diff --git a/.travis.yml b/.travis.yml
index 8c8c6699d3d9abddd65a3a224c2bceedc7d88348..b4b83fcdbc84ce0fb0c91c816ebc3c964acfa590 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -37,8 +37,8 @@ before_install:
   - if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
   # Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
   # protobuf version.
-  - pip install numpy wheel 'protobuf==3.1' sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit requests==2.9.2 LinkChecker
-  - pip install rarfile nltk==3.2.2 scipy==0.19.0 recordio matplotlib Pillow
+  - pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
+  - pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
   - curl https://glide.sh/get | bash
   - eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
   - go get -u github.com/alecthomas/gometalinter
diff --git a/doc/getstarted/build_and_install/docker_install_cn.rst b/doc/getstarted/build_and_install/docker_install_cn.rst
index 02b96bb413156786db6dc77696c5640b97c10aa4..84e33177740ca1652efc09c8081c2519b4366906 100644
--- a/doc/getstarted/build_and_install/docker_install_cn.rst
+++ b/doc/getstarted/build_and_install/docker_install_cn.rst
@@ -74,13 +74,13 @@ PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以
 
    .. code-block:: bash
 
-      docker run -it --rm paddlepaddle/paddle:0.10.0-dev /bin/bash
+      docker run -it --rm -v $(pwd):/paddle  paddlepaddle/paddle:0.10.0-dev /bin/bash
 
    或者，可以以后台进程方式运行容器：
 
    .. code-block:: bash
 
-      docker run -d -p 2202:22 -p 8888:8888 paddledev/paddle:0.10.0-dev
+      docker run -d -p 2202:22 -p 8888:8888 -v $(pwd):/paddle paddlepaddle/paddle:0.10.0-dev /usr/sbin/sshd -D
 
    然后用密码 :code:`root` SSH进入容器：
 
diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt
index 9142ecb02a741d74569566e9b040e974675c7873..03985260241689a099ae9ebc136bd04831a44167 100644
--- a/paddle/framework/CMakeLists.txt
+++ b/paddle/framework/CMakeLists.txt
@@ -7,7 +7,7 @@ cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context)
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
 cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
 
-cc_library(lod_tensor SRCS lod_tensor.cc details/lod_tensor.cc DEPS ddim place tensor)
+cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor)
 cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor)
 
 cc_test(variable_test SRCS variable_test.cc)
diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc
index d7cb1787065f72463f6cb05416612c7787788bb4..ebe52d5f284a8d271b666483001544a805d598ac 100644
--- a/paddle/framework/backward_test.cc
+++ b/paddle/framework/backward_test.cc
@@ -39,9 +39,9 @@ class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
  public:
   RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "Input X of Add").IgnoreGradient();
-    AddInput("b", "Bias of Add").IgnoreGradient();
-    AddOutput("Out", "Out of Add").IgnoreGradient();
+    AddInput("X", "Input X of Add").AsNoGradient();
+    AddInput("b", "Bias of Add").AsNoGradient();
+    AddOutput("Out", "Out of Add").AsNoGradient();
     AddComment("Add Op");
   }
 };
@@ -112,8 +112,8 @@ class FcOpMaker : public OpProtoAndCheckerMaker {
     AddInput("X", "x");
     AddInput("W", "w");
     AddInput("b", "b");
-    AddOutput("mul_result", "").SetTemporary();
-    AddOutput("add_result", "").SetTemporary();
+    AddOutput("mul_result", "").AsIntermediate();
+    AddOutput("add_result", "").AsIntermediate();
     AddOutput("Out", "");
     AddComment("");
   }
@@ -144,7 +144,7 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
  public:
   AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "x").SetMultiple();
+    AddInput("X", "x").AsDuplicable();
     AddOutput("Y", "y");
     AddComment("");
   }
@@ -393,18 +393,20 @@ TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
   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_["all"].size(),
+
+  const char *all = paddle::operators::NetOp::kAll;
+  EXPECT_EQ(grad_fc.inputs_[all].size(),
             2UL       /* external input number */
                 + 1UL /* external output number*/
                 + 1UL /* number of gradient of external output*/
                 + 2U /* internal variable number*/);
-  EXPECT_EQ(grad_fc.outputs_["all"].size(),
+  EXPECT_EQ(grad_fc.outputs_[all].size(),
             2UL       /* input number of mul*/
                 + 2UL /* input number of rowwise_add
                        */
                 + 1UL /* input number of sigmod */);
-  EXPECT_EQ(bwd_net->ops_[1]->inputs_["all"].size(), 0UL);
-  EXPECT_EQ(bwd_net->ops_[1]->outputs_["all"].size(), 0UL);
-  EXPECT_EQ(bwd_net->ops_[2]->inputs_["all"].size(), 0UL);
-  EXPECT_EQ(bwd_net->ops_[2]->outputs_["all"].size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[1]->inputs_[all].size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[1]->outputs_[all].size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[2]->inputs_[all].size(), 0UL);
+  EXPECT_EQ(bwd_net->ops_[2]->outputs_[all].size(), 0UL);
 }
diff --git a/paddle/framework/ddim.cc b/paddle/framework/ddim.cc
index 0b76a4fdb701ea7435d0da865e928b8da15335c9..cfd3e8dfdec0e92620aef5cd246b4622b779ce19 100644
--- a/paddle/framework/ddim.cc
+++ b/paddle/framework/ddim.cc
@@ -283,12 +283,5 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
 DDim::DDim(std::initializer_list<int> init_list) {
   *this = make_ddim(init_list);
 }
-
-std::string DDim::DebugString() const {
-  std::ostringstream ss;
-  ss << *this;
-  return ss.str();
-}
-
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/ddim.h b/paddle/framework/ddim.h
index 1627bcb269258bf8e3d9db86c716ea8c4b7406eb..95f294b62737be5c3eac39303148ac35da29fe7d 100644
--- a/paddle/framework/ddim.h
+++ b/paddle/framework/ddim.h
@@ -72,8 +72,6 @@ struct DDim {
   DDim operator*(DDim d) const;
 
   ssize_t size() const;
-
-  std::string DebugString() const;
 };
 
 /**
diff --git a/paddle/framework/details/lod_tensor.cc b/paddle/framework/details/lod_tensor.cc
deleted file mode 100644
index 9ad3979e5b511517f75d2d43004f97ee1576953b..0000000000000000000000000000000000000000
--- a/paddle/framework/details/lod_tensor.cc
+++ /dev/null
@@ -1,62 +0,0 @@
-/* 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
-
-   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. */
-
-#include "paddle/framework/lod_tensor.h"
-
-#include <memory>
-
-namespace paddle {
-namespace framework {
-namespace details {
-
-using LOD = LODTensor::LOD;
-
-std::shared_ptr<LOD> SliceLOD(const LOD &lod, size_t level_begin,
-                              size_t level_end) {
-  auto new_lod = std::make_shared<LOD>();
-  new_lod->reserve(level_end - level_begin);
-  for (size_t i = level_begin; i < level_end; i++) {
-    new_lod->emplace_back(lod[i]);
-  }
-  return new_lod;
-}
-
-std::shared_ptr<LOD> SliceLOD(const LOD &lod, size_t level, size_t elem_begin,
-                              size_t elem_end, bool tensor_shared) {
-  // slice the lod.
-  auto new_lod = std::make_shared<LOD>();
-  new_lod->reserve(lod.size() - level);
-  auto start = lod.at(level)[elem_begin];
-  auto end = lod.at(level)[elem_end];
-
-  for (auto it = lod.begin() + level; it != lod.end(); it++) {
-    auto it_begin = std::find(it->begin(), it->end(), start);
-    auto it_end = std::find(it_begin, it->end(), end);
-    PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
-    PADDLE_ENFORCE(it_end != it->end(), "error in parsing lod info");
-    new_lod->emplace_back(it_begin, it_end + 1);
-    if (!tensor_shared) {
-      // reset offset if tensor is copyed and sliced.
-      std::transform(new_lod->back().begin(), new_lod->back().end(),
-                     new_lod->back().begin(),
-                     [start](int v) { return v - start; });
-      PADDLE_ENFORCE(new_lod->back().front() == 0, "error in slice LOD");
-    }
-  }
-  return new_lod;
-}
-
-}  // namespace details
-}  // namespace framework
-}  // namespace paddle
diff --git a/paddle/framework/details/lod_tensor.h b/paddle/framework/details/lod_tensor.h
deleted file mode 100644
index 9a6a6cd2ea41f02db991bdc0a2b917433dafed99..0000000000000000000000000000000000000000
--- a/paddle/framework/details/lod_tensor.h
+++ /dev/null
@@ -1,46 +0,0 @@
-/* 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
-
-   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. */
-
-#pragma once
-
-#include <memory>
-
-namespace paddle {
-namespace framework {
-namespace details {
-
-/*
- * Slice levels from LOD.
- *
- * @lod: LOD to slice.
- * @level_begin: level to begin slice.
- * @level_end: level to end slice.
- */
-std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
-                                         size_t level_begin, size_t level_end);
-
-/*
- * Slice elements from a level of LOD.
- *
- * @lod: LOD to slice.
- * @level: which level to slice.
- * @elem_begin: element's index to begin slice.
- * @elem_end: element's index to end slice.
- */
-std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
-                                         size_t level, size_t elem_begin,
-                                         size_t elem_end, bool tensor_shared);
-}  // namespace details
-}  // namespace framework
-}  // namespace paddle
diff --git a/paddle/framework/grad_op_builder.cc b/paddle/framework/grad_op_builder.cc
index c2855d3a589a8a0aea43e59acd671925b5506c8e..21bc30d1fbdae31548547bccf39e78fe16eedfaa 100644
--- a/paddle/framework/grad_op_builder.cc
+++ b/paddle/framework/grad_op_builder.cc
@@ -18,9 +18,6 @@ permissions and limitations under the License. */
 
 namespace paddle {
 namespace framework {
-
-class OpRegistry;
-
 enum class OpArgType { IN, OUT };
 
 static void TransOpArg(const OperatorBase* src_op,
diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc
index a351e86c5d10e8471285cf29e269b67d2d51dcaa..ebaf84545fce0d281d8821861264cddc8854893d 100644
--- a/paddle/framework/grad_op_builder_test.cc
+++ b/paddle/framework/grad_op_builder_test.cc
@@ -21,10 +21,10 @@ class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
   MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("In1", "a single input");
-    AddInput("In2_mult", "a multiple input").SetMultiple();
+    AddInput("In2_mult", "a multiple input").AsDuplicable();
     AddInput("In3", "another single input");
     AddOutput("Out1", "a single output");
-    AddOutput("Out2_mult", "a multiple output").SetMultiple();
+    AddOutput("Out2_mult", "a multiple output").AsDuplicable();
     AddComment("test op with multiple inputs and outputs");
   }
 };
@@ -34,10 +34,10 @@ class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
   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();
+    AddInput("In2_mult", "a multiple input").AsDuplicable().AsNoGradient();
+    AddInput("In3_mult", "another multiple input").AsDuplicable();
+    AddOutput("Out1_mult", "a multiple output").AsDuplicable();
+    AddOutput("Out2", "a single output").AsNoGradient();
     AddComment("op with inputs and outputs ignored in gradient calculating");
   }
 };
diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc
index 70045dbf7afd0935e4df852b2f0e3ecd163a9316..2b178907747b3911292b070b65160a24c120b726 100644
--- a/paddle/framework/lod_tensor.cc
+++ b/paddle/framework/lod_tensor.cc
@@ -19,32 +19,59 @@
 namespace paddle {
 namespace framework {
 
-LODTensor LODTensor::SliceShared(size_t level_begin, size_t level_end) const {
-  PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
-  auto new_lod = details::SliceLOD(*lod_start_pos_, level_begin, level_end);
-  // slice levels just need to update LOD info, each level will contains the
-  // whole tensor_, so no need to modify tensor_.
-  return LODTensor(tensor_, new_lod);
+LODTensor::LOD LODTensor::LOD::SliceLevels(size_t level_begin,
+                                           size_t level_end) const {
+  LOD new_lod;
+  new_lod.reserve(level_end - level_begin);
+  for (size_t i = level_begin; i < level_end; i++) {
+    new_lod.emplace_back(at(i));
+  }
+  return new_lod;
 }
 
-LODTensor LODTensor::SliceShared(size_t level, size_t elem_begin,
-                                 size_t elem_end) const {
-  PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
-  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
-                 NumLevels());
-  PADDLE_ENFORCE(elem_begin < NumElements(level),
-                 "element begin [%d] out of range [%d]", elem_begin,
-                 NumElements(level));
-  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
-                 "element end [%d] out of range [%d]", elem_end,
-                 NumElements(level));
-
-  auto new_lod = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end,
-                                   true /*tensor_shared*/);
-
-  // slice elements just need to update LOD info, because offsets are not
-  // changed, so the original tensor_ can be reused.
-  return LODTensor(tensor_, new_lod);
+LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin,
+                                            size_t elem_end) const {
+  // slice the lod.
+  LOD new_lod;
+  new_lod.reserve(size() - level);
+  auto start = this->at(level)[elem_begin];
+  auto end = this->at(level)[elem_end];
+
+  for (auto it = this->begin() + level; it != this->end(); it++) {
+    auto it_begin = std::find(it->begin(), it->end(), start);
+    auto it_end = std::find(it_begin, it->end(), end);
+    PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
+    PADDLE_ENFORCE(it_end != it->end(), "error in parsing lod info");
+    new_lod.emplace_back(it_begin, it_end + 1);
+    // reset offset if tensor is copyed and sliced.
+    std::transform(new_lod.back().begin(), new_lod.back().end(),
+                   new_lod.back().begin(),
+                   [start](int v) { return v - start; });
+    PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LOD");
+  }
+  PADDLE_ENFORCE_LE(new_lod.size(), this->size());
+  return new_lod;
+}
+
+bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) {
+  if (a.size() != b.size()) {
+    return false;
+  }
+
+  for (size_t i = 0; i < a.size(); i++) {
+    const auto& a_level = a[i];
+    const auto& b_level = b[i];
+    if (a_level.size() != b_level.size()) {
+      return false;
+    }
+    for (size_t j = 0; j < a_level.size(); j++) {
+      if (a_level[j] != b_level[j]) {
+        return false;
+      }
+    }
+  }
+
+  return true;
 }
 
 }  // namespace framework
diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h
index 4933479b109694312e99595dc8ad6db70259efa6..9e27aec38d336db8a4f0adbed098d299aa741356 100644
--- a/paddle/framework/lod_tensor.h
+++ b/paddle/framework/lod_tensor.h
@@ -15,7 +15,7 @@
 #pragma once
 
 #include <memory>
-#if (!PADDLE_ONLY_CPU)
+#if !defined(PADDLE_ONLY_CPU)
 #include <thrust/device_vector.h>
 #include <thrust/host_vector.h>
 #endif
@@ -31,30 +31,29 @@ namespace framework {
  * LODTensor (Level of details Tensor)
  * see https://en.wikipedia.org/wiki/Level_of_details for reference.
  */
-class LODTensor {
+class LODTensor : public Tensor {
  public:
 // Level save offsets of each unit.
 #ifdef PADDLE_ONLY_CPU
-  using Level = std::vector<size_t>;
+  template <typename T>
+  using Vector = std::vector<T>;
 #else
-  using Level = thrust::device_vector<size_t>;
+  template <typename T>
+  using Vector = thrust::host_vector<T>;
 #endif
-  // LOD stores offsets of each level of units, the largest units level first,
+  // LoD stores offsets of each level of units, the largest units level first,
   // then the smaller units level. Each Level stores the offsets of units in
   // Tesor.
-  typedef std::vector<Level> LOD;
+  class LOD : public std::vector<Vector<size_t>> {
+   public:
+    LOD SliceLevels(size_t level_begin, size_t level_end) const;
+    LOD SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) const;
+  };
 
   LODTensor() {}
-  LODTensor(const std::shared_ptr<Tensor> &tensor,
-            const std::shared_ptr<LOD> &lod) {
-    Reset(tensor, lod);
-  }
+  explicit LODTensor(const LOD &lod) : lod_(lod) {}
 
-  void Reset(const std::shared_ptr<Tensor> &tensor,
-             const std::shared_ptr<LOD> &lod) {
-    tensor_ = tensor;
-    lod_start_pos_ = lod;
-  }
+  virtual Tensor *Clone() const { return new LODTensor(lod_); }
 
   /*
    * Get a element from LOD.
@@ -65,16 +64,14 @@ class LODTensor {
     PADDLE_ENFORCE(elem < NumElements(level),
                    "element begin [%d] out of range [%d]", elem,
                    NumElements(level));
-    return (*lod_start_pos_)[level][elem];
+    return (lod_)[level][elem];
   }
 
   /*
    * Number of LODTensor's levels, each level has units of data, for example,
    * in the sentence's view, article, paragraph, sentence are 3 levels.
    */
-  size_t NumLevels() const {
-    return lod_start_pos_ ? lod_start_pos_->size() : 0UL;
-  }
+  size_t NumLevels() const { return lod_.size(); }
   /*
    * Number of elements in a level.
    */
@@ -82,64 +79,71 @@ class LODTensor {
     PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
                    NumLevels());
     // the last offset is the end of last element
-    return lod_start_pos_->at(level).size() - 1;
+    return lod_[level].size() - 1;
   }
 
-  /*
-   * Slice of levels[level_begin:level_end], with tensor copied.
-   */
-  template <typename T>
-  LODTensor SliceCopied(size_t level_begin, size_t level_end,
-                        const platform::Place &dst_place) const;
-
   /*
    * Slice of levels[level_begin:level_end], with tensor shared.
    */
-  LODTensor SliceShared(size_t level_begin, size_t level_end) const;
-
-  /*
-   * Slice of elements of a level, [elem_begin: elem_end], with tensor copied.
-   * @note: low performance in slice lod_start_pos_.
-   */
   template <typename T>
-  LODTensor SliceCopied(size_t level, size_t elem_begin, size_t elem_end,
-                        const platform::Place &dst_place) const;
+  LODTensor SliceLevels(size_t level_begin, size_t level_end) const;
 
   /*
    * Slice of elements of a level, [elem_begin: elem_end], with tensor shared.
-   * @note: low performance in slice lod_start_pos_.
-   */
-  LODTensor SliceShared(size_t level, size_t elem_begin, size_t elem_end) const;
-
-  /*
-   * Copy other's lod_start_pos_, to share LOD info.
-   * @note: the LOD info should not be changed.
+   * @note: low performance in slice lod_.
    */
-  void ShareLOD(const LODTensor &other) {
-    lod_start_pos_ = other.lod_start_pos_;
-  }
+  template <typename T>
+  LODTensor SliceInLevel(size_t level, size_t elem_begin,
+                         size_t elem_end) const;
 
   /*
-   * Copy other's lod_start_pos_'s content, free to mutate.
+   * Copy other's lod_'s content, free to mutate.
    */
-  void CopyLOD(const LODTensor &other) {
-    lod_start_pos_ = std::make_shared<LOD>(*other.lod_start_pos_);
-  }
+  void CopyLOD(const LODTensor &other) { lod_ = other.lod_; }
   /*
    * Determine whether LODTensor has a valid LOD info.
    */
-  bool HasLOD() const { return bool(lod_start_pos_); }
-  LOD *lod() const { return lod_start_pos_.get(); }
+  const LOD &lod() const { return lod_; }
+  LOD *mutable_lod() { return &lod_; }
 
-  std::shared_ptr<Tensor> &tensor() { return tensor_; }
-  Tensor *raw_tensor() { return tensor_.get(); }
+  virtual ~LODTensor() {}
 
  private:
-  std::shared_ptr<LOD> lod_start_pos_;
-  std::shared_ptr<Tensor> tensor_;
+  LOD lod_;
 };
 
+bool operator==(const LODTensor::LOD &a, const LODTensor::LOD &b);
+
+template <typename T>
+LODTensor LODTensor::SliceLevels(size_t level_begin, size_t level_end) const {
+  auto new_lod = lod_.SliceLevels(level_begin, level_end);
+  // slice levels just need to update LOD info, each level will contains the
+  // whole tensor_, so no need to modify tensor_.
+  LODTensor new_tensor(new_lod);
+  new_tensor.ShareDataWith<T>(*this);
+  return new_tensor;
+}
+
+template <typename T>
+LODTensor LODTensor::SliceInLevel(size_t level, size_t elem_begin,
+                                  size_t elem_end) const {
+  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
+                 NumLevels());
+  PADDLE_ENFORCE(elem_begin < NumElements(level),
+                 "element begin [%d] out of range [%d]", elem_begin,
+                 NumElements(level));
+  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
+                 "element end [%d] out of range [%d]", elem_end,
+                 NumElements(level));
+
+  auto new_lod = lod_.SliceInLevel(level, elem_begin, elem_end);
+
+  // slice elements just need to update LOD info, because offsets are not
+  // changed, so the original tensor_ can be reused.
+  LODTensor new_tensor(new_lod);
+  new_tensor.ShareDataWith<T>(*this);
+  return new_tensor;
+}
+
 }  // namespace framework
 }  // namespace paddle
-
-#include "paddle/framework/lod_tensor_impl.h"
diff --git a/paddle/framework/lod_tensor_impl.h b/paddle/framework/lod_tensor_impl.h
deleted file mode 100644
index 0eb6469aea3ae25f035751da985b5bebb489d961..0000000000000000000000000000000000000000
--- a/paddle/framework/lod_tensor_impl.h
+++ /dev/null
@@ -1,60 +0,0 @@
-/* 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
-
-   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. */
-
-#pragma once
-
-#include "paddle/framework/details/lod_tensor.h"
-
-namespace paddle {
-namespace framework {
-
-template <typename T>
-LODTensor LODTensor::SliceCopied(size_t level_begin, size_t level_end,
-                                 const platform::Place &dst_place) const {
-  PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
-  auto new_lod = details::SliceLOD(*lod_start_pos_, level_begin, level_end);
-  auto new_tensor = std::make_shared<Tensor>();
-  new_tensor->CopyFrom<T>(*tensor_, dst_place);
-
-  return LODTensor(new_tensor, new_lod);
-}
-
-template <typename T>
-LODTensor LODTensor::SliceCopied(size_t level, size_t elem_begin,
-                                 size_t elem_end,
-                                 const platform::Place &dst_place) const {
-  PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
-  PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
-                 NumLevels());
-  PADDLE_ENFORCE(elem_begin < NumElements(level),
-                 "element begin [%d] out of range [%d]", elem_begin,
-                 NumElements(level));
-  PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
-                 "element end [%d] out of range [%d]", elem_end,
-                 NumElements(level));
-
-  auto new_lod = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end,
-                                   false /*tensor_shared*/);
-
-  auto start_idx = new_lod->front().front();
-  auto end_idx = new_lod->front().back() - 1 /*the next element's start*/;
-  auto sliced_tensor = tensor_->Slice<T>(start_idx, end_idx);
-  auto new_tensor = std::make_shared<Tensor>();
-  new_tensor->CopyFrom<T>(sliced_tensor, dst_place);
-
-  return LODTensor(new_tensor, new_lod);
-}
-
-}  // namespace framework
-}  // namespace paddle
diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc
index 511716375e81e8fd89b071c940ee97327c268b8b..2881136ced6ef957a192e303e529b9b2867b3dda 100644
--- a/paddle/framework/lod_tensor_test.cc
+++ b/paddle/framework/lod_tensor_test.cc
@@ -15,6 +15,7 @@
 
 #include <glog/logging.h>
 #include <gtest/gtest.h>
+#include <algorithm>
 #include <memory>
 
 namespace paddle {
@@ -29,22 +30,28 @@ class LODTensorTester : public ::testing::Test {
     // 0 10 20
     // 0 5 10 15 20
     // 0 2 5 7 10 12 15 20
-    auto lod = std::make_shared<LODTensor::LOD>();
-    lod->push_back(std::vector<size_t>{0, 10, 20});
-    lod->push_back(std::vector<size_t>{0, 5, 10, 15, 20});
-    lod->push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
+    LODTensor::LOD lod;
+    lod.push_back(std::vector<size_t>{0, 10, 20});
+    lod.push_back(std::vector<size_t>{0, 5, 10, 15, 20});
+    lod.push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
 
-    auto tensor = std::make_shared<Tensor>();
-    tensor->Resize({20 /*batch size*/, 128 /*dim*/});
+    ASSERT_EQ(lod.size(), 3UL);
+
+    tensor.Resize({20 /*batch size*/, 128 /*dim*/});
     // malloc memory
-    tensor->mutable_data<float>(place);
+    tensor.mutable_data<float>(place);
+
+    lod_tensor.reset(new LODTensor(lod));
+    lod_tensor->Resize({20 /*batch size*/, 128 /*dim*/});
 
-    lod_tensor->Reset(tensor, lod);
+    lod_tensor->ShareDataWith<float>(tensor);
+    // lod_tensor->ShareDataWith<Tensor>(tensor);
   }
 
  protected:
   std::unique_ptr<LODTensor> lod_tensor;
   platform::CPUPlace place;
+  Tensor tensor;
 };
 
 TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor->NumLevels(), 3UL); }
@@ -55,110 +62,54 @@ TEST_F(LODTensorTester, NumElements) {
   ASSERT_EQ(lod_tensor->NumElements(2), 8UL);
 }
 
-TEST_F(LODTensorTester, SliceShared_Level) {
-  // slice 1 level
-  for (size_t level = 0; level < 3UL; ++level) {
-    auto new_lod_tensor = lod_tensor->SliceShared(level, level + 1);
-    ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level));
-    ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
-  }
-  // slice 2 level
-  for (size_t level = 0; level < 2UL; ++level) {
-    auto new_lod_tensor = lod_tensor->SliceShared(level, level + 2);
-    ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-    ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor->NumElements(level));
-    ASSERT_EQ(new_lod_tensor.NumElements(1),
-              lod_tensor->NumElements(level + 1));
-    ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
-  }
-}
-
-TEST_F(LODTensorTester, SliceCopied_Level) {
+TEST_F(LODTensorTester, SliceLevels) {
   // slice 1 level
   for (size_t level = 0; level < 3UL; ++level) {
-    auto new_lod_tensor =
-        lod_tensor->SliceCopied<float>(level, level + 1, place);
+    auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 1);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
     ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level));
-    // ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
-    // TODO(superjom) add tensor comparation here.
+    // ASSERT_EQ(new_lod_tensor, *lod_tensor);
   }
   // slice 2 level
   for (size_t level = 0; level < 2UL; ++level) {
-    auto new_lod_tensor =
-        lod_tensor->SliceCopied<float>(level, level + 2, place);
+    auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 2);
     ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
     ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor->NumElements(level));
     ASSERT_EQ(new_lod_tensor.NumElements(1),
               lod_tensor->NumElements(level + 1));
-    // ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
-    // TODO(superjom) add tensor comparation here.
+    ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
   }
 }
 
-TEST_F(LODTensorTester, SliceShared_Element) {
-  size_t level = 0;
-  auto new_lod_tensor = lod_tensor->SliceShared(level, 0, 2);
-  ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL);
-  ASSERT_EQ(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
-
-  level = 1;
-  new_lod_tensor = lod_tensor->SliceShared(level, 0, 2);
-  ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
-}
-
-TEST_F(LODTensorTester, SliceCopied_Element) {
+TEST_F(LODTensorTester, SliceInLevel) {
   size_t level = 0;
-  auto new_lod_tensor = lod_tensor->SliceCopied<float>(level, 0, 2, place);
-  ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL);
-  ASSERT_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
+  auto new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
+  EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
+  EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
+  EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
+  EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
 
   level = 1;
-  new_lod_tensor = lod_tensor->SliceCopied<float>(level, 0, 2, place);
+  new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
   ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
   ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-  ASSERT_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
-
-  level = 1;
-  // LOD is
-  //    0 5 10
-  //    0 2 5 7 10
-  new_lod_tensor = lod_tensor->SliceCopied<float>(level, 1, 3, place);
-  ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
-  ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
-
-  ASSERT_EQ(new_lod_tensor.lod_element(0, 0), 0UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(0, 1), 5UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(1, 0), 0UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(1, 1), 2UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(1, 2), 5UL);
-  ASSERT_EQ(new_lod_tensor.lod_element(1, 3), 7UL);
-
-  // TODO(superjom) compare the content of these tensors
+  ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
 }
 
 TEST_F(LODTensorTester, ShareLOD) {
   LODTensor new_lod_tensor;
-  new_lod_tensor.ShareLOD(*lod_tensor);
+  new_lod_tensor.CopyLOD(*lod_tensor);
   ASSERT_EQ(new_lod_tensor.lod(), lod_tensor->lod());
 }
 
 TEST_F(LODTensorTester, CopyLOD) {
   LODTensor new_lod_tensor;
   new_lod_tensor.CopyLOD(*lod_tensor);
-  ASSERT_NE(new_lod_tensor.lod(), lod_tensor->lod());
+  bool equals = std::equal(lod_tensor->lod().begin(), lod_tensor->lod().end(),
+                           new_lod_tensor.lod().begin());
+  ASSERT_TRUE(equals);
 }
 
 }  // namespace framework
diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
index 0fbda936c640c91385e289d6728fd2d5cdf23363..3b793628aa6fdb08544ba90274736c9d29262a8b 100644
--- a/paddle/framework/op_registry.h
+++ b/paddle/framework/op_registry.h
@@ -47,17 +47,20 @@ class OpProtoAndCheckerMaker {
   struct VariableBuilder {
     OpProto::Var* var_;
 
-    VariableBuilder& SetMultiple() {
+    VariableBuilder& AsDuplicable() {
       var_->set_duplicable(true);
       return *this;
     }
 
-    VariableBuilder& SetTemporary() {
+    VariableBuilder& AsIntermediate() {
       var_->set_intermediate(true);
       return *this;
     }
 
-    VariableBuilder& IgnoreGradient() {
+    // TODO(FengJiayi, yuyang18): `AsNoGradient` is a very bad name, because it
+    // means that input/output is not needed when calculate gradient. It does
+    // not mean no gradient when backward. It should be changed soon.
+    VariableBuilder& AsNoGradient() {
       var_->set_no_gradient(true);
       return *this;
     }
@@ -166,25 +169,22 @@ class OpRegistry {
     return std::shared_ptr<OperatorBase>(op);
   }
 
-  static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc) {
-    VarNameMap inputs;
-    for (auto& input : op_desc.inputs()) {
-      auto& var_names = inputs[input.parameter()];
-      auto& var_names_in_proto = input.arguments();
-      var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
-      std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
-                std::back_inserter(var_names));
-    }
-
-    VarNameMap outputs;
-    for (auto& output : op_desc.outputs()) {
-      auto& var_names = outputs[output.parameter()];
-      auto& var_names_in_proto = output.arguments();
+  static VarNameMap ConvertOpDescVarsToVarNameMap(
+      const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
+    VarNameMap ret_val;
+    for (auto& var : op_desc_vars) {
+      auto& var_names = ret_val[var.parameter()];
+      auto& var_names_in_proto = var.arguments();
       var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
       std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
                 std::back_inserter(var_names));
     }
+    return ret_val;
+  }
 
+  static std::shared_ptr<OperatorBase> CreateOp(const OpDesc& op_desc) {
+    VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
+    VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
     AttributeMap attrs;
     for (auto& attr : op_desc.attrs()) {
       attrs[attr.name()] = GetAttrValue(attr);
diff --git a/paddle/framework/op_registry_test.cc b/paddle/framework/op_registry_test.cc
index 42361c718bc1a5ed29d31f175216697643f08ef7..0b8f8289490135b8976c38fa3fb3c2995c50416f 100644
--- a/paddle/framework/op_registry_test.cc
+++ b/paddle/framework/op_registry_test.cc
@@ -38,8 +38,8 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
  public:
   MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("input", "input of cosine op").SetMultiple();
-    AddOutput("output", "output of cosine op").SetTemporary();
+    AddInput("input", "input of cosine op").AsDuplicable();
+    AddOutput("output", "output of cosine op").AsIntermediate();
     auto my_checker = [](int i) {
       PADDLE_ENFORCE(i % 2 == 0, "'test_attr' must be even!");
     };
@@ -51,6 +51,15 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 }  // namespace framework
 }  // namespace paddle
 
+static void BuildVar(const std::string& param_name,
+                     std::initializer_list<const char*> arguments,
+                     paddle::framework::OpDesc::Var* var) {
+  var->set_parameter(param_name);
+  for (auto& arg_name : arguments) {
+    var->add_arguments(arg_name);
+  }
+}
+
 REGISTER_OP(cos_sim, paddle::framework::CosineOp,
             paddle::framework::CosineOpProtoAndCheckerMaker);
 REGISTER_OP(my_test_op, paddle::framework::MyTestOp,
@@ -59,13 +68,8 @@ REGISTER_OP(my_test_op, paddle::framework::MyTestOp,
 TEST(OpRegistry, CreateOp) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("cos_sim");
-  auto input = op_desc.add_inputs();
-  input->set_parameter("input");
-  *input->mutable_arguments()->Add() = "aa";
-
-  auto output = op_desc.add_outputs();
-  output->set_parameter("output");
-  *output->mutable_arguments()->Add() = "bb";
+  BuildVar("input", {"aa"}, op_desc.add_inputs());
+  BuildVar("output", {"bb"}, op_desc.add_outputs());
 
   float scale = 3.3;
   auto attr = op_desc.mutable_attrs()->Add();
@@ -85,13 +89,8 @@ TEST(OpRegistry, CreateOp) {
 TEST(OpRegistry, IllegalAttr) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("cos_sim");
-  auto input = op_desc.add_inputs();
-  input->set_parameter("input");
-  *input->mutable_arguments()->Add() = "aa";
-
-  auto output = op_desc.add_outputs();
-  output->set_parameter("output");
-  *output->mutable_arguments()->Add() = "bb";
+  BuildVar("input", {"aa"}, op_desc.add_inputs());
+  BuildVar("output", {"bb"}, op_desc.add_outputs());
 
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
@@ -115,13 +114,8 @@ TEST(OpRegistry, IllegalAttr) {
 TEST(OpRegistry, DefaultValue) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("cos_sim");
-  auto input = op_desc.add_inputs();
-  input->set_parameter("input");
-  *input->mutable_arguments()->Add() = "aa";
-
-  auto output = op_desc.add_outputs();
-  output->set_parameter("output");
-  *output->mutable_arguments()->Add() = "bb";
+  BuildVar("input", {"aa"}, op_desc.add_inputs());
+  BuildVar("output", {"bb"}, op_desc.add_outputs());
 
   ASSERT_TRUE(op_desc.IsInitialized());
 
@@ -136,13 +130,8 @@ TEST(OpRegistry, DefaultValue) {
 TEST(OpRegistry, CustomChecker) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("my_test_op");
-  auto input = op_desc.add_inputs();
-  input->set_parameter("input");
-  *input->mutable_arguments()->Add() = "ii";
-
-  auto output = op_desc.add_outputs();
-  output->set_parameter("output");
-  *output->mutable_arguments()->Add() = "oo";
+  BuildVar("input", {"ii"}, op_desc.add_inputs());
+  BuildVar("output", {"oo"}, op_desc.add_outputs());
 
   // attr 'test_attr' is not set
   bool caught = false;
diff --git a/paddle/framework/operator.cc b/paddle/framework/operator.cc
index 59593cb6bd7027364cf154efa6eff3a8af9815b2..13442a72b9d77a4858b5d91dd7690e089ec7ed49 100644
--- a/paddle/framework/operator.cc
+++ b/paddle/framework/operator.cc
@@ -42,33 +42,35 @@ std::unordered_map<std::string, OpProto>& OpProtos() {
 }
 
 const std::string& OperatorBase::Input(const std::string& name) const {
-  auto it = inputs_.find(name);
-  PADDLE_ENFORCE(it != inputs_.end(), "Op %s does not have input %s", type_,
-                 name);
-  PADDLE_ENFORCE_EQ(it->second.size(), 1UL,
+  auto& ins = Inputs(name);
+  PADDLE_ENFORCE_EQ(ins.size(), 1UL,
                     "Op %s input %s should contain only one variable", type_,
                     name);
-  return it->second[0];
+  return ins[0];
 }
 
 const std::vector<std::string>& OperatorBase::Inputs(
     const std::string& name) const {
-  return inputs_.at(name);
+  auto it = inputs_.find(name);
+  PADDLE_ENFORCE(it != inputs_.end(), "Op %s do not have input %s", type_,
+                 name);
+  return it->second;
 }
 
 const std::string& OperatorBase::Output(const std::string& name) const {
-  auto it = outputs_.find(name);
-  PADDLE_ENFORCE(it != outputs_.end(), "Op %s does not have output %s", type_,
-                 name);
-  PADDLE_ENFORCE_EQ(it->second.size(), 1UL,
-                    "Op %s input %s should contain only one variable", type_,
+  auto& outs = Outputs(name);
+  PADDLE_ENFORCE_EQ(outs.size(), 1UL,
+                    "Op %s output %s should contain only one variable", type_,
                     name);
-  return it->second[0];
+  return outs[0];
 }
 
 const std::vector<std::string>& OperatorBase::Outputs(
     const std::string& name) const {
-  return outputs_.at(name);
+  auto it = outputs_.find(name);
+  PADDLE_ENFORCE(it != outputs_.end(), "Op %s does not have output %s", type_,
+                 name);
+  return it->second;
 }
 
 std::string OperatorBase::DebugString() const {
@@ -136,5 +138,35 @@ OperatorBase::OperatorBase(const std::string& type,
     }
   }
 }
+
+std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
+  std::vector<std::string> ret_val;
+  if (has_intermediate) {
+    // push all outputs into ret_val
+    for (auto& o : outputs_) {
+      ret_val.reserve(ret_val.size() + o.second.size());
+      ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
+    }
+    return ret_val;
+  }
+  auto it = OpProtos().find(type_);
+  PADDLE_ENFORCE(
+      it != OpProtos().end(),
+      "Operator %s not registered, cannot figure out intermediate outputs",
+      type_);
+
+  // get all OpProto::Var for outputs
+  for (auto& o : it->second.outputs()) {
+    // ignore all intermediate output
+    if (o.intermediate()) continue;
+    auto out = outputs_.find(o.name());
+    if (out != outputs_.end()) {
+      ret_val.reserve(ret_val.size() + out->second.size());
+      ret_val.insert(ret_val.end(), out->second.begin(), out->second.end());
+    }
+  }
+  return ret_val;
+}
+
 }  // namespace framework
 }  // namespace paddle
diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h
index 292847f1f0f5343ece3381522d7e52e0fece097c..161b6c5c50467ef54848f6bf848dfff61f4800b6 100644
--- a/paddle/framework/operator.h
+++ b/paddle/framework/operator.h
@@ -110,34 +110,7 @@ class OperatorBase {
   //! TODO add a vector_view to prevent memory copy.
   const std::vector<std::string>& Outputs(const std::string& name) const;
 
-  virtual std::vector<std::string> OutputVars(bool has_intermediate) const {
-    std::vector<std::string> ret_val;
-    if (has_intermediate) {
-      // push all outputs into ret_val
-      for (auto& o : outputs_) {
-        ret_val.reserve(ret_val.size() + o.second.size());
-        ret_val.insert(ret_val.end(), o.second.begin(), o.second.end());
-      }
-      return ret_val;
-    }
-    auto it = OpProtos().find(type_);
-    PADDLE_ENFORCE(
-        it != OpProtos().end(),
-        "Operator %s not registered, cannot figure out intermediate outputs",
-        type_);
-
-    // get all OpProto::Var for outputs
-    for (auto& o : it->second.outputs()) {
-      // ignore all intermediate output
-      if (o.intermediate()) continue;
-      auto out = outputs_.find(o.name());
-      if (out != outputs_.end()) {
-        ret_val.reserve(ret_val.size() + out->second.size());
-        ret_val.insert(ret_val.end(), out->second.begin(), out->second.end());
-      }
-    }
-    return ret_val;
-  }
+  virtual std::vector<std::string> OutputVars(bool has_intermediate) const;
 
   std::string Type() const { return type_; }
   const AttributeMap& Attrs() const { return attrs_; }
@@ -162,11 +135,11 @@ class InferShapeContext {
       : op_(op), scope_(scope) {}
 
   size_t InputSize(const std::string& name) const {
-    return op_.inputs_.at(name).size();
+    return op_.Inputs(name).size();
   }
 
   size_t OutputSize(const std::string& name) const {
-    return op_.outputs_.at(name).size();
+    return op_.Outputs(name).size();
   }
 
   const Variable* InputVar(const std::string& name) const {
diff --git a/paddle/framework/operator_test.cc b/paddle/framework/operator_test.cc
index 6a6ee10f213b22f7a7914b678fd09a4e509e6148..6804841587730d51d9cfad30a9de81401d36695b 100644
--- a/paddle/framework/operator_test.cc
+++ b/paddle/framework/operator_test.cc
@@ -56,19 +56,24 @@ class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
 }  // namespace framework
 }  // namespace paddle
 
+static void BuildVar(const std::string& param_name,
+                     std::initializer_list<const char*> arguments,
+                     paddle::framework::OpDesc::Var* var) {
+  var->set_parameter(param_name);
+  for (auto& arg_name : arguments) {
+    *var->mutable_arguments()->Add() = arg_name;
+  }
+}
+
 REGISTER_OP(test_operator, paddle::framework::OpWithoutKernelTest,
             paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
 
 TEST(OperatorBase, all) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("test_operator");
-  auto* ipt = op_desc.mutable_inputs()->Add();
-  *ipt->mutable_arguments()->Add() = "IN1";
-  ipt->set_parameter("input");
+  BuildVar("input", {"IN1"}, op_desc.add_inputs());
+  BuildVar("output", {"OUT1"}, op_desc.add_outputs());
 
-  auto* output = op_desc.mutable_outputs()->Add();
-  *output->mutable_arguments()->Add() = "OUT1";
-  output->set_parameter("output");
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
   attr->set_type(paddle::framework::AttrType::FLOAT);
@@ -129,9 +134,9 @@ class OpKernelTestMultiInputsProtoAndCheckerMaker
   OpKernelTestMultiInputsProtoAndCheckerMaker(OpProto* proto,
                                               OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("xs", "inputs of test op").SetMultiple();
+    AddInput("xs", "inputs of test op").AsDuplicable();
     AddInput("k", "input of test op");
-    AddOutput("ys", "outputs of test op").SetMultiple();
+    AddOutput("ys", "outputs of test op").AsDuplicable();
     AddAttr<float>("scale", "scale of cosine op")
         .SetDefault(1.0)
         .LargerThan(0.0);
@@ -188,13 +193,8 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel,
 TEST(OpKernel, all) {
   paddle::framework::OpDesc op_desc;
   op_desc.set_type("op_with_kernel");
-  auto* ipt = op_desc.mutable_inputs()->Add();
-  *ipt->mutable_arguments()->Add() = "IN1";
-  ipt->set_parameter("x");
-
-  auto* output = op_desc.mutable_outputs()->Add();
-  *output->mutable_arguments()->Add() = "OUT1";
-  output->set_parameter("y");
+  BuildVar("x", {"IN1"}, op_desc.add_inputs());
+  BuildVar("y", {"OUT1"}, op_desc.add_outputs());
 
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
@@ -221,18 +221,9 @@ TEST(OpKernel, multi_inputs) {
 
   OpDesc op_desc;
   op_desc.set_type("op_multi_inputs_with_kernel");
-  auto x = op_desc.mutable_inputs()->Add();
-  x->set_parameter("xs");
-  *x->mutable_arguments()->Add() = "x0";
-  *x->mutable_arguments()->Add() = "x1";
-  *x->mutable_arguments()->Add() = "x2";
-  auto k = op_desc.mutable_inputs()->Add();
-  k->set_parameter("k");
-  *k->mutable_arguments()->Add() = "k0";
-  auto y = op_desc.mutable_outputs()->Add();
-  y->set_parameter("ys");
-  *y->mutable_arguments()->Add() = "y0";
-  *y->mutable_arguments()->Add() = "y1";
+  BuildVar("xs", {"x0", "x1", "x2"}, op_desc.add_inputs());
+  BuildVar("k", {"k0"}, op_desc.add_inputs());
+  BuildVar("ys", {"y0", "y1"}, op_desc.add_outputs());
 
   auto attr = op_desc.mutable_attrs()->Add();
   attr->set_name("scale");
diff --git a/paddle/gserver/tests/LayerGradUtil.cpp b/paddle/gserver/tests/LayerGradUtil.cpp
index fd9cfa1dc7a9028cb2c5c98baca98ffb2a837bac..a38880e14cdfcef05461dae567d198e5400c6bb1 100644
--- a/paddle/gserver/tests/LayerGradUtil.cpp
+++ b/paddle/gserver/tests/LayerGradUtil.cpp
@@ -388,14 +388,23 @@ void initDataLayer(TestConfig testConf,
         data.grad->zeroMem();
         break;
       case INPUT_SELF_DEFINE_DATA: {
-        size_t height = testConf.inputDefs[i].selfDefinedData->getHeight();
-        size_t width = testConf.inputDefs[i].selfDefinedData->getWidth();
-        CHECK_GT(static_cast<int>(height), 0);
-        CHECK_GT(static_cast<int>(width), 0);
-        data.value = Matrix::create(height, width, false, useGpu);
-        data.grad = Matrix::create(height, width, false, useGpu);
-        data.value->copyFrom(*testConf.inputDefs[i].selfDefinedData);
-        data.grad->zeroMem();
+        if (testConf.inputDefs[i].ids.size()) {
+          data.ids = IVector::create(testConf.inputDefs[i].ids.size(), useGpu);
+          data.ids->copyFrom(testConf.inputDefs[i].ids.data(),
+                             testConf.inputDefs[i].ids.size());
+        } else if (testConf.inputDefs[i].selfDefinedData) {
+          size_t height = testConf.inputDefs[i].selfDefinedData->getHeight();
+          size_t width = testConf.inputDefs[i].selfDefinedData->getWidth();
+          CHECK_GT(static_cast<int>(height), 0);
+          CHECK_GT(static_cast<int>(width), 0);
+          data.value = Matrix::create(height, width, false, useGpu);
+          data.grad = Matrix::create(height, width, false, useGpu);
+          data.value->copyFrom(*testConf.inputDefs[i].selfDefinedData);
+          data.grad->zeroMem();
+        } else {
+          LOG(FATAL) << "No self-defined data are given.";
+          return;
+        }
 
         const std::vector<int>& labelSeqStartPositions =
             testConf.inputDefs[i].labelSeqStartPositions;
diff --git a/paddle/gserver/tests/LayerGradUtil.h b/paddle/gserver/tests/LayerGradUtil.h
index 5debedf5ef6a3262578ca01b335e664f9a334d35..88e831f78bd165f63806df6c081d84411be51502 100644
--- a/paddle/gserver/tests/LayerGradUtil.h
+++ b/paddle/gserver/tests/LayerGradUtil.h
@@ -68,6 +68,7 @@ struct InputDef {
   std::vector<int> labelInitValue;
   std::vector<int> labelSeqStartPositions;
   std::vector<int> labelSubSeqStartPositions;
+  std::vector<int> ids;
   MatrixPtr selfDefinedData;
 
   InputDef(InputType type, string nameIn, size_t dimIn, size_t sizeIn) {
@@ -95,6 +96,23 @@ struct InputDef {
     isStatic = false;
   }
 
+  InputDef(InputType type,
+           string nameIn,
+           const std::vector<int>& ids,
+           const std::vector<int>& selfDefinedSeqStartPos = {},
+           const std::vector<int>& selfDefinedSubSeqStartPos = {})
+      : labelSeqStartPositions(selfDefinedSeqStartPos),
+        labelSubSeqStartPositions(selfDefinedSubSeqStartPos),
+        ids(ids) {
+    selfDefinedData = nullptr;
+    inputType = type;
+    name = nameIn;
+    dim = 0;
+    sparse = {""};
+    paraSize = 0;
+    isStatic = false;
+  }
+
   InputDef(InputType type,
            string nameIn,
            size_t dimIn,
diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc
index 8e3f0111662296cf9c50bf081edd974c6a8b3488..35e7212dde210a50285272cfd94118fa34fb7cd9 100644
--- a/paddle/operators/mean_op.cc
+++ b/paddle/operators/mean_op.cc
@@ -34,7 +34,7 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
   MeanOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input of mean op");
-    AddOutput("Out", "The output of mean op").IgnoreGradient();
+    AddOutput("Out", "The output of mean op").AsNoGradient();
     AddComment("Mean Operator");
   }
 };
diff --git a/paddle/operators/recurrent_op.cc b/paddle/operators/recurrent_op.cc
index bb30ae6894fd9161c85f8071711e6e2d102fe4a9..5ddee75581824996fd312f8ddf13007759fd9a67 100644
--- a/paddle/operators/recurrent_op.cc
+++ b/paddle/operators/recurrent_op.cc
@@ -155,13 +155,13 @@ class RecurrentAlgorithmProtoAndCheckerMaker
     // inputs and outputs stored in proto
     AddInput(name.inlinks,
              "the inputs that need to be segmented for each step.")
-        .SetMultiple();
+        .AsDuplicable();
     AddInput(name.boot_memories, "variables to initialize memories.")
-        .SetMultiple();
+        .AsDuplicable();
     AddInput(name.step_net, "network shared by all steps.");
 
     AddOutput(name.outlinks, "the outputs that need to concated for all steps.")
-        .SetMultiple();
+        .AsDuplicable();
     AddOutput(name.step_scopes, "step scopes");
 
     // Attributes stored in AttributeMap
diff --git a/paddle/operators/recurrent_op_test.cc b/paddle/operators/recurrent_op_test.cc
index 40c212d6b765e4d48fbba00e82837257219caeb1..2f6eff0720847fdfa6443d2fc233e92dac2d0fab 100644
--- a/paddle/operators/recurrent_op_test.cc
+++ b/paddle/operators/recurrent_op_test.cc
@@ -26,8 +26,6 @@ namespace paddle {
 namespace operators {
 
 using namespace paddle::framework;
-// using framework::make_ddim;
-// using framework::DDim;
 
 class RecurrentGradientAlgorithmTest : public ::testing::Test {
  protected:
diff --git a/python/paddle/v2/framework/tests/CMakeLists.txt b/python/paddle/v2/framework/tests/CMakeLists.txt
index b76c05dc8142af40d9872b42cc51b3c317e095be..96fad9b42e04a88fdcbda093683b57451b2a3e41 100644
--- a/python/paddle/v2/framework/tests/CMakeLists.txt
+++ b/python/paddle/v2/framework/tests/CMakeLists.txt
@@ -24,3 +24,4 @@ py_test(test_default_scope_funcs SRCS test_default_scope_funcs.py)
 py_test(test_operator SRCS test_operator.py)
 # py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
 py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
+py_test(test_recurrent_op SRCS test_recurrent_op.py)
diff --git a/python/paddle/v2/framework/tests/test_add_two_op.py b/python/paddle/v2/framework/tests/test_add_two_op.py
index 019784a8b4d6138d8f8a45f308513c376cd41ebb..0def484eddb88604398ee10390d3f28058714a57 100644
--- a/python/paddle/v2/framework/tests/test_add_two_op.py
+++ b/python/paddle/v2/framework/tests/test_add_two_op.py
@@ -19,13 +19,5 @@ class TestAddOp(unittest.TestCase):
         self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
 
 
-#class TestAddGradOp(unittest.TestCase):
-#    def test_add_grad(self):
-#        op = Operator('add_two', X="X", Y="Y", Out="Out")
-#        backward_op = core.Operator.backward(op, set())
-#        self.assertEqual(backward_op.type(), "add_two_grad")
-#        expected = '''Op(add_two_grad), inputs:(X, Y, Out, Out@GRAD), outputs:(X@GRAD, Y@GRAD).'''
-#        self.assertEqual(expected, str(backward_op))
-
 if __name__ == '__main__':
     unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_recurrent_op.py b/python/paddle/v2/framework/tests/test_recurrent_op.py
index 5c77c477b347f4713e4af2a8cb462b243d7a779c..0db66cc4e181fde10f161a323ea749fd84a5f963 100644
--- a/python/paddle/v2/framework/tests/test_recurrent_op.py
+++ b/python/paddle/v2/framework/tests/test_recurrent_op.py
@@ -2,19 +2,74 @@ import logging
 import paddle.v2.framework.core as core
 import unittest
 import numpy as np
-import paddle.v2.framework.create_op_creation_methods as creation
+from paddle.v2.framework.op import Operator
 
-ops = creation.op_creations
 
+def py_sigmoid(x):
+    return 1. / (1. + np.exp(-x))
 
-def create_tensor(scope, name, shape):
+
+class PySimpleRNN(object):
+    '''
+    A simple implementation of RNN based on numpy, to futhur test RecurrentOp's alogorithm
+    '''
+
+    def __init__(self, input_dim=30, batch_size=50, weight_dim=15, sent_len=11):
+        self.x = np.random.normal(size=(sent_len, batch_size, input_dim))
+        self.W = np.random.normal(size=(input_dim, input_dim))
+        self.U = np.random.normal(size=(input_dim, input_dim))
+        self.h_boot = np.random.normal(size=(batch_size, input_dim))
+
+        # memories
+        self.mems = [
+            np.zeros(shape=(batch_size, input_dim)) for i in range(sent_len)
+        ]
+
+    def forward(self):
+        xs = self.segment_inputs()
+        for step_id in range(self.x.shape[0]):
+            self.step(step_id, xs[step_id])
+        return self.concat_outputs()
+
+    def segment_inputs(self):
+        return [self.x[i] for i in range(self.x.shape[0])]
+
+    def concat_outputs(self):
+        return np.array(self.mems)
+
+    def step(self, step_id, x):
+        '''
+        run a step
+        '''
+        mem = self.mems[step_id]
+        if step_id > 0:
+            pre_mem = self.mems[step_id - 1]
+        else:
+            pre_mem = self.h_boot
+        xW = np.matmul(x, self.W)
+        hU = np.matmul(mem, self.U)
+
+        sum = xW + hU
+        self.mems[step_id] = py_sigmoid(sum)
+
+
+class PySimpleRNNTest(unittest.TestCase):
+    def setUp(self):
+        self.rnn = PySimpleRNN()
+
+    def test_forward(self):
+        output = self.rnn.forward()
+        print 'output', output
+
+
+def create_tensor(scope, name, shape, np_data):
     tensor = scope.new_var(name).get_tensor()
     tensor.set_dims(shape)
-    tensor.set(np.random.random(shape), core.CPUPlace())
+    tensor.set(np_data, core.CPUPlace())
     return tensor
 
 
-class TestRNN(unittest.TestCase):
+class TestRecurrentOp(unittest.TestCase):
     '''
     Test RNNOp
 
@@ -28,7 +83,7 @@ class TestRNN(unittest.TestCase):
     memories:
         - h
     outputs:
-        - h
+       - h
     '''
 
     input_dim = 30
@@ -36,33 +91,45 @@ class TestRNN(unittest.TestCase):
     weight_dim = 15
     sent_len = 11
 
-    def init(self):
+    def setUp(self):
+        self.py_rnn = PySimpleRNN(self.input_dim, self.batch_size,
+                                  self.weight_dim, self.sent_len)
 
+    def forward(self):
         self.scope = core.Scope()
-
         self.create_global_variables()
         self.create_step_net()
         rnn_op = self.create_rnn_op()
         ctx = core.DeviceContext.create(core.CPUPlace())
-        print 'infer_shape'
         rnn_op.infer_shape(self.scope)
-
         rnn_op.run(self.scope, ctx)
+        return np.array(self.scope.find_var("h").get_tensor())
 
     def create_global_variables(self):
         # create inlink
+        x_np_data = self.py_rnn.x
         create_tensor(self.scope, "x",
-                      [self.sent_len, self.batch_size, self.input_dim])
-        create_tensor(self.scope, "W", [self.input_dim, self.input_dim])
-        create_tensor(self.scope, "U", [self.input_dim, self.input_dim])
-        create_tensor(self.scope, "h_boot", [self.batch_size, self.input_dim])
+                      [self.sent_len, self.batch_size, self.input_dim],
+                      x_np_data)
+        W_np_data = self.py_rnn.W
+        create_tensor(self.scope, "W", [self.input_dim, self.input_dim],
+                      W_np_data)
+
+        U_np_data = self.py_rnn.U
+        create_tensor(self.scope, "U", [self.input_dim, self.input_dim],
+                      U_np_data)
+
+        h_boot_np_data = self.py_rnn.h_boot
+        create_tensor(self.scope, "h_boot", [self.batch_size, self.input_dim],
+                      h_boot_np_data)
         self.scope.new_var("step_scopes")
         self.scope.new_var("h@alias")
         self.scope.new_var("h")
 
     def create_rnn_op(self):
         # create RNNOp
-        rnnop = ops.recurrent_op(
+        rnnop = Operator(
+            "recurrent_op",
             # inputs
             inlinks=["x"],
             boot_memories=["h_boot"],
@@ -81,17 +148,25 @@ class TestRNN(unittest.TestCase):
         var = self.scope.new_var("stepnet")
         stepnet = var.get_net()
 
-        x_fc_op = ops.fc(X="x@alias", W="W", Y="Wx")
-        h_fc_op = ops.fc(X="h@pre", W="U", Y="Uh")
-        sum_op = ops.add_two(X="Wx", Y="Uh", Out="sum")
-        sig_op = ops.sigmoid(X="sum", Y="h@alias")
+        # x_fc_op = Operator("fc", X="x@alias", W="W", Y="Wx")
+        # h_fc_op = Operator("fc", X="h@pre", W="U", Y="Uh")
+        x_fc_op = Operator("mul", X="x@alias", Y="W", Out="Wx")
+        h_fc_op = Operator("mul", X="h@pre", Y="U", Out="Uh")
+        sum_op = Operator("add_two", X="Wx", Y="Uh", Out="sum")
+        sig_op = Operator("sigmoid", X="sum", Y="h@alias")
 
         for op in [x_fc_op, h_fc_op, sum_op, sig_op]:
             stepnet.add_op(op)
         stepnet.complete_add_op(True)
 
-    def test_recurrent(self):
-        self.init()
+    def test_forward(self):
+        print 'test recurrent op forward'
+        pd_output = self.forward()
+        py_output = self.py_rnn.forward()
+        print 'pd_output', pd_output
+        print
+        print 'py_output', py_output
+        self.assertEqual(pd_output.shape, py_output.shape)
 
 
 if __name__ == '__main__':
diff --git a/python/requirements.txt b/python/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..3df822bd76d2a64a0a35f84b4ec309ce7150c221
--- /dev/null
+++ b/python/requirements.txt
@@ -0,0 +1,9 @@
+requests==2.9.2
+numpy>=1.12
+protobuf==3.1
+recordio
+matplotlib
+rarfile
+scipy>=0.19.0
+Pillow
+nltk>=3.2.2
diff --git a/python/setup.py.in b/python/setup.py.in
index 4110c983180937e86716324f8c92c37b7d2cc3ea..38728aa2fd77cf3c882479ed83e99688b9ffa541 100644
--- a/python/setup.py.in
+++ b/python/setup.py.in
@@ -1,5 +1,4 @@
 from setuptools import setup, Distribution
-
 class BinaryDistribution(Distribution):
     def has_ext_modules(foo):
         return True
@@ -18,15 +17,8 @@ packages=['paddle',
           'paddle.v2.framework.proto',
           'py_paddle']
 
-setup_requires=["requests",
-                "numpy>=1.12",
-                "protobuf==3.1",
-                "recordio",
-                "matplotlib",
-                "rarfile",
-                "scipy>=0.19.0",
-                "Pillow",
-                "nltk>=3.2.2"]
+with open('@PADDLE_SOURCE_DIR@/python/requirements.txt') as f:
+    setup_requires = f.read().splitlines()
 
 if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
     setup_requires+=["opencv-python"]