParser {
 
   int InitSaveFP16();
 
+  int InitNoFusion();
+
   void InitAclDefaultOption();
 
   int InitExportMindIR();
@@ -90,6 +92,7 @@ class Flags : public virtual mindspore::lite::FlagParser {
   TypeId outputDataType;
   std::string saveFP16Str = "off";
   bool saveFP16 = false;
+  std::string noFusionStr = "false";
   std::string inputDataTypeStr;
   std::string outputDataTypeStr;
   ParallelSplitConfig parallel_split_config_{};
@@ -134,4 +137,4 @@ std::string GetStrFromConfigFile(const std::string &file, const std::string &tar
 }  // namespace converter
 }  // namespace mindspore
 
-#endif
+#endif  // MINDSPORE_LITE_TOOLS_CONVERTER_CONVERTER_FLAGS_H_