!1261 Incremental subgraph initialization

Merge pull request !1261 from changzherui/add_tensor_init

mindspore/ccsrc/ir/meta_tensor.cc

@@ -374,6 +374,10 @@ TypeId Tensor::set_data_type(const TypeId data_type) {
   return data_type_;
 }
 
+bool Tensor::is_init() { return init_flag_; }
+
+void Tensor::set_init_flag(bool flag) { init_flag_ = flag; }
+
 bool Tensor::convert_data(const py::array &in, const TypeId in_data_type, py::array *const out,
                           const TypeId out_data_type) {
   if (out == nullptr) {
@@ -499,6 +503,24 @@ REGISTER_PYBIND_DEFINE(Tensor, ([](const py::module *m) {
                                  >>> data.size()
                                  6
                              )mydelimiter")
+                           .def("is_init", &Tensor::is_init, R"mydelimiter(
+                             Get tensor init_flag.
+
+                             Returns:
+                                 bool, whether the tensor init.
+
+                             Examples:
+                                 >>> data = mindspore.Tensor(np.ones((2, 3)))
+                                 >>> data.is_init()
+                                 False
+                             )mydelimiter")
+                           .def("set_init_flag", &Tensor::set_init_flag, R"mydelimiter(
+                             Set tensor init_flag.
+
+                             Examples:
+                                 >>> data = mindspore.Tensor(np.ones((2, 3)))
+                                 >>> data.set_init_flag(True)
+                             )mydelimiter")
                            .def("dim", &Tensor::DataDim, R"mydelimiter(
                              Get tensor's data dimension.
 

mindspore/ccsrc/ir/meta_tensor.h

@@ -389,6 +389,8 @@ class Tensor : public MetaTensor {
   std::string ToStringRepr() const;
   py::array data_;  // < Tensor's data value
   const bool parse_info_ = true;
+  bool is_init();
+  void set_init_flag(bool flag);
 
  private:
   // brief init tensor
@@ -398,7 +400,7 @@ class Tensor : public MetaTensor {
   // return true if succeed, false if failed.
   void init(const py::array &input, const TypeId &data_type);
   void init(const py::array &input, const TypePtr &type_ptr);
-
+  bool init_flag_{false};
   // brief init tensor attribute
   //
   // param data_type [TypeId] Data type of the tensor.

mindspore/ccsrc/transform/convert.cc

@@ -649,7 +649,6 @@ void DfGraphConvertor::InitParamWithData(const TensorOrderMap &tensors) {
     if (adpt == nullptr) continue;
     auto param_op = adpt->generate(name + "_data");
     MS_LOG(INFO) << "Add parameter " << name << " as input, index " << index << ".";
-    (void)std::static_pointer_cast<Data>(param_op)->set_attr_index(index++);
 
     if (!training_) {
       auto adpt_const = FindAdapter(kNameConst, training_);
@@ -678,14 +677,17 @@ void DfGraphConvertor::InitParamWithData(const TensorOrderMap &tensors) {
 
     // we need three variable ops for each graph with same name
     // build init subgraph
-    auto init_var = std::make_shared<Variable>(name);
-    auto assign_op = std::make_shared<Assign>("assign_" + name);
-    (void)init_var->update_output_desc_y(*desc);
-    (void)assign_op->set_input_ref(*init_var).set_input_value(*param_op);
-    init_input.push_back(*init_var);
-    init_ops_.push_back(param_op);
-    init_ops_.push_back(assign_op);
-    init_ops_.push_back(init_var);
+    if (it.second->is_init() == 0) {
+      (void)std::static_pointer_cast<Data>(param_op)->set_attr_index(index++);
+      auto init_var = std::make_shared<Variable>(name);
+      auto assign_op = std::make_shared<Assign>("assign_" + name);
+      (void)init_var->update_output_desc_y(*desc);
+      (void)assign_op->set_input_ref(*init_var).set_input_value(*param_op);
+      init_input.push_back(*init_var);
+      init_ops_.push_back(param_op);
+      init_ops_.push_back(assign_op);
+      init_ops_.push_back(init_var);
+    }
 
     auto variable = std::make_shared<Variable>(name);
     (void)variable->update_output_desc_y(*desc);

mindspore/common/api.py

@@ -82,14 +82,15 @@ def _wrap_func(fn):
 def _exec_init_graph(obj, init_phase):
     """Execute the parameter initializer graph."""
     inst_executor = Executor_.get_instance()
-    exec_init_graph = False
-    for param in obj.get_parameters():
+    param_dict = OrderedDict()
+    for name, param in obj.parameters_dict().items():
         if not param.is_init:
+            param_dict[name] = param
             param.is_init = True
-            exec_init_graph = True
+            param.data.init_flag = True
 
-    if exec_init_graph:
-        inst_executor.run_init_graph(obj.parameters_dict(), init_phase)
+    if param_dict:
+        inst_executor.run_init_graph(param_dict, init_phase)
 
 
 class _MindSporeFunction:

mindspore/common/parameter.py

@@ -188,11 +188,14 @@ class Parameter:
         if isinstance(data, Tensor):
             # make a copy of Tensor to init the parameter
             data = Tensor(data.asnumpy().copy())
+            data.init_flag = False
         elif isinstance(data, Initializer):
             self.init_mode = data
             data = MetaTensor(self.init_mode.dtype, self.init_mode.shape)
         else:
             data = Tensor(data)
+            data.init_flag = False
+
         self.default_input = data
 
 

mindspore/common/tensor.py

@@ -65,6 +65,7 @@ class Tensor(Tensor_):
         else:
             super(Tensor, self).__init__(input_data, dtype)
         self._virtual_flag = False
+        self._init_flag = False
 
     def __repr__(self):
         return str(self.__str__())
@@ -153,3 +154,16 @@ class Tensor(Tensor_):
         if not isinstance(value, bool):
             raise TypeError("virtual_flag must be bool.")
         self._virtual_flag = value
+
+    @property
+    def init_flag(self):
+        """whether the tensor is init."""
+        return self._init_flag
+
+    @init_flag.setter
+    def init_flag(self, value):
+        """Set the tensor is init_flag."""
+        if not isinstance(value, bool):
+            raise TypeError("init_flag must be bool.")
+        self.set_init_flag(value)
+        self._init_flag = value