diff --git a/paddle/operators/add_op.cc b/paddle/operators/add_op.cc
index 3a43dbfbada87e458109d8ca22effdb4407b4c1d..85269a5f7445a1745d9be68417789e33eb725d5c 100644
--- a/paddle/operators/add_op.cc
+++ b/paddle/operators/add_op.cc
@@ -50,10 +50,6 @@ The equation is: Out = X + Y
 class AddOpGrad : public OperatorWithKernel {
 protected:
   void InferShape(const InferShapeContext &ctx) const override {}
-  std::string DebugString() const override {
-    LOG(INFO) << "AddOpGrad";
-    return "";
-  }
 };
 
 }  // namespace operators
diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt
index 845589dcb1997b662b5175e5cce320eec4be4a8d..ac12b504b5cbee778c7c0a74a84a7729f210e01e 100644
--- a/paddle/pybind/CMakeLists.txt
+++ b/paddle/pybind/CMakeLists.txt
@@ -1,6 +1,6 @@
 cc_library(paddle_pybind SHARED
     SRCS pybind.cc
-    DEPS pybind python
+    DEPS pybind python backward
 	fc_op
 	sgd_op
 	add_op
diff --git a/paddle/pybind/pybind.cc b/paddle/pybind/pybind.cc
index d3cde07bd045dfe46be2363c14720c1393221157..40ff164497f627c0b562b6d33bfb4bec590e4c85 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/pybind/pybind.cc
@@ -16,6 +16,7 @@ limitations under the License. */
 #include <fstream>
 #include <vector>
 
+#include "paddle/framework/backward.h"
 #include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
@@ -45,6 +46,10 @@ template <typename ClassType>
 void ExposeOperator(ClassType& m) {
   m.def("infer_shape", &ClassType::type::InferShape)
       .def("run", &ClassType::type::Run)
+      .def("type",
+           [](const typename ClassType::type& op) -> std::string {
+             return op.type_;
+           })
       .def("outputs",
            [](const typename ClassType::type& op) -> std::vector<std::string> {
              return op.outputs_;
@@ -192,6 +197,13 @@ All parameter, weight, gradient are variables in Paddle.
                    desc.InitializationErrorString());
     return pd::OpRegistry::CreateOp(desc);
   });
+
+  operator_base.def("backward",
+                    [](const pd::OperatorBase& forwardOp,
+                       const std::unordered_set<std::string>& no_grad_vars) {
+                      return pd::Backward(forwardOp, no_grad_vars);
+                    });
+
   ExposeOperator(operator_base);
 
   py::class_<pd::NetOp, std::shared_ptr<pd::NetOp>> net(m, "Net");
diff --git a/python/paddle/v2/framework/tests/test_add_two_op.py b/python/paddle/v2/framework/tests/test_add_two_op.py
index 73b373490956c741f06be2473e70592dc06fbc25..6e6643201bf361fce1bad7de10b2562f0525e00a 100644
--- a/python/paddle/v2/framework/tests/test_add_two_op.py
+++ b/python/paddle/v2/framework/tests/test_add_two_op.py
@@ -1,6 +1,10 @@
 import unittest
-from op_test_util import OpTestMeta
+
 import numpy
+import paddle.v2.framework.core as core
+import paddle.v2.framework.create_op_creation_methods as creation
+
+from op_test_util import OpTestMeta
 
 
 class TestAddOp(unittest.TestCase):
@@ -13,5 +17,14 @@ class TestAddOp(unittest.TestCase):
         self.Out = self.X + self.Y
 
 
+class TestAddGradOp(unittest.TestCase):
+    def test_add_grad(self):
+        op = creation.op_creations.add_two(X="X", Y="Y", Out="Out")
+        backward_op = core.Operator.backward(op, set())
+        self.assertEqual(backward_op.type(), "add_two_grad")
+        expected = '''Op(add_two_grad), inputs:(X, Y, Out, Out@GRAD), outputs:(X@GRAD, Y@GRAD).'''
+        self.assertEqual(expected, str(backward_op))
+
+
 if __name__ == '__main__':
     unittest.main()