make variable->Grad() a weak_ptr (#11453)

* fix #11416

* make sgd check tape has been backwarded_

* add error message

paddle/contrib/tape/function.h

@@ -112,6 +112,8 @@ class SGD {
   }
 
   void operator()(VariableHandle input) {
+    PADDLE_ENFORCE(get_global_tape().HasBeenBackwarded(),
+                   "optimization must happen after the backward");
     Tape temp_tape;
     temp_tape.AddOp("sgd",
                     {{"Param", {input}},
@@ -120,7 +122,6 @@ class SGD {
                     {{"ParamOut", {input}}},
                     {});
     temp_tape.Forward();
-    input->ResetGrad();
   }
 
  private:

paddle/contrib/tape/tape.h

@@ -47,6 +47,8 @@ class Tape {
   void Forward();
   void Backward(VariableHandle target);
 
+  bool HasBeenBackwarded() { return has_been_backwarded_; }
+
  private:
   bool has_been_backwarded_ = false;
   size_t current_position_ = 0;

paddle/contrib/tape/variable.h

@@ -45,15 +45,15 @@ class Variable {
   void InitializeVariable();
 
   VariableHandle Grad() {
-    if (grad_ == nullptr) {
-      grad_.reset(new Variable(desc_.Name(), true));
+    if (grad_.expired()) {
+      VariableHandle new_grad(new Variable(desc_.Name(), true));
+      grad_ = new_grad;
+      return new_grad;
+    } else {
+      return VariableHandle(grad_);
     }
-
-    return grad_;
   }
 
-  void ResetGrad() { grad_ = nullptr; }
-
   // Stochastic Gradient Descent with Momentum
   //  VariableHandle Momentum ();
 
@@ -79,7 +79,7 @@ class Variable {
   framework::VarDesc desc_;
   framework::Variable var_;
 
-  VariableHandle grad_;
+  std::weak_ptr<Variable> grad_;
 };
 }
 }