Merge branch 'develop' into core_add_inference_unittest

cmake/external/boost.cmake

@@ -15,9 +15,9 @@
 include(ExternalProject)
 
 set(BOOST_PROJECT       "extern_boost")
-set(BOOST_VER           "1.66.0")
-set(BOOST_TAR           "boost_1_66_0")
-set(BOOST_URL           "https://dl.bintray.com/boostorg/release/${BOOST_VER}/source/${BOOST_TAR}.tar.gz")
+set(BOOST_VER           "1.41.0")
+set(BOOST_TAR           "boost_1_41_0")
+set(BOOST_URL           "http://sourceforge.net/projects/boost/files/boost/${BOOST_VER}/${BOOST_TAR}.tar.gz")
 set(BOOST_SOURCES_DIR ${THIRD_PARTY_PATH}/boost)
 set(BOOST_DOWNLOAD_DIR  "${BOOST_SOURCES_DIR}/src/${BOOST_PROJECT}")
 set(BOOST_INCLUDE_DIR "${BOOST_DOWNLOAD_DIR}/${BOOST_TAR}" CACHE PATH "boost include directory." FORCE)

doc/api/v2/fluid/layers.rst

@@ -18,6 +18,11 @@ dynamic_lstm
 ..  autofunction:: paddle.v2.fluid.layers.dynamic_lstm
     :noindex:
 
+dynamic_lstmp
+-------------
+..  autofunction:: paddle.v2.fluid.layers.dynamic_lstmp
+    :noindex:
+
 dynamic_gru
 -----------
 ..  autofunction:: paddle.v2.fluid.layers.dynamic_gru

doc/design/support_new_device.md

@@ -2,9 +2,9 @@
 
 ## Background
 
-Deep learning has a high demand for computing resources. New high-performance devices and computing libraries are appearing very frequently. Deep learning frameworks have to integrate these high-performance devices and computing libraries flexibly and efficiently.
+Deep learning has a high demand for computing resources. New high-performance devices and computing libraries are appearing very frequently. Deep learning frameworks have to integrate these high-performance devices and computing libraries in a flexible and efficient manner.
 
-On one hand, hardware and computing libraries usually do not have a one-to-one correspondence. For example,Intel CPUs support Eigen and MKL computing libraries while Nvidia GPUs support Eigen and cuDNN computing libraries. We have to implement operator specific kernels for each computing library.
+On one hand, hardware and computing libraries usually do not have a one-to-one correspondence. For example, Intel CPUs support Eigen and MKL computing libraries while Nvidia GPUs support Eigen and cuDNN computing libraries. We have to implement operator specific kernels for each computing library.
 
 On the other hand, users usually do not want to care about the low-level hardware and computing libraries when writing a neural network configuration. In Fluid, `Layer` is exposed in `Python`, and `Operator` is exposed in `C++`. Both `Layer` and `Operator` are hardware independent.
 
@@ -17,7 +17,7 @@ For a general overview of fluid, please refer to the [overview doc](https://gith
 
 There are mainly three parts that we have to consider while integrating a new device/library:
 
-- Place and DeviceContext: indicates the device id and manages hardware resources
+- Place and DeviceContext: indicate the device id and manage hardware resources
 
 - Memory and Tensor: malloc/free data on certain device
 
@@ -25,10 +25,10 @@ There are mainly three parts that we have to consider while integrating a new de
 
 ### Place and DeviceContext
 
-Please remind that device and computing library are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.
+Please note that device and computing library are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.
 
 #### Place
-Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent the device memory where data is located. If we add another device, we have to add corresponding `DevicePlace`.
+Fluid uses class [Place](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/platform/place.h#L55) to represent the device memory where data is located. If we add another device, we have to add the corresponding `DevicePlace`.
 
 ```
         |   CPUPlace
@@ -144,7 +144,7 @@ class Tensor {
 };
 ```
 
-`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configure its shape, and then call `mutuable_data` to allocate the actual memory.
+`Placeholder` is used to delay memory allocation; that is, we can first define a tensor, using `Resize` to configurate its shape, and then call `mutuable_data` to allocate the actual memory.
 
 ```cpp
 paddle::framework::Tensor t;
@@ -163,7 +163,7 @@ Fluid implements computing units based on different DeviceContexts. Some computi
 
 Let's take [MaxOutFunctor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/math/maxouting.h#L27) as an example:
 
-The interface is defined in header file.
+The interface is defined in the header file.
 
 ```
 template <typename DeviceContext, typename T>
@@ -174,7 +174,7 @@ class MaxOutFunctor {
 };
 ```
 
-CPU implemention is in .cc file
+CPU implementation is in .cc file
 
 ```
 template <typename T>
@@ -188,7 +188,7 @@ class MaxOutFunctor<platform::CPUDeviceContext, T> {
 };
 ```
 
-CUDA implemention is in .cu file
+CUDA implementation is in .cu file
 
 ```
 template <typename T>
@@ -203,9 +203,9 @@ class MaxOutFunctor<platform::CUDADeviceContext, T> {
 ```
 
 
-We get computing handle from a concrete DeviceContext, and make compution on tensors.
+We first obtain the computing handle from a concrete DeviceContext and then compute on tensors.
 
-The implemention of `OpKernel` is similar to math functors, the extra thing we need to do is to register the OpKernel in a global map.
+The implementation of `OpKernel` is similar to math functors, the extra thing we need to do is to register the OpKernel in a global map.
 
 Fluid provides different register interfaces in op_registry.h
 
@@ -231,7 +231,7 @@ REGISTER_OP_CUDA_KERNEL(
 
 ## Advanced topics: How to switch between different Device/Library
 
-Generally, we will impelement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not sutibale on a specific Device. For example, crf operator can only run on CPU, whereas most other operators can run at GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
+Generally, we will implement OpKernel for all Device/Library of an Operator. We can easily train a Convolutional Neural Network in GPU. However, some OpKernel is not suitable on a specific Device. For example, crf operator can only run on CPU, whereas most other operators can run on GPU. To achieve high performance in such circumstance, we have to switch between different Device/Library.
 
 
 For more details, please refer to following docs:

paddle/framework/CMakeLists.txt

@@ -26,7 +26,7 @@ nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
 
 cc_test(variable_test SRCS variable_test.cc)
 
-cc_library(threadpool SRCS threadpool.cc)
+cc_library(threadpool SRCS threadpool.cc DEPS enforce)
 cc_test(threadpool_test SRCS threadpool_test.cc DEPS threadpool)
 
 cc_library(scope SRCS scope.cc DEPS glog threadpool)
@@ -98,3 +98,5 @@ if(NOT WITH_C_API AND WITH_FLUID)
   install(FILES ${CMAKE_CURRENT_BINARY_DIR}/framework.pb.h DESTINATION include/paddle/framework)
   install(FILES details/cow_ptr.h details/op_registry.h DESTINATION include/paddle/framework/details)
 endif()
+
+cc_test(channel_test SRCS channel_test.cc)

paddle/framework/channel.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 <stddef.h>  // for size_t
+
+namespace paddle {
+namespace framework {
+
+// Channel is the abstract class of buffered and un-buffered channels.
+template <typename T>
+class Channel {
+ public:
+  virtual void Send(T*) = 0;
+  virtual void Receive(T*) = 0;
+  virtual size_t Cap() = 0;
+
+  // Don't delete channels; instead, call Channel::Close.
+ protected:
+  virtual ~Channel() {}
+};
+
+// Forward declaration of channel implementations.
+namespace details {
+template <typename T>
+class Buffered;
+template <typename T>
+class UnBuffered;
+}  // namespace details
+
+template <typename T>
+Channel<T>* MakeChannel(size_t buffer_size) {
+  if (buffer_size > 0) {
+    return new details::Buffered<T>(buffer_size);
+  }
+  return new details::UnBuffered<T>();
+}
+
+template <typename T>
+void CloseChannel(Channel<T>* ch) {
+  if (ch->Cap() > 0) {
+    delete dynamic_cast<details::Buffered<T>*>(ch);
+  } else {
+    delete dynamic_cast<details::UnBuffered<T>*>(ch);
+  }
+}
+
+}  // namespace framework
+}  // namespace paddle
+
+#include "paddle/framework/details/buffered_channel.h"
+#include "paddle/framework/details/unbuffered_channel.h"

paddle/framework/channel_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/framework/channel.h"
+
+#include "gtest/gtest.h"
+
+TEST(Channel, MakeAndClose) {
+  using paddle::framework::Channel;
+  using paddle::framework::MakeChannel;
+  using paddle::framework::CloseChannel;
+
+  Channel<int>* ch = MakeChannel<int>(10);
+  CloseChannel(ch);
+}

paddle/framework/data_type.h

@@ -79,5 +79,33 @@ inline void VisitDataType(proto::DataType type, Visitor visitor) {
   }
 }
 
+inline std::string DataTypeToString(const proto::DataType type) {
+  using namespace paddle::framework::proto;
+  switch (type) {
+    case DataType::FP16:
+      return "float16";
+    case DataType::FP32:
+      return "float32";
+    case DataType::FP64:
+      return "float64";
+    case DataType::INT16:
+      return "int16";
+    case DataType::INT32:
+      return "int32";
+    case DataType::INT64:
+      return "int64";
+    case DataType::BOOL:
+      return "bool";
+    default:
+      PADDLE_THROW("Not support type %d", type);
+  }
+}
+
+inline std::ostream& operator<<(std::ostream& out,
+                                const proto::DataType& type) {
+  out << DataTypeToString(type);
+  return out;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/details/buffered_channel.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 <condition_variable>
+#include <deque>
+#include <mutex>
+
+#include "paddle/framework/channel.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+template <typename T>
+class Buffered : public paddle::framework::Channel<T> {
+  friend Channel<T>* paddle::framework::MakeChannel<T>(size_t);
+  friend void paddle::framework::CloseChannel<T>(Channel<T>*);
+
+ public:
+  virtual void Send(T*);
+  virtual void Receive(T*);
+  virtual size_t Cap() { return cap_; }
+
+ private:
+  size_t cap_;
+  std::mutex mu_;
+  std::condition_variable empty_cond_var_;
+  std::condition_variable full_cond_var_;
+  std::deque<T> channel_;
+
+  Buffered(size_t cap) : cap_(cap) {}
+  virtual ~Buffered();
+
+  void NotifyAllSenders(std::unique_lock<std::mutex>*);
+};
+
+template <typename T>
+void Buffered<T>::Send(T* item) {
+  std::unique_lock<std::mutex> lock(mu_);
+  full_cond_var_.wait(lock, [this]() { return channel_.size() < cap_; });
+  channel_.push_back(std::move(*item));
+  lock.unlock();
+  empty_cond_var_.notify_one();
+}
+
+template <typename T>
+void Buffered<T>::Receive(T* item) {
+  std::unique_lock<std::mutex> lock(mu_);
+  empty_cond_var_.wait(lock, [this]() { return !channel_.empty(); });
+  *item = std::move(channel_.front());
+  channel_.pop_front();
+  NotifyAllSenders(&lock);
+}
+
+template <typename T>
+Buffered<T>::~Buffered() {
+  std::unique_lock<std::mutex> lock(mu_);
+  channel_.clear();
+  NotifyAllSenders(&lock);
+}
+
+template <typename T>
+void Buffered<T>::NotifyAllSenders(std::unique_lock<std::mutex>* lock) {
+  lock->unlock();
+  full_cond_var_.notify_one();
+}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/framework/details/unbuffered_channel.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 <condition_variable>
+#include <deque>
+#include <mutex>
+
+#include "paddle/framework/channel.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+template <typename T>
+class UnBuffered : public paddle::framework::Channel<T> {
+  friend Channel<T>* paddle::framework::MakeChannel<T>(size_t);
+  friend void paddle::framework::CloseChannel<T>(Channel<T>*);
+
+ public:
+  virtual void Send(T*);
+  virtual void Receive(T*);
+  virtual size_t Cap() { return 0; }
+
+ private:
+  UnBuffered() {}
+  virtual ~UnBuffered();
+};
+
+template <typename T>
+void UnBuffered<T>::Send(T* channel_element) {}
+
+template <typename T>
+void UnBuffered<T>::Receive(T*) {}
+
+template <typename T>
+UnBuffered<T>::~UnBuffered() {}
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/framework/executor.cc

@@ -170,8 +170,6 @@ static bool has_feed_operators(
           feed_targets.find(feed_target_name) != feed_targets.end(),
           "Feed operator output name '%s' cannot be found in 'feed_targets'",
           feed_target_name);
-    } else {
-      break;
     }
   }
 
@@ -270,8 +268,6 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
       int idx = boost::get<int>(op->GetAttr("col"));
       SetFeedVariable(scope, *feed_targets[feed_target_name], feed_holder_name,
                       idx);
-    } else {
-      break;
     }
   }
 

paddle/framework/op_kernel_type_test.cc

@@ -26,9 +26,9 @@ TEST(OpKernelType, ToString) {
   OpKernelType op_kernel_type(DataType::FP32, CPUPlace(), DataLayout::kNCHW,
                               LibraryType::kCUDNN);
 
-  ASSERT_EQ(
-      paddle::framework::KernelTypeToString(op_kernel_type),
-      "data_type[5]:data_layout[NCHW]:place[CPUPlace]:library_type[CUDNN]");
+  ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type),
+            "data_type[float32]:data_layout[NCHW]:place[CPUPlace]:library_type["
+            "CUDNN]");
 }
 
 TEST(OpKernelType, Hash) {

paddle/framework/program_desc.cc

@@ -18,6 +18,9 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+const std::string kFeedOpType = "feed";
+const std::string kFetchOpType = "fetch";
+
 BlockDesc *ProgramDesc::AppendBlock(const BlockDesc &parent) {
   auto *b = desc_.add_blocks();
   b->set_parent_idx(parent.ID());
@@ -64,5 +67,27 @@ ProgramDesc::ProgramDesc(const std::string &binary_str) {
   }
 }
 
+const std::vector<std::string> ProgramDesc::GetFeedVarNames() {
+  BlockDesc *global_block = blocks_[0].get();
+  std::vector<std::string> feed_var_names;
+  for (auto *op : global_block->AllOps()) {
+    if (op->Type() == "feed") {
+      feed_var_names.insert(feed_var_names.begin(), op->Output("Out")[0]);
+    }
+  }
+  return feed_var_names;
+}
+
+const std::vector<std::string> ProgramDesc::GetFetchVarNames() {
+  BlockDesc *global_block = blocks_[0].get();
+  std::vector<std::string> fetch_var_names;
+  for (auto *op : global_block->AllOps()) {
+    if (op->Type() == "fetch") {
+      fetch_var_names.push_back(op->Input("X")[0]);
+    }
+  }
+  return fetch_var_names;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/program_desc.h

@@ -45,6 +45,10 @@ class ProgramDesc {
 
   proto::ProgramDesc *Proto();
 
+  const std::vector<std::string> GetFeedVarNames();
+
+  const std::vector<std::string> GetFetchVarNames();
+
  private:
   proto::ProgramDesc desc_;
 

paddle/framework/prune.cc

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <algorithm>
 #include <set>
 #include <string>
+#include <unordered_map>
 #include <vector>
 
 #include <glog/logging.h>
@@ -102,6 +103,32 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
       *op_field->Add() = input.blocks(block_id).ops(i);
     }
   }
+
+  // remove the VarDescs in BlockDesc that are not referenced in
+  // the pruned OpDescs
+  std::unordered_map<std::string, proto::VarDesc> var_map;
+  auto* var_field = output->mutable_blocks(block_id)->mutable_vars();
+  for (const auto& var : *var_field) {
+    var_map[var.name()] = var;
+  }
+
+  var_field->Clear();
+  for (const auto& op : *op_field) {
+    // add VarDescs of all input arguments for each OpDesc
+    auto& input_field = op.inputs();
+    for (auto& input_var : input_field) {
+      for (auto& arg : input_var.arguments()) {
+        *var_field->Add() = var_map[arg];
+      }
+    }
+    // add VarDescs of all output arguments for each OpDesc
+    auto& output_field = op.outputs();
+    for (auto& output_var : output_field) {
+      for (auto& arg : output_var.arguments()) {
+        *var_field->Add() = var_map[arg];
+      }
+    }
+  }
 }
 
 // TODO(fengjiayi): Prune() could be inplaced to avoid unnecessary copies

paddle/framework/threadpool.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
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
 
-    http://www.apache.org/licenses/LICENSE-2.0
+   http://www.apache.org/licenses/LICENSE-2.0
 
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License. */
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License. */
 
 #include "paddle/framework/threadpool.h"
 
+#include "paddle/platform/enforce.h"
+
 namespace paddle {
 namespace framework {
 
-std::unique_ptr<ThreadPool> ThreadPool::threadpool(nullptr);
-std::once_flag ThreadPool::init_flag;
+std::unique_ptr<ThreadPool> ThreadPool::threadpool_(nullptr);
+std::once_flag ThreadPool::init_flag_;
+
+ThreadPool* ThreadPool::GetInstance() {
+  std::call_once(init_flag_, &ThreadPool::Init);
+  return threadpool_.get();
+}
+
+void ThreadPool::Init() {
+  if (threadpool_.get() == nullptr) {
+    // TODO(Yancey1989): specify the max threads number
+    int num_threads = std::thread::hardware_concurrency();
+    PADDLE_ENFORCE_GT(num_threads, 0);
+    threadpool_.reset(new ThreadPool(num_threads));
+  }
+}
+
+ThreadPool::ThreadPool(int num_threads)
+    : total_threads_(num_threads), idle_threads_(num_threads), running_(true) {
+  threads_.resize(num_threads);
+  for (auto& thread : threads_) {
+    // TODO(Yancey1989): binding the thread on the specify CPU number
+    thread.reset(new std::thread(std::bind(&ThreadPool::TaskLoop, this)));
+  }
+}
+
+ThreadPool::~ThreadPool() {
+  {
+    // notify all threads to stop running
+    running_ = false;
+    scheduled_.notify_all();
+  }
+
+  for (auto& t : threads_) {
+    t->join();
+    t.reset(nullptr);
+  }
+}
+
+void ThreadPool::Wait() {
+  std::unique_lock<std::mutex> lock(mutex_);
+  completed_.wait(lock, [=] { return Done() == true; });
+}
+
+void ThreadPool::TaskLoop() {
+  while (running_) {
+    std::unique_lock<std::mutex> lock(mutex_);
+    scheduled_.wait(lock, [=] { return !tasks_.empty() || !running_; });
+
+    if (!running_) {
+      break;
+    }
+    // pop a task from the task queue
+    auto task = std::move(tasks_.front());
+    tasks_.pop();
+
+    --idle_threads_;
+    lock.unlock();
+
+    // run the task
+    task();
+
+    {
+      std::unique_lock<std::mutex> lock(mutex_);
+      ++idle_threads_;
+      if (Done()) {
+        completed_.notify_all();
+      }
+    }
+  }
+}
 
 }  // namespace framework
 }  // namespace paddle

paddle/framework/threadpool.h

@@ -20,52 +20,36 @@ limitations under the License. */
 #include <mutex>
 #include <queue>
 #include <thread>
+#include <vector>
 
-#include "paddle/platform/enforce.h"
+#include "paddle/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
 
 namespace paddle {
 namespace framework {
 
+// ThreadPool maintains a queue of tasks, and runs them using a fixed
+// number of threads.
 class ThreadPool {
  public:
   typedef std::packaged_task<void()> Task;
 
-  /**
-   * @brief   Get a instance of threadpool, the thread number will
-   *          be specified as the number of hardware thread contexts
-   */
-  static ThreadPool* GetInstance() {
-    std::call_once(init_flag, &ThreadPool::Init);
-    return threadpool.get();
-  }
+  // Returns the singleton of ThreadPool.
+  static ThreadPool* GetInstance();
 
-  ~ThreadPool() {
-    {
-      // notify all threads to stop running
-      running_ = false;
-      scheduled_.notify_all();
-    }
-
-    for (auto& t : threads_) {
-      t->join();
-      t.reset(nullptr);
-    }
-  }
+  ~ThreadPool();
 
-  int GetNumThreads() const { return num_threads_; }
+  // Returns the number of threads created by the constructor.
+  size_t Threads() const { return total_threads_; }
 
-  int GetAvailable() {
+  // Returns the number of currently idle threads.
+  size_t IdleThreads() {
     std::unique_lock<std::mutex> lock(mutex_);
-    return available_;
+    return idle_threads_;
   }
 
-  /**
-   * @brief   Push a function to the queue, and will be scheduled and
-   *          executed if a thread is available.
-   * @param[in] Task, will be pushed to the task queue.
-   * @return    std::future<void>, we could wait for the task finished by
-   *            f.wait().
-   */
+  // Run pushes a function to the task queue and returns a std::future
+  // object.  To wait for the completion of the task, call
+  // std::future::wait().
   template <typename Callback>
   std::future<void> Run(Callback fn) {
     std::unique_lock<std::mutex> lock(mutex_);
@@ -77,84 +61,40 @@ class ThreadPool {
     return f;
   }
 
-  /**
-   * @brief   Wait until all the tasks are completed.
-   */
-  void Wait() {
-    std::unique_lock<std::mutex> lock(mutex_);
-    completed_.wait(lock, [=] { return Done() == true; });
-  }
+  // Wait until all the tasks are completed.
+  void Wait();
 
  private:
   DISABLE_COPY_AND_ASSIGN(ThreadPool);
 
-  explicit ThreadPool(int num_threads)
-      : num_threads_(num_threads), available_(num_threads), running_(true) {
-    threads_.resize(num_threads);
-    for (auto& thread : threads_) {
-      // TODO(Yancey1989): binding the thread on the specify CPU number
-      thread.reset(new std::thread(std::bind(&ThreadPool::TaskLoop, this)));
-    }
-  }
+  explicit ThreadPool(int num_threads);
 
-  /**
-   * @brief   If the task queue is empty and avaialbe
-   *          is equal to the number of threads, means that
-   *          all tasks are completed.
-   *
-   *          Note: this function is not thread-safe.
-   *
-   * @return true if all tasks are completed.
-   */
-  bool Done() { return tasks_.empty() && available_ == num_threads_; }
-
-  void TaskLoop() {
-    while (running_) {
-      std::unique_lock<std::mutex> lock(mutex_);
-      scheduled_.wait(lock, [=] { return !tasks_.empty() || !running_; });
-
-      if (!running_) {
-        break;
-      }
-      // pop a task from the task queue
-      auto task = std::move(tasks_.front());
-      tasks_.pop();
-
-      --available_;
-      lock.unlock();
-
-      // run the task
-      task();
-
-      {
-        std::unique_lock<std::mutex> lock(mutex_);
-        ++available_;
-        if (Done()) {
-          completed_.notify_all();
-        }
-      }
-    }
-  }
+  // If the task queue is empty and avaialbe is equal to the number of
+  // threads, means that all tasks are completed.  Note: this function
+  // is not thread-safe.  Returns true if all tasks are completed.
+  // Note: don't delete the data member total_threads_ and use
+  // threads_.size() instead; because you'd need to lock the mutex
+  // before accessing threads_.
+  bool Done() { return tasks_.empty() && idle_threads_ == total_threads_; }
 
-  static void Init() {
-    if (threadpool.get() == nullptr) {
-      // TODO(Yancey1989): specify the max threads number
-      int num_threads = std::thread::hardware_concurrency();
-      PADDLE_ENFORCE_GT(num_threads, 0);
-      threadpool.reset(new ThreadPool(num_threads));
-    }
-  }
+  // The constructor starts threads to run TaskLoop, which retrieves
+  // and runs tasks from the queue.
+  void TaskLoop();
+
+  // Init is called by GetInstance.
+  static void Init();
 
  private:
-  static std::unique_ptr<ThreadPool> threadpool;
-  static std::once_flag init_flag;
+  static std::unique_ptr<ThreadPool> threadpool_;
+  static std::once_flag init_flag_;
 
-  int num_threads_;
-  int available_;
-  bool running_;
-  std::queue<Task> tasks_;
   std::vector<std::unique_ptr<std::thread>> threads_;
+  const size_t total_threads_;
+  size_t idle_threads_;
+
+  std::queue<Task> tasks_;
   std::mutex mutex_;
+  bool running_;
   std::condition_variable scheduled_;
   std::condition_variable completed_;
 };

paddle/framework/threadpool_test.cc

@@ -22,11 +22,7 @@ namespace framework = paddle::framework;
 void do_sum(framework::ThreadPool* pool, std::atomic<int>& sum, int cnt) {
   std::vector<std::future<void>> fs;
   for (int i = 0; i < cnt; ++i) {
-    auto f = pool->Run([&sum]() { sum.fetch_add(1); });
-    fs.push_back(std::move(f));
-  }
-  for (auto& f : fs) {
-    f.wait();
+    fs.push_back(framework::Async([&sum]() { sum.fetch_add(1); }));
   }
 }
 

paddle/inference/CMakeLists.txt

 set(FLUID_CORE_MODULES proto_desc paddle_memory executor prune init)
 
 cc_library(paddle_fluid_api
-    SRCS inference.cc
+    SRCS io.cc
     DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
 
 # Merge all modules into a single static library
 cc_library(paddle_fluid DEPS paddle_fluid_api ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
 
 # Create shared library
-add_library(paddle_fluid_shared SHARED inference.cc)
+add_library(paddle_fluid_shared SHARED io.cc)
 
 target_circle_link_libraries(paddle_fluid_shared
   ARCHIVE_START
@@ -20,7 +20,7 @@ SET_TARGET_PROPERTIES(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
 
 # install library & headers
 if(NOT WITH_C_API AND WITH_FLUID)
-  install(FILES inference.h DESTINATION include/paddle/inference)
+  install(FILES io.h DESTINATION include/paddle/inference)
   install(TARGETS paddle_fluid_shared DESTINATION lib)
 endif()
 

paddle/inference/inference.cc → paddle/inference/io.cc

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* Copyright (c) 2018 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.
@@ -12,51 +12,22 @@ 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 "inference.h"
+#include "paddle/inference/io.h"
 #include <fstream>
 
 namespace paddle {
+namespace inference {
 
-framework::ProgramDesc* InferenceEngine::LoadInferenceModel(
-    framework::Executor& exe,
-    framework::Scope* scope,
-    const std::string& dirname) {
-  std::string model_filename = dirname + "/__model__";
-  LOG(INFO) << "loading model from " << model_filename;
-  std::ifstream inputfs(model_filename, std::ios::in | std::ios::binary);
-  std::string program_desc_str;
-  inputfs.seekg(0, std::ios::end);
-  program_desc_str.resize(inputfs.tellg());
-  inputfs.seekg(0, std::ios::beg);
-  LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
-  inputfs.read(&program_desc_str[0], program_desc_str.size());
-  inputfs.close();
-
-  program_ = new framework::ProgramDesc(program_desc_str);
-  GenerateLoadProgram(dirname);
-  exe.Run(*load_program_, scope, 0, true, true);
-
-  framework::BlockDesc* global_block = program_->MutableBlock(0);
-  feed_var_names_.clear();
-  fetch_var_names_.clear();
-  for (auto* op : global_block->AllOps()) {
-    if (op->Type() == "feed") {
-      feed_var_names_.insert(feed_var_names_.begin(), op->Output("Out")[0]);
-    } else if (op->Type() == "fetch") {
-      fetch_var_names_.push_back(op->Input("X")[0]);
-    }
-  }
-
-  return program_;
-}
+const std::string kFeedOpType = "feed";
 
-bool InferenceEngine::IsParameter(const framework::VarDesc* var) {
+bool IsParameter(const framework::VarDesc* var,
+                 const framework::ProgramDesc* main_program) {
   if (var->Persistable()) {
     // There are many unreachable variables in the program
-    for (size_t i = 0; i < program_->Size(); ++i) {
-      const framework::BlockDesc& block = program_->Block(i);
+    for (size_t i = 0; i < main_program->Size(); ++i) {
+      const framework::BlockDesc& block = main_program->Block(i);
       for (auto* op : block.AllOps()) {
-        if (op->Type() == "feed") {
+        if (op->Type() == kFeedOpType) {
           continue;
         }
         for (auto input_argument_name : op->InputArgumentNames()) {
@@ -70,13 +41,16 @@ bool InferenceEngine::IsParameter(const framework::VarDesc* var) {
   return false;
 }
 
-void InferenceEngine::GenerateLoadProgram(const std::string& dirname) {
-  framework::BlockDesc* global_block = program_->MutableBlock(0);
+void LoadPersistables(framework::Executor& executor,
+                      framework::Scope& scope,
+                      const std::string& dirname,
+                      framework::ProgramDesc* main_program) {
+  framework::BlockDesc* global_block = main_program->MutableBlock(0);
 
-  load_program_ = new framework::ProgramDesc();
-  framework::BlockDesc* load_block = load_program_->MutableBlock(0);
+  framework::ProgramDesc* load_program = new framework::ProgramDesc();
+  framework::BlockDesc* load_block = load_program->MutableBlock(0);
   for (auto* var : global_block->AllVars()) {
-    if (IsParameter(var)) {
+    if (IsParameter(var, main_program)) {
       VLOG(3) << "parameter's name: " << var->Name();
 
       framework::VarDesc* new_var = load_block->Var(var->Name());
@@ -94,5 +68,30 @@ void InferenceEngine::GenerateLoadProgram(const std::string& dirname) {
       op->CheckAttrs();
     }
   }
+  executor.Run(*load_program, &scope, 0, true, true);
+  delete load_program;
 }
+
+framework::ProgramDesc* Load(framework::Executor& executor,
+                             framework::Scope& scope,
+                             const std::string& dirname) {
+  std::string model_filename = dirname + "/__model__";
+  LOG(INFO) << "loading model from " << model_filename;
+  std::ifstream inputfs(model_filename, std::ios::in | std::ios::binary);
+  std::string program_desc_str;
+  inputfs.seekg(0, std::ios::end);
+  program_desc_str.resize(inputfs.tellg());
+  inputfs.seekg(0, std::ios::beg);
+  LOG(INFO) << "program_desc_str's size: " << program_desc_str.size();
+  inputfs.read(&program_desc_str[0], program_desc_str.size());
+  inputfs.close();
+
+  framework::ProgramDesc* main_program =
+      new framework::ProgramDesc(program_desc_str);
+
+  LoadPersistables(executor, scope, dirname, main_program);
+  return main_program;
+}
+
+}  // namespace inference
 }  // namespace paddle

paddle/inference/inference.h → paddle/inference/io.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* Copyright (c) 2018 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.
@@ -14,43 +14,28 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
+#include <vector>
 #include "paddle/framework/block_desc.h"
 #include "paddle/framework/executor.h"
-#include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/program_desc.h"
 #include "paddle/framework/scope.h"
+#include "paddle/framework/var_desc.h"
 
 namespace paddle {
+namespace inference {
 
-class InferenceEngine {
-public:
-  InferenceEngine() : program_(nullptr), load_program_(nullptr) {}
-  ~InferenceEngine() {
-    delete program_;
-    delete load_program_;
-  }
-
-  framework::ProgramDesc* LoadInferenceModel(framework::Executor& exe,
-                                             framework::Scope* scope,
-                                             const std::string& dirname);
-
-  const std::vector<std::string>& GetFeedVarNames() const {
-    return feed_var_names_;
-  }
-
-  const std::vector<std::string>& GetFetchVarNames() const {
-    return fetch_var_names_;
-  }
-
-private:
-  bool IsParameter(const framework::VarDesc* var);
-  void GenerateLoadProgram(const std::string& dirname);
-
-private:
-  framework::ProgramDesc* program_;
-  framework::ProgramDesc* load_program_;
-  std::vector<std::string> feed_var_names_;
-  std::vector<std::string> fetch_var_names_;
-};
+bool IsParameter(const framework::VarDesc* var,
+                 const framework::ProgramDesc* main_program);
 
+void LoadPersistables(framework::Executor& executor,
+                      framework::Scope& scope,
+                      const std::string& dirname,
+                      framework::ProgramDesc* main_program);
+
+framework::ProgramDesc* Load(framework::Executor& executor,
+                             framework::Scope& scope,
+                             const std::string& dirname);
+
+}  // namespace inference
 }  // namespace paddle

paddle/inference/tests/book/test_inference_recognize_digits.cc

@@ -16,8 +16,8 @@ limitations under the License. */
 #include <time.h>
 #include <sstream>
 #include "gflags/gflags.h"
-#include "paddle/framework/init.h"
-#include "paddle/inference/inference.h"
+#include "paddle/framework/lod_tensor.h"
+#include "paddle/inference/io.h"
 
 DEFINE_string(dirname, "", "Directory of the inference model.");
 
@@ -31,14 +31,13 @@ void TestInference(const std::string& dirname,
   auto* scope = new paddle::framework::Scope();
 
   // 2. Initialize the inference_program and load all parameters from file
-  paddle::InferenceEngine* engine = new paddle::InferenceEngine();
-  paddle::framework::ProgramDesc* inference_program =
-      engine->LoadInferenceModel(executor, scope, dirname);
+  auto* inference_program = paddle::inference::Load(executor, *scope, dirname);
 
   // 3. Get the feed_var_names and fetch_var_names
-  const std::vector<std::string>& feed_target_names = engine->GetFeedVarNames();
+  const std::vector<std::string>& feed_target_names =
+      inference_program->GetFeedVarNames();
   const std::vector<std::string>& fetch_target_names =
-      engine->GetFetchVarNames();
+      inference_program->GetFetchVarNames();
 
   // 4. Prepare inputs
   std::map<std::string, const paddle::framework::LoDTensor*> feed_targets;
@@ -56,8 +55,8 @@ void TestInference(const std::string& dirname,
   // 6. Run the inference program
   executor.Run(*inference_program, scope, feed_targets, fetch_targets);
 
+  delete inference_program;
   delete scope;
-  delete engine;
 }
 
 TEST(inference, recognize_digits) {

paddle/operators/CMakeLists.txt

@@ -147,6 +147,7 @@ op_library(max_sequence_len_op DEPS lod_rank_table)
 op_library(sequence_conv_op DEPS context_project)
 op_library(sequence_pool_op DEPS sequence_pooling)
 op_library(lstm_op DEPS sequence2batch lstm_compute)
+op_library(lstmp_op DEPS sequence2batch lstm_compute)
 op_library(gru_op DEPS sequence2batch gru_compute)
 op_library(recurrent_op DEPS executor)
 op_library(warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale math_function)

paddle/operators/activation_op.h

@@ -323,7 +323,7 @@ template <typename T>
 struct FloorFunctor : public BaseActivationFunctor<T> {
   template <typename Device, typename X, typename Out>
   void operator()(Device d, X x, Out out) const {
-    out.device(d) = x.ceil();
+    out.device(d) = x.floor();
   }
 };
 

paddle/operators/detail/grpc_client.cc

@@ -97,12 +97,27 @@ bool RPCClient::AsyncGetVariable(const std::string& ep,
   return true;
 }
 
+bool RPCClient::AsyncSendBatchBarrier(const std::string& ep, int64_t time_out) {
+  const auto ch = GetChannel(ep);
+
+  BatchBarrierProcessor* s = new BatchBarrierProcessor(ch);
+  s->Prepare(time_out);
+
+  sendrecv::VariableMessage req;
+  req.set_varname(BATCH_BARRIER_MESSAGE);
+  auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
+  rpc->Finish(&s->reply_, &s->status_, (void*)s);
+  req_count_++;
+
+  return true;
+}
+
 bool RPCClient::Wait() {
   if (req_count_ <= 0) {
     return true;
   }
-
-  std::vector<bool> a(req_count_);
+  const size_t kReqCnt = req_count_;
+  bool a[kReqCnt];
   std::vector<std::future<void>> waits(req_count_);
 
   for (int i = 0; i < req_count_; i++) {

paddle/operators/detail/grpc_client.h

@@ -71,6 +71,15 @@ class ClientBase {
     context_->set_deadline(deadline);
   }
 
+  virtual void Prepare(int64_t time_out) {
+    context_.reset(new grpc::ClientContext());
+
+    std::chrono::system_clock::time_point deadline =
+        std::chrono::system_clock::now() + std::chrono::milliseconds(time_out);
+
+    context_->set_deadline(deadline);
+  }
+
   virtual void Process() = 0;
 
   std::unique_ptr<sendrecv::SendRecvService::Stub> stub_;
@@ -117,6 +126,17 @@ class GetProcessor : public ClientBase {
   RequestGetCallBack response_call_back_ = ProcGetResponse;
 };
 
+class BatchBarrierProcessor : public ClientBase {
+ public:
+  explicit BatchBarrierProcessor(std::shared_ptr<grpc::Channel> ch)
+      : ClientBase(ch) {}
+
+  virtual ~BatchBarrierProcessor() {}
+
+  virtual void Process() {}
+  sendrecv::VoidMessage reply_;
+};
+
 class RPCClient {
  public:
   bool AsyncSendVariable(const std::string& ep,
@@ -130,6 +150,10 @@ class RPCClient {
                         const framework::Scope& scope,
                         const std::string& var_name,
                         int64_t time_out = 600 * 1000);
+
+  bool AsyncSendBatchBarrier(const std::string& ep,
+                             int64_t time_out = 600 * 1000);
+
   bool Wait();
 
  private:

paddle/operators/detail/grpc_server.cc

@@ -132,6 +132,7 @@ void AsyncGRPCServer::RunSyncUpdate() {
 
   cq_send_ = builder.AddCompletionQueue();
   cq_get_ = builder.AddCompletionQueue();
+
   server_ = builder.BuildAndStart();
   LOG(INFO) << "Server listening on " << address_ << std::endl;
 
@@ -141,11 +142,11 @@ void AsyncGRPCServer::RunSyncUpdate() {
       std::bind(&AsyncGRPCServer::TryToRegisterNewGetOne, this);
 
   t_send_.reset(
-      new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, false,
+      new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this,
                                 cq_send_.get(), "cq_send", send_register)));
 
   t_get_.reset(
-      new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this, true,
+      new std::thread(std::bind(&AsyncGRPCServer::HandleRequest, this,
                                 cq_get_.get(), "cq_get", get_register)));
 
   // wait server
@@ -174,7 +175,7 @@ void AsyncGRPCServer::TryToRegisterNewSendOne() {
   }
   RequestSend* send =
       new RequestSend(&service_, cq_send_.get(), &var_recv_queue_);
-  VLOG(4) << "create RequestSend status:" << send->Status();
+  VLOG(4) << "Create RequestSend status:" << send->Status();
 }
 
 void AsyncGRPCServer::TryToRegisterNewGetOne() {
@@ -184,11 +185,11 @@ void AsyncGRPCServer::TryToRegisterNewGetOne() {
   }
   RequestGet* get = new RequestGet(&service_, cq_get_.get(), scope_, dev_ctx_,
                                    &var_get_queue_);
-  VLOG(4) << "create Requestget status:" << get->Status();
+  VLOG(4) << "Create RequestGet status:" << get->Status();
 }
 
-// FIXME(typhoonzero): remove wait argument and change cq_name to enum.
-void AsyncGRPCServer::HandleRequest(bool wait, grpc::ServerCompletionQueue* cq,
+// FIXME(typhoonzero): change cq_name to enum.
+void AsyncGRPCServer::HandleRequest(grpc::ServerCompletionQueue* cq,
                                     std::string cq_name,
                                     std::function<void()> TryToRegisterNewOne) {
   TryToRegisterNewOne();

paddle/operators/detail/grpc_server.h

@@ -57,8 +57,7 @@ class AsyncGRPCServer final : public sendrecv::SendRecvService::Service {
   void ShutDown();
 
  protected:
-  void HandleRequest(bool wait, grpc::ServerCompletionQueue *cq,
-                     std::string cq_name,
+  void HandleRequest(grpc::ServerCompletionQueue *cq, std::string cq_name,
                      std::function<void()> TryToRegisterNewOne);
   void TryToRegisterNewSendOne();
   void TryToRegisterNewGetOne();

paddle/operators/detail/sendrecvop_utils.h

@@ -30,6 +30,9 @@ namespace paddle {
 namespace operators {
 namespace detail {
 
+#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV"
+#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV"
+
 void SerializeToMessage(const std::string& name, const framework::Variable* var,
                         const platform::DeviceContext& ctx,
                         sendrecv::VariableMessage* msg);

paddle/operators/gru_op.cc

@@ -135,14 +135,14 @@ class GRUOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 GRU Operator implements part calculations of the complete GRU as following:
 
-\f[
-update \ gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\
-reset \ gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r)  \\
-output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\
+$$
+update\_gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\
+reset\_gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r)  \\
+output\_candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\
 output: h_t = dot((1 - u_t), h_{t-1}) + dot(u_t, {h}_t)
-\f]
+$$
 
-@note To implement the complete GRU, fully-connected operator must be used  
+@note To implement the complete GRU, fully-connected operator must be used
 before to feed xu, xr and xc as the Input of GRU operator.
 )DOC");
   }

paddle/operators/one_hot_op.cc

+//   Copyright (c) 2018 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/one_hot_op.h"
+#include "paddle/framework/framework.pb.h"
+
+namespace paddle {
+namespace operators {
+
+class OneHotOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of OneHotOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of OneHotOp should not be null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    PADDLE_ENFORCE_GE(x_dims.size(), 2,
+                      "Rank of Input(X) should be at least 2.");
+    PADDLE_ENFORCE_GE(x_dims[x_dims.size() - 1], 1U,
+                      "Last dimension of Input(X) should be 1.");
+
+    int depth = ctx->Attrs().Get<int>("depth");
+
+    PADDLE_ENFORCE_GT(depth, 0, "Should provide a positive depth (%d).", depth);
+
+    framework::DDim out_dims(x_dims);
+    out_dims[out_dims.size() - 1] = depth;
+    ctx->SetOutputDim("Out", out_dims);
+    ctx->ShareLoD("X", /* --> */ "Out");
+  }
+};
+
+class OneHotOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  OneHotOpMaker(OpProto* proto, OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+             "(LoDTensor, LoDTensor<int>) Input variable with rank at least 2. "
+             "The last dimension of X should be 1. Each value of X is an index "
+             "to indicate the position.");
+    AddOutput("Out",
+              "(Tensor, Tensor<float>) Output tensor with same rank as X. "
+              "The tensor consists of one-hot representations of values in X.");
+    AddAttr<int>("depth",
+                 "A positive integer to specify the length of one-hot vector.");
+    AddAttr<int>("dtype",
+                 "An integer to specify the data type of one-hot "
+                 "vector. The default value is FP32.")
+        .SetDefault(paddle::framework::proto::DataType::FP32);
+    AddComment(R"DOC(
+One Hot Operator. This operator creates the one-hot representations for input
+index values. The following example will help to explain the function of this
+operator:
+
+X is a LoDTensor:
+  X.lod = [[0, 1, 4]]
+  X.shape = [4, 1]
+  X.data = [[1], [1], [3], [0]]
+
+set depth = 4
+
+Out is a LoDTensor:
+  Out.lod = [[0, 1, 4]]
+  Out.shape = [4, 4]
+  Out.data = [[0., 1., 0., 0.],
+              [0., 1., 0., 0.],
+              [0., 0., 0., 1.],
+              [1., 0., 0., 0.]]
+)DOC");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OPERATOR(one_hot, ops::OneHotOp, ops::OneHotOpMaker,
+                  paddle::framework::EmptyGradOpMaker);
+REGISTER_OP_CPU_KERNEL(
+    one_hot, ops::OneHotKernel<paddle::platform::CPUDeviceContext, int>,
+    ops::OneHotKernel<paddle::platform::CPUDeviceContext, int64_t>);

paddle/operators/one_hot_op.cu

+//   Copyright (c) 2018 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/one_hot_op.h"
+#include "paddle/platform/cuda_helper.h"
+#include "paddle/platform/gpu_info.h"
+
+namespace paddle {
+namespace operators {
+using platform::PADDLE_CUDA_NUM_THREADS;
+
+template <typename InT, typename OutT>
+__global__ void FillOutputKernel(const InT* p_in_data, OutT* p_out_data,
+                                 const int64_t numel, const int depth) {
+  int idx = blockIdx.x * blockDim.x + threadIdx.x;
+  if (idx < numel) {
+    *(p_out_data + (idx * depth) + p_in_data[idx]) = 1.0;
+  }
+}
+
+template <typename DeviceContext, typename InT>
+struct OneHotOpCUDAFunctor {
+  const framework::LoDTensor* in_;
+  framework::LoDTensor* out_;
+  const DeviceContext& ctx_;
+  int depth_;
+
+  OneHotOpCUDAFunctor(const framework::LoDTensor* in, framework::LoDTensor* out,
+                      int depth, const DeviceContext& ctx)
+      : in_(in), out_(out), depth_(depth), ctx_(ctx) {}
+
+  template <typename OutT>
+  void operator()() const {
+    auto* p_in_data = in_->data<InT>();
+    auto numel = in_->numel();
+    auto* p_out_data = out_->mutable_data<OutT>(ctx_.GetPlace());
+    auto stream = ctx_.stream();
+    math::set_constant(ctx_, out_, 0.0);
+
+    FillOutputKernel<<<(numel + PADDLE_CUDA_NUM_THREADS - 1) /
+                           PADDLE_CUDA_NUM_THREADS,
+                       PADDLE_CUDA_NUM_THREADS, 0, stream>>>(
+        p_in_data, p_out_data, numel, depth_);
+  }
+};
+
+using LoDTensor = framework::LoDTensor;
+template <typename DeviceContext, typename T>
+class OneHotCUDAKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in = context.Input<LoDTensor>("X");
+    auto* out = context.Output<LoDTensor>("Out");
+    int depth = context.Attr<int>("depth");
+
+    framework::VisitDataType(
+        static_cast<framework::proto::DataType>(context.Attr<int>("dtype")),
+        OneHotOpCUDAFunctor<DeviceContext, T>(
+            in, out, depth, context.template device_context<DeviceContext>()));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_CUDA_KERNEL(
+    one_hot, ops::OneHotCUDAKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::OneHotCUDAKernel<paddle::platform::CUDADeviceContext, int64_t>);

paddle/operators/one_hot_op.h

+//   Copyright (c) 2018 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/op_registry.h"
+#include "paddle/operators/math/math_function.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename InT>
+struct OneHotOpFunctor {
+  const framework::LoDTensor* in_;
+  framework::LoDTensor* out_;
+  int depth_;
+  const DeviceContext& ctx_;
+
+  OneHotOpFunctor(const framework::LoDTensor* in, framework::LoDTensor* out,
+                  int depth, const DeviceContext& ctx)
+      : in_(in), out_(out), depth_(depth), ctx_(ctx) {}
+
+  template <typename OutT>
+  void operator()() const {
+    auto* p_in_data = in_->data<InT>();
+    auto numel = in_->numel();
+    auto* p_out_data = out_->mutable_data<OutT>(ctx_.GetPlace());
+    math::set_constant(ctx_, out_, 0.0);
+
+    for (int i = 0; i < numel; ++i) {
+      PADDLE_ENFORCE_GE(p_in_data[i], 0,
+                        "Illegal index value, should be at least 0.");
+      PADDLE_ENFORCE_LT(p_in_data[i], depth_,
+                        "Illegal index value, should be less than depth (%d).",
+                        depth_);
+      *(p_out_data + i * depth_ + p_in_data[i]) = 1.0;
+    }
+  }
+};
+
+using LoDTensor = framework::LoDTensor;
+template <typename DeviceContext, typename T>
+class OneHotKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* in = context.Input<LoDTensor>("X");
+    auto* out = context.Output<LoDTensor>("Out");
+    int depth = context.Attr<int>("depth");
+
+    framework::VisitDataType(
+        static_cast<framework::proto::DataType>(context.Attr<int>("dtype")),
+        OneHotOpFunctor<DeviceContext, T>(
+            in, out, depth, context.template device_context<DeviceContext>()));
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/recv_op.cc

@@ -29,8 +29,6 @@ limitations under the License. */
 #include "paddle/operators/detail/simple_block_queue.h"
 #include "paddle/string/printf.h"
 
-#define LISTEN_TERMINATE_MESSAGE "TERMINATE@RECV"
-
 namespace paddle {
 namespace operators {
 
@@ -95,7 +93,6 @@ class RecvOp : public framework::OperatorBase {
     auto param_list = Attr<std::vector<std::string>>("ParamList");
     auto grad_list = Attr<std::vector<std::string>>("GradList");
     auto fan_in = Attr<int>("Fanin");
-    size_t param_count = param_list.size();
 
     auto *block = Attr<framework::BlockDesc *>(kOptimizeBlock);
     auto *program = block->Program();
@@ -103,38 +100,50 @@ class RecvOp : public framework::OperatorBase {
 
     // TODO(typhoonzero): change this to a while_op for every cluster-batch.
     bool exit_flag = false;
-    size_t barrier_size = param_count * fan_in;
     while (!exit_flag) {
       // Get from multiple trainers, we don't care about the order in which
       // the gradients arrives, just add suffix 0~n and merge the gradient.
       rpc_service_->SetCond(0);
-      for (size_t i = 0; i < barrier_size; ++i) {
+      size_t recv_var_cnt = 0;
+      int batch_barrier = 0;
+      while (batch_barrier != fan_in) {
         const detail::MessageWithName &v = rpc_service_->Get();
         auto grad_var_name = v.first;
         if (grad_var_name == LISTEN_TERMINATE_MESSAGE) {
           LOG(INFO) << "received terminate message and exit";
           exit_flag = true;
           break;
-        }
-        auto it = std::find(grad_list.begin(), grad_list.end(), grad_var_name);
-        std::string param_var_name;
-        if (it != grad_list.end()) {
-          param_var_name = param_list[it - grad_list.begin()];
+        } else if (grad_var_name == BATCH_BARRIER_MESSAGE) {
+          VLOG(3) << "recv batch barrier message";
+          batch_barrier++;
+          continue;
         } else {
-          LOG(ERROR) << "grad has no paired param:" << grad_var_name;
-        }
-        VLOG(3) << "received grad: " << grad_var_name
-                << " updating param: " << param_var_name;
-        if (fan_in > 1) {
-          grad_var_name = this->GetGradVarNameForTrainer(grad_var_name);
-        }
-        auto *var = recv_scope.FindVar(grad_var_name);
-        if (var == nullptr) {
-          LOG(ERROR) << "Can not find server side var: " << grad_var_name;
-          PADDLE_THROW("Can not find server side var");
+          // receive a variable
+          recv_var_cnt++;
+          auto it =
+              std::find(grad_list.begin(), grad_list.end(), grad_var_name);
+          std::string param_var_name;
+          if (it != grad_list.end()) {
+            param_var_name = param_list[it - grad_list.begin()];
+          } else {
+            LOG(ERROR) << "grad has no paired param:" << grad_var_name;
+          }
+          VLOG(3) << "received grad: " << grad_var_name
+                  << " updating param: " << param_var_name;
+
+          if (fan_in > 1) {
+            grad_var_name = this->GetGradVarNameForTrainer(grad_var_name);
+          }
+          auto *var = recv_scope.FindVar(grad_var_name);
+          if (var == nullptr) {
+            LOG(ERROR) << "Can not find server side var: " << grad_var_name;
+            PADDLE_THROW("Can not find server side var");
+          }
+          detail::DeserializeFromMessage(v.second, dev_ctx, var);
         }
-        detail::DeserializeFromMessage(v.second, dev_ctx, var);
       }
+      VLOG(3) << "recv " << recv_var_cnt << " parmeters for one barrier.";
+      // TODO(Yancey1989): merge SelectedRows variables here
       if (exit_flag) {
         break;
       }
@@ -146,7 +155,7 @@ class RecvOp : public framework::OperatorBase {
         LOG(ERROR) << "run sub program error " << e.what();
       }
       rpc_service_->SetCond(1);
-      rpc_service_->WaitClientGet(barrier_size);
+      rpc_service_->WaitClientGet(recv_var_cnt);
       grads_counter_.clear();
     }  // while(true)
   }
@@ -161,7 +170,6 @@ class RecvOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   RecvOpMaker(OpProto *proto, OpAttrChecker *op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("RX", "(Tensor) Input tensor to be optimized").AsDuplicable();
     AddComment(R"DOC(
 Recv operator
 

paddle/operators/reduce_op.cc

@@ -13,7 +13,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/operators/reduce_op.h"
-#include "paddle/operators/net_op.h"
 
 namespace paddle {
 namespace operators {
@@ -38,10 +37,14 @@ class ReduceOp : public framework::OperatorWithKernel {
         dim, x_rank,
         "The dim should be in the range [-rank(input), rank(input)).");
     bool reduce_all = ctx->Attrs().Get<bool>("reduce_all");
+    bool keep_dim = ctx->Attrs().Get<bool>("keep_dim");
     if (reduce_all) {
-      ctx->SetOutputDim("Out", {1});
+      if (keep_dim)
+        ctx->SetOutputDim(
+            "Out", framework::make_ddim(std::vector<int64_t>(x_rank, 1)));
+      else
+        ctx->SetOutputDim("Out", {1});
     } else {
-      bool keep_dim = ctx->Attrs().Get<bool>("keep_dim");
       auto dims_vector = vectorize(x_dims);
       if (keep_dim || x_rank == 1) {
         dims_vector[dim] = 1;

paddle/operators/send_op.cc

@@ -37,17 +37,25 @@ class SendOp : public framework::OperatorBase {
     auto ins = Inputs("X");
     auto outs = Outputs("Out");
     std::vector<std::string> epmap = Attr<std::vector<std::string>>("epmap");
+    std::vector<std::string> endpoints =
+        Attr<std::vector<std::string>>("endpoints");
 
     platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
     auto& ctx = *pool.Get(place);
     for (size_t i = 0; i < ins.size(); i++) {
-      VLOG(3) << "sending " << ins[i];
+      VLOG(3) << "sending " << ins[i] << " to " << epmap[i];
       client_.AsyncSendVariable(epmap[i], ctx, scope, ins[i]);
     }
     PADDLE_ENFORCE(client_.Wait());
 
+    for (auto& ep : endpoints) {
+      VLOG(3) << "batch barrier, ep: " << ep;
+      client_.AsyncSendBatchBarrier(ep);
+    }
+    PADDLE_ENFORCE(client_.Wait());
+
     for (size_t i = 0; i < outs.size(); i++) {
-      VLOG(3) << "getting " << outs[i];
+      VLOG(3) << "getting " << outs[i] << " from " << epmap[i];
       client_.AsyncGetVariable(epmap[i], ctx, scope, outs[i]);
     }
 

paddle/operators/sequence_reshape_op.cc

@@ -30,8 +30,13 @@ class SequenceReshapeOp : public framework::OperatorWithKernel {
     auto x_numel = product(x_dims);
     PADDLE_ENFORCE_EQ(x_dims.size(), 2U, "Rank of Input(X) should be 2.");
     int new_dim = ctx->Attrs().Get<int>("new_dim");
-    ctx->SetOutputDim("Out",
-                      {x_numel / new_dim, static_cast<int64_t>(new_dim)});
+    if (ctx->IsRuntime()) {
+      ctx->SetOutputDim("Out",
+                        {x_numel / new_dim, static_cast<int64_t>(new_dim)});
+    } else {
+      // when compiling, the batch size is undetermined, just set to -1
+      ctx->SetOutputDim("Out", {-1, static_cast<int64_t>(new_dim)});
+    }
   }
 };
 

paddle/scripts/docker/build.sh

@@ -32,7 +32,7 @@ function cmake_gen() {
     cat <<EOF
     ========================================
     Configuring cmake in /paddle/build ...
-        -DCMAKE_BUILD_TYPE=Release
+        -DCMAKE_BUILD_TYPE=${BUILD_TYPE:Release}
         ${PYTHON_FLAGS}
         -DWITH_DOC=OFF
         -DWITH_GPU=${WITH_GPU:-OFF}
@@ -54,7 +54,7 @@ EOF
     # docker environment is fully controlled by this script.
     # See /Paddle/CMakeLists.txt, UNITTEST_USE_VIRTUALENV option.
     cmake .. \
-        -DCMAKE_BUILD_TYPE=Release \
+        -DCMAKE_BUILD_TYPE=${BUILD_TYPE:Release} \
         ${PYTHON_FLAGS} \
         -DWITH_DOC=OFF \
         -DWITH_GPU=${WITH_GPU:-OFF} \

python/paddle/trainer/config_parser.py

@@ -140,8 +140,13 @@ def init_config_environment(
         g_submodel_stack=[],
         g_add_submodel_suffix=False, ):
 
-    for k, v in locals().iteritems():
-        globals()[k] = copy.deepcopy(v)
+    # directly iterate through locals().iteritems() will change
+    # the size of locals() due to introducing k, v into scope
+    # which will break the process in some env
+
+    local_vars = copy.deepcopy(locals())
+    for k, v in local_vars.iteritems():
+        globals()[k] = v
 
 
 # Because type is widely used as a variable name in this code.

python/paddle/v2/fluid/clip.py

@@ -30,6 +30,9 @@ __all__ = [
 
 
 class BaseErrorClipAttr(object):
+    def __str__(self):
+        raise NotImplementedError()
+
     def append_clip_op(self, block, grad_name):
         raise NotImplementedError()
 
@@ -44,6 +47,9 @@ class ErrorClipByValue(BaseErrorClipAttr):
         self.max = max
         self.min = min
 
+    def __str__(self):
+        return "ByValue, min=%f, max=%f" % (self.min, self.max)
+
     def append_clip_op(self, block, grad_name):
         clip_op_desc = block.desc.append_op()
         clip_op_desc.set_type("clip")
@@ -71,6 +77,9 @@ def error_clip_callback(block, context):
 
 
 class BaseGradientClipAttr(object):
+    def __str__(self):
+        raise NotImplementedError()
+
     def process_context(self, context, param, grad):
         raise NotImplementedError()
 
@@ -79,6 +88,9 @@ class BaseGradientClipAttr(object):
 
 
 class NullGradientClipAttr(BaseGradientClipAttr):
+    def __str__(self):
+        return "Null"
+
     def process_context(self, context, param, grad):
         pass
 
@@ -96,6 +108,9 @@ class GradientClipByValue(BaseGradientClipAttr):
         self.max = max
         self.min = min
 
+    def __str__(self):
+        return "ByValue, min=%f, max=%f" % (self.min, self.max)
+
     def process_context(self, context, param, grad):
         pass
 
@@ -108,6 +123,9 @@ class GradientClipByNorm(BaseGradientClipAttr):
     def __init__(self, clip_norm):
         self.clip_norm = clip_norm
 
+    def __str__(self):
+        return "ByNorm, clip_norm=%f" % self.clip_norm
+
     def process_context(self, context, param, grad):
         pass
 
@@ -124,6 +142,10 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
         self.clip_norm = clip_norm
         self.group_name = group_name
 
+    def __str__(self):
+        return "ByGlobalNorm, group_name=%s, clip_norm=%f" % (self.group_name,
+                                                              self.clip_norm)
+
     def process_context(self, context, param, grad):
         if self.group_name not in context:
             context[self.group_name] = []
@@ -160,6 +182,17 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
 
 
 def set_gradient_clip(clip, param_list=None, program=None):
+    """
+        To specify parameters that require gradient clip.
+        Args:
+            clip(BaseGradientClipAttr): An instance of some derived class of BaseGradientClipAttr, 
+                    which describes the type and detailed attributes of required gradient clip.
+            param_list(list, None by default): Parameters that require gradient clip. 
+                    It can be a list of parameter or a list of parameter's name. 
+                    When it's None, all parameters in the program will be included. 
+            program(Program, None by default): The program where parameters are. 
+                    Will be the default main program when assigned with None.
+    """
     if not isinstance(clip, BaseGradientClipAttr):
         raise TypeError(
             "'clip' should be an instance of BaseGradientClipAttr's derived class"
@@ -199,3 +232,5 @@ def append_gradient_clip_ops(param_grad):
 
 
 ClipByValue = GradientClipByValue
+ClipByNorm = GradientClipByNorm
+ClipByGlobalNorm = GradientClipByGlobalNorm

python/paddle/v2/fluid/distribute_transpiler.py

@@ -474,8 +474,7 @@ class DistributeTranspiler:
         # Append the recv op
         pserver_program.global_block().append_op(
             type="recv",
-            inputs={"RX": self.param_grad_ep_mapping[endpoint]["grads"]
-                    },  # grads to recv
+            inputs={},
             outputs={},
             attrs={
                 "OptimizeBlock": optimize_sub_program.global_block(),

python/paddle/v2/fluid/framework.py

@@ -14,6 +14,7 @@
 
 import collections
 import contextlib
+import re
 
 import numpy as np
 
@@ -239,20 +240,30 @@ class Variable(object):
     def __str__(self):
         return self.to_string(True)
 
-    def to_string(self, throw_on_error):
+    def to_string(self, throw_on_error, with_details=False):
         """
         Get debug string.
 
         Args:
             throw_on_error(bool): True if raise an exception when self is not
                 intialized.
+            with_details(bool): more details about variables and parameters
+                (e.g. trainable, optimize_attr, ...) will be printed when with_details is True
 
         Returns(str): The debug string.
 
         """
+        assert isinstance(throw_on_error, bool) and isinstance(with_details,
+                                                               bool)
         protostr = self.desc.serialize_to_string()
         proto = framework_pb2.VarDesc.FromString(str(protostr))
-        return _debug_string_(proto, throw_on_error)
+        res_str = _debug_string_(proto, throw_on_error)
+        if with_details:
+            additional_attr = ("error_clip", "stop_gradient")
+            for attr_name in additional_attr:
+                res_str += "%s: %s\n" % (attr_name,
+                                         str(getattr(self, attr_name)))
+        return res_str
 
     __repr__ = __str__
 
@@ -629,10 +640,36 @@ class Block(object):
     def __str__(self):
         return self.to_string(True)
 
-    def to_string(self, throw_on_error):
-        protostr = self.desc.serialize_to_string()
-        proto = framework_pb2.BlockDesc.FromString(str(protostr))
-        return _debug_string_(proto, throw_on_error)
+    def to_string(self, throw_on_error, with_details=False):
+        """
+        To debug string.
+        Args:
+            throw_on_error(bool): raise exception when self is not initialized
+                when throw_on_error is True
+            with_details(bool): more details about variables and parameters
+                (e.g. trainable, optimize_attr, ...) will be printed when with_details is True
+
+        Returns(str): The debug string.
+
+        """
+        assert isinstance(throw_on_error, bool) and isinstance(with_details,
+                                                               bool)
+        if with_details:
+            re_add_indent = re.compile(r"\n(.)")
+            res_str = "blocks {\n  idx: %d\n  parent_idx: %d" % (
+                self.idx, self.parent_idx)
+            for var in self.vars.itervalues():
+                res_str += "\n  vars {\n    %s  }" % re_add_indent.sub(
+                    r"\n    \1", var.to_string(throw_on_error, with_details))
+            for op in self.ops:
+                res_str += "\n  ops {\n    %s  }" % re_add_indent.sub(
+                    r"\n    \1", op.to_string(throw_on_error))
+            res_str += "\n}"
+        else:
+            protostr = self.desc.serialize_to_string()
+            proto = framework_pb2.BlockDesc.FromString(str(protostr))
+            res_str = _debug_string_(proto, throw_on_error)
+        return res_str
 
     __repr__ = __str__
 
@@ -796,10 +833,29 @@ class Program(object):
     def __str__(self):
         return self.to_string(True)
 
-    def to_string(self, throw_on_error):
-        protostr = self.desc.serialize_to_string()
-        proto = framework_pb2.ProgramDesc.FromString(str(protostr))
-        return _debug_string_(proto, throw_on_error)
+    def to_string(self, throw_on_error, with_details=False):
+        """
+        To debug string.
+        Args:
+            throw_on_error(bool): raise exception when self is not initialized
+                when throw_on_error is True
+            with_details(bool): more details about variables and parameters
+                (e.g. trainable, optimize_attr, ...) will be printed when with_details is True
+
+        Returns(str): The debug string.
+
+        """
+        assert isinstance(throw_on_error, bool) and isinstance(with_details,
+                                                               bool)
+        if with_details:
+            res_str = ""
+            for block in self.blocks:
+                res_str += block.to_string(throw_on_error, with_details)
+        else:
+            protostr = self.desc.serialize_to_string()
+            proto = framework_pb2.ProgramDesc.FromString(str(protostr))
+            res_str = _debug_string_(proto, throw_on_error)
+        return res_str
 
     def get_desc(self):
         return self.desc
@@ -950,6 +1006,36 @@ class Parameter(Variable):
 
         self.gradient_clip_attr = kwargs.get('gradient_clip_attr', None)
 
+    def __str__(self):
+        return self.to_string(True)
+
+    def to_string(self, throw_on_error, with_details=False):
+        """
+        To debug string.
+        Args:
+            throw_on_error(bool): raise exception when self is not initialized
+                when throw_on_error is True
+            with_details(bool): more details about variables and parameters
+                (e.g. trainable, optimize_attr, ...) will be printed when with_details is True
+
+        Returns(str): The debug string.
+
+        """
+        assert isinstance(throw_on_error, bool) and isinstance(with_details,
+                                                               bool)
+        if with_details:
+            res_str = Variable.to_string(self, throw_on_error, True)
+            additional_attr = ("trainable", "optimize_attr", "regularizer",
+                               "gradient_clip_attr")
+            for attr_name in additional_attr:
+                res_str += "%s: %s\n" % (attr_name,
+                                         str(getattr(self, attr_name)))
+        else:
+            res_str = Variable.to_string(self, throw_on_error, False)
+        return res_str
+
+    __repr__ = __str__
+
 
 # program is a global instance.
 _main_program_ = Program()

python/paddle/v2/fluid/layer_helper.py

@@ -18,7 +18,7 @@ import itertools
 from framework import Variable, Parameter, default_main_program, default_startup_program, \
     unique_name, dtype_is_floating
 from paddle.v2.fluid.initializer import Constant, Xavier
-from param_attr import ParamAttr
+from param_attr import ParamAttr, WeightNormParamAttr
 
 
 class LayerHelper(object):
@@ -104,6 +104,177 @@ class LayerHelper(object):
                                  (dtype, each.dtype))
         return dtype
 
+    def _create_weight_normalize(self, attr, shape, dtype):
+        from .layers import elementwise_mul, elementwise_div, reshape
+
+        # Remove these ops when LayerHelper and layers support indicating
+        # program and block.
+        def __norm_op(x,
+                      out=None,
+                      p=2,
+                      dim=None,
+                      keep_dim=False,
+                      block=self.startup_program.global_block()):
+            if out is None:
+                out = block.create_var(
+                    name=unique_name(".".join([self.name, 'weight_norm_norm'])),
+                    dtype=dtype,
+                    persistable=False)
+            abs_out = block.create_var(
+                name=unique_name(".".join([self.name, 'weight_norm_abs'])),
+                dtype=dtype,
+                persistable=False)
+            block.append_op(
+                type='abs', inputs={'X': x}, outputs={'Out': abs_out})
+            pow_out = block.create_var(
+                name=unique_name(".".join([self.name, 'weight_norm_pow'])),
+                dtype=dtype,
+                persistable=False)
+            block.append_op(
+                type='pow',
+                inputs={'X': abs_out},
+                outputs={'Out': pow_out},
+                attrs={'factor': float(p)})
+            sum_out = block.create_var(
+                name=unique_name(".".join([self.name, 'weight_norm_sum'])),
+                dtype=dtype,
+                persistable=False)
+            block.append_op(
+                type='reduce_sum',
+                inputs={'X': pow_out},
+                outputs={'Out': sum_out},
+                attrs={
+                    'dim': dim,
+                    'keep_dim': keep_dim,
+                    'reduce_all': True if dim is None else False
+                })
+            block.append_op(
+                type='pow',
+                inputs={'X': sum_out},
+                outputs={'Out': out},
+                attrs={'factor': 1. / p})
+            return out
+
+        def __reshape_op(x,
+                         shape,
+                         out=None,
+                         block=self.startup_program.global_block()):
+            if out is None:
+                out = block.create_var(
+                    name=unique_name(".".join(
+                        [self.name, 'weight_norm_reshape'])),
+                    dtype=dtype,
+                    persistable=False)
+            block.append_op(
+                type='reshape',
+                inputs={'X': x},
+                outputs={'Out': out},
+                attrs={'shape': shape})
+            return out
+
+        def __transpose_op(x,
+                           axis,
+                           out=None,
+                           block=self.startup_program.global_block()):
+            if out is None:
+                out = block.create_var(
+                    name=unique_name(".".join(
+                        [self.name, 'weight_norm_transpose'])),
+                    dtype=dtype,
+                    persistable=False)
+            block.append_op(
+                type='transpose',
+                inputs={'X': x},
+                outputs={'Out': out},
+                attrs={'axis': axis})
+            return out
+
+        def __norm_except_dim(x,
+                              out=None,
+                              dim=None,
+                              block=self.startup_program.global_block()):
+            """Computes the norm over all dimensions except dim"""
+            if out is None:
+                out = block.create_var(
+                    name=unique_name(".".join([self.name, 'weight_norm_norm'])),
+                    dtype=dtype,
+                    persistable=False)
+            if dim is None:
+                __norm_op(x, out, dim=dim, block=block)
+            elif dim == 0:
+                out_shape = [x.shape[0]] + [1] * (len(x.shape) - 1)
+                reshape = __reshape_op(x, shape=[x.shape[0], -1], block=block)
+                norm = __norm_op(reshape, dim=1, block=block)
+                __reshape_op(norm, out=out, shape=out_shape, block=block)
+            elif dim == len(x.shape) - 1:
+                out_shape = [1] * (len(x.shape) - 1) + [x.shape[-1]]
+                reshape = __reshape_op(x, shape=[-1, x.shape[-1]], block=block)
+                norm = __norm_op(reshape, dim=0, block=block)
+                __reshape_op(norm, out=out, shape=out_shape, block=block)
+            else:
+                perm = range(len(x.shape))
+                perm[0], perm[dim] = dim, 0
+                transpose = __transpose_op(x, perm, block=block)
+                norm = __norm_op(transpose, dim=0, block=block)
+                __transpose_op(norm, perm, out=out, block=block)
+            return out
+
+        def __weight_normalize(g, v, dim):
+            """Calculations for weight normalization"""
+            norm = __norm_except_dim(
+                v, dim=dim, block=self.main_program.current_block())
+            scale = elementwise_div(
+                x=g, y=norm)  # The shapes of g and norm are the same.
+            # Currently, elementwise_mul only support broadcast when the shape
+            # of y is a subset of the shape of x. Thus, we reshape y to squeeze
+            # to achive the subset.
+            w = elementwise_mul(
+                x=v,
+                y=scale if dim is None else reshape(
+                    x=scale, shape=[v.shape[dim]]),
+                axis=-1 if dim is None else dim)
+            # To serialize the original parameter for inference, maybe a
+            # parameter rather than a variable should be returned.
+            return w
+
+        g_param_attr = copy.deepcopy(attr)
+        g_param_attr.name = attr.name + '_g'
+        g_param_shape = [1] * len(shape)
+        if attr.dim is not None:
+            g_param_shape[attr.dim] = shape[attr.dim]
+        v_param_attr = copy.deepcopy(attr)
+        v_param_attr.name = attr.name + '_v'
+        v_param_shape = shape
+
+        # Add to startup_program to initialize g and v.
+        # Try to reconstruct the initializer of w by initializing g and v.
+        # Set the initializers of g and v as below, then the distribution
+        # of w is the same as initializing w with the given initializer.
+        # For Data-Dependent Initialization, please compute the init-values
+        # of g and v in external and then feed the values to g and v by
+        # executing an extra program.
+        g_param = self.startup_program.global_block().create_parameter(
+            dtype=dtype,
+            shape=g_param_shape,
+            **g_param_attr.to_kwargs(with_initializer=False))
+        v_param = self.startup_program.global_block().create_parameter(
+            dtype=dtype,
+            shape=v_param_shape,
+            **v_param_attr.to_kwargs(with_initializer=True))
+        __norm_except_dim(
+            x=v_param,
+            out=g_param,
+            dim=attr.dim,
+            block=self.startup_program.global_block())
+
+        # Add weight normalization to main_program
+        g_param = self.main_program.global_block().create_parameter(
+            dtype=dtype, shape=g_param_shape, **g_param_attr.to_kwargs())
+        v_param = self.main_program.global_block().create_parameter(
+            dtype=dtype, shape=v_param_shape, **v_param_attr.to_kwargs())
+        w_param = __weight_normalize(g_param, v_param, dim=attr.dim)
+        return w_param
+
     def create_parameter(self,
                          attr,
                          shape,
@@ -114,16 +285,23 @@ class LayerHelper(object):
         attr = copy.deepcopy(attr)
         assert isinstance(attr, ParamAttr)
         suffix = 'b' if is_bias else 'w'
+        if attr.name is None:
+            attr.name = unique_name(".".join([self.name, suffix]))
 
-        if default_initializer is None:
+        if default_initializer is None and attr.initializer is None:
             if is_bias:
                 attr.set_default_bias_initializer()
             else:
                 attr.set_default_param_initializer()
         else:
             attr.set_default_initializer(default_initializer)
-        if attr.name is None:
-            attr.name = unique_name(".".join([self.name, suffix]))
+
+        # If weight normalization is set, insert extra parameters and ops.
+        # Refer to https://arxiv.org/pdf/1602.07868.pdf
+        if isinstance(attr, WeightNormParamAttr):
+            param = self._create_weight_normalize(attr, shape, dtype)
+            WeightNormParamAttr.params_with_weight_norm.append(param)
+            return param
 
         self.startup_program.global_block().create_parameter(
             dtype=dtype, shape=shape, **attr.to_kwargs(with_initializer=True))

python/paddle/v2/fluid/layers/io.py

@@ -14,8 +14,10 @@
 
 from .. import core
 from ..layer_helper import LayerHelper
+from control_flow import BlockGuard
+from ..layer_helper import LayerHelper
 
-__all__ = ['data']
+__all__ = ['data', 'BlockGuardServ', 'ListenAndServ', 'Send']
 
 
 def data(name,
@@ -74,3 +76,123 @@ def data(name,
         type=type,
         stop_gradient=stop_gradient,
         lod_level=lod_level)
+
+
+class BlockGuardServ(BlockGuard):
+    """
+    BlockGuardServ class.
+
+    BlockGuardServ class is used to create an op with a block in a program.
+    """
+
+    def __init__(self, server):
+        if not (isinstance(server, ListenAndServ)):
+            raise TypeError("BlockGuardServ takes a ListenAndServ")
+        super(BlockGuardServ, self).__init__(server.helper.main_program)
+        self.server = server
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if exc_type is not None:
+            return False
+
+        self.server.complete_op()
+        return super(BlockGuardServ, self).__exit__(exc_type, exc_val, exc_tb)
+
+
+class ListenAndServ(object):
+    """
+    ListenAndServ class.
+
+    ListenAndServ class is used to wrap listen_and_serv op to create a server
+    which can receive variables from clients and run a block.
+    """
+
+    def __init__(self, endpoint, fan_in=1, optimizer_mode=True):
+        self.helper = LayerHelper("recv")
+        self.inputs = []
+        self.outputs = []
+        self.endpoint = endpoint
+        self.fan_in = fan_in
+        # FIXME(typhoonzero): add optimizer_mode is stupid, should make it more
+        # general.
+        self.optimizer_mode = optimizer_mode
+
+    def do(self):
+        return BlockGuardServ(self)
+
+    def get_params_and_grads(self):
+        main_program = self.helper.main_program
+        current_block = main_program.current_block()
+        parent_block = self.parent_block()
+        # params and grads in the same order.
+        params = list()
+        grads = list()
+        for op in current_block.ops:
+            # FIXME(typhoonzero): op.inputs is None if it's cloned.
+            if self.optimizer_mode:
+                if "Grad" in op.inputs and "Param" in op.inputs:
+                    params.append(op.inputs["Param"].name)
+                    grads.append(op.inputs["Grad"].name)
+            else:
+                # simple recv mode, recv operators inputs.
+                for iname in op.input_names:
+                    for in_var_name in op.input(iname):
+                        params.append(parent_block.var(in_var_name))
+                        grads.append(parent_block.var(in_var_name))
+
+        return params, grads
+
+    def parent_block(self):
+        prog = self.helper.main_program
+        parent_idx = prog.current_block().parent_idx
+        assert parent_idx >= 0
+        parent_block = prog.block(parent_idx)
+        return parent_block
+
+    def complete_op(self):
+        main_program = self.helper.main_program
+        current_block = main_program.current_block()
+        parent_block = self.parent_block()
+
+        params, grads = self.get_params_and_grads()
+        param_names = [p.name for p in params]
+        grad_names = [g.name for g in grads]
+        parent_block.append_op(
+            type='recv',
+            inputs={},
+            outputs={},
+            attrs={
+                'endpoint': self.endpoint,
+                'Fanin': self.fan_in,
+                'ParamList': param_names,
+                'GradList': grad_names,
+                'OptimizeBlock': current_block
+            })
+
+
+def Send(endpoints, send_vars, get_vars):
+    """
+    Send layer
+
+    Args:
+        endpoints: comma seperated IP:PORT pairs in the order
+                   of send_vars to send
+        send_vars: vars to send
+        get_vars: vars to get from server after send completes.
+
+    Send variables to the server side, and get vars from server
+    side when server have finished running server side program.
+    """
+    assert (type(send_vars) == list)
+    assert (type(get_vars) == list)
+
+    epmap = endpoints.split(",")
+    endpoints = list(set(epmap))
+
+    helper = LayerHelper("Send", **locals())
+    helper.append_op(
+        type="send",
+        inputs={"X": send_vars},
+        outputs={"Out": get_vars},
+        attrs={"endpoints": endpoints,
+               "epmap": epmap})

python/paddle/v2/fluid/layers/nn.py

@@ -26,6 +26,7 @@ __all__ = [
     'fc',
     'embedding',
     'dynamic_lstm',
+    'dynamic_lstmp',
     'dynamic_gru',
     'gru_unit',
     'linear_chain_crf',
@@ -256,7 +257,8 @@ def dynamic_lstm(input,
                  gate_activation='sigmoid',
                  cell_activation='tanh',
                  candidate_activation='tanh',
-                 dtype='float32'):
+                 dtype='float32',
+                 name=None):
     """
     **Dynamic LSTM Layer**
 
@@ -282,7 +284,7 @@ def dynamic_lstm(input,
     W_{fc}, W_{oc}` are diagonal weight matrices for peephole connections. In
     our implementation, we use vectors to reprenset these diagonal weight
     matrices. The :math:`b` terms denote bias vectors (:math:`b_i` is the input
-    gate bias vector), :math:`\sigma` is the non-line activations, such as
+    gate bias vector), :math:`\sigma` is the non-linear activations, such as
     logistic sigmoid function, and :math:`i, f, o` and :math:`c` are the input
     gate, forget gate, output gate, and cell activation vectors, respectively,
     all of which have the same size as the cell output activation vector :math:`h`.
@@ -308,25 +310,25 @@ def dynamic_lstm(input,
                          (T X 4D), where T is the total time steps in this
                          mini-batch, D is the hidden size.
         size(int): 4 * hidden size.
-        param_attr(ParamAttr): The parameter attribute for the learnable
+        param_attr(ParamAttr|None): The parameter attribute for the learnable
                                hidden-hidden weights.
 
-                               - The shape is (D x 4D), where D is the hidden
-                                 size.
                                - Weights = {:math:`W_{ch}, W_{ih}, \
                                                 W_{fh}, W_{oh}`}
-        bias_attr(ParamAttr): The bias attribute for the learnable bias
+                               - The shape is (D x 4D), where D is the hidden
+                                 size.
+        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
                               weights, which contains two parts, input-hidden
                               bias weights and peephole connections weights if
                               setting `use_peepholes` to `True`.
 
                               1. `use_peepholes = False`
-                                - The shape is (1 x 4D).
                                 - Biases = {:math:`b_c, b_i, b_f, b_o`}.
+                                - The shape is (1 x 4D).
                               2. `use_peepholes = True`
-                                - The shape is (1 x 7D).
                                 - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
                                                  W_{fc}, W_{oc}`}.
+                                - The shape is (1 x 7D).
         use_peepholes(bool): Whether to enable diagonal/peephole connections,
                              default `True`.
         is_reverse(bool): Whether to compute reversed LSTM, default `False`.
@@ -339,6 +341,8 @@ def dynamic_lstm(input,
                               Choices = ["sigmoid", "tanh", "relu", "identity"],
                               default "tanh".
         dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
 
     Returns:
         tuple: The hidden state, and cell state of LSTM. The shape of both \
@@ -353,6 +357,7 @@ def dynamic_lstm(input,
             forward, _ = fluid.layers.dynamic_lstm(
                 input=forward_proj, size=hidden_dim * 4, use_peepholes=False)
     """
+
     helper = LayerHelper('lstm', **locals())
     size = size / 4
     weight = helper.create_parameter(
@@ -389,6 +394,192 @@ def dynamic_lstm(input,
     return hidden, cell
 
 
+def dynamic_lstmp(input,
+                  size,
+                  proj_size,
+                  param_attr=None,
+                  bias_attr=None,
+                  use_peepholes=True,
+                  is_reverse=False,
+                  gate_activation='sigmoid',
+                  cell_activation='tanh',
+                  candidate_activation='tanh',
+                  proj_activation='tanh',
+                  dtype='float32',
+                  name=None):
+    """
+    **Dynamic LSTMP Layer**
+
+    LSTMP (LSTM with recurrent projection) layer has a separate projection 
+    layer after the LSTM layer, projecting the original hidden state to a 
+    lower-dimensional one, which is proposed to reduce the number of total 
+    parameters and furthermore computational complexity for the LSTM, 
+    espeacially for the case that the size of output units is relative 
+    large (https://research.google.com/pubs/archive/43905.pdf). 
+
+    The formula is as follows:
+
+    .. math::
+
+        i_t & = \sigma(W_{ix}x_{t} + W_{ir}r_{t-1} + W_{ic}c_{t-1} + b_i)
+
+        f_t & = \sigma(W_{fx}x_{t} + W_{fr}r_{t-1} + W_{fc}c_{t-1} + b_f)
+
+        \\tilde{c_t} & = act_g(W_{cx}x_t + W_{cr}r_{t-1} + b_c)
+
+        o_t & = \sigma(W_{ox}x_{t} + W_{or}r_{t-1} + W_{oc}c_t + b_o)
+
+        c_t & = f_t \odot c_{t-1} + i_t \odot \\tilde{c_t}
+
+        h_t & = o_t \odot act_h(c_t)
+
+        r_t & = \overline{act_h}(W_{rh}h_t)
+
+    In the above formula:
+
+    * :math:`W`: Denotes weight matrices (e.g. :math:`W_{xi}` is \
+          the matrix of weights from the input gate to the input).
+    * :math:`W_{ic}`, :math:`W_{fc}`, :math:`W_{oc}`: Diagonal weight \
+          matrices for peephole connections. In our implementation, \
+          we use vectors to reprenset these diagonal weight matrices. 
+    * :math:`b`: Denotes bias vectors (e.g. :math:`b_i` is the input gate \
+          bias vector). 
+    * :math:`\sigma`: The activation, such as logistic sigmoid function.
+    * :math:`i, f, o` and :math:`c`: The input gate, forget gate, output \
+          gate, and cell activation vectors, respectively, all of which have \
+          the same size as the cell output activation vector :math:`h`. 
+    * :math:`h`: The hidden state.
+    * :math:`r`: The recurrent projection of the hidden state. 
+    * :math:`\\tilde{c_t}`: The candidate hidden state, whose \
+          computation is based on the current input and previous hidden state.
+    * :math:`\odot`: The element-wise product of the vectors. 
+    * :math:`act_g` and :math:`act_h`: The cell input and cell output \
+          activation functions and `tanh` is usually used for them. 
+    * :math:`\overline{act_h}`: The activation function for the projection \
+          output, usually using `identity` or same as :math:`act_h`.
+
+    Set `use_peepholes` to `False` to disable peephole connection. The formula
+    is omitted here, please refer to the paper
+    http://www.bioinf.jku.at/publications/older/2604.pdf for details.
+    
+    Note that these :math:`W_{xi}x_{t}, W_{xf}x_{t}, W_{xc}x_{t}, W_{xo}x_{t}`
+    operations on the input :math:`x_{t}` are NOT included in this operator.
+    Users can choose to use fully-connected layer before LSTMP layer.
+
+    Args:
+        input(Variable): The input of dynamic_lstmp layer, which supports
+                         variable-time length input sequence. The underlying
+                         tensor in this Variable is a matrix with shape
+                         (T X 4D), where T is the total time steps in this
+                         mini-batch, D is the hidden size.
+        size(int): 4 * hidden size.
+        proj_size(int): The size of projection output.
+        param_attr(ParamAttr|None): The parameter attribute for the learnable
+                               hidden-hidden weight and projection weight.
+
+                               - Hidden-hidden weight = {:math:`W_{ch}, W_{ih}, \
+                                                W_{fh}, W_{oh}`}.
+                               - The shape of hidden-hidden weight is (P x 4D), 
+                                 where P is the projection size and D the hidden 
+                                 size.
+                               - Projection weight = {:math:`W_{rh}`}.
+                               - The shape of projection weight is (D x P).
+        bias_attr(ParamAttr|None): The bias attribute for the learnable bias
+                              weights, which contains two parts, input-hidden
+                              bias weights and peephole connections weights if
+                              setting `use_peepholes` to `True`.
+
+                              1. `use_peepholes = False`
+                                - Biases = {:math:`b_c, b_i, b_f, b_o`}.
+                                - The shape is (1 x 4D).
+                              2. `use_peepholes = True`
+                                - Biases = { :math:`b_c, b_i, b_f, b_o, W_{ic}, \
+                                                 W_{fc}, W_{oc}`}.
+                                - The shape is (1 x 7D).
+        use_peepholes(bool): Whether to enable diagonal/peephole connections,
+                             default `True`.
+        is_reverse(bool): Whether to compute reversed LSTM, default `False`.
+        gate_activation(str): The activation for input gate, forget gate and
+                              output gate. Choices = ["sigmoid", "tanh", "relu",
+                              "identity"], default "sigmoid".
+        cell_activation(str): The activation for cell output. Choices = ["sigmoid",
+                              "tanh", "relu", "identity"], default "tanh".
+        candidate_activation(str): The activation for candidate hidden state.
+                              Choices = ["sigmoid", "tanh", "relu", "identity"],
+                              default "tanh".
+        proj_activation(str): The activation for projection output.
+                              Choices = ["sigmoid", "tanh", "relu", "identity"],
+                              default "tanh".
+        dtype(str): Data type. Choices = ["float32", "float64"], default "float32".
+        name(str|None): A name for this layer(optional). If set None, the layer
+                        will be named automatically.
+
+    Returns:
+        tuple: The projection of hidden state, and cell state of LSTMP. The \
+               shape of projection is (T x P), for the cell state which is \
+               (T x D), and both LoD is the same with the `input`.
+
+    Examples:
+        .. code-block:: python
+
+            hidden_dim, proj_dim = 512, 256
+            fc_out = fluid.layers.fc(input=input_seq, size=hidden_dim * 4,
+                                     act=None, bias_attr=None)
+            proj_out, _ = fluid.layers.dynamic_lstmp(input=fc_out, 
+                                                     size=hidden_dim * 4, 
+                                                     proj_size=proj_dim, 
+                                                     use_peepholes=False,
+                                                     is_reverse=True,
+                                                     cell_activation="tanh",
+                                                     proj_activation="tanh")
+    """
+
+    helper = LayerHelper('lstmp', **locals())
+    size = size / 4
+    weight = helper.create_parameter(
+        attr=helper.param_attr, shape=[proj_size, 4 * size], dtype=dtype)
+    proj_weight = helper.create_parameter(
+        attr=helper.param_attr, shape=[size, proj_size], dtype=dtype)
+    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=dtype, is_bias=True)
+
+    projection = helper.create_tmp_variable(dtype)
+    cell = helper.create_tmp_variable(dtype)
+    ordered_proj0 = helper.create_tmp_variable(dtype)
+    batch_hidden = helper.create_tmp_variable(dtype)
+    batch_gate = helper.create_tmp_variable(dtype)
+    batch_cell_pre_act = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type='lstmp',
+        inputs={
+            'Input': input,
+            'Weight': weight,
+            'ProjWeight': proj_weight,
+            'Bias': bias
+        },
+        outputs={
+            'Projection': projection,
+            'Cell': cell,
+            'OrderedP0': ordered_proj0,
+            'BatchHidden': batch_hidden,
+            '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,
+            'proj_activation': proj_activation
+        })
+    return projection, cell
+
+
 def dynamic_gru(input,
                 size,
                 param_attr=None,

python/paddle/v2/fluid/param_attr.py

@@ -15,7 +15,10 @@
 from initializer import Initializer, Xavier, Constant
 from regularizer import WeightDecayRegularizer
 
-__all__ = ['ParamAttr']
+__all__ = [
+    'ParamAttr',
+    'WeightNormParamAttr',
+]
 
 
 class ParamAttr(object):
@@ -82,3 +85,20 @@ class ParamAttr(object):
         if with_initializer:
             kwargs['initializer'] = self.initializer
         return kwargs
+
+
+class WeightNormParamAttr(ParamAttr):
+    """
+    Used for weight normalization. Any field in ParamAttr can also be set here.
+    Besides, an extra field dim can be set to indicate the dimension except 
+    which to normalize.
+    """
+    # List to record the parameters reparameterized by weight normalization.
+    # If these parameters are treated as Variable rather than Parameter,
+    # it can be used to discriminate these parameters and help to serialize
+    # these paramters for inference.
+    params_with_weight_norm = []
+
+    def __init__(self, dim=None, **kwargs):
+        super(WeightNormParamAttr, self).__init__(**kwargs)
+        self.dim = dim

python/paddle/v2/fluid/regularizer.py

@@ -87,6 +87,11 @@ class WeightDecayRegularizer(object):
         """
         raise NotImplementedError()
 
+    def __str__(self):
+        """Debug string
+        """
+        raise NotImplementedError()
+
 
 class L2DecayRegularizer(WeightDecayRegularizer):
     """Implements the L2 Weight Decay Regularization
@@ -123,6 +128,9 @@ class L2DecayRegularizer(WeightDecayRegularizer):
 
         return decay
 
+    def __str__(self):
+        return "L2Decay, regularization_coeff=%f" % self._regularization_coeff
+
 
 class L1DecayRegularizer(WeightDecayRegularizer):
     """Implements the L1 Weight Decay Regularization
@@ -163,6 +171,9 @@ class L1DecayRegularizer(WeightDecayRegularizer):
 
         return decay
 
+    def __str__(self):
+        return "L1Decay, regularization_coeff=%f" % self._regularization_coeff
+
 
 # We short the class name, since users will use the regulaizer with the package
 # name. The sample code:

python/paddle/v2/fluid/tests/CMakeLists.txt

 file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
 string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+if(NOT WITH_DISTRIBUTE)
+    list(REMOVE_ITEM TEST_OPS test_recv_op)
+endif(NOT WITH_DISTRIBUTE)
+
 foreach(src ${TEST_OPS})
     py_test(${src} SRCS ${src}.py)
 endforeach()

python/paddle/v2/fluid/tests/book/test_rnn_encoder_decoder.py

@@ -49,7 +49,11 @@ def bi_lstm_encoder(input_seq, hidden_size):
         size=hidden_size * 4,
         is_reverse=True,
         use_peepholes=USE_PEEPHOLES)
-    return forward, backward
+
+    forward_last = fluid.layers.sequence_last_step(input=forward)
+    backward_first = fluid.layers.sequence_first_step(input=backward)
+
+    return forward_last, backward_first
 
 
 # FIXME(peterzhang2029): Replace this function with the lstm_unit_op.
@@ -115,16 +119,13 @@ def seq_to_seq_net():
         size=[source_dict_dim, embedding_dim],
         dtype='float32')
 
-    src_forward, src_backward = bi_lstm_encoder(
+    src_forward_last, src_backward_first = bi_lstm_encoder(
         input_seq=src_embedding, hidden_size=encoder_size)
 
-    src_forward_last = fluid.layers.sequence_last_step(input=src_forward)
-    src_backward_first = fluid.layers.sequence_first_step(input=src_backward)
-
     encoded_vector = fluid.layers.concat(
         input=[src_forward_last, src_backward_first], axis=1)
 
-    decoder_boot = fluid.layers.fc(input=encoded_vector,
+    decoder_boot = fluid.layers.fc(input=src_backward_first,
                                    size=decoder_size,
                                    bias_attr=False,
                                    act='tanh')

python/paddle/v2/fluid/tests/book_distribute/notest_dist_image_classification.py

-#Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
 #
-#Licensed under the Apache License, Version 2.0 (the "License");
-#you may not use this file except in compliance with the License.
-#You may obtain a copy of the License at
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
 #
-#    http://www.apache.org/licenses/LICENSE-2.0
+#     http://www.apache.org/licenses/LICENSE-2.0
 #
-#Unless required by applicable law or agreed to in writing, software
-#distributed under the License is distributed on an "AS IS" BASIS,
-#WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-#See the License for the specific language governing permissions and
-#limitations under the License.
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
 
 from __future__ import print_function
 
-import sys
-
 import paddle.v2 as paddle
 import paddle.v2.fluid as fluid
 import os
@@ -106,10 +104,10 @@ if len(sys.argv) >= 2:
     net_type = sys.argv[1]
 
 if net_type == "vgg":
-    print("train vgg net")
+    print("training vgg net")
     net = vgg16_bn_drop(images)
 elif net_type == "resnet":
-    print("train resnet")
+    print("training resnet")
     net = resnet_cifar10(images, 32)
 else:
     raise ValueError("%s network is not supported" % net_type)
@@ -129,6 +127,7 @@ train_reader = paddle.batch(
     batch_size=BATCH_SIZE)
 
 place = fluid.CPUPlace()
+feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
 exe = fluid.Executor(place)
 
 t = fluid.DistributeTranspiler()
@@ -146,17 +145,14 @@ if training_role == "PSERVER":
     if not current_endpoint:
         print("need env SERVER_ENDPOINT")
         exit(1)
-    print("start pserver at:", current_endpoint)
     pserver_prog = t.get_pserver_program(current_endpoint)
     pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
     exe.run(pserver_startup)
     exe.run(pserver_prog)
-    print("pserver run end")
 elif training_role == "TRAINER":
-    print("start trainer")
     trainer_prog = t.get_trainer_program()
-    feeder = fluid.DataFeeder(place=place, feed_list=[images, label])
     exe.run(fluid.default_startup_program())
+
     for pass_id in range(PASS_NUM):
         accuracy.reset(exe)
         for data in train_reader():
@@ -164,9 +160,10 @@ elif training_role == "TRAINER":
                                 feed=feeder.feed(data),
                                 fetch_list=[avg_cost] + accuracy.metrics)
             pass_acc = accuracy.eval(exe)
-            print("loss:" + str(loss) + " acc:" + str(acc) + " pass_acc:" + str(
-                pass_acc))
-            # this model is slow, so if we can train two mini batch, we think it works properly.
+            print("pass_id:" + str(pass_id) + "loss:" + str(loss) + " pass_acc:"
+                  + str(pass_acc))
+            # this model is slow, so if we can train two mini batches,
+            # we think it works properly.
     print("trainer run end")
 else:
     print("environment var TRAINER_ROLE should be TRAINER os PSERVER")

python/paddle/v2/fluid/tests/book_distribute/notest_understand_sentiment_dynamic_lstm.py

+#   Copyright (c) 2018 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.
+
+import numpy as np
+import os
+import paddle.v2 as paddle
+import paddle.v2.fluid as fluid
+
+
+def stacked_lstm_net(data,
+                     label,
+                     input_dim,
+                     class_dim=2,
+                     emb_dim=128,
+                     hid_dim=512,
+                     stacked_num=3):
+    assert stacked_num % 2 == 1
+
+    emb = fluid.layers.embedding(input=data, size=[input_dim, emb_dim])
+    # add bias attr
+
+    # TODO(qijun) linear act
+    fc1 = fluid.layers.fc(input=emb, size=hid_dim)
+    lstm1, cell1 = fluid.layers.dynamic_lstm(input=fc1, size=hid_dim)
+
+    inputs = [fc1, lstm1]
+
+    for i in range(2, stacked_num + 1):
+        fc = fluid.layers.fc(input=inputs, size=hid_dim)
+        lstm, cell = fluid.layers.dynamic_lstm(
+            input=fc, size=hid_dim, is_reverse=(i % 2) == 0)
+        inputs = [fc, lstm]
+
+    fc_last = fluid.layers.sequence_pool(input=inputs[0], pool_type='max')
+    lstm_last = fluid.layers.sequence_pool(input=inputs[1], pool_type='max')
+
+    prediction = fluid.layers.fc(input=[fc_last, lstm_last],
+                                 size=class_dim,
+                                 act='softmax')
+    cost = fluid.layers.cross_entropy(input=prediction, label=label)
+    avg_cost = fluid.layers.mean(x=cost)
+    adam_optimizer = fluid.optimizer.Adam(learning_rate=0.002)
+    optimize_ops, params_grads = adam_optimizer.minimize(avg_cost)
+    accuracy = fluid.evaluator.Accuracy(input=prediction, label=label)
+    return avg_cost, accuracy, accuracy.metrics[0], optimize_ops, params_grads
+
+
+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 = fluid.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 "loaded word dict successfully"
+    dict_dim = len(word_dict)
+    class_dim = 2
+
+    data = fluid.layers.data(
+        name="words", shape=[1], dtype="int64", lod_level=1)
+    label = fluid.layers.data(name="label", shape=[1], dtype="int64")
+    cost, accuracy, acc_out, optimize_ops, params_grads = stacked_lstm_net(
+        data, label, 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 = fluid.CPUPlace()
+    exe = fluid.Executor(place)
+    feeder = fluid.DataFeeder(feed_list=[data, label], place=place)
+
+    t = fluid.DistributeTranspiler()
+    # all parameter server endpoints list for spliting parameters
+    pserver_endpoints = os.getenv("PSERVERS")
+    # server endpoint for current node
+    current_endpoint = os.getenv("SERVER_ENDPOINT")
+    # run as trainer or parameter server
+    training_role = os.getenv(
+        "TRAINING_ROLE", "TRAINER")  # get the training role: trainer/pserver
+    t.transpile(
+        optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
+
+    if training_role == "PSERVER":
+        if not current_endpoint:
+            print("need env SERVER_ENDPOINT")
+            exit(1)
+        pserver_prog = t.get_pserver_program(current_endpoint)
+        pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
+        exe.run(pserver_startup)
+        exe.run(pserver_prog)
+    elif training_role == "TRAINER":
+        exe.run(fluid.default_startup_program())
+        trainer_prog = t.get_trainer_program()
+        for pass_id in xrange(PASS_NUM):
+            accuracy.reset(exe)
+            for data in train_data():
+                cost_val, acc_val = exe.run(trainer_prog,
+                                            feed=feeder.feed(data),
+                                            fetch_list=[cost, acc_out])
+                pass_acc = accuracy.eval(exe)
+                print("cost=" + str(cost_val) + " acc=" + str(acc_val) +
+                      " pass_acc=" + str(pass_acc))
+                if cost_val < 1.0 and acc_val > 0.8:
+                    exit(0)
+    else:
+        print("environment var TRAINER_ROLE should be TRAINER os PSERVER")
+
+
+if __name__ == '__main__':
+    main()

python/paddle/v2/fluid/tests/test_activation_op.py

@@ -186,8 +186,7 @@ class TestFloor(OpTest):
         self.op_type = "floor"
         x = np.random.uniform(-1, 1, [4, 4]).astype("float32")
         self.inputs = {'X': x}
-        # numpy floor need +1
-        self.outputs = {'Out': np.floor(self.inputs['X']) + 1.0}
+        self.outputs = {'Out': np.floor(self.inputs['X'])}
 
     def test_check_output(self):
         self.check_output()

python/paddle/v2/fluid/tests/test_layers.py

@@ -202,6 +202,18 @@ class TestBook(unittest.TestCase):
                     x_t=x_t, hidden_t_prev=prev_hidden, cell_t_prev=prev_cell))
         print(str(program))
 
+    def test_dynamic_lstmp(self):
+        program = Program()
+        with program_guard(program):
+            hidden_dim, proj_dim = 16, 8
+            seq_data = layers.data(
+                name='seq_data', shape=[10, 10], dtype='float32', lod_level=1)
+            fc_out = layers.fc(input=seq_data, size=4 * hidden_dim)
+            self.assertIsNotNone(
+                layers.dynamic_lstmp(
+                    input=fc_out, size=4 * hidden_dim, proj_size=proj_dim))
+        print(str(program))
+
     def test_sequence_softmax(self):
         program = Program()
         with program_guard(program):

python/paddle/v2/fluid/tests/test_one_hot_op.py

+#   Copyright (c) 2018 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.
+
+import unittest
+import numpy as np
+import math
+from op_test import OpTest
+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.core as core
+import paddle.v2.fluid.framework as framework
+from paddle.v2.fluid.framework import Program, program_guard
+
+
+class TestOneHotOp(OpTest):
+    def setUp(self):
+        self.op_type = 'one_hot'
+        depth = 10
+        dimension = 12
+        x_lod = [[0, 4, 5, 8, 11]]
+        x = [np.random.randint(0, depth - 1) for i in xrange(x_lod[0][-1])]
+        x = np.array(x).astype('int').reshape([x_lod[0][-1], 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        for i in xrange(np.product(x.shape)):
+            out[i, x[i]] = 1.0
+
+        self.inputs = {'X': (x, x_lod)}
+        self.attrs = {'depth': depth, 'dtype': int(core.DataType.FP32)}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestOneHotOp_default_dtype(OpTest):
+    def setUp(self):
+        self.op_type = 'one_hot'
+        depth = 10
+        dimension = 12
+        x_lod = [[0, 4, 5, 8, 11]]
+        x = [np.random.randint(0, depth - 1) for i in xrange(x_lod[0][-1])]
+        x = np.array(x).astype('int').reshape([x_lod[0][-1], 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        for i in xrange(np.product(x.shape)):
+            out[i, x[i]] = 1.0
+
+        self.inputs = {'X': (x, x_lod)}
+        self.attrs = {'depth': depth}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output()
+
+
+class TestOneHotOp_exception(OpTest):
+    def setUp(self):
+        self.op_type = 'one_hot'
+        self.depth = 10
+        self.place = core.CPUPlace()
+        self.dimension = 12
+        self.x = core.LoDTensor()
+        x_lod = [[0, 4, 5, 8, 11]]
+        data = [np.random.randint(11, 20) for i in xrange(x_lod[0][-1])]
+        data = np.array(data).astype('int').reshape([x_lod[0][-1], 1])
+        self.x.set(data, self.place)
+        self.x.set_lod(x_lod)
+
+    def test_check_output(self):
+        program = Program()
+        with program_guard(program):
+            x = fluid.layers.data(
+                name='x', shape=[self.dimension], dtype='float32', lod_level=1)
+            block = program.current_block()
+            one_hot_out = block.create_var(
+                name="one_hot_out",
+                type=core.VarDesc.VarType.LOD_TENSOR,
+                dtype='float32')
+            block.append_op(
+                type='one_hot',
+                inputs={'X': x},
+                attrs={'depth': self.depth},
+                outputs={'Out': one_hot_out})
+            exe = fluid.Executor(self.place)
+
+            def run():
+                exe.run(feed={'x': self.x},
+                        fetch_list=[one_hot_out],
+                        return_numpy=False)
+
+            self.assertRaises(core.EnforceNotMet, run)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/fluid/tests/test_recv_op.py

+#   Copyright (c) 2018 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.
+
+import unittest
+
+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.layers as layers
+import numpy
+from multiprocessing import Process
+import os, sys
+
+
+class TestRecvOp(unittest.TestCase):
+    def test_send(self):
+        # Run init_serv in a thread
+        place = fluid.CPUPlace()
+        p = Process(target=self.init_serv, args=(place, ))
+        p.daemon = True
+        p.start()
+        self.init_client(place)
+        # FIXME(typhoonzero): find a way to gracefully shutdown the server.
+        os.system("kill -9 %d" % p.pid)
+        p.join()
+
+    def init_serv(self, place):
+        main = fluid.Program()
+        with fluid.program_guard(main):
+            x = layers.data(
+                shape=[32, 32],
+                dtype='float32',
+                name="X",
+                append_batch_size=False)
+            fluid.initializer.Constant(value=1.0)(x, main.global_block())
+            serv = layers.ListenAndServ("127.0.0.1:6174", optimizer_mode=False)
+            with serv.do():
+                o = layers.scale(x=x, scale=10.0)
+            main.global_block().create_var(
+                name=o.name, psersistable=False, dtype=o.dtype, shape=o.shape)
+        exe = fluid.Executor(place)
+        exe.run(main)
+
+    def init_client(self, place):
+        main = fluid.Program()
+        with fluid.program_guard(main):
+            x = layers.data(
+                shape=[32, 32],
+                dtype='float32',
+                name='X',
+                append_batch_size=False)
+            fluid.initializer.Constant(value=1.0)(x, main.global_block())
+            layers.Send("127.0.0.1:6174", [x], [x])
+        exe = fluid.Executor(place)
+        exe.run(main)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/v2/fluid/tests/test_weight_normalization.py

+#   Copyright (c) 2018 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.
+
+import unittest
+import numpy
+import collections
+import paddle.v2.fluid as fluid
+import paddle.v2.fluid.core as core
+from paddle.v2.fluid.initializer import ConstantInitializer
+from paddle.v2.fluid.param_attr import WeightNormParamAttr
+
+
+class TestWeightNormalization(unittest.TestCase):
+    batch_size = 3
+    hidden_size = 5
+    data_desc = (['x', [10], 0], )
+
+    @classmethod
+    def setUpClass(cls):
+        cls.set_program()
+
+    @classmethod
+    def set_program(cls):
+        data = fluid.layers.data(
+            name=cls.data_desc[0][0], shape=cls.data_desc[0][1])
+        out = fluid.layers.fc(input=data,
+                              size=cls.hidden_size,
+                              param_attr=WeightNormParamAttr(
+                                  dim=None,
+                                  name='weight_norm_param',
+                                  initializer=ConstantInitializer(1.0)),
+                              bias_attr=False,
+                              act=None)
+        loss = fluid.layers.reduce_sum(out)
+        fluid.backward.append_backward(loss=loss)
+        cls.fetch_list = [
+            'weight_norm_param_g', 'weight_norm_param_v',
+            'weight_norm_param_g@GRAD'
+        ]
+
+    def run_program(self):
+        outputs = []
+        places = [core.CPUPlace()]
+        if core.is_compiled_with_cuda():
+            places.append(core.CUDAPlace(0))
+        for place in places:
+            self.set_inputs(place)
+            exe = fluid.Executor(place)
+            exe.run(fluid.default_startup_program())
+            output = exe.run(fluid.default_main_program(),
+                             feed=self.inputs,
+                             fetch_list=self.fetch_list,
+                             return_numpy=False)
+            outputs.append(output)
+        self.actual_outputs = outputs
+
+    def set_data(self):
+        self.data = collections.OrderedDict()
+        for desc in self.data_desc:
+            data_name = desc[0]
+            data_shape = desc[1]
+            data_lod_level = desc[2]
+            data_lod = []
+            for i in range(data_lod_level):
+                lod_level_i = numpy.random.randint(
+                    low=1,
+                    high=5,
+                    size=self.batch_size if i == 0 else lod_level_i[-1])
+                lod_level_i = [0] + numpy.cumsum(lod_level_i).tolist()
+                data_lod.append(lod_level_i)
+            data_value = numpy.random.random(
+                size=[data_lod[-1][-1] if data_lod else self.batch_size
+                      ] + data_shape).astype('float32')
+            self.data[data_name] = (data_value, data_lod)
+
+    def set_inputs(self, place):
+        self.inputs = {}
+        for desc in self.data_desc:
+            tensor = fluid.Tensor()
+            tensor.set(self.data[desc[0]][0], place)
+            if self.data[desc[0]][1]:
+                tensor.set_lod(self.data[desc[0]][1])
+            self.inputs[desc[0]] = tensor
+
+    def weight_normalize(self):
+        v = numpy.ones((self.data[self.data_desc[0][0]][0].shape[-1],
+                        self.hidden_size))
+        g = numpy.linalg.norm(v, axis=None, keepdims=True)
+        w = g * v / numpy.linalg.norm(v, axis=None, keepdims=True)
+        x = self.data[self.data_desc[0][0]][0]
+        out = numpy.dot(x, w)
+        g_grad = (numpy.dot(x.T, numpy.ones_like(out)) * (v / numpy.linalg.norm(
+            v, axis=None, keepdims=True))).sum(axis=None, keepdims=True)
+        return g, v, g_grad
+
+    def test_weight_normalization(self):
+        self.set_data()
+        self.run_program()
+        expect_output = self.weight_normalize()
+        for actual_output in self.actual_outputs:
+            [
+                self.assertTrue(
+                    numpy.allclose(
+                        numpy.array(actual), expect, atol=0.001))
+                for expect, actual in zip(expect_output, actual_output)
+            ]
+
+
+if __name__ == '__main__':
+    unittest.main()