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

doc/getstarted/quickstart_cn.rst

@@ -1,6 +1,9 @@
 快速开始
 ========
 
+快速安装
+--------
+
 PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14.04以及MacOS 10.12，并安装有Python2.7。
 执行下面的命令完成快速安装，版本为cpu_avx_openblas：
 
@@ -16,6 +19,9 @@ PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14.
 
 更详细的安装和编译方法参考：:ref:`install_steps` 。
 
+快速使用
+--------
+
 创建一个 housing.py 并粘贴此Python代码：
 
   .. code-block:: python

doc/getstarted/quickstart_en.rst

@@ -1,6 +1,9 @@
 Quick Start
 ============
 
+Quick Install
+-------------
+
 You can use pip to install PaddlePaddle with a single command, supports
 CentOS 6 above, Ubuntu 14.04 above or MacOS 10.12, with Python 2.7 installed.
 Simply run the following command to install, the version is cpu_avx_openblas:
@@ -17,6 +20,9 @@ If you need to install GPU version (cuda7.5_cudnn5_avx_openblas), run:
 
 For more details about installation and build: :ref:`install_steps` .
 
+Quick Use
+---------
+
 Create a new file called housing.py, and paste this Python
 code:
 

doc/howto/cluster/index_cn.rst

@@ -1,10 +1,22 @@
 分布式训练
 ==========
 
+本节将介绍如何使用PaddlePaddle在不同的集群框架下完成分布式训练。分布式训练架构如下图所示：
+
+.. image:: src/ps_cn.png
+   :width: 500
+
+- 数据分片（Data shard): 用于训练神经网络的数据，被切分成多个部分，每个部分分别给每个trainer使用。
+- 计算节点（Trainer）: 每个trainer启动后读取切分好的一部分数据，开始神经网络的“前馈”和“后馈”计算，并和参数服务器通信。在完成一定量数据的训练后，上传计算得出的梯度（gradients），然后下载优化更新后的神经网络参数（parameters）。
+- 参数服务器（Parameter server）:每个参数服务器只保存整个神经网络所有参数的一部分。参数服务器接收从计算节点上传的梯度，并完成参数优化更新，再将更新后的参数下发到每个计算节点。
+
+这样，通过计算节点和参数服务器的分布式协作，可以完成神经网络的SGD方法的训练。PaddlePaddle可以同时支持同步随机梯度下降（SGD）和异步随机梯度下降。
+
+在使用同步SGD训练神经网络时，PaddlePaddle使用同步屏障（barrier），使梯度的提交和参数的更新按照顺序方式执行。在异步SGD中，则并不会等待所有trainer提交梯度才更新参数，这样极大地提高了计算的并行性：参数服务器之间不相互依赖，并行地接收梯度和更新参数，参数服务器也不会等待计算节点全部都提交梯度之后才开始下一步，计算节点之间也不会相互依赖，并行地执行模型的训练。可以看出，虽然异步SGD方式会提高参数更新并行度, 但是并不能保证参数同步更新，在任意时间某一台参数服务器上保存的参数可能比另一台要更新，与同步SGD相比，梯度会有噪声。
+
 ..  toctree::
   :maxdepth: 1
 
-  introduction_cn.md
   preparations_cn.md
   cmd_argument_cn.md
   multi_cluster/index_cn.rst

doc/howto/cluster/index_en.rst

@@ -1,10 +1,22 @@
 Distributed Training
 ====================
 
+In this section, we'll explain how to run distributed training jobs with PaddlePaddle on different types of clusters. The diagram below shows the main architecture of a distributed trainning job:
+
+.. image:: src/ps_en.png
+   :width: 500
+
+- Data shard: training data will be split into multiple partitions, trainers use the partitions of the whole dataset to do the training job.
+- Trainer: each trainer reads the data shard, and train the neural network. Then the trainer will upload calculated "gradients" to parameter servers, and wait for parameters to be optimized on the parameter server side. When that finishes, the trainer download optimized parameters and continues its training.
+- Parameter server: every parameter server stores part of the whole neural network model data. They will do optimization calculations when gradients are uploaded from trainers, and then send updated parameters to trainers.
+
+PaddlePaddle can support both synchronize stochastic gradient descent (SGD) and asynchronous SGD.
+
+When training with synchronize SGD, PaddlePaddle uses an internal "synchronize barrier" which makes gradients update and parameter download in strict order. On the other hand, asynchronous SGD won't wait for all trainers to finish upload at a single step, this will increase the parallelism of distributed training: parameter servers do not depend on each other, they'll do parameter optimization concurrently. Parameter servers will not wait for trainers, so trainers will also do their work concurrently. But asynchronous SGD will introduce more randomness and noises in the gradient.
+
 ..  toctree::
   :maxdepth: 1
 
-  introduction_en.md
   preparations_en.md
   cmd_argument_en.md
   multi_cluster/index_en.rst

doc/howto/cluster/introduction_cn.md

@@ -1,13 +0,0 @@
-## 概述
-
-本节将介绍如何使用PaddlePaddle在不同的集群框架下完成分布式训练。分布式训练架构如下图所示：
-
-<img src="https://user-images.githubusercontent.com/13348433/31772175-5f419eca-b511-11e7-9db7-5231fe3d9ccb.png" width="500">
-
-- 数据分片（Data shard): 用于训练神经网络的数据，被切分成多个部分，每个部分分别给每个trainer使用。
-- 计算节点（Trainer）: 每个trainer启动后读取切分好的一部分数据，开始神经网络的“前馈”和“后馈”计算，并和参数服务器通信。在完成一定量数据的训练后，上传计算得出的梯度（gradients），然后下载优化更新后的神经网络参数（parameters）。
-- 参数服务器（Parameter server）:每个参数服务器只保存整个神经网络所有参数的一部分。参数服务器接收从计算节点上传的梯度，并完成参数优化更新，再将更新后的参数下发到每个计算节点。
-
-这样，通过计算节点和参数服务器的分布式协作，可以完成神经网络的SGD方法的训练。PaddlePaddle可以同时支持同步随机梯度下降（SGD）和异步随机梯度下降。
-
-在使用同步SGD训练神经网络时，PaddlePaddle使用同步屏障（barrier），使梯度的提交和参数的更新按照顺序方式执行。在异步SGD中，则并不会等待所有trainer提交梯度才更新参数，这样极大地提高了计算的并行性：参数服务器之间不相互依赖，并行地接收梯度和更新参数，参数服务器也不会等待计算节点全部都提交梯度之后才开始下一步，计算节点之间也不会相互依赖，并行地执行模型的训练。可以看出，虽然异步SGD方式会提高参数更新并行度, 但是并不能保证参数同步更新，在任意时间某一台参数服务器上保存的参数可能比另一台要更新，与同步SGD相比，梯度会有噪声。

doc/howto/cluster/introduction_en.md

@@ -1,13 +0,0 @@
-## Introduction
-
-In this section, we'll explain how to run distributed training jobs with PaddlePaddle on different types of clusters. The diagram below shows the main architecture of a distributed trainning job:
-
-<img src="https://user-images.githubusercontent.com/13348433/31772146-41523d84-b511-11e7-8a12-a69fd136c283.png" width="500">
-
-- Data shard: training data will be split into multiple partitions, trainers use the partitions of the whole dataset to do the training job.
-- Trainer: each trainer reads the data shard, and train the neural network. Then the trainer will upload calculated "gradients" to parameter servers, and wait for parameters to be optimized on the parameter server side. When that finishes, the trainer download optimized parameters and continues its training.
-- Parameter server: every parameter server stores part of the whole neural network model data. They will do optimization calculations when gradients are uploaded from trainers, and then send updated parameters to trainers.
-
-PaddlePaddle can support both synchronize stochastic gradient descent (SGD) and asynchronous SGD.
-
-When training with synchronize SGD, PaddlePaddle uses an internal "synchronize barrier" which makes gradients update and parameter download in strict order. On the other hand, asynchronous SGD won't wait for all trainers to finish upload at a single step, this will increase the parallelism of distributed training: parameter servers do not depend on each other, they'll do parameter optimization concurrently. Parameter servers will not wait for trainers, so trainers will also do their work concurrently. But asynchronous SGD will introduce more randomness and noises in the gradient.

doc/howto/rnn/index_cn.rst

@@ -1,4 +1,4 @@
-RNN相关模型
+RNN模型
 ===========
 
 ..  toctree::

paddle/framework/block_desc.cc

@@ -162,9 +162,8 @@ BlockDesc::BlockDesc(const BlockDesc &other, proto::BlockDesc *desc,
     : prog_(prog), desc_(desc) {
   need_update_ = true;
   for (auto &op : other.ops_) {
-    ops_.emplace_back(new OpDesc(*op, this));
+    ops_.emplace_back(new OpDesc(*op->Proto(), prog, this));
   }
-
   for (auto &it : other.vars_) {
     auto *var = new VarDesc(*it.second);
     vars_[it.first].reset(var);

paddle/framework/channel.h

@@ -1,4 +1,4 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* 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.

paddle/framework/details/buffered_channel.h

@@ -1,4 +1,4 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* 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.
@@ -25,6 +25,14 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+// Four of the properties of Buffered Channel:
+// - A send to a full channel blocks temporarily until a receive from the
+// channel or the channel is closed.
+// - A receive from an empty channel blocks temporarily until a send to the
+// channel or the channel is closed.
+// - A send to a closed channel returns false immediately.
+// - A receive from a closed channel returns false immediately.
+
 template <typename T>
 class Buffered : public paddle::framework::Channel<T> {
   friend Channel<T>* paddle::framework::MakeChannel<T>(size_t);
@@ -42,8 +50,11 @@ class Buffered : public paddle::framework::Channel<T> {
   std::mutex mu_;
   std::condition_variable empty_cond_var_;
   std::condition_variable full_cond_var_;
+  std::condition_variable destructor_cond_var_;
   std::deque<T> channel_;
   std::atomic<bool> closed_{false};
+  std::atomic<unsigned> send_ctr{0};
+  std::atomic<unsigned> recv_ctr{0};
 
   Buffered(size_t cap) : cap_(cap), closed_(false) {
     PADDLE_ENFORCE_GT(cap, 0);
@@ -58,6 +69,7 @@ bool Buffered<T>::Send(T* item) {
   if (closed_) {
     return ret;
   }
+  send_ctr++;
   std::unique_lock<std::mutex> lock(mu_);
   full_cond_var_.wait(lock,
                       [this]() { return channel_.size() < cap_ || closed_; });
@@ -67,20 +79,30 @@ bool Buffered<T>::Send(T* item) {
     empty_cond_var_.notify_one();
     ret = true;
   }
+  send_ctr--;
+  destructor_cond_var_.notify_one();
   return ret;
 }
 
 template <typename T>
 bool Buffered<T>::Receive(T* item) {
+  bool ret = false;
+  // Once the channel has been closed and all data has been consumed,
+  // just return false. Don't even try acquiring the mutex.
+  if (closed_ && channel_.empty()) {
+    return false;
+  }
+  recv_ctr++;
   std::unique_lock<std::mutex> lock(mu_);
   empty_cond_var_.wait(lock, [this]() { return !channel_.empty() || closed_; });
-  bool ret = false;
   if (!channel_.empty()) {
     *item = std::move(channel_.front());
     channel_.pop_front();
     full_cond_var_.notify_one();
     ret = true;
   }
+  recv_ctr--;
+  destructor_cond_var_.notify_one();
   return ret;
 }
 
@@ -100,6 +122,12 @@ Buffered<T>::~Buffered() {
   closed_ = true;
   channel_.clear();
   NotifyAllParticipants(&lock);
+
+  // The destructor must wait for all readers and writers to complete their task
+  // The channel has been closed, so we will not accept new readers and writers
+  lock.lock();
+  destructor_cond_var_.wait(
+      lock, [this]() { return send_ctr == 0 && recv_ctr == 0; });
 }
 
 template <typename T>

paddle/framework/details/cow_ptr.h

@@ -1,4 +1,4 @@
-/* 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.

paddle/framework/details/cow_ptr_test.cc

@@ -1,4 +1,4 @@
-/* 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.

paddle/framework/details/op_registry.h

@@ -1,4 +1,4 @@
-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+/* Copyright (c) 2017 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.

paddle/framework/details/unbuffered_channel.h

@@ -23,6 +23,13 @@ namespace paddle {
 namespace framework {
 namespace details {
 
+// Four of the properties of UnBuffered Channel:
+// - A send to a channel blocks temporarily until a receive from the
+// channel or the channel is closed.
+// - A receive from a channel blocks temporarily until a send to the
+// channel or the channel is closed.
+// - A send to a closed channel returns false immediately.
+// - A receive from a closed channel returns false immediately.
 template <typename T>
 class UnBuffered : public paddle::framework::Channel<T> {
   friend Channel<T>* paddle::framework::MakeChannel<T>(size_t);
@@ -45,9 +52,11 @@ class UnBuffered : public paddle::framework::Channel<T> {
   // A transaction occurs only when both are true
   std::atomic<bool> reader_found_{false}, writer_found_{false};
   std::condition_variable cv_channel_;
-  std::condition_variable_any cv_reader_, cv_writer_;
+  std::condition_variable_any cv_reader_, cv_writer_, cv_destructor_;
   T* item{nullptr};
   std::atomic<bool> closed_{false};
+  std::atomic<unsigned> send_ctr{0};
+  std::atomic<unsigned> recv_ctr{0};
 
   UnBuffered() : closed_(false) {}
 
@@ -62,6 +71,7 @@ bool UnBuffered<T>::Send(T* data) {
   if (closed_) {
     return ret;
   }
+  send_ctr++;
   // Prevent other writers from entering
   std::unique_lock<std::recursive_mutex> writer_lock(mu_write_);
   writer_found_ = true;
@@ -81,6 +91,8 @@ bool UnBuffered<T>::Send(T* data) {
     ret = true;
   }
   writer_found_ = false;
+  send_ctr--;
+  cv_destructor_.notify_one();
   return ret;
 }
 
@@ -88,6 +100,12 @@ bool UnBuffered<T>::Send(T* data) {
 // data that was sent by a writer is read from a reader.
 template <typename T>
 bool UnBuffered<T>::Receive(T* data) {
+  bool ret = false;
+  // If channel is closed, we don't even want any reader to enter.
+  // Unlike a buffered channel, an unbuffered channel does not allow
+  // readers to read after closing because there is no buffer to be consumed.
+  if (closed_) return ret;
+  recv_ctr++;
   // Prevent other readers from entering
   std::unique_lock<std::recursive_mutex> read_lock{mu_read_};
   reader_found_ = true;
@@ -96,7 +114,6 @@ bool UnBuffered<T>::Receive(T* data) {
   cv_reader_.wait(cv_lock,
                   [this]() { return writer_found_ == true || closed_; });
   cv_writer_.notify_one();
-  bool ret = false;
   if (!closed_) {
     std::unique_lock<std::mutex> lock_ch{mu_ch_};
     // Reader should wait for the writer to first write its data
@@ -110,6 +127,8 @@ bool UnBuffered<T>::Receive(T* data) {
     cv_channel_.notify_one();
   }
   reader_found_ = false;
+  recv_ctr--;
+  cv_destructor_.notify_one();
   return ret;
 }
 
@@ -135,6 +154,9 @@ UnBuffered<T>::~UnBuffered() {
   item = nullptr;
   closed_ = true;
   NotifyAllParticipants(&lock);
+  lock.lock();
+  cv_destructor_.wait(lock,
+                      [this]() { return send_ctr == 0 && recv_ctr == 0; });
 }
 
 // This function notifies all the readers, writers and

paddle/framework/op_desc.cc

@@ -125,11 +125,10 @@ OpDesc::OpDesc(const proto::OpDesc &desc, ProgramDesc *prog, BlockDesc *block)
   // restore attrs_
   for (const proto::OpDesc::Attr &attr : desc_.attrs()) {
     std::string attr_name = attr.name();
+    // The sub_block referred to by the BLOCK attr hasn't been added
+    // to ProgramDesc class yet, we skip setting BLOCK attr here.
     if (attr.type() != proto::AttrType::BLOCK) {
       attrs_[attr_name] = GetAttrValue(attr);
-    } else {
-      auto bid = attr.block_idx();
-      attrs_[attr_name] = prog->MutableBlock(bid);
     }
   }
   this->block_ = block;

paddle/framework/program_desc.cc

@@ -43,11 +43,20 @@ ProgramDesc::ProgramDesc() {
 
 ProgramDesc::ProgramDesc(const ProgramDesc &o) {
   desc_ = o.desc_;
-
   for (int i = 0; i < desc_.blocks_size(); ++i) {
     auto *block = desc_.mutable_blocks(i);
     blocks_.emplace_back(new BlockDesc(*o.blocks_[i], block, this));
   }
+  for (auto &block : blocks_) {
+    for (auto *op : block->AllOps()) {
+      for (const auto &attr : op->Proto()->attrs()) {
+        if (attr.type() == proto::AttrType::BLOCK) {
+          size_t blk_idx = attr.block_idx();
+          op->SetBlockAttr(attr.name(), *this->MutableBlock(blk_idx));
+        }
+      }
+    }
+  }
 }
 
 ProgramDesc::ProgramDesc(const proto::ProgramDesc &desc) {
@@ -55,6 +64,16 @@ ProgramDesc::ProgramDesc(const proto::ProgramDesc &desc) {
   for (auto &block_desc : *desc_.mutable_blocks()) {
     blocks_.emplace_back(new BlockDesc(this, &block_desc));
   }
+  for (auto &block : blocks_) {
+    for (auto *op : block->AllOps()) {
+      for (const auto &attr : op->Proto()->attrs()) {
+        if (attr.type() == proto::AttrType::BLOCK) {
+          size_t blk_idx = attr.block_idx();
+          op->SetBlockAttr(attr.name(), *this->MutableBlock(blk_idx));
+        }
+      }
+    }
+  }
 }
 
 ProgramDesc::ProgramDesc(const std::string &binary_str) {

paddle/framework/prune.cc

@@ -49,11 +49,28 @@ bool IsTarget(const proto::OpDesc& op_desc) {
   return false;
 }
 
-void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
-                int block_id) {
-  // TODO(tonyyang-svail):
-  //    - will change to use multiple blocks for RNN op and Cond Op
+int GetSubBlockIndex(const proto::OpDesc& op_desc) {
+  for (auto& attr : op_desc.attrs()) {
+    if (attr.type() == proto::AttrType::BLOCK) {
+      PADDLE_ENFORCE(attr.has_block_idx());
+      return attr.block_idx();
+    }
+  }
+  return -1;
+}
+
+bool HasSubBlock(const proto::OpDesc& op_desc) {
+  return GetSubBlockIndex(op_desc) > 0;
+}
 
+// block_id is the idx of the current block in the input desc
+// parent_block_id is the idx of the parent of the current block
+// in the output desc, -1 means the current block is global block
+// dependent_vars is passed recursively from the parent block to
+// the child block to help pruning
+void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
+                int block_id, int parent_block_id,
+                std::set<std::string>& dependent_vars) {
   auto& block = input.blocks(block_id);
   auto& ops = block.ops();
 
@@ -72,11 +89,9 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
     expect_fetch = (op_desc.type() == kFetchOpType);
   }
 
-  std::set<std::string> dependent_vars;
   std::vector<bool> should_run;
   for (auto op_iter = ops.rbegin(); op_iter != ops.rend(); ++op_iter) {
     auto& op_desc = *op_iter;
-
     if (IsTarget(op_desc) || HasDependentVar(op_desc, dependent_vars)) {
       // insert its input to the dependency graph
       for (auto& var : op_desc.inputs()) {
@@ -84,7 +99,6 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
           dependent_vars.insert(argu);
         }
       }
-
       should_run.push_back(true);
     } else {
       should_run.push_back(false);
@@ -95,45 +109,81 @@ void prune_impl(const proto::ProgramDesc& input, proto::ProgramDesc* output,
   // we reverse the should_run vector
   std::reverse(should_run.begin(), should_run.end());
 
-  *output = input;
-  auto* op_field = output->mutable_blocks(block_id)->mutable_ops();
+  // copy the current block from input to output
+  auto* block_field = output->mutable_blocks();
+  *block_field->Add() = input.blocks(block_id);
+
+  int output_block_id = output->blocks_size() - 1;
+  auto* output_block = output->mutable_blocks(output_block_id);
+  output_block->set_idx(output_block_id);
+  output_block->set_parent_idx(parent_block_id);
+
+  auto* op_field = output_block->mutable_ops();
   op_field->Clear();
   for (size_t i = 0; i < should_run.size(); ++i) {
     if (should_run[i]) {
-      *op_field->Add() = input.blocks(block_id).ops(i);
+      auto* op = op_field->Add();
+      *op = input.blocks(block_id).ops(i);
+      if (HasSubBlock(*op)) {
+        // create sub_block_dependent_vars here to help prune the sub block
+        std::set<std::string> sub_block_dependent_vars;
+        for (auto& var : op->inputs()) {
+          for (auto& argu : var.arguments()) {
+            sub_block_dependent_vars.insert(argu);
+          }
+        }
+        for (auto& var : op->outputs()) {
+          for (auto& argu : var.arguments()) {
+            sub_block_dependent_vars.insert(argu);
+          }
+        }
+        // GetSubBlockIndex(*op) is the idx of the sub_block in the input desc
+        // output_block_id is the idx of the current block in the output desc
+        prune_impl(input, output, GetSubBlockIndex(*op), output_block_id,
+                   sub_block_dependent_vars);
+      }
     }
   }
 
   // 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();
+  auto* var_field = output->mutable_blocks(output_block_id)->mutable_vars();
   for (const auto& var : *var_field) {
     var_map[var.name()] = var;
   }
 
-  var_field->Clear();
+  std::set<std::string> var_names;
   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];
+        if (var_map.count(arg) != 0) {
+          var_names.insert(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];
+        if (var_map.count(arg) != 0) {
+          var_names.insert(arg);
+        }
       }
     }
   }
+
+  var_field->Clear();
+  for (const auto& name : var_names) {
+    *var_field->Add() = var_map[name];
+  }
 }
 
 // TODO(fengjiayi): Prune() could be inplaced to avoid unnecessary copies
 void Prune(const proto::ProgramDesc& input, proto::ProgramDesc* output) {
-  prune_impl(input, output, 0);
+  std::set<std::string> dependent_vars;
+  output->clear_blocks();
+  prune_impl(input, output, 0, -1, dependent_vars);
 }
 
 void inference_optimize_impl(const proto::ProgramDesc& input,

paddle/inference/io.cc

@@ -21,6 +21,17 @@ limitations under the License. */
 namespace paddle {
 namespace inference {
 
+void ReadBinaryFile(const std::string& filename, std::string& contents) {
+  VLOG(3) << "loading model from " << filename;
+  std::ifstream inputfs(filename, std::ios::in | std::ios::binary);
+  inputfs.seekg(0, std::ios::end);
+  contents.clear();
+  contents.resize(inputfs.tellg());
+  inputfs.seekg(0, std::ios::beg);
+  inputfs.read(&contents[0], contents.size());
+  inputfs.close();
+}
+
 bool IsParameter(const framework::VarDesc* var,
                  const framework::ProgramDesc& main_program) {
   if (var->Persistable()) {
@@ -44,12 +55,15 @@ bool IsParameter(const framework::VarDesc* var,
 
 void LoadPersistables(framework::Executor& executor,
                       framework::Scope& scope,
+                      const framework::ProgramDesc& main_program,
                       const std::string& dirname,
-                      const framework::ProgramDesc& main_program) {
+                      const std::string& param_filename) {
   const framework::BlockDesc& global_block = main_program.Block(0);
 
   framework::ProgramDesc* load_program = new framework::ProgramDesc();
   framework::BlockDesc* load_block = load_program->MutableBlock(0);
+  std::vector<std::string> paramlist;
+
   for (auto* var : global_block.AllVars()) {
     if (IsParameter(var, main_program)) {
       VLOG(3) << "parameter's name: " << var->Name();
@@ -61,15 +75,33 @@ void LoadPersistables(framework::Executor& executor,
       new_var->SetLoDLevel(var->GetLoDLevel());
       new_var->SetPersistable(true);
 
-      // append_op
-      framework::OpDesc* op = load_block->AppendOp();
-      op->SetType("load");
-      op->SetOutput("Out", {new_var->Name()});
-      op->SetAttr("file_path", {dirname + "/" + new_var->Name()});
-      op->CheckAttrs();
+      if (!param_filename.empty()) {
+        paramlist.push_back(new_var->Name());
+      } else {
+        // append_op
+        framework::OpDesc* op = load_block->AppendOp();
+        op->SetType("load");
+        op->SetOutput("Out", {new_var->Name()});
+        op->SetAttr("file_path", {dirname + "/" + new_var->Name()});
+        op->CheckAttrs();
+      }
     }
   }
+
+  if (!param_filename.empty()) {
+    // sort paramlist to have consistent ordering
+    std::sort(paramlist.begin(), paramlist.end());
+    // append just the load_combine op
+    framework::OpDesc* op = load_block->AppendOp();
+    op->SetType("load_combine");
+    op->SetOutput("Out", paramlist);
+    op->SetAttr("file_path", {param_filename});
+    op->CheckAttrs();
+  }
+
   executor.Run(*load_program, &scope, 0, true, true);
+
+  VLOG(3) << "Ran loading successfully";
   delete load_program;
 }
 
@@ -77,20 +109,29 @@ std::unique_ptr<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();
+  ReadBinaryFile(model_filename, program_desc_str);
+
+  std::unique_ptr<framework::ProgramDesc> main_program(
+      new framework::ProgramDesc(program_desc_str));
+
+  LoadPersistables(executor, scope, *main_program, dirname, "");
+  return main_program;
+}
+
+std::unique_ptr<framework::ProgramDesc> Load(
+    framework::Executor& executor,
+    framework::Scope& scope,
+    const std::string& prog_filename,
+    const std::string& param_filename) {
+  std::string model_filename = prog_filename;
+  std::string program_desc_str;
+  ReadBinaryFile(model_filename, program_desc_str);
 
   std::unique_ptr<framework::ProgramDesc> main_program(
       new framework::ProgramDesc(program_desc_str));
 
-  LoadPersistables(executor, scope, dirname, *main_program);
+  LoadPersistables(executor, scope, *main_program, "", param_filename);
   return main_program;
 }
 

paddle/inference/io.h

@@ -26,12 +26,18 @@ namespace inference {
 
 void LoadPersistables(framework::Executor& executor,
                       framework::Scope& scope,
+                      const framework::ProgramDesc& main_program,
                       const std::string& dirname,
-                      const framework::ProgramDesc& main_program);
+                      const std::string& param_filename);
 
 std::unique_ptr<framework::ProgramDesc> Load(framework::Executor& executor,
                                              framework::Scope& scope,
                                              const std::string& dirname);
 
+std::unique_ptr<framework::ProgramDesc> Load(framework::Executor& executor,
+                                             framework::Scope& scope,
+                                             const std::string& prog_filename,
+                                             const std::string& param_filename);
+
 }  // namespace inference
 }  // namespace paddle

paddle/inference/tests/book/CMakeLists.txt

@@ -27,3 +27,4 @@ endfunction(inference_test)
 inference_test(recognize_digits ARGS mlp)
 inference_test(image_classification ARGS vgg resnet)
 inference_test(label_semantic_roles)
+inference_test(rnn_encoder_decoder)

paddle/inference/tests/book/test_helper.h

@@ -67,17 +67,28 @@ void CheckError(paddle::framework::LoDTensor& output1,
   EXPECT_EQ(count, 0) << "There are " << count << " different elements.";
 }
 
-template <typename Place, typename T>
+template <typename Place, typename T, bool IsCombined = false>
 void TestInference(const std::string& dirname,
                    const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
                    std::vector<paddle::framework::LoDTensor*>& cpu_fetchs) {
-  // 1. Define place, executor and scope
+  // 1. Define place, executor, scope and inference_program
   auto place = Place();
   auto executor = paddle::framework::Executor(place);
   auto* scope = new paddle::framework::Scope();
+  std::unique_ptr<paddle::framework::ProgramDesc> inference_program;
 
   // 2. Initialize the inference_program and load all parameters from file
-  auto inference_program = paddle::inference::Load(executor, *scope, dirname);
+  if (IsCombined) {
+    // Hard-coding the names for combined params case
+    std::string prog_filename = "__model_combined__";
+    std::string param_filename = "__params_combined__";
+    inference_program = paddle::inference::Load(executor,
+                                                *scope,
+                                                dirname + "/" + prog_filename,
+                                                dirname + "/" + param_filename);
+  } else {
+    inference_program = paddle::inference::Load(executor, *scope, dirname);
+  }
 
   // 3. Get the feed_target_names and fetch_target_names
   const std::vector<std::string>& feed_target_names =

paddle/inference/tests/book/test_inference_recognize_digits.cc

@@ -59,3 +59,45 @@ TEST(inference, recognize_digits) {
   CheckError<float>(output1, output2);
 #endif
 }
+
+TEST(inference, recognize_digits_combine) {
+  if (FLAGS_dirname.empty()) {
+    LOG(FATAL) << "Usage: ./example --dirname=path/to/your/model";
+  }
+
+  LOG(INFO) << "FLAGS_dirname: " << FLAGS_dirname << std::endl;
+  std::string dirname = FLAGS_dirname;
+
+  // 0. Call `paddle::framework::InitDevices()` initialize all the devices
+  // In unittests, this is done in paddle/testing/paddle_gtest_main.cc
+
+  paddle::framework::LoDTensor input;
+  // Use normilized image pixels as input data,
+  // which should be in the range [-1.0, 1.0].
+  SetupTensor<float>(
+      input, {1, 28, 28}, static_cast<float>(-1), static_cast<float>(1));
+  std::vector<paddle::framework::LoDTensor*> cpu_feeds;
+  cpu_feeds.push_back(&input);
+
+  paddle::framework::LoDTensor output1;
+  std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
+  cpu_fetchs1.push_back(&output1);
+
+  // Run inference on CPU
+  TestInference<paddle::platform::CPUPlace, float, true>(
+      dirname, cpu_feeds, cpu_fetchs1);
+  LOG(INFO) << output1.dims();
+
+#ifdef PADDLE_WITH_CUDA
+  paddle::framework::LoDTensor output2;
+  std::vector<paddle::framework::LoDTensor*> cpu_fetchs2;
+  cpu_fetchs2.push_back(&output2);
+
+  // Run inference on CUDA GPU
+  TestInference<paddle::platform::CUDAPlace, float, true>(
+      dirname, cpu_feeds, cpu_fetchs2);
+  LOG(INFO) << output2.dims();
+
+  CheckError<float>(output1, output2);
+#endif
+}