diff --git a/CMakeLists.txt b/CMakeLists.txt
index c7d743e193e7d32dbc0b56f3bcb05b6c61f85f1d..b174831109372cb014741d63032fa6a470e74042 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -36,8 +36,8 @@ include(simd)
 ################################ Configurations #######################################
 option(WITH_GPU         "Compile PaddlePaddle with NVIDIA GPU"          ${CUDA_FOUND})
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    OFF)
-option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      OFF)
+option(WITH_MKLDNN      "Compile PaddlePaddle with mkl-dnn support."    ${AVX_FOUND})
+option(WITH_MKLML       "Compile PaddlePaddle with mklml package."      ${AVX_FOUND})
 option(WITH_DSO         "Compile PaddlePaddle with dynamic linked CUDA" ON)
 option(WITH_TESTING     "Compile PaddlePaddle with unit testing"        ON)
 option(WITH_SWIG_PY     "Compile PaddlePaddle with inference api"       ON)
diff --git a/README.md b/README.md
index 2a6beeb342b34f8e91ef509d7d41f286a666480c..b9793c3eab5d40c28f01cc67ad607b97261b3235 100644
--- a/README.md
+++ b/README.md
@@ -72,7 +72,7 @@ We provide [English](http://doc.paddlepaddle.org/develop/doc/) and
 
 - [Deep Learning 101](http://book.paddlepaddle.org/index.html)
 
-  You might want to start from the this online interactive book that can run in Jupyter Notebook.
+  You might want to start from this online interactive book that can run in Jupyter Notebook.
 
 - [Distributed Training](http://doc.paddlepaddle.org/develop/doc/howto/usage/cluster/cluster_train_en.html)
 
diff --git a/cmake/external/mkldnn.cmake b/cmake/external/mkldnn.cmake
index eff15de73f23db6dea3a7b79006bfec90d712ae5..25c6b4ef52d3f8ebff1572ae8d348be7c577c08c 100644
--- a/cmake/external/mkldnn.cmake
+++ b/cmake/external/mkldnn.cmake
@@ -20,34 +20,30 @@ INCLUDE(ExternalProject)
 
 SET(MKLDNN_PROJECT        "extern_mkldnn")
 SET(MKLDNN_SOURCES_DIR    ${THIRD_PARTY_PATH}/mkldnn)
-SET(MKLDNN_INSTALL_ROOT   ${CMAKE_INSTALL_PREFIX})
-IF(NOT "$ENV{HOME}" STREQUAL "/root")
-    SET(MKLDNN_INSTALL_ROOT  "$ENV{HOME}")
-ENDIF()
-
-SET(MKLDNN_INSTALL_DIR    "${MKLDNN_INSTALL_ROOT}/opt/paddle/third_party/mkldnn")
-SET(MKLDNN_INCLUDE_DIR    "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE)
+SET(MKLDNN_INSTALL_DIR    ${THIRD_PARTY_PATH}/install/mkldnn)
+SET(MKLDNN_INC_DIR        "${MKLDNN_INSTALL_DIR}/include" CACHE PATH "mkldnn include directory." FORCE)
 
-IF(WIN32)
-    MESSAGE(WARNING "It is not supported compiling with mkldnn in windows Paddle yet."
-      "Force WITH_MKLDNN=OFF")
-    SET(WITH_MKLDNN OFF)
+IF(WIN32 OR APPLE)
+    MESSAGE(WARNING 
+        "Windows or Mac is not supported with MKLDNN in Paddle yet."
+        "Force WITH_MKLDNN=OFF")
+    SET(WITH_MKLDNN OFF CACHE STRING "Disable MKLDNN in Windows and MacOS" FORCE)
     return()
-ELSE(WIN32)
-    SET(MKLDNN_LIBRARY "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
-    MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
-    SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
-    #SET(CMAKE_MACOSX_RPATH 1) # hold for MacOS
-    SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
-ENDIF(WIN32)
+ENDIF()
+
+SET(MKLDNN_LIB "${MKLDNN_INSTALL_DIR}/lib/libmkldnn.so" CACHE FILEPATH "mkldnn library." FORCE)
+MESSAGE(STATUS "Set ${MKLDNN_INSTALL_DIR}/lib to runtime path")
+SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
+SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${MKLDNN_INSTALL_DIR}/lib")
 
-INCLUDE_DIRECTORIES(${MKLDNN_INCLUDE_DIR})
+INCLUDE_DIRECTORIES(${MKLDNN_INC_DIR})
 
 IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
     SET(MKLDNN_DEPENDS   ${MKLML_PROJECT})
     SET(MKLDNN_MKLROOT   ${MKLML_ROOT})
     SET(MKLDNN_IOMP_LIB  ${MKLML_IOMP_LIB})
     SET(MKLDNN_IOMP_DIR  ${MKLML_LIB_DIR})
+    MESSAGE(STATUS "Build MKLDNN with ${MKLDNN_MKLROOT}")
 ENDIF()
 
 ExternalProject_Add(
@@ -57,16 +53,15 @@ ExternalProject_Add(
     GIT_REPOSITORY      "https://github.com/01org/mkl-dnn.git"
     GIT_TAG             "v0.9"
     PREFIX              ${MKLDNN_SOURCES_DIR}
-    CONFIGURE_COMMAND   mkdir -p <SOURCE_DIR>/build
-    BUILD_COMMAND       cd <SOURCE_DIR>/build
-                        && cmake .. -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR} -DMKLROOT=${MKLDNN_MKLROOT}
-                        && $(MAKE)
-    INSTALL_COMMAND     cd <SOURCE_DIR>/build && $(MAKE) install
     UPDATE_COMMAND      ""
+    CMAKE_ARGS          -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
+    CMAKE_ARGS          -DMKLROOT=${MKLDNN_MKLROOT}
+    CMAKE_CACHE_ARGS    -DCMAKE_INSTALL_PREFIX:PATH=${MKLDNN_INSTALL_DIR}
+                        -DMKLROOT:PATH=${MKLDNN_MKLROOT}
 )
 
 ADD_LIBRARY(mkldnn SHARED IMPORTED GLOBAL)
-SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIBRARY})
+SET_PROPERTY(TARGET mkldnn PROPERTY IMPORTED_LOCATION ${MKLDNN_LIB})
 ADD_DEPENDENCIES(mkldnn ${MKLDNN_PROJECT})
-MESSAGE(STATUS "Mkldnn library: ${MKLDNN_LIBRARY}")
+MESSAGE(STATUS "Mkldnn library: ${MKLDNN_LIB}")
 LIST(APPEND external_project_dependencies mkldnn)
diff --git a/cmake/external/mklml.cmake b/cmake/external/mklml.cmake
index 3f940756a4abb79aba7d3561db19db8532a0b673..17a1ca4ed04dce85ae3c7fdd5f22d6eeed03db59 100644
--- a/cmake/external/mklml.cmake
+++ b/cmake/external/mklml.cmake
@@ -16,19 +16,23 @@ IF(NOT ${WITH_MKLML})
   return()
 ENDIF(NOT ${WITH_MKLML})
 
+IF(WIN32 OR APPLE)
+    MESSAGE(WARNING 
+        "Windows or Mac is not supported with MKLML in Paddle yet."
+        "Force WITH_MKLML=OFF")
+    SET(WITH_MKLML OFF CACHE STRING "Disable MKLML package in Windows and MacOS" FORCE)
+    return()
+ENDIF()
+
 INCLUDE(ExternalProject)
 
 SET(MKLML_PROJECT       "extern_mklml")
-SET(MKLML_VER           "mklml_lnx_2018.0.20170425")
+SET(MKLML_VER           "mklml_lnx_2018.0.20170720")
 SET(MKLML_URL           "https://github.com/01org/mkl-dnn/releases/download/v0.9/${MKLML_VER}.tgz")
 SET(MKLML_SOURCE_DIR    "${THIRD_PARTY_PATH}/mklml")
 SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
-SET(MKLML_DST_DIR       "opt/paddle/third_party/mklml")
-SET(MKLML_INSTALL_ROOT  "${CMAKE_INSTALL_PREFIX}")
-IF(NOT "$ENV{HOME}" STREQUAL "/root")
-    SET(MKLML_INSTALL_ROOT  "$ENV{HOME}")
-ENDIF()
-
+SET(MKLML_DST_DIR       "mklml")
+SET(MKLML_INSTALL_ROOT  "${THIRD_PARTY_PATH}/install")
 SET(MKLML_INSTALL_DIR   ${MKLML_INSTALL_ROOT}/${MKLML_DST_DIR})
 SET(MKLML_ROOT          ${MKLML_INSTALL_DIR}/${MKLML_VER})
 SET(MKLML_INC_DIR       ${MKLML_ROOT}/include)
diff --git a/paddle/CMakeLists.txt b/paddle/CMakeLists.txt
index 4b06966fba2bc9f92756be0cb8110bbcd5272423..f8a88cf317aee6c5dd25e4cc25d588c6c50fcbce 100644
--- a/paddle/CMakeLists.txt
+++ b/paddle/CMakeLists.txt
@@ -15,7 +15,6 @@ if(Boost_FOUND)
   add_subdirectory(platform)
   add_subdirectory(framework)
   add_subdirectory(operators)
-  add_subdirectory(pybind)
 endif()
 
 if(WITH_C_API)
diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt
index 12a3a00bba35d476fca9c9fb47ac20b87e6f53f2..9c39430835d37d5dfbe4031f29e5a6216ed8b67f 100644
--- a/paddle/framework/CMakeLists.txt
+++ b/paddle/framework/CMakeLists.txt
@@ -31,8 +31,14 @@ py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.
 add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
 add_dependencies(framework_py_proto framework_py_proto_init)
 
-cc_library(net SRCS net.cc DEPS op_registry)
-cc_test(net_op_test SRCS net_op_test.cc DEPS net)
-
-cc_library(backward SRCS backward.cc DEPS net)
+cc_library(backward SRCS backward.cc DEPS net_op)
 cc_test(backward_test SRCS backward_test.cc DEPS backward)
+cc_library(paddle_pybind SHARED
+    SRCS pybind.cc
+    DEPS pybind python backward
+	fc_op
+	sgd_op
+	add_op
+	mean_op
+	cross_entropy_op
+	recurrent_op)
diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc
index 0da11b91a7fe4a98e0832f70095c3200956ff001..c034e265fe4837ca22ab969b0e6952677904e05c 100644
--- a/paddle/framework/backward.cc
+++ b/paddle/framework/backward.cc
@@ -14,8 +14,8 @@
 
 #include "paddle/framework/backward.h"
 #include <list>
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace framework {
@@ -32,7 +32,7 @@ static bool AllInSet(const std::vector<std::string>& names,
 }
 
 static std::shared_ptr<OperatorBase> NOP() {
-  auto net_op = std::make_shared<NetOp>();
+  auto net_op = std::make_shared<operators::NetOp>();
   net_op->type_ = "@NOP@";
   net_op->CompleteAddOp();
   return net_op;
@@ -42,9 +42,9 @@ static std::shared_ptr<OperatorBase> NOP() {
 //
 //  no_grad_names the gradient variable names without gradient calculating.
 //
-//  uniq_id is a unique index used inside recursively calling BackwardRecursive.
-//  use `uid = uniq_id++;` to get the unique index, and pass `uniq_id` through
-//  recursive calling.
+//  uniq_id is a unique index used inside recursively calling
+//  BackwardRecursive. use `uid = uniq_id++;` to get the unique index, and
+//  pass `uniq_id` through recursive calling.
 //
 //  returns The backward operator. For simple situation, it is a simple
 //  operator. For complex situation, it is a NetOp.
@@ -64,8 +64,8 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
     return NOP();
   }
 
-  //  All output gradients of forwarding operator do not need to calculate. Then
-  //  all input gradients cannot be computed at all, and we put them into
+  //  All output gradients of forwarding operator do not need to calculate.
+  //  Then all input gradients cannot be computed at all, and we put them into
   //  `no_grad_names` set. Return an NOP.
   if (AllInSet(forwardOp.outputs_, OperatorBase::GRAD_VAR_SUFFIX(),
                no_grad_names)) {
@@ -77,14 +77,14 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
   }
 
   // Returned gradient network
-  auto net = std::make_shared<NetOp>();
+  auto net = std::make_shared<operators::NetOp>();
 
   if (forwardOp.IsNetOp()) {
     // Because forwardOp is a net op, it can static_cast.
-    auto& forwardNet = static_cast<const NetOp&>(forwardOp);
+    auto& forwardNet = static_cast<const operators::NetOp&>(forwardOp);
 
-    // Map from output gradient variable name to operator's indices in backward
-    // net. That operator generates that variable.
+    // Map from output gradient variable name to operator's indices in
+    // backward net. That operator generates that variable.
     std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;
 
     size_t local_op_id = 0;
@@ -168,6 +168,9 @@ std::shared_ptr<OperatorBase> Backward(
   std::unordered_set<std::string> no_grad_names;
   no_grad_names.reserve(no_grad_vars.size());
 
+  no_grad_names.insert(OperatorBase::EMPTY_VAR_NAME() +
+                       OperatorBase::GRAD_VAR_SUFFIX());
+
   for (auto& name : no_grad_vars) {
     no_grad_names.insert(name + OperatorBase::GRAD_VAR_SUFFIX());
   }
diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc
index b095c2c3d5dbf21b5ea70e17475a4aaad9b1db44..8f437e68041188831a17217099e0b0c96432cda4 100644
--- a/paddle/framework/backward_test.cc
+++ b/paddle/framework/backward_test.cc
@@ -15,8 +15,9 @@
 #include "paddle/framework/backward.h"
 
 #include <gtest/gtest.h>
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/type_alias.h"
 
 namespace paddle {
 namespace framework {
@@ -70,7 +71,7 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker {
   }
 };
 
-class FcOp : public NetOp {
+class FcOp : public ops::NetOp {
  public:
   void Init() override {
     AddOp(OpRegistry::CreateOp("mul", {Input("X"), Input("W")},
@@ -161,8 +162,8 @@ TEST(Backward, simple_op_grad) {
   auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
   ASSERT_NE(fwd, nullptr);
   auto gop = f::OpRegistry::CreateGradOp(*fwd);
-  ASSERT_EQ(1UL, gop->inputs_.size());
-  ASSERT_EQ("Out" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->inputs_[0]);
+  ASSERT_EQ(4UL, gop->inputs_.size());
+  ASSERT_EQ(f::OperatorBase::EMPTY_VAR_NAME(), gop->inputs_[0]);
   ASSERT_EQ("rowwise_add_grad", gop->type_);
   ASSERT_EQ("X" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[0]);
   ASSERT_EQ("b" + f::OperatorBase::GRAD_VAR_SUFFIX(), gop->outputs_[1]);
@@ -182,7 +183,8 @@ TEST(Backward, simple_op_not_need_grad) {
   auto no_input_gop = f::Backward(*fwd, {"X", "b"});
   ASSERT_NE(no_input_gop, nullptr);
   ASSERT_TRUE(no_input_gop->IsNetOp());
-  ASSERT_EQ(0UL, std::static_pointer_cast<f::NetOp>(no_input_gop)->ops_.size());
+  ASSERT_EQ(0UL,
+            std::static_pointer_cast<ops::NetOp>(no_input_gop)->ops_.size());
 }
 
 TEST(Backward, net_fc_backward_normal) {
@@ -191,7 +193,7 @@ TEST(Backward, net_fc_backward_normal) {
   ASSERT_NE(fwd, nullptr);
   std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
   ASSERT_TRUE(gop->IsNetOp());
-  auto net = static_cast<f::NetOp *>(gop.get());
+  auto net = static_cast<ops::NetOp *>(gop.get());
 
   ASSERT_NO_THROW(net->DebugString());
 
@@ -214,7 +216,7 @@ TEST(Backward, net_fc_backward_not_have_b) {
   ASSERT_NE(fwd, nullptr);
   std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
   ASSERT_TRUE(gop->IsNetOp());
-  auto net = static_cast<f::NetOp *>(gop.get());
+  auto net = static_cast<ops::NetOp *>(gop.get());
 
   ASSERT_NO_THROW(net->DebugString());
 
@@ -228,7 +230,7 @@ TEST(Backward, net_fc_backward_not_have_b) {
 }
 
 TEST(Backward, net_input_of_network_not_need_grad) {
-  f::NetOp net;
+  ops::NetOp net;
   net.AddOp(f::OpRegistry::CreateOp("fc", {"X", "W1", "b1"},
                                     {"mul_tmp_0", "add_tmp_0", "hidden0"}, {}));
   net.AddOp(f::OpRegistry::CreateOp("fc", {"hidden0", "W2", "b2"},
@@ -236,7 +238,7 @@ TEST(Backward, net_input_of_network_not_need_grad) {
   net.CompleteAddOp();
   auto bwd = Backward(net, {"X"});  // X@GRAD is not need.
   ASSERT_TRUE(bwd->IsNetOp());
-  auto bwd_net = static_cast<f::NetOp *>(bwd.get());
+  auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
 
   std::unordered_set<std::string> all_output = std::unordered_set<std::string>(
       bwd_net->outputs_.begin(), bwd_net->outputs_.end());
@@ -253,7 +255,7 @@ TEST(Backward, net_input_of_network_not_need_grad) {
 
   ASSERT_EQ(2UL, bwd_net->ops_.size());
   ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
-  auto first_fc_grad = static_cast<f::NetOp *>(bwd_net->ops_[1].get());
+  auto first_fc_grad = static_cast<ops::NetOp *>(bwd_net->ops_[1].get());
   ASSERT_EQ(3UL, first_fc_grad->ops_.size());
   ASSERT_EQ(
       f::OperatorBase::EMPTY_VAR_NAME(),
@@ -261,14 +263,14 @@ TEST(Backward, net_input_of_network_not_need_grad) {
 }
 
 TEST(Backward, net_shared_weight) {
-  f::NetOp net;
+  ops::NetOp net;
   net.AddOp(f::OpRegistry::CreateOp("mul", {"X", "W"}, {"Out"}, {}));
   net.AddOp(f::OpRegistry::CreateOp("mul", {"Out", "W"}, {"FinalOut"}, {}));
   net.CompleteAddOp();
 
   auto bwd = f::Backward(net, {});
   ASSERT_TRUE(bwd->IsNetOp());
-  auto bwd_net = static_cast<f::NetOp *>(bwd.get());
+  auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
   ASSERT_EQ(3UL, bwd_net->ops_.size());
   ASSERT_EQ("add", bwd_net->ops_[2]->type_);
 }
@@ -285,7 +287,7 @@ TEST(Backward, op_all_input_are_not_need) {
   auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
   auto backward = f::Backward(*fwd, {"X", "b"});
   ASSERT_TRUE(backward->IsNetOp());
-  auto net = static_cast<f::NetOp *>(backward.get());
+  auto net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_TRUE(net->ops_.empty());
 }
 
@@ -293,7 +295,7 @@ TEST(Backward, op_all_output_are_not_need) {
   auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
   auto backward = f::Backward(*fwd, {"Out"});
   ASSERT_TRUE(backward->IsNetOp());
-  auto net = static_cast<f::NetOp *>(backward.get());
+  auto net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_TRUE(net->ops_.empty());
 }
 
@@ -301,7 +303,7 @@ TEST(Backward, op_part_of_output_are_not_need) {
   auto fwd = f::OpRegistry::CreateOp("many_output_op", {"X"}, {"Y", "Z"}, {});
   auto backward = f::Backward(*fwd, {"Z"});
   ASSERT_TRUE(backward->IsNetOp());
-  auto net = static_cast<f::NetOp *>(backward.get());
+  auto net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_EQ(net->ops_.size(), 2UL);
 
   auto &fill_zero = *net->ops_[0];
@@ -341,7 +343,7 @@ TEST(Backward, op_part_of_input_are_not_need) {
 }
 
 TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
-  f::NetOp net;
+  ops::NetOp net;
   net.AddOp(f::OpRegistry::CreateOp("fc", {"x1", "w1", "b1"},
                                     {"mul_out1", "add_out1", "out1"}, {}));
   net.AddOp(f::OpRegistry::CreateOp("fc", {"out1", "w2", "b2"},
@@ -351,14 +353,13 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
   net.CompleteAddOp();
   auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"});
   ASSERT_TRUE(backward->IsNetOp());
-  auto bwd_net = static_cast<f::NetOp *>(backward.get());
+  auto bwd_net = static_cast<ops::NetOp *>(backward.get());
   ASSERT_EQ(bwd_net->ops_.size(), 3UL);
   auto &grad_fc = *bwd_net->ops_[0];
   EXPECT_EQ(grad_fc.inputs_.size(),
             3UL       /* external input number */
                 + 1UL /* external output number*/
                 + 1UL /* number of gradient of external output*/
-                - 1UL /*ignoreGradient varable number*/
                 + 2U /* internal variable number*/);
   EXPECT_EQ(grad_fc.outputs_.size(), 2UL       /* input number of mul*/
                                          + 2UL /* input number of rowwise_add */
diff --git a/paddle/framework/grad_op_builder.cc b/paddle/framework/grad_op_builder.cc
index dd686cc78246f06cdc3ec7d013086863d7e8fac0..ea5e939c6e26514c2f3c515da5581b29103f75b6 100644
--- a/paddle/framework/grad_op_builder.cc
+++ b/paddle/framework/grad_op_builder.cc
@@ -8,107 +8,97 @@ You may obtain a copy of the License at
 
 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. */
+WITHOpArgType::OUT WARRANTIES OR CONDITIONS OF ANY KOpArgType::IND, either
+express or implied. See the License for the specific language governing
+permissions and limitations under the License. */
 
 #include "paddle/framework/grad_op_builder.h"
+#include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
 namespace framework {
 
-OperatorBase* GradOpBuilder::Build() {
-  BuildOpInOutArgList();
-  std::string grad_op_type = OpRegistry::grad_ops().at(op_.type_);
-  OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)();
-  grad_op->type_ = grad_op_type;
-  CompleteGradOp(grad_op);
-  return grad_op;
-}
+class OpRegistry;
+
+using VarIndexMap = std::unordered_map<std::string, int>;
 
-OpInOutArg* GradOpBuilder::BuildArg(const VarProto& var,
-                                    const VarIndexMap& var_map,
-                                    const std::vector<int>& format,
-                                    InOutType type) {
-  int idx = var_map.at(var.name());
-  int begin_idx = format.empty() ? idx : format.at(idx);
-  int end_idx = format.empty() ? idx + 1 : format.at(idx + 1);
-  return new OpInOutArg(var.name(), type, !var.ignore_gradient(), begin_idx,
-                        end_idx);
+enum class OpArgType { IN, OUT };
+
+static std::vector<int>* GetOpFormat(OperatorBase* op, const OpArgType& type) {
+  std::string key = type == OpArgType::IN ? "input_format" : "output_format";
+  return op->attrs_.count(key)
+             ? &boost::get<std::vector<int>>(op->attrs_.at(key))
+             : nullptr;
 }
 
-void GradOpBuilder::BuildOpInOutArgList() {
-  const OpProto& op_proto = OpRegistry::protos().at(op_.type_);
-  const auto& var_map = *(OpRegistry::VarIndexMaps().at(op_.type_));
-  const std::vector<int>& in_format =
-      op_.attrs_.count("input_format")
-          ? op_.GetAttr<std::vector<int>>("input_format")
-          : std::vector<int>();
-  const std::vector<int>& out_format =
-      op_.attrs_.count("output_format")
-          ? op_.GetAttr<std::vector<int>>("output_format")
-          : std::vector<int>();
-  for (const auto& var : op_proto.inputs()) {
-    arg_list_.emplace_back(
-        std::shared_ptr<OpInOutArg>(BuildArg(var, var_map, in_format, IN)));
-  }
-  for (const auto& var : op_proto.outputs()) {
-    arg_list_.emplace_back(
-        std::shared_ptr<OpInOutArg>(BuildArg(var, var_map, out_format, OUT)));
-  }
+static const std::vector<int>* GetOpFormat(const OperatorBase* op,
+                                           const OpArgType& type) {
+  std::string key = type == OpArgType::IN ? "input_format" : "output_format";
+  return op->attrs_.count(key)
+             ? &boost::get<std::vector<int>>(op->attrs_.at(key))
+             : nullptr;
 }
 
-void GradOpBuilder::AddArgIntoGradOp(const OpInOutArg* arg,
-                                     std::vector<std::string>& in_out,
-                                     std::vector<int>& format,
-                                     VarIndexMap* varmap, int& idx,
-                                     bool is_grad) const {
-  std::string var_name = arg->proto_name_;
-  if (is_grad) {
-    var_name += OperatorBase::GRAD_VAR_SUFFIX();
-  }
-  (*varmap)[var_name] = idx++;
-  size_t pre_sz = in_out.size();
-  auto base_it = arg->type_ == IN ? op_.inputs_.begin() : op_.outputs_.begin();
-  std::copy(base_it + arg->begin_idx_, base_it + arg->end_idx_,
-            std::back_inserter(in_out));
-  if (is_grad) {
-    for (size_t i = pre_sz; i < in_out.size(); ++i) {
-      in_out[i] += OperatorBase::GRAD_VAR_SUFFIX();
+static void TransOpArg(const OperatorBase* src_op, OperatorBase* dst_op,
+                       const OpArgType& src_type, const OpArgType& dst_type,
+                       int& idx, bool is_grad) {
+  const std::vector<std::string>& src_inout =
+      src_type == OpArgType::IN ? src_op->inputs_ : src_op->outputs_;
+  const std::vector<int>* src_format = GetOpFormat(src_op, src_type);
+
+  std::vector<std::string>& dst_inout =
+      dst_type == OpArgType::IN ? dst_op->inputs_ : dst_op->outputs_;
+  std::vector<int>* dst_format = GetOpFormat(dst_op, dst_type);
+  const OpProto& proto = OpRegistry::protos().at(src_op->type_);
+  const auto& src_arg_list =
+      src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
+
+  for (const auto& arg : src_arg_list) {
+    std::string src_name = arg.name();
+    std::string dst_name =
+        is_grad ? src_name + OperatorBase::GRAD_VAR_SUFFIX() : src_name;
+    (*dst_op->in_out_idxs_)[dst_name] = idx++;
+    int src_arg_idx = src_op->in_out_idxs_->at(src_name);
+    int src_begin =
+        src_format == nullptr ? src_arg_idx : src_format->at(src_arg_idx);
+    int src_end = src_format == nullptr ? src_arg_idx + 1
+                                        : src_format->at(src_arg_idx + 1);
+    for (int i = src_begin; i < src_end; ++i) {
+      std::string s = is_grad ? src_inout[i] + OperatorBase::GRAD_VAR_SUFFIX()
+                              : arg.ignore_gradient()
+                                    ? OperatorBase::EMPTY_VAR_NAME()
+                                    : src_inout[i];
+      dst_inout.emplace_back(s);
+    }
+    if (dst_format != nullptr) {
+      dst_format->push_back(dst_inout.size());
     }
   }
-  format.push_back(in_out.size());
 }
 
-void GradOpBuilder::CompleteGradOp(OperatorBase* grad_op) const {
-  grad_op->attrs_ = op_.attrs_;
+OperatorBase* BuildGradOp(const OperatorBase* op) {
+  std::string grad_op_type = OpRegistry::grad_ops().at(op->type_);
+  OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)();
+  grad_op->type_ = grad_op_type;
+  grad_op->attrs_ = op->attrs_;
   grad_op->attrs_.erase("input_format");
   grad_op->attrs_.erase("output_format");
-  VarIndexMap* grad_varmap = new VarIndexMap();
+  if (GetOpFormat(op, OpArgType::IN) != nullptr) {
+    grad_op->attrs_["output_format"] = std::vector<int>({0});
+  }
+  if (GetOpFormat(op, OpArgType::IN) != nullptr ||
+      GetOpFormat(op, OpArgType::OUT) != nullptr) {
+    grad_op->attrs_["input_format"] = std::vector<int>({0});
+  }
+  grad_op->in_out_idxs_.reset(new VarIndexMap());
   int in_idx = 0;
   int out_idx = 0;
-  std::vector<int> in_format({0});
-  std::vector<int> out_format({0});
-  for (const auto& arg : arg_list_) {
-    // op_'s inputs_ and outputs_
-    if (arg->needed_in_grad_) {
-      AddArgIntoGradOp(arg.get(), grad_op->inputs_, in_format, grad_varmap,
-                       in_idx, false);
-    }
-    if (arg->type_ == IN) {
-      // gradients of op_'s inputs_
-      AddArgIntoGradOp(arg.get(), grad_op->outputs_, out_format, grad_varmap,
-                       out_idx, true);
-    } else {
-      // gradients of op_'s outputs_
-      AddArgIntoGradOp(arg.get(), grad_op->inputs_, in_format, grad_varmap,
-                       in_idx, true);
-    }
-  }
-  grad_op->attrs_["input_format"] = in_format;
-  grad_op->attrs_["output_format"] = out_format;
-  grad_op->in_out_idxs_.reset(grad_varmap);
+  TransOpArg(op, grad_op, OpArgType::IN, OpArgType::IN, in_idx, false);   // I
+  TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, false);  // G
+  TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, true);   // OG
+  TransOpArg(op, grad_op, OpArgType::IN, OpArgType::OUT, out_idx, true);  // IG
+  return grad_op;
 }
 
 }  // namespace framework
diff --git a/paddle/framework/grad_op_builder.h b/paddle/framework/grad_op_builder.h
index cc7a76f3726e00a08fbe06bca4c9b9f5bad466b4..cf235de6c267a4a1feb7afd3e4dbe7a6a668ee5e 100644
--- a/paddle/framework/grad_op_builder.h
+++ b/paddle/framework/grad_op_builder.h
@@ -1,48 +1,11 @@
 #pragma once
 
-#include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/operator.h"
 
 namespace paddle {
 namespace framework {
-class OpRegistry;
 
-enum InOutType { IN, OUT };
-
-struct OpInOutArg {
-  OpInOutArg(const std::string& proto_name, const InOutType& type,
-             bool needed_in_grad, size_t begin_idx, size_t end_idx)
-      : proto_name_(proto_name),
-        type_(type),
-        needed_in_grad_(needed_in_grad),
-        begin_idx_(begin_idx),
-        end_idx_(end_idx) {}
-
-  std::string proto_name_;
-  InOutType type_;
-  bool needed_in_grad_;
-  size_t begin_idx_;
-  size_t end_idx_;
-};
-
-class GradOpBuilder {
-  using VarIndexMap = std::unordered_map<std::string, int>;
-
- public:
-  GradOpBuilder(const OperatorBase& op) : op_(op) {}
-  OperatorBase* Build();
-
- private:
-  OpInOutArg* BuildArg(const VarProto& var, const VarIndexMap& var_map,
-                       const std::vector<int>& format, InOutType type);
-  void BuildOpInOutArgList();
-  void AddArgIntoGradOp(const OpInOutArg* arg, std::vector<std::string>& in_out,
-                        std::vector<int>& format, VarIndexMap* varmap, int& idx,
-                        bool is_grad) const;
-  void CompleteGradOp(OperatorBase* grad_op) const;
-  const OperatorBase& op_;
-  std::vector<std::shared_ptr<OpInOutArg>> arg_list_;
-};
+OperatorBase* BuildGradOp(const OperatorBase* op);
 
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc
index e9cf3b9798db2cbfb8d26259ae9a6741fbae8278..96d7f309d67b15c000ab8ce3769931322fbca880 100644
--- a/paddle/framework/grad_op_builder_test.cc
+++ b/paddle/framework/grad_op_builder_test.cc
@@ -8,10 +8,49 @@ USE_OP(add_two);
 namespace paddle {
 namespace framework {
 
+class NOP : public OperatorBase {
+ public:
+  void InferShape(const Scope &scope) const override {}
+  void Run(const Scope &scope,
+           const platform::DeviceContext &dev_ctx) const override {}
+};
+
+class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("In1", "a single input");
+    AddInput("In2_mult", "a multiple input").SetMultiple();
+    AddInput("In3", "another single input");
+    AddOutput("Out1", "a single output");
+    AddOutput("Out2_mult", "a multiple output").SetMultiple();
+    AddComment("test op with multiple inputs and outputs");
+  }
+};
+
+class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
+ public:
+  IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("In1", "a single input");
+    AddInput("In2_mult", "a multiple input").SetMultiple().IgnoreGradient();
+    AddInput("In3_mult", "another multiple input").SetMultiple();
+    AddOutput("Out1_mult", "a multiple output").SetMultiple();
+    AddOutput("Out2", "a single output").IgnoreGradient();
+    AddComment("op with inputs and outputs ignored in gradient calculating");
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle
+
+namespace f = paddle::framework;
+
 TEST(GradOpBuilder, AddTwo) {
-  std::shared_ptr<OperatorBase> add_op(
-      OpRegistry::CreateOp("add_two", {"x", "y"}, {"out"}, {}));
-  std::shared_ptr<OperatorBase> grad_add_op = OpRegistry::CreateGradOp(*add_op);
+  std::shared_ptr<f::OperatorBase> add_op(
+      f::OpRegistry::CreateOp("add_two", {"x", "y"}, {"out"}, {}));
+  std::shared_ptr<f::OperatorBase> grad_add_op =
+      f::OpRegistry::CreateGradOp(*add_op);
   EXPECT_EQ(static_cast<int>(grad_add_op->inputs_.size()), 4);
   EXPECT_EQ(static_cast<int>(grad_add_op->outputs_.size()), 2);
   EXPECT_EQ(grad_add_op->Input("X"), "x");
@@ -22,5 +61,85 @@ TEST(GradOpBuilder, AddTwo) {
   EXPECT_EQ(grad_add_op->Output("Y@GRAD"), "y@GRAD");
 }
 
-}  // namespace framework
-}  // namespace paddle
\ No newline at end of file
+REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker);
+REGISTER_GRADIENT_OP(mult_io, mult_io_grad, f::NOP);
+REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker);
+REGISTER_GRADIENT_OP(io_ignored, io_ignored_grad, f::NOP);
+
+TEST(GradOpBuilder, MutiInOut) {
+  f::AttributeMap attrs{{"input_format", std::vector<int>{0, 1, 4, 5}},
+                        {"output_format", std::vector<int>{0, 1, 3}}};
+  std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
+      "mult_io", {"in1", "in2_1", "in2_2", "in2_3", "in3"},
+      {"out1", "out2_1", "out2_2"}, attrs));
+  std::shared_ptr<f::OperatorBase> grad_test_op =
+      f::OpRegistry::CreateGradOp(*test_op);
+
+  ASSERT_EQ(grad_test_op->inputs_.size(), 5UL + 3UL + 3UL);
+  EXPECT_EQ(grad_test_op->Input("In1"), "in1");
+  EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
+            std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
+  EXPECT_EQ(grad_test_op->Input("In3"), "in3");
+  EXPECT_EQ(grad_test_op->Input("Out1"), "out1");
+  EXPECT_EQ(grad_test_op->Inputs("Out2_mult"),
+            std::vector<std::string>({"out2_1", "out2_2"}));
+  EXPECT_EQ(grad_test_op->Input("Out1" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            "out1" + f::OperatorBase::GRAD_VAR_SUFFIX());
+  EXPECT_EQ(
+      grad_test_op->Inputs("Out2_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+      std::vector<std::string>(
+          {"out2_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+           "out2_2" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
+
+  ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
+  EXPECT_EQ(grad_test_op->Output("In1" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            "in1" + f::OperatorBase::GRAD_VAR_SUFFIX());
+  EXPECT_EQ(
+      grad_test_op->Outputs("In2_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+      std::vector<std::string>({"in2_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+                                "in2_2" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+                                "in2_3" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
+  EXPECT_EQ(grad_test_op->Output("In3" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            "in3" + f::OperatorBase::GRAD_VAR_SUFFIX());
+}
+
+TEST(GradOpBuilder, IOIgnoredInGradient) {
+  f::AttributeMap attrs{{"input_format", std::vector<int>{0, 1, 3, 5}},
+                        {"output_format", std::vector<int>{0, 2, 3}}};
+  std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
+      "io_ignored", {"in1", "in2_1", "in2_2", "in3_1", "in3_2"},
+      {"out1_1", "out1_2", "out2"}, attrs));
+  std::shared_ptr<f::OperatorBase> grad_test_op =
+      f::OpRegistry::CreateGradOp(*test_op);
+
+  // 'In2' and 'Out2' are ignored in gradient calculating
+  ASSERT_EQ(grad_test_op->inputs_.size(), 5UL + 3UL + 3UL);
+  EXPECT_EQ(grad_test_op->Input("In1"), "in1");
+  EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
+            std::vector<std::string>({f::OperatorBase::EMPTY_VAR_NAME(),
+                                      f::OperatorBase::EMPTY_VAR_NAME()}));
+  EXPECT_EQ(grad_test_op->Inputs("In3_mult"),
+            std::vector<std::string>({"in3_1", "in3_2"}));
+  EXPECT_EQ(grad_test_op->Inputs("Out1_mult"),
+            std::vector<std::string>({"out1_1", "out1_2"}));
+  EXPECT_EQ(grad_test_op->Input("Out2"), f::OperatorBase::EMPTY_VAR_NAME());
+  EXPECT_EQ(
+      grad_test_op->Inputs("Out1_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+      std::vector<std::string>(
+          {"out1_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+           "out1_2" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
+  EXPECT_EQ(grad_test_op->Input("Out2" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            "out2" + f::OperatorBase::GRAD_VAR_SUFFIX());
+
+  ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
+  EXPECT_EQ(grad_test_op->Output("In1" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+            "in1" + f::OperatorBase::GRAD_VAR_SUFFIX());
+  EXPECT_EQ(
+      grad_test_op->Outputs("In2_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+      std::vector<std::string>({"in2_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+                                "in2_2" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
+  EXPECT_EQ(
+      grad_test_op->Outputs("In3_mult" + f::OperatorBase::GRAD_VAR_SUFFIX()),
+      std::vector<std::string>({"in3_1" + f::OperatorBase::GRAD_VAR_SUFFIX(),
+                                "in3_2" + f::OperatorBase::GRAD_VAR_SUFFIX()}));
+}
diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
index 3e72e391266066de9e4114e68b43b066c15254db..9a975185f04da8df5ba22e457936218756e7c4bc 100644
--- a/paddle/framework/op_registry.h
+++ b/paddle/framework/op_registry.h
@@ -306,8 +306,7 @@ class OpRegistry {
   static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op) {
     PADDLE_ENFORCE(!op.IsNetOp(),
                    "Use framework::Backward to get backward ops");
-    GradOpBuilder builder(op);
-    std::shared_ptr<OperatorBase> grad_op(builder.Build());
+    std::shared_ptr<OperatorBase> grad_op(BuildGradOp(&op));
     grad_op->Init();
     return grad_op;
   }
diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index 55435103489ace11868eed61c38018d8ba357e65..0b588297169540417586d7c167a1265827b683ac 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -55,6 +55,10 @@ class OperatorBase {
   /// e.g. Variable "x@GRAD" is the gradient of varibale "x".
   static std::string GRAD_VAR_SUFFIX() { return "@GRAD"; }
 
+  static std::string GRAD_VAR_NAME(const std::string& name) {
+    return name + GRAD_VAR_SUFFIX();
+  }
+
   /// Variables with this suffix are supposed to be filled up with zeros.
   static std::string ZERO_VAR_SUFFIX() { return "@ZERO"; }
 
diff --git a/paddle/pybind/pybind.cc b/paddle/framework/pybind.cc
similarity index 60%
rename from paddle/pybind/pybind.cc
rename to paddle/framework/pybind.cc
index 4df02d40fe08290d1165e032f0ad013c5986feaa..b9889e483e27e9dad3310a34b5306f073ad1887d 100644
--- a/paddle/pybind/pybind.cc
+++ b/paddle/framework/pybind.cc
@@ -4,7 +4,7 @@ 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
+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,
@@ -17,19 +17,19 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/framework/backward.h"
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
+#include "paddle/framework/tensor_py.h"
+#include "paddle/operators/net_op.h"
+#include "paddle/operators/type_alias.h"
 #include "paddle/platform/enforce.h"
 #include "paddle/platform/place.h"
-#include "paddle/pybind/tensor_bind.h"
 #include "pybind11/numpy.h"
 #include "pybind11/pybind11.h"
 #include "pybind11/stl.h"
 
 namespace py = pybind11;
-namespace pd = paddle::framework;
 
 USE_OP(add_two);
 USE_OP(onehot_cross_entropy);
@@ -41,17 +41,18 @@ USE_OP(sigmoid);
 USE_OP(softmax);
 USE_OP(rowwise_add);
 USE_OP_WITHOUT_KERNEL(recurrent_op);
-
+namespace paddle {
+namespace framework {
 template <typename ClassType>
-void ExposeOperator(ClassType& m) {
+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 {
+           [](const typename ClassType::type &op) -> std::string {
              return op.type_;
            })
       .def("outputs",
-           [](const typename ClassType::type& op) -> std::vector<std::string> {
+           [](const typename ClassType::type &op) -> std::vector<std::string> {
              return op.outputs_;
            })
       .def("__str__", &ClassType::type::DebugString);
@@ -73,88 +74,81 @@ bool IsCompileGPU() {
 PYBIND11_PLUGIN(core) {
   py::module m("core", "C++ core of PaddlePaddle");
 
-  py::class_<pd::Tensor>(m, "Tensor", py::buffer_protocol())
-      .def_buffer([](pd::Tensor& self) -> py::buffer_info {
-        return paddle::pybind::CastToPyBuffer(self);
-      })
+  py::class_<Tensor>(m, "Tensor", py::buffer_protocol())
+      .def_buffer(
+          [](Tensor &self) -> py::buffer_info { return CastToPyBuffer(self); })
       .def("get_dims",
-           [](const pd::Tensor& self) { return pd::vectorize(self.dims()); })
+           [](const Tensor &self) { return vectorize(self.dims()); })
       .def("set_dims",
-           [](pd::Tensor& self, const std::vector<int>& dim) {
-             self.Resize(pd::make_ddim(dim));
+           [](Tensor &self, const std::vector<int> &dim) {
+             self.Resize(make_ddim(dim));
            })
       .def("alloc_float",
-           [](pd::Tensor& self, paddle::platform::GPUPlace& place) {
+           [](Tensor &self, paddle::platform::GPUPlace &place) {
              self.mutable_data<float>(place);
            })
       .def("alloc_float",
-           [](pd::Tensor& self, paddle::platform::CPUPlace& place) {
+           [](Tensor &self, paddle::platform::CPUPlace &place) {
              self.mutable_data<float>(place);
            })
       .def("alloc_int",
-           [](pd::Tensor& self, paddle::platform::CPUPlace& place) {
+           [](Tensor &self, paddle::platform::CPUPlace &place) {
              self.mutable_data<int>(place);
            })
       .def("alloc_int",
-           [](pd::Tensor& self, paddle::platform::GPUPlace& place) {
+           [](Tensor &self, paddle::platform::GPUPlace &place) {
              self.mutable_data<int>(place);
            })
-      .def("set", paddle::pybind::PyCPUTensorSetFromArray<float>)
-      .def("set", paddle::pybind::PyCPUTensorSetFromArray<int>)
+      .def("set", PyCPUTensorSetFromArray<float>)
+      .def("set", PyCPUTensorSetFromArray<int>)
 #ifndef PADDLE_ONLY_CPU
-      .def("set", paddle::pybind::PyCUDATensorSetFromArray<float>)
-      .def("set", paddle::pybind::PyCUDATensorSetFromArray<int>)
+      .def("set", PyCUDATensorSetFromArray<float>)
+      .def("set", PyCUDATensorSetFromArray<int>)
 #endif
-      .def("shape", [](pd::Tensor& self) { return pd::vectorize(self.dims()); })
+      .def("shape", [](Tensor &self) { return vectorize(self.dims()); })
       .def("set_float_element",
-           [](pd::Tensor& self, size_t offset, float f) {
+           [](Tensor &self, size_t offset, float f) {
              // TODO(yuyang18): Only support GPU now.
              self.data<float>()[offset] = f;
            })
-      .def("get_float_element", [](pd::Tensor& self, size_t offset) -> float {
+      .def("get_float_element", [](Tensor &self, size_t offset) -> float {
         // TODO(yuyang18): Only support GPU now.
         return self.data<float>()[offset];
       });
 
-  py::class_<pd::Variable>(m, "Variable", R"DOC(Variable Class.
+  py::class_<Variable>(m, "Variable", R"DOC(Variable Class.
 
 All parameter, weight, gradient are variables in Paddle.
 )DOC")
-      .def("is_int", [](const pd::Variable& var) { return var.IsType<int>(); })
+      .def("is_int", [](const Variable &var) { return var.IsType<int>(); })
       .def("set_int",
-           [](pd::Variable& var, int val) -> void {
-             *var.GetMutable<int>() = val;
-           })
-      .def("get_int",
-           [](const pd::Variable& var) -> int { return var.Get<int>(); })
+           [](Variable &var, int val) -> void { *var.GetMutable<int>() = val; })
+      .def("get_int", [](const Variable &var) -> int { return var.Get<int>(); })
       .def("get_tensor",
-           [](pd::Variable& self) -> pd::Tensor* {
-             return self.GetMutable<pd::Tensor>();
-           },
+           [](Variable &self) -> Tensor * { return self.GetMutable<Tensor>(); },
            py::return_value_policy::reference)
       .def("get_net",
-           [](pd::Variable& self) -> pd::NetOp* {
-             return self.GetMutable<pd::NetOp>();
+           [](Variable &self) -> ops::NetOp * {
+             return self.GetMutable<ops::NetOp>();
            },
            py::return_value_policy::reference);
 
-  py::class_<pd::Scope>(m, "Scope", "")
+  py::class_<Scope>(m, "Scope", "")
       .def("new_var",
-           [](pd::Scope& self, const std::string& name) -> pd::Variable* {
+           [](Scope &self, const std::string &name) -> Variable * {
              return self.NewVar(name);
            },
            py::return_value_policy::reference)
-      .def("find_var", &pd::Scope::FindVar, py::return_value_policy::reference)
+      .def("find_var", &Scope::FindVar, py::return_value_policy::reference)
       .def(py::init<>())
-      .def("new_scope",
-           [](pd::Scope& self) -> pd::Scope* { return &self.NewScope(); },
+      .def("new_scope", [](Scope &self) -> Scope * { return &self.NewScope(); },
            py::return_value_policy::reference)
-      .def("drop_kids", &pd::Scope::DropKids);
+      .def("drop_kids", &Scope::DropKids);
 
   //! @note: Be careful! PyBind will return std::string as an unicode, not
   //! Python str. If you want a str object, you should cast them in Python.
   m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
-    auto& protos = pd::OpRegistry::protos();
+    auto &protos = OpRegistry::protos();
     std::vector<py::bytes> ret_values;
     for (auto it = protos.begin(); it != protos.end(); ++it) {
       PADDLE_ENFORCE(it->second.IsInitialized(),
@@ -169,8 +163,8 @@ All parameter, weight, gradient are variables in Paddle.
   m.def_submodule(
        "var_names",
        "The module will return special predefined variable name in Paddle")
-      .def("empty", pd::OperatorBase::EMPTY_VAR_NAME)
-      .def("temp", pd::OperatorBase::TMP_VAR_NAME);
+      .def("empty", OperatorBase::EMPTY_VAR_NAME)
+      .def("temp", OperatorBase::TMP_VAR_NAME);
   // clang-format off
   py::class_<paddle::platform::DeviceContext>(m, "DeviceContext")
       .def_static("create",
@@ -193,43 +187,45 @@ All parameter, weight, gradient are variables in Paddle.
 
   py::class_<paddle::platform::CPUPlace>(m, "CPUPlace").def(py::init<>());
 
-  py::class_<pd::OperatorBase, std::shared_ptr<pd::OperatorBase>> operator_base(
+  py::class_<OperatorBase, std::shared_ptr<OperatorBase>> operator_base(
       m, "Operator");
 
   operator_base.def_static("create", [](py::bytes protobin) {
-    pd::OpDesc desc;
+    OpDesc desc;
     PADDLE_ENFORCE(desc.ParsePartialFromString(protobin),
                    "Cannot parse user input to OpDesc");
     PADDLE_ENFORCE(desc.IsInitialized(),
                    "User OpDesc is not initialized, reason %s",
                    desc.InitializationErrorString());
-    return pd::OpRegistry::CreateOp(desc);
+    return 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);
+                    [](const OperatorBase &forwardOp,
+                       const std::unordered_set<std::string> &no_grad_vars) {
+                      return Backward(forwardOp, no_grad_vars);
                     });
 
   ExposeOperator(operator_base);
 
-  py::class_<pd::NetOp, std::shared_ptr<pd::NetOp>> net(m, "Net");
+  py::class_<ops::NetOp, std::shared_ptr<ops::NetOp>> net(m, "Net");
 
   net.def_static("create",
-                 []() -> std::shared_ptr<pd::NetOp> {
-                   auto retv = std::make_shared<pd::NetOp>();
+                 []() -> std::shared_ptr<ops::NetOp> {
+                   auto retv = std::make_shared<ops::NetOp>();
                    retv->type_ = "plain_net";
                    return retv;
                  })
-      .def("add_op", &pd::NetOp::AddOp)
-      .def("add_op",
-           [](pd::NetOp& self, const std::shared_ptr<pd::NetOp>& net) -> void {
-             self.AddOp(std::static_pointer_cast<pd::OperatorBase>(net));
-           })
-      .def("complete_add_op", &pd::NetOp::CompleteAddOp)
+      .def("add_op", &ops::NetOp::AddOp)
+      .def(
+          "add_op",
+          [](ops::NetOp &self, const std::shared_ptr<ops::NetOp> &net) -> void {
+            self.AddOp(std::static_pointer_cast<OperatorBase>(net));
+          })
+      .def("complete_add_op", &ops::NetOp::CompleteAddOp)
       .def("complete_add_op",
-           [](std::shared_ptr<pd::NetOp>& self) { self->CompleteAddOp(); });
+           [](std::shared_ptr<ops::NetOp> &self) { self->CompleteAddOp(); });
+
   ExposeOperator(net);
 
   m.def("unique_integer", UniqueIntegerGenerator);
@@ -238,3 +234,5 @@ All parameter, weight, gradient are variables in Paddle.
 
   return m.ptr();
 }
+}  // namespace framework
+}  // namespace paddle
diff --git a/paddle/framework/tensor.h b/paddle/framework/tensor.h
index 76070f636b0971f4a136042e056c59adb5dc2d40..4c3b14b83d841e88683a13634c93f51c012128b6 100644
--- a/paddle/framework/tensor.h
+++ b/paddle/framework/tensor.h
@@ -26,19 +26,17 @@ limitations under the License. */
 #include "unsupported/Eigen/CXX11/Tensor"
 
 namespace paddle {
-namespace pybind {
-namespace details {  // forward declare
-template <bool less, size_t i, typename... args>
-struct CastToPyBufferImpl;
-}  // namespace details
-}  // namespace pybind
 
 namespace framework {
+namespace details {
+template <bool less, size_t i, typename... args>
+struct CastToPyBufferImpl;
+}
 
 class Tensor {
  public:
   template <bool less, size_t i, typename... args>
-  friend struct paddle::pybind::details::CastToPyBufferImpl;
+  friend struct details::CastToPyBufferImpl;
 
   template <typename T, size_t D, int MajorType, typename IndexType>
   friend struct EigenTensor;
@@ -167,4 +165,4 @@ class Tensor {
 }  // namespace framework
 }  // namespace paddle
 
-#include "paddle/framework/detail/tensor-inl.h"
+#include "paddle/framework/tensor_impl.h"
diff --git a/paddle/framework/detail/tensor-inl.h b/paddle/framework/tensor_impl.h
similarity index 100%
rename from paddle/framework/detail/tensor-inl.h
rename to paddle/framework/tensor_impl.h
diff --git a/paddle/pybind/tensor_bind.h b/paddle/framework/tensor_py.h
similarity index 92%
rename from paddle/pybind/tensor_bind.h
rename to paddle/framework/tensor_py.h
index def37219ccefd5435f1212c4e4daac5a351d76f4..4e1ab77b157fe1adaeac55c271c056236f2d40de 100644
--- a/paddle/pybind/tensor_bind.h
+++ b/paddle/framework/tensor_py.h
@@ -23,7 +23,7 @@ namespace py = pybind11;
 
 namespace paddle {
 
-namespace pybind {
+namespace framework {
 
 namespace details {
 
@@ -63,11 +63,8 @@ struct CastToPyBufferImpl<true, I, ARGS...> {
       }
       return py::buffer_info(
           dst_tensor.mutable_data<CUR_TYPE>(dst_tensor.holder_->place()),
-          sizeof(CUR_TYPE),
-          py::format_descriptor<CUR_TYPE>::format(),
-          (size_t)framework::arity(dst_tensor.dims()),
-          dims_outside,
-          strides);
+          sizeof(CUR_TYPE), py::format_descriptor<CUR_TYPE>::format(),
+          (size_t)framework::arity(dst_tensor.dims()), dims_outside, strides);
     } else {
       constexpr bool less = I + 1 < std::tuple_size<std::tuple<ARGS...>>::value;
       return CastToPyBufferImpl<less, I + 1, ARGS...>()(tensor);
@@ -110,8 +107,8 @@ void PyCUDATensorSetFromArray(
 
   self.Resize(framework::make_ddim(dims));
   auto *dst = self.mutable_data<T>(place);
-  paddle::platform::GpuMemcpySync(
-      dst, array.data(), sizeof(T) * array.size(), cudaMemcpyHostToDevice);
+  paddle::platform::GpuMemcpySync(dst, array.data(), sizeof(T) * array.size(),
+                                  cudaMemcpyHostToDevice);
 }
 #endif
 
diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt
index b910bee836ed488aeb34f28d0503b5efba396583..96c76e22e9814682008f2e6c7ae98e2599d391c2 100644
--- a/paddle/operators/CMakeLists.txt
+++ b/paddle/operators/CMakeLists.txt
@@ -41,6 +41,9 @@ function(op_library TARGET)
     endif()
 endfunction()
 
+cc_library(net_op SRCS net_op.cc DEPS op_registry)
+cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
+
 op_library(add_op SRCS add_op.cc add_op.cu)
 cc_test(add_op_test SRCS add_op_test.cc DEPS add_op)
 
@@ -59,11 +62,6 @@ op_library(sgd_op SRCS sgd_op.cc sgd_op.cu)
 
 op_library(fc_op
     SRCS fc_op.cc
-    DEPS mul_op rowwise_add_op sigmoid_op softmax_op net)
-
-op_library(recurrent_network_op
-    SRCS recurrent_network_op.cc
-    DEPS op_desc tensor net)
-cc_test(recurrent_network_op_test
-    SRCS recurrent_network_op_test.cc
-    DEPS recurrent_network_op mul_op add_op)
+    DEPS mul_op rowwise_add_op sigmoid_op softmax_op net_op)
+op_library(recurrent_op SRCS recurrent_op.cc DEPS op_desc tensor op_registry operator net_op)
+cc_test(recurrent_op_test SRCS recurrent_op_test.cc DEPS recurrent_op gtest mul_op add_op)
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index fe34d6ad4015620cac520146850e10563d4c50e0..78131b26808b183ee107313374493ae870f1b641 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -33,13 +33,23 @@ public:
   MeanOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input of mean op");
-    AddOutput("Out", "The output of mean op");
+    AddOutput("Out", "The output of mean op").IgnoreGradient();
     AddComment("Mean Operator");
   }
 };
 
+class MeanGradOp : public OperatorWithKernel {
+protected:
+  void InferShape(const InferShapeContext &ctx) const override {
+    ctx.Output<Tensor>("X" + GRAD_VAR_SUFFIX())
+        ->Resize(ctx.Input<Tensor>("X")->dims());
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker);
 REGISTER_OP_CPU_KERNEL(mean, ops::MeanKernel<ops::CPUPlace, float>);
+REGISTER_GRADIENT_OP(mean, mean_grad, ops::MeanGradOp);
+REGISTER_OP_CPU_KERNEL(mean_grad, ops::MeanGradKernel<ops::CPUPlace, float>);
diff --git a/paddle/operators/mean_op.cu b/paddle/operators/mean_op.cu
index 740157cbc57a64cafcf109186c630691620f542b..e15de2fd0dd84e4015ee0e3b5343d7651b027a88 100644
--- a/paddle/operators/mean_op.cu
+++ b/paddle/operators/mean_op.cu
@@ -3,3 +3,4 @@
 #include "paddle/operators/mean_op.h"
 
 REGISTER_OP_GPU_KERNEL(mean, ops::MeanKernel<ops::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(mean_grad, ops::MeanGradKernel<ops::GPUPlace, float>);
\ No newline at end of file
diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h
index 5c339bffbf8e39f36ee9b4f857ab380cbac82879..e712dee6a785749e51be7b233e85dbf39c835218 100644
--- a/paddle/operators/mean_op.h
+++ b/paddle/operators/mean_op.h
@@ -35,5 +35,22 @@ public:
   }
 };
 
+template <typename Place, typename T>
+class MeanGradKernel : public OpKernel {
+public:
+  void Compute(const ExecutionContext& context) const override {
+    auto OG = context.Input<Tensor>("Out" + OperatorBase::GRAD_VAR_SUFFIX());
+    PADDLE_ENFORCE(framework::product(OG->dims()) == 1,
+                   "Mean Gradient should be scalar");
+    auto IG = context.Output<Tensor>("X" + OperatorBase::GRAD_VAR_SUFFIX());
+    IG->mutable_data<T>(context.GetPlace());
+
+    T ig_size = (T)framework::product(IG->dims());
+
+    EigenVector<T>::Flatten(*IG).device(context.GetEigenDevice<Place>()) =
+        EigenScalar<T>::From(*OG) / ig_size;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/framework/net.cc b/paddle/operators/net_op.cc
similarity index 96%
rename from paddle/framework/net.cc
rename to paddle/operators/net_op.cc
index 2cd378c6b21303d1a24206ba3010b0d035aaa766..fbc98e09923bda7f3baee04e02df9076247bff0b 100644
--- a/paddle/framework/net.cc
+++ b/paddle/operators/net_op.cc
@@ -14,11 +14,11 @@
   limitations under the License.
 */
 
-#include "paddle/framework/net.h"
+#include "paddle/operators/net_op.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
-namespace framework {
+namespace operators {
 
 void NetOp::CompleteAddOp(bool calc) {
   add_op_done_ = true;
@@ -74,5 +74,5 @@ std::string NetOp::DebugString() const {
 
 bool NetOp::IsNetOp() const { return true; }
 
-}  // namespace framework
+}  // namespace operators
 }  // namespace paddle
diff --git a/paddle/framework/net.h b/paddle/operators/net_op.h
similarity index 89%
rename from paddle/framework/net.h
rename to paddle/operators/net_op.h
index acf1a69da9fd8adce1bd89367c882eade052e725..13611e1ee83170db43e17d6088e4b04588ce6255 100644
--- a/paddle/framework/net.h
+++ b/paddle/operators/net_op.h
@@ -14,15 +14,17 @@ limitations under the License. */
 
 #pragma once
 
-#include <paddle/framework/op_desc.pb.h>
-#include <paddle/framework/operator.h>
+#include "paddle/framework/op_desc.pb.h"
 #include "paddle/framework/op_proto.pb.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
+#include "paddle/operators/type_alias.h"
 #include "paddle/platform/device_context.h"
 
 namespace paddle {
-namespace framework {
+namespace operators {
+
 /**
  * @brief Network is also a type of Operator
  *
@@ -37,13 +39,13 @@ namespace framework {
  * This is the base class of network, all the networks should implement the APIs
  * it defines.
  */
-class NetOp : public OperatorBase {
- public:
+class NetOp : public framework::OperatorBase {
+public:
   /**
    * Infer all the operators' input and output variables' shapes, will be called
    * before every mini-batch
    */
-  void InferShape(const Scope& scope) const override {
+  void InferShape(const framework::Scope& scope) const override {
     for (auto& op : ops_) {
       op->InferShape(scope);
     }
@@ -56,7 +58,7 @@ class NetOp : public OperatorBase {
    * scope will be used instead. If no OpContext is provicded, default context
    * will be used.
    */
-  void Run(const Scope& scope,
+  void Run(const framework::Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {
     for (auto& op : ops_) {
       op->Run(scope, dev_ctx);
@@ -88,7 +90,7 @@ class NetOp : public OperatorBase {
 
   std::vector<std::shared_ptr<OperatorBase>> ops_;
 
- private:
+private:
   bool add_op_done_{false};
 
   template <typename T, typename KeyType>
@@ -97,5 +99,5 @@ class NetOp : public OperatorBase {
   }
 };
 
-}  // namespace framework
+}  // namespace operators
 }  // namespace paddle
diff --git a/paddle/framework/net_design.md b/paddle/operators/net_op_design.md
similarity index 100%
rename from paddle/framework/net_design.md
rename to paddle/operators/net_op_design.md
diff --git a/paddle/framework/net_op_test.cc b/paddle/operators/net_op_test.cc
similarity index 91%
rename from paddle/framework/net_op_test.cc
rename to paddle/operators/net_op_test.cc
index f32e456e5d142bf8203f9ec03e8059772c4f5c99..18c5c60eb43250c23e2819a3c79ab8a96fec103e 100644
--- a/paddle/framework/net_op_test.cc
+++ b/paddle/operators/net_op_test.cc
@@ -1,16 +1,18 @@
+#include "paddle/operators/net_op.h"
+
 #include <gtest/gtest.h>
-#include <paddle/framework/net.h>
-#include <paddle/framework/op_registry.h>
-#include <paddle/framework/operator.h>
+
+#include "paddle/framework/op_registry.h"
+#include "paddle/framework/operator.h"
 
 namespace paddle {
-namespace framework {
+namespace operators {
 
 static int infer_shape_cnt = 0;
 static int run_cnt = 0;
 
 class TestOp : public OperatorBase {
- public:
+public:
   void InferShape(const framework::Scope& scope) const override {
     ++infer_shape_cnt;
   }
@@ -21,7 +23,7 @@ class TestOp : public OperatorBase {
 };
 
 class EmptyOp : public OperatorBase {
- public:
+public:
   void InferShape(const Scope& scope) const override {}
   void Run(const Scope& scope,
            const platform::DeviceContext& dev_ctx) const override {}
@@ -73,7 +75,7 @@ TEST(OpKernel, all) {
   ASSERT_THROW(net->AddOp(op2), paddle::platform::EnforceNotMet);
 }
 
-TEST(Net, insert_op) {
+TEST(NetOp, insert_op) {
   NetOp net;
   auto op1 = std::make_shared<EmptyOp>();
   op1->inputs_ = {"x", "w1", "b1"};
@@ -85,5 +87,5 @@ TEST(Net, insert_op) {
   ASSERT_EQ(3UL, net.ops_.size());
 }
 
-}  // namespace framework
+}  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/recurrent_network_op.cc b/paddle/operators/recurrent_op.cc
similarity index 67%
rename from paddle/operators/recurrent_network_op.cc
rename to paddle/operators/recurrent_op.cc
index 60d065fc4789f76370840328870165579aa73b67..aeb95569b728f53b288a0c9a28220be8b5f7aaa4 100644
--- a/paddle/operators/recurrent_network_op.cc
+++ b/paddle/operators/recurrent_op.cc
@@ -12,14 +12,14 @@
    See the License for the specific language governing permissions and
    limitations under the License. */
 
-#include "paddle/operators/recurrent_network_op.h"
+#include "paddle/operators/recurrent_op.h"
 
 #include <glog/logging.h>
 #include <cstring>
 #include <sstream>
 
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
 #include "paddle/platform/enforce.h"
 
 namespace paddle {
@@ -29,11 +29,15 @@ namespace rnn {
 
 void SegmentInputs(const std::vector<Scope*>& step_scopes,
                    const std::vector<Link>& inlinks,
-                   const size_t seq_len) {
+                   const size_t seq_len,
+                   bool infer_shape_mode) {
   PADDLE_ENFORCE(!inlinks.empty(), "no in links are provided.");
   for (size_t i = 0; i < inlinks.size(); ++i) {
-    Tensor* input =
-        step_scopes[0]->FindVar(inlinks[i].external)->GetMutable<Tensor>();
+    auto input_var = step_scopes[0]->FindVar(inlinks[i].external);
+    PADDLE_ENFORCE(input_var != nullptr,
+                   "input link [%s] is not in scope.",
+                   inlinks[i].external);
+    Tensor* input = input_var->GetMutable<Tensor>();
     DDim dims = input->dims();
     PADDLE_ENFORCE(static_cast<size_t>(dims[0]) == seq_len,
                    "all the inlinks must have same length");
@@ -41,7 +45,9 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
     for (size_t j = 0; j < seq_len; j++) {
       Tensor* step_input =
           step_scopes[j]->NewVar(inlinks[i].internal)->GetMutable<Tensor>();
-      *step_input = input->Slice<float>(j, j + 1);
+      if (!infer_shape_mode) {
+        *step_input = input->Slice<float>(j, j + 1);
+      }
       step_input->Resize(step_dims);
     }
   }
@@ -49,36 +55,41 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
 
 void ConcatOutputs(const std::vector<Scope*>& step_scopes,
                    const std::vector<Link>& outlinks,
-                   const size_t seq_len) {
+                   const size_t seq_len,
+                   bool infer_shape_mode) {
   for (size_t i = 0; i < outlinks.size(); i++) {
-    Tensor* output =
-        step_scopes[0]->FindVar(outlinks[i].external)->GetMutable<Tensor>();
-
-    // TODO(qingiqng) remove following code after adding
-    // InferShape in RecurrentGradientOp
-    DDim step_dims = step_scopes[0]
-                         ->FindVar(outlinks[i].internal)
-                         ->GetMutable<Tensor>()
-                         ->dims();
-    std::vector<int> dims_vec = vectorize(step_dims);
-    dims_vec.insert(dims_vec.begin(), seq_len);
-    output->mutable_data<float>(make_ddim(dims_vec), platform::CPUPlace());
-
-    for (size_t j = 0; j < seq_len; j++) {
-      Tensor* step_output =
-          step_scopes[j]->FindVar(outlinks[i].internal)->GetMutable<Tensor>();
-      // TODO(luotao02) data type and platform::DeviceContext() should set
-      // correctly
-      (output->Slice<float>(j, j + 1))
-          .CopyFrom<float>(*step_output, platform::CPUPlace());
+    auto output_var = step_scopes[0]->FindVar(outlinks[i].external);
+    PADDLE_ENFORCE(output_var != nullptr,
+                   "output link [%s] is not in scope.",
+                   outlinks[i].external);
+    Tensor* output = output_var->GetMutable<Tensor>();
+    if (infer_shape_mode) {
+      DDim step_dims = step_scopes[0]
+                           ->FindVar(outlinks[i].internal)
+                           ->GetMutable<Tensor>()
+                           ->dims();
+      std::vector<int> dims_vec = vectorize(step_dims);
+      dims_vec.insert(dims_vec.begin(), seq_len);
+      output->Resize(make_ddim(dims_vec));
+    } else {
+      output->mutable_data<float>(platform::CPUPlace());
+      for (size_t j = 0; j < seq_len; j++) {
+        Tensor* step_output =
+            step_scopes[j]->FindVar(outlinks[i].internal)->GetMutable<Tensor>();
+        // TODO(luotao02) data type and platform::DeviceContext() should set
+        // correctly
+        (output->Slice<float>(j, j + 1))
+            .CopyFrom<float>(*step_output, platform::CPUPlace());
+      }
     }
   }
 }
 
 void LinkMemories(const std::vector<Scope*>& scopes,
                   const std::vector<rnn::MemoryAttr>& memories,
-                  size_t step_id,
-                  int offset) {
+                  const size_t step_id,
+                  const int offset,
+                  bool infer_shape_mode) {
   PADDLE_ENFORCE(step_id < scopes.size(),
                  "step [%d] is out of range of step scopes' size [%d]",
                  step_id,
@@ -95,18 +106,13 @@ void LinkMemories(const std::vector<Scope*>& scopes,
   auto scope = scopes[step_id];
   auto linked_scope = scopes[step_id + offset];
   for (auto& attr : memories) {
-    auto mem = scope->NewVar(attr.pre_var)->GetMutable<Tensor>();
-    // maybe share variable is better?
+    auto mem = scope->FindVar(attr.pre_var)->GetMutable<Tensor>();
     auto linked_mem = linked_scope->FindVar(attr.var)->GetMutable<Tensor>();
-    mem->ShareDataWith<float>(*linked_mem);
-
-    // TODO(qingqing) remove following code
-    // the memory of current step should be allocated in step net
-    auto m = scope->NewVar(attr.var)->GetMutable<Tensor>();
-    // for unit test, as addOp and mulOp are null currently, if not
-    // mutable_data, mem.data() in output will be error. We will
-    // remove this line after merge the correct addOp and mulOp.
-    m->mutable_data<float>(mem->dims(), platform::CPUPlace());
+    if (infer_shape_mode) {
+      mem->Resize(linked_mem->dims());
+    } else {
+      mem->ShareDataWith<float>(*linked_mem);
+    }
   }
 }
 
@@ -175,60 +181,39 @@ void RecurrentAlgorithm::InferShape(const Scope& scope) const {
                  ->dims()[0];
   CreateScopes(scope);
   auto step_scopes = GetStepScopes(scope);
-
-  // SegmentInputs is called in InferShape. The input must hold memory in
-  // SegmentInputs. But the other op only set dimension for the output in
-  // InferShape. That's a problem. Wether the RNN op needs InferShape or not?
-  // Wether the following functions (SegmentInputs, InitMemories, ...) need
-  // to rewrite for RNN op?
-  rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_);
-
-  InitMemories(step_scopes[0]);
-
-  PADDLE_ENFORCE(scope.FindVar(arg_->step_net) != nullptr,
-                 "stepnet [%s] is not in scope.",
-                 arg_->step_net);
+  rnn::SegmentInputs(
+      step_scopes, arg_->inlinks, seq_len_, true /*infer_shape_mode*/);
+  InitMemories(step_scopes[0], true /*infer_shape_mode*/);
   Variable* net = scope.FindVar(arg_->step_net);
   PADDLE_ENFORCE(net != nullptr, "failed to get step net");
-  // If the InferShape is called in OperatorBase's run function,
-  // the rnn op only needs to do InferShape for the first time step
   for (size_t i = 0; i < seq_len_; i++) {
     if (i > 0) {
-      rnn::LinkMemories(step_scopes, arg_->memories, i, -1);
+      rnn::LinkMemories(
+          step_scopes, arg_->memories, i, -1, true /*infer_shape_mode*/);
     }
     net->GetMutable<NetOp>()->InferShape(*step_scopes[i]);
   }
-
-  auto outlinks = arg_->outlinks;
-  for (size_t i = 0; i < outlinks.size(); i++) {
-    DDim step_dims = step_scopes[0]
-                         ->FindVar(outlinks[i].internal)
-                         ->GetMutable<Tensor>()
-                         ->dims();
-    std::vector<int> dims_vec = vectorize(step_dims);
-    // now only support fixed length
-    dims_vec.insert(dims_vec.begin(), seq_len_);
-    Tensor* output =
-        step_scopes[0]->FindVar(outlinks[i].external)->GetMutable<Tensor>();
-    output->Resize(make_ddim(dims_vec));
-  }
+  rnn::ConcatOutputs(
+      step_scopes, arg_->outlinks, seq_len_, true /*infer_shape_mode*/);
 }
 
 void RecurrentAlgorithm::Run(const Scope& scope,
                              const platform::DeviceContext& dev_ctx) const {
   auto step_scopes = GetStepScopes(scope);
-
+  rnn::SegmentInputs(
+      step_scopes, arg_->inlinks, seq_len_, false /*infer_shape_mode*/);
+  InitMemories(step_scopes[0], false /*infer_shape_mode*/);
   Variable* net = scope.FindVar(arg_->step_net);
+
   for (size_t step_id = 0; step_id < seq_len_; step_id++) {
-    // the link memory is done in InferShape
-    // maybe remove following code after testing
     if (step_id > 0) {
-      rnn::LinkMemories(step_scopes, arg_->memories, step_id, -1);
+      rnn::LinkMemories(
+          step_scopes, arg_->memories, step_id, -1, false /*infer_shape_mode*/);
     }
     net->GetMutable<NetOp>()->Run(*step_scopes[step_id], dev_ctx);
   }
-
-  rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_);
+  rnn::ConcatOutputs(
+      step_scopes, arg_->outlinks, seq_len_, false /*infer_shape_mode*/);
 }
 
 void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
@@ -244,18 +229,19 @@ void RecurrentAlgorithm::CreateScopes(const Scope& scope) const {
       // Now all variables in scope must be created outside of op.
       auto net_op = scope.FindVar(arg_->step_net)->GetMutable<NetOp>();
       for (auto& input : net_op->inputs_) {
+        // the weight are located in parent scope
         if (!step_scope.FindVar(input)) step_scope.NewVar(input);
       }
       for (auto& output : net_op->outputs_) {
         step_scope.NewVar(output);
       }
-
       step_scopes->emplace_back(&step_scope);
     }
   }
 }
 
-void RecurrentAlgorithm::InitMemories(Scope* step_scope) const {
+void RecurrentAlgorithm::InitMemories(Scope* step_scope,
+                                      bool infer_shape_mode) const {
   for (auto& attr : arg_->memories) {
     Tensor* pre_mem = step_scope->NewVar(attr.pre_var)->GetMutable<Tensor>();
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
@@ -263,13 +249,11 @@ void RecurrentAlgorithm::InitMemories(Scope* step_scope) const {
                    attr.var,
                    attr.boot_var);
     Tensor* boot_mem = step_scope->FindVar(attr.boot_var)->GetMutable<Tensor>();
-    pre_mem->ShareDataWith<float>(*boot_mem);
-
-    // TODO(qingqing) remove following code
-    // the memory of current step should be allocated in step net
-    // here for unit test
-    auto cur_step_mem = step_scope->NewVar(attr.var)->GetMutable<Tensor>();
-    cur_step_mem->mutable_data<float>(boot_mem->dims(), platform::CPUPlace());
+    if (infer_shape_mode) {
+      pre_mem->Resize(boot_mem->dims());
+    } else {
+      pre_mem->ShareDataWith<float>(*boot_mem);
+    }
   }
 }
 
@@ -307,13 +291,14 @@ public:
       : OpProtoAndCheckerMaker(proto, op_checker) {
     const auto& name = RecurrentOp::kArgName;
     // inputs and outputs stored in proto
-    AddInput(name.inlinks, "the input that need to be segmented for each step.")
+    AddInput(name.inlinks,
+             "the inputs that need to be segmented for each step.")
         .SetMultiple();
     AddInput(name.boot_memories, "variables to initialize memories.")
         .SetMultiple();
     AddInput(name.step_net, "network shared by all steps.");
 
-    AddOutput(name.outlinks, "the output that need to concated for all steps.")
+    AddOutput(name.outlinks, "the outputs that need to concated for all steps.")
         .SetMultiple();
     AddOutput(name.step_scopes, "step scopes");
 
@@ -331,34 +316,39 @@ public:
 void RecurrentGradientAlgorithm::Run(
     const Scope& scope, const platform::DeviceContext& dev_ctx) const {
   auto step_scopes = GetStepScopes(scope);
-  rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_);
-  PADDLE_ENFORCE(scope.FindVar(arg_->step_net) != nullptr,
-                 "step net is not in scope.");
+  rnn::SegmentInputs(
+      step_scopes, arg_->inlinks, seq_len_, false /*infer_shape_mode*/);
   Variable* net = scope.FindVar(arg_->step_net);
   PADDLE_ENFORCE(net != nullptr, "failed to get step net");
   for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) {
     if (static_cast<size_t>(step_id) != seq_len_ - 1) {
-      rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1);
+      rnn::LinkMemories(
+          step_scopes, arg_->memories, step_id, 1, false /*infer_shape_mode*/);
     }
     net->GetMutable<NetOp>()->Run(*step_scopes[step_id], dev_ctx);
   }
-  LinkBootMemoryGradients(step_scopes[0]);
-  rnn::ConcatOutputs(step_scopes, arg_->outlinks, seq_len_);
+  LinkBootMemoryGradients(step_scopes[0], false);
+  rnn::ConcatOutputs(
+      step_scopes, arg_->outlinks, seq_len_, false /*infer_shape_mode*/);
 }
 
 void RecurrentGradientAlgorithm::LinkBootMemoryGradients(
-    Scope* step_scope) const {
+    Scope* step_scope, bool infer_shape_mode) const {
   for (auto& attr : arg_->memories) {
-    Tensor* mem_grad = step_scope->NewVar(attr.var)->GetMutable<Tensor>();
-    PADDLE_ENFORCE(mem_grad != nullptr,
-                   "boot_tensor should be retrieved before");
+    PADDLE_ENFORCE(step_scope->FindVar(attr.var) != nullptr,
+                   "memory variable [%s] does not exists",
+                   attr.var);
     PADDLE_ENFORCE(step_scope->FindVar(attr.boot_var) != nullptr,
-                   "memory [%s]'s boot variable [%s] not exists",
-                   attr.var,
+                   "boot variable [%s] does not exists",
                    attr.boot_var);
+    Tensor* mem_grad = step_scope->NewVar(attr.var)->GetMutable<Tensor>();
     Tensor* boot_mem_grad =
         step_scope->NewVar(attr.boot_var)->GetMutable<Tensor>();
-    boot_mem_grad->ShareDataWith<float>(*mem_grad);
+    if (infer_shape_mode) {
+      boot_mem_grad->Resize(mem_grad->dims());
+    } else {
+      boot_mem_grad->ShareDataWith<float>(*mem_grad);
+    }
   }
 }
 
@@ -367,34 +357,20 @@ void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const {
                  ->GetMutable<Tensor>()
                  ->dims()[0];
   auto step_scopes = GetStepScopes(scope);
-  rnn::SegmentInputs(step_scopes, arg_->inlinks, seq_len_);
-
-  PADDLE_ENFORCE(scope.FindVar(arg_->step_net) != nullptr,
-                 "step net is not in scope.");
+  rnn::SegmentInputs(
+      step_scopes, arg_->inlinks, seq_len_, true /*infer_shape_mode*/);
   Variable* net = scope.FindVar(arg_->step_net);
   PADDLE_ENFORCE(net != nullptr, "failed to get step net");
-
   for (int step_id = seq_len_ - 1; step_id >= 0; --step_id) {
     if (static_cast<size_t>(step_id) != seq_len_ - 1) {
-      rnn::LinkMemories(step_scopes, arg_->memories, step_id, 1);
+      rnn::LinkMemories(
+          step_scopes, arg_->memories, step_id, 1, true /*infer_shape_mode*/);
     }
     net->GetMutable<NetOp>()->InferShape(*step_scopes[step_id]);
   }
-
-  auto outlinks = arg_->outlinks;
-  for (size_t i = 0; i < outlinks.size(); i++) {
-    DDim step_dims = step_scopes[0]
-                         ->FindVar(outlinks[i].internal)
-                         ->GetMutable<Tensor>()
-                         ->dims();
-    std::vector<int> dims_vec = vectorize(step_dims);
-    // now only support fixed length
-    dims_vec.insert(dims_vec.begin(), seq_len_);
-    Tensor* output =
-        step_scopes[0]->FindVar(outlinks[i].external)->GetMutable<Tensor>();
-    output->Resize(make_ddim(dims_vec));
-  }
-  LinkBootMemoryGradients(step_scopes[0]);
+  rnn::ConcatOutputs(
+      step_scopes, arg_->outlinks, seq_len_, true /*infer_shape_mode*/);
+  LinkBootMemoryGradients(step_scopes[0], true /*infer_shape_mode*/);
 }
 
 void RecurrentGradientOp::Init() {
diff --git a/paddle/operators/recurrent_network_op.h b/paddle/operators/recurrent_op.h
similarity index 92%
rename from paddle/operators/recurrent_network_op.h
rename to paddle/operators/recurrent_op.h
index d57a1a2e51cbed22549ab6ebce79223e2d4e3bcf..2a0964fff326500b6215dd4afac63c75d64c4a06 100644
--- a/paddle/operators/recurrent_network_op.h
+++ b/paddle/operators/recurrent_op.h
@@ -72,19 +72,22 @@ struct ArgumentName {
  */
 void SegmentInputs(const std::vector<Scope*>& step_scopes,
                    const std::vector<Link>& inlinks,
-                   const size_t seq_len);
+                   const size_t seq_len,
+                   bool infer_shape_mode);
 
 /**
  * Process outputs of step nets and merge to variables.
  */
 void ConcatOutputs(const std::vector<Scope*>& step_scopes,
                    const std::vector<Link>& outlinks,
-                   const size_t seq_len);
+                   const size_t seq_len,
+                   bool infer_shape_mode);
 
 void LinkMemories(const std::vector<Scope*>& step_scopes,
                   const std::vector<MemoryAttr>& memories,
-                  size_t step_id,
-                  int offset);
+                  const size_t step_id,
+                  const int offset,
+                  bool infer_shape_mode);
 
 void InitArgument(const ArgumentName& name, Argument* arg);
 
@@ -122,7 +125,7 @@ protected:
     return *scope.FindVar(arg_->step_scopes)->GetMutable<std::vector<Scope*>>();
   }
 
-  void InitMemories(Scope* step_scopes) const;
+  void InitMemories(Scope* step_scopes, bool infer_shape_mode) const;
 
 private:
   std::unique_ptr<rnn::Argument> arg_;
@@ -145,7 +148,7 @@ public:
 
   void Run(const Scope& scope, const platform::DeviceContext& dev_ctx) const;
 
-  void LinkBootMemoryGradients(Scope* step_scopes) const;
+  void LinkBootMemoryGradients(Scope* step_scopes, bool infer_shape_mode) const;
 
   /**
    * InferShape must be called before Run.
diff --git a/paddle/operators/recurrent_network_op_test.cc b/paddle/operators/recurrent_op_test.cc
similarity index 90%
rename from paddle/operators/recurrent_network_op_test.cc
rename to paddle/operators/recurrent_op_test.cc
index b0e61fbee611744adb85b498b1c3540f059afc8c..08a6d9fe5681fdea180de2e9361734ade8564775 100644
--- a/paddle/operators/recurrent_network_op_test.cc
+++ b/paddle/operators/recurrent_op_test.cc
@@ -11,14 +11,15 @@
   limitations under the License.
 */
 
+#include "paddle/operators/recurrent_op.h"
+
 #include <glog/logging.h>
 #include <gtest/gtest.h>
 
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/tensor.h"
-#include "paddle/operators/recurrent_network_op.h"
+#include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace operators {
@@ -55,7 +56,7 @@ protected:
     w->GetMutable<Tensor>()->mutable_data<float>(
         make_ddim(std::vector<int>{30, 30}), platform::CPUPlace());
 
-    for (auto boot : std::vector<std::string>{"x_boot", "h_boot"}) {
+    for (auto boot : std::vector<std::string>{"h_boot"}) {
       LOG(INFO) << "create global variable " << boot;
       Variable* h_boot = scope_.NewVar(boot);
       h_boot->GetMutable<Tensor>()->mutable_data<float>(
@@ -79,7 +80,6 @@ protected:
     op_desc.add_inputs("x0");
     op_desc.add_inputs("x1");
     // boot_memories 3
-    op_desc.add_inputs("x_boot");
     op_desc.add_inputs("h_boot");
     // step net 5
     op_desc.add_inputs("step_net");
@@ -91,7 +91,7 @@ protected:
     auto _input_format = std::vector<int>{
         0,  // in_link
         3,  // memories
-        5   // step_net
+        4   // step_net
     };
     auto input_format = op_desc.add_attrs();
     input_format->set_name("input_format");
@@ -129,12 +129,11 @@ protected:
       inlink_alias->add_strings(item);
     }
     // pre memories
-    for (const auto& item :
-         std::vector<std::string>{"rnn/x@pre", "rnn/h@pre"}) {
+    for (const auto& item : std::vector<std::string>{"rnn/h@pre"}) {
       pre_memories->add_strings(item);
     }
     // memories
-    for (const auto& item : std::vector<std::string>{"rnn/x", "rnn/h"}) {
+    for (const auto& item : std::vector<std::string>{"rnn/h"}) {
       memories->add_strings(item);
     }
     // output alias
@@ -151,14 +150,11 @@ protected:
     LOG(INFO) << "create variable step_net";
     Variable* var = scope_.NewVar("step_net");
     auto net = var->GetMutable<NetOp>();
-    // rnn/s is net's input or output?
-    net->inputs_ = {"rnn/h@pre", "rnn/w", "rnn/x"};
-    net->inputs_ = {"rnn/s", "rnn/h"};
     net->AddOp(
         OpRegistry::CreateOp("mul", {"rnn/h@pre", "rnn/w"}, {"rnn/s"}, {}));
 
     net->AddOp(
-        OpRegistry::CreateOp("add_two", {"rnn/x", "rnn/s"}, {"rnn/h"}, {}));
+        OpRegistry::CreateOp("add_two", {"x@alias", "rnn/s"}, {"rnn/h"}, {}));
     net->CompleteAddOp();
   }
 
@@ -297,7 +293,10 @@ protected:
     inlink.internal = "rnn/x";
     auto step_scopes =
         scope_.FindVar("step_scopes")->GetMutable<std::vector<Scope*>>();
-    rnn::SegmentInputs(*step_scopes, std::vector<rnn::Link>{inlink}, 10);
+    rnn::SegmentInputs(*step_scopes,
+                       std::vector<rnn::Link>{inlink},
+                       10,
+                       true /*infer_shape_mode*/);
   }
 
   void LinkeMemories() {
@@ -311,7 +310,8 @@ protected:
     auto step_scopes =
         scope_.FindVar("step_scopes")->GetMutable<std::vector<Scope*>>();
     for (int i = 1; i < 10; ++i) {
-      rnn::LinkMemories(*step_scopes, memories, i, -1);
+      rnn::LinkMemories(
+          *step_scopes, memories, i, -1, true /*infer_shape_mode*/);
     }
   }
 
@@ -333,14 +333,14 @@ TEST(RecurrentOp, LinkMemories) {
   using namespace paddle::operators;
 
   // create and init step scopes
-  int len = 10;
+  size_t len = 10;
   std::vector<Scope*> step_scopes;
-  for (int i = 0; i < len; ++i) {
+  for (size_t i = 0; i < len; ++i) {
     auto scope = new Scope();
     scope->NewVar("pre_h");
     auto tensor = scope->NewVar("h")->GetMutable<Tensor>();
     float* data = tensor->mutable_data<float>({15, 20}, CPUPlace());
-    for (int j = 0; j < 15 * 20; ++j) {
+    for (size_t j = 0; j < 15 * 20; ++j) {
       data[j] = rand() * (1. / (double)RAND_MAX);
     }
     step_scopes.push_back(scope);
@@ -354,24 +354,24 @@ TEST(RecurrentOp, LinkMemories) {
   std::vector<rnn::MemoryAttr> memories;
   memories.push_back(mem_attr);
 
-  for (int i = 1; i < len; ++i) {
-    rnn::LinkMemories(step_scopes, memories, i, -1);
+  for (size_t i = 1; i < len; ++i) {
+    rnn::LinkMemories(step_scopes, memories, i, -1, false /*infer_shape_mode*/);
   }
   // check
-  for (int i = 0; i < len - 1; ++i) {
+  for (size_t i = 0; i < len - 1; ++i) {
     const float* a =
         step_scopes[i]->FindVar("h")->GetMutable<Tensor>()->data<float>();
     const float* b = step_scopes[i + 1]
                          ->FindVar("pre_h")
                          ->GetMutable<Tensor>()
                          ->data<float>();
-    for (size_t i = 0; i < 15 * 20; ++i) {
-      ASSERT_FLOAT_EQ(a[i], b[i]);
+    for (size_t j = 0; j < 15 * 20; ++j) {
+      ASSERT_FLOAT_EQ(a[j], b[j]);
     }
   }
 
   for (int i = len - 2; i >= 0; --i) {
-    rnn::LinkMemories(step_scopes, memories, i, 1);
+    rnn::LinkMemories(step_scopes, memories, i, 1, false /*infer_shape_mode*/);
   }
   // check
   for (int i = len - 2; i >= 0; --i) {
@@ -379,8 +379,8 @@ TEST(RecurrentOp, LinkMemories) {
         step_scopes[i]->FindVar("pre_h")->GetMutable<Tensor>()->data<float>();
     const float* b =
         step_scopes[i + 1]->FindVar("h")->GetMutable<Tensor>()->data<float>();
-    for (size_t i = 0; i < 15 * 20; ++i) {
-      ASSERT_FLOAT_EQ(a[i], b[i]);
+    for (size_t j = 0; j < 15 * 20; ++j) {
+      ASSERT_FLOAT_EQ(a[j], b[j]);
     }
   }
 
@@ -391,9 +391,3 @@ TEST(RecurrentOp, LinkMemories) {
 
 USE_OP(add_two);
 USE_OP(mul);
-
-// int main() {
-//  //! TODO(yuyang18): Temporary disable this unit-test because implementation
-//  //! error.
-//  return 0;
-//}
\ No newline at end of file
diff --git a/paddle/operators/softmax_op.cc b/paddle/operators/softmax_op.cc
index 5b59fad7d5f9729b0862f8cd78cb32f94f87f513..5cbb96ab754467ea6ddab9380ca25987c9376980 100644
--- a/paddle/operators/softmax_op.cc
+++ b/paddle/operators/softmax_op.cc
@@ -1,16 +1,17 @@
 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
 
-   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
+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
+    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. */
 
-   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 "paddle/operators/softmax_op.h"
 
 namespace paddle {
@@ -19,12 +20,13 @@ namespace operators {
 class SoftmaxOp : public OperatorWithKernel {
 protected:
   void InferShape(const InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.InputSize() == 1, "Only one input is need for softmax");
-    PADDLE_ENFORCE(ctx.Input<Tensor>(0)->dims().size() == 2,
+    PADDLE_ENFORCE(ctx.InputSize() == 1UL,
+                   "Only one input is need for softmax");
+    PADDLE_ENFORCE(ctx.Input<Tensor>("X")->dims().size() == 2UL,
                    "The input of softmax op must be matrix");
-    PADDLE_ENFORCE(ctx.OutputSize() == 1,
+    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
                    "Only one output is need for softmax");
-    ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
+    ctx.Output<Tensor>("Y")->Resize(ctx.Input<Tensor>("X")->dims());
   }
 };
 
@@ -40,10 +42,19 @@ public:
 
 class SoftmaxOpGrad : public OperatorWithKernel {
 protected:
-  void InferShape(const InferShapeContext &ctx) const override {}
-  std::string DebugString() const override {
-    LOG(INFO) << "SoftmaxOpGrad";
-    return "";
+  void InferShape(const InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE(ctx.InputSize() == 3UL,
+                   "Input of SoftmaxOpGrad should be 3, X, Y, YG");
+    PADDLE_ENFORCE(ctx.OutputSize() == 1UL,
+                   "Output of SoftmaxOpGrad should be 1");
+    PADDLE_ENFORCE(ctx.InputVar("Y") != nullptr, "Input(Y) should not be null");
+    PADDLE_ENFORCE(ctx.InputVar(GRAD_VAR_NAME("Y")) != nullptr,
+                   "Input(Y@GRAD) should not be null");
+    PADDLE_ENFORCE(ctx.Input<Tensor>("Y")->dims() ==
+                       ctx.Input<Tensor>(GRAD_VAR_NAME("Y"))->dims(),
+                   "the shape of Input(0) and Input(1) should be the same");
+    ctx.Output<Tensor>(GRAD_VAR_NAME("X"))
+        ->Resize(ctx.Input<Tensor>("Y")->dims());
   }
 };
 
@@ -51,5 +62,7 @@ protected:
 }  // namespace paddle
 
 REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker);
-REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad);
 REGISTER_OP_CPU_KERNEL(softmax, ops::SoftmaxKernel<ops::CPUPlace, float>);
+REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad);
+REGISTER_OP_CPU_KERNEL(softmax_grad,
+                       ops::SoftmaxGradKernel<ops::CPUPlace, float>);
diff --git a/paddle/operators/softmax_op.cu b/paddle/operators/softmax_op.cu
index ddf8f6e913ccf450185f377f531bf978f69ed1fc..8c652213f2e4c0e0ea1a31987fcb37c86374cd2a 100644
--- a/paddle/operators/softmax_op.cu
+++ b/paddle/operators/softmax_op.cu
@@ -3,3 +3,4 @@
 #include "paddle/operators/softmax_op.h"
 
 REGISTER_OP_GPU_KERNEL(softmax, ops::SoftmaxKernel<ops::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(softmax_grad, ops::SoftmaxGradKernel<ops::GPUPlace, float>);
diff --git a/paddle/operators/softmax_op.h b/paddle/operators/softmax_op.h
index 75c5197697dada58e09f4cda41cea13af56e79a3..13e74a79077982e9fba5d90f40986e699c1ed897 100644
--- a/paddle/operators/softmax_op.h
+++ b/paddle/operators/softmax_op.h
@@ -1,19 +1,22 @@
 /* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
 
-   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
+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
+    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. */
+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. */
 
 #pragma once
 
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/operator.h"
+#include "paddle/framework/tensor.h"
 #include "paddle/operators/type_alias.h"
 
 namespace paddle {
@@ -23,8 +26,8 @@ template <typename Place, typename T>
 class SoftmaxKernel : public OpKernel {
 public:
   void Compute(const ExecutionContext& context) const override {
-    auto input = context.Input<Tensor>(0);
-    auto output = context.Output<Tensor>(0);
+    auto input = context.Input<Tensor>("X");
+    auto output = context.Output<Tensor>("Y");
     output->mutable_data<T>(context.GetPlace());
 
     auto logits = EigenMatrix<T>::From(*input);
@@ -57,5 +60,38 @@ public:
              .broadcast(one_by_class));
   }
 };
+
+template <typename Place, typename T>
+class SoftmaxGradKernel : public OpKernel {
+public:
+  void Compute(const ExecutionContext& context) const override {
+    std::shared_ptr<Tensor> scale_ = std::make_shared<Tensor>();
+
+    auto Y = context.Input<Tensor>("Y");
+    auto dY = context.Input<Tensor>(OperatorBase::GRAD_VAR_NAME("Y"));
+    auto dX = context.Output<Tensor>(OperatorBase::GRAD_VAR_NAME("X"));
+    dX->mutable_data<T>(context.GetPlace());
+
+    const int batch_size = Y->dims()[0];
+    const int class_num = Y->dims()[1];
+
+    Eigen::DSizes<int, 1> along_class(1);
+    Eigen::DSizes<int, 2> batch_by_one(batch_size, 1);
+    Eigen::DSizes<int, 2> one_by_class(1, class_num);
+
+    auto Y_eigen = EigenMatrix<T>::From(*Y);
+    auto dY_eigen = EigenMatrix<T>::From(*dY);
+    auto dX_eigen = EigenMatrix<T>::From(*dX);
+    auto place = context.GetEigenDevice<Place>();
+
+    auto dot = (Y_eigen * dY_eigen)
+                   .sum(along_class)
+                   .eval()
+                   .reshape(batch_by_one)
+                   .broadcast(one_by_class);
+    dX_eigen.device(place) = (dY_eigen - dot) * Y_eigen;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
diff --git a/paddle/operators/type_alias.h b/paddle/operators/type_alias.h
index 93b62cddc819e0d1fd48323e474a294ff0d327e1..931740e150946a939b8656be5a30185c6ee1cb8f 100644
--- a/paddle/operators/type_alias.h
+++ b/paddle/operators/type_alias.h
@@ -15,13 +15,14 @@
 #pragma once
 
 #include "paddle/framework/eigen.h"
-#include "paddle/framework/net.h"
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace operators {
 
 using OpKernel = framework::OpKernel;
+using OperatorBase = framework::OperatorBase;
 using InferShapeContext = framework::InferShapeContext;
 using ExecutionContext = framework::ExecutionContext;
 using Variable = framework::Variable;
@@ -43,14 +44,16 @@ template <typename T,
           typename IndexType = Eigen::DenseIndex>
 using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
 using Tensor = framework::Tensor;
+using Scope = framework::Scope;
 using OperatorWithKernel = framework::OperatorWithKernel;
+using OperatorBase = framework::OperatorBase;
 using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker;
 using OpProto = framework::OpProto;
 using OpAttrChecker = framework::OpAttrChecker;
 using CPUPlace = platform::CPUPlace;
 using GPUPlace = platform::GPUPlace;
-using NetOp = framework::NetOp;
 using OpRegistry = framework::OpRegistry;
+
 }  // namespace operators
 }  // namespace paddle
 
diff --git a/paddle/pybind/CMakeLists.txt b/paddle/pybind/CMakeLists.txt
index ac12b504b5cbee778c7c0a74a84a7729f210e01e..29dd0ded0ac75893da7e244d92725cd5e285efce 100644
--- a/paddle/pybind/CMakeLists.txt
+++ b/paddle/pybind/CMakeLists.txt
@@ -6,4 +6,4 @@ cc_library(paddle_pybind SHARED
 	add_op
 	mean_op
 	cross_entropy_op
-	recurrent_network_op)
+	recurrent_op)
diff --git a/paddle/scripts/docker/build.sh b/paddle/scripts/docker/build.sh
index 69ae0ea2d72c199a8e17c0595693e5e0b2f79ee1..8de0e608c1f482e4553c07ff7ffd572d65a772aa 100644
--- a/paddle/scripts/docker/build.sh
+++ b/paddle/scripts/docker/build.sh
@@ -69,7 +69,7 @@ cat <<EOF
 Installing ...
 ========================================
 EOF
-make install
+make install -j `nproc`
 pip install /usr/local/opt/paddle/share/wheels/*.whl
 paddle version
 
@@ -122,7 +122,7 @@ cat <<EOF
 Generating .deb package ...
 ========================================
 EOF
-cpack -D CPACK_GENERATOR='DEB' ..
+cpack -D CPACK_GENERATOR='DEB' -j `nproc` ..
 
 
 cat <<EOF
diff --git a/paddle/scripts/travis/build_doc.sh b/paddle/scripts/travis/build_doc.sh
index a44385158042a23eca175df261852148642f7fa0..33fb5d84e2701c163b5d1b1bb3362ee81ebb34ea 100755
--- a/paddle/scripts/travis/build_doc.sh
+++ b/paddle/scripts/travis/build_doc.sh
@@ -6,14 +6,14 @@ mkdir -p $TRAVIS_BUILD_DIR/build
 cd $TRAVIS_BUILD_DIR/build
 
 # Compile paddle binaries first
-cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_DOC=OFF -DWITH_GOLANG=ON -DWITH_STYLE_CHECK=OFF
+cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_DOC=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_GOLANG=ON -DWITH_STYLE_CHECK=OFF
 
 mkdir output
 make -j `nproc`
 find .. -name '*whl' | xargs pip install  # install all wheels.
 rm -rf *
 # Compile Documentation only.
-cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_DOC=ON
+cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_MKLDNN=OFF -DWITH_MKLML=OFF -DWITH_DOC=ON
 make -j `nproc` paddle_docs paddle_docs_cn
 
 # check websites for broken links
diff --git a/python/paddle/v2/framework/tests/test_softmax_op.py b/python/paddle/v2/framework/tests/test_softmax_op.py
index 191b698c1cdec9b86b4ded6b1f743586867ca62f..c80888128781d98e4ed30d845a30b39121f66459 100644
--- a/python/paddle/v2/framework/tests/test_softmax_op.py
+++ b/python/paddle/v2/framework/tests/test_softmax_op.py
@@ -1,6 +1,10 @@
 import unittest
-from op_test_util import OpTestMeta
+
 import numpy as np
+import paddle.v2.framework.core as core
+import paddle.v2.framework.create_op_creation_methods as creation
+
+from op_test_util import OpTestMeta
 
 
 def stable_softmax(x):
@@ -19,5 +23,63 @@ class TestSoftmaxOp(unittest.TestCase):
         self.Y = np.apply_along_axis(stable_softmax, 1, self.X)
 
 
+class TestSoftmaxGradOp(unittest.TestCase):
+    def test_softmax_grad(self):
+        op = creation.op_creations.softmax(X="X", Y="Y")
+        backward_op = core.Operator.backward(op, set())
+        self.assertEqual(backward_op.type(), "softmax_grad")
+        expected = '''Op(softmax_grad), inputs:(X, Y, Y@GRAD), outputs:(X@GRAD).'''
+        self.assertEqual(expected, str(backward_op))
+
+        batch_size = 3
+        class_num = 5
+        # Initialize X and add 1e-2 for numerical stability
+        Y = np.random.rand(batch_size, class_num).astype(np.float32)
+        Y = Y + 1e-2
+        dY = np.random.rand(batch_size, class_num).astype(np.float32)
+
+        # Reference implementation of cross entropy with soft labels
+        def label_softmax_grad(Y, dY):
+            dX = Y * 0.0
+            for i in range(batch_size):
+                d = np.dot(Y[i, :], dY[i, :])
+                dX[i, :] = Y[i, :] * (dY[i, :] - d)
+            return dX
+
+        expected = label_softmax_grad(Y, dY)
+
+        scope = core.Scope()
+        places = []
+        places.append(core.CPUPlace())
+        if core.is_compile_gpu():
+            places.append(core.GPUPlace(0))
+
+        for place in places:
+            y = scope.new_var("Y")
+            y_tensor = y.get_tensor()
+            y_tensor.set_dims([batch_size, class_num])
+            y_tensor.alloc_float(place)
+            y_tensor.set(Y, place)
+
+            dy = scope.new_var("Y@GRAD")
+            dy_tensor = dy.get_tensor()
+            dy_tensor.set_dims([batch_size, class_num])
+            dy_tensor.alloc_float(place)
+            dy_tensor.set(dY, place)
+
+            x = scope.new_var("X")
+            dx = scope.new_var("X@GRAD")
+
+            tensor = scope.find_var("X@GRAD").get_tensor()
+            backward_op.infer_shape(scope)
+            self.assertEqual([batch_size, class_num], tensor.shape())
+
+            ctx = core.DeviceContext.create(place)
+            backward_op.run(scope, ctx)
+            actual = np.array(tensor)
+
+            np.testing.assert_almost_equal(actual, expected, decimal=3)
+
+
 if __name__ == '__main__':
     unittest.main()