Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into cmake_speed

paddle/gserver/layers/MKLDNNAddtoLayer.cpp

@@ -54,7 +54,6 @@ void MKLDNNAddtoLayer::reshape(
   ow = iw;
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc * oh * ow);
-  printSizeInfo();
 }
 
 void MKLDNNAddtoLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNBatchNormLayer.cpp

@@ -125,7 +125,6 @@ void MKLDNNBatchNormLayer::reshape(
       << "Input channel can not be changed";
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc * oh * ow);
-  printSizeInfo();
 }
 
 void MKLDNNBatchNormLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -102,8 +102,6 @@ void MKLDNNConvLayer::reshape(
 
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc * oh * ow);
-
-  printSizeInfo();
 }
 
 void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNConvLayer.h

@@ -92,7 +92,7 @@ public:
   void printSizeInfo() override {
     MKLDNNLayer::printSizeInfo();
     VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
-                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
+                       << ", ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
                        << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
   }
 

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -84,8 +84,6 @@ void MKLDNNFcLayer::reshape(
 
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc);
-
-  printSizeInfo();
 }
 
 void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNPoolLayer.cpp

@@ -71,8 +71,6 @@ void MKLDNNPoolLayer::reshape(
   reshapeOutput(oh, ow);
 
   resizeOutput(bs, oc * oh * ow);
-
-  printSizeInfo();
 }
 
 void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/ROIPoolLayer.cpp

@@ -98,7 +98,7 @@ void ROIPoolLayer::forward(PassType passType) {
     size_t roiStartH = round(bottomROIs[2] * spatialScale_);
     size_t roiEndW = round(bottomROIs[3] * spatialScale_);
     size_t roiEndH = round(bottomROIs[4] * spatialScale_);
-    CHECK_GE(roiBatchIdx, 0);
+    CHECK_GE(roiBatchIdx, 0UL);
     CHECK_LT(roiBatchIdx, batchSize);
     size_t roiHeight = std::max(roiEndH - roiStartH + 1, 1UL);
     size_t roiWidth = std::max(roiEndW - roiStartW + 1, 1UL);

paddle/gserver/tests/test_MKLDNN.cpp

@@ -297,7 +297,7 @@ static void getAddtoConfig(TestConfig& cfg,
 }
 
 void testAddtoLayer(const testImageDesc& pm, const size_t nInputs) {
-  CHECK_GE(nInputs, 1);
+  CHECK_GE(nInputs, 1UL);
   TestConfig dnnConfig;
   getAddtoConfig(dnnConfig, pm, nInputs);
   dnnConfig.layerConfig.set_type("mkldnn_addto");

paddle/math/MKLDNNMatrix.cpp

@@ -152,12 +152,7 @@ void MKLDNNMatrix::downSpatial() {
   }
   memory::desc md = memory::desc(dstDims, getDtype(), dstFmt);
   memory::primitive_desc pd = memory::primitive_desc(md, getEngine());
-  mkldnn_primitive_t result;
-  mkldnn::error::wrap_c_api(
-      mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
-      "could not create a memory primitive");
-  reset(result);
-  set_data_handle(data_);
+  resetMKLDNNMemory(pd, data_);
 }
 
 }  // namespace paddle

paddle/math/MKLDNNMatrix.h

@@ -145,6 +145,27 @@ public:
     m_.reset();
   }
 
+  /**
+   * override the CpuMatrix::resize
+   */
+  void resize(size_t newHeight, size_t newWidth) override {
+    m_->resize(newHeight, newWidth);
+    if (data_ == m_->getData() && elementCnt_ == newHeight * newWidth) {
+      return;
+    }
+    CpuMatrix::setData(data_);
+    height_ = newHeight;
+    width_ = newWidth;
+    elementCnt_ = newHeight * newWidth;
+    stride_ = width_;
+    auto pd = mkldnn::memory::primitive_desc(
+        mkldnn::memory::desc({(int)newHeight, (int)newWidth},
+                             getDtype(),
+                             mkldnn::memory::format::nc),
+        getEngine());
+    resetMKLDNNMemory(pd, data_);
+  }
+
   /**
    * override Matrix::getData
    * check data before return
@@ -215,6 +236,17 @@ protected:
                    memory::format srcFmt,
                    memory::format dstFmt,
                    memory::dims dm);
+  /**
+   * reset this MKLDNN Memory from primitve desc
+   */
+  void resetMKLDNNMemory(memory::primitive_desc pd, real* data) {
+    mkldnn_primitive_t result;
+    mkldnn::error::wrap_c_api(
+        mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
+        "could not create a memory primitive");
+    reset(result);
+    set_data_handle(data);
+  }
 
 private:
   // save the CpuMatrixPtr in case the buffer released outside

paddle/operators/CMakeLists.txt

@@ -225,6 +225,7 @@ set(GLOB_OP_LIB ${OP_LIBRARY} CACHE INTERNAL "Global OP library")
 cc_test(gather_test SRCS gather_test.cc DEPS tensor)
 cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
 cc_test(scatter_test SRCS scatter_test.cc DEPS tensor)
+cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_tensor)
 cc_test(strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor paddle_memory)
 cc_test(dynamic_recurrent_op_test SRCS dynamic_recurrent_op_test.cc
         rnn/recurrent_op_utils.cc

paddle/operators/beam_search_decode_op.cc

+/* 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/operators/beam_search_decode_op.h"
+
+namespace paddle {
+namespace operators {
+
+class BeamSearchDecodeOp : public framework::OperatorBase {
+ public:
+  BeamSearchDecodeOp(const std::string& type,
+                     const framework::VariableNameMap& inputs,
+                     const framework::VariableNameMap& outputs,
+                     const framework::AttributeMap& attrs)
+      : OperatorBase(type, inputs, outputs, attrs) {}
+  void Run(const framework::Scope& scope,
+           const platform::DeviceContext& dev_ctx) const override {
+    framework::ExecutionContext ctx(*this, scope, dev_ctx);
+    const LoDTensorArray* ids = ctx.Input<LoDTensorArray>("Ids");
+    const LoDTensorArray* scores = ctx.Input<LoDTensorArray>("Scores");
+    const size_t step_num = ids->size();
+    PADDLE_ENFORCE_GT(step_num, 0UL,
+                      "beam search steps should be larger than 0");
+    const size_t source_num = ids->at(0).lod().at(0).size() - 1;
+    PADDLE_ENFORCE_GT(source_num, 0UL, "source num should be larger than 0");
+
+    for (size_t i = 0; i < step_num; ++i) {
+      PADDLE_ENFORCE_EQ(ids->at(i).lod().size(), 2UL,
+                        "Level of LodTensor should be 2");
+    }
+
+    // prepare output
+    LoDTensor* sentenceIds = ctx.Output<LoDTensor>("SentenceIds");
+    LoDTensor* sentenceScores = ctx.Output<LoDTensor>("SentenceScores");
+
+    BeamSearchDecoder<float> beam_search_decoder;
+    beam_search_decoder.PackAllSteps(*ids, *scores, sentenceIds,
+                                     sentenceScores);
+  }
+};
+
+class BeamSearchDecodeOpProtoMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  BeamSearchDecodeOpProtoMaker(framework::OpProto* proto,
+                               framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("Ids",
+             "(LodTensorArray)"
+             "score of the candidate words in each step");
+    AddInput("Scores",
+             "(LodTensorArray)"
+             "score of the candidate words in each step");
+    AddOutput("SentenceIds",
+              "(LodTensor)"
+              "All possible result sentences of word ids");
+    AddOutput("SentenceScores",
+              "(LodTensor)"
+              "All possible result sentences of word scores");
+    AddComment(R"DOC(
+Pack the result of Beam search op into SentenceIds and SentenceScores.
+)DOC");
+  }
+};
+
+class BeamSearchDecodeInferShape : public framework::InferShapeBase {
+ public:
+  void operator()(framework::InferShapeContext* context) const override {
+    PADDLE_ENFORCE(context->HasInput("Ids"),
+                   "BeamSearchDecodeOp must has input Ids");
+    PADDLE_ENFORCE(context->HasInput("Scores"),
+                   "BeamSearchDecodeOp must has input Scores");
+    PADDLE_ENFORCE(context->HasOutput("SentenceIds"),
+                   "BeamSearchDecodeOp must has output SentenceIds");
+    PADDLE_ENFORCE(context->HasOutput("SentenceScores"),
+                   "BeamSearchDecodeOp must has output SentenceScores");
+  }
+};
+
+class BeamSearchDecodeInferVarType : public framework::VarTypeInference {
+ public:
+  void operator()(const framework::OpDescBind& op_desc,
+                  framework::BlockDescBind* block) const override {
+    for (auto& o : op_desc.Output("SentenceIds")) {
+      block->Var(o)->SetType(framework::VarDesc::LOD_TENSOR);
+    }
+    for (auto& o : op_desc.Output("SentenceScores")) {
+      block->Var(o)->SetType(framework::VarDesc::LOD_TENSOR);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+REGISTER_OPERATOR(beam_search_decode, paddle::operators::BeamSearchDecodeOp,
+                  paddle::operators::BeamSearchDecodeOpProtoMaker,
+                  paddle::operators::BeamSearchDecodeInferShape,
+                  paddle::operators::BeamSearchDecodeInferVarType,
+                  paddle::framework::EmptyGradOpMaker);

paddle/operators/beam_search_decode_op.h

+/* 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/lod_tensor_array.h"
+#include "paddle/framework/op_registry.h"
+
+namespace paddle {
+namespace operators {
+
+using LoDTensor = framework::LoDTensor;
+using LoDTensorArray = framework::LoDTensorArray;
+
+// all the lod have 2 levels.
+// The First is source level, the second is sentence level.
+// source level describe how many candidate words for this source.
+// sentence level describe these candidates belong to which prefix
+const size_t kSourceLevel = 0;
+const size_t kSentenceLevel = 1;
+
+template <typename T>
+struct BeamNode {
+  BeamNode(int64_t word_id, T score) : word_id_(word_id), score_(score) {}
+
+  ~BeamNode() {
+    if (parent_) {
+      parent_->DropKid(this);
+      if (parent_->kids_.size() == 0UL) {
+        delete parent_;
+      }
+    }
+    VLOG(3) << "Delete BeamNode root with word_id:" << this->word_id_;
+  }
+
+  void AppendTo(BeamNode* parent) {
+    parent_ = parent;
+    parent->kids_.insert(this);
+  }
+
+  void DropKid(BeamNode* kid) { kids_.erase(kid); }
+
+  BeamNode* parent_ = nullptr;
+  std::unordered_set<BeamNode*> kids_;
+  int64_t word_id_;
+  T score_;
+};
+
+template <typename T>
+using BeamNodeVector = std::vector<std::unique_ptr<BeamNode<T>>>;
+
+template <typename T>
+struct Sentence {
+  std::vector<int64_t> word_ids;
+  std::vector<T> scores;
+};
+
+template <typename T>
+using SentenceVector = std::vector<Sentence<T>>;
+
+template <typename T>
+struct BeamSearchDecoder {
+  /**
+   * make a BeamNode and all it's related prefix BeanNode into a Sentence.
+   */
+  Sentence<T> MakeSentence(const BeamNode<T>* node) const;
+
+  /**
+   * Param:
+   *  cur_ids: LoDTensor of One step for word ID
+   *  cur_scores: LoDTensor of One Step for word score
+   *  prefixes_list: prefixes for each source sentence.
+   *  sentence_vector_list: result sentence_vector for each source sentence.
+   * Return:
+   *  a new prefixes list for each source of current step
+   */
+  std::vector<BeamNodeVector<T>> PackTwoSteps(
+      const LoDTensor& cur_ids, const LoDTensor& cur_scores,
+      std::vector<BeamNodeVector<T>>& prefixes_list,
+      std::vector<SentenceVector<T>>* sentence_vector_list) const;
+
+  /**
+   * convert the result sentence_vector for each source sentence into two
+   * LodTensor.
+   * One is all candidate sentences with word id, one is all candidate sentences
+   * with word score.
+   * Param:
+   *  sentence_vector_list: sentence_vector for each source sentence.
+   *  id_tensor: result LoDTensor for sentences of id.
+   *  score_tensor: result LoDTensor for sentences of score.
+   */
+  void ConvertSentenceVectorToLodTensor(
+      std::vector<SentenceVector<T>> sentence_vector_list, LoDTensor* id_tensor,
+      LoDTensor* score_tensor) const;
+
+  /**
+   * Pack all steps of id/score LodTensor into sentence LoDTensor
+   * it's main logic is:
+   * ```python
+   *   prefix
+   *   result_sentence
+   *   result_lod_tensor
+   *
+   *   for (step in steps):
+   *     prefix = PackTwoSteps(prefix, step, &result_sentence)
+   *   ConvertSentenceVector<T>ToLodTensor(result_sentence, &result_lod_tensor)
+   * ```
+   */
+  void PackAllSteps(const LoDTensorArray& step_ids,
+                    const LoDTensorArray& step_scores, LoDTensor* id_tensor,
+                    LoDTensor* score_tensor) const;
+};
+
+template <typename T>
+Sentence<T> BeamSearchDecoder<T>::MakeSentence(const BeamNode<T>* node) const {
+  Sentence<T> sentence;
+  while (node != nullptr) {
+    sentence.word_ids.emplace_back(node->word_id_);
+    sentence.scores.emplace_back(node->score_);
+    node = node->parent_;
+  }
+
+  std::reverse(std::begin(sentence.word_ids), std::end(sentence.word_ids));
+  std::reverse(std::begin(sentence.scores), std::end(sentence.scores));
+
+  return sentence;
+}
+
+template <typename T>
+std::vector<BeamNodeVector<T>> BeamSearchDecoder<T>::PackTwoSteps(
+    const LoDTensor& cur_ids, const LoDTensor& cur_scores,
+    std::vector<BeamNodeVector<T>>& prefixes_list,
+    std::vector<SentenceVector<T>>* sentence_vector_list) const {
+  std::vector<BeamNodeVector<T>> result;
+
+  for (size_t src_idx = 0; src_idx < cur_ids.lod()[kSourceLevel].size() - 1;
+       ++src_idx) {
+    size_t src_start = cur_ids.lod().at(kSourceLevel)[src_idx];
+    size_t src_end = cur_ids.lod().at(kSourceLevel)[src_idx + 1];
+
+    BeamNodeVector<T> beam_nodes;
+
+    // if prefixes size is 0, it means this is the first step. In this step,
+    // all candidate id is the start of candidate sentences.
+    if (prefixes_list.empty()) {
+      PADDLE_ENFORCE_EQ(cur_ids.lod().at(kSourceLevel).back(),
+                        cur_ids.lod().at(kSentenceLevel).back(),
+                        "in the first step");
+      for (size_t id_idx = src_start; id_idx < src_end; ++id_idx) {
+        beam_nodes.push_back(std::unique_ptr<BeamNode<T>>(new BeamNode<T>(
+            cur_ids.data<int64_t>()[id_idx], cur_scores.data<T>()[id_idx])));
+      }
+    } else {
+      BeamNodeVector<T>& prefixes = prefixes_list[src_idx];
+      SentenceVector<T>& sentence_vector = (*sentence_vector_list)[src_idx];
+
+      PADDLE_ENFORCE_EQ(src_end - src_start, prefixes.size(),
+                        "prefix and candidate set number should be the same");
+
+      auto candidate_offset = cur_ids.lod()[kSentenceLevel];
+      for (size_t prefix_idx = 0; prefix_idx < prefixes.size(); ++prefix_idx) {
+        std::unique_ptr<BeamNode<T>>& prefix = prefixes[prefix_idx];
+        size_t candidate_start = candidate_offset[src_start + prefix_idx];
+        size_t candidate_end = candidate_offset[src_start + prefix_idx + 1];
+        if (candidate_start == candidate_end) {
+          VLOG(3) << "this sentence has no more candidate, "
+                     "add to result sentence and rm it from beam tree";
+          sentence_vector.push_back(MakeSentence(prefix.get()));
+          prefix.reset();
+        } else {
+          for (size_t candidate_idx = candidate_start;
+               candidate_idx < candidate_end; ++candidate_idx) {
+            auto* candidate =
+                new BeamNode<T>(cur_ids.data<int64_t>()[candidate_idx],
+                                cur_scores.data<T>()[candidate_idx]);
+            candidate->AppendTo(prefix.get());
+            beam_nodes.push_back(std::unique_ptr<BeamNode<T>>(candidate));
+          }
+          prefix.release();
+        }
+      }
+    }
+    result.push_back(std::move(beam_nodes));
+  }
+  return result;
+}
+
+template <typename T>
+void BeamSearchDecoder<T>::ConvertSentenceVectorToLodTensor(
+    std::vector<SentenceVector<T>> sentence_vector_list, LoDTensor* id_tensor,
+    LoDTensor* score_tensor) const {
+  size_t src_num = sentence_vector_list.size();
+
+  PADDLE_ENFORCE_NE(src_num, 0, "src_num should not be 0");
+
+  std::vector<size_t> source_level_lod = {0};
+  std::vector<size_t> sentence_level_lod = {0};
+  std::vector<int64_t> id_data;
+  std::vector<T> score_data;
+
+  for (size_t src_idx = 0; src_idx < src_num; ++src_idx) {
+    for (Sentence<T>& sentence : sentence_vector_list[src_idx]) {
+      id_data.insert(id_data.end(), sentence.word_ids.begin(),
+                     sentence.word_ids.end());
+      score_data.insert(score_data.end(), sentence.scores.begin(),
+                        sentence.scores.end());
+      sentence_level_lod.push_back(sentence_level_lod.back() +
+                                   sentence.word_ids.size());
+    }
+    source_level_lod.push_back(source_level_lod.back() +
+                               sentence_vector_list[src_idx].size());
+  }
+
+  auto cpu_place = new paddle::platform::CPUPlace();
+  paddle::platform::CPUDeviceContext cpu_ctx(*cpu_place);
+
+  framework::LoD lod;
+  lod.push_back(source_level_lod);
+  lod.push_back(sentence_level_lod);
+
+  id_tensor->set_lod(lod);
+  id_tensor->Resize({static_cast<int64_t>(id_data.size())});
+  id_tensor->mutable_data<int64_t>(paddle::platform::CPUPlace());
+  id_tensor->CopyFromVector<int64_t>(id_data, cpu_ctx);
+
+  score_tensor->set_lod(lod);
+  score_tensor->Resize({static_cast<int64_t>(score_data.size())});
+  score_tensor->mutable_data<T>(paddle::platform::CPUPlace());
+  score_tensor->CopyFromVector<T>(score_data, cpu_ctx);
+}
+
+template <typename T>
+void BeamSearchDecoder<T>::PackAllSteps(const LoDTensorArray& step_ids,
+                                        const LoDTensorArray& step_scores,
+                                        LoDTensor* id_tensor,
+                                        LoDTensor* score_tensor) const {
+  PADDLE_ENFORCE(!step_ids.empty(), "step num should be larger than 0");
+  PADDLE_ENFORCE_EQ(step_ids.size(), step_scores.size(),
+                    "step_ids and step_scores should be the same");
+  const size_t step_num = step_ids.size();
+  const size_t src_num = step_ids.at(0).lod().at(kSourceLevel).size() - 1;
+
+  PADDLE_ENFORCE_GT(src_num, 0UL, "source num should be larger than 0");
+
+  // previous prefixes for each step,
+  // the init length is 0, means this is the first step.
+  std::vector<BeamNodeVector<T>> beamnode_vector_list(0);
+  std::vector<SentenceVector<T>> sentence_vector_list(src_num);
+
+  // pack all steps for one batch first, then another batch
+  for (size_t step_id = 0; step_id < step_num; ++step_id) {
+    beamnode_vector_list =
+        PackTwoSteps(step_ids.at(step_id), step_scores.at(step_id),
+                     beamnode_vector_list, &sentence_vector_list);
+  }
+  // append last beam_node to result
+  for (size_t src_idx = 0; src_idx < src_num; ++src_idx) {
+    for (auto& beam_node : beamnode_vector_list.at(src_idx)) {
+      sentence_vector_list[src_idx].push_back(MakeSentence(beam_node.get()));
+      beam_node.reset();
+    }
+  }
+
+  ConvertSentenceVectorToLodTensor(sentence_vector_list, id_tensor,
+                                   score_tensor);
+}
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/beam_search_decode_op_test.cc

+/* 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/operators/beam_search_decode_op.h"
+#include "gtest/gtest.h"
+
+using CPUPlace = paddle::platform::CPUPlace;
+using LoD = paddle::framework::LoD;
+using LoDTensor = paddle::framework::LoDTensor;
+using LoDTensorArray = paddle::framework::LoDTensorArray;
+
+template <typename T>
+using BeamNode = paddle::operators::BeamNode<T>;
+template <typename T>
+using BeamSearchDecoder = paddle::operators::BeamSearchDecoder<T>;
+template <typename T>
+using Sentence = paddle::operators::Sentence<T>;
+template <typename T>
+using BeamNodeVector = paddle::operators::BeamNodeVector<T>;
+template <typename T>
+using SentenceVector = paddle::operators::SentenceVector<T>;
+
+namespace paddle {
+namespace test {
+
+void GenerateExample(const std::vector<size_t>& level_0,
+                     const std::vector<size_t>& level_1,
+                     const std::vector<int>& data, LoDTensorArray* ids,
+                     LoDTensorArray* scores) {
+  PADDLE_ENFORCE_EQ(level_0.back(), level_1.size() - 1,
+                    "source level is used to describe candidate set");
+  PADDLE_ENFORCE_EQ(level_1.back(), data.size(),
+                    "the lowest level is used to describe data"
+                    ", so it's last element should be data length");
+
+  CPUPlace place;
+
+  LoD lod;
+  lod.push_back(level_0);
+  lod.push_back(level_1);
+
+  // Ids
+  LoDTensor tensor_id;
+  tensor_id.set_lod(lod);
+  tensor_id.Resize({static_cast<int64_t>(data.size())});
+  // malloc memory
+  int64_t* id_ptr = tensor_id.mutable_data<int64_t>(place);
+  for (size_t i = 0; i < data.size(); ++i) {
+    id_ptr[i] = static_cast<int64_t>(data.at(i));
+  }
+
+  // Scores
+  LoDTensor tensor_score;
+  tensor_score.set_lod(lod);
+  tensor_score.Resize({static_cast<int64_t>(data.size())});
+  // malloc memory
+  float* score_ptr = tensor_score.mutable_data<float>(place);
+  for (size_t i = 0; i < data.size(); ++i) {
+    score_ptr[i] = static_cast<float>(data.at(i));
+  }
+
+  ids->push_back(tensor_id);
+  scores->push_back(tensor_score);
+}
+
+}  // namespace test
+}  // namespace paddle
+
+TEST(BeamSearchDecodeOp, DeleteBeamNode) {
+  auto* root = new BeamNode<float>(0, 0);
+  auto* b1 = new BeamNode<float>(1, 1);
+  auto* b2 = new BeamNode<float>(2, 2);
+  auto* b3 = new BeamNode<float>(3, 3);
+
+  b1->AppendTo(root);
+  b2->AppendTo(root);
+  b3->AppendTo(b1);
+
+  delete b3;
+  delete b2;
+}
+
+TEST(BeamSearchDecodeOp, MakeSentence) {
+  auto* root = new BeamNode<float>(0, 0);
+  auto* b1 = new BeamNode<float>(1, 1);
+  auto* end = new BeamNode<float>(2, 2);
+  b1->AppendTo(root);
+  end->AppendTo(b1);
+
+  BeamSearchDecoder<float> helper;
+  Sentence<float> sentence = helper.MakeSentence(end);
+  delete end;
+
+  std::vector<int64_t> expect_ids = {0, 1, 2};
+  ASSERT_EQ(sentence.word_ids, expect_ids);
+
+  std::vector<float> expect_scores = {0, 1, 2};
+  ASSERT_EQ(sentence.scores, expect_scores);
+}
+
+TEST(BeamSearchDecodeOp, PackTwoStepsFistStep) {
+  CPUPlace place;
+
+  LoDTensorArray ids;
+  LoDTensorArray scores;
+
+  paddle::test::GenerateExample(
+      std::vector<size_t>{0, 2, 6}, std::vector<size_t>{0, 1, 2, 3, 4, 5, 6},
+      std::vector<int>{1, 2, 3, 4, 5, 6}, &ids, &scores);
+
+  std::vector<BeamNodeVector<float>> beamnode_vector_list;
+  std::vector<SentenceVector<float>> sentence_vector_list(
+      2, SentenceVector<float>());
+
+  BeamSearchDecoder<float> helper;
+  beamnode_vector_list = helper.PackTwoSteps(
+      ids[0], scores[0], beamnode_vector_list, &sentence_vector_list);
+  ASSERT_EQ(beamnode_vector_list.size(), 2UL);
+  ASSERT_EQ(beamnode_vector_list[0].size(), 2UL);
+  ASSERT_EQ(beamnode_vector_list[1].size(), 4UL);
+}
+
+TEST(BeamSearchDecodeOp, PackTwoSteps) {
+  CPUPlace place;
+
+  // first source has three prefix
+  BeamNodeVector<float> source0_prefixes;
+  source0_prefixes.push_back(
+      std::unique_ptr<BeamNode<float>>(new BeamNode<float>(1, 1)));
+  source0_prefixes.push_back(
+      std::unique_ptr<BeamNode<float>>(new BeamNode<float>(0, 0)));
+  source0_prefixes.push_back(
+      std::unique_ptr<BeamNode<float>>(new BeamNode<float>(3, 3)));
+
+  // second source has two prefix
+  BeamNodeVector<float> source1_prefixes;
+  source1_prefixes.push_back(
+      std::unique_ptr<BeamNode<float>>(new BeamNode<float>(4, 4)));
+  source1_prefixes.push_back(
+      std::unique_ptr<BeamNode<float>>(new BeamNode<float>(5, 5)));
+
+  std::vector<BeamNodeVector<float>> beamnode_vector_list;
+  std::vector<SentenceVector<float>> sentence_vector_list(
+      2, SentenceVector<float>());
+
+  beamnode_vector_list.push_back(std::move(source0_prefixes));
+  beamnode_vector_list.push_back(std::move(source1_prefixes));
+
+  // generate data for one step
+  LoDTensorArray ids;
+  LoDTensorArray scores;
+
+  paddle::test::GenerateExample(std::vector<size_t>{0, 3, 5},
+                                std::vector<size_t>{0, 1, 1, 3, 4, 5},
+                                std::vector<int>{0, 1, 2, 3, 4}, &ids, &scores);
+
+  BeamSearchDecoder<float> helper1;
+  beamnode_vector_list = helper1.PackTwoSteps(
+      ids[0], scores[0], beamnode_vector_list, &sentence_vector_list);
+
+  ASSERT_EQ(sentence_vector_list[0].size(), 1UL);
+  ASSERT_EQ(sentence_vector_list[1].size(), 0UL);
+  ASSERT_EQ(beamnode_vector_list[0].size(), 3UL);
+  ASSERT_EQ(beamnode_vector_list[1].size(), 2UL);
+}
+
+TEST(BeamSearchDecodeOp, PackAllSteps) {
+  CPUPlace place;
+
+  // we will constuct a sample data with 3 steps and 2 source sentences
+  LoDTensorArray ids;
+  LoDTensorArray scores;
+
+  paddle::test::GenerateExample(
+      std::vector<size_t>{0, 3, 6}, std::vector<size_t>{0, 1, 2, 3, 4, 5, 6},
+      std::vector<int>{1, 2, 3, 4, 5, 6}, &ids, &scores);
+  paddle::test::GenerateExample(
+      std::vector<size_t>{0, 3, 6}, std::vector<size_t>{0, 1, 1, 3, 5, 5, 6},
+      std::vector<int>{0, 1, 2, 3, 4, 5}, &ids, &scores);
+  paddle::test::GenerateExample(std::vector<size_t>{0, 3, 6},
+                                std::vector<size_t>{0, 0, 1, 2, 3, 4, 5},
+                                std::vector<int>{0, 1, 2, 3, 4}, &ids, &scores);
+
+  ASSERT_EQ(ids.size(), 3UL);
+  ASSERT_EQ(scores.size(), 3UL);
+
+  BeamSearchDecoder<float> helper;
+
+  LoDTensor id_tensor;
+  LoDTensor score_tensor;
+  helper.PackAllSteps(ids, scores, &id_tensor, &score_tensor);
+
+  LoD lod = id_tensor.lod();
+  std::vector<size_t> expect_source_lod = {0, 4, 8};
+  EXPECT_EQ(lod[0], expect_source_lod);
+  std::vector<size_t> expect_sentence_lod = {0, 1, 3, 6, 9, 10, 13, 16, 19};
+  EXPECT_EQ(lod[1], expect_sentence_lod);
+  // 2| 1, 0| 3, 1, 0| 3, 2, 1| 5| 4, 3, 2| 4, 4, 3| 6, 5, 4
+  std::vector<int> expect_data = {2, 1, 0, 3, 1, 0, 3, 2, 1, 5,
+                                  4, 3, 2, 4, 4, 3, 6, 5, 4};
+  ASSERT_EQ(id_tensor.dims()[0], static_cast<int64_t>(expect_data.size()));
+  for (size_t i = 0; i < expect_data.size(); ++i) {
+    ASSERT_EQ(id_tensor.data<int64_t>()[i],
+              static_cast<int64_t>(expect_data[i]));
+  }
+  for (int64_t i = 0; i < id_tensor.dims()[0]; ++i) {
+    ASSERT_EQ(score_tensor.data<float>()[i],
+              static_cast<float>(id_tensor.data<int64_t>()[i]));
+  }
+}

paddle/operators/sequence_concat_op.cc

@@ -47,7 +47,7 @@ class SequenceConcatOpMaker : public framework::OpProtoAndCheckerMaker {
                         framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X",
-             "(vector<LoDTensor>) Input is a vector of LoDTensor, "
+             "(LodTensorArray) Input is a vector of LoDTensor, "
              "each of which is a variable-length sequence or nested sequence.")
         .AsDuplicable();
     AddOutput("Out",

paddle/platform/call_once.h

@@ -27,20 +27,22 @@ namespace platform {
 
  This wrap is a hack to avoid this bug.
 */
-template <class Callable, class... Args>
+template <typename Callable, typename... Args>
 inline void call_once(std::once_flag& flag, Callable&& f, Args&&... args) {
   bool good = false;
   std::exception ex;
-  std::call_once(flag, [&]() {
-    try {
-      f(args...);
-      good = true;
-    } catch (const std::exception& e) {
-      ex = e;
-    } catch (...) {
-      ex = std::runtime_error("excption caught in call_once");
-    }
-  });
+  std::call_once(flag,
+                 [&](Args&&... args) {
+                   try {
+                     f(args...);
+                     good = true;
+                   } catch (const std::exception& e) {
+                     ex = e;
+                   } catch (...) {
+                     ex = std::runtime_error("excption caught in call_once");
+                   }
+                 },
+                 args...);
   if (!good) {
     throw std::exception(ex);
   }

python/paddle/v2/framework/layer_helper.py

@@ -4,7 +4,7 @@ import itertools
 from paddle.v2.framework.framework import Variable, g_main_program, \
     g_startup_program, unique_name, Program
 from paddle.v2.framework.initializer import ConstantInitializer, \
-    UniformInitializer
+    UniformInitializer, XavierInitializer
 
 
 class LayerHelper(object):
@@ -61,7 +61,7 @@ class LayerHelper(object):
 
     @property
     def param_attr(self):
-        default = {'name': None, 'initializer': UniformInitializer()}
+        default = {'name': None, 'initializer': XavierInitializer()}
         actual = self.kwargs.get('param_attr', None)
         if actual is None:
             actual = default
@@ -70,10 +70,11 @@ class LayerHelper(object):
                 actual[default_field] = default[default_field]
         return actual
 
+    @property
     def bias_attr(self):
-        default = {'name': None, 'initializer': ConstantInitializer()}
+        default = {'name': None, 'initializer': XavierInitializer()}
         bias_attr = self.kwargs.get('bias_attr', None)
-        if bias_attr is True:
+        if bias_attr is None:
             bias_attr = default
 
         if isinstance(bias_attr, dict):
@@ -166,7 +167,7 @@ class LayerHelper(object):
                 num_flatten_dims = 1
 
         size = list(input_var.shape[num_flatten_dims:])
-        bias_attr = self.bias_attr()
+        bias_attr = self.bias_attr
         if not bias_attr:
             return input_var
 

python/paddle/v2/framework/layers.py

 import paddle.v2.framework.core as core
+import paddle.v2.framework.proto.framework_pb2 as framework_pb2
 from paddle.v2.framework.framework import OpProtoHolder, Variable, Program, \
     Operator
 from paddle.v2.framework.initializer import ConstantInitializer, \
     NormalInitializer
 from paddle.v2.framework.layer_helper import LayerHelper, unique_name
 import re
+import cStringIO
 
 __all__ = [
     'fc', 'data', 'cross_entropy', 'conv2d', 'pool2d', 'embedding', 'concat',
@@ -16,7 +18,7 @@ __all__ = [
 def fc(input,
        size,
        param_attr=None,
-       bias_attr=True,
+       bias_attr=None,
        name=None,
        act=None,
        num_flatten_dims=1,
@@ -125,6 +127,55 @@ def embedding(input,
     return tmp
 
 
+# TODO(qijun): expose H0 and C0
+def dynamic_lstm(input,
+                 size,
+                 data_type='float32',
+                 param_attr=None,
+                 bias_attr=None,
+                 use_peepholes=True,
+                 is_reverse=False,
+                 gate_activation='sigmoid',
+                 cell_activation='tanh',
+                 candidate_activation='tanh',
+                 main_program=None,
+                 startup_program=None):
+    helper = LayerHelper('lstm', **locals())
+    size = size / 4
+    weight = helper.create_parameter(
+        attr=helper.param_attr, shape=[size, 4 * size], dtype=data_type)
+    bias_size = [1, 7 * size]
+    if not use_peepholes:
+        bias_size[1] = 4 * size
+    bias = helper.create_parameter(
+        attr=helper.bias_attr, shape=bias_size, dtype=data_type, suffix='b')
+
+    hidden = helper.create_tmp_variable(data_type)
+    cell = helper.create_tmp_variable(data_type)
+    batch_gate = helper.create_tmp_variable(data_type)
+    batch_cell_pre_act = helper.create_tmp_variable(data_type)
+
+    helper.append_op(
+        type='lstm',
+        inputs={'Input': input,
+                'Weight': weight,
+                'Bias': bias},
+        outputs={
+            'Hidden': hidden,
+            'Cell': cell,
+            'BatchGate': batch_gate,
+            'BatchCellPreAct': batch_cell_pre_act
+        },
+        attrs={
+            'use_peepholes': use_peepholes,
+            'is_reverse': is_reverse,
+            'gate_activation': gate_activation,
+            'cell_activation': cell_activation,
+            'candidate_activation': candidate_activation
+        })
+    return hidden, cell
+
+
 def data(name,
          shape,
          data_type='float32',
@@ -191,6 +242,58 @@ def _convert_(name):
     return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower()
 
 
+def _generate_doc_string_(op_proto):
+    """
+    Generate docstring by OpProto
+    
+    Args:
+        op_proto (framework_pb2.OpProto): a protobuf message typed OpProto
+
+    Returns:
+        str: the document string
+    """
+
+    def _type_to_str_(tp):
+        return framework_pb2.AttrType.Name(tp)
+
+    if not isinstance(op_proto, framework_pb2.OpProto):
+        raise TypeError("OpProto should be `framework_pb2.OpProto`")
+
+    buf = cStringIO.StringIO()
+    buf.write(op_proto.comment)
+    buf.write('\nArgs:\n')
+    for each_input in op_proto.inputs:
+        line_begin = '    {0}: '.format(_convert_(each_input.name))
+        buf.write(line_begin)
+        buf.write(each_input.comment)
+        buf.write('\n')
+        buf.write(' ' * len(line_begin))
+        buf.write('Duplicable: ')
+        buf.write(str(each_input.duplicable))
+        buf.write('  Optional: ')
+        buf.write(str(each_input.dispensable))
+        buf.write('\n')
+
+    for each_attr in op_proto.attrs:
+        buf.write('    ')
+        buf.write(each_attr.name)
+        buf.write(' (')
+        buf.write(_type_to_str_(each_attr.type))
+        buf.write('): ')
+        buf.write(each_attr.comment)
+        buf.write('\n')
+
+    if len(op_proto.outputs) != 0:
+        buf.write('\nReturns:\n')
+        buf.write('    ')
+        for each_opt in op_proto.outputs:
+            if not each_opt.intermediate:
+                break
+        buf.write(each_opt.comment)
+
+    return buf.getvalue()
+
+
 def _create_op_func_(op_type):
     """
     Create an Operator for a Function.
@@ -249,11 +352,6 @@ def _create_op_func_(op_type):
         return dtype
 
     def func(**kwargs):
-        """
-        This function implements the function for the operator. This process
-        involves doing the sanity check (using the function above), reading
-        inputs from protobuf and applying the activations on top.
-        """
         helper = LayerHelper(op_type, **kwargs)
 
         dtype = infer_and_check_data_type(op_proto, **kwargs)
@@ -277,6 +375,7 @@ def _create_op_func_(op_type):
 
     func.__name__ = op_type
     globals()[op_type] = func
+    func.__doc__ = _generate_doc_string_(op_proto)
     global __all__
     __all__.append(op_type)
 

python/paddle/v2/framework/optimizer.py

@@ -35,15 +35,21 @@ class Optimizer(object):
         """
         raise NotImplementedError()
 
-    def _initialize_tensors(self, block):
-        """Create all necessary tensors, that will be shared for all parameter updates.
-
-        Tensors like learning rate should be initialized here.
-
-        Args:
-            block: the block in which the loss variable is present
-        """
-        pass
+    def _create_param_lr(self, param_and_grad):
+        # create learning rate variable for every parameter
+        param = param_and_grad[0]
+        param_lr = param.optimize_attr['learning_rate']
+        param_lr_shape = [1]
+        param_lr_var = self.helper.create_global_variable(
+            name=unique_name("learning_rate"),
+            dtype='float32',
+            shape=param_lr_shape,
+            lod_level=1,
+            persistable=True)
+        param_lr = param_lr * self._learning_rate
+        self.helper.set_variable_initializer(
+            var=param_lr_var, initializer=ConstantInitializer(param_lr))
+        return param_lr_var
 
     def _create_accumulators(self, block, parameters):
         """Create all accumulators needed by the parameters
@@ -161,8 +167,6 @@ class Optimizer(object):
             startup_program=startup_program)
         self._create_accumulators(loss.block,
                                   [p[0] for p in parameters_and_grads])
-        # Create any necessary tensors
-        self._initialize_tensors(loss.block)
 
         optimize_ops = []
         for param_and_grad in parameters_and_grads:
@@ -214,27 +218,16 @@ class SGDOptimizer(Optimizer):
         self.type = "sgd"
         self._learning_rate = learning_rate
 
-    def _initialize_tensors(self, block):
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _append_optimize_op(self, block, param_and_grad):
         assert isinstance(block, framework.Block)
+
         # create the optimize op
         sgd_op = block.append_op(
             type=self.type,
             inputs={
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
-                "LearningRate": self._lr
+                "LearningRate": self._create_param_lr(param_and_grad)
             },
             outputs={"ParamOut": param_and_grad[0]})
 
@@ -259,19 +252,6 @@ class MomentumOptimizer(Optimizer):
         self._momentum = momentum
         self._use_nesterov = bool(use_nesterov)
 
-    def _initialize_tensors(self, block):
-        assert isinstance(block, framework.Block)
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _create_accumulators(self, block, parameters):
         assert isinstance(block, framework.Block)
 
@@ -290,7 +270,7 @@ class MomentumOptimizer(Optimizer):
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
                 "Velocity": velocity_acc,
-                "LearningRate": self._lr
+                "LearningRate": self._create_param_lr(param_and_grad)
             },
             outputs={
                 "ParamOut": param_and_grad[0],
@@ -315,18 +295,6 @@ class AdagradOptimizer(Optimizer):
         self._learning_rate = learning_rate
         self._epsilon = epsilon
 
-    def _initialize_tensors(self, block):
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _create_accumulators(self, block, parameters):
         assert isinstance(block, framework.Block)
 
@@ -346,7 +314,7 @@ class AdagradOptimizer(Optimizer):
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
                 "Moment": moment_acc,
-                "LearningRate": self._lr
+                "LearningRate": self._create_param_lr(param_and_grad)
             },
             outputs={"ParamOut": param_and_grad[0],
                      "MomentOut": moment_acc},
@@ -378,18 +346,6 @@ class AdamOptimizer(Optimizer):
         self._beta2 = beta2
         self._epsilon = epsilon
 
-    def _initialize_tensors(self, block):
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _create_accumulators(self, block, parameters):
         assert isinstance(block, framework.Block)
 
@@ -433,7 +389,7 @@ class AdamOptimizer(Optimizer):
             inputs={
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
-                "LearningRate": self._lr,
+                "LearningRate": self._create_param_lr(param_and_grad),
                 "Moment1": moment1,
                 "Moment2": moment2,
                 "Beta1Pow": self._beta1_pow_acc,
@@ -495,18 +451,6 @@ class AdamaxOptimizer(Optimizer):
         self._beta2 = beta2
         self._epsilon = epsilon
 
-    def _initialize_tensors(self, block):
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _create_accumulators(self, block, parameters):
         # Create beta1 power accumulator tensor
         beta_shape = [1]
@@ -536,7 +480,7 @@ class AdamaxOptimizer(Optimizer):
             inputs={
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
-                "LearningRate": self._lr,
+                "LearningRate": self._create_param_lr(param_and_grad),
                 "Moment": moment,
                 "InfNorm": inf_norm,
                 "Beta1Pow": self._beta1_pow_acc

python/paddle/v2/framework/tests/test_create_op_doc_string.py

+import unittest
+import paddle.v2.framework.layers as layers
+
+
+class TestDocString(unittest.TestCase):
+    def test_layer_doc_string(self):
+        print layers.dropout.__doc__
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/framework/tests/test_understand_sentiment_dynamic_lstm.py

+import paddle.v2 as paddle
+import paddle.v2.framework.layers as layers
+import paddle.v2.framework.nets as nets
+import paddle.v2.framework.core as core
+import paddle.v2.framework.optimizer as optimizer
+
+from paddle.v2.framework.framework import Program, g_main_program, g_startup_program
+from paddle.v2.framework.executor import Executor
+
+import numpy as np
+
+
+def stacked_lstm_net(input_dim,
+                     class_dim=2,
+                     emb_dim=128,
+                     hid_dim=512,
+                     stacked_num=3):
+    assert stacked_num % 2 == 1
+    data = layers.data(name="words", shape=[1], data_type="int64")
+    label = layers.data(name="label", shape=[1], data_type="int64")
+
+    emb = layers.embedding(input=data, size=[input_dim, emb_dim])
+    # add bias attr
+
+    # TODO(qijun) linear act
+    fc1 = layers.fc(input=emb, size=hid_dim)
+    lstm1, cell1 = layers.dynamic_lstm(input=fc1, size=hid_dim)
+
+    inputs = [fc1, lstm1]
+
+    for i in range(2, stacked_num + 1):
+        fc = layers.fc(input=inputs, size=hid_dim)
+        lstm, cell = layers.dynamic_lstm(
+            input=fc, size=hid_dim, is_reverse=(i % 2) == 0)
+        inputs = [fc, lstm]
+
+    fc_last = layers.sequence_pool(input=inputs[0], pool_type='max')
+    lstm_last = layers.sequence_pool(input=inputs[1], pool_type='max')
+
+    prediction = layers.fc(input=[fc_last, lstm_last],
+                           size=class_dim,
+                           act='softmax')
+    cost = layers.cross_entropy(input=prediction, label=label)
+    avg_cost = layers.mean(x=cost)
+    adam_optimizer = optimizer.AdamOptimizer(learning_rate=0.002)
+    opts = adam_optimizer.minimize(avg_cost)
+    acc = layers.accuracy(input=prediction, label=label)
+    return avg_cost, acc
+
+
+def to_lodtensor(data, place):
+    seq_lens = [len(seq) for seq in data]
+    cur_len = 0
+    lod = [cur_len]
+    for l in seq_lens:
+        cur_len += l
+        lod.append(cur_len)
+    flattened_data = np.concatenate(data, axis=0).astype("int64")
+    flattened_data = flattened_data.reshape([len(flattened_data), 1])
+    res = core.LoDTensor()
+    res.set(flattened_data, place)
+    res.set_lod([lod])
+    return res
+
+
+def main():
+    BATCH_SIZE = 100
+    PASS_NUM = 5
+
+    word_dict = paddle.dataset.imdb.word_dict()
+    print "load word dict successfully"
+    dict_dim = len(word_dict)
+    class_dim = 2
+
+    cost, acc = stacked_lstm_net(input_dim=dict_dim, class_dim=class_dim)
+
+    train_data = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.imdb.train(word_dict), buf_size=1000),
+        batch_size=BATCH_SIZE)
+    place = core.CPUPlace()
+    exe = Executor(place)
+
+    exe.run(g_startup_program)
+
+    for pass_id in xrange(PASS_NUM):
+        for data in train_data():
+            tensor_words = to_lodtensor(map(lambda x: x[0], data), place)
+
+            label = np.array(map(lambda x: x[1], data)).astype("int64")
+            label = label.reshape([BATCH_SIZE, 1])
+
+            tensor_label = core.LoDTensor()
+            tensor_label.set(label, place)
+
+            outs = exe.run(g_main_program,
+                           feed={"words": tensor_words,
+                                 "label": tensor_label},
+                           fetch_list=[cost, acc])
+            cost_val = np.array(outs[0])
+            acc_val = np.array(outs[1])
+
+            print("cost=" + str(cost_val) + " acc=" + str(acc_val))
+            if cost_val < 1.0 and acc_val > 0.7:
+                exit(0)
+    exit(1)
+
+
+if __name__ == '__main__':
+    main()