!4283 add InsertGradientOf operator support in pynative mode

Merge pull request !4283 from wangqiuliang/add-insert-gradient-of-operator-support

mindspore/ccsrc/pipeline/pynative/pynative_execute.cc

@@ -58,7 +58,8 @@ using mindspore::tensor::TensorPy;
 
 const char SINGLE_OP_GRAPH[] = "single_op_graph";
 // primitive unable to infer value for constant input in PyNative mode
-const std::set<std::string> vm_operators = {"make_ref", "HookBackward", "stop_gradient", "mixed_precision_cast"};
+const std::set<std::string> vm_operators = {"make_ref", "HookBackward", "InsertGradientOf", "stop_gradient",
+                                            "mixed_precision_cast"};
 
 namespace mindspore {
 namespace pynative {
@@ -346,7 +347,7 @@ py::object RunOpInVM(const OpExecInfoPtr &op_exec_info, PynativeStatusCode *stat
   MS_EXCEPTION_IF_NULL(op_exec_info->py_primitive);
 
   auto &op_inputs = op_exec_info->op_inputs;
-  if (op_exec_info->op_name == "HookBackward") {
+  if (op_exec_info->op_name == "HookBackward" || op_exec_info->op_name == "InsertGradientOf") {
     py::tuple result(op_inputs.size());
     for (size_t i = 0; i < op_inputs.size(); i++) {
       py::object input = op_inputs[i];

mindspore/ops/operations/debug_ops.py

@@ -238,10 +238,6 @@ class InsertGradientOf(PrimitiveWithInfer):
     def __init__(self, f):
         self.f = f
 
-    def __call__(self, x):
-        """run in PyNative mode."""
-        return x
-
     def infer_shape(self, x_shape):
         return x_shape