fix conflicts

deploy/README.md

@@ -0,0 +1,3 @@
+# 模型部署
+
+本目录为PaddleX模型部署代码。

deploy/cpp/.clang-format

@@ -0,0 +1,27 @@
+# This file is used by clang-format to autoformat paddle source code
+#
+# The clang-format is part of llvm toolchain.
+# It need to install llvm and clang to format source code style.
+#
+# The basic usage is,
+#   clang-format -i -style=file PATH/TO/SOURCE/CODE
+#
+# The -style=file implicit use ".clang-format" file located in one of 
+# parent directory. 
+# The -i means inplace change.
+#
+# The document of clang-format is 
+#   http://clang.llvm.org/docs/ClangFormat.html
+#   http://clang.llvm.org/docs/ClangFormatStyleOptions.html
+---
+Language:        Cpp
+BasedOnStyle:  Google
+IndentWidth:     2
+TabWidth:        2
+ContinuationIndentWidth: 4
+AccessModifierOffset: -1  # The private/protected/public has no indent in class
+Standard:  Cpp11 
+AllowAllParametersOfDeclarationOnNextLine: true
+BinPackParameters: false
+BinPackArguments: false
+...

deploy/cpp/CMakeLists.txt

@@ -0,0 +1,254 @@
+cmake_minimum_required(VERSION 3.0)
+project(PaddleX CXX C)
+
+option(WITH_MKL        "Compile demo with MKL/OpenBlas support,defaultuseMKL."          ON)
+option(WITH_GPU        "Compile demo with GPU/CPU, default use CPU."                    ON)
+option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static."   ON)
+option(WITH_TENSORRT "Compile demo with TensorRT."   OFF)
+
+SET(PADDLE_DIR "" CACHE PATH "Location of libraries")
+SET(OPENCV_DIR "" CACHE PATH "Location of libraries")
+SET(CUDA_LIB "" CACHE PATH "Location of libraries")
+
+include(cmake/yaml-cpp.cmake)
+
+include_directories("${CMAKE_SOURCE_DIR}/")
+include_directories("${CMAKE_CURRENT_BINARY_DIR}/ext/yaml-cpp/src/ext-yaml-cpp/include")
+link_directories("${CMAKE_CURRENT_BINARY_DIR}/ext/yaml-cpp/lib")
+
+macro(safe_set_static_flag)
+    foreach(flag_var
+        CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+        CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
+      if(${flag_var} MATCHES "/MD")
+        string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
+      endif(${flag_var} MATCHES "/MD")
+    endforeach(flag_var)
+endmacro()
+
+if (WITH_MKL)
+    ADD_DEFINITIONS(-DUSE_MKL)
+endif()
+
+if (NOT DEFINED PADDLE_DIR OR ${PADDLE_DIR} STREQUAL "")
+    message(FATAL_ERROR "please set PADDLE_DIR with -DPADDLE_DIR=/path/paddle_influence_dir")
+endif()
+
+if (NOT DEFINED OPENCV_DIR OR ${OPENCV_DIR} STREQUAL "")
+    message(FATAL_ERROR "please set OPENCV_DIR with -DOPENCV_DIR=/path/opencv")
+endif()
+
+include_directories("${CMAKE_SOURCE_DIR}/")
+include_directories("${PADDLE_DIR}/")
+include_directories("${PADDLE_DIR}/third_party/install/protobuf/include")
+include_directories("${PADDLE_DIR}/third_party/install/glog/include")
+include_directories("${PADDLE_DIR}/third_party/install/gflags/include")
+include_directories("${PADDLE_DIR}/third_party/install/xxhash/include")
+if (EXISTS "${PADDLE_DIR}/third_party/install/snappy/include")
+    include_directories("${PADDLE_DIR}/third_party/install/snappy/include")
+endif()
+if(EXISTS "${PADDLE_DIR}/third_party/install/snappystream/include")
+    include_directories("${PADDLE_DIR}/third_party/install/snappystream/include")
+endif()
+include_directories("${PADDLE_DIR}/third_party/install/zlib/include")
+include_directories("${PADDLE_DIR}/third_party/boost")
+include_directories("${PADDLE_DIR}/third_party/eigen3")
+
+if (EXISTS "${PADDLE_DIR}/third_party/install/snappy/lib")
+    link_directories("${PADDLE_DIR}/third_party/install/snappy/lib")
+endif()
+if(EXISTS "${PADDLE_DIR}/third_party/install/snappystream/lib")
+    link_directories("${PADDLE_DIR}/third_party/install/snappystream/lib")
+endif()
+
+link_directories("${PADDLE_DIR}/third_party/install/zlib/lib")
+link_directories("${PADDLE_DIR}/third_party/install/protobuf/lib")
+link_directories("${PADDLE_DIR}/third_party/install/glog/lib")
+link_directories("${PADDLE_DIR}/third_party/install/gflags/lib")
+link_directories("${PADDLE_DIR}/third_party/install/xxhash/lib")
+link_directories("${PADDLE_DIR}/paddle/lib/")
+link_directories("${CMAKE_CURRENT_BINARY_DIR}")
+
+if (WIN32)
+  include_directories("${PADDLE_DIR}/paddle/fluid/inference")
+  include_directories("${PADDLE_DIR}/paddle/include")
+  link_directories("${PADDLE_DIR}/paddle/fluid/inference")
+  find_package(OpenCV REQUIRED PATHS ${OPENCV_DIR}/build/ NO_DEFAULT_PATH)
+  unset(OpenCV_DIR CACHE)
+else ()
+  find_package(OpenCV REQUIRED PATHS ${OPENCV_DIR}/share/OpenCV NO_DEFAULT_PATH)
+  include_directories("${PADDLE_DIR}/paddle/include")
+  link_directories("${PADDLE_DIR}/paddle/lib")
+endif ()
+include_directories(${OpenCV_INCLUDE_DIRS})
+
+if (WIN32)
+    add_definitions("/DGOOGLE_GLOG_DLL_DECL=")
+    set(CMAKE_C_FLAGS_DEBUG   "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd")
+    set(CMAKE_C_FLAGS_RELEASE  "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT")
+    set(CMAKE_CXX_FLAGS_DEBUG  "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd")
+    set(CMAKE_CXX_FLAGS_RELEASE   "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
+    if (WITH_STATIC_LIB)
+        safe_set_static_flag()
+        add_definitions(-DSTATIC_LIB)
+    endif()
+else()
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -o2 -fopenmp -std=c++11")
+    set(CMAKE_STATIC_LIBRARY_PREFIX "")
+endif()
+
+if (WITH_GPU)
+    if (NOT DEFINED CUDA_LIB OR ${CUDA_LIB} STREQUAL "")
+        message(FATAL_ERROR "please set CUDA_LIB with -DCUDA_LIB=/path/cuda/lib64")
+    endif()
+    if (NOT WIN32)
+        if (NOT DEFINED CUDNN_LIB)
+            message(FATAL_ERROR "please set CUDNN_LIB with -DCUDNN_LIB=/path/cudnn/")
+        endif()
+    endif(NOT WIN32)
+endif()
+
+
+if (NOT WIN32)
+  if (WITH_TENSORRT AND WITH_GPU)
+      include_directories("${PADDLE_DIR}/third_party/install/tensorrt/include")
+      link_directories("${PADDLE_DIR}/third_party/install/tensorrt/lib")
+  endif()
+endif(NOT WIN32)
+
+if (NOT WIN32)
+    set(NGRAPH_PATH "${PADDLE_DIR}/third_party/install/ngraph")
+    if(EXISTS ${NGRAPH_PATH})
+        include(GNUInstallDirs)
+        include_directories("${NGRAPH_PATH}/include")
+        link_directories("${NGRAPH_PATH}/${CMAKE_INSTALL_LIBDIR}")
+        set(NGRAPH_LIB ${NGRAPH_PATH}/${CMAKE_INSTALL_LIBDIR}/libngraph${CMAKE_SHARED_LIBRARY_SUFFIX})
+    endif()
+endif()
+
+if(WITH_MKL)
+  include_directories("${PADDLE_DIR}/third_party/install/mklml/include")
+  if (WIN32)
+    set(MATH_LIB ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.lib
+            ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.lib)
+  else ()
+    set(MATH_LIB ${PADDLE_DIR}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX}
+            ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5${CMAKE_SHARED_LIBRARY_SUFFIX})
+    execute_process(COMMAND cp -r ${PADDLE_DIR}/third_party/install/mklml/lib/libmklml_intel${CMAKE_SHARED_LIBRARY_SUFFIX} /usr/lib)
+  endif ()
+  set(MKLDNN_PATH "${PADDLE_DIR}/third_party/install/mkldnn")
+  if(EXISTS ${MKLDNN_PATH})
+    include_directories("${MKLDNN_PATH}/include")
+    if (WIN32)
+      set(MKLDNN_LIB ${MKLDNN_PATH}/lib/mkldnn.lib)
+    else ()
+      set(MKLDNN_LIB ${MKLDNN_PATH}/lib/libmkldnn.so.0)
+    endif ()
+  endif()
+else()
+  set(MATH_LIB ${PADDLE_DIR}/third_party/install/openblas/lib/libopenblas${CMAKE_STATIC_LIBRARY_SUFFIX})
+endif()
+
+if (WIN32)
+    if(EXISTS "${PADDLE_DIR}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX}")
+        set(DEPS
+            ${PADDLE_DIR}/paddle/fluid/inference/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
+    else()
+        set(DEPS
+            ${PADDLE_DIR}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
+    endif()
+endif()
+
+if(WITH_STATIC_LIB)
+    set(DEPS
+        ${PADDLE_DIR}/paddle/lib/libpaddle_fluid${CMAKE_STATIC_LIBRARY_SUFFIX})
+else()
+    set(DEPS
+        ${PADDLE_DIR}/paddle/lib/libpaddle_fluid${CMAKE_SHARED_LIBRARY_SUFFIX})
+endif()
+
+if (NOT WIN32)
+    set(DEPS ${DEPS}
+        ${MATH_LIB} ${MKLDNN_LIB} 
+        glog gflags protobuf z xxhash yaml-cpp
+        )
+    if(EXISTS "${PADDLE_DIR}/third_party/install/snappystream/lib")
+        set(DEPS ${DEPS} snappystream)
+    endif()
+    if (EXISTS "${PADDLE_DIR}/third_party/install/snappy/lib")
+        set(DEPS ${DEPS} snappy)
+    endif()
+else()
+    set(DEPS ${DEPS}
+        ${MATH_LIB} ${MKLDNN_LIB}
+        glog gflags_static libprotobuf zlibstatic xxhash libyaml-cppmt)
+    set(DEPS ${DEPS} libcmt shlwapi)
+    if (EXISTS "${PADDLE_DIR}/third_party/install/snappy/lib")
+        set(DEPS ${DEPS} snappy)
+    endif()
+    if(EXISTS "${PADDLE_DIR}/third_party/install/snappystream/lib")
+        set(DEPS ${DEPS} snappystream)
+    endif()
+endif(NOT WIN32)
+
+if(WITH_GPU)
+  if(NOT WIN32)
+    if (WITH_TENSORRT)
+      set(DEPS ${DEPS} ${PADDLE_DIR}/third_party/install/tensorrt/lib/libnvinfer${CMAKE_STATIC_LIBRARY_SUFFIX})
+      set(DEPS ${DEPS} ${PADDLE_DIR}/third_party/install/tensorrt/lib/libnvinfer_plugin${CMAKE_STATIC_LIBRARY_SUFFIX})
+    endif()
+    set(DEPS ${DEPS} ${CUDA_LIB}/libcudart${CMAKE_SHARED_LIBRARY_SUFFIX})
+    set(DEPS ${DEPS} ${CUDNN_LIB}/libcudnn${CMAKE_SHARED_LIBRARY_SUFFIX})
+  else()
+    set(DEPS ${DEPS} ${CUDA_LIB}/cudart${CMAKE_STATIC_LIBRARY_SUFFIX} )
+    set(DEPS ${DEPS} ${CUDA_LIB}/cublas${CMAKE_STATIC_LIBRARY_SUFFIX} )
+    set(DEPS ${DEPS} ${CUDA_LIB}/cudnn${CMAKE_STATIC_LIBRARY_SUFFIX})
+  endif()
+endif()
+
+if (NOT WIN32)
+    set(EXTERNAL_LIB "-ldl -lrt -lgomp -lz -lm -lpthread")
+    set(DEPS ${DEPS} ${EXTERNAL_LIB})
+endif()
+
+set(DEPS ${DEPS} ${OpenCV_LIBS}) 
+add_executable(classifier src/classifier.cpp src/transforms.cpp src/paddlex.cpp)
+ADD_DEPENDENCIES(classifier ext-yaml-cpp)
+target_link_libraries(classifier ${DEPS})
+
+add_executable(detector src/detector.cpp src/transforms.cpp src/paddlex.cpp src/visualize.cpp)
+ADD_DEPENDENCIES(detector ext-yaml-cpp)
+target_link_libraries(detector ${DEPS})
+
+add_executable(segmenter src/segmenter.cpp src/transforms.cpp src/paddlex.cpp src/visualize.cpp)
+ADD_DEPENDENCIES(segmenter ext-yaml-cpp)
+target_link_libraries(segmenter ${DEPS})
+
+if (WIN32 AND WITH_MKL)
+    add_custom_command(TARGET classifier POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./mklml.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./libiomp5md.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./mkldnn.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./release/mklml.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./release/libiomp5md.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./release/mkldnn.dll
+    )
+    add_custom_command(TARGET detector POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./mklml.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./libiomp5md.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./mkldnn.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./release/mklml.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./release/libiomp5md.dll
+    )
+    add_custom_command(TARGET segmenter POST_BUILD
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./mklml.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./libiomp5md.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./mkldnn.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/mklml.dll ./release/mklml.dll
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mklml/lib/libiomp5md.dll ./release/libiomp5md.dll
+
+        COMMAND ${CMAKE_COMMAND} -E copy_if_different ${PADDLE_DIR}/third_party/install/mkldnn/lib/mkldnn.dll ./release/mkldnn.dll
+    )
+
+endif()
+

deploy/cpp/CMakeSettings.json

@@ -0,0 +1,47 @@
+{
+    "configurations": [
+        {
+            "name": "x64-Release",
+            "generator": "Ninja",
+            "configurationType": "RelWithDebInfo",
+            "inheritEnvironments": [ "msvc_x64_x64" ],
+            "buildRoot": "${projectDir}\\out\\build\\${name}",
+            "installRoot": "${projectDir}\\out\\install\\${name}",
+            "cmakeCommandArgs": "",
+            "buildCommandArgs": "-v",
+            "ctestCommandArgs": "",
+            "variables": [
+                {
+                    "name": "OPENCV_DIR",
+                    "value": "C:/projects/opencv",
+                    "type": "PATH"
+                },
+                {
+                    "name": "PADDLE_DIR",
+                    "value": "C:/projects/fluid_install_dir_win_cpu_1.6/fluid_install_dir_win_cpu_1.6",
+                    "type": "PATH"
+                },
+                {
+                    "name": "CMAKE_BUILD_TYPE",
+                    "value": "Release",
+                    "type": "STRING"
+                },
+                {
+                    "name": "WITH_STATIC_LIB",
+                    "value": "True",
+                    "type": "BOOL"
+                },
+                {
+                    "name": "WITH_MKL",
+                    "value": "True",
+                    "type": "BOOL"
+                },
+                {
+                    "name": "WITH_GPU",
+                    "value": "False",
+                    "type": "BOOL"
+                }
+            ]
+        }
+    ]
+}

deploy/cpp/cmake/yaml-cpp.cmake

@@ -0,0 +1,30 @@
+find_package(Git REQUIRED)
+
+include(ExternalProject)
+
+message("${CMAKE_BUILD_TYPE}")
+
+ExternalProject_Add(
+        ext-yaml-cpp
+        URL https://bj.bcebos.com/paddlex/deploy/deps/yaml-cpp.zip
+        URL_MD5 9542d6de397d1fbd649ed468cb5850e6
+        CMAKE_ARGS
+        -DYAML_CPP_BUILD_TESTS=OFF
+        -DYAML_CPP_BUILD_TOOLS=OFF
+        -DYAML_CPP_INSTALL=OFF
+        -DYAML_CPP_BUILD_CONTRIB=OFF
+        -DMSVC_SHARED_RT=OFF
+        -DBUILD_SHARED_LIBS=OFF
+        -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
+        -DCMAKE_CXX_FLAGS=${CMAKE_CXX_FLAGS}
+        -DCMAKE_CXX_FLAGS_DEBUG=${CMAKE_CXX_FLAGS_DEBUG}
+        -DCMAKE_CXX_FLAGS_RELEASE=${CMAKE_CXX_FLAGS_RELEASE}
+        -DCMAKE_LIBRARY_OUTPUT_DIRECTORY=${CMAKE_BINARY_DIR}/ext/yaml-cpp/lib
+        -DCMAKE_ARCHIVE_OUTPUT_DIRECTORY=${CMAKE_BINARY_DIR}/ext/yaml-cpp/lib
+        PREFIX "${CMAKE_BINARY_DIR}/ext/yaml-cpp"
+        # Disable install step
+        INSTALL_COMMAND ""
+        LOG_DOWNLOAD ON
+        LOG_BUILD 1
+)
+

deploy/cpp/include/paddlex/config_parser.h

@@ -0,0 +1,57 @@
+//   Copyright (c) 2020 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 <iostream>
+#include <map>
+#include <string>
+#include <vector>
+
+#include "yaml-cpp/yaml.h"
+
+#ifdef _WIN32
+#define OS_PATH_SEP "\\"
+#else
+#define OS_PATH_SEP "/"
+#endif
+
+namespace PaddleX {
+
+// Inference model configuration parser
+class ConfigPaser {
+ public:
+  ConfigPaser() {}
+
+  ~ConfigPaser() {}
+
+  bool load_config(const std::string& model_dir,
+                   const std::string& cfg = "model.yml") {
+    // Load as a YAML::Node
+    YAML::Node config;
+    config = YAML::LoadFile(model_dir + OS_PATH_SEP + cfg);
+
+    if (config["Transforms"].IsDefined()) {
+      YAML::Node transforms_ = config["Transforms"];
+    } else {
+      std::cerr << "There's no field 'Transforms' in model.yml" << std::endl;
+      return false;
+    }
+    return true;
+  }
+
+  YAML::Node Transforms_;
+};
+
+}  // namespace PaddleDetection

deploy/cpp/include/paddlex/paddlex.h

@@ -0,0 +1,71 @@
+//   Copyright (c) 2020 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 <functional>
+#include <iostream>
+#include <numeric>
+
+#include "yaml-cpp/yaml.h"
+
+#ifdef _WIN32
+#define OS_PATH_SEP "\\"
+#else
+#define OS_PATH_SEP "/"
+#endif
+
+#include "paddle_inference_api.h"  // NOLINT
+
+#include "include/paddlex/config_parser.h"
+#include "include/paddlex/results.h"
+#include "include/paddlex/transforms.h"
+
+namespace PaddleX {
+
+class Model {
+ public:
+  void Init(const std::string& model_dir,
+            bool use_gpu = false,
+            int gpu_id = 0) {
+    create_predictor(model_dir, use_gpu, gpu_id);
+  }
+
+  void create_predictor(const std::string& model_dir,
+                        bool use_gpu = false,
+                        int gpu_id = 0);
+
+  bool load_config(const std::string& model_dir);
+
+  bool preprocess(const cv::Mat& input_im, ImageBlob* blob);
+
+  bool predict(const cv::Mat& im, ClsResult* result);
+
+  bool predict(const cv::Mat& im, DetResult* result);
+
+  bool predict(const cv::Mat& im, SegResult* result);
+
+  bool postprocess(SegResult* result);
+
+  bool postprocess(DetResult* result);
+
+  std::string type;
+  std::string name;
+  std::map<int, std::string> labels;
+  Transforms transforms_;
+  ImageBlob inputs_;
+  std::vector<float> outputs_;
+  std::unique_ptr<paddle::PaddlePredictor> predictor_;
+};
+}  // namespce of PaddleX

deploy/cpp/include/paddlex/results.h

@@ -0,0 +1,71 @@
+//   Copyright (c) 2020 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 <iostream>
+#include <string>
+#include <vector>
+
+namespace PaddleX {
+
+template <class T>
+struct Mask {
+  std::vector<T> data;
+  std::vector<int> shape;
+  void clear() {
+    data.clear();
+    shape.clear();
+  }
+};
+
+struct Box {
+  int category_id;
+  std::string category;
+  float score;
+  std::vector<float> coordinate;
+  Mask<float> mask;
+};
+
+class BaseResult {
+ public:
+  std::string type = "base";
+};
+
+class ClsResult : public BaseResult {
+ public:
+  int category_id;
+  std::string category;
+  float score;
+  std::string type = "cls";
+};
+
+class DetResult : public BaseResult {
+ public:
+  std::vector<Box> boxes;
+  int mask_resolution;
+  std::string type = "det";
+  void clear() { boxes.clear(); }
+};
+
+class SegResult : public BaseResult {
+ public:
+  Mask<int64_t> label_map;
+  Mask<float> score_map;
+  void clear() {
+    label_map.clear();
+    score_map.clear();
+  }
+};
+}  // namespce of PaddleX

deploy/cpp/include/paddlex/transforms.h

@@ -0,0 +1,196 @@
+//   Copyright (c) 2020 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 <yaml-cpp/yaml.h>
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <opencv2/core/core.hpp>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/imgproc/imgproc.hpp>
+
+namespace PaddleX {
+
+// Object for storing all preprocessed data
+class ImageBlob {
+ public:
+  // Original image height and width
+  std::vector<int> ori_im_size_ = std::vector<int>(2);
+  // Newest image height and width after process
+  std::vector<int> new_im_size_ = std::vector<int>(2);
+  // Image height and width before padding
+  std::vector<int> im_size_before_padding_ = std::vector<int>(2);
+  // Image height and width before resize
+  std::vector<int> im_size_before_resize_ = std::vector<int>(2);
+  // Reshape order
+  std::vector<std::string> reshape_order_;
+  // Resize scale
+  float scale = 1.0;
+  // Buffer for image data after preprocessing
+  std::vector<float> im_data_;
+
+  void clear() {
+    ori_im_size_.clear();
+    new_im_size_.clear();
+    im_size_before_padding_.clear();
+    im_size_before_resize_.clear();
+    reshape_order_.clear();
+    im_data_.clear();
+  }
+};
+
+// Abstraction of preprocessing opration class
+class Transform {
+ public:
+  virtual void Init(const YAML::Node& item) = 0;
+  virtual bool Run(cv::Mat* im, ImageBlob* data) = 0;
+};
+
+class Normalize : public Transform {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    mean_ = item["mean"].as<std::vector<float>>();
+    std_ = item["std"].as<std::vector<float>>();
+  }
+
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  std::vector<float> mean_;
+  std::vector<float> std_;
+};
+
+class ResizeByShort : public Transform {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    short_size_ = item["short_size"].as<int>();
+    if (item["max_size"].IsDefined()) {
+      max_size_ = item["max_size"].as<int>();
+    } else {
+      max_size_ = -1;
+    }
+  };
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  float GenerateScale(const cv::Mat& im);
+  int short_size_;
+  int max_size_;
+};
+
+class ResizeByLong : public Transform {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    long_size_ = item["long_size"].as<int>();
+  };
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  int long_size_;
+};
+
+class Resize : public Transform {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    if (item["target_size"].IsScalar()) {
+      height_ = item["target_size"].as<int>();
+      width_ = item["target_size"].as<int>();
+      interp_ = item["interp"].as<std::string>();
+    } else if (item["target_size"].IsSequence()) {
+      std::vector<int> target_size = item["target_size"].as<std::vector<int>>();
+      width_ = target_size[0];
+      height_ = target_size[1];
+    }
+    if (height_ <= 0 || width_ <= 0) {
+      std::cerr << "[Resize] target_size should greater than 0" << std::endl;
+      exit(-1);
+    }
+  }
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  int height_;
+  int width_;
+  std::string interp_;
+};
+
+class CenterCrop : public Transform {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    if (item["crop_size"].IsScalar()) {
+      height_ = item["crop_size"].as<int>();
+      width_ = item["crop_size"].as<int>();
+    } else if (item["crop_size"].IsSequence()) {
+      std::vector<int> crop_size = item["crop_size"].as<std::vector<int>>();
+      width_ = crop_size[0];
+      height_ = crop_size[1];
+    }
+  }
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  int height_;
+  int width_;
+};
+
+class Padding : public Transform {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    if (item["coarsest_stride"].IsDefined()) {
+      coarsest_stride_ = item["coarsest_stride"].as<int>();
+      if (coarsest_stride_ <= 1) {
+        std::cerr << "[Padding] coarest_stride should greater than 0"
+                  << std::endl;
+        exit(-1);
+      }
+    } else {
+      if (item["target_size"].IsScalar()) {
+        width_ = item["target_size"].as<int>();
+        height_ = item["target_size"].as<int>();
+      } else if (item["target_size"].IsSequence()) {
+        width_ = item["target_size"].as<std::vector<int>>()[0];
+        height_ = item["target_size"].as<std::vector<int>>()[1];
+      }
+    }
+    if (item["im_padding_value"].IsDefined()) {
+      value_ = item["im_padding_value"].as<std::vector<float>>();
+    }
+  }
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  int coarsest_stride_ = -1;
+  int width_ = 0;
+  int height_ = 0;
+  std::vector<float> value_;
+};
+
+class Transforms {
+ public:
+  void Init(const YAML::Node& node, bool to_rgb = true);
+  std::shared_ptr<Transform> CreateTransform(const std::string& name);
+  bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  std::vector<std::shared_ptr<Transform>> transforms_;
+  bool to_rgb_ = true;
+};
+
+}  // namespace PaddleX

deploy/cpp/include/paddlex/visualize.h

@@ -0,0 +1,62 @@
+//   Copyright (c) 2020 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 <iostream>
+#include <map>
+#include <vector>
+#ifdef _WIN32
+#include <direct.h>
+#include <io.h>
+#else  // Linux/Unix
+#include <dirent.h>
+#include <sys/io.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+#endif
+#include <string>
+
+#include <opencv2/core/core.hpp>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/imgproc/imgproc.hpp>
+
+#include "include/paddlex/results.h"
+
+#ifdef _WIN32
+#define OS_PATH_SEP "\\"
+#else
+#define OS_PATH_SEP "/"
+#endif
+
+namespace PaddleX {
+
+// Generate visualization colormap for each class
+std::vector<int> GenerateColorMap(int num_class);
+
+cv::Mat VisualizeDet(const cv::Mat& img,
+                     const DetResult& results,
+                     const std::map<int, std::string>& labels,
+                     const std::vector<int>& colormap,
+                     float threshold = 0.5);
+
+cv::Mat VisualizeSeg(const cv::Mat& img,
+                     const SegResult& result,
+                     const std::map<int, std::string>& labels,
+                     const std::vector<int>& colormap);
+
+std::string generate_save_path(const std::string& save_dir,
+                               const std::string& file_path);
+}  // namespce of PaddleX

deploy/cpp/scripts/bootstrap.sh

@@ -0,0 +1,10 @@
+# download pre-compiled opencv lib
+OPENCV_URL=https://paddleseg.bj.bcebos.com/deploy/docker/opencv3gcc4.8.tar.bz2
+if [ ! -d "./deps/opencv3gcc4.8" ]; then
+    mkdir -p deps
+    cd deps
+    wget -c ${OPENCV_URL}
+    tar xvfj opencv3gcc4.8.tar.bz2
+    rm -rf opencv3gcc4.8.tar.bz2
+    cd ..
+fi

deploy/cpp/scripts/build.sh

@@ -0,0 +1,27 @@
+# 是否使用GPU(即是否使用 CUDA)
+WITH_GPU=ON
+# 是否集成 TensorRT(仅WITH_GPU=ON 有效)
+WITH_TENSORRT=OFF
+# Paddle 预测库路径
+PADDLE_DIR=/path/to/fluid_inference/
+# CUDA 的 lib 路径
+CUDA_LIB=/path/to/cuda/lib/
+# CUDNN 的 lib 路径
+CUDNN_LIB=/path/to/cudnn/lib/
+
+# OPENCV 路径, 如果使用自带预编译版本可不修改
+OPENCV_DIR=$(pwd)/deps/opencv3gcc4.8/
+sh $(pwd)/scripts/bootstrap.sh
+
+# 以下无需改动
+rm -rf build
+mkdir -p build
+cd build
+cmake .. \
+    -DWITH_GPU=${WITH_GPU} \
+    -DWITH_TENSORRT=${WITH_TENSORRT} \
+    -DPADDLE_DIR=${PADDLE_DIR} \
+    -DCUDA_LIB=${CUDA_LIB} \
+    -DCUDNN_LIB=${CUDNN_LIB} \
+    -DOPENCV_DIR=${OPENCV_DIR}
+make

deploy/cpp/src/classifier.cpp

@@ -0,0 +1,73 @@
+//   Copyright (c) 2020 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 <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "include/paddlex/paddlex.h"
+
+DEFINE_string(model_dir, "", "Path of inference model");
+DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
+DEFINE_int32(gpu_id, 0, "GPU card id");
+DEFINE_string(image, "", "Path of test image file");
+DEFINE_string(image_list, "", "Path of test image list file");
+
+int main(int argc, char** argv) {
+  // Parsing command-line
+  google::ParseCommandLineFlags(&argc, &argv, true);
+
+  if (FLAGS_model_dir == "") {
+    std::cerr << "--model_dir need to be defined" << std::endl;
+    return -1;
+  }
+  if (FLAGS_image == "" & FLAGS_image_list == "") {
+    std::cerr << "--image or --image_list need to be defined" << std::endl;
+    return -1;
+  }
+
+  // 加载模型
+  PaddleX::Model model;
+  model.Init(FLAGS_model_dir, FLAGS_use_gpu, FLAGS_gpu_id);
+
+  // 进行预测
+  if (FLAGS_image_list != "") {
+    std::ifstream inf(FLAGS_image_list);
+    if (!inf) {
+      std::cerr << "Fail to open file " << FLAGS_image_list << std::endl;
+      return -1;
+    }
+    std::string image_path;
+    while (getline(inf, image_path)) {
+      PaddleX::ClsResult result;
+      cv::Mat im = cv::imread(image_path, 1);
+      model.predict(im, &result);
+      std::cout << "Predict label: " << result.category
+                << ", label_id:" << result.category_id
+                << ", score: " << result.score << std::endl;
+    }
+  } else {
+    PaddleX::ClsResult result;
+    cv::Mat im = cv::imread(FLAGS_image, 1);
+    model.predict(im, &result);
+    std::cout << "Predict label: " << result.category
+              << ", label_id:" << result.category_id
+              << ", score: " << result.score << std::endl;
+  }
+
+  return 0;
+}

deploy/cpp/src/detector.cpp

@@ -0,0 +1,108 @@
+//   Copyright (c) 2020 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 <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "include/paddlex/paddlex.h"
+#include "include/paddlex/visualize.h"
+
+DEFINE_string(model_dir, "", "Path of inference model");
+DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
+DEFINE_int32(gpu_id, 0, "GPU card id");
+DEFINE_string(image, "", "Path of test image file");
+DEFINE_string(image_list, "", "Path of test image list file");
+DEFINE_string(save_dir, "output", "Path to save visualized image");
+
+int main(int argc, char** argv) {
+  // 解析命令行参数
+  google::ParseCommandLineFlags(&argc, &argv, true);
+
+  if (FLAGS_model_dir == "") {
+    std::cerr << "--model_dir need to be defined" << std::endl;
+    return -1;
+  }
+  if (FLAGS_image == "" & FLAGS_image_list == "") {
+    std::cerr << "--image or --image_list need to be defined" << std::endl;
+    return -1;
+  }
+
+  // 加载模型
+  PaddleX::Model model;
+  model.Init(FLAGS_model_dir, FLAGS_use_gpu, FLAGS_gpu_id);
+
+  auto colormap = PaddleX::GenerateColorMap(model.labels.size());
+  std::string save_dir = "output";
+  // 进行预测
+  if (FLAGS_image_list != "") {
+    std::ifstream inf(FLAGS_image_list);
+    if (!inf) {
+      std::cerr << "Fail to open file " << FLAGS_image_list << std::endl;
+      return -1;
+    }
+    std::string image_path;
+    while (getline(inf, image_path)) {
+      PaddleX::DetResult result;
+      cv::Mat im = cv::imread(image_path, 1);
+      model.predict(im, &result);
+      for (int i = 0; i < result.boxes.size(); ++i) {
+        std::cout << "image file: " << image_path
+                  << ", predict label: " << result.boxes[i].category
+                  << ", label_id:" << result.boxes[i].category_id
+                  << ", score: " << result.boxes[i].score << ", box:("
+                  << result.boxes[i].coordinate[0] << ", "
+                  << result.boxes[i].coordinate[1] << ", "
+                  << result.boxes[i].coordinate[2] << ", "
+                  << result.boxes[i].coordinate[3] << std::endl;
+      }
+
+      // 可视化
+      cv::Mat vis_img =
+          PaddleX::VisualizeDet(im, result, model.labels, colormap, 0.5);
+      std::string save_path =
+          PaddleX::generate_save_path(FLAGS_save_dir, image_path);
+      cv::imwrite(save_path, vis_img);
+      result.clear();
+      std::cout << "Visualized output saved as " << save_path << std::endl;
+    }
+  } else {
+    PaddleX::DetResult result;
+    cv::Mat im = cv::imread(FLAGS_image, 1);
+    model.predict(im, &result);
+    for (int i = 0; i < result.boxes.size(); ++i) {
+      std::cout << ", predict label: " << result.boxes[i].category
+                << ", label_id:" << result.boxes[i].category_id
+                << ", score: " << result.boxes[i].score << ", box:("
+                << result.boxes[i].coordinate[0] << ", "
+                << result.boxes[i].coordinate[1] << ", "
+                << result.boxes[i].coordinate[2] << ", "
+                << result.boxes[i].coordinate[3] << std::endl;
+    }
+
+    // 可视化
+    cv::Mat vis_img =
+        PaddleX::VisualizeDet(im, result, model.labels, colormap, 0.5);
+    std::string save_path =
+        PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_image);
+    cv::imwrite(save_path, vis_img);
+    result.clear();
+    std::cout << "Visualized output saved as " << save_path << std::endl;
+  }
+
+  return 0;
+}

deploy/cpp/src/paddlex.cpp

@@ -0,0 +1,324 @@
+//   Copyright (c) 2020 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 "include/paddlex/paddlex.h"
+
+namespace PaddleX {
+
+void Model::create_predictor(const std::string& model_dir,
+                             bool use_gpu,
+                             int gpu_id) {
+  // 读取配置文件
+  if (!load_config(model_dir)) {
+    std::cerr << "Parse file 'model.yml' failed!" << std::endl;
+    exit(-1);
+  }
+  paddle::AnalysisConfig config;
+  std::string model_file = model_dir + OS_PATH_SEP + "__model__";
+  std::string params_file = model_dir + OS_PATH_SEP + "__params__";
+  config.SetModel(model_file, params_file);
+  if (use_gpu) {
+    config.EnableUseGpu(100, gpu_id);
+  } else {
+    config.DisableGpu();
+  }
+  config.SwitchUseFeedFetchOps(false);
+  config.SwitchSpecifyInputNames(true);
+  // 开启内存优化
+  config.EnableMemoryOptim();
+  predictor_ = std::move(CreatePaddlePredictor(config));
+}
+
+bool Model::load_config(const std::string& model_dir) {
+  std::string yaml_file = model_dir + OS_PATH_SEP + "model.yml";
+  YAML::Node config = YAML::LoadFile(yaml_file);
+  type = config["_Attributes"]["model_type"].as<std::string>();
+  name = config["Model"].as<std::string>();
+  bool to_rgb = true;
+  if (config["TransformsMode"].IsDefined()) {
+    std::string mode = config["TransformsMode"].as<std::string>();
+    if (mode == "BGR") {
+      to_rgb = false;
+    } else if (mode != "RGB") {
+      std::cerr << "[Init] Only 'RGB' or 'BGR' is supported for TransformsMode"
+                << std::endl;
+      return false;
+    }
+  }
+  // 构建数据处理流
+  transforms_.Init(config["Transforms"], to_rgb);
+  // 读入label list
+  labels.clear();
+  for (const auto& item : config["_Attributes"]["labels"]) {
+    int index = labels.size();
+    labels[index] = item.as<std::string>();
+  }
+  return true;
+}
+
+bool Model::preprocess(const cv::Mat& input_im, ImageBlob* blob) {
+  cv::Mat im = input_im.clone();
+  if (!transforms_.Run(&im, &inputs_)) {
+    return false;
+  }
+  return true;
+}
+
+bool Model::predict(const cv::Mat& im, ClsResult* result) {
+  inputs_.clear();
+  if (type == "detector") {
+    std::cerr << "Loading model is a 'detector', DetResult should be passed to "
+                 "function predict()!"
+              << std::endl;
+    return false;
+  } else if (type == "segmenter") {
+    std::cerr << "Loading model is a 'segmenter', SegResult should be passed "
+                 "to function predict()!"
+              << std::endl;
+    return false;
+  }
+  // 处理输入图像
+  if (!preprocess(im, &inputs_)) {
+    std::cerr << "Preprocess failed!" << std::endl;
+    return false;
+  }
+  // 使用加载的模型进行预测
+  auto in_tensor = predictor_->GetInputTensor("image");
+  int h = inputs_.new_im_size_[0];
+  int w = inputs_.new_im_size_[1];
+  in_tensor->Reshape({1, 3, h, w});
+  in_tensor->copy_from_cpu(inputs_.im_data_.data());
+  predictor_->ZeroCopyRun();
+  // 取出模型的输出结果
+  auto output_names = predictor_->GetOutputNames();
+  auto output_tensor = predictor_->GetOutputTensor(output_names[0]);
+  std::vector<int> output_shape = output_tensor->shape();
+  int size = 1;
+  for (const auto& i : output_shape) {
+    size *= i;
+  }
+  outputs_.resize(size);
+  output_tensor->copy_to_cpu(outputs_.data());
+  // 对模型输出结果进行后处理
+  auto ptr = std::max_element(std::begin(outputs_), std::end(outputs_));
+  result->category_id = std::distance(std::begin(outputs_), ptr);
+  result->score = *ptr;
+  result->category = labels[result->category_id];
+}
+
+bool Model::predict(const cv::Mat& im, DetResult* result) {
+  result->clear();
+  inputs_.clear();
+  if (type == "classifier") {
+    std::cerr << "Loading model is a 'classifier', ClsResult should be passed "
+                 "to function predict()!"
+              << std::endl;
+    return false;
+  } else if (type == "segmenter") {
+    std::cerr << "Loading model is a 'segmenter', SegResult should be passed "
+                 "to function predict()!"
+              << std::endl;
+    return false;
+  }
+
+  // 处理输入图像
+  if (!preprocess(im, &inputs_)) {
+    std::cerr << "Preprocess failed!" << std::endl;
+    return false;
+  }
+
+  int h = inputs_.new_im_size_[0];
+  int w = inputs_.new_im_size_[1];
+  auto im_tensor = predictor_->GetInputTensor("image");
+  im_tensor->Reshape({1, 3, h, w});
+  im_tensor->copy_from_cpu(inputs_.im_data_.data());
+  if (name == "YOLOv3") {
+    auto im_size_tensor = predictor_->GetInputTensor("im_size");
+    im_size_tensor->Reshape({1, 2});
+    im_size_tensor->copy_from_cpu(inputs_.ori_im_size_.data());
+  } else if (name == "FasterRCNN" || name == "MaskRCNN") {
+    auto im_info_tensor = predictor_->GetInputTensor("im_info");
+    auto im_shape_tensor = predictor_->GetInputTensor("im_shape");
+    im_info_tensor->Reshape({1, 3});
+    im_shape_tensor->Reshape({1, 3});
+    float ori_h = static_cast<float>(inputs_.ori_im_size_[0]);
+    float ori_w = static_cast<float>(inputs_.ori_im_size_[1]);
+    float new_h = static_cast<float>(inputs_.new_im_size_[0]);
+    float new_w = static_cast<float>(inputs_.new_im_size_[1]);
+    float im_info[] = {new_h, new_w, inputs_.scale};
+    float im_shape[] = {ori_h, ori_w, 1.0};
+    im_info_tensor->copy_from_cpu(im_info);
+    im_shape_tensor->copy_from_cpu(im_shape);
+  }
+  // 使用加载的模型进行预测
+  predictor_->ZeroCopyRun();
+
+  std::vector<float> output_box;
+  auto output_names = predictor_->GetOutputNames();
+  auto output_box_tensor = predictor_->GetOutputTensor(output_names[0]);
+  std::vector<int> output_box_shape = output_box_tensor->shape();
+  int size = 1;
+  for (const auto& i : output_box_shape) {
+    size *= i;
+  }
+  output_box.resize(size);
+  output_box_tensor->copy_to_cpu(output_box.data());
+  if (size < 6) {
+    std::cerr << "[WARNING] There's no object detected." << std::endl;
+    return true;
+  }
+  int num_boxes = size / 6;
+  // 解析预测框box
+  for (int i = 0; i < num_boxes; ++i) {
+    Box box;
+    box.category_id = static_cast<int>(round(output_box[i * 6]));
+    box.category = labels[box.category_id];
+    box.score = output_box[i * 6 + 1];
+    float xmin = output_box[i * 6 + 2];
+    float ymin = output_box[i * 6 + 3];
+    float xmax = output_box[i * 6 + 4];
+    float ymax = output_box[i * 6 + 5];
+    float w = xmax - xmin + 1;
+    float h = ymax - ymin + 1;
+    box.coordinate = {xmin, ymin, w, h};
+    result->boxes.push_back(std::move(box));
+  }
+  // 实例分割需解析mask
+  if (name == "MaskRCNN") {
+    std::vector<float> output_mask;
+    auto output_mask_tensor = predictor_->GetOutputTensor(output_names[1]);
+    std::vector<int> output_mask_shape = output_mask_tensor->shape();
+    int masks_size = 1;
+    for (const auto& i : output_mask_shape) {
+      masks_size *= i;
+    }
+    int mask_pixels = output_mask_shape[2] * output_mask_shape[3];
+    int classes = output_mask_shape[1];
+    output_mask.resize(masks_size);
+    output_mask_tensor->copy_to_cpu(output_mask.data());
+    result->mask_resolution = output_mask_shape[2];
+    for (int i = 0; i < result->boxes.size(); ++i) {
+      Box* box = &result->boxes[i];
+      auto begin_mask =
+          output_mask.begin() + (i * classes + box->category_id) * mask_pixels;
+      auto end_mask = begin_mask + mask_pixels;
+      box->mask.data.assign(begin_mask, end_mask);
+      box->mask.shape = {static_cast<int>(box->coordinate[2]),
+                         static_cast<int>(box->coordinate[3])};
+    }
+  }
+}
+
+bool Model::predict(const cv::Mat& im, SegResult* result) {
+  result->clear();
+  inputs_.clear();
+  if (type == "classifier") {
+    std::cerr << "Loading model is a 'classifier', ClsResult should be passed "
+                 "to function predict()!"
+              << std::endl;
+    return false;
+  } else if (type == "detector") {
+    std::cerr << "Loading model is a 'detector', DetResult should be passed to "
+                 "function predict()!"
+              << std::endl;
+    return false;
+  }
+
+  // 处理输入图像
+  if (!preprocess(im, &inputs_)) {
+    std::cerr << "Preprocess failed!" << std::endl;
+    return false;
+  }
+
+  int h = inputs_.new_im_size_[0];
+  int w = inputs_.new_im_size_[1];
+  auto im_tensor = predictor_->GetInputTensor("image");
+  im_tensor->Reshape({1, 3, h, w});
+  im_tensor->copy_from_cpu(inputs_.im_data_.data());
+  std::cout << "input image: " << h << " " << w << std::endl;
+
+  // 使用加载的模型进行预测
+  predictor_->ZeroCopyRun();
+
+  // 获取预测置信度，经过argmax后的labelmap
+  auto output_names = predictor_->GetOutputNames();
+  auto output_label_tensor = predictor_->GetOutputTensor(output_names[0]);
+  std::vector<int> output_label_shape = output_label_tensor->shape();
+  int size = 1;
+  for (const auto& i : output_label_shape) {
+    size *= i;
+    result->label_map.shape.push_back(i);
+  }
+  result->label_map.data.resize(size);
+  output_label_tensor->copy_to_cpu(result->label_map.data.data());
+
+  // 获取预测置信度scoremap
+  auto output_score_tensor = predictor_->GetOutputTensor(output_names[1]);
+  std::vector<int> output_score_shape = output_score_tensor->shape();
+  size = 1;
+  for (const auto& i : output_score_shape) {
+    size *= i;
+    result->score_map.shape.push_back(i);
+  }
+  result->score_map.data.resize(size);
+  output_score_tensor->copy_to_cpu(result->score_map.data.data());
+
+  // 解析输出结果到原图大小
+  std::vector<uint8_t> label_map(result->label_map.data.begin(),
+                                 result->label_map.data.end());
+  cv::Mat mask_label(result->label_map.shape[1],
+                     result->label_map.shape[2],
+                     CV_8UC1,
+                     label_map.data());
+
+  cv::Mat mask_score(result->score_map.shape[2],
+                     result->score_map.shape[3],
+                     CV_32FC1,
+                     result->score_map.data.data());
+
+  for (std::vector<std::string>::reverse_iterator iter =
+           inputs_.reshape_order_.rbegin();
+       iter != inputs_.reshape_order_.rend();
+       ++iter) {
+    if (*iter == "padding") {
+      auto padding_w = inputs_.im_size_before_padding_[0];
+      auto padding_h = inputs_.im_size_before_padding_[1];
+      mask_label = mask_label(cv::Rect(0, 0, padding_w, padding_h));
+      mask_score = mask_score(cv::Rect(0, 0, padding_w, padding_h));
+    } else if (*iter == "resize") {
+      auto resize_w = inputs_.im_size_before_resize_[0];
+      auto resize_h = inputs_.im_size_before_resize_[1];
+      cv::resize(mask_label,
+                 mask_label,
+                 cv::Size(resize_h, resize_w),
+                 0,
+                 0,
+                 cv::INTER_NEAREST);
+      cv::resize(mask_score,
+                 mask_score,
+                 cv::Size(resize_h, resize_w),
+                 0,
+                 0,
+                 cv::INTER_NEAREST);
+    }
+  }
+  result->label_map.data.assign(mask_label.begin<uint8_t>(),
+                                mask_label.end<uint8_t>());
+  result->label_map.shape = {mask_label.rows, mask_label.cols};
+  result->score_map.data.assign(mask_score.begin<float>(),
+                                mask_score.end<float>());
+  result->score_map.shape = {mask_score.rows, mask_score.cols};
+}
+
+}  // namespce of PaddleX

deploy/cpp/src/segmenter.cpp

@@ -0,0 +1,84 @@
+//   Copyright (c) 2020 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 <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "include/paddlex/paddlex.h"
+#include "include/paddlex/visualize.h"
+
+DEFINE_string(model_dir, "", "Path of inference model");
+DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
+DEFINE_int32(gpu_id, 0, "GPU card id");
+DEFINE_string(image, "", "Path of test image file");
+DEFINE_string(image_list, "", "Path of test image list file");
+DEFINE_string(save_dir, "output", "Path to save visualized image");
+
+int main(int argc, char** argv) {
+  // 解析命令行参数
+  google::ParseCommandLineFlags(&argc, &argv, true);
+
+  if (FLAGS_model_dir == "") {
+    std::cerr << "--model_dir need to be defined" << std::endl;
+    return -1;
+  }
+  if (FLAGS_image == "" & FLAGS_image_list == "") {
+    std::cerr << "--image or --image_list need to be defined" << std::endl;
+    return -1;
+  }
+
+  // 加载模型
+  PaddleX::Model model;
+  model.Init(FLAGS_model_dir, FLAGS_use_gpu, FLAGS_gpu_id);
+  auto colormap = PaddleX::GenerateColorMap(model.labels.size());
+  // 进行预测
+  if (FLAGS_image_list != "") {
+    std::ifstream inf(FLAGS_image_list);
+    if (!inf) {
+      std::cerr << "Fail to open file " << FLAGS_image_list << std::endl;
+      return -1;
+    }
+    std::string image_path;
+    while (getline(inf, image_path)) {
+      PaddleX::SegResult result;
+      cv::Mat im = cv::imread(image_path, 1);
+      model.predict(im, &result);
+      // 可视化
+      cv::Mat vis_img =
+          PaddleX::VisualizeSeg(im, result, model.labels, colormap);
+      std::string save_path =
+          PaddleX::generate_save_path(FLAGS_save_dir, image_path);
+      cv::imwrite(save_path, vis_img);
+      result.clear();
+      std::cout << "Visualized output saved as " << save_path << std::endl;
+    }
+  } else {
+    PaddleX::SegResult result;
+    cv::Mat im = cv::imread(FLAGS_image, 1);
+    model.predict(im, &result);
+    // 可视化
+    cv::Mat vis_img = PaddleX::VisualizeSeg(im, result, model.labels, colormap);
+    std::string save_path =
+        PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_image);
+    cv::imwrite(save_path, vis_img);
+    result.clear();
+    std::cout << "Visualized output saved as " << save_path << std::endl;
+  }
+
+  return 0;
+}

deploy/cpp/src/transforms.cpp

@@ -0,0 +1,225 @@
+//   Copyright (c) 2020 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 <iostream>
+#include <string>
+#include <vector>
+
+#include "include/paddlex/transforms.h"
+
+namespace PaddleX {
+
+std::map<std::string, int> interpolations = {{"LINEAR", cv::INTER_LINEAR},
+                                             {"NEAREST", cv::INTER_NEAREST},
+                                             {"AREA", cv::INTER_AREA},
+                                             {"CUBIC", cv::INTER_CUBIC},
+                                             {"LANCZOS4", cv::INTER_LANCZOS4}};
+
+bool Normalize::Run(cv::Mat* im, ImageBlob* data) {
+  for (int h = 0; h < im->rows; h++) {
+    for (int w = 0; w < im->cols; w++) {
+      im->at<cv::Vec3f>(h, w)[0] =
+          (im->at<cv::Vec3f>(h, w)[0] / 255.0 - mean_[0]) / std_[0];
+      im->at<cv::Vec3f>(h, w)[1] =
+          (im->at<cv::Vec3f>(h, w)[1] / 255.0 - mean_[1]) / std_[1];
+      im->at<cv::Vec3f>(h, w)[2] =
+          (im->at<cv::Vec3f>(h, w)[2] / 255.0 - mean_[2]) / std_[2];
+    }
+  }
+  return true;
+}
+
+float ResizeByShort::GenerateScale(const cv::Mat& im) {
+  int origin_w = im.cols;
+  int origin_h = im.rows;
+  int im_size_max = std::max(origin_w, origin_h);
+  int im_size_min = std::min(origin_w, origin_h);
+  float scale =
+      static_cast<float>(short_size_) / static_cast<float>(im_size_min);
+  if (max_size_ > 0) {
+    if (round(scale * im_size_max) > max_size_) {
+      scale = static_cast<float>(max_size_) / static_cast<float>(im_size_max);
+    }
+  }
+  return scale;
+}
+
+bool ResizeByShort::Run(cv::Mat* im, ImageBlob* data) {
+  data->im_size_before_resize_[0] = im->rows;
+  data->im_size_before_resize_[1] = im->cols;
+  data->reshape_order_.push_back("resize");
+
+  float scale = GenerateScale(*im);
+  int width = static_cast<int>(scale * im->cols);
+  int height = static_cast<int>(scale * im->rows);
+  cv::resize(*im, *im, cv::Size(width, height), 0, 0, cv::INTER_LINEAR);
+
+  data->new_im_size_[0] = im->rows;
+  data->new_im_size_[1] = im->cols;
+  data->scale = scale;
+  return true;
+}
+
+bool CenterCrop::Run(cv::Mat* im, ImageBlob* data) {
+  int height = static_cast<int>(im->rows);
+  int width = static_cast<int>(im->cols);
+  if (height < height_ || width < width_) {
+    std::cerr << "[CenterCrop] Image size less than crop size" << std::endl;
+    return false;
+  }
+  int offset_x = static_cast<int>((width - width_) / 2);
+  int offset_y = static_cast<int>((height - height_) / 2);
+  cv::Rect crop_roi(offset_x, offset_y, width_, height_);
+  *im = (*im)(crop_roi);
+  data->new_im_size_[0] = im->rows;
+  data->new_im_size_[1] = im->cols;
+  return true;
+}
+
+bool Padding::Run(cv::Mat* im, ImageBlob* data) {
+  data->im_size_before_padding_[0] = im->rows;
+  data->im_size_before_padding_[1] = im->cols;
+  data->reshape_order_.push_back("padding");
+
+  int padding_w = 0;
+  int padding_h = 0;
+  if (width_ > 0 & height_ > 0) {
+    padding_w = width_ - im->cols;
+    padding_h = height_ - im->rows;
+  } else if (coarsest_stride_ > 0) {
+    padding_h =
+        ceil(im->rows * 1.0 / coarsest_stride_) * coarsest_stride_ - im->rows;
+    padding_w =
+        ceil(im->cols * 1.0 / coarsest_stride_) * coarsest_stride_ - im->cols;
+  }
+  if (padding_h < 0 || padding_w < 0) {
+    std::cerr << "[Padding] Computed padding_h=" << padding_h
+              << ", padding_w=" << padding_w
+              << ", but they should be greater than 0." << std::endl;
+    return false;
+  }
+  cv::copyMakeBorder(
+      *im, *im, 0, padding_h, 0, padding_w, cv::BORDER_CONSTANT, cv::Scalar(0));
+  data->new_im_size_[0] = im->rows;
+  data->new_im_size_[1] = im->cols;
+  return true;
+}
+
+bool ResizeByLong::Run(cv::Mat* im, ImageBlob* data) {
+  if (long_size_ <= 0) {
+    std::cerr << "[ResizeByLong] long_size should be greater than 0"
+              << std::endl;
+    return false;
+  }
+  data->im_size_before_resize_[0] = im->rows;
+  data->im_size_before_resize_[1] = im->cols;
+  data->reshape_order_.push_back("resize");
+  int origin_w = im->cols;
+  int origin_h = im->rows;
+
+  int im_size_max = std::max(origin_w, origin_h);
+  float scale =
+      static_cast<float>(long_size_) / static_cast<float>(im_size_max);
+  cv::resize(*im, *im, cv::Size(), scale, scale, cv::INTER_NEAREST);
+  data->new_im_size_[0] = im->rows;
+  data->new_im_size_[1] = im->cols;
+  data->scale = scale;
+  return true;
+}
+
+bool Resize::Run(cv::Mat* im, ImageBlob* data) {
+  if (width_ <= 0 || height_ <= 0) {
+    std::cerr << "[Resize] width and height should be greater than 0"
+              << std::endl;
+    return false;
+  }
+  if (interpolations.count(interp_) <= 0) {
+    std::cerr << "[Resize] Invalid interpolation method: '" << interp_ << "'"
+              << std::endl;
+    return false;
+  }
+  data->im_size_before_resize_[0] = im->rows;
+  data->im_size_before_resize_[1] = im->cols;
+  data->reshape_order_.push_back("resize");
+
+  cv::resize(
+      *im, *im, cv::Size(width_, height_), 0, 0, interpolations[interp_]);
+  data->new_im_size_[0] = im->rows;
+  data->new_im_size_[1] = im->cols;
+  return true;
+}
+
+void Transforms::Init(const YAML::Node& transforms_node, bool to_rgb) {
+  transforms_.clear();
+  to_rgb_ = to_rgb;
+  for (const auto& item : transforms_node) {
+    std::string name = item.begin()->first.as<std::string>();
+    std::cout << "trans name: " << name << std::endl;
+    std::shared_ptr<Transform> transform = CreateTransform(name);
+    transform->Init(item.begin()->second);
+    transforms_.push_back(transform);
+  }
+}
+
+std::shared_ptr<Transform> Transforms::CreateTransform(
+    const std::string& transform_name) {
+  if (transform_name == "Normalize") {
+    return std::make_shared<Normalize>();
+  } else if (transform_name == "ResizeByShort") {
+    return std::make_shared<ResizeByShort>();
+  } else if (transform_name == "CenterCrop") {
+    return std::make_shared<CenterCrop>();
+  } else if (transform_name == "Resize") {
+    return std::make_shared<Resize>();
+  } else if (transform_name == "Padding") {
+    return std::make_shared<Padding>();
+  } else if (transform_name == "ResizeByLong") {
+    return std::make_shared<ResizeByLong>();
+  } else {
+    std::cerr << "There's unexpected transform(name='" << transform_name
+              << "')." << std::endl;
+    exit(-1);
+  }
+}
+
+bool Transforms::Run(cv::Mat* im, ImageBlob* data) {
+  // 按照transforms中预处理算子顺序处理图像
+  if (to_rgb_) {
+    cv::cvtColor(*im, *im, cv::COLOR_BGR2RGB);
+  }
+  (*im).convertTo(*im, CV_32FC3);
+  data->ori_im_size_[0] = im->rows;
+  data->ori_im_size_[1] = im->cols;
+  data->new_im_size_[0] = im->rows;
+  data->new_im_size_[1] = im->cols;
+  for (int i = 0; i < transforms_.size(); ++i) {
+    if (!transforms_[i]->Run(im, data)) {
+      std::cerr << "Apply transforms to image failed!" << std::endl;
+      return false;
+    }
+  }
+
+  // 将图像由NHWC转为NCHW格式
+  // 同时转为连续的内存块存储到ImageBlob
+  int h = im->rows;
+  int w = im->cols;
+  int c = im->channels();
+  (data->im_data_).resize(c * h * w);
+  float* ptr = (data->im_data_).data();
+  for (int i = 0; i < c; ++i) {
+    cv::extractChannel(*im, cv::Mat(h, w, CV_32FC1, ptr + i * h * w), i);
+  }
+  return true;
+}
+}  // namespace PaddleX

deploy/cpp/src/visualize.cpp

@@ -0,0 +1,148 @@
+//   Copyright (c) 2020 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 "include/paddlex/visualize.h"
+
+namespace PaddleX {
+std::vector<int> GenerateColorMap(int num_class) {
+  auto colormap = std::vector<int>(3 * num_class, 0);
+  for (int i = 0; i < num_class; ++i) {
+    int j = 0;
+    int lab = i;
+    while (lab) {
+      colormap[i * 3] |= (((lab >> 0) & 1) << (7 - j));
+      colormap[i * 3 + 1] |= (((lab >> 1) & 1) << (7 - j));
+      colormap[i * 3 + 2] |= (((lab >> 2) & 1) << (7 - j));
+      ++j;
+      lab >>= 3;
+    }
+  }
+  return colormap;
+}
+
+cv::Mat VisualizeDet(const cv::Mat& img,
+                     const DetResult& result,
+                     const std::map<int, std::string>& labels,
+                     const std::vector<int>& colormap,
+                     float threshold) {
+  cv::Mat vis_img = img.clone();
+  auto boxes = result.boxes;
+  for (int i = 0; i < boxes.size(); ++i) {
+    if (boxes[i].score < threshold) {
+      continue;
+    }
+    cv::Rect roi = cv::Rect(boxes[i].coordinate[0],
+                            boxes[i].coordinate[1],
+                            boxes[i].coordinate[2],
+                            boxes[i].coordinate[3]);
+
+    // 生成预测框和标题
+    std::string text = boxes[i].category;
+    int c1 = colormap[3 * boxes[i].category_id + 0];
+    int c2 = colormap[3 * boxes[i].category_id + 1];
+    int c3 = colormap[3 * boxes[i].category_id + 2];
+    cv::Scalar roi_color = cv::Scalar(c1, c2, c3);
+    text += std::to_string(static_cast<int>(boxes[i].score * 100)) + "%";
+    int font_face = cv::FONT_HERSHEY_SIMPLEX;
+    double font_scale = 0.5f;
+    float thickness = 0.5;
+    cv::Size text_size =
+        cv::getTextSize(text, font_face, font_scale, thickness, nullptr);
+    cv::Point origin;
+    origin.x = roi.x;
+    origin.y = roi.y;
+
+    // 生成预测框标题的背景
+    cv::Rect text_back = cv::Rect(boxes[i].coordinate[0],
+                                  boxes[i].coordinate[1] - text_size.height,
+                                  text_size.width,
+                                  text_size.height);
+
+    // 绘图和文字
+    cv::rectangle(vis_img, roi, roi_color, 2);
+    cv::rectangle(vis_img, text_back, roi_color, -1);
+    cv::putText(vis_img,
+                text,
+                origin,
+                font_face,
+                font_scale,
+                cv::Scalar(255, 255, 255),
+                thickness);
+
+    // 生成实例分割mask
+    if (boxes[i].mask.data.size() == 0) {
+      continue;
+    }
+    cv::Mat bin_mask(result.mask_resolution,
+                     result.mask_resolution,
+                     CV_32FC1,
+                     boxes[i].mask.data.data());
+    cv::resize(bin_mask,
+               bin_mask,
+               cv::Size(boxes[i].mask.shape[0], boxes[i].mask.shape[1]));
+    cv::threshold(bin_mask, bin_mask, 0.5, 1, cv::THRESH_BINARY);
+    cv::Mat full_mask = cv::Mat::zeros(vis_img.size(), CV_8UC1);
+    bin_mask.copyTo(full_mask(roi));
+    cv::Mat mask_ch[3];
+    mask_ch[0] = full_mask * c1;
+    mask_ch[1] = full_mask * c2;
+    mask_ch[2] = full_mask * c3;
+    cv::Mat mask;
+    cv::merge(mask_ch, 3, mask);
+    cv::addWeighted(vis_img, 1, mask, 0.5, 0, vis_img);
+  }
+  return vis_img;
+}
+
+cv::Mat VisualizeSeg(const cv::Mat& img,
+                     const SegResult& result,
+                     const std::map<int, std::string>& labels,
+                     const std::vector<int>& colormap) {
+  std::vector<uint8_t> label_map(result.label_map.data.begin(),
+                                 result.label_map.data.end());
+  cv::Mat mask(result.label_map.shape[0],
+               result.label_map.shape[1],
+               CV_8UC1,
+               label_map.data());
+  cv::Mat color_mask = cv::Mat::zeros(
+      result.label_map.shape[0], result.label_map.shape[1], CV_8UC3);
+  int rows = img.rows;
+  int cols = img.cols;
+  for (int i = 0; i < rows; i++) {
+    for (int j = 0; j < cols; j++) {
+      int category_id = static_cast<int>(mask.at<uchar>(i, j));
+      color_mask.at<cv::Vec3b>(i, j)[0] = colormap[3 * category_id + 0];
+      color_mask.at<cv::Vec3b>(i, j)[1] = colormap[3 * category_id + 1];
+      color_mask.at<cv::Vec3b>(i, j)[2] = colormap[3 * category_id + 2];
+    }
+  }
+  return color_mask;
+}
+
+std::string generate_save_path(const std::string& save_dir,
+                               const std::string& file_path) {
+  if (access(save_dir.c_str(), 0) < 0) {
+#ifdef _WIN32
+    mkdir(save_dir.c_str());
+#else
+    if (mkdir(save_dir.c_str(), S_IRWXU) < 0) {
+      std::cerr << "Fail to create " << save_dir << "directory." << std::endl;
+    }
+#endif
+  }
+  int pos = file_path.find_last_of(OS_PATH_SEP);
+  std::string image_name(file_path.substr(pos + 1));
+  return save_dir + OS_PATH_SEP + image_name;
+}
+}  // namespace of PaddleX

docs/apis/models.md

@@ -17,7 +17,7 @@ paddlex.cls.ResNet50(num_classes=1000)
 #### 分类器训练函数接口
 
 > ```python
-> train(self, num_epochs, train_dataset, train_batch_size=64, eval_dataset=None, save_interval_epochs=1, log_interval_steps=2, save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=0.025, lr_decay_epochs=[30, 60, 90], lr_decay_gamma=0.1, use_vdl=False, sensitivities_file=None, eval_metric_loss=0.05)
+> train(self, num_epochs, train_dataset, train_batch_size=64, eval_dataset=None, save_interval_epochs=1, log_interval_steps=2, save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=0.025, lr_decay_epochs=[30, 60, 90], lr_decay_gamma=0.1, use_vdl=False, sensitivities_file=None, eval_metric_loss=0.05, early_stop=False, early_stop_patience=5)
 > ```
 >
 > **参数：**
@@ -37,6 +37,8 @@ paddlex.cls.ResNet50(num_classes=1000)
 > > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认值为False。
 > > - **sensitivities_file** (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
 > > - **eval_metric_loss** (float): 可容忍的精度损失。默认为0.05。
+> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
 
 #### 分类器评估函数接口
 
@@ -75,7 +77,7 @@ paddlex.cls.ResNet50(num_classes=1000)
 
 ### 其它分类器类
 
-除`ResNet50`外，`paddlex.cls`下还提供了`ResNet18`、`ResNet34`、`ResNet101`、`ResNet50_vd`、`ResNet101_vd`、`DarkNet53`、`MobileNetV1`、`MobileNetV2`、`MobileNetV3_small`、`MobileNetV3_large`、`Xception41`、`Xception65`、`Xception71`、`ShuffleNetV2`,  使用方式（包括函数接口和参数）均与`ResNet50`一致，各模型效果可参考[模型库](../model_zoo.md)中列表。
+除`ResNet50`外，`paddlex.cls`下还提供了`ResNet18`、`ResNet34`、`ResNet101`、`ResNet50_vd`、`ResNet101_vd`、`ResNet50_vd_ssld`、`ResNet101_vd_ssld`、`DarkNet53`、`MobileNetV1`、`MobileNetV2`、`MobileNetV3_small`、`MobileNetV3_large`、`MobileNetV3_small_ssld`、`MobileNetV3_large_ssld`、`Xception41`、`Xception65`、`Xception71`、`ShuffleNetV2`,  使用方式（包括函数接口和参数）均与`ResNet50`一致，各模型效果可参考[模型库](../model_zoo.md)中列表。
 
 
 
@@ -109,7 +111,7 @@ paddlex.det.YOLOv3(num_classes=80, backbone='MobileNetV1', anchors=None, anchor_
 #### YOLOv3训练函数接口
 
 > ```python
-> train(self, num_epochs, train_dataset, train_batch_size=8, eval_dataset=None, save_interval_epochs=20, log_interval_steps=2, save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=1.0/8000, warmup_steps=1000, warmup_start_lr=0.0, lr_decay_epochs=[213, 240], lr_decay_gamma=0.1, metric=None, use_vdl=False, sensitivities_file=None, eval_metric_loss=0.05)
+> train(self, num_epochs, train_dataset, train_batch_size=8, eval_dataset=None, save_interval_epochs=20, log_interval_steps=2, save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=1.0/8000, warmup_steps=1000, warmup_start_lr=0.0, lr_decay_epochs=[213, 240], lr_decay_gamma=0.1, metric=None, use_vdl=False, sensitivities_file=None, eval_metric_loss=0.05, early_stop=False, early_stop_patience=5)
 > ```
 >
 > **参数：**
@@ -132,6 +134,8 @@ paddlex.det.YOLOv3(num_classes=80, backbone='MobileNetV1', anchors=None, anchor_
 > > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认值为False。
 > > - **sensitivities_file** (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，则自动下载在PascalVOC数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
 > > - **eval_metric_loss** (float): 可容忍的精度损失。默认为0.05。
+> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
 
 #### YOLOv3评估函数接口
 
@@ -186,7 +190,7 @@ paddlex.det.FasterRCNN(num_classes=81, backbone='ResNet50', with_fpn=True, aspec
 #### FasterRCNN训练函数接口
 
 > ```python
-> train(self, num_epochs, train_dataset, train_batch_size=2, eval_dataset=None, save_interval_epochs=1, log_interval_steps=2,save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=0.0025, warmup_steps=500, warmup_start_lr=1.0/1200, lr_decay_epochs=[8, 11], lr_decay_gamma=0.1, metric=None, use_vdl=False)
+> train(self, num_epochs, train_dataset, train_batch_size=2, eval_dataset=None, save_interval_epochs=1, log_interval_steps=2,save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=0.0025, warmup_steps=500, warmup_start_lr=1.0/1200, lr_decay_epochs=[8, 11], lr_decay_gamma=0.1, metric=None, use_vdl=False, early_stop=False, early_stop_patience=5)
 >
 > ```
 >
@@ -208,6 +212,8 @@ paddlex.det.FasterRCNN(num_classes=81, backbone='ResNet50', with_fpn=True, aspec
 > > - **lr_decay_gamma** (float): 默认优化器的学习率衰减率。默认为0.1。
 > > - **metric** (bool): 训练过程中评估的方式，取值范围为['COCO', 'VOC']。默认值为None。
 > > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认值为False。
+> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
 
 #### FasterRCNN评估函数接口
 
@@ -264,7 +270,7 @@ paddlex.det.MaskRCNN(num_classes=81, backbone='ResNet50', with_fpn=True, aspect_
 #### MaskRCNN训练函数接口
 
 > ```python
-> train(self, num_epochs, train_dataset, train_batch_size=1, eval_dataset=None, save_interval_epochs=1, log_interval_steps=20, save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=1.0/800, warmup_steps=500, warmup_start_lr=1.0 / 2400, lr_decay_epochs=[8, 11], lr_decay_gamma=0.1, metric=None, use_vdl=False)
+> train(self, num_epochs, train_dataset, train_batch_size=1, eval_dataset=None, save_interval_epochs=1, log_interval_steps=20, save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=1.0/800, warmup_steps=500, warmup_start_lr=1.0 / 2400, lr_decay_epochs=[8, 11], lr_decay_gamma=0.1, metric=None, use_vdl=False, early_stop=False, early_stop_patience=5)
 >
 > ```
 >
@@ -286,6 +292,8 @@ paddlex.det.MaskRCNN(num_classes=81, backbone='ResNet50', with_fpn=True, aspect_
 > > - **lr_decay_gamma** (float): 默认优化器的学习率衰减率。默认为0.1。
 > > - **metric** (bool): 训练过程中评估的方式，取值范围为['COCO', 'VOC']。默认值为None。
 > > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认值为False。
+> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
 
 #### MaskRCNN评估函数接口
 
@@ -350,7 +358,7 @@ paddlex.seg.DeepLabv3p(num_classes=2, backbone='MobileNetV2_x1.0', output_stride
 #### DeepLabv3训练函数接口
 
 > ```python
-> train(self, num_epochs, train_dataset, train_batch_size=2, eval_dataset=None, eval_batch_size=1, save_interval_epochs=1, log_interval_steps=2, save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=0.01, lr_decay_power=0.9, use_vdl=False, sensitivities_file=None, eval_metric_loss=0.05):
+> train(self, num_epochs, train_dataset, train_batch_size=2, eval_dataset=None, eval_batch_size=1, save_interval_epochs=1, log_interval_steps=2, save_dir='output', pretrain_weights='IMAGENET', optimizer=None, learning_rate=0.01, lr_decay_power=0.9, use_vdl=False, sensitivities_file=None, eval_metric_loss=0.05, early_stop=False, early_stop_patience=5):
 >
 > ```
 >
@@ -370,6 +378,8 @@ paddlex.seg.DeepLabv3p(num_classes=2, backbone='MobileNetV2_x1.0', output_stride
 > > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认False。
 > > - **sensitivities_file** (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
 > > - **eval_metric_loss** (float): 可容忍的精度损失。默认为0.05。
+> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
 
 #### DeepLabv3评估函数接口
 
@@ -427,7 +437,7 @@ paddlex.seg.UNet(num_classes=2, upsample_mode='bilinear', use_bce_loss=False, us
 #### Unet训练函数接口
 
 > ```python
-> train(self, num_epochs, train_dataset, train_batch_size=2, eval_dataset=None, eval_batch_size=1, save_interval_epochs=1, log_interval_steps=2, save_dir='output', pretrain_weights='COCO', optimizer=None, learning_rate=0.01, lr_decay_power=0.9, use_vdl=False, sensitivities_file=None, eval_metric_loss=0.05):
+> train(self, num_epochs, train_dataset, train_batch_size=2, eval_dataset=None, eval_batch_size=1, save_interval_epochs=1, log_interval_steps=2, save_dir='output', pretrain_weights='COCO', optimizer=None, learning_rate=0.01, lr_decay_power=0.9, use_vdl=False, sensitivities_file=None, eval_metric_loss=0.05, early_stop=False, early_stop_patience=5):
 > ```
 >
 > **参数：**
@@ -446,6 +456,8 @@ paddlex.seg.UNet(num_classes=2, upsample_mode='bilinear', use_bce_loss=False, us
 > > - **use_vdl** (bool): 是否使用VisualDL进行可视化。默认False。
 > > - **sensitivities_file** (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
 > > - **eval_metric_loss** (float): 可容忍的精度损失。默认为0.05。
+> > - **early_stop** (float): 是否使用提前终止训练策略。默认值为False。
+> > - **early_stop_patience** (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内连续下降或持平，则终止训练。默认值为5。
 
 #### Unet评估函数接口
 

docs/deploy.md

@@ -1,13 +0,0 @@
-# 模型部署导出
-
-### 导出inference模型
-
-在服务端部署的模型需要首先将模型导出为inference格式模型，导出的模型将包括`__model__`、`__params__`和`model.yml`三个文名，分别为模型的网络结构，模型权重和模型的配置文件（包括数据预处理参数等等）。在安装完PaddleX后，在命令行终端使用如下命令导出模型到当前目录`inferece_model`下。
-
-> 可直接下载小度熊分拣模型测试本文档的流程[xiaoduxiong_epoch_12.tar.gz](https://bj.bcebos.com/paddlex/models/xiaoduxiong_epoch_12.tar.gz)
-
-```
-paddlex --export_inference --model_dir=./xiaoduxiong_epoch_12 --save_dir=./inference_model
-```
-
-## 模型C++和Python部署方案预计一周内推出...

docs/deploy/deploy.md

@@ -0,0 +1,59 @@
+# 模型预测部署
+
+本文档指引用户如何采用更高性能地方式来部署使用PaddleX训练的模型。使用本文档模型部署方式，会在模型运算过程中，对模型计算图进行优化，同时减少内存操作，相对比普通的paddlepaddle模型加载和预测方式，预测速度平均可提升1倍，具体各模型性能对比见[预测性能对比](#预测性能对比)
+
+## 服务端部署
+
+### 导出inference模型
+
+在服务端部署的模型需要首先将模型导出为inference格式模型，导出的模型将包括`__model__`、`__params__`和`model.yml`三个文名，分别为模型的网络结构，模型权重和模型的配置文件（包括数据预处理参数等等）。在安装完PaddleX后，在命令行终端使用如下命令导出模型到当前目录`inferece_model`下。
+
+> 可直接下载垃圾检测模型测试本文档的流程[garbage_epoch_12.tar.gz](https://bj.bcebos.com/paddlex/models/garbage_epoch_12.tar.gz)
+
+```
+paddlex --export_inference --model_dir=./garbage_epoch_12 --save_dir=./inference_model
+```
+
+### Python部署
+PaddleX已经集成了基于Python的高性能预测接口，在安装PaddleX后，可参照如下代码示例，进行预测。相关的接口文档可参考[paddlex.deploy](apis/deploy.md)
+> 点击下载测试图片 [garbage.bmp](https://bj.bcebos.com/paddlex/datasets/garbage.bmp)
+```
+import paddlex as pdx
+predictorpdx.deploy.create_predictor('./inference_model')
+result = predictor.predict(image='garbage.bmp')
+```
+
+### C++部署
+
+C++部署方案位于目录`deploy/cpp/`下，且独立于PaddleX其他模块。该方案支持在 Windows 和 Linux 完成编译、二次开发集成和部署运行。具体使用方法和编译：
+
+- Linux平台：[linux](deploy_cpp_linux.md)
+- window平台：[windows](deploy_cpp_win_vs2019.md)
+
+### 预测性能对比
+
+#### 测试环境
+
+- CUDA 9.0
+- CUDNN 7.5
+- PaddlePaddle 1.71
+- GPU: Tesla P40
+- AnalysisPredictor 指采用Python的高性能预测方式
+- Executor 指采用paddlepaddle普通的python预测方式
+- Batch Size均为1，耗时单位为ms/image，只计算模型运行时间，不包括数据的预处理和后处理
+
+| 模型 | AnalysisPredictor耗时 | Executor耗时 | 输入图像大小 |
+| :---- | :--------------------- | :------------ | :------------ |
+| resnet50 | 4.84 | 7.57 | 224*224 |
+| mobilenet_v2 | 3.27 | 5.76 | 224*224 |
+| unet | 22.51 | 34.60 |513*513 |
+| deeplab_mobile | 63.44 | 358.31 |1025*2049 |
+| yolo_mobilenetv2 | 15.20 | 19.54 |  608*608 |
+| faster_rcnn_r50_fpn_1x | 50.05 | 69.58 |800*1088 |
+| faster_rcnn_r50_1x | 326.11 | 347.22 | 800*1067 |
+| mask_rcnn_r50_fpn_1x | 67.49 | 91.02 | 800*1088 |
+| mask_rcnn_r50_1x | 326.11 | 350.94 | 800*1067 |
+
+## 移动端部署
+
+> Lite模型导出正在集成中，即将开源...

docs/deploy/deploy_cpp_linux.md

@@ -0,0 +1,116 @@
+# Linux平台编译指南
+
+## 说明
+本文档在 `Linux`平台使用`GCC 4.8.5` 和 `GCC 4.9.4`测试过，如果需要使用更高G++版本编译使用，则需要重新编译Paddle预测库，请参考: [从源码编译Paddle预测库](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html#id12)。
+
+## 前置条件
+* G++ 4.8.2 ~ 4.9.4
+* CUDA 9.0 / CUDA 10.0, CUDNN 7+ （仅在使用GPU版本的预测库时需要）
+* CMake 3.0+
+
+请确保系统已经安装好上述基本软件，**下面所有示例以工作目录 `/root/projects/`演示**。
+
+### Step1: 下载代码
+
+ `git clone https://github.com/PaddlePaddle/PaddleX.git`
+
+**说明**：其中`C++`预测代码在`/root/projects/PaddleX/deploy/cpp` 目录，该目录不依赖任何`PaddleX`下其他目录。
+
+
+### Step2: 下载PaddlePaddle C++ 预测库 fluid_inference
+
+PaddlePaddle C++ 预测库针对不同的`CPU`，`CUDA`，以及是否支持TensorRT，提供了不同的预编译版本，请根据实际情况下载:  [C++预测库下载列表](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html#id1)
+
+
+下载并解压后`/root/projects/fluid_inference`目录包含内容为：
+```
+fluid_inference
+├── paddle # paddle核心库和头文件
+|
+├── third_party # 第三方依赖库和头文件
+|
+└── version.txt # 版本和编译信息
+```
+
+**注意:** 预编译版本除`nv-jetson-cuda10-cudnn7.5-trt5` 以外其它包都是基于`GCC 4.8.5`编译，使用高版本`GCC`可能存在 `ABI`兼容性问题，建议降级或[自行编译预测库](https://www.paddlepaddle.org.cn/documentation/docs/zh/advanced_guide/inference_deployment/inference/build_and_install_lib_cn.html#id12)。
+
+
+### Step4: 编译
+
+编译`cmake`的命令在`scripts/build.sh`中，请根据实际情况修改主要参数，其主要内容说明如下：
+```
+# 是否使用GPU(即是否使用 CUDA)
+WITH_GPU=ON
+# 是否集成 TensorRT(仅WITH_GPU=ON 有效)
+WITH_TENSORRT=OFF
+# 上一步下载的 Paddle 预测库路径
+PADDLE_DIR=/root/projects/deps/fluid_inference/
+# CUDA 的 lib 路径
+CUDA_LIB=/usr/local/cuda/lib64/
+# CUDNN 的 lib 路径
+CUDNN_LIB=/usr/local/cudnn/lib64/
+
+# OPENCV 路径, 如果使用自带预编译版本可不设置
+OPENCV_DIR=$(pwd)/deps/opencv3gcc4.8/
+sh $(pwd)/scripts/bootstrap.sh
+
+# 以下无需改动
+rm -rf build
+mkdir -p build
+cd build
+cmake .. \
+    -DWITH_GPU=${WITH_GPU} \
+    -DWITH_TENSORRT=${WITH_TENSORRT} \
+    -DPADDLE_DIR=${PADDLE_DIR} \
+    -DCUDA_LIB=${CUDA_LIB} \
+    -DCUDNN_LIB=${CUDNN_LIB} \
+    -DOPENCV_DIR=${OPENCV_DIR}
+make
+
+```
+
+修改脚本设置好主要参数后，执行`build`脚本：
+ ```shell
+ sh ./scripts/build.sh
+ ```
+
+### Step5: 预测及可视化
+编译成功后，预测demo的可执行程序分别为`build/detector`，`build/classifer`，`build/segmenter`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
+
+|  参数   | 说明  |
+|  ----  | ----  |
+| model_dir  | 导出的预测模型所在路径 |
+| image  | 要预测的图片文件路径 |
+| image_list  | 按行存储图片路径的.txt文件 |
+| use_gpu  | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
+| gpu_id  | GPU 设备ID, 默认值为0 |
+| save_dir | 保存可视化结果的路径, 默认值为"output"，classfier无该参数 |
+
+## 样例
+
+可使用[垃圾检测模型](deploy.md#导出inference模型)中生成的`inference_model`模型和测试图片进行预测。
+
+`样例一`：
+
+不使用`GPU`测试图片 `/path/to/garbage.bmp`  
+
+```shell
+./build/detector --model_dir=/path/to/inference_model --image=/path/to/garbage.bmp --save_dir=output
+```
+图片文件`可视化预测结果`会保存在`save_dir`参数设置的目录下。
+
+
+`样例二`:
+
+使用`GPU`预测多个图片`/path/to/image_list.txt`，image_list.txt内容的格式如下：
+```
+/path/to/images/garbage1.jpeg
+/path/to/images/garbage2.jpeg
+...
+/path/to/images/garbagen.jpeg
+```
+```shell
+./build/detector --model_dir=/path/to/models/inference_model --image_list=/root/projects/images_list.txt --use_gpu=1 --save_dir=output
+```
+图片文件`可视化预测结果`会保存在`save_dir`参数设置的目录下。
+

docs/deploy/deploy_cpp_win_vs2019.md

@@ -0,0 +1,138 @@
+# Visual Studio 2019 Community CMake 编译指南
+
+## 说明
+Windows 平台下，我们使用`Visual Studio 2019 Community` 进行了测试。微软从`Visual Studio 2017`开始即支持直接管理`CMake`跨平台编译项目，但是直到`2019`才提供了稳定和完全的支持，所以如果你想使用CMake管理项目编译构建，我们推荐你使用`Visual Studio 2019`环境下构建。
+
+## 前置条件
+* Visual Studio 2019
+* CUDA 9.0 / CUDA 10.0, CUDNN 7+ （仅在使用GPU版本的预测库时需要）
+* CMake 3.0+
+
+请确保系统已经安装好上述基本软件，我们使用的是`VS2019`的社区版。
+
+**下面所有示例以工作目录为 `D:\projects`演示**。
+
+### Step1: 下载代码
+
+下载源代码
+```shell
+d:
+mkdir projects
+cd projects
+git clone https://github.com/PaddlePaddle/PaddleX.git
+```
+
+**说明**：其中`C++`预测代码在`PaddleX/deploy/cpp` 目录，该目录不依赖任何`PaddleX`下其他目录。
+
+
+### Step2: 下载PaddlePaddle C++ 预测库 fluid_inference
+
+PaddlePaddle C++ 预测库针对不同的`CPU`和`CUDA`版本提供了不同的预编译版本，请根据实际情况下载:  [C++预测库下载列表](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/advanced_guide/inference_deployment/inference/windows_cpp_inference.html)
+
+解压后`D:\projects\fluid_inference*\`目录下主要包含的内容为：
+```
+├── \paddle\ # paddle核心库和头文件
+|
+├── \third_party\ # 第三方依赖库和头文件
+|
+└── \version.txt # 版本和编译信息
+```
+
+### Step3: 安装配置OpenCV
+
+1. 在OpenCV官网下载适用于Windows平台的3.4.6版本， [下载地址](https://sourceforge.net/projects/opencvlibrary/files/3.4.6/opencv-3.4.6-vc14_vc15.exe/download)  
+2. 运行下载的可执行文件，将OpenCV解压至指定目录，如`D:\projects\opencv`
+3. 配置环境变量，如下流程所示  
+    - 我的电脑->属性->高级系统设置->环境变量
+    - 在系统变量中找到Path（如没有，自行创建），并双击编辑
+    - 新建，将opencv路径填入并保存，如`D:\projects\opencv\build\x64\vc14\bin`
+
+### Step4: 使用Visual Studio 2019直接编译CMake
+
+1. 打开Visual Studio 2019 Community，点击`继续但无需代码`
+
+![step2](images/vs2019_step1.png)
+
+2. 点击： `文件`->`打开`->`CMake`
+
+![step2.1](images/vs2019_step2.png)
+
+选择项目代码所在路径，并打开`CMakeList.txt`：
+
+![step2.2](images/vs2019_step3.png)
+
+3. 点击：`项目`->`PADDLEX_INFERENCE的CMake设置`
+
+![step3](images/vs2019_step4.png)
+
+4. 点击`浏览`，分别设置编译选项指定`CUDA`、`OpenCV`、`Paddle预测库`的路径
+
+依赖库路径的含义说明如下（带*表示仅在使用**GPU版本**预测库时指定, 其中CUDA库版本尽量对齐，**使用9.0、10.0版本，不使用9.2、10.1等版本CUDA库**）：
+
+|  参数名   | 含义  |
+|  ----  | ----  |
+| *CUDA_LIB  | CUDA的库路径, 注：请将CUDNN的cudnn.lib文件拷贝到CUDA_LIB路径下 |
+| OPENCV_DIR  | OpenCV的安装路径， |
+| PADDLE_DIR | Paddle c++预测库的路径 |
+
+**注意：** 1. 使用`CPU`版预测库，请把`WITH_GPU`的`值`去掉勾 2. 如果使用的是`openblas`版本，请把`WITH_MKL`的`值`去掉勾
+
+![step4](images/vs2019_step5.png)
+
+**设置完成后**, 点击上图中`保存并生成CMake缓存以加载变量`。
+
+5. 点击`生成`->`全部生成`
+
+![step6](images/vs2019_step6.png)
+
+
+### Step5: 预测及可视化
+
+上述`Visual Studio 2019`编译产出的可执行文件在`out\build\x64-Release`目录下，打开`cmd`，并切换到该目录：
+
+```
+d:
+cd D:\projects\PaddleX\deploy\cpp\out\build\x64-Release
+```
+
+编译成功后，预测demo的入口程序为`detector`，`classifer`，`segmenter`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
+
+|  参数   | 说明  |
+|  ----  | ----  |
+| model_dir  | 导出的预测模型所在路径 |
+| image  | 要预测的图片文件路径 |
+| image_list  | 按行存储图片路径的.txt文件 |
+| use_gpu  | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
+| gpu_id  | GPU 设备ID, 默认值为0 |
+| save_dir | 保存可视化结果的路径, 默认值为"output"，classfier无该参数 |
+
+
+## 样例
+
+可使用[垃圾检测模型](deploy.md#导出inference模型)中生成的`inference_model`模型和测试图片进行预测。
+
+`样例一`：
+
+不使用`GPU`测试图片  `\\path\\to\\garbage.bmp`  
+
+```shell
+.\detector --model_dir=\\path\\to\\inference_model --image=D:\\images\\garbage.bmp --save_dir=output
+
+```
+图片文件`可视化预测结果`会保存在`save_dir`参数设置的目录下。
+
+
+`样例二`:
+
+使用`GPU`预测多个图片`\\path\\to\\image_list.txt`，image_list.txt内容的格式如下：
+```
+\\path\\to\\images\\garbage1.jpeg
+\\path\\to\\images\\garbage2.jpeg
+...
+\\path\\to\\images\\garbagen.jpeg
+```
+```shell
+.\detector --model_dir=\\path\\to\\inference_model --image_list=\\path\\to\\images_list.txt --use_gpu=1 --save_dir=output
+```
+图片文件`可视化预测结果`会保存在`save_dir`参数设置的目录下。
+

docs/model_zoo.md

@@ -4,35 +4,38 @@
 表中相关模型也可下载好作为相应模型的预训练模型，通过`pretrain_weights`指定目录加载使用。
 
 ## 图像分类模型
-> 表中模型相关指标均为在ImageNet数据集上使用PaddlePaddle Python预测接口测试得到（测试GPU型号为Nvidia Tesla P4），预测速度为每张图片预测用时（不包括预处理和后处理）,表中符号`-`表示相关指标暂未测试。
+> 表中模型相关指标均为在ImageNet数据集上使用PaddlePaddle Python预测接口测试得到（测试GPU型号为Nvidia Tesla P40），预测速度为每张图片预测用时（不包括预处理和后处理）,表中符号`-`表示相关指标暂未测试。
 
 
-| 模型  | 模型大小 | 预测速度（毫秒） | Top1准确率 | Top5准确率 |
+| 模型  | 模型大小 | 预测速度（毫秒） | Top1准确率（%） | Top5准确率（%） |
 | :----|  :------- | :----------- | :--------- | :--------- |
-| ResNet18| 46.9MB   | 3.456        | 70.98%     | 89.92%     |
-| ResNet34| 87.5MB   | 5.668        | 74.57%     | 92.14%     |
-| ResNet50| 102.7MB  | 8.787        | 76.50%     | 93.00%     |
-| ResNet101 |179.1MB  | 15.447      | 77.56%     | 93.64%    |
-| ResNet50_vd |102.8MB  | 9.058        | 79.12%     | 94.44%     |
-| ResNet101_vd| 179.2MB  | 15.685       | 80.17%     | 94.97%     |
-| DarkNet53|166.9MB  | 11.969       | 78.04%     | 94.05%     |
-| MobileNetV1 | 16.4MB   | 2.609        | 70.99%     | 89.68%     |
-| MobileNetV2 | 14.4MB   | 4.546        | 72.15%     | 90.65%     |
-| MobileNetV3_large|  22.8MB   | -        | 75.3%     | 75.3%     |
-| MobileNetV3_small |  12.5MB   | 6.809        | 67.46%     | 87.12%     |
-| Xception41 |92.4MB   | 13.757       | 79.30%     | 94.53%     |
-| Xception65 | 144.6MB  | 19.216       | 81.00%     | 95.49%     |
-| Xception71| 151.9MB  | 23.291       | 81.11%     | 95.45%     |
-| DenseNet121 | 32.8MB   | 12.437       | 75.66%     | 92.58%     |
-| DenseNet161|116.3MB  | 27.717       | 78.57%     | 94.14%     |
-| DenseNet201|  84.6MB   | 26.583       | 77.63%     | 93.66%     |
-| ShuffleNetV2 | 10.2MB   | 6.101        | 68.8%     | 88.5%     |
+| ResNet18| 46.9MB   | 1.499        | 71.0     | 89.9     |
+| ResNet34| 87.5MB   | 2.272        | 74.6    | 92.1    |
+| ResNet50| 102.7MB  | 2.939        | 76.5     | 93.0     |
+| ResNet101 |179.1MB  | 5.314      | 77.6     | 93.6  |
+| ResNet50_vd |102.8MB  | 3.165        | 79.1     | 94.4     |
+| ResNet101_vd| 179.2MB  | 5.252       | 80.2   | 95.0     |
+| ResNet50_vd_ssld |102.8MB  | 3.165        | 82.4     | 96.1     |
+| ResNet101_vd_ssld| 179.2MB  | 5.252       | 83.7   | 96.7     |
+| DarkNet53|166.9MB  | 3.139       | 78.0     | 94.1     |
+| MobileNetV1 | 16.0MB   | 32.523        | 71.0     | 89.7    |
+| MobileNetV2 | 14.0MB   | 23.318        | 72.2     | 90.7    |
+| MobileNetV3_large|  21.0MB   | 19.308        | 75.3    | 93.2   |
+| MobileNetV3_small |  12.0MB   | 6.546        | 68.2    | 88.1     |
+| MobileNetV3_large_ssld|  21.0MB   | 19.308        | 79.0     | 94.5     |
+| MobileNetV3_small_ssld |  12.0MB   | 6.546        | 71.3     | 90.1     |
+| Xception41 |92.4MB   | 4.408       | 79.6    | 94.4     |
+| Xception65 | 144.6MB  | 6.464       | 80.3     | 94.5     |
+| DenseNet121 | 32.8MB   | 4.371       | 75.7     | 92.6     |
+| DenseNet161|116.3MB  | 8.863       | 78.6     | 94.1     |
+| DenseNet201|  84.6MB   | 8.173       | 77.6     | 93.7     |
+| ShuffleNetV2 | 9.0MB   | 10.941        | 68.8     | 88.5     |
 
 ## 目标检测模型
 
 > 表中模型相关指标均为在MSCOCO数据集上使用PaddlePaddle Python预测接口测试得到（测试GPU型号为Nvidia Tesla V100测试得到,表中符号`-`表示相关指标暂未测试。
 
-| 模型    | 模型大小    | 预测时间(毫秒) | BoxAP |
+| 模型    | 模型大小    | 预测时间(毫秒) | BoxAP（%） |
 |:-------|:-----------|:-------------|:----------|
 |FasterRCNN-ResNet50|135.6MB| 78.450 | 35.2 |
 |FasterRCNN-ResNet50_vd| 135.7MB | 79.523 | 36.4 |
@@ -50,7 +53,7 @@
 
 > 表中模型相关指标均为在MSCOCO数据集上测试得到。
 
-| 模型 |模型大小 | 预测时间(毫秒) | BoxAP | SegAP |
+| 模型 |模型大小 | 预测时间(毫秒) | BoxAP | SegAP（%） |
 |:---------|:---------|:----------|:---------|:--------|
 |MaskRCNN-ResNet50|51.2MB| 86.096 | 36.5 |32.2|
 |MaskRCNN-ResNet50-FPN|184.6MB | 65.859 | 37.9 |34.2|

paddlex/__init__.py

@@ -20,6 +20,12 @@ from . import seg
 from . import cls
 from . import slim
 
+try:
+    import pycocotools
+except:
+    print("[WARNING] pycocotools is not installed, detection model is not available now.")
+    print("[WARNING] pycocotools install: https://github.com/PaddlePaddle/PaddleX/blob/develop/docs/install.md")
+
 env_info = get_environ_info()
 load_model = cv.models.load_model
 datasets = cv.datasets

paddlex/cv/datasets/coco.py

@@ -19,7 +19,6 @@ import random
 import numpy as np
 import paddlex.utils.logging as logging
 import paddlex as pst
-from pycocotools.coco import COCO
 from .voc import VOCDetection
 from .dataset import is_pic
 
@@ -47,6 +46,8 @@ class CocoDetection(VOCDetection):
                  buffer_size=100,
                  parallel_method='process',
                  shuffle=False):
+        from pycocotools.coco import COCO
+
         super(VOCDetection, self).__init__(
             transforms=transforms,
             num_workers=num_workers,

paddlex/cv/datasets/voc.py

@@ -18,7 +18,6 @@ import os.path as osp
 import random
 import numpy as np
 import xml.etree.ElementTree as ET
-from pycocotools.coco import COCO
 import paddlex.utils.logging as logging
 from .dataset import Dataset
 from .dataset import is_pic
@@ -51,6 +50,7 @@ class VOCDetection(Dataset):
                  buffer_size=100,
                  parallel_method='process',
                  shuffle=False):
+        from pycocotools.coco import COCO
         super(VOCDetection, self).__init__(
             transforms=transforms,
             num_workers=num_workers,

paddlex/cv/models/base.py

@@ -24,6 +24,7 @@ import json
 import functools
 import paddlex.utils.logging as logging
 from paddlex.utils import seconds_to_hms
+from paddlex.utils.utils import EarlyStop
 import paddlex
 from collections import OrderedDict
 from os import path as osp
@@ -334,7 +335,9 @@ class BaseAPI:
                    save_interval_epochs=1,
                    log_interval_steps=10,
                    save_dir='output',
-                   use_vdl=False):
+                   use_vdl=False,
+                   early_stop=False,
+                   early_stop_patience=5):
         if not osp.isdir(save_dir):
             if osp.exists(save_dir):
                 os.remove(save_dir)
@@ -396,6 +399,9 @@ class BaseAPI:
             train_step_component = OrderedDict()
             eval_component = OrderedDict()
 
+        thresh = 0.0001
+        if early_stop:
+            earlystop = EarlyStop(early_stop_patience, thresh)
         best_accuracy_key = ""
         best_accuracy = -1.0
         best_model_epoch = 1
@@ -507,3 +513,6 @@ class BaseAPI:
                     'Current evaluated best model in eval_dataset is epoch_{}, {}={}'
                     .format(best_model_epoch, best_accuracy_key,
                             best_accuracy))
+                if eval_dataset is not None and early_stop:
+                    if earlystop(current_accuracy):
+                        break

paddlex/cv/models/classifier.py

@@ -102,7 +102,9 @@ class BaseClassifier(BaseAPI):
               lr_decay_gamma=0.1,
               use_vdl=False,
               sensitivities_file=None,
-              eval_metric_loss=0.05):
+              eval_metric_loss=0.05,
+              early_stop=False,
+              early_stop_patience=5):
         """训练。
 
         Args:
@@ -124,6 +126,9 @@ class BaseClassifier(BaseAPI):
             sensitivities_file (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，
                 则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
             eval_metric_loss (float): 可容忍的精度损失。默认为0.05。
+            early_stop (bool): 是否使用提前终止训练策略。默认值为False。
+            early_stop_patience (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内
+                连续下降或持平，则终止训练。默认值为5。
 
         Raises:
             ValueError: 模型从inference model进行加载。
@@ -158,7 +163,9 @@ class BaseClassifier(BaseAPI):
             save_interval_epochs=save_interval_epochs,
             log_interval_steps=log_interval_steps,
             save_dir=save_dir,
-            use_vdl=use_vdl)
+            use_vdl=use_vdl,
+            early_stop=early_stop,
+            early_stop_patience=early_stop_patience)
 
     def evaluate(self,
                  eval_dataset,

paddlex/cv/models/deeplabv3p.py

@@ -231,7 +231,9 @@ class DeepLabv3p(BaseAPI):
               lr_decay_power=0.9,
               use_vdl=False,
               sensitivities_file=None,
-              eval_metric_loss=0.05):
+              eval_metric_loss=0.05,
+              early_stop=False,
+              early_stop_patience=5):
         """训练。
 
         Args:
@@ -252,6 +254,9 @@ class DeepLabv3p(BaseAPI):
             sensitivities_file (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，
                 则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
             eval_metric_loss (float): 可容忍的精度损失。默认为0.05。
+            early_stop (bool): 是否使用提前终止训练策略。默认值为False。
+            early_stop_patience (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内
+                连续下降或持平，则终止训练。默认值为5。
 
         Raises:
             ValueError: 模型从inference model进行加载。
@@ -288,7 +293,9 @@ class DeepLabv3p(BaseAPI):
             save_interval_epochs=save_interval_epochs,
             log_interval_steps=log_interval_steps,
             save_dir=save_dir,
-            use_vdl=use_vdl)
+            use_vdl=use_vdl,
+            early_stop=early_stop,
+            early_stop_patience=early_stop_patience)
 
     def evaluate(self,
                  eval_dataset,

paddlex/cv/models/faster_rcnn.py

@@ -163,7 +163,9 @@ class FasterRCNN(BaseAPI):
               lr_decay_epochs=[8, 11],
               lr_decay_gamma=0.1,
               metric=None,
-              use_vdl=False):
+              use_vdl=False,
+              early_stop=False,
+              early_stop_patience=5):
         """训练。
 
         Args:
@@ -186,6 +188,9 @@ class FasterRCNN(BaseAPI):
             lr_decay_gamma (float): 默认优化器的学习率衰减率。默认为0.1。
             metric (bool): 训练过程中评估的方式，取值范围为['COCO', 'VOC']。默认值为None。
             use_vdl (bool): 是否使用VisualDL进行可视化。默认值为False。
+            early_stop (bool): 是否使用提前终止训练策略。默认值为False。
+            early_stop_patience (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内
+                连续下降或持平，则终止训练。默认值为5。
 
         Raises:
             ValueError: 评估类型不在指定列表中。
@@ -233,7 +238,9 @@ class FasterRCNN(BaseAPI):
             save_interval_epochs=save_interval_epochs,
             log_interval_steps=log_interval_steps,
             save_dir=save_dir,
-            use_vdl=use_vdl)
+            use_vdl=use_vdl,
+            early_stop=early_stop,
+            early_stop_patience=early_stop_patience)
 
     def evaluate(self,
                  eval_dataset,

paddlex/cv/models/mask_rcnn.py

@@ -128,7 +128,9 @@ class MaskRCNN(FasterRCNN):
               lr_decay_epochs=[8, 11],
               lr_decay_gamma=0.1,
               metric=None,
-              use_vdl=False):
+              use_vdl=False,
+              early_stop=False,
+              early_stop_patience=5):
         """训练。
 
         Args:
@@ -151,6 +153,9 @@ class MaskRCNN(FasterRCNN):
             lr_decay_gamma (float): 默认优化器的学习率衰减率。默认为0.1。
             metric (bool): 训练过程中评估的方式，取值范围为['COCO', 'VOC']。
             use_vdl (bool): 是否使用VisualDL进行可视化。默认值为False。
+            early_stop (bool): 是否使用提前终止训练策略。默认值为False。
+            early_stop_patience (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内
+                连续下降或持平，则终止训练。默认值为5。
 
         Raises:
             ValueError: 评估类型不在指定列表中。
@@ -199,7 +204,9 @@ class MaskRCNN(FasterRCNN):
             save_interval_epochs=save_interval_epochs,
             log_interval_steps=log_interval_steps,
             save_dir=save_dir,
-            use_vdl=use_vdl)
+            use_vdl=use_vdl,
+            early_stop=early_stop,
+            early_stop_patience=early_stop_patience)
 
     def evaluate(self,
                  eval_dataset,

paddlex/cv/models/slim/visualize.py

@@ -15,9 +15,6 @@
 import os.path as osp
 import tqdm
 import numpy as np
-import matplotlib
-matplotlib.use('Agg')
-import matplotlib.pyplot as plt
 from .prune import cal_model_size
 from paddleslim.prune import load_sensitivities
 
@@ -30,6 +27,10 @@ def visualize(model, sensitivities_file, save_dir='./'):
         model (paddlex.cv.models): paddlex中的模型。
         sensitivities_file (str): 敏感度文件存储路径。
     """
+    import matplotlib
+    matplotlib.use('Agg')
+    import matplotlib.pyplot as plt
+
     program = model.test_prog
     place = model.places[0]
     fig = plt.figure()

paddlex/cv/models/unet.py

@@ -117,7 +117,9 @@ class UNet(DeepLabv3p):
               lr_decay_power=0.9,
               use_vdl=False,
               sensitivities_file=None,
-              eval_metric_loss=0.05):
+              eval_metric_loss=0.05,
+              early_stop=False,
+              early_stop_patience=5):
         """训练。
 
         Args:
@@ -138,12 +140,17 @@ class UNet(DeepLabv3p):
             sensitivities_file (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，
                 则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
             eval_metric_loss (float): 可容忍的精度损失。默认为0.05。
+            early_stop (bool): 是否使用提前终止训练策略。默认值为False。
+            early_stop_patience (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内
+                连续下降或持平，则终止训练。默认值为5。
 
         Raises:
             ValueError: 模型从inference model进行加载。
         """
-        return super(UNet, self).train(
-            num_epochs, train_dataset, train_batch_size, eval_dataset,
-            save_interval_epochs, log_interval_steps, save_dir,
-            pretrain_weights, optimizer, learning_rate, lr_decay_power,
-            use_vdl, sensitivities_file, eval_metric_loss)
+        return super(
+            UNet,
+            self).train(num_epochs, train_dataset, train_batch_size,
+                        eval_dataset, save_interval_epochs, log_interval_steps,
+                        save_dir, pretrain_weights, optimizer, learning_rate,
+                        lr_decay_power, use_vdl, sensitivities_file,
+                        eval_metric_loss, early_stop, early_stop_patience)

paddlex/cv/models/utils/visualize.py

@@ -333,6 +333,7 @@ def draw_pr_curve(eval_details_file=None,
         return mean_s
 
     def cal_pr(coco_gt, coco_dt, iou_thresh, save_dir, style='bbox'):
+        import matplotlib.pyplot as plt
         from pycocotools.cocoeval import COCOeval
         coco_dt = loadRes(coco_gt, coco_dt)
         np.linspace = fixed_linspace

paddlex/cv/models/yolo_v3.py

@@ -162,7 +162,9 @@ class YOLOv3(BaseAPI):
               metric=None,
               use_vdl=False,
               sensitivities_file=None,
-              eval_metric_loss=0.05):
+              eval_metric_loss=0.05,
+              early_stop=False,
+              early_stop_patience=5):
         """训练。
 
         Args:
@@ -188,6 +190,9 @@ class YOLOv3(BaseAPI):
             sensitivities_file (str): 若指定为路径时，则加载路径下敏感度信息进行裁剪；若为字符串'DEFAULT'，
                 则自动下载在ImageNet图片数据上获得的敏感度信息进行裁剪；若为None，则不进行裁剪。默认为None。
             eval_metric_loss (float): 可容忍的精度损失。默认为0.05。
+            early_stop (bool): 是否使用提前终止训练策略。默认值为False。
+            early_stop_patience (int): 当使用提前终止训练策略时，如果验证集精度在`early_stop_patience`个epoch内
+                连续下降或持平，则终止训练。默认值为5。
 
         Raises:
             ValueError: 评估类型不在指定列表中。
@@ -238,7 +243,9 @@ class YOLOv3(BaseAPI):
             save_interval_epochs=save_interval_epochs,
             log_interval_steps=log_interval_steps,
             save_dir=save_dir,
-            use_vdl=use_vdl)
+            use_vdl=use_vdl,
+            early_stop=early_stop,
+            early_stop_patience=early_stop_patience)
 
     def evaluate(self,
                  eval_dataset,

paddlex/utils/utils.py

@@ -220,3 +220,39 @@ def load_pretrain_weights(exe, main_prog, weights_dir, fuse_bn=False):
             len(vars_to_load), weights_dir))
     if fuse_bn:
         fuse_bn_weights(exe, main_prog, weights_dir)
+
+
+class EarlyStop:
+    def __init__(self, patience, thresh):
+        self.patience = patience
+        self.counter = 0
+        self.score = None
+        self.max = 0
+        self.thresh = thresh
+        if patience < 1:
+            raise Exception("Argument patience should be a positive integer.")
+
+    def __call__(self, current_score):
+        if self.score is None:
+            self.score = current_score
+            return False
+        elif current_score > self.max:
+            self.counter = 0
+            self.score = current_score
+            self.max = current_score
+            return False
+        else:
+            if (abs(self.score - current_score) < self.thresh
+                    or current_score < self.score):
+                self.counter += 1
+                self.score = current_score
+                logging.debug(
+                    "EarlyStopping: %i / %i" % (self.counter, self.patience))
+                if self.counter >= self.patience:
+                    logging.info("EarlyStopping: Stop training")
+                    return True
+                return False
+            else:
+                self.counter = 0
+                self.score = current_score
+                return False

setup.py

@@ -29,7 +29,8 @@ setuptools.setup(
     packages=setuptools.find_packages(),
     setup_requires=['cython', 'numpy', 'sklearn'],
     install_requires=[
-        'pycocotools', 'pyyaml', 'colorama', 'tqdm', 'visualdl==1.3.0',
+        "pycocotools;platform_system!='Windows'", 
+        'pyyaml', 'colorama', 'tqdm', 'visualdl==1.3.0',
         'paddleslim==1.0.1', 'paddlehub>=1.6.2'
     ],
     classifiers=[