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

CONTRIBUTING.md

@@ -58,6 +58,8 @@ PaddlePaddle uses this [Git branching model](http://nvie.com/posts/a-successful-
     create mode 100644 233
    ```
 
+	NOTE: The `yapf` installed by `pip install pre-commit` and `conda install -c conda-forge pre-commit` is slightly different. Paddle developers use `pip install pre-commit`.
+
 1. Build and test
 
    Users can build PaddlePaddle natively on Linux and Mac OS X.  But to unify the building environment and to make it easy for debugging, the recommended way is [using Docker](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/dev/build_en.md).

Dockerfile

@@ -29,7 +29,7 @@ RUN apt-get update && \
     wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
     curl sed grep graphviz libjpeg-dev zlib1g-dev  \
     python-matplotlib gcc-4.8 g++-4.8 \
-    automake locales clang-format swig doxygen cmake  \
+    automake locales clang-format swig cmake  \
     liblapack-dev liblapacke-dev \
     clang-3.8 llvm-3.8 libclang-3.8-dev \
     net-tools libtool ccache && \

benchmark/fluid/fluid_benchmark.py

@@ -98,6 +98,8 @@ def parse_args():
         '--use_fake_data',
         action='store_true',
         help='If set ommit the actual read data operators.')
+    parser.add_argument(
+        '--profile', action='store_true', help='If set, profile a few steps.')
     parser.add_argument(
         '--update_method',
         type=str,
@@ -108,8 +110,8 @@ def parse_args():
     return args
 
 
-def append_nccl2_prepare():
-    if os.getenv("PADDLE_TRAINER_ID", None) != None:
+def append_nccl2_prepare(trainer_id):
+    if trainer_id >= 0:
         # append gen_nccl_id at the end of startup program
         trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
         port = os.getenv("PADDLE_PSERVER_PORT")
@@ -136,12 +138,12 @@ def append_nccl2_prepare():
             })
         return nccl_id_var, num_trainers, trainer_id
     else:
-        raise Exception(
-            "must set PADDLE_TRAINER_ID env variables for dist train.")
+        raise Exception("must set positive PADDLE_TRAINER_ID env variables for "
+                        "nccl-based dist train.")
 
 
-def dist_transpile():
-    if "PADDLE_TRAINING_ROLE" not in os.environ:
+def dist_transpile(trainer_id):
+    if trainer_id < 0:
         return None, None
 
     # the port of all pservers, needed by both trainer and pserver
@@ -158,9 +160,6 @@ def dist_transpile():
     trainers = int(os.getenv("PADDLE_TRAINERS"))
     # the IP of the local machine, needed by pserver only
     current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port
-    # the unique trainer id, starting from 0, needed by trainer
-    # only
-    trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
     # the role, should be either PSERVER or TRAINER
     training_role = os.getenv("PADDLE_TRAINING_ROLE")
 
@@ -295,6 +294,11 @@ def train_parallel(avg_loss, infer_prog, optimizer, train_reader, test_reader,
         iters = 0
         start_time = time.time()
         for batch_id, data in enumerate(train_reader()):
+            if args.profile and pass_id == 0 and batch_id == 5:
+                profiler.start_profiler("All")
+            elif args.profile and pass_id == 0 and batch_id == 10:
+                profiler.stop_profiler("total", "/tmp/profile_%d" % trainer_id)
+
             if iters == args.skip_batch_num:
                 start_time = time.time()
                 num_samples = 0
@@ -334,7 +338,11 @@ def print_arguments(args):
 def main():
     args = parse_args()
     print_arguments(args)
-    nccl_id_var, num_trainers, trainer_id = None, 1, 0
+
+    # the unique trainer id, starting from 0, needed by trainer
+    # only
+    nccl_id_var, num_trainers, trainer_id = (
+        None, 1, int(os.getenv("PADDLE_TRAINER_ID", "-1")))
 
     if args.use_cprof:
         pr = cProfile.Profile()
@@ -348,7 +356,7 @@ def main():
         fluid.memory_optimize(fluid.default_main_program())
 
     if args.update_method == "pserver":
-        train_prog, startup_prog = dist_transpile()
+        train_prog, startup_prog = dist_transpile(trainer_id)
         if not train_prog:
             raise Exception(
                 "Must configure correct environments to run dist train.")
@@ -364,7 +372,7 @@ def main():
     train_args.append(fluid.default_startup_program())
 
     if args.update_method == "nccl2":
-        nccl_id_var, num_trainers, trainer_id = append_nccl2_prepare()
+        nccl_id_var, num_trainers, trainer_id = append_nccl2_prepare(trainer_id)
     if args.gpus == 1:
         # NOTE: parallel executor use profiler interanlly
         if args.use_nvprof and args.device == 'GPU':

benchmark/fluid/kube_gen_job.py

@@ -49,7 +49,7 @@ def parse_args():
     parser.add_argument(
         '--fluid', default=1, type=int, help='whether is fluid job')
     parser.add_argument(
-        '--rdma', action='store_ture', help='whether mount rdma libs')
+        '--rdma', action='store_true', help='whether mount rdma libs')
     parser.add_argument(
         '--disttype',
         default="pserver",

doc/fluid/getstarted/Developer's_Guide_to_Paddle_Fluid.md

@@ -86,7 +86,7 @@
 <br>
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/fluid_compiler.png" width=100%>
+  <img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/fluid-compiler.png" width=100%>
 </p>
 
 ---

doc/fluid/getstarted/index_cn.rst

@@ -17,3 +17,4 @@
   :maxdepth: 1
 
   concepts/use_concepts_cn.rst
+  developer's_guide_to_paddle_fluid.md

doc/fluid/getstarted/index_en.rst

@@ -16,3 +16,4 @@ Here is an example of linear regression. It introduces workflow of PaddlePaddle,
   :maxdepth: 1
 
   concepts/index_en.rst
+  developer's_guide_to_paddle_fluid.md

doc/fluid/getstarted/quickstart_cn.rst

@@ -11,7 +11,7 @@ PaddlePaddle支持使用pip快速安装，目前支持CentOS 6以上, Ubuntu 14.
 
      pip install paddlepaddle
 
-如果需要安装支持GPU的版本（cuda7.5_cudnn5_avx_openblas），需要执行：
+如果需要安装支持GPU的版本（cuda8.0_cudnn5_avx_openblas），需要执行：
 
   .. code-block:: bash
 

doc/fluid/getstarted/quickstart_en.rst

@@ -12,7 +12,7 @@ Simply run the following command to install, the version is cpu_avx_openblas:
 
      pip install paddlepaddle
 
-If you need to install GPU version (cuda7.5_cudnn5_avx_openblas), run:
+If you need to install GPU version (cuda8.0_cudnn5_avx_openblas), run:
 
   .. code-block:: bash
 

doc/v2/dev/contribute_to_paddle_cn.md

@@ -51,6 +51,8 @@ Paddle 开发人员使用 [pre-commit](http://pre-commit.com/) 工具来管理 G
 
 Paddle 使用 `clang-format` 来调整 C/C++ 源代码格式，请确保 `clang-format` 版本在 3.8 以上。
 
+注：通过`pip install pre-commit`和`conda install -c conda-forge pre-commit`安装的`yapf`稍有不同的，Paddle 开发人员使用的是`pip install pre-commit`。
+
 ## 开始开发
 
 在本例中，我删除了 README.md 中的一行，并创建了一个新文件。

paddle/contrib/inference/CMakeLists.txt

@@ -13,7 +13,11 @@
 # limitations under the License.
 #
 
-function(inference_api_test TARGET_NAME TEST_SRC DEP_TEST)
+if(APPLE)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
+endif(APPLE)
+
+function(inference_api_test TARGET_NAME TEST_SRC)
     set(options "")
     set(oneValueArgs "")
     set(multiValueArgs ARGS)
@@ -32,8 +36,10 @@ function(inference_api_test TARGET_NAME TEST_SRC DEP_TEST)
         string(REGEX REPLACE "^_$" "" arg "${arg}")
         cc_test(${TARGET_NAME}
                 SRCS ${TEST_SRC}
-                DEPS paddle_fluid_api paddle_inference_api paddle_inference_api_impl
+                DEPS paddle_fluid_api paddle_inference_api
                 ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
+        # TODO(panyx0178): Figure out how to add word2vec and image_classification
+        # as deps.
         # set_tests_properties(${TARGET_NAME}
         #         PROPERTIES DEPENDS ${DEP_TEST})
     endforeach()
@@ -41,17 +47,12 @@ endfunction(inference_api_test)
 
 
 cc_library(paddle_inference_api
-    SRCS paddle_inference_api.cc
+    SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
     DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
 
-cc_library(paddle_inference_api_impl
-           SRCS paddle_inference_api_impl.cc
-           DEPS paddle_inference_api paddle_fluid_api)
-
 cc_test(test_paddle_inference_api
         SRCS test_paddle_inference_api.cc
         DEPS paddle_inference_api)
 
 inference_api_test(test_paddle_inference_api_impl
-                   test_paddle_inference_api_impl.cc
-                   test_word2vec)
+                   test_paddle_inference_api_impl.cc)

paddle/contrib/inference/paddle_inference_api.h

@@ -45,9 +45,9 @@ struct PaddleTensor {
 };
 
 /*
-* A simple Inference API for Paddle. Currently this API might just be used by
+ * A simple Inference API for Paddle. Currently this API can be used by
  * non-sequence scenerios.
-* TODO(Superjomn) Prepare another API for NLP-related usages.
+ * TODO(Superjomn) Support another API for NLP-related usages.
  */
 class PaddlePredictor {
  public:
@@ -66,34 +66,38 @@ class PaddlePredictor {
   // be thread-safe.
   virtual std::unique_ptr<PaddlePredictor> Clone() = 0;
 
-  virtual bool InitShared() { return false; }
   // Destroy the Predictor.
   virtual ~PaddlePredictor() {}
 
-  friend std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(
-      const PaddlePredictor::Config& config);
+  enum class EngineKind {
+    kNative = -1,  // Use the native Fluid facility.
+    // TODO(Superjomn) support latter.
+    // kAnakin,             // Use Anakin for inference.
+    // kTensorRT,           // Use TensorRT for inference.
+    // kAutoMixedAnakin,    // Automatically mix Fluid with Anakin.
+    // kAutoMixedTensorRT,  // Automatically mix Fluid with TensorRT.
+  };
 
   // The common configs for all the predictors.
   struct Config {
-    enum class EngineKind;
-
     std::string model_dir;      // path to the model directory.
     bool enable_engine{false};  // Enable to execute (part of) the model on
-    // third-party engines.
-    EngineKind engine_kind{Config::EngineKind::kNone};
-
-    enum class EngineKind {
-      kNone = -1,          // Use the native Fluid facility.
-      kAnakin,             // Use Anakin for inference.
-      kTensorRT,           // Use TensorRT for inference.
-      kAutoMixedAnakin,    // Automatically mix Fluid with Anakin.
-      kAutoMixedTensorRT,  // Automatically mix Fluid with TensorRT.
   };
 };
+
+struct NativeConfig : public PaddlePredictor::Config {
+  bool use_gpu{false};
+  int device;
+  float fraction_of_gpu_memory;
+  std::string prog_file;
+  std::string param_file;
+  bool share_variables;
 };
 
 // A factory to help create difference predictor.
-template <typename ConfigT>
+template <
+    typename ConfigT,
+    PaddlePredictor::EngineKind engine = PaddlePredictor::EngineKind::kNative>
 std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT& config);
 
 }  // namespace paddle

paddle/contrib/inference/paddle_inference_api_impl.cc

@@ -54,7 +54,7 @@ std::string num2str(T a) {
 }
 }  // namespace
 
-bool PaddlePredictorImpl::Init() {
+bool NativePaddlePredictor::Init() {
   VLOG(3) << "Predictor::init()";
 
   // TODO(panyx0718): Should CPU vs GPU device be decided by id?
@@ -96,14 +96,14 @@ bool PaddlePredictorImpl::Init() {
   return true;
 }
 
-bool PaddlePredictorImpl::Run(const std::vector<PaddleTensor> &inputs,
+bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
                                 std::vector<PaddleTensor> *output_data) {
   VLOG(3) << "Predictor::predict";
   Timer timer;
   timer.tic();
   // set feed variable
-  std::map<std::string, const paddle::framework::LoDTensor *> feed_targets;
-  std::vector<paddle::framework::LoDTensor> feeds;
+  std::map<std::string, const framework::LoDTensor *> feed_targets;
+  std::vector<framework::LoDTensor> feeds;
   if (!SetFeed(inputs, &feeds)) {
     LOG(ERROR) << "fail to set feed";
     return false;
@@ -112,8 +112,8 @@ bool PaddlePredictorImpl::Run(const std::vector<PaddleTensor> &inputs,
     feed_targets[feed_target_names_[i]] = &feeds[i];
   }
   // get fetch variable
-  std::map<std::string, paddle::framework::LoDTensor *> fetch_targets;
-  std::vector<paddle::framework::LoDTensor> fetchs;
+  std::map<std::string, framework::LoDTensor *> fetch_targets;
+  std::vector<framework::LoDTensor> fetchs;
   fetchs.resize(fetch_target_names_.size());
   for (size_t i = 0; i < fetch_target_names_.size(); ++i) {
     fetch_targets[fetch_target_names_[i]] = &fetchs[i];
@@ -133,76 +133,33 @@ bool PaddlePredictorImpl::Run(const std::vector<PaddleTensor> &inputs,
   return true;
 }
 
-std::unique_ptr<PaddlePredictor> PaddlePredictorImpl::Clone() {
+std::unique_ptr<PaddlePredictor> NativePaddlePredictor::Clone() {
   VLOG(3) << "Predictor::clone";
-  std::unique_ptr<PaddlePredictor> cls(new PaddlePredictorImpl(config_));
-  if (!cls->InitShared()) {
-    LOG(ERROR) << "fail to call InitShared";
+  std::unique_ptr<PaddlePredictor> cls(new NativePaddlePredictor(config_));
+
+  if (!dynamic_cast<NativePaddlePredictor *>(cls.get())->Init()) {
+    LOG(ERROR) << "fail to call Init";
     return nullptr;
   }
   // fix manylinux compile error.
   return std::move(cls);
 }
 
-// TODO(panyx0718): Consider merge with Init()?
-bool PaddlePredictorImpl::InitShared() {
-  VLOG(3) << "Predictor::init_shared";
-  // 1. Define place, executor, scope
-  if (this->config_.device >= 0) {
-    place_ = paddle::platform::CUDAPlace();
-  } else {
-    place_ = paddle::platform::CPUPlace();
-  }
-  this->executor_.reset(new paddle::framework::Executor(this->place_));
-  this->scope_.reset(new paddle::framework::Scope());
-  // Initialize the inference program
-  if (!this->config_.model_dir.empty()) {
-    // Parameters are saved in separate files sited in
-    // the specified `dirname`.
-    this->inference_program_ = paddle::inference::Load(
-        this->executor_.get(), this->scope_.get(), this->config_.model_dir);
-  } else if (!this->config_.prog_file.empty() &&
-             !this->config_.param_file.empty()) {
-    // All parameters are saved in a single file.
-    // The file names should be consistent with that used
-    // in Python API `fluid.io.save_inference_model`.
-    this->inference_program_ =
-        paddle::inference::Load(this->executor_.get(),
-                                this->scope_.get(),
-                                this->config_.prog_file,
-                                this->config_.param_file);
-  }
-  this->ctx_ = this->executor_->Prepare(*this->inference_program_, 0);
-  // 3. create variables
-  // TODO(panyx0718): why test share_variables.
-  if (config_.share_variables) {
-    this->executor_->CreateVariables(
-        *this->inference_program_, this->scope_.get(), 0);
-  }
-  // 4. Get the feed_target_names and fetch_target_names
-  this->feed_target_names_ = this->inference_program_->GetFeedTargetNames();
-  this->fetch_target_names_ = this->inference_program_->GetFetchTargetNames();
-  return true;
-}
-
-bool PaddlePredictorImpl::SetFeed(
-    const std::vector<PaddleTensor> &inputs,
-    std::vector<paddle::framework::LoDTensor> *feeds) {
+bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
+                                    std::vector<framework::LoDTensor> *feeds) {
   VLOG(3) << "Predictor::set_feed";
   if (inputs.size() != feed_target_names_.size()) {
     LOG(ERROR) << "wrong feed input size.";
     return false;
   }
   for (size_t i = 0; i < feed_target_names_.size(); ++i) {
-    paddle::framework::LoDTensor input;
-    paddle::framework::DDim ddim =
-        paddle::framework::make_ddim(inputs[i].shape);
+    framework::LoDTensor input;
+    framework::DDim ddim = framework::make_ddim(inputs[i].shape);
     void *input_ptr;
     if (inputs[i].dtype == PaddleDType::INT64) {
-      input_ptr =
-          input.mutable_data<int64_t>(ddim, paddle::platform::CPUPlace());
+      input_ptr = input.mutable_data<int64_t>(ddim, platform::CPUPlace());
     } else if (inputs[i].dtype == PaddleDType::FLOAT32) {
-      input_ptr = input.mutable_data<float>(ddim, paddle::platform::CPUPlace());
+      input_ptr = input.mutable_data<float>(ddim, platform::CPUPlace());
     } else {
       LOG(ERROR) << "unsupported feed type " << inputs[i].dtype;
       return false;
@@ -213,13 +170,12 @@ bool PaddlePredictorImpl::SetFeed(
                 inputs[i].data.data,
                 inputs[i].data.length);
     feeds->push_back(input);
-    LOG(ERROR) << "Actual feed type " << feeds->back().type().name();
   }
   return true;
 }
 
-bool PaddlePredictorImpl::GetFetch(
-    const std::vector<paddle::framework::LoDTensor> &fetchs,
+bool NativePaddlePredictor::GetFetch(
+    const std::vector<framework::LoDTensor> &fetchs,
     std::vector<PaddleTensor> *outputs) {
   VLOG(3) << "Predictor::get_fetch";
   outputs->resize(fetchs.size());
@@ -284,9 +240,12 @@ bool PaddlePredictorImpl::GetFetch(
   return true;
 }
 
-std::unique_ptr<PaddlePredictorImpl> CreatePaddlePredictorImpl(
-    const VisConfig &config) {
-  VLOG(3) << "create PaddlePredictorImpl";
+template <>
+std::unique_ptr<PaddlePredictor>
+CreatePaddlePredictor<NativeConfig, PaddlePredictor::EngineKind::kNative>(
+    const NativeConfig &config) {
+  VLOG(3) << "create NativePaddlePredictor";
+  if (config.use_gpu) {
     // 1. GPU memeroy
     std::vector<std::string> flags;
     if (config.fraction_of_gpu_memory >= 0.0f ||
@@ -298,13 +257,13 @@ std::unique_ptr<PaddlePredictorImpl> CreatePaddlePredictorImpl(
       VLOG(3) << "set flag: " << flag;
       framework::InitGflags(flags);
     }
+  }
 
-  std::unique_ptr<PaddlePredictorImpl> predictor(
-      new PaddlePredictorImpl(config));
-  if (!predictor->Init()) {
+  std::unique_ptr<PaddlePredictor> predictor(new NativePaddlePredictor(config));
+  if (!dynamic_cast<NativePaddlePredictor *>(predictor.get())->Init()) {
     return nullptr;
   }
-  return predictor;
+  return std::move(predictor);
 }
 
 }  // namespace paddle

paddle/contrib/inference/paddle_inference_api_impl.h

@@ -29,20 +29,10 @@
 
 namespace paddle {
 
-struct VisConfig : public PaddlePredictor::Config {
-  int device;
-  float fraction_of_gpu_memory;
-  std::string prog_file;
-  std::string param_file;
-  bool share_variables;
-};
-
-/*
- * Do not use this, just a demo indicating how to customize a Predictor.
- */
-class PaddlePredictorImpl : public PaddlePredictor {
+class NativePaddlePredictor : public PaddlePredictor {
  public:
-  explicit PaddlePredictorImpl(const VisConfig &config) : config_(config) {}
+  explicit NativePaddlePredictor(const NativeConfig &config)
+      : config_(config) {}
 
   bool Init();
 
@@ -51,26 +41,22 @@ class PaddlePredictorImpl : public PaddlePredictor {
 
   std::unique_ptr<PaddlePredictor> Clone() override;
 
-  ~PaddlePredictorImpl() override{};
+  ~NativePaddlePredictor() override{};
 
  private:
-  bool InitShared() override;
   bool SetFeed(const std::vector<PaddleTensor> &input_datas,
-               std::vector<paddle::framework::LoDTensor> *feeds);
-  bool GetFetch(const std::vector<paddle::framework::LoDTensor> &fetchs,
+               std::vector<framework::LoDTensor> *feeds);
+  bool GetFetch(const std::vector<framework::LoDTensor> &fetchs,
                 std::vector<PaddleTensor> *output_data);
 
-  VisConfig config_;
-  paddle::platform::Place place_;
-  std::unique_ptr<paddle::framework::Executor> executor_;
-  std::unique_ptr<paddle::framework::Scope> scope_;
-  std::unique_ptr<paddle::framework::ExecutorPrepareContext> ctx_;
-  std::unique_ptr<paddle::framework::ProgramDesc> inference_program_;
+  NativeConfig config_;
+  platform::Place place_;
+  std::unique_ptr<framework::Executor> executor_;
+  std::unique_ptr<framework::Scope> scope_;
+  std::unique_ptr<framework::ExecutorPrepareContext> ctx_;
+  std::unique_ptr<framework::ProgramDesc> inference_program_;
   std::vector<std::string> feed_target_names_;
   std::vector<std::string> fetch_target_names_;
 };
 
-std::unique_ptr<PaddlePredictorImpl> CreatePaddlePredictorImpl(
-    const VisConfig &config);
-
 }  // namespace paddle

paddle/contrib/inference/test_paddle_inference_api_impl.cc

@@ -40,16 +40,20 @@ PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
   return pt;
 }
 
-TEST(paddle_inference_api_impl, word2vec) {
-  VisConfig config;
+NativeConfig GetConfig() {
+  NativeConfig config;
   config.model_dir = FLAGS_dirname + "word2vec.inference.model";
   LOG(INFO) << "dirname  " << config.model_dir;
   config.fraction_of_gpu_memory = 0.15;
+  config.use_gpu = true;
   config.device = 0;
   config.share_variables = true;
+  return config;
+}
 
-  std::unique_ptr<PaddlePredictorImpl> predictor =
-      CreatePaddlePredictorImpl(config);
+TEST(paddle_inference_api_impl, word2vec) {
+  NativeConfig config = GetConfig();
+  auto predictor = CreatePaddlePredictor<NativeConfig>(config);
 
   framework::LoDTensor first_word, second_word, third_word, fourth_word;
   framework::LoD lod{{0, 1}};
@@ -60,24 +64,90 @@ TEST(paddle_inference_api_impl, word2vec) {
   SetupLoDTensor(&third_word, lod, static_cast<int64_t>(0), dict_size - 1);
   SetupLoDTensor(&fourth_word, lod, static_cast<int64_t>(0), dict_size - 1);
 
-  std::vector<PaddleTensor> cpu_feeds;
-  cpu_feeds.push_back(LodTensorToPaddleTensor(&first_word));
-  cpu_feeds.push_back(LodTensorToPaddleTensor(&second_word));
-  cpu_feeds.push_back(LodTensorToPaddleTensor(&third_word));
-  cpu_feeds.push_back(LodTensorToPaddleTensor(&fourth_word));
+  std::vector<PaddleTensor> paddle_tensor_feeds;
+  paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&first_word));
+  paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&second_word));
+  paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&third_word));
+  paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&fourth_word));
 
   std::vector<PaddleTensor> outputs;
-  ASSERT_TRUE(predictor->Run(cpu_feeds, &outputs));
+  ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs));
   ASSERT_EQ(outputs.size(), 1UL);
-  for (size_t i = 0; i < outputs.size(); ++i) {
-    size_t len = outputs[i].data.length;
-    float* data = static_cast<float*>(outputs[i].data.data);
-    for (size_t j = 0; j < len / sizeof(float); ++j) {
+  size_t len = outputs[0].data.length;
+  float* data = static_cast<float*>(outputs[0].data.data);
+  for (int j = 0; j < len / sizeof(float); ++j) {
     ASSERT_LT(data[j], 1.0);
     ASSERT_GT(data[j], -1.0);
   }
-    free(outputs[i].data.data);
+
+  std::vector<paddle::framework::LoDTensor*> cpu_feeds;
+  cpu_feeds.push_back(&first_word);
+  cpu_feeds.push_back(&second_word);
+  cpu_feeds.push_back(&third_word);
+  cpu_feeds.push_back(&fourth_word);
+
+  framework::LoDTensor output1;
+  std::vector<paddle::framework::LoDTensor*> cpu_fetchs1;
+  cpu_fetchs1.push_back(&output1);
+
+  TestInference<platform::CPUPlace>(config.model_dir, cpu_feeds, cpu_fetchs1);
+
+  float* lod_data = output1.data<float>();
+  for (size_t i = 0; i < output1.numel(); ++i) {
+    EXPECT_LT(lod_data[i] - data[i], 1e-3);
+    EXPECT_GT(lod_data[i] - data[i], -1e-3);
+  }
+
+  free(outputs[0].data.data);
+}
+
+TEST(paddle_inference_api_impl, image_classification) {
+  int batch_size = 2;
+  bool use_mkldnn = false;
+  bool repeat = false;
+  NativeConfig config = GetConfig();
+  config.model_dir =
+      FLAGS_dirname + "image_classification_resnet.inference.model";
+
+  const bool is_combined = false;
+  std::vector<std::vector<int64_t>> feed_target_shapes =
+      GetFeedTargetShapes(config.model_dir, is_combined);
+
+  framework::LoDTensor input;
+  // Use normilized image pixels as input data,
+  // which should be in the range [0.0, 1.0].
+  feed_target_shapes[0][0] = batch_size;
+  framework::DDim input_dims = framework::make_ddim(feed_target_shapes[0]);
+  SetupTensor<float>(
+      &input, input_dims, static_cast<float>(0), static_cast<float>(1));
+  std::vector<framework::LoDTensor*> cpu_feeds;
+  cpu_feeds.push_back(&input);
+
+  framework::LoDTensor output1;
+  std::vector<framework::LoDTensor*> cpu_fetchs1;
+  cpu_fetchs1.push_back(&output1);
+
+  TestInference<platform::CPUPlace, false, true>(config.model_dir,
+                                                 cpu_feeds,
+                                                 cpu_fetchs1,
+                                                 repeat,
+                                                 is_combined,
+                                                 use_mkldnn);
+
+  auto predictor = CreatePaddlePredictor(config);
+  std::vector<PaddleTensor> paddle_tensor_feeds;
+  paddle_tensor_feeds.push_back(LodTensorToPaddleTensor(&input));
+
+  std::vector<PaddleTensor> outputs;
+  ASSERT_TRUE(predictor->Run(paddle_tensor_feeds, &outputs));
+  ASSERT_EQ(outputs.size(), 1UL);
+  size_t len = outputs[0].data.length;
+  float* data = static_cast<float*>(outputs[0].data.data);
+  float* lod_data = output1.data<float>();
+  for (size_t j = 0; j < len / sizeof(float); ++j) {
+    EXPECT_NEAR(lod_data[j], data[j], 1e-3);
   }
+  free(data);
 }
 
 }  // namespace paddle

paddle/fluid/framework/operator.cc

@@ -469,6 +469,7 @@ class RuntimeInferShapeContext : public InferShapeContext {
  protected:
   DDim GetDim(const std::string& name) const override {
     Variable* var = scope_.FindVar(name);
+    PADDLE_ENFORCE_NOT_NULL(var);
     if (var->IsType<LoDTensor>()) {
       return var->Get<LoDTensor>().dims();
     } else if (var->IsType<SelectedRows>()) {

paddle/fluid/framework/reader.cc

@@ -25,8 +25,10 @@ void FileReader::ReadNext(std::vector<LoDTensor> *out) {
   if (out->empty()) {
     return;
   }
+
+  PADDLE_ENFORCE_EQ(out->size(), dims_.size());
   for (size_t i = 0; i < dims_.size(); ++i) {
-    auto &actual = out->at(i).dims();
+    auto &actual = (*out)[i].dims();
     auto &expect = dims_[i];
 
     PADDLE_ENFORCE_EQ(actual.size(), expect.size());

paddle/fluid/framework/selected_rows.cc

@@ -18,8 +18,8 @@ namespace paddle {
 namespace framework {
 
 struct ReAllocateVisitor {
-  ReAllocateVisitor(framework::Tensor* tensor, const framework::DDim& dims)
-      : tensor_(tensor), dims_(dims) {}
+  ReAllocateVisitor(const framework::DDim& dims, framework::Tensor* tensor)
+      : dims_(dims), tensor_(tensor) {}
 
   template <typename T>
   void operator()() const {
@@ -34,8 +34,8 @@ struct ReAllocateVisitor {
     tensor_->ShareDataWith(cpu_tensor);
   }
 
-  framework::Tensor* tensor_;
   framework::DDim dims_;
+  framework::Tensor* tensor_;
 };
 
 struct TensorCopyVisitor {
@@ -158,6 +158,7 @@ bool SelectedRows::Set(int64_t key, const framework::Tensor& value) {
   }
   PADDLE_ENFORCE_EQ(value.dims()[0], static_cast<size_t>(1),
                     "The first dim of value should be 1.");
+  std::lock_guard<std::mutex> lock(*auto_grown_mutex_.get());
   auto index = Index(key);
   bool is_new_key = false;
   if (index == -1) {
@@ -169,7 +170,7 @@ bool SelectedRows::Set(int64_t key, const framework::Tensor& value) {
       auto dims = value_->dims();
       dims[0] = (dims[0] + 1) << 1;
       framework::VisitDataType(framework::ToDataType(value.type()),
-                               ReAllocateVisitor(value_.get(), dims));
+                               ReAllocateVisitor(dims, value_.get()));
     }
   }
 

paddle/fluid/framework/selected_rows.h

@@ -15,6 +15,8 @@ limitations under the License. */
 #pragma once
 
 #include <algorithm>
+#include <memory>
+#include <mutex>  // NOLINT
 #include <utility>
 #include <vector>
 
@@ -46,11 +48,13 @@ class SelectedRows {
   SelectedRows(const std::vector<int64_t>& rows, const int64_t& height)
       : rows_(rows), height_(height) {
     value_.reset(new Tensor());
+    auto_grown_mutex_.reset(new std::mutex);
   }
 
   SelectedRows() {
     height_ = 0;
     value_.reset(new Tensor());
+    auto_grown_mutex_.reset(new std::mutex);
   }
 
   platform::Place place() const { return value_->place(); }
@@ -125,6 +129,7 @@ class SelectedRows {
   Vector<int64_t> rows_;
   std::unique_ptr<Tensor> value_{nullptr};
   int64_t height_;
+  std::unique_ptr<std::mutex> auto_grown_mutex_{nullptr};
 };
 
 /*

paddle/fluid/framework/tensor_impl.h

@@ -39,7 +39,7 @@ template <typename T>
 inline const T* Tensor::data() const {
   check_memory_size();
   PADDLE_ENFORCE(std::is_same<T, void>::value ||
-                     holder_->type().hash_code() == typeid(T).hash_code(),
+                     holder_->type() == std::type_index(typeid(T)),
                  "Tensor holds the wrong type, it holds %s",
                  this->holder_->type().name());
 
@@ -53,7 +53,7 @@ template <typename T>
 inline T* Tensor::data() {
   check_memory_size();
   PADDLE_ENFORCE(std::is_same<T, void>::value ||
-                     holder_->type().hash_code() == typeid(T).hash_code(),
+                     holder_->type() == std::type_index(typeid(T)),
                  "Tensor holds the wrong type, it holds %s",
                  this->holder_->type().name());
   return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +

paddle/fluid/inference/CMakeLists.txt

@@ -5,14 +5,19 @@ cc_library(paddle_fluid_api
     SRCS io.cc
     DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
 
-# Create static library
 get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
-cc_library(paddle_fluid DEPS ${fluid_modules})
 
+if(WITH_CONTRIB)
+  set(fluid_modules "${fluid_modules}" paddle_inference_api)
+endif()
+
+# Create static library
+cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api)
 # Create shared library
 cc_library(paddle_fluid_shared SHARED
     SRCS io.cc
-    DEPS ${fluid_modules})
+    DEPS ${fluid_modules} paddle_fluid_api)
+
 set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
 if(NOT APPLE)
   # TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.

paddle/fluid/inference/analysis/data_flow_graph.h

@@ -21,7 +21,10 @@ limitations under the License. */
 
 #include <deque>
 #include <stack>
+#include <string>
 #include <unordered_set>
+#include <utility>
+#include <vector>
 
 #include "paddle/fluid/inference/analysis/graph_traits.h"
 #include "paddle/fluid/inference/analysis/node.h"

paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc

@@ -44,6 +44,6 @@ TEST_F(DFG_Tester, Test) {
   LOG(INFO) << graph.nodes.size();
 }
 
-}  // analysis
-}  // inference
-}  // paddle
+};  // namespace analysis
+};  // namespace inference
+};  // namespace paddle

paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc

@@ -12,9 +12,11 @@ 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/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
+#include <string>
 #include <vector>
 
+#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
+
 namespace paddle {
 namespace inference {
 namespace analysis {

paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h

@@ -19,6 +19,8 @@
 
 #pragma once
 
+#include <string>
+
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/inference/analysis/data_flow_graph.h"
 #include "paddle/fluid/inference/analysis/pass.h"