Merge pull request #10956 from panyx0718/infer2

paddle inference interface implementation

paddle/contrib/inference/CMakeLists.txt

@@ -13,10 +13,45 @@
 # limitations under the License.
 #
 
+function(inference_api_test TARGET_NAME TEST_SRC DEP_TEST)
+    set(options "")
+    set(oneValueArgs "")
+    set(multiValueArgs ARGS)
+    cmake_parse_arguments(inference_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+
+    set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
+    set(arg_list "")
+    if(inference_test_ARGS)
+        foreach(arg ${inference_test_ARGS})
+            list(APPEND arg_list "_${arg}")
+        endforeach()
+    else()
+        list(APPEND arg_list "_")
+    endif()
+    foreach(arg ${arg_list})
+        string(REGEX REPLACE "^_$" "" arg "${arg}")
+        cc_test(${TARGET_NAME}
+                SRCS ${TEST_SRC}
+                DEPS paddle_fluid_api paddle_inference_api paddle_inference_api_impl
+                ARGS --dirname=${PYTHON_TESTS_DIR}/book/)
+        # set_tests_properties(${TARGET_NAME}
+        #         PROPERTIES DEPENDS ${DEP_TEST})
+    endforeach()
+endfunction(inference_api_test)
+
+
 cc_library(paddle_inference_api
     SRCS paddle_inference_api.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)

paddle/contrib/inference/paddle_inference_api.h

@@ -27,29 +27,38 @@
 
 namespace paddle {
 
+enum PaddleDType {
+  FLOAT32,
+  INT64,
+};
+
+struct PaddleBuf {
+  void* data;     // pointer to the data memory.
+  size_t length;  // number of memory bytes.
+};
+
 struct PaddleTensor {
   std::string name;  // variable name.
   std::vector<int> shape;
-  std::vector<unsigned char> data;         // bytes of data.
+  PaddleBuf data;  // blob of data.
-  size_t type{typeid(float).hash_code()};  // hash of type
+  PaddleDType dtype;
 };
 
 /*
- * A simple Inference API for Paddle. Currently this API might just be used by
+* A simple Inference API for Paddle. Currently this API might just be used by
- * non-sequence scenerios.
+* non-sequence scenerios.
- * TODO(Superjomn) Prepare another API for NLP-related usages.
+* TODO(Superjomn) Prepare another API for NLP-related usages.
- */
+*/
 class PaddlePredictor {
 public:
   struct Config;
   PaddlePredictor() = default;
   PaddlePredictor(const PaddlePredictor&) = delete;
 
-  // One drived class should has such a constructor
-  // PaddlePredictor(const XConfig& config);
-  // The XConfig is a derived class of Config.
-
   // Predict an record.
+  // The caller should be responsible for allocating and releasing the memory of
+  // `inputs`. `inputs` should be alive until Run returns. caller should be
+  // responsible for releasing the memory of `output_data`.
   virtual bool Run(const std::vector<PaddleTensor>& inputs,
                    std::vector<PaddleTensor>* output_data) = 0;
 

paddle/contrib/inference/paddle_inference_api_impl.h

@@ -0,0 +1,76 @@
+/* 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
+
+   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 <glog/logging.h>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "paddle/contrib/inference/paddle_inference_api.h"
+
+#include "paddle/fluid/framework/ddim.h"
+#include "paddle/fluid/framework/init.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/inference/io.h"
+#include "paddle/fluid/platform/profiler.h"
+
+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 {
+public:
+  explicit PaddlePredictorImpl(const VisConfig &config) : config_(config) {}
+
+  bool Init();
+
+  bool Run(const std::vector<PaddleTensor> &inputs,
+           std::vector<PaddleTensor> *output_data) override;
+
+  std::unique_ptr<PaddlePredictor> Clone() override;
+
+  ~PaddlePredictorImpl() override{};
+
+private:
+  bool InitShared(PaddlePredictorImpl *cls);
+  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<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_;
+  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

@@ -0,0 +1,83 @@
+/* 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
+
+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 <glog/logging.h>
+#include <gtest/gtest.h>
+
+#include "gflags/gflags.h"
+#include "paddle/contrib/inference/paddle_inference_api_impl.h"
+#include "paddle/fluid/inference/tests/test_helper.h"
+
+DEFINE_string(dirname, "", "Directory of the inference model.");
+
+namespace paddle {
+
+PaddleTensor LodTensorToPaddleTensor(framework::LoDTensor* t) {
+  PaddleTensor pt;
+  pt.data.data = t->data<void>();
+
+  if (t->type() == typeid(int64_t)) {
+    pt.data.length = t->numel() * sizeof(int64_t);
+    pt.dtype = PaddleDType::INT64;
+  } else if (t->type() == typeid(float)) {
+    pt.data.length = t->numel() * sizeof(float);
+    pt.dtype = PaddleDType::FLOAT32;
+  } else {
+    LOG(FATAL) << "unsupported type.";
+  }
+  pt.shape = framework::vectorize2int(t->dims());
+  return pt;
+}
+
+TEST(paddle_inference_api_impl, word2vec) {
+  VisConfig config;
+  config.model_dir = FLAGS_dirname + "word2vec.inference.model";
+  LOG(INFO) << "dirname  " << config.model_dir;
+  config.fraction_of_gpu_memory = 0.85;
+  config.device = 0;
+  config.share_variables = true;
+
+  std::unique_ptr<PaddlePredictorImpl> predictor =
+      CreatePaddlePredictorImpl(config);
+
+  framework::LoDTensor first_word, second_word, third_word, fourth_word;
+  framework::LoD lod{{0, 1}};
+  int64_t dict_size = 2073;  // The size of dictionary
+
+  SetupLoDTensor(&first_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  SetupLoDTensor(&second_word, lod, static_cast<int64_t>(0), dict_size - 1);
+  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> outputs;
+  ASSERT_TRUE(predictor->Run(cpu_feeds, &outputs));
+  ASSERT_EQ(outputs.size(), 1);
+  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 (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);
+  }
+}
+
+}  // namespace paddle

paddle/fluid/inference/CMakeLists.txt

@@ -1,5 +1,6 @@
 set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor init)
 
+# TODO(panyx0718): Should this be called paddle_fluid_inference_api_internal?
 cc_library(paddle_fluid_api
     SRCS io.cc
     DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})