add openvino supported

deploy/openvino/CMakeLists.txt

@@ -8,7 +8,9 @@ SET(CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake" ${CMAKE_MODULE_PATH})
 SET(OPENVINO_DIR "" CACHE PATH "Location of libraries")
 SET(OPENCV_DIR "" CACHE PATH "Location of libraries")
 SET(GFLAGS_DIR "" CACHE PATH "Location of libraries")
+SET(GLOG_DIR "" CACHE PATH "Location of libraries")
 SET(NGRAPH_LIB "" CACHE PATH "Location of libraries")
+SET(ARCH "" CACHE PATH "Location of libraries")
 
 include(cmake/yaml-cpp.cmake)
 
@@ -27,6 +29,12 @@ macro(safe_set_static_flag)
     endforeach(flag_var)
 endmacro()
 
+if(NOT WIN32)
+    if (NOT DEFINED ARCH OR ${ARCH} STREQUAL "")
+        message(FATAL_ERROR "please set ARCH with -DARCH=x86 OR armv7")
+    endif()
+endif()
+
 if (NOT DEFINED OPENVINO_DIR OR ${OPENVINO_DIR} STREQUAL "")
     message(FATAL_ERROR "please set OPENVINO_DIR with -DOPENVINO_DIR=/path/influence_engine")
 endif()
@@ -39,19 +47,32 @@ if (NOT DEFINED GFLAGS_DIR OR ${GFLAGS_DIR} STREQUAL "")
     message(FATAL_ERROR "please set GFLAGS_DIR with -DGFLAGS_DIR=/path/gflags")
 endif()
 
+if (NOT DEFINED GLOG_DIR OR ${GLOG_DIR} STREQUAL "")
+    message(FATAL_ERROR "please set GLOG_DIR with -DLOG_DIR=/path/glog")
+endif()
+
 if (NOT DEFINED NGRAPH_LIB OR ${NGRAPH_LIB} STREQUAL "")
     message(FATAL_ERROR "please set NGRAPH_DIR with -DNGRAPH_DIR=/path/ngraph")
 endif()
 
 include_directories("${OPENVINO_DIR}")
-link_directories("${OPENVINO_DIR}/lib")
 include_directories("${OPENVINO_DIR}/include")
-link_directories("${OPENVINO_DIR}/external/tbb/lib")
 include_directories("${OPENVINO_DIR}/external/tbb/include/tbb")
+link_directories("${OPENVINO_DIR}/lib")
+link_directories("${OPENVINO_DIR}/external/tbb/lib")
+if(WIN32)
+    link_directories("${OPENVINO_DIR}/lib/intel64/Release")
+    link_directories("${OPENVINO_DIR}/bin/intel64/Release")
+endif()
+
+
 
 link_directories("${GFLAGS_DIR}/lib")
 include_directories("${GFLAGS_DIR}/include")
 
+link_directories("${GLOG_DIR}/lib")
+include_directories("${GLOG_DIR}/include")
+
 link_directories("${NGRAPH_LIB}")
 link_directories("${NGRAPH_LIB}/lib")
 
@@ -79,14 +100,29 @@ else()
     set(CMAKE_STATIC_LIBRARY_PREFIX "")
 endif()
 
-
-if(WITH_STATIC_LIB)
+if(WIN32)
+    set(DEPS ${OPENVINO_DIR}/lib/intel64/Release/inference_engine${CMAKE_STATIC_LIBRARY_SUFFIX})
+    set(DEPS ${DEPS} ${OPENVINO_DIR}/lib/intel64/Release/inference_engine_legacy${CMAKE_STATIC_LIBRARY_SUFFIX})
+else()
+    if (ARCH STREQUAL "armv7")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -march=armv7-a")
+        if(WITH_STATIC_LIB)
+            set(DEPS ${OPENVINO_DIR}/lib/armv7l/libinference_engine${CMAKE_STATIC_LIBRARY_SUFFIX})
+            set(DEPS ${DEPS} ${OPENVINO_DIR}/lib/armv7l/libinference_engine_legacy${CMAKE_STATIC_LIBRARY_SUFFIX})
+        else()
+            set(DEPS ${OPENVINO_DIR}/lib/armv7l/libinference_engine${CMAKE_SHARED_LIBRARY_SUFFIX})
+            set(DEPS ${DEPS} ${OPENVINO_DIR}/lib/armv7l/libinference_engine_legacy${CMAKE_SHARED_LIBRARY_SUFFIX})
+        endif()
+    else()
+        if(WITH_STATIC_LIB)
             set(DEPS ${OPENVINO_DIR}/lib/intel64/libinference_engine${CMAKE_STATIC_LIBRARY_SUFFIX})
             set(DEPS ${DEPS} ${OPENVINO_DIR}/lib/intel64/libinference_engine_legacy${CMAKE_STATIC_LIBRARY_SUFFIX})
-else()
+        else()
             set(DEPS ${OPENVINO_DIR}/lib/intel64/libinference_engine${CMAKE_SHARED_LIBRARY_SUFFIX})
             set(DEPS ${DEPS} ${OPENVINO_DIR}/lib/intel64/libinference_engine_legacy${CMAKE_SHARED_LIBRARY_SUFFIX})
-endif()
+        endif()
+    endif()
+endif(WIN32)
 
 if (NOT WIN32)
     set(DEPS ${DEPS}
@@ -94,7 +130,7 @@ if (NOT WIN32)
         )
 else()
     set(DEPS ${DEPS}
-        glog gflags_static libprotobuf zlibstatic xxhash libyaml-cppmt)
+        glog gflags_static  libyaml-cppmt)
     set(DEPS ${DEPS} libcmt shlwapi)
 endif(NOT WIN32)
 
@@ -105,7 +141,14 @@ if (NOT WIN32)
 endif()
 
 set(DEPS ${DEPS} ${OpenCV_LIBS})
-add_executable(classifier src/classifier.cpp src/transforms.cpp src/paddlex.cpp)
+add_executable(classifier demo/classifier.cpp src/transforms.cpp src/paddlex.cpp)
 ADD_DEPENDENCIES(classifier ext-yaml-cpp)
 target_link_libraries(classifier ${DEPS})
 
+add_executable(segmenter demo/segmenter.cpp src/transforms.cpp src/paddlex.cpp src/visualize.cpp)
+ADD_DEPENDENCIES(segmenter ext-yaml-cpp)
+target_link_libraries(segmenter ${DEPS})
+
+add_executable(detector demo/detector.cpp src/transforms.cpp src/paddlex.cpp src/visualize.cpp)
+ADD_DEPENDENCIES(detector ext-yaml-cpp)
+target_link_libraries(detector ${DEPS})

deploy/openvino/CMakeSettings.json

@@ -13,12 +13,32 @@
       "variables": [
         {
           "name": "OPENCV_DIR",
-                    "value": "C:/projects/opencv",
+          "value": "/path/to/opencv",
           "type": "PATH"
         },
         {
-                    "name": "OPENVINO_LIB",
-                    "value": "C:/projetcs/inference_engine",
+          "name": "OPENVINO_DIR",
+          "value": "C:/Program Files (x86)/IntelSWTools/openvino/deployment_tools/inference_engine",
+          "type": "PATH"
+        },
+        {
+          "name": "NGRAPH_LIB",
+          "value": "C:/Program Files (x86)/IntelSWTools/openvino/deployment_tools/ngraph/lib",
+          "type": "PATH"
+        },
+        {
+          "name": "GFLAGS_DIR",
+          "value": "/path/to/gflags",
+          "type": "PATH"
+        },
+        {
+          "name": "WITH_STATIC_LIB",
+          "value": "True",
+          "type": "BOOL"
+        },
+        {
+          "name": "GLOG_DIR",
+          "value": "/path/to/glog",
           "type": "PATH"
         }
       ]

deploy/openvino/cmake/yaml-cpp.cmake

-find_package(Git REQUIRED)
 
 include(ExternalProject)
 

deploy/openvino/src/classifier.cpp → deploy/openvino/demo/classifier.cpp

@@ -22,7 +22,7 @@
 #include "include/paddlex/paddlex.h"
 
 DEFINE_string(model_dir, "", "Path of inference model");
-DEFINE_string(cfg_dir, "", "Path of inference model");
+DEFINE_string(cfg_dir, "", "Path of PaddelX model yml file");
 DEFINE_string(device, "CPU", "Device name");
 DEFINE_string(image, "", "Path of test image file");
 DEFINE_string(image_list, "", "Path of test image list file");
@@ -62,7 +62,8 @@ int main(int argc, char** argv) {
       model.predict(im, &result);
       std::cout << "Predict label: " << result.category
                 << ", label_id:" << result.category_id
-                << ", score: " << result.score << std::endl;
+                << ", score: " << result.score 
+                << ", num_img: " << model.count_num_ << std::endl;
     }
   } else {
     PaddleX::ClsResult result;

deploy/openvino/demo/detector.cpp

+#include <glog/logging.h>
+#include <omp.h>
+
+#include <algorithm>
+#include <chrono>  // NOLINT
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+#include <utility>
+
+#include "include/paddlex/paddlex.h"
+#include "include/paddlex/visualize.h"
+
+using namespace std::chrono;  // NOLINT
+
+DEFINE_string(model_dir, "", "Path of openvino model xml file");
+DEFINE_string(cfg_dir, "", "Path of PaddleX model yaml file");
+DEFINE_string(image, "", "Path of test image file");
+DEFINE_string(image_list, "", "Path of test image list file");
+DEFINE_string(device, "CPU", "Device name");
+DEFINE_string(save_dir, "", "Path to save visualized image");
+DEFINE_int32(batch_size, 1, "Batch size of infering");
+DEFINE_double(threshold,
+              0.5,
+              "The minimum scores of target boxes which are shown");
+
+int main(int argc, char** argv) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  if (FLAGS_model_dir == "") {
+    std::cerr << "--model_dir need to be defined" << std::endl;
+    return -1;
+  }
+  if (FLAGS_cfg_dir == "") {
+    std::cerr << "--cfg_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;
+  }
+
+  //
+  PaddleX::Model model; 
+  model.Init(FLAGS_model_dir, FLAGS_cfg_dir, FLAGS_device);
+
+  int imgs = 1;
+  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::DetResult result;
+      cv::Mat im = cv::imread(image_path, 1);
+      model.predict(im, &result);
+      if(FLAGS_save_dir != ""){
+        cv::Mat vis_img =
+          PaddleX::Visualize(im, result, model.labels, colormap, FLAGS_threshold);  
+        std::string save_path =
+          PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_image);      
+        cv::imwrite(save_path, vis_img);
+        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 << "image file: " << FLAGS_image << std::endl;
+      std::cout << ", predict label: " << result.boxes[i].category
+                << ", label_id:" << result.boxes[i].category_id
+                << ", score: " << result.boxes[i].score
+                << ", box(xmin, ymin, w, h):(" << result.boxes[i].coordinate[0]
+                << ", " << result.boxes[i].coordinate[1] << ", "
+                << result.boxes[i].coordinate[2] << ", "
+                << result.boxes[i].coordinate[3] << ")" << std::endl;
+    }
+    if(FLAGS_save_dir != ""){
+    // 可视化
+      cv::Mat vis_img =
+          PaddleX::Visualize(im, result, model.labels, colormap, FLAGS_threshold);
+      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/openvino/demo/segmenter.cpp

+//   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 <algorithm>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+#include <utility>
+#include "include/paddlex/paddlex.h"
+#include "include/paddlex/visualize.h"
+
+
+DEFINE_string(model_dir, "", "Path of openvino model xml file");
+DEFINE_string(cfg_dir, "", "Path of PaddleX model yaml file");
+DEFINE_string(image, "", "Path of test image file");
+DEFINE_string(image_list, "", "Path of test image list file");
+DEFINE_string(device, "CPU", "Device name");
+DEFINE_string(save_dir, "", "Path to save visualized image");
+DEFINE_int32(batch_size, 1, "Batch size of infering");
+
+
+int main(int argc, char** argv) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  if (FLAGS_model_dir == "") {
+    std::cerr << "--model_dir need to be defined" << std::endl;
+    return -1;
+  }
+  if (FLAGS_cfg_dir == "") {
+    std::cerr << "--cfg_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;
+  }
+
+  //
+  std::cout << "init start" << std::endl;
+  PaddleX::Model model; 
+  model.Init(FLAGS_model_dir, FLAGS_cfg_dir, FLAGS_device);
+  std::cout << "init done" << std::endl;
+  int imgs = 1;
+  auto colormap = PaddleX::GenerateColorMap(model.labels.size());
+  
+  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;
+    model.total_time_ = 0.0f;
+    model.count_num_ = 0;
+    while (getline(inf, image_path)) {
+      PaddleX::SegResult result;
+      cv::Mat im = cv::imread(image_path, 1);
+      model.predict(im, &result);
+      if(FLAGS_save_dir != ""){
+      cv::Mat vis_img = PaddleX::Visualize(im, result, model.labels, colormap);
+        std::string save_path =
+          PaddleX::generate_save_path(FLAGS_save_dir, image_path);
+        cv::imwrite(save_path, vis_img);
+        std::cout << "Visualized output saved as " << save_path << std::endl;
+      }
+    }
+  }else{
+    PaddleX::SegResult result;
+    cv::Mat im = cv::imread(FLAGS_image, 1);
+    std::cout << "predict start" << std::endl;
+    model.predict(im, &result);
+    std::cout << "predict done" << std::endl; 
+    if(FLAGS_save_dir != ""){
+      cv::Mat vis_img = PaddleX::Visualize(im, result, model.labels, colormap);
+      std::string save_path =
+          PaddleX::generate_save_path(FLAGS_save_dir, FLAGS_image);
+      cv::imwrite(save_path, vis_img);
+      std::cout << "Visualized` output saved as " << save_path << std::endl;
+    }
+    result.clear();
+  }
+  return 0;
+}

deploy/openvino/include/paddlex/paddlex.h

@@ -17,6 +17,7 @@
 #include <functional>
 #include <iostream>
 #include <numeric>
+#include <chrono> 
 
 #include "yaml-cpp/yaml.h"
 
@@ -48,15 +49,20 @@ class Model {
 
   bool load_config(const std::string& model_dir);
 
-  bool preprocess(cv::Mat* input_im);
+  bool preprocess(cv::Mat* input_im, ImageBlob* inputs);
 
   bool predict(const cv::Mat& im, ClsResult* result);
   
+  bool predict(const cv::Mat& im, DetResult* result);  
+
+  bool predict(const cv::Mat& im, SegResult* result);
+
+
   std::string type;
   std::string name;
-  std::vector<std::string> labels;
+  std::map<int, std::string> labels;
   Transforms transforms_;
-  Blob::Ptr inputs_;
+  ImageBlob inputs_;
   Blob::Ptr output_;
   CNNNetwork network_;
   ExecutableNetwork executable_network_;

deploy/openvino/include/paddlex/results.h

@@ -61,7 +61,7 @@ class DetResult : public BaseResult {
 
 class SegResult : public BaseResult {
  public:
-  Mask<int64_t> label_map;
+  Mask<int> label_map;
   Mask<float> score_map;
   void clear() {
     label_map.clear();

deploy/openvino/include/paddlex/transforms.h

@@ -31,11 +31,40 @@ using namespace InferenceEngine;
 
 namespace PaddleX {
 
+/*
+ * @brief
+ * This class represents object for storing all preprocessed data
+ * */
+class ImageBlob {
+ public:
+  // Original image height and width
+  //std::vector<int> ori_im_size_ = std::vector<int>(2);
+  Blob::Ptr ori_im_size_;
+
+  // Newest image height and width after process
+  std::vector<int> new_im_size_ = std::vector<int>(2);
+  // Image height and width before resize
+  std::vector<std::vector<int>> im_size_before_resize_;
+  // Reshape order
+  std::vector<std::string> reshape_order_;
+  // Resize scale
+  float scale = 1.0;
+  // Buffer for image data after preprocessing
+  Blob::Ptr blob;
+
+  void clear() {
+    im_size_before_resize_.clear();
+    reshape_order_.clear();
+  }
+};
+
+
+
 // Abstraction of preprocessing opration class
 class Transform {
  public:
   virtual void Init(const YAML::Node& item) = 0;
-  virtual bool Run(cv::Mat* im) = 0;
+  virtual bool Run(cv::Mat* im, ImageBlob* data) = 0;
 };
 
 class Normalize : public Transform {
@@ -45,7 +74,7 @@ class Normalize : public Transform {
     std_ = item["std"].as<std::vector<float>>();
   }
 
-  virtual bool Run(cv::Mat* im);
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
 
  private:
   std::vector<float> mean_;
@@ -62,7 +91,7 @@ class ResizeByShort : public Transform {
       max_size_ = -1;
     }
   };
-  virtual bool Run(cv::Mat* im);
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
 
  private:
   float GenerateScale(const cv::Mat& im);
@@ -70,6 +99,55 @@ class ResizeByShort : public Transform {
   int max_size_;
 };
 
+/*
+ * @brief
+ * This class execute resize by long operation on image matrix. At first, it resizes
+ * the long side of image matrix to specified length. Accordingly, the short side
+ * will be resized in the same proportion.
+ * */
+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_;
+};
+
+/*
+ * @brief
+ * This class execute resize operation on image matrix. It resizes width and height
+ * to specified length.
+ * */
+class Resize : public Transform {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    if (item["interp"].IsDefined()) {
+      interp_ = item["interp"].as<std::string>();
+    }
+    if (item["target_size"].IsScalar()) {
+      height_ = item["target_size"].as<int>();
+      width_ = item["target_size"].as<int>();
+    } 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:
@@ -83,22 +161,66 @@ class CenterCrop : public Transform {
       height_ = crop_size[1];
     }
   }
-  virtual bool Run(cv::Mat* im);
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
 
  private:
   int height_;
   int width_;
 };
 
+
+/*
+ * @brief
+ * This class execute padding operation on image matrix. It makes border on edge
+ * of image matrix.
+ * */
+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);
+      }
+    }
+    if (item["target_size"].IsDefined()) {
+      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()) {
+      im_value_ = item["im_padding_value"].as<std::vector<float>>();
+    }
+    else {
+      im_value_ = {0, 0, 0};
+    }
+  }
+  
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  int coarsest_stride_ = -1;
+  int width_ = 0;
+  int height_ = 0;
+  std::vector<float> im_value_;
+};
+
 class Transforms {
  public:
-  void Init(const YAML::Node& node, bool to_rgb = true);
+  void Init(const YAML::Node& node, std::string type, bool to_rgb = true);
   std::shared_ptr<Transform> CreateTransform(const std::string& name);
-  bool Run(cv::Mat* im, Blob::Ptr blob);
+  bool Run(cv::Mat* im, ImageBlob* data);
 
  private:
   std::vector<std::shared_ptr<Transform>> transforms_;
   bool to_rgb_ = true;
+  std::string type_;
 };
 
 }  // namespace PaddleX

deploy/openvino/include/paddlex/visualize.h

+//   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 {
+
+/*
+ * @brief
+ * Generate visualization colormap for each class
+ *
+ * @param number of class
+ * @return color map, the size of vector is 3 * num_class
+ * */
+std::vector<int> GenerateColorMap(int num_class);
+
+
+/*
+ * @brief
+ * Visualize the detection result
+ *
+ * @param img: initial image matrix
+ * @param results: the detection result
+ * @param labels: label map
+ * @param colormap: visualization color map
+ * @return visualized image matrix
+ * */
+cv::Mat Visualize(const cv::Mat& img,
+                     const DetResult& results,
+                     const std::map<int, std::string>& labels,
+                     const std::vector<int>& colormap,
+                     float threshold = 0.5);
+
+/*
+ * @brief
+ * Visualize the segmentation result
+ *
+ * @param img: initial image matrix
+ * @param results: the detection result
+ * @param labels: label map
+ * @param colormap: visualization color map
+ * @return visualized image matrix
+ * */
+cv::Mat Visualize(const cv::Mat& img,
+                     const SegResult& result,
+                     const std::map<int, std::string>& labels,
+                     const std::vector<int>& colormap);
+
+/*
+ * @brief
+ * generate save path for visualized image matrix
+ *
+ * @param save_dir: directory for saving visualized image matrix
+ * @param file_path: sourcen image file path
+ * @return path of saving visualized result
+ * */
+std::string generate_save_path(const std::string& save_dir,
+                               const std::string& file_path);
+}  // namespace PaddleX

deploy/openvino/python/convertor.py

+#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from six import text_type as _text_type
+import argparse
+import sys
+from utils import logging 
+import paddlex as pdx
+
+def arg_parser():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_dir",
+        "-m",
+        type=_text_type,
+        default=None,
+        help="define model directory path")
+    parser.add_argument(
+        "--save_dir",
+        "-s",
+        type=_text_type,
+        default=None,
+        help="path to save inference model")
+    parser.add_argument(
+        "--fixed_input_shape",
+        "-fs",
+        default=None,
+        help="export openvino model with  input shape:[h,w]")
+    return parser
+
+
+    
+
+
+def export_openvino_model(model, args):
+    if model.model_type == "detector" or model.__class__.__name__ == "FastSCNN":
+        logging.error(
+            "Only image classifier models and semantic segmentation models(except FastSCNN) are supported to export to openvino")
+    try:
+        import x2paddle
+        if x2paddle.__version__ < '0.7.4':
+            logging.error("You need to upgrade x2paddle >= 0.7.4")
+    except:
+        logging.error(
+            "You need to install x2paddle first, pip install x2paddle>=0.7.4")
+        
+    from x2paddle.op_mapper.paddle_op_mapper import PaddleOpMapper
+    mapper = PaddleOpMapper()
+    mapper.convert(model.test_prog, args.save_dir)
+
+    import mo.main as mo
+    from mo.utils.cli_parser import get_onnx_cli_parser
+    onnx_parser = get_onnx_cli_parser()
+    onnx_parser.add_argument("--model_dir",type=_text_type)
+    onnx_parser.add_argument("--save_dir",type=_text_type)
+    onnx_parser.add_argument("--fixed_input_shape")
+    onnx_input = os.path.join(args.save_dir, 'x2paddle_model.onnx')
+    onnx_parser.set_defaults(input_model=onnx_input)
+    onnx_parser.set_defaults(output_dir=args.save_dir)
+    shape = '[1,3,'
+    shape =  shape + args.fixed_input_shape[1:]
+    onnx_parser.set_defaults(input_shape = shape)
+    mo.main(onnx_parser,'onnx')
+
+
+def main():
+    parser = arg_parser()
+    args = parser.parse_args()
+    assert args.model_dir is not None, "--model_dir should be defined while exporting openvino model"
+    assert args.save_dir is not None, "--save_dir should be defined to create openvino model"
+    model = pdx.load_model(args.model_dir)
+    if model.status == "Normal" or model.status == "Prune":
+        logging.error(
+            "Only support inference model, try to export model first as below,",
+            exit=False)
+    export_openvino_model(model, args)
+
+if  __name__ == "__main__":
+    main()
+
+

deploy/openvino/python/demo.py

+#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import os
+import argparse
+import deploy
+
+
+def arg_parser():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_dir",
+        "-m",
+        type=str,
+        default=None,
+        help="path to openvino model .xml file")
+    parser.add_argument(
+        "--device",
+        "-d",
+        type=str,
+        default='CPU',
+        help="Specify the target device to infer on:[CPU, GPU, FPGA, HDDL, MYRIAD,HETERO]"
+             "Default value is CPU")
+    parser.add_argument(
+        "--img",
+        "-i",
+        type=str,
+        default=None,
+        help="path to an image files")
+
+    parser.add_argument(
+        "--img_list",
+        "-l",
+        type=str,
+        default=None,
+        help="Path to a imglist")
+
+
+    parser.add_argument(
+        "--cfg_dir",
+        "-c",
+        type=str,
+        default=None,
+        help="Path to PaddelX model yml file")
+
+
+    return parser
+
+
+def main():
+    parser = arg_parser()
+    args = parser.parse_args()
+    model_xml = args.model_dir
+    model_yaml = args.cfg_dir
+
+    #model init
+    if("CPU" not in args.device):
+        predictor = deploy.Predictor(model_xml,model_yaml,args.device)
+    else:
+        predictor = deploy.Predictor(model_xml,model_yaml)
+    
+    #predict
+    if(args.img_list != None):
+        f = open(args.img_list)
+        lines = f.readlines()
+        for im_path in lines:
+            print(im_path)
+            predictor.predict(im_path.strip('\n'))
+        f.close()
+    else:
+        im_path = args.img
+        predictor.predict(im_path)
+
+if __name__ == "__main__":
+    main()

deploy/openvino/python/deploy.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import sys
+import os
+import os.path as osp
+import time
+import cv2
+import numpy as np
+import yaml
+from six import text_type as _text_type
+from openvino.inference_engine import IECore
+from utils import logging 
+
+
+
+
+class Predictor:
+    def __init__(self,
+                 model_xml,
+                 model_yaml,
+                 device="CPU"):
+        self.device = device
+        if not osp.exists(model_xml):
+            logging.error("model xml file is not exists in {}".format(model_xml))
+        self.model_xml = model_xml
+        self.model_bin = osp.splitext(model_xml)[0] + ".bin"
+        if not osp.exists(model_yaml):
+            logging,error("model yaml file is not exists in {}".format(model_yaml))
+        with open(model_yaml) as f:
+            self.info = yaml.load(f.read(), Loader=yaml.Loader)
+        self.model_type = self.info['_Attributes']['model_type']
+        self.model_name = self.info['Model']
+        self.num_classes = self.info['_Attributes']['num_classes']
+        self.labels = self.info['_Attributes']['labels']
+        if self.info['Model'] == 'MaskRCNN':
+            if self.info['_init_params']['with_fpn']:
+                self.mask_head_resolution = 28
+            else:
+                self.mask_head_resolution = 14
+        transforms_mode = self.info.get('TransformsMode', 'RGB')
+        if transforms_mode == 'RGB':
+            to_rgb = True
+        else:
+            to_rgb = False
+        self.transforms = self.build_transforms(self.info['Transforms'], to_rgb)
+        self.predictor, self.net = self.create_predictor()
+        self.total_time = 0
+        self.count_num = 0
+
+
+
+
+    def create_predictor(self):
+
+        #initialization for specified device
+        logging.info("Creating Inference Engine")
+        ie = IECore()
+        logging.info("Loading network files:\n\t{}\n\t{}".format(self.model_xml, self.model_bin))
+        net = ie.read_network(model=self.model_xml, weights=self.model_bin)
+        net.batch_size = 1
+        network_config = {}
+        if self.device == "MYRIAD":
+            network_config = {'VPU_HW_STAGES_OPTIMIZATION':'NO'}
+        exec_net = ie.load_network(network=net, device_name=self.device, network_config)
+        return exec_net, net
+
+
+    def build_transforms(self, transforms_info, to_rgb=True):
+        if self.model_type == "classifier":
+            import transforms.cls_transforms as transforms
+        elif self.model_type == "detector":
+            import transforms.det_transforms as transforms
+        elif self.model_type == "segmenter":
+            import transforms.seg_transforms as transforms
+        op_list = list()
+        for op_info in transforms_info:
+            op_name = list(op_info.keys())[0]
+            op_attr = op_info[op_name]
+            if not hasattr(transforms, op_name):
+                raise Exception(
+                    "There's no operator named '{}' in transforms of {}".
+                    format(op_name, self.model_type))
+            op_list.append(getattr(transforms, op_name)(**op_attr))
+        eval_transforms = transforms.Compose(op_list)
+        if hasattr(eval_transforms, 'to_rgb'):
+            eval_transforms.to_rgb = to_rgb
+        self.arrange_transforms(eval_transforms)
+        return eval_transforms    
+ 
+    def arrange_transforms(self, eval_transforms):
+        if self.model_type == 'classifier':
+            import transforms.cls_transforms as transforms
+            arrange_transform = transforms.ArrangeClassifier
+        elif self.model_type == 'segmenter':
+            import transforms.seg_transforms as transforms
+            arrange_transform = transforms.ArrangeSegmenter
+        elif self.model_type == 'detector':
+            import transforms.det_transforms as transforms
+            arrange_name = 'Arrange{}'.format(self.model_name)
+            arrange_transform = getattr(transforms, arrange_name)
+        else:
+            raise Exception("Unrecognized model type: {}".format(
+                self.model_type))
+        if type(eval_transforms.transforms[-1]).__name__.startswith('Arrange'):
+            eval_transforms.transforms[-1] = arrange_transform(mode='test')
+        else:
+            eval_transforms.transforms.append(arrange_transform(mode='test'))
+
+
+    def raw_predict(self, preprocessed_input):
+        self.count_num += 1
+        feed_dict = {}
+        if self.model_name == "YOLOv3":
+            inputs = self.net.inputs 
+            for name in inputs:
+                if(len(inputs[name].shape) == 2):
+                    feed_dict[name] = preprocessed_input['im_size']
+                elif(len(inputs[name].shape) == 4):
+                    feed_dict[name] = preprocessed_input['image']
+                else:
+                    pass
+        else:
+            input_blob = next(iter(self.net.inputs))
+            feed_dict[input_blob] = preprocessed_input['image']
+        #Start sync inference
+        logging.info("Starting inference in synchronous mode")
+        res = self.predictor.infer(inputs=feed_dict)
+ 
+        #Processing output blob
+        logging.info("Processing output blob")
+        return res
+        
+
+    def preprocess(self, image):
+        res = dict() 
+        if self.model_type == "classifier":
+            im, = self.transforms(image)
+            im = np.expand_dims(im, axis=0).copy()
+            res['image'] = im
+        elif self.model_type == "detector":
+            if self.model_name == "YOLOv3":
+                im, im_shape = self.transforms(image)
+                im = np.expand_dims(im, axis=0).copy()
+                im_shape = np.expand_dims(im_shape, axis=0).copy()
+                res['image'] = im
+                res['im_size'] = im_shape
+            if self.model_name.count('RCNN') > 0:
+                im, im_resize_info, im_shape = self.transforms(image)
+                im = np.expand_dims(im, axis=0).copy()
+                im_resize_info = np.expand_dims(im_resize_info, axis=0).copy()
+                im_shape = np.expand_dims(im_shape, axis=0).copy()
+                res['image'] = im
+                res['im_info'] = im_resize_info
+                res['im_shape'] = im_shape
+        elif self.model_type == "segmenter":
+            im, im_info = self.transforms(image)
+            im = np.expand_dims(im, axis=0).copy()
+            #np.savetxt('./input_data.txt',im.flatten())
+            res['image'] = im
+            res['im_info'] = im_info
+        return res
+    
+
+    def classifier_postprocess(self, preds, topk=1):
+        """ 对分类模型的预测结果做后处理
+        """
+        true_topk = min(self.num_classes, topk)
+        output_name = next(iter(self.net.outputs))
+        pred_label = np.argsort(-preds[output_name][0])[:true_topk]
+        result = [{
+            'category_id': l,
+            'category': self.labels[l],
+            'score': preds[output_name][0][l],
+        } for l in pred_label]
+        print(result)
+        return result 
+
+    def segmenter_postprocess(self, preds, preprocessed_inputs):
+        """ 对语义分割结果做后处理
+        """
+        it = iter(self.net.outputs)
+        next(it)
+        score_name = next(it)
+        #np.savetxt('./score_map.txt',preds[score_name].flatten())
+        score_map = np.squeeze(preds[score_name])
+        score_map = np.transpose(score_map, (1, 2, 0))
+        label_name = next(it)
+        #np.savetxt('./label_map.txt',preds[label_name].flatten())
+        label_map = np.squeeze(preds[label_name]).astype('uint8')
+        im_info = preprocessed_inputs['im_info']
+        for info in im_info[::-1]:
+            if info[0] == 'resize':
+                w, h = info[1][1], info[1][0]
+                label_map = cv2.resize(label_map, (w, h), cv2.INTER_NEAREST)
+                score_map = cv2.resize(score_map, (w, h), cv2.INTER_LINEAR)
+            elif info[0] == 'padding':
+                w, h = info[1][1], info[1][0]
+                label_map = label_map[0:h, 0:w]
+                score_map = score_map[0:h, 0:w, :]
+            else:
+                raise Exception("Unexpected info '{}' in im_info".format(info[
+                    0]))
+        return {'label_map': label_map, 'score_map': score_map}
+        
+    def detector_postprocess(self, preds, preprocessed_inputs):
+        """对图像检测结果做后处理
+        """
+        output_name = next(iter(self.net.outputs))
+        outputs = preds[output_name][0]
+        result = []
+        for out in outputs:
+            if(out[0] > 0):
+                result.append(out.tolist())
+            else:
+                pass
+        print(result)
+        return result
+
+
+    def predict(self, image, topk=1, threshold=0.5):
+        preprocessed_input = self.preprocess(image)
+        model_pred = self.raw_predict(preprocessed_input)
+        if self.model_type == "classifier":
+            results = self.classifier_postprocess(model_pred, topk)
+        elif self.model_type == "detector":
+            results = self.detector_postprocess(model_pred, preprocessed_input)
+        elif self.model_type == "segmenter":
+            results = self.segmenter_postprocess(model_pred,
+                                                 preprocessed_input) 
+

deploy/openvino/python/transforms/__init__.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from . import cls_transforms
+from . import det_transforms
+from . import seg_transforms

deploy/openvino/python/transforms/box_utils.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import random
+import math
+import cv2
+import scipy
+
+
+def bbox_area(src_bbox):
+    if src_bbox[2] < src_bbox[0] or src_bbox[3] < src_bbox[1]:
+        return 0.
+    else:
+        width = src_bbox[2] - src_bbox[0]
+        height = src_bbox[3] - src_bbox[1]
+        return width * height
+
+
+def jaccard_overlap(sample_bbox, object_bbox):
+    if sample_bbox[0] >= object_bbox[2] or \
+        sample_bbox[2] <= object_bbox[0] or \
+        sample_bbox[1] >= object_bbox[3] or \
+        sample_bbox[3] <= object_bbox[1]:
+        return 0
+    intersect_xmin = max(sample_bbox[0], object_bbox[0])
+    intersect_ymin = max(sample_bbox[1], object_bbox[1])
+    intersect_xmax = min(sample_bbox[2], object_bbox[2])
+    intersect_ymax = min(sample_bbox[3], object_bbox[3])
+    intersect_size = (intersect_xmax - intersect_xmin) * (
+        intersect_ymax - intersect_ymin)
+    sample_bbox_size = bbox_area(sample_bbox)
+    object_bbox_size = bbox_area(object_bbox)
+    overlap = intersect_size / (
+        sample_bbox_size + object_bbox_size - intersect_size)
+    return overlap
+
+
+def iou_matrix(a, b):
+    tl_i = np.maximum(a[:, np.newaxis, :2], b[:, :2])
+    br_i = np.minimum(a[:, np.newaxis, 2:], b[:, 2:])
+
+    area_i = np.prod(br_i - tl_i, axis=2) * (tl_i < br_i).all(axis=2)
+    area_a = np.prod(a[:, 2:] - a[:, :2], axis=1)
+    area_b = np.prod(b[:, 2:] - b[:, :2], axis=1)
+    area_o = (area_a[:, np.newaxis] + area_b - area_i)
+    return area_i / (area_o + 1e-10)
+
+
+def crop_box_with_center_constraint(box, crop):
+    cropped_box = box.copy()
+
+    cropped_box[:, :2] = np.maximum(box[:, :2], crop[:2])
+    cropped_box[:, 2:] = np.minimum(box[:, 2:], crop[2:])
+    cropped_box[:, :2] -= crop[:2]
+    cropped_box[:, 2:] -= crop[:2]
+
+    centers = (box[:, :2] + box[:, 2:]) / 2
+    valid = np.logical_and(crop[:2] <= centers, centers < crop[2:]).all(axis=1)
+    valid = np.logical_and(
+        valid, (cropped_box[:, :2] < cropped_box[:, 2:]).all(axis=1))
+
+    return cropped_box, np.where(valid)[0]
+
+
+def is_poly(segm):
+    if not isinstance(segm, (list, dict)):
+        raise Exception("Invalid segm type: {}".format(type(segm)))
+    return isinstance(segm, list)
+
+
+def crop_image(img, crop):
+    x1, y1, x2, y2 = crop
+    return img[y1:y2, x1:x2, :]
+
+
+def crop_segms(segms, valid_ids, crop, height, width):
+    def _crop_poly(segm, crop):
+        xmin, ymin, xmax, ymax = crop
+        crop_coord = [xmin, ymin, xmin, ymax, xmax, ymax, xmax, ymin]
+        crop_p = np.array(crop_coord).reshape(4, 2)
+        crop_p = Polygon(crop_p)
+
+        crop_segm = list()
+        for poly in segm:
+            poly = np.array(poly).reshape(len(poly) // 2, 2)
+            polygon = Polygon(poly)
+            if not polygon.is_valid:
+                exterior = polygon.exterior
+                multi_lines = exterior.intersection(exterior)
+                polygons = shapely.ops.polygonize(multi_lines)
+                polygon = MultiPolygon(polygons)
+            multi_polygon = list()
+            if isinstance(polygon, MultiPolygon):
+                multi_polygon = copy.deepcopy(polygon)
+            else:
+                multi_polygon.append(copy.deepcopy(polygon))
+            for per_polygon in multi_polygon:
+                inter = per_polygon.intersection(crop_p)
+                if not inter:
+                    continue
+                if isinstance(inter, (MultiPolygon, GeometryCollection)):
+                    for part in inter:
+                        if not isinstance(part, Polygon):
+                            continue
+                        part = np.squeeze(
+                            np.array(part.exterior.coords[:-1]).reshape(1, -1))
+                        part[0::2] -= xmin
+                        part[1::2] -= ymin
+                        crop_segm.append(part.tolist())
+                elif isinstance(inter, Polygon):
+                    crop_poly = np.squeeze(
+                        np.array(inter.exterior.coords[:-1]).reshape(1, -1))
+                    crop_poly[0::2] -= xmin
+                    crop_poly[1::2] -= ymin
+                    crop_segm.append(crop_poly.tolist())
+                else:
+                    continue
+        return crop_segm
+
+    def _crop_rle(rle, crop, height, width):
+        if 'counts' in rle and type(rle['counts']) == list:
+            rle = mask_util.frPyObjects(rle, height, width)
+        mask = mask_util.decode(rle)
+        mask = mask[crop[1]:crop[3], crop[0]:crop[2]]
+        rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
+        return rle
+
+    crop_segms = []
+    for id in valid_ids:
+        segm = segms[id]
+        if is_poly(segm):
+            import copy
+            import shapely.ops
+            import logging
+            from shapely.geometry import Polygon, MultiPolygon, GeometryCollection
+            logging.getLogger("shapely").setLevel(logging.WARNING)
+            # Polygon format
+            crop_segms.append(_crop_poly(segm, crop))
+        else:
+            # RLE format
+            import pycocotools.mask as mask_util
+            crop_segms.append(_crop_rle(segm, crop, height, width))
+    return crop_segms
+
+
+def expand_segms(segms, x, y, height, width, ratio):
+    def _expand_poly(poly, x, y):
+        expanded_poly = np.array(poly)
+        expanded_poly[0::2] += x
+        expanded_poly[1::2] += y
+        return expanded_poly.tolist()
+
+    def _expand_rle(rle, x, y, height, width, ratio):
+        if 'counts' in rle and type(rle['counts']) == list:
+            rle = mask_util.frPyObjects(rle, height, width)
+        mask = mask_util.decode(rle)
+        expanded_mask = np.full((int(height * ratio), int(width * ratio)),
+                                0).astype(mask.dtype)
+        expanded_mask[y:y + height, x:x + width] = mask
+        rle = mask_util.encode(
+            np.array(expanded_mask, order='F', dtype=np.uint8))
+        return rle
+
+    expanded_segms = []
+    for segm in segms:
+        if is_poly(segm):
+            # Polygon format
+            expanded_segms.append([_expand_poly(poly, x, y) for poly in segm])
+        else:
+            # RLE format
+            import pycocotools.mask as mask_util
+            expanded_segms.append(
+                _expand_rle(segm, x, y, height, width, ratio))
+    return expanded_segms
+
+
+def box_horizontal_flip(bboxes, width):
+    oldx1 = bboxes[:, 0].copy()
+    oldx2 = bboxes[:, 2].copy()
+    bboxes[:, 0] = width - oldx2 - 1
+    bboxes[:, 2] = width - oldx1 - 1
+    if bboxes.shape[0] != 0 and (bboxes[:, 2] < bboxes[:, 0]).all():
+        raise ValueError(
+            "RandomHorizontalFlip: invalid box, x2 should be greater than x1")
+    return bboxes
+
+
+def segms_horizontal_flip(segms, height, width):
+    def _flip_poly(poly, width):
+        flipped_poly = np.array(poly)
+        flipped_poly[0::2] = width - np.array(poly[0::2]) - 1
+        return flipped_poly.tolist()
+
+    def _flip_rle(rle, height, width):
+        if 'counts' in rle and type(rle['counts']) == list:
+            rle = mask_util.frPyObjects([rle], height, width)
+        mask = mask_util.decode(rle)
+        mask = mask[:, ::-1]
+        rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
+        return rle
+
+    flipped_segms = []
+    for segm in segms:
+        if is_poly(segm):
+            # Polygon format
+            flipped_segms.append([_flip_poly(poly, width) for poly in segm])
+        else:
+            # RLE format
+            import pycocotools.mask as mask_util
+            flipped_segms.append(_flip_rle(segm, height, width))
+    return flipped_segms

deploy/openvino/python/transforms/cls_transforms.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .ops import *
+from .imgaug_support import execute_imgaug
+import random
+import os.path as osp
+import numpy as np
+from PIL import Image, ImageEnhance
+import utils.logging as logging
+
+
+class ClsTransform:
+    """分类Transform的基类
+    """
+
+    def __init__(self):
+        pass
+
+
+class Compose(ClsTransform):
+    """根据数据预处理/增强算子对输入数据进行操作。
+       所有操作的输入图像流形状均是[H, W, C]，其中H为图像高，W为图像宽，C为图像通道数。
+
+    Args:
+        transforms (list): 数据预处理/增强算子。
+
+    Raises:
+        TypeError: 形参数据类型不满足需求。
+        ValueError: 数据长度不匹配。
+    """
+
+    def __init__(self, transforms):
+        if not isinstance(transforms, list):
+            raise TypeError('The transforms must be a list!')
+        if len(transforms) < 1:
+            raise ValueError('The length of transforms ' + \
+                            'must be equal or larger than 1!')
+        self.transforms = transforms
+
+        # 检查transforms里面的操作，目前支持PaddleX定义的或者是imgaug操作
+        for op in self.transforms:
+            if not isinstance(op, ClsTransform):
+                import imgaug.augmenters as iaa
+                if not isinstance(op, iaa.Augmenter):
+                    raise Exception(
+                        "Elements in transforms should be defined in 'paddlex.cls.transforms' or class of imgaug.augmenters.Augmenter, see docs here: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/"
+                    )
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (str/np.ndarray): 图像路径/图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+        Returns:
+            tuple: 根据网络所需字段所组成的tuple；
+                字段由transforms中的最后一个数据预处理操作决定。
+        """
+        if isinstance(im, np.ndarray):
+            if len(im.shape) != 3:
+                raise Exception(
+                    "im should be 3-dimension, but now is {}-dimensions".
+                    format(len(im.shape)))
+        else:
+            try:
+                im = cv2.imread(im).astype('float32')
+            except:
+                raise TypeError('Can\'t read The image file {}!'.format(im))
+        im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
+        for op in self.transforms:
+            if isinstance(op, ClsTransform):
+                outputs = op(im, label)
+                im = outputs[0]
+                if len(outputs) == 2:
+                    label = outputs[1]
+            else:
+                import imgaug.augmenters as iaa
+                if isinstance(op, iaa.Augmenter):
+                    im = execute_imgaug(op, im)
+                outputs = (im, )
+                if label is not None:
+                    outputs = (im, label)
+        return outputs
+
+    def add_augmenters(self, augmenters):
+        if not isinstance(augmenters, list):
+            raise Exception(
+                "augmenters should be list type in func add_augmenters()")
+        transform_names = [type(x).__name__ for x in self.transforms]
+        for aug in augmenters:
+            if type(aug).__name__ in transform_names:
+                logging.error("{} is already in ComposedTransforms, need to remove it from add_augmenters().".format(type(aug).__name__))
+        self.transforms = augmenters + self.transforms
+
+
+class RandomCrop(ClsTransform):
+    """对图像进行随机剪裁，模型训练时的数据增强操作。
+
+    1. 根据lower_scale、lower_ratio、upper_ratio计算随机剪裁的高、宽。
+    2. 根据随机剪裁的高、宽随机选取剪裁的起始点。
+    3. 剪裁图像。
+    4. 调整剪裁后的图像的大小到crop_size*crop_size。
+
+    Args:
+        crop_size (int): 随机裁剪后重新调整的目标边长。默认为224。
+        lower_scale (float): 裁剪面积相对原面积比例的最小限制。默认为0.08。
+        lower_ratio (float): 宽变换比例的最小限制。默认为3. / 4。
+        upper_ratio (float): 宽变换比例的最大限制。默认为4. / 3。
+    """
+
+    def __init__(self,
+                 crop_size=224,
+                 lower_scale=0.08,
+                 lower_ratio=3. / 4,
+                 upper_ratio=4. / 3):
+        self.crop_size = crop_size
+        self.lower_scale = lower_scale
+        self.lower_ratio = lower_ratio
+        self.upper_ratio = upper_ratio
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, )，对应图像np.ndarray数据；
+                   当label不为空时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、图像类别id。
+        """
+        im = random_crop(im, self.crop_size, self.lower_scale,
+                         self.lower_ratio, self.upper_ratio)
+        if label is None:
+            return (im, )
+        else:
+            return (im, label)
+
+
+class RandomHorizontalFlip(ClsTransform):
+    """以一定的概率对图像进行随机水平翻转，模型训练时的数据增强操作。
+
+    Args:
+        prob (float): 随机水平翻转的概率。默认为0.5。
+    """
+
+    def __init__(self, prob=0.5):
+        self.prob = prob
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, )，对应图像np.ndarray数据；
+                   当label不为空时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、图像类别id。
+        """
+        if random.random() < self.prob:
+            im = horizontal_flip(im)
+        if label is None:
+            return (im, )
+        else:
+            return (im, label)
+
+
+class RandomVerticalFlip(ClsTransform):
+    """以一定的概率对图像进行随机垂直翻转，模型训练时的数据增强操作。
+
+    Args:
+        prob (float): 随机垂直翻转的概率。默认为0.5。
+    """
+
+    def __init__(self, prob=0.5):
+        self.prob = prob
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, )，对应图像np.ndarray数据；
+                   当label不为空时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、图像类别id。
+        """
+        if random.random() < self.prob:
+            im = vertical_flip(im)
+        if label is None:
+            return (im, )
+        else:
+            return (im, label)
+
+
+class Normalize(ClsTransform):
+    """对图像进行标准化。
+
+    1. 对图像进行归一化到区间[0.0, 1.0]。
+    2. 对图像进行减均值除以标准差操作。
+
+    Args:
+        mean (list): 图像数据集的均值。默认为[0.485, 0.456, 0.406]。
+        std (list): 图像数据集的标准差。默认为[0.229, 0.224, 0.225]。
+
+    """
+
+    def __init__(self, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]):
+        self.mean = mean
+        self.std = std
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, )，对应图像np.ndarray数据；
+                   当label不为空时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、图像类别id。
+        """
+        mean = np.array(self.mean)[np.newaxis, np.newaxis, :]
+        std = np.array(self.std)[np.newaxis, np.newaxis, :]
+        im = normalize(im, mean, std)
+        if label is None:
+            return (im, )
+        else:
+            return (im, label)
+
+
+class ResizeByShort(ClsTransform):
+    """根据图像短边对图像重新调整大小（resize）。
+
+    1. 获取图像的长边和短边长度。
+    2. 根据短边与short_size的比例，计算长边的目标长度，
+       此时高、宽的resize比例为short_size/原图短边长度。
+    3. 如果max_size>0，调整resize比例：
+       如果长边的目标长度>max_size，则高、宽的resize比例为max_size/原图长边长度；
+    4. 根据调整大小的比例对图像进行resize。
+
+    Args:
+        short_size (int): 调整大小后的图像目标短边长度。默认为256。
+        max_size (int): 长边目标长度的最大限制。默认为-1。
+    """
+
+    def __init__(self, short_size=256, max_size=-1):
+        self.short_size = short_size
+        self.max_size = max_size
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, )，对应图像np.ndarray数据；
+                   当label不为空时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、图像类别id。
+        """
+        im_short_size = min(im.shape[0], im.shape[1])
+        im_long_size = max(im.shape[0], im.shape[1])
+        scale = float(self.short_size) / im_short_size
+        if self.max_size > 0 and np.round(scale *
+                                          im_long_size) > self.max_size:
+            scale = float(self.max_size) / float(im_long_size)
+        resized_width = int(round(im.shape[1] * scale))
+        resized_height = int(round(im.shape[0] * scale))
+        im = cv2.resize(
+            im, (resized_width, resized_height),
+            interpolation=cv2.INTER_LINEAR)
+
+        if label is None:
+            return (im, )
+        else:
+            return (im, label)
+
+
+class CenterCrop(ClsTransform):
+    """以图像中心点扩散裁剪长宽为`crop_size`的正方形
+
+    1. 计算剪裁的起始点。
+    2. 剪裁图像。
+
+    Args:
+        crop_size (int): 裁剪的目标边长。默认为224。
+    """
+
+    def __init__(self, crop_size=224):
+        self.crop_size = crop_size
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, )，对应图像np.ndarray数据；
+                   当label不为空时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、图像类别id。
+        """
+        im = center_crop(im, self.crop_size)
+        if label is None:
+            return (im, )
+        else:
+            return (im, label)
+
+
+class RandomRotate(ClsTransform):
+    def __init__(self, rotate_range=30, prob=0.5):
+        """以一定的概率对图像在[-rotate_range, rotaterange]角度范围内进行旋转，模型训练时的数据增强操作。
+
+        Args:
+            rotate_range (int): 旋转度数的范围。默认为30。
+            prob (float): 随机旋转的概率。默认为0.5。
+        """
+        self.rotate_range = rotate_range
+        self.prob = prob
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, )，对应图像np.ndarray数据；
+                   当label不为空时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、图像类别id。
+        """
+        rotate_lower = -self.rotate_range
+        rotate_upper = self.rotate_range
+        im = im.astype('uint8')
+        im = Image.fromarray(im)
+        if np.random.uniform(0, 1) < self.prob:
+            im = rotate(im, rotate_lower, rotate_upper)
+        im = np.asarray(im).astype('float32')
+        if label is None:
+            return (im, )
+        else:
+            return (im, label)
+
+
+class RandomDistort(ClsTransform):
+    """以一定的概率对图像进行随机像素内容变换，模型训练时的数据增强操作。
+
+    1. 对变换的操作顺序进行随机化操作。
+    2. 按照1中的顺序以一定的概率对图像在范围[-range, range]内进行随机像素内容变换。
+
+    Args:
+        brightness_range (float): 明亮度因子的范围。默认为0.9。
+        brightness_prob (float): 随机调整明亮度的概率。默认为0.5。
+        contrast_range (float): 对比度因子的范围。默认为0.9。
+        contrast_prob (float): 随机调整对比度的概率。默认为0.5。
+        saturation_range (float): 饱和度因子的范围。默认为0.9。
+        saturation_prob (float): 随机调整饱和度的概率。默认为0.5。
+        hue_range (int): 色调因子的范围。默认为18。
+        hue_prob (float): 随机调整色调的概率。默认为0.5。
+    """
+
+    def __init__(self,
+                 brightness_range=0.9,
+                 brightness_prob=0.5,
+                 contrast_range=0.9,
+                 contrast_prob=0.5,
+                 saturation_range=0.9,
+                 saturation_prob=0.5,
+                 hue_range=18,
+                 hue_prob=0.5):
+        self.brightness_range = brightness_range
+        self.brightness_prob = brightness_prob
+        self.contrast_range = contrast_range
+        self.contrast_prob = contrast_prob
+        self.saturation_range = saturation_range
+        self.saturation_prob = saturation_prob
+        self.hue_range = hue_range
+        self.hue_prob = hue_prob
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, )，对应图像np.ndarray数据；
+                   当label不为空时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、图像类别id。
+        """
+        brightness_lower = 1 - self.brightness_range
+        brightness_upper = 1 + self.brightness_range
+        contrast_lower = 1 - self.contrast_range
+        contrast_upper = 1 + self.contrast_range
+        saturation_lower = 1 - self.saturation_range
+        saturation_upper = 1 + self.saturation_range
+        hue_lower = -self.hue_range
+        hue_upper = self.hue_range
+        ops = [brightness, contrast, saturation, hue]
+        random.shuffle(ops)
+        params_dict = {
+            'brightness': {
+                'brightness_lower': brightness_lower,
+                'brightness_upper': brightness_upper
+            },
+            'contrast': {
+                'contrast_lower': contrast_lower,
+                'contrast_upper': contrast_upper
+            },
+            'saturation': {
+                'saturation_lower': saturation_lower,
+                'saturation_upper': saturation_upper
+            },
+            'hue': {
+                'hue_lower': hue_lower,
+                'hue_upper': hue_upper
+            }
+        }
+        prob_dict = {
+            'brightness': self.brightness_prob,
+            'contrast': self.contrast_prob,
+            'saturation': self.saturation_prob,
+            'hue': self.hue_prob,
+        }
+        for id in range(len(ops)):
+            params = params_dict[ops[id].__name__]
+            prob = prob_dict[ops[id].__name__]
+            params['im'] = im
+            if np.random.uniform(0, 1) < prob:
+                im = ops[id](**params)
+        if label is None:
+            return (im, )
+        else:
+            return (im, label)
+
+
+class ArrangeClassifier(ClsTransform):
+    """获取训练/验证/预测所需信息。注意：此操作不需用户自己显示调用
+
+    Args:
+        mode (str): 指定数据用于何种用途，取值范围为['train', 'eval', 'test', 'quant']。
+
+    Raises:
+        ValueError: mode的取值不在['train', 'eval', 'test', 'quant']之内。
+    """
+
+    def __init__(self, mode=None):
+        if mode not in ['train', 'eval', 'test', 'quant']:
+            raise ValueError(
+                "mode must be in ['train', 'eval', 'test', 'quant']!")
+        self.mode = mode
+
+    def __call__(self, im, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            label (int): 每张图像所对应的类别序号。
+
+        Returns:
+            tuple: 当mode为'train'或'eval'时，返回(im, label)，分别对应图像np.ndarray数据、
+                图像类别id；当mode为'test'或'quant'时，返回(im, )，对应图像np.ndarray数据。
+        """
+        im = permute(im, False).astype('float32')
+        if self.mode == 'train' or self.mode == 'eval':
+            outputs = (im, label)
+        else:
+            outputs = (im, )
+        return outputs
+
+
+class ComposedClsTransforms(Compose):
+    """ 分类模型的基础Transforms流程，具体如下
+        训练阶段：
+        1. 随机从图像中crop一块子图，并resize成crop_size大小
+        2. 将1的输出按0.5的概率随机进行水平翻转
+        3. 将图像进行归一化
+        验证/预测阶段：
+        1. 将图像按比例Resize，使得最小边长度为crop_size[0] * 1.14
+        2. 从图像中心crop出一个大小为crop_size的图像
+        3. 将图像进行归一化
+
+        Args:
+            mode(str): 图像处理流程所处阶段，训练/验证/预测，分别对应'train', 'eval', 'test'
+            crop_size(int|list): 输入模型里的图像大小
+            mean(list): 图像均值
+            std(list): 图像方差
+    """
+
+    def __init__(self,
+                 mode,
+                 crop_size=[224, 224],
+                 mean=[0.485, 0.456, 0.406],
+                 std=[0.229, 0.224, 0.225]):
+        width = crop_size
+        if isinstance(crop_size, list):
+            if crop_size[0] != crop_size[1]:
+                raise Exception(
+                    "In classifier model, width and height should be equal, please modify your parameter `crop_size`"
+                )
+            width = crop_size[0]
+        if width % 32 != 0:
+            raise Exception(
+                "In classifier model, width and height should be multiple of 32, e.g 224、256、320...., please modify your parameter `crop_size`"
+            )
+
+        if mode == 'train':
+            # 训练时的transforms，包含数据增强
+            transforms = [
+                RandomCrop(crop_size=width), RandomHorizontalFlip(prob=0.5),
+                Normalize(
+                    mean=mean, std=std)
+            ]
+        else:
+            # 验证/预测时的transforms
+            transforms = [
+                ResizeByShort(short_size=int(width * 1.14)),
+                CenterCrop(crop_size=width), Normalize(
+                    mean=mean, std=std)
+            ]
+
+        super(ComposedClsTransforms, self).__init__(transforms)

deploy/openvino/python/transforms/det_transforms.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+try:
+    from collections.abc import Sequence
+except Exception:
+    from collections import Sequence
+
+import random
+import os.path as osp
+import numpy as np
+
+import cv2
+from PIL import Image, ImageEnhance
+
+from .imgaug_support import execute_imgaug
+from .ops import *
+from .box_utils import *
+import utils.logging as logging
+
+
+class DetTransform:
+    """检测数据处理基类
+    """
+
+    def __init__(self):
+        pass
+
+
+class Compose(DetTransform):
+    """根据数据预处理/增强列表对输入数据进行操作。
+       所有操作的输入图像流形状均是[H, W, C]，其中H为图像高，W为图像宽，C为图像通道数。
+
+    Args:
+        transforms (list): 数据预处理/增强列表。
+
+    Raises:
+        TypeError: 形参数据类型不满足需求。
+        ValueError: 数据长度不匹配。
+    """
+
+    def __init__(self, transforms):
+        if not isinstance(transforms, list):
+            raise TypeError('The transforms must be a list!')
+        if len(transforms) < 1:
+            raise ValueError('The length of transforms ' + \
+                            'must be equal or larger than 1!')
+        self.transforms = transforms
+        self.use_mixup = False
+        for t in self.transforms:
+            if type(t).__name__ == 'MixupImage':
+                self.use_mixup = True
+        # 检查transforms里面的操作，目前支持PaddleX定义的或者是imgaug操作
+        for op in self.transforms:
+            if not isinstance(op, DetTransform):
+                import imgaug.augmenters as iaa
+                if not isinstance(op, iaa.Augmenter):
+                    raise Exception(
+                        "Elements in transforms should be defined in 'paddlex.det.transforms' or class of imgaug.augmenters.Augmenter, see docs here: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/"
+                    )
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (str/np.ndarray): 图像路径/图像np.ndarray数据。
+            im_info (dict): 存储与图像相关的信息，dict中的字段如下：
+                - im_id (np.ndarray): 图像序列号，形状为(1,)。
+                - image_shape (np.ndarray): 图像原始大小，形状为(2,)，
+                                        image_shape[0]为高，image_shape[1]为宽。
+                - mixup (list): list为[im, im_info, label_info]，分别对应
+                                与当前图像进行mixup的图像np.ndarray数据、图像相关信息、标注框相关信息；
+                                注意，当前epoch若无需进行mixup，则无该字段。
+            label_info (dict): 存储与标注框相关的信息，dict中的字段如下：
+                - gt_bbox (np.ndarray): 真实标注框坐标[x1, y1, x2, y2]，形状为(n, 4)，
+                                   其中n代表真实标注框的个数。
+                - gt_class (np.ndarray): 每个真实标注框对应的类别序号，形状为(n, 1)，
+                                    其中n代表真实标注框的个数。
+                - gt_score (np.ndarray): 每个真实标注框对应的混合得分，形状为(n, 1)，
+                                    其中n代表真实标注框的个数。
+                - gt_poly (list): 每个真实标注框内的多边形分割区域，每个分割区域由点的x、y坐标组成，
+                                  长度为n，其中n代表真实标注框的个数。
+                - is_crowd (np.ndarray): 每个真实标注框中是否是一组对象，形状为(n, 1)，
+                                    其中n代表真实标注框的个数。
+                - difficult (np.ndarray): 每个真实标注框中的对象是否为难识别对象，形状为(n, 1)，
+                                     其中n代表真实标注框的个数。
+        Returns:
+            tuple: 根据网络所需字段所组成的tuple；
+                字段由transforms中的最后一个数据预处理操作决定。
+        """
+
+        def decode_image(im_file, im_info, label_info):
+            if im_info is None:
+                im_info = dict()
+            if isinstance(im_file, np.ndarray):
+                if len(im_file.shape) != 3:
+                    raise Exception(
+                        "im should be 3-dimensions, but now is {}-dimensions".
+                        format(len(im_file.shape)))
+                im = im_file
+            else:
+                try:
+                    im = cv2.imread(im_file).astype('float32')
+                except:
+                    raise TypeError('Can\'t read The image file {}!'.format(
+                        im_file))
+            im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
+            # make default im_info with [h, w, 1]
+            im_info['im_resize_info'] = np.array(
+                [im.shape[0], im.shape[1], 1.], dtype=np.float32)
+            im_info['image_shape'] = np.array([im.shape[0],
+                                               im.shape[1]]).astype('int32')
+            if not self.use_mixup:
+                if 'mixup' in im_info:
+                    del im_info['mixup']
+            # decode mixup image
+            if 'mixup' in im_info:
+                im_info['mixup'] = \
+                  decode_image(im_info['mixup'][0],
+                               im_info['mixup'][1],
+                               im_info['mixup'][2])
+            if label_info is None:
+                return (im, im_info)
+            else:
+                return (im, im_info, label_info)
+
+        outputs = decode_image(im, im_info, label_info)
+        im = outputs[0]
+        im_info = outputs[1]
+        if len(outputs) == 3:
+            label_info = outputs[2]
+        for op in self.transforms:
+            if im is None:
+                return None
+            if isinstance(op, DetTransform):
+                outputs = op(im, im_info, label_info)
+                im = outputs[0]
+            else:
+                im = execute_imgaug(op, im)
+                if label_info is not None:
+                    outputs = (im, im_info, label_info)
+                else:
+                    outputs = (im, im_info)
+        return outputs
+
+    def add_augmenters(self, augmenters):
+        if not isinstance(augmenters, list):
+            raise Exception(
+                "augmenters should be list type in func add_augmenters()")
+        transform_names = [type(x).__name__ for x in self.transforms]
+        for aug in augmenters:
+            if type(aug).__name__ in transform_names:
+                logging.error("{} is already in ComposedTransforms, need to remove it from add_augmenters().".format(type(aug).__name__))
+        self.transforms = augmenters + self.transforms
+
+
+class ResizeByShort(DetTransform):
+    """根据图像的短边调整图像大小（resize）。
+
+    1. 获取图像的长边和短边长度。
+    2. 根据短边与short_size的比例，计算长边的目标长度，
+       此时高、宽的resize比例为short_size/原图短边长度。
+    3. 如果max_size>0，调整resize比例：
+       如果长边的目标长度>max_size，则高、宽的resize比例为max_size/原图长边长度。
+    4. 根据调整大小的比例对图像进行resize。
+
+    Args:
+        target_size (int): 短边目标长度。默认为800。
+        max_size (int): 长边目标长度的最大限制。默认为1333。
+
+     Raises:
+        TypeError: 形参数据类型不满足需求。
+    """
+
+    def __init__(self, short_size=800, max_size=1333):
+        self.max_size = int(max_size)
+        if not isinstance(short_size, int):
+            raise TypeError(
+                "Type of short_size is invalid. Must be Integer, now is {}".
+                format(type(short_size)))
+        self.short_size = short_size
+        if not (isinstance(self.max_size, int)):
+            raise TypeError("max_size: input type is invalid.")
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (numnp.ndarraypy): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当label_info为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label_info不为空时，返回的tuple为(im, im_info, label_info)，分别对应图像np.ndarray数据、
+                   存储与标注框相关信息的字典。
+                   其中，im_info更新字段为：
+                       - im_resize_info (np.ndarray): resize后的图像高、resize后的图像宽、resize后的图像相对原始图的缩放比例
+                                                 三者组成的np.ndarray，形状为(3,)。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+            ValueError: 数据长度不匹配。
+        """
+        if im_info is None:
+            im_info = dict()
+        if not isinstance(im, np.ndarray):
+            raise TypeError("ResizeByShort: image type is not numpy.")
+        if len(im.shape) != 3:
+            raise ValueError('ResizeByShort: image is not 3-dimensional.')
+        im_short_size = min(im.shape[0], im.shape[1])
+        im_long_size = max(im.shape[0], im.shape[1])
+        scale = float(self.short_size) / im_short_size
+        if self.max_size > 0 and np.round(scale *
+                                          im_long_size) > self.max_size:
+            scale = float(self.max_size) / float(im_long_size)
+        resized_width = int(round(im.shape[1] * scale))
+        resized_height = int(round(im.shape[0] * scale))
+        im_resize_info = [resized_height, resized_width, scale]
+        im = cv2.resize(
+            im, (resized_width, resized_height),
+            interpolation=cv2.INTER_LINEAR)
+        im_info['im_resize_info'] = np.array(im_resize_info).astype(np.float32)
+        if label_info is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label_info)
+
+
+class Padding(DetTransform):
+    """1.将图像的长和宽padding至coarsest_stride的倍数。如输入图像为[300, 640],
+       `coarest_stride`为32，则由于300不为32的倍数，因此在图像最右和最下使用0值
+       进行padding，最终输出图像为[320, 640]。
+       2.或者，将图像的长和宽padding到target_size指定的shape，如输入的图像为[300，640]，
+         a. `target_size` = 960，在图像最右和最下使用0值进行padding，最终输出
+            图像为[960, 960]。
+         b. `target_size` = [640, 960]，在图像最右和最下使用0值进行padding，最终
+            输出图像为[640, 960]。
+
+    1. 如果coarsest_stride为1，target_size为None则直接返回。
+    2. 获取图像的高H、宽W。
+    3. 计算填充后图像的高H_new、宽W_new。
+    4. 构建大小为(H_new, W_new, 3)像素值为0的np.ndarray，
+       并将原图的np.ndarray粘贴于左上角。
+
+    Args:
+        coarsest_stride (int): 填充后的图像长、宽为该参数的倍数，默认为1。
+        target_size (int|list|tuple): 填充后的图像长、宽，默认为None，coarset_stride优先级更高。
+
+    Raises:
+        TypeError: 形参`target_size`数据类型不满足需求。
+        ValueError: 形参`target_size`为(list|tuple)时，长度不满足需求。
+    """
+
+    def __init__(self, coarsest_stride=1, target_size=None):
+        self.coarsest_stride = coarsest_stride
+        if target_size is not None:
+            if not isinstance(target_size, int):
+                if not isinstance(target_size, tuple) and not isinstance(
+                        target_size, list):
+                    raise TypeError(
+                        "Padding: Type of target_size must in (int|list|tuple)."
+                    )
+                elif len(target_size) != 2:
+                    raise ValueError(
+                        "Padding: Length of target_size must equal 2.")
+        self.target_size = target_size
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (numnp.ndarraypy): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当label_info为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label_info不为空时，返回的tuple为(im, im_info, label_info)，分别对应图像np.ndarray数据、
+                   存储与标注框相关信息的字典。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+            ValueError: 数据长度不匹配。
+            ValueError: coarsest_stride，target_size需有且只有一个被指定。
+            ValueError: target_size小于原图的大小。
+        """
+        if im_info is None:
+            im_info = dict()
+        if not isinstance(im, np.ndarray):
+            raise TypeError("Padding: image type is not numpy.")
+        if len(im.shape) != 3:
+            raise ValueError('Padding: image is not 3-dimensional.')
+        im_h, im_w, im_c = im.shape[:]
+
+        if isinstance(self.target_size, int):
+            padding_im_h = self.target_size
+            padding_im_w = self.target_size
+        elif isinstance(self.target_size, list) or isinstance(self.target_size,
+                                                              tuple):
+            padding_im_w = self.target_size[0]
+            padding_im_h = self.target_size[1]
+        elif self.coarsest_stride > 0:
+            padding_im_h = int(
+                np.ceil(im_h / self.coarsest_stride) * self.coarsest_stride)
+            padding_im_w = int(
+                np.ceil(im_w / self.coarsest_stride) * self.coarsest_stride)
+        else:
+            raise ValueError(
+                "coarsest_stridei(>1) or target_size(list|int) need setting in Padding transform"
+            )
+        pad_height = padding_im_h - im_h
+        pad_width = padding_im_w - im_w
+        if pad_height < 0 or pad_width < 0:
+            raise ValueError(
+                'the size of image should be less than target_size, but the size of image ({}, {}), is larger than target_size ({}, {})'
+                .format(im_w, im_h, padding_im_w, padding_im_h))
+        padding_im = np.zeros(
+            (padding_im_h, padding_im_w, im_c), dtype=np.float32)
+        padding_im[:im_h, :im_w, :] = im
+        if label_info is None:
+            return (padding_im, im_info)
+        else:
+            return (padding_im, im_info, label_info)
+
+
+class Resize(DetTransform):
+    """调整图像大小（resize）。
+
+    - 当目标大小（target_size）类型为int时，根据插值方式，
+      将图像resize为[target_size, target_size]。
+    - 当目标大小（target_size）类型为list或tuple时，根据插值方式，
+      将图像resize为target_size。
+    注意：当插值方式为“RANDOM”时，则随机选取一种插值方式进行resize。
+
+    Args:
+        target_size (int/list/tuple): 短边目标长度。默认为608。
+        interp (str): resize的插值方式，与opencv的插值方式对应，取值范围为
+            ['NEAREST', 'LINEAR', 'CUBIC', 'AREA', 'LANCZOS4', 'RANDOM']。默认为"LINEAR"。
+
+    Raises:
+        TypeError: 形参数据类型不满足需求。
+        ValueError: 插值方式不在['NEAREST', 'LINEAR', 'CUBIC',
+                    'AREA', 'LANCZOS4', 'RANDOM']中。
+    """
+
+    # The interpolation mode
+    interp_dict = {
+        'NEAREST': cv2.INTER_NEAREST,
+        'LINEAR': cv2.INTER_LINEAR,
+        'CUBIC': cv2.INTER_CUBIC,
+        'AREA': cv2.INTER_AREA,
+        'LANCZOS4': cv2.INTER_LANCZOS4
+    }
+
+    def __init__(self, target_size=608, interp='LINEAR'):
+        self.interp = interp
+        if not (interp == "RANDOM" or interp in self.interp_dict):
+            raise ValueError("interp should be one of {}".format(
+                self.interp_dict.keys()))
+        if isinstance(target_size, list) or isinstance(target_size, tuple):
+            if len(target_size) != 2:
+                raise TypeError(
+                    'when target is list or tuple, it should include 2 elements, but it is {}'
+                    .format(target_size))
+        elif not isinstance(target_size, int):
+            raise TypeError(
+                "Type of target_size is invalid. Must be Integer or List or tuple, now is {}"
+                .format(type(target_size)))
+
+        self.target_size = target_size
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当label_info为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label_info不为空时，返回的tuple为(im, im_info, label_info)，分别对应图像np.ndarray数据、
+                   存储与标注框相关信息的字典。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+            ValueError: 数据长度不匹配。
+        """
+        if im_info is None:
+            im_info = dict()
+        if not isinstance(im, np.ndarray):
+            raise TypeError("Resize: image type is not numpy.")
+        if len(im.shape) != 3:
+            raise ValueError('Resize: image is not 3-dimensional.')
+        if self.interp == "RANDOM":
+            interp = random.choice(list(self.interp_dict.keys()))
+        else:
+            interp = self.interp
+        im = resize(im, self.target_size, self.interp_dict[interp])
+        if label_info is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label_info)
+
+
+class RandomHorizontalFlip(DetTransform):
+    """随机翻转图像、标注框、分割信息，模型训练时的数据增强操作。
+
+    1. 随机采样一个0-1之间的小数，当小数小于水平翻转概率时，
+       执行2-4步操作，否则直接返回。
+    2. 水平翻转图像。
+    3. 计算翻转后的真实标注框的坐标，更新label_info中的gt_bbox信息。
+    4. 计算翻转后的真实分割区域的坐标，更新label_info中的gt_poly信息。
+
+    Args:
+        prob (float): 随机水平翻转的概率。默认为0.5。
+
+    Raises:
+        TypeError: 形参数据类型不满足需求。
+    """
+
+    def __init__(self, prob=0.5):
+        self.prob = prob
+        if not isinstance(self.prob, float):
+            raise TypeError("RandomHorizontalFlip: input type is invalid.")
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当label_info为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label_info不为空时，返回的tuple为(im, im_info, label_info)，分别对应图像np.ndarray数据、
+                   存储与标注框相关信息的字典。
+                   其中，im_info更新字段为：
+                       - gt_bbox (np.ndarray): 水平翻转后的标注框坐标[x1, y1, x2, y2]，形状为(n, 4)，
+                                          其中n代表真实标注框的个数。
+                       - gt_poly (list): 水平翻转后的多边形分割区域的x、y坐标，长度为n，
+                                         其中n代表真实标注框的个数。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+            ValueError: 数据长度不匹配。
+        """
+        if not isinstance(im, np.ndarray):
+            raise TypeError(
+                "RandomHorizontalFlip: image is not a numpy array.")
+        if len(im.shape) != 3:
+            raise ValueError(
+                "RandomHorizontalFlip: image is not 3-dimensional.")
+        if im_info is None or label_info is None:
+            raise TypeError(
+                'Cannot do RandomHorizontalFlip! ' +
+                'Becasuse the im_info and label_info can not be None!')
+        if 'gt_bbox' not in label_info:
+            raise TypeError('Cannot do RandomHorizontalFlip! ' + \
+                            'Becasuse gt_bbox is not in label_info!')
+        image_shape = im_info['image_shape']
+        gt_bbox = label_info['gt_bbox']
+        height = image_shape[0]
+        width = image_shape[1]
+
+        if np.random.uniform(0, 1) < self.prob:
+            im = horizontal_flip(im)
+            if gt_bbox.shape[0] == 0:
+                if label_info is None:
+                    return (im, im_info)
+                else:
+                    return (im, im_info, label_info)
+            label_info['gt_bbox'] = box_horizontal_flip(gt_bbox, width)
+            if 'gt_poly' in label_info and \
+                    len(label_info['gt_poly']) != 0:
+                label_info['gt_poly'] = segms_horizontal_flip(
+                    label_info['gt_poly'], height, width)
+        if label_info is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label_info)
+
+
+class Normalize(DetTransform):
+    """对图像进行标准化。
+
+    1. 归一化图像到到区间[0.0, 1.0]。
+    2. 对图像进行减均值除以标准差操作。
+
+    Args:
+        mean (list): 图像数据集的均值。默认为[0.485, 0.456, 0.406]。
+        std (list): 图像数据集的标准差。默认为[0.229, 0.224, 0.225]。
+
+    Raises:
+        TypeError: 形参数据类型不满足需求。
+    """
+
+    def __init__(self, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]):
+        self.mean = mean
+        self.std = std
+        if not (isinstance(self.mean, list) and isinstance(self.std, list)):
+            raise TypeError("NormalizeImage: input type is invalid.")
+        from functools import reduce
+        if reduce(lambda x, y: x * y, self.std) == 0:
+            raise TypeError('NormalizeImage: std is invalid!')
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (numnp.ndarraypy): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当label_info为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label_info不为空时，返回的tuple为(im, im_info, label_info)，分别对应图像np.ndarray数据、
+                   存储与标注框相关信息的字典。
+        """
+        mean = np.array(self.mean)[np.newaxis, np.newaxis, :]
+        std = np.array(self.std)[np.newaxis, np.newaxis, :]
+        im = normalize(im, mean, std)
+        if label_info is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label_info)
+
+
+class RandomDistort(DetTransform):
+    """以一定的概率对图像进行随机像素内容变换，模型训练时的数据增强操作
+
+    1. 对变换的操作顺序进行随机化操作。
+    2. 按照1中的顺序以一定的概率在范围[-range, range]对图像进行随机像素内容变换。
+
+    Args:
+        brightness_range (float): 明亮度因子的范围。默认为0.5。
+        brightness_prob (float): 随机调整明亮度的概率。默认为0.5。
+        contrast_range (float): 对比度因子的范围。默认为0.5。
+        contrast_prob (float): 随机调整对比度的概率。默认为0.5。
+        saturation_range (float): 饱和度因子的范围。默认为0.5。
+        saturation_prob (float): 随机调整饱和度的概率。默认为0.5。
+        hue_range (int): 色调因子的范围。默认为18。
+        hue_prob (float): 随机调整色调的概率。默认为0.5。
+    """
+
+    def __init__(self,
+                 brightness_range=0.5,
+                 brightness_prob=0.5,
+                 contrast_range=0.5,
+                 contrast_prob=0.5,
+                 saturation_range=0.5,
+                 saturation_prob=0.5,
+                 hue_range=18,
+                 hue_prob=0.5):
+        self.brightness_range = brightness_range
+        self.brightness_prob = brightness_prob
+        self.contrast_range = contrast_range
+        self.contrast_prob = contrast_prob
+        self.saturation_range = saturation_range
+        self.saturation_prob = saturation_prob
+        self.hue_range = hue_range
+        self.hue_prob = hue_prob
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当label_info为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label_info不为空时，返回的tuple为(im, im_info, label_info)，分别对应图像np.ndarray数据、
+                   存储与标注框相关信息的字典。
+        """
+        brightness_lower = 1 - self.brightness_range
+        brightness_upper = 1 + self.brightness_range
+        contrast_lower = 1 - self.contrast_range
+        contrast_upper = 1 + self.contrast_range
+        saturation_lower = 1 - self.saturation_range
+        saturation_upper = 1 + self.saturation_range
+        hue_lower = -self.hue_range
+        hue_upper = self.hue_range
+        ops = [brightness, contrast, saturation, hue]
+        random.shuffle(ops)
+        params_dict = {
+            'brightness': {
+                'brightness_lower': brightness_lower,
+                'brightness_upper': brightness_upper
+            },
+            'contrast': {
+                'contrast_lower': contrast_lower,
+                'contrast_upper': contrast_upper
+            },
+            'saturation': {
+                'saturation_lower': saturation_lower,
+                'saturation_upper': saturation_upper
+            },
+            'hue': {
+                'hue_lower': hue_lower,
+                'hue_upper': hue_upper
+            }
+        }
+        prob_dict = {
+            'brightness': self.brightness_prob,
+            'contrast': self.contrast_prob,
+            'saturation': self.saturation_prob,
+            'hue': self.hue_prob
+        }
+        for id in range(4):
+            params = params_dict[ops[id].__name__]
+            prob = prob_dict[ops[id].__name__]
+            params['im'] = im
+
+            if np.random.uniform(0, 1) < prob:
+                im = ops[id](**params)
+        if label_info is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label_info)
+
+
+class MixupImage(DetTransform):
+    """对图像进行mixup操作,模型训练时的数据增强操作，目前仅YOLOv3模型支持该transform。
+
+    当label_info中不存在mixup字段时，直接返回，否则进行下述操作：
+    1. 从随机beta分布中抽取出随机因子factor。
+    2.
+        - 当factor>=1.0时，去除label_info中的mixup字段，直接返回。
+        - 当factor<=0.0时，直接返回label_info中的mixup字段，并在label_info中去除该字段。
+        - 其余情况，执行下述操作：
+            （1）原图像乘以factor，mixup图像乘以(1-factor)，叠加2个结果。
+            （2）拼接原图像标注框和mixup图像标注框。
+            （3）拼接原图像标注框类别和mixup图像标注框类别。
+            （4）原图像标注框混合得分乘以factor，mixup图像标注框混合得分乘以(1-factor)，叠加2个结果。
+    3. 更新im_info中的image_shape信息。
+
+    Args:
+        alpha (float): 随机beta分布的下限。默认为1.5。
+        beta (float): 随机beta分布的上限。默认为1.5。
+        mixup_epoch (int): 在前mixup_epoch轮使用mixup增强操作；当该参数为-1时，该策略不会生效。
+            默认为-1。
+
+    Raises:
+        ValueError: 数据长度不匹配。
+    """
+
+    def __init__(self, alpha=1.5, beta=1.5, mixup_epoch=-1):
+        self.alpha = alpha
+        self.beta = beta
+        if self.alpha <= 0.0:
+            raise ValueError("alpha shold be positive in MixupImage")
+        if self.beta <= 0.0:
+            raise ValueError("beta shold be positive in MixupImage")
+        self.mixup_epoch = mixup_epoch
+
+    def _mixup_img(self, img1, img2, factor):
+        h = max(img1.shape[0], img2.shape[0])
+        w = max(img1.shape[1], img2.shape[1])
+        img = np.zeros((h, w, img1.shape[2]), 'float32')
+        img[:img1.shape[0], :img1.shape[1], :] = \
+            img1.astype('float32') * factor
+        img[:img2.shape[0], :img2.shape[1], :] += \
+            img2.astype('float32') * (1.0 - factor)
+        return img.astype('float32')
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当label_info为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label_info不为空时，返回的tuple为(im, im_info, label_info)，分别对应图像np.ndarray数据、
+                   存储与标注框相关信息的字典。
+                   其中，im_info更新字段为：
+                       - image_shape (np.ndarray): mixup后的图像高、宽二者组成的np.ndarray，形状为(2,)。
+                   im_info删除的字段：
+                       - mixup (list): 与当前字段进行mixup的图像相关信息。
+                   label_info更新字段为：
+                       - gt_bbox (np.ndarray): mixup后真实标注框坐标，形状为(n, 4)，
+                                          其中n代表真实标注框的个数。
+                       - gt_class (np.ndarray): mixup后每个真实标注框对应的类别序号，形状为(n, 1)，
+                                           其中n代表真实标注框的个数。
+                       - gt_score (np.ndarray): mixup后每个真实标注框对应的混合得分，形状为(n, 1)，
+                                           其中n代表真实标注框的个数。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+        """
+        if im_info is None:
+            raise TypeError('Cannot do MixupImage! ' +
+                            'Becasuse the im_info can not be None!')
+        if 'mixup' not in im_info:
+            if label_info is None:
+                return (im, im_info)
+            else:
+                return (im, im_info, label_info)
+        factor = np.random.beta(self.alpha, self.beta)
+        factor = max(0.0, min(1.0, factor))
+        if im_info['epoch'] > self.mixup_epoch \
+                or factor >= 1.0:
+            im_info.pop('mixup')
+            if label_info is None:
+                return (im, im_info)
+            else:
+                return (im, im_info, label_info)
+        if factor <= 0.0:
+            return im_info.pop('mixup')
+        im = self._mixup_img(im, im_info['mixup'][0], factor)
+        if label_info is None:
+            raise TypeError('Cannot do MixupImage! ' +
+                            'Becasuse the label_info can not be None!')
+        if 'gt_bbox' not in label_info or \
+                'gt_class' not in label_info or \
+                'gt_score' not in label_info:
+            raise TypeError('Cannot do MixupImage! ' + \
+                            'Becasuse gt_bbox/gt_class/gt_score is not in label_info!')
+        gt_bbox1 = label_info['gt_bbox']
+        gt_bbox2 = im_info['mixup'][2]['gt_bbox']
+        gt_class1 = label_info['gt_class']
+        gt_class2 = im_info['mixup'][2]['gt_class']
+        gt_score1 = label_info['gt_score']
+        gt_score2 = im_info['mixup'][2]['gt_score']
+        if 'gt_poly' in label_info:
+            gt_poly1 = label_info['gt_poly']
+            gt_poly2 = im_info['mixup'][2]['gt_poly']
+        is_crowd1 = label_info['is_crowd']
+        is_crowd2 = im_info['mixup'][2]['is_crowd']
+        
+        if 0 not in gt_class1 and 0 not in gt_class2:
+            gt_bbox = np.concatenate((gt_bbox1, gt_bbox2), axis=0)
+            gt_class = np.concatenate((gt_class1, gt_class2), axis=0)
+            gt_score = np.concatenate(
+                (gt_score1 * factor, gt_score2 * (1. - factor)), axis=0)
+            if 'gt_poly' in label_info:
+                label_info['gt_poly'] = gt_poly1 + gt_poly2
+            is_crowd = np.concatenate((is_crowd1, is_crowd2), axis=0)
+        elif 0 in gt_class1:
+            gt_bbox = gt_bbox2
+            gt_class = gt_class2
+            gt_score = gt_score2 * (1. - factor)
+            if 'gt_poly' in label_info:
+                label_info['gt_poly'] = gt_poly2
+            is_crowd = is_crowd2
+        else:
+            gt_bbox = gt_bbox1
+            gt_class = gt_class1
+            gt_score = gt_score1 * factor
+            if 'gt_poly' in label_info:
+                label_info['gt_poly'] = gt_poly1
+            is_crowd = is_crowd1
+        label_info['gt_bbox'] = gt_bbox
+        label_info['gt_score'] = gt_score
+        label_info['gt_class'] = gt_class
+        label_info['is_crowd'] = is_crowd
+        im_info['image_shape'] = np.array([im.shape[0],
+                                           im.shape[1]]).astype('int32')
+        im_info.pop('mixup')
+        if label_info is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label_info)
+
+
+class RandomExpand(DetTransform):
+    """随机扩张图像，模型训练时的数据增强操作。
+    1. 随机选取扩张比例（扩张比例大于1时才进行扩张）。
+    2. 计算扩张后图像大小。
+    3. 初始化像素值为输入填充值的图像，并将原图像随机粘贴于该图像上。
+    4. 根据原图像粘贴位置换算出扩张后真实标注框的位置坐标。
+    5. 根据原图像粘贴位置换算出扩张后真实分割区域的位置坐标。
+    Args:
+        ratio (float): 图像扩张的最大比例。默认为4.0。
+        prob (float): 随机扩张的概率。默认为0.5。
+        fill_value (list): 扩张图像的初始填充值（0-255）。默认为[123.675, 116.28, 103.53]。
+    """
+
+    def __init__(self,
+                 ratio=4.,
+                 prob=0.5,
+                 fill_value=[123.675, 116.28, 103.53]):
+        super(RandomExpand, self).__init__()
+        assert ratio > 1.01, "expand ratio must be larger than 1.01"
+        self.ratio = ratio
+        self.prob = prob
+        assert isinstance(fill_value, Sequence), \
+            "fill value must be sequence"
+        if not isinstance(fill_value, tuple):
+            fill_value = tuple(fill_value)
+        self.fill_value = fill_value
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+        Returns:
+            tuple: 当label_info为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label_info不为空时，返回的tuple为(im, im_info, label_info)，分别对应图像np.ndarray数据、
+                   存储与标注框相关信息的字典。
+                   其中，im_info更新字段为：
+                       - image_shape (np.ndarray): 扩张后的图像高、宽二者组成的np.ndarray，形状为(2,)。
+                   label_info更新字段为：
+                       - gt_bbox (np.ndarray): 随机扩张后真实标注框坐标，形状为(n, 4)，
+                                          其中n代表真实标注框的个数。
+                       - gt_class (np.ndarray): 随机扩张后每个真实标注框对应的类别序号，形状为(n, 1)，
+                                           其中n代表真实标注框的个数。
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+        """
+        if im_info is None or label_info is None:
+            raise TypeError(
+                'Cannot do RandomExpand! ' +
+                'Becasuse the im_info and label_info can not be None!')
+        if 'gt_bbox' not in label_info or \
+                'gt_class' not in label_info:
+            raise TypeError('Cannot do RandomExpand! ' + \
+                            'Becasuse gt_bbox/gt_class is not in label_info!')
+        if np.random.uniform(0., 1.) < self.prob:
+            return (im, im_info, label_info)
+
+        if 'gt_class' in label_info and 0 in label_info['gt_class']:
+            return (im, im_info, label_info)
+        image_shape = im_info['image_shape']
+        height = int(image_shape[0])
+        width = int(image_shape[1])
+
+        expand_ratio = np.random.uniform(1., self.ratio)
+        h = int(height * expand_ratio)
+        w = int(width * expand_ratio)
+        if not h > height or not w > width:
+            return (im, im_info, label_info)
+        y = np.random.randint(0, h - height)
+        x = np.random.randint(0, w - width)
+        canvas = np.ones((h, w, 3), dtype=np.float32)
+        canvas *= np.array(self.fill_value, dtype=np.float32)
+        canvas[y:y + height, x:x + width, :] = im
+
+        im_info['image_shape'] = np.array([h, w]).astype('int32')
+        if 'gt_bbox' in label_info and len(label_info['gt_bbox']) > 0:
+            label_info['gt_bbox'] += np.array([x, y] * 2, dtype=np.float32)
+        if 'gt_poly' in label_info and len(label_info['gt_poly']) > 0:
+            label_info['gt_poly'] = expand_segms(label_info['gt_poly'], x, y,
+                                                 height, width, expand_ratio)
+        return (canvas, im_info, label_info)
+
+
+class RandomCrop(DetTransform):
+    """随机裁剪图像。
+    1. 若allow_no_crop为True，则在thresholds加入’no_crop’。
+    2. 随机打乱thresholds。
+    3. 遍历thresholds中各元素：
+        (1) 如果当前thresh为’no_crop’，则返回原始图像和标注信息。
+        (2) 随机取出aspect_ratio和scaling中的值并由此计算出候选裁剪区域的高、宽、起始点。
+        (3) 计算真实标注框与候选裁剪区域IoU，若全部真实标注框的IoU都小于thresh，则继续第3步。
+        (4) 如果cover_all_box为True且存在真实标注框的IoU小于thresh，则继续第3步。
+        (5) 筛选出位于候选裁剪区域内的真实标注框，若有效框的个数为0，则继续第3步，否则进行第4步。
+    4. 换算有效真值标注框相对候选裁剪区域的位置坐标。
+    5. 换算有效分割区域相对候选裁剪区域的位置坐标。
+
+    Args:
+        aspect_ratio (list): 裁剪后短边缩放比例的取值范围，以[min, max]形式表示。默认值为[.5, 2.]。
+        thresholds (list): 判断裁剪候选区域是否有效所需的IoU阈值取值列表。默认值为[.0, .1, .3, .5, .7, .9]。
+        scaling (list): 裁剪面积相对原面积的取值范围，以[min, max]形式表示。默认值为[.3, 1.]。
+        num_attempts (int): 在放弃寻找有效裁剪区域前尝试的次数。默认值为50。
+        allow_no_crop (bool): 是否允许未进行裁剪。默认值为True。
+        cover_all_box (bool): 是否要求所有的真实标注框都必须在裁剪区域内。默认值为False。
+    """
+
+    def __init__(self,
+                 aspect_ratio=[.5, 2.],
+                 thresholds=[.0, .1, .3, .5, .7, .9],
+                 scaling=[.3, 1.],
+                 num_attempts=50,
+                 allow_no_crop=True,
+                 cover_all_box=False):
+        self.aspect_ratio = aspect_ratio
+        self.thresholds = thresholds
+        self.scaling = scaling
+        self.num_attempts = num_attempts
+        self.allow_no_crop = allow_no_crop
+        self.cover_all_box = cover_all_box
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当label_info为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label_info不为空时，返回的tuple为(im, im_info, label_info)，分别对应图像np.ndarray数据、
+                   存储与标注框相关信息的字典。
+                   其中，im_info更新字段为：
+                           - image_shape (np.ndarray): 扩裁剪的图像高、宽二者组成的np.ndarray，形状为(2,)。
+                       label_info更新字段为：
+                           - gt_bbox (np.ndarray): 随机裁剪后真实标注框坐标，形状为(n, 4)，
+                                          其中n代表真实标注框的个数。
+                           - gt_class (np.ndarray): 随机裁剪后每个真实标注框对应的类别序号，形状为(n, 1)，
+                                           其中n代表真实标注框的个数。
+                           - gt_score (np.ndarray): 随机裁剪后每个真实标注框对应的混合得分，形状为(n, 1)，
+                                           其中n代表真实标注框的个数。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+        """
+        if im_info is None or label_info is None:
+            raise TypeError(
+                'Cannot do RandomCrop! ' +
+                'Becasuse the im_info and label_info can not be None!')
+        if 'gt_bbox' not in label_info or \
+                'gt_class' not in label_info:
+            raise TypeError('Cannot do RandomCrop! ' + \
+                            'Becasuse gt_bbox/gt_class is not in label_info!')
+
+        if len(label_info['gt_bbox']) == 0:
+            return (im, im_info, label_info)
+        if 'gt_class' in label_info and 0 in label_info['gt_class']:
+            return (im, im_info, label_info)
+
+        image_shape = im_info['image_shape']
+        w = image_shape[1]
+        h = image_shape[0]
+        gt_bbox = label_info['gt_bbox']
+        thresholds = list(self.thresholds)
+        if self.allow_no_crop:
+            thresholds.append('no_crop')
+        np.random.shuffle(thresholds)
+
+        for thresh in thresholds:
+            if thresh == 'no_crop':
+                return (im, im_info, label_info)
+
+            found = False
+            for i in range(self.num_attempts):
+                scale = np.random.uniform(*self.scaling)
+                min_ar, max_ar = self.aspect_ratio
+                aspect_ratio = np.random.uniform(
+                    max(min_ar, scale**2), min(max_ar, scale**-2))
+                crop_h = int(h * scale / np.sqrt(aspect_ratio))
+                crop_w = int(w * scale * np.sqrt(aspect_ratio))
+                crop_y = np.random.randint(0, h - crop_h)
+                crop_x = np.random.randint(0, w - crop_w)
+                crop_box = [crop_x, crop_y, crop_x + crop_w, crop_y + crop_h]
+                iou = iou_matrix(
+                    gt_bbox, np.array(
+                        [crop_box], dtype=np.float32))
+                if iou.max() < thresh:
+                    continue
+
+                if self.cover_all_box and iou.min() < thresh:
+                    continue
+
+                cropped_box, valid_ids = crop_box_with_center_constraint(
+                    gt_bbox, np.array(
+                        crop_box, dtype=np.float32))
+                if valid_ids.size > 0:
+                    found = True
+                    break
+
+            if found:
+                if 'gt_poly' in label_info and len(label_info['gt_poly']) > 0:
+                    crop_polys = crop_segms(
+                        label_info['gt_poly'],
+                        valid_ids,
+                        np.array(
+                            crop_box, dtype=np.int64),
+                        h,
+                        w)
+                    if [] in crop_polys:
+                        delete_id = list()
+                        valid_polys = list()
+                        for id, crop_poly in enumerate(crop_polys):
+                            if crop_poly == []:
+                                delete_id.append(id)
+                            else:
+                                valid_polys.append(crop_poly)
+                        valid_ids = np.delete(valid_ids, delete_id)
+                        if len(valid_polys) == 0:
+                            return (im, im_info, label_info)
+                        label_info['gt_poly'] = valid_polys
+                    else:
+                        label_info['gt_poly'] = crop_polys
+                im = crop_image(im, crop_box)
+                label_info['gt_bbox'] = np.take(cropped_box, valid_ids, axis=0)
+                label_info['gt_class'] = np.take(
+                    label_info['gt_class'], valid_ids, axis=0)
+                im_info['image_shape'] = np.array(
+                    [crop_box[3] - crop_box[1],
+                     crop_box[2] - crop_box[0]]).astype('int32')
+                if 'gt_score' in label_info:
+                    label_info['gt_score'] = np.take(
+                        label_info['gt_score'], valid_ids, axis=0)
+
+                if 'is_crowd' in label_info:
+                    label_info['is_crowd'] = np.take(
+                        label_info['is_crowd'], valid_ids, axis=0)
+                return (im, im_info, label_info)
+
+        return (im, im_info, label_info)
+
+
+class ArrangeFasterRCNN(DetTransform):
+    """获取FasterRCNN模型训练/验证/预测所需信息。
+
+    Args:
+        mode (str): 指定数据用于何种用途，取值范围为['train', 'eval', 'test', 'quant']。
+
+    Raises:
+        ValueError: mode的取值不在['train', 'eval', 'test', 'quant']之内。
+    """
+
+    def __init__(self, mode=None):
+        if mode not in ['train', 'eval', 'test', 'quant']:
+            raise ValueError(
+                "mode must be in ['train', 'eval', 'test', 'quant']!")
+        self.mode = mode
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当mode为'train'时，返回(im, im_resize_info, gt_bbox, gt_class, is_crowd)，分别对应
+                图像np.ndarray数据、图像相当对于原图的resize信息、真实标注框、真实标注框对应的类别、真实标注框内是否是一组对象；
+                当mode为'eval'时，返回(im, im_resize_info, im_id, im_shape, gt_bbox, gt_class, is_difficult)，
+                分别对应图像np.ndarray数据、图像相当对于原图的resize信息、图像id、图像大小信息、真实标注框、真实标注框对应的类别、
+                真实标注框是否为难识别对象；当mode为'test'或'quant'时，返回(im, im_resize_info, im_shape)，分别对应图像np.ndarray数据、
+                图像相当对于原图的resize信息、图像大小信息。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+            ValueError: 数据长度不匹配。
+        """
+        im = permute(im, False)
+        if self.mode == 'train':
+            if im_info is None or label_info is None:
+                raise TypeError(
+                    'Cannot do ArrangeFasterRCNN! ' +
+                    'Becasuse the im_info and label_info can not be None!')
+            if len(label_info['gt_bbox']) != len(label_info['gt_class']):
+                raise ValueError("gt num mismatch: bbox and class.")
+            im_resize_info = im_info['im_resize_info']
+            gt_bbox = label_info['gt_bbox']
+            gt_class = label_info['gt_class']
+            is_crowd = label_info['is_crowd']
+            outputs = (im, im_resize_info, gt_bbox, gt_class, is_crowd)
+        elif self.mode == 'eval':
+            if im_info is None or label_info is None:
+                raise TypeError(
+                    'Cannot do ArrangeFasterRCNN! ' +
+                    'Becasuse the im_info and label_info can not be None!')
+            im_resize_info = im_info['im_resize_info']
+            im_id = im_info['im_id']
+            im_shape = np.array(
+                (im_info['image_shape'][0], im_info['image_shape'][1], 1),
+                dtype=np.float32)
+            gt_bbox = label_info['gt_bbox']
+            gt_class = label_info['gt_class']
+            is_difficult = label_info['difficult']
+            outputs = (im, im_resize_info, im_id, im_shape, gt_bbox, gt_class,
+                       is_difficult)
+        else:
+            if im_info is None:
+                raise TypeError('Cannot do ArrangeFasterRCNN! ' +
+                                'Becasuse the im_info can not be None!')
+            im_resize_info = im_info['im_resize_info']
+            im_shape = np.array(
+                (im_info['image_shape'][0], im_info['image_shape'][1], 1),
+                dtype=np.float32)
+            outputs = (im, im_resize_info, im_shape)
+        return outputs
+
+
+class ArrangeMaskRCNN(DetTransform):
+    """获取MaskRCNN模型训练/验证/预测所需信息。
+
+    Args:
+        mode (str): 指定数据用于何种用途，取值范围为['train', 'eval', 'test', 'quant']。
+
+    Raises:
+        ValueError: mode的取值不在['train', 'eval', 'test', 'quant']之内。
+    """
+
+    def __init__(self, mode=None):
+        if mode not in ['train', 'eval', 'test', 'quant']:
+            raise ValueError(
+                "mode must be in ['train', 'eval', 'test', 'quant']!")
+        self.mode = mode
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当mode为'train'时，返回(im, im_resize_info, gt_bbox, gt_class, is_crowd, gt_masks)，分别对应
+                图像np.ndarray数据、图像相当对于原图的resize信息、真实标注框、真实标注框对应的类别、真实标注框内是否是一组对象、
+                真实分割区域；当mode为'eval'时，返回(im, im_resize_info, im_id, im_shape)，分别对应图像np.ndarray数据、
+                图像相当对于原图的resize信息、图像id、图像大小信息；当mode为'test'或'quant'时，返回(im, im_resize_info, im_shape)，
+                分别对应图像np.ndarray数据、图像相当对于原图的resize信息、图像大小信息。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+            ValueError: 数据长度不匹配。
+        """
+        im = permute(im, False)
+        if self.mode == 'train':
+            if im_info is None or label_info is None:
+                raise TypeError(
+                    'Cannot do ArrangeTrainMaskRCNN! ' +
+                    'Becasuse the im_info and label_info can not be None!')
+            if len(label_info['gt_bbox']) != len(label_info['gt_class']):
+                raise ValueError("gt num mismatch: bbox and class.")
+            im_resize_info = im_info['im_resize_info']
+            gt_bbox = label_info['gt_bbox']
+            gt_class = label_info['gt_class']
+            is_crowd = label_info['is_crowd']
+            assert 'gt_poly' in label_info
+            segms = label_info['gt_poly']
+            if len(segms) != 0:
+                assert len(segms) == is_crowd.shape[0]
+            gt_masks = []
+            valid = True
+            for i in range(len(segms)):
+                segm = segms[i]
+                gt_segm = []
+                if is_crowd[i]:
+                    gt_segm.append([[0, 0]])
+                else:
+                    for poly in segm:
+                        if len(poly) == 0:
+                            valid = False
+                            break
+                        gt_segm.append(np.array(poly).reshape(-1, 2))
+                if (not valid) or len(gt_segm) == 0:
+                    break
+                gt_masks.append(gt_segm)
+            outputs = (im, im_resize_info, gt_bbox, gt_class, is_crowd,
+                       gt_masks)
+        else:
+            if im_info is None:
+                raise TypeError('Cannot do ArrangeMaskRCNN! ' +
+                                'Becasuse the im_info can not be None!')
+            im_resize_info = im_info['im_resize_info']
+            im_shape = np.array(
+                (im_info['image_shape'][0], im_info['image_shape'][1], 1),
+                dtype=np.float32)
+            if self.mode == 'eval':
+                im_id = im_info['im_id']
+                outputs = (im, im_resize_info, im_id, im_shape)
+            else:
+                outputs = (im, im_resize_info, im_shape)
+        return outputs
+
+
+class ArrangeYOLOv3(DetTransform):
+    """获取YOLOv3模型训练/验证/预测所需信息。
+
+    Args:
+        mode (str): 指定数据用于何种用途，取值范围为['train', 'eval', 'test', 'quant']。
+
+    Raises:
+        ValueError: mode的取值不在['train', 'eval', 'test', 'quant']之内。
+    """
+
+    def __init__(self, mode=None):
+        if mode not in ['train', 'eval', 'test', 'quant']:
+            raise ValueError(
+                "mode must be in ['train', 'eval', 'test', 'quant']!")
+        self.mode = mode
+
+    def __call__(self, im, im_info=None, label_info=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (dict, 可选): 存储与图像相关的信息。
+            label_info (dict, 可选): 存储与标注框相关的信息。
+
+        Returns:
+            tuple: 当mode为'train'时，返回(im, gt_bbox, gt_class, gt_score, im_shape)，分别对应
+                图像np.ndarray数据、真实标注框、真实标注框对应的类别、真实标注框混合得分、图像大小信息；
+                当mode为'eval'时，返回(im, im_shape, im_id, gt_bbox, gt_class, difficult)，
+                分别对应图像np.ndarray数据、图像大小信息、图像id、真实标注框、真实标注框对应的类别、
+                真实标注框是否为难识别对象；当mode为'test'或'quant'时，返回(im, im_shape)，
+                分别对应图像np.ndarray数据、图像大小信息。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+            ValueError: 数据长度不匹配。
+        """
+        im = permute(im, False)
+        if self.mode == 'train':
+            if im_info is None or label_info is None:
+                raise TypeError(
+                    'Cannot do ArrangeYolov3! ' +
+                    'Becasuse the im_info and label_info can not be None!')
+            im_shape = im_info['image_shape']
+            if len(label_info['gt_bbox']) != len(label_info['gt_class']):
+                raise ValueError("gt num mismatch: bbox and class.")
+            if len(label_info['gt_bbox']) != len(label_info['gt_score']):
+                raise ValueError("gt num mismatch: bbox and score.")
+            gt_bbox = np.zeros((50, 4), dtype=im.dtype)
+            gt_class = np.zeros((50, ), dtype=np.int32)
+            gt_score = np.zeros((50, ), dtype=im.dtype)
+            gt_num = min(50, len(label_info['gt_bbox']))
+            if gt_num > 0:
+                label_info['gt_class'][:gt_num, 0] = label_info[
+                    'gt_class'][:gt_num, 0] - 1
+                if -1 not in label_info['gt_class']:
+                    gt_bbox[:gt_num, :] = label_info['gt_bbox'][:gt_num, :]
+                    gt_class[:gt_num] = label_info['gt_class'][:gt_num, 0]
+                    gt_score[:gt_num] = label_info['gt_score'][:gt_num, 0]
+            # parse [x1, y1, x2, y2] to [x, y, w, h]
+            gt_bbox[:, 2:4] = gt_bbox[:, 2:4] - gt_bbox[:, :2]
+            gt_bbox[:, :2] = gt_bbox[:, :2] + gt_bbox[:, 2:4] / 2.
+            outputs = (im, gt_bbox, gt_class, gt_score, im_shape)
+        elif self.mode == 'eval':
+            if im_info is None or label_info is None:
+                raise TypeError(
+                    'Cannot do ArrangeYolov3! ' +
+                    'Becasuse the im_info and label_info can not be None!')
+            im_shape = im_info['image_shape']
+            if len(label_info['gt_bbox']) != len(label_info['gt_class']):
+                raise ValueError("gt num mismatch: bbox and class.")
+            im_id = im_info['im_id']
+            gt_bbox = np.zeros((50, 4), dtype=im.dtype)
+            gt_class = np.zeros((50, ), dtype=np.int32)
+            difficult = np.zeros((50, ), dtype=np.int32)
+            gt_num = min(50, len(label_info['gt_bbox']))
+            if gt_num > 0:
+                label_info['gt_class'][:gt_num, 0] = label_info[
+                    'gt_class'][:gt_num, 0] - 1
+                gt_bbox[:gt_num, :] = label_info['gt_bbox'][:gt_num, :]
+                gt_class[:gt_num] = label_info['gt_class'][:gt_num, 0]
+                difficult[:gt_num] = label_info['difficult'][:gt_num, 0]
+            outputs = (im, im_shape, im_id, gt_bbox, gt_class, difficult)
+        else:
+            if im_info is None:
+                raise TypeError('Cannot do ArrangeYolov3! ' +
+                                'Becasuse the im_info can not be None!')
+            im_shape = im_info['image_shape']
+            outputs = (im, im_shape)
+        return outputs
+
+
+class ComposedRCNNTransforms(Compose):
+    """ RCNN模型(faster-rcnn/mask-rcnn)图像处理流程，具体如下，
+        训练阶段：
+        1. 随机以0.5的概率将图像水平翻转
+        2. 图像归一化
+        3. 图像按比例Resize，scale计算方式如下
+            scale = min_max_size[0] / short_size_of_image
+            if max_size_of_image * scale > min_max_size[1]:
+                scale = min_max_size[1] / max_size_of_image
+        4. 将3步骤的长宽进行padding，使得长宽为32的倍数
+        验证阶段：
+        1. 图像归一化
+        2. 图像按比例Resize，scale计算方式同上训练阶段
+        3. 将2步骤的长宽进行padding，使得长宽为32的倍数
+
+        Args:
+            mode(str): 图像处理流程所处阶段，训练/验证/预测，分别对应'train', 'eval', 'test'
+            min_max_size(list): 图像在缩放时，最小边和最大边的约束条件
+            mean(list): 图像均值
+            std(list): 图像方差
+    """
+
+    def __init__(self,
+                 mode,
+                 min_max_size=[800, 1333],
+                 mean=[0.485, 0.456, 0.406],
+                 std=[0.229, 0.224, 0.225]):
+        if mode == 'train':
+            # 训练时的transforms，包含数据增强
+            transforms = [
+                RandomHorizontalFlip(prob=0.5), Normalize(
+                    mean=mean, std=std), ResizeByShort(
+                        short_size=min_max_size[0], max_size=min_max_size[1]),
+                Padding(coarsest_stride=32)
+            ]
+        else:
+            # 验证/预测时的transforms
+            transforms = [
+                Normalize(
+                    mean=mean, std=std), ResizeByShort(
+                        short_size=min_max_size[0], max_size=min_max_size[1]),
+                Padding(coarsest_stride=32)
+            ]
+
+        super(ComposedRCNNTransforms, self).__init__(transforms)
+
+
+class ComposedYOLOv3Transforms(Compose):
+    """YOLOv3模型的图像预处理流程，具体如下，
+        训练阶段：
+        1. 在前mixup_epoch轮迭代中，使用MixupImage策略，见https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/det_transforms.html#mixupimage
+        2. 对图像进行随机扰动，包括亮度，对比度，饱和度和色调
+        3. 随机扩充图像，见https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/det_transforms.html#randomexpand
+        4. 随机裁剪图像
+        5. 将4步骤的输出图像Resize成shape参数的大小
+        6. 随机0.5的概率水平翻转图像
+        7. 图像归一化
+        验证/预测阶段：
+        1. 将图像Resize成shape参数大小
+        2. 图像归一化
+
+        Args:
+            mode(str): 图像处理流程所处阶段，训练/验证/预测，分别对应'train', 'eval', 'test'
+            shape(list): 输入模型中图像的大小，输入模型的图像会被Resize成此大小
+            mixup_epoch(int): 模型训练过程中，前mixup_epoch会使用mixup策略
+            mean(list): 图像均值
+            std(list): 图像方差
+    """
+
+    def __init__(self,
+                 mode,
+                 shape=[608, 608],
+                 mixup_epoch=250,
+                 mean=[0.485, 0.456, 0.406],
+                 std=[0.229, 0.224, 0.225]):
+        width = shape
+        if isinstance(shape, list):
+            if shape[0] != shape[1]:
+                raise Exception(
+                    "In YOLOv3 model, width and height should be equal")
+            width = shape[0]
+        if width % 32 != 0:
+            raise Exception(
+                "In YOLOv3 model, width and height should be multiple of 32, e.g 224、256、320...."
+            )
+
+        if mode == 'train':
+            # 训练时的transforms，包含数据增强
+            transforms = [
+                MixupImage(mixup_epoch=mixup_epoch), RandomDistort(),
+                RandomExpand(), RandomCrop(), Resize(
+                    target_size=width,
+                    interp='RANDOM'), RandomHorizontalFlip(), Normalize(
+                        mean=mean, std=std)
+            ]
+        else:
+            # 验证/预测时的transforms
+            transforms = [
+                Resize(
+                    target_size=width, interp='CUBIC'), Normalize(
+                        mean=mean, std=std)
+            ]
+        super(ComposedYOLOv3Transforms, self).__init__(transforms)

deploy/openvino/python/transforms/imgaug_support.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import copy
+
+
+def execute_imgaug(augmenter, im, bboxes=None, polygons=None,
+                   segment_map=None):
+    # 预处理，将bboxes, polygons转换成imgaug格式
+    import imgaug.augmentables.kps as kps
+    import imgaug.augmentables.bbs as bbs
+
+    aug_im = im.astype('uint8')
+    aug_im = augmenter.augment(image=aug_im).astype('float32')
+    return aug_im
+
+    # TODO imgaug的标注处理逻辑与paddlex已存的transform存在部分差异
+    # 目前仅支持对原图进行处理，因此只能使用pixlevel的imgaug增强操作
+    # 以下代码暂不会执行
+    aug_bboxes = None
+    if bboxes is not None:
+        aug_bboxes = list()
+        for i in range(len(bboxes)):
+            x1 = bboxes[i, 0]
+            y1 = bboxes[i, 1]
+            x2 = bboxes[i, 2]
+            y2 = bboxes[i, 3]
+            aug_bboxes.append(bbs.BoundingBox(x1, y1, x2, y2))
+
+    aug_points = None
+    if polygons is not None:
+        aug_points = list()
+        for i in range(len(polygons)):
+            num = len(polygons[i])
+            for j in range(num):
+                tmp = np.reshape(polygons[i][j], (-1, 2))
+                for k in range(len(tmp)):
+                    aug_points.append(kps.Keypoint(tmp[k, 0], tmp[k, 1]))
+
+    aug_segment_map = None
+    if segment_map is not None:
+        if len(segment_map.shape) == 2:
+            h, w = segment_map.shape
+            aug_segment_map = np.reshape(segment_map, (1, h, w, 1))
+        elif len(segment_map.shape) == 3:
+            h, w, c = segment_map.shape
+            aug_segment_map = np.reshape(segment_map, (1, h, w, c))
+        else:
+            raise Exception(
+                "Only support 2-dimensions for 3-dimensions for segment_map")
+
+    unnormalized_batch = augmenter.augment(
+        image=aug_im,
+        bounding_boxes=aug_bboxes,
+        keypoints=aug_points,
+        segmentation_maps=aug_segment_map,
+        return_batch=True)
+    aug_im = unnormalized_batch.images_aug[0]
+    aug_bboxes = unnormalized_batch.bounding_boxes_aug
+    aug_points = unnormalized_batch.keypoints_aug
+    aug_seg_map = unnormalized_batch.segmentation_maps_aug
+
+    aug_im = aug_im.astype('float32')
+
+    if aug_bboxes is not None:
+        converted_bboxes = list()
+        for i in range(len(aug_bboxes)):
+            converted_bboxes.append([
+                aug_bboxes[i].x1, aug_bboxes[i].y1, aug_bboxes[i].x2,
+                aug_bboxes[i].y2
+            ])
+        aug_bboxes = converted_bboxes
+
+    aug_polygons = None
+    if aug_points is not None:
+        aug_polygons = copy.deepcopy(polygons)
+        idx = 0
+        for i in range(len(aug_polygons)):
+            num = len(aug_polygons[i])
+            for j in range(num):
+                num_points = len(aug_polygons[i][j]) // 2
+                for k in range(num_points):
+                    aug_polygons[i][j][k * 2] = aug_points[idx].x
+                    aug_polygons[i][j][k * 2 + 1] = aug_points[idx].y
+                    idx += 1
+
+    result = [aug_im]
+    if aug_bboxes is not None:
+        result.append(np.array(aug_bboxes))
+    if aug_polygons is not None:
+        result.append(aug_polygons)
+    if aug_seg_map is not None:
+        n, h, w, c = aug_seg_map.shape
+        if len(segment_map.shape) == 2:
+            aug_seg_map = np.reshape(aug_seg_map, (h, w))
+        elif len(segment_map.shape) == 3:
+            aug_seg_map = np.reshape(aug_seg_map, (h, w, c))
+        result.append(aug_seg_map)
+    return result

deploy/openvino/python/transforms/ops.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import cv2
+import math
+import numpy as np
+from PIL import Image, ImageEnhance
+
+
+def normalize(im, mean, std):
+    im = im / 255.0
+    im -= mean
+    im /= std
+    return im
+
+
+def permute(im, to_bgr=False):
+    im = np.swapaxes(im, 1, 2)
+    im = np.swapaxes(im, 1, 0)
+    if to_bgr:
+        im = im[[2, 1, 0], :, :]
+    return im
+
+
+def resize_long(im, long_size=224, interpolation=cv2.INTER_LINEAR):
+    value = max(im.shape[0], im.shape[1])
+    scale = float(long_size) / float(value)
+    resized_width = int(round(im.shape[1] * scale))
+    resized_height = int(round(im.shape[0] * scale))
+
+    im = cv2.resize(
+        im, (resized_width, resized_height), interpolation=interpolation)
+    return im
+
+
+def resize(im, target_size=608, interp=cv2.INTER_LINEAR):
+    if isinstance(target_size, list) or isinstance(target_size, tuple):
+        w = target_size[0]
+        h = target_size[1]
+    else:
+        w = target_size
+        h = target_size
+    im = cv2.resize(im, (w, h), interpolation=interp)
+    return im
+
+
+def random_crop(im,
+                crop_size=224,
+                lower_scale=0.08,
+                lower_ratio=3. / 4,
+                upper_ratio=4. / 3):
+    scale = [lower_scale, 1.0]
+    ratio = [lower_ratio, upper_ratio]
+    aspect_ratio = math.sqrt(np.random.uniform(*ratio))
+    w = 1. * aspect_ratio
+    h = 1. / aspect_ratio
+    bound = min((float(im.shape[0]) / im.shape[1]) / (h**2),
+                (float(im.shape[1]) / im.shape[0]) / (w**2))
+    scale_max = min(scale[1], bound)
+    scale_min = min(scale[0], bound)
+    target_area = im.shape[0] * im.shape[1] * np.random.uniform(
+        scale_min, scale_max)
+    target_size = math.sqrt(target_area)
+    w = int(target_size * w)
+    h = int(target_size * h)
+    i = np.random.randint(0, im.shape[0] - h + 1)
+    j = np.random.randint(0, im.shape[1] - w + 1)
+    im = im[i:i + h, j:j + w, :]
+    im = cv2.resize(im, (crop_size, crop_size))
+    return im
+
+
+def center_crop(im, crop_size=224):
+    height, width = im.shape[:2]
+    w_start = (width - crop_size) // 2
+    h_start = (height - crop_size) // 2
+    w_end = w_start + crop_size
+    h_end = h_start + crop_size
+    im = im[h_start:h_end, w_start:w_end, :]
+    return im
+
+
+def horizontal_flip(im):
+    if len(im.shape) == 3:
+        im = im[:, ::-1, :]
+    elif len(im.shape) == 2:
+        im = im[:, ::-1]
+    return im
+
+
+def vertical_flip(im):
+    if len(im.shape) == 3:
+        im = im[::-1, :, :]
+    elif len(im.shape) == 2:
+        im = im[::-1, :]
+    return im
+
+
+def bgr2rgb(im):
+    return im[:, :, ::-1]
+
+
+def hue(im, hue_lower, hue_upper):
+    delta = np.random.uniform(hue_lower, hue_upper)
+    u = np.cos(delta * np.pi)
+    w = np.sin(delta * np.pi)
+    bt = np.array([[1.0, 0.0, 0.0], [0.0, u, -w], [0.0, w, u]])
+    tyiq = np.array([[0.299, 0.587, 0.114], [0.596, -0.274, -0.321],
+                     [0.211, -0.523, 0.311]])
+    ityiq = np.array([[1.0, 0.956, 0.621], [1.0, -0.272, -0.647],
+                      [1.0, -1.107, 1.705]])
+    t = np.dot(np.dot(ityiq, bt), tyiq).T
+    im = np.dot(im, t)
+    return im
+
+
+def saturation(im, saturation_lower, saturation_upper):
+    delta = np.random.uniform(saturation_lower, saturation_upper)
+    gray = im * np.array([[[0.299, 0.587, 0.114]]], dtype=np.float32)
+    gray = gray.sum(axis=2, keepdims=True)
+    gray *= (1.0 - delta)
+    im *= delta
+    im += gray
+    return im
+
+
+def contrast(im, contrast_lower, contrast_upper):
+    delta = np.random.uniform(contrast_lower, contrast_upper)
+    im *= delta
+    return im
+
+
+def brightness(im, brightness_lower, brightness_upper):
+    delta = np.random.uniform(brightness_lower, brightness_upper)
+    im += delta
+    return im
+
+def rotate(im, rotate_lower, rotate_upper):
+    rotate_delta = np.random.uniform(rotate_lower, rotate_upper)
+    im = im.rotate(int(rotate_delta))
+    return im
+
+
+def resize_padding(im, max_side_len=2400):
+    '''
+    resize image to a size multiple of 32 which is required by the network
+    :param im: the resized image
+    :param max_side_len: limit of max image size to avoid out of memory in gpu
+    :return: the resized image and the resize ratio
+    '''
+    h, w, _ = im.shape
+
+    resize_w = w
+    resize_h = h
+
+    # limit the max side
+    if max(resize_h, resize_w) > max_side_len:
+        ratio = float(
+            max_side_len) / resize_h if resize_h > resize_w else float(
+                max_side_len) / resize_w
+    else:
+        ratio = 1.
+    resize_h = int(resize_h * ratio)
+    resize_w = int(resize_w * ratio)
+
+    resize_h = resize_h if resize_h % 32 == 0 else (resize_h // 32 - 1) * 32
+    resize_w = resize_w if resize_w % 32 == 0 else (resize_w // 32 - 1) * 32
+    resize_h = max(32, resize_h)
+    resize_w = max(32, resize_w)
+    im = cv2.resize(im, (int(resize_w), int(resize_h)))
+    #im = cv2.resize(im, (512, 512))
+    ratio_h = resize_h / float(h)
+    ratio_w = resize_w / float(w)
+    _ratio = np.array([ratio_h, ratio_w]).reshape(-1, 2)
+    return im, _ratio

deploy/openvino/python/transforms/seg_transforms.py

+# coding: utf8
+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .ops import *
+from .imgaug_support import execute_imgaug
+import random
+import os.path as osp
+import numpy as np
+from PIL import Image
+import cv2
+from collections import OrderedDict
+import utils.logging as logging
+
+
+class SegTransform:
+    """ 分割transform基类
+    """
+
+    def __init__(self):
+        pass
+
+
+class Compose(SegTransform):
+    """根据数据预处理/增强算子对输入数据进行操作。
+       所有操作的输入图像流形状均是[H, W, C]，其中H为图像高，W为图像宽，C为图像通道数。
+
+    Args:
+        transforms (list): 数据预处理/增强算子。
+
+    Raises:
+        TypeError: transforms不是list对象
+        ValueError: transforms元素个数小于1。
+
+    """
+
+    def __init__(self, transforms):
+        if not isinstance(transforms, list):
+            raise TypeError('The transforms must be a list!')
+        if len(transforms) < 1:
+            raise ValueError('The length of transforms ' + \
+                            'must be equal or larger than 1!')
+        self.transforms = transforms
+        self.to_rgb = False
+        # 检查transforms里面的操作，目前支持PaddleX定义的或者是imgaug操作
+        for op in self.transforms:
+            if not isinstance(op, SegTransform):
+                import imgaug.augmenters as iaa
+                if not isinstance(op, iaa.Augmenter):
+                    raise Exception(
+                        "Elements in transforms should be defined in 'paddlex.seg.transforms' or class of imgaug.augmenters.Augmenter, see docs here: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/"
+                    )
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (str/np.ndarray): 图像路径/图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (str/np.ndarray): 标注图像路径/标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 根据网络所需字段所组成的tuple；字段由transforms中的最后一个数据预处理操作决定。
+        """
+
+        if im_info is None:
+            im_info = list()
+        if isinstance(im, np.ndarray):
+            if len(im.shape) != 3:
+                raise Exception(
+                    "im should be 3-dimensions, but now is {}-dimensions".
+                    format(len(im.shape)))
+        else:
+            try:
+                im = cv2.imread(im).astype('float32')
+            except:
+                raise ValueError('Can\'t read The image file {}!'.format(im))
+        if self.to_rgb:
+            im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
+        if label is not None:
+            if not isinstance(label, np.ndarray):
+                label = np.asarray(Image.open(label))
+        for op in self.transforms:
+            if isinstance(op, SegTransform):
+                outputs = op(im, im_info, label)
+                im = outputs[0]
+                if len(outputs) >= 2:
+                    im_info = outputs[1]
+                if len(outputs) == 3:
+                    label = outputs[2]
+            else:
+                im = execute_imgaug(op, im)
+                if label is not None:
+                    outputs = (im, im_info, label)
+                else:
+                    outputs = (im, im_info)
+        return outputs
+
+    def add_augmenters(self, augmenters):
+        if not isinstance(augmenters, list):
+            raise Exception(
+                "augmenters should be list type in func add_augmenters()")
+        transform_names = [type(x).__name__ for x in self.transforms]
+        for aug in augmenters:
+            if type(aug).__name__ in transform_names:
+                logging.error("{} is already in ComposedTransforms, need to remove it from add_augmenters().".format(type(aug).__name__))
+        self.transforms = augmenters + self.transforms
+
+
+class RandomHorizontalFlip(SegTransform):
+    """以一定的概率对图像进行水平翻转。当存在标注图像时，则同步进行翻转。
+
+    Args:
+        prob (float): 随机水平翻转的概率。默认值为0.5。
+
+    """
+
+    def __init__(self, prob=0.5):
+        self.prob = prob
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+        if random.random() < self.prob:
+            im = horizontal_flip(im)
+            if label is not None:
+                label = horizontal_flip(label)
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class RandomVerticalFlip(SegTransform):
+    """以一定的概率对图像进行垂直翻转。当存在标注图像时，则同步进行翻转。
+
+    Args:
+        prob (float): 随机垂直翻转的概率。默认值为0.1。
+    """
+
+    def __init__(self, prob=0.1):
+        self.prob = prob
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+        if random.random() < self.prob:
+            im = vertical_flip(im)
+            if label is not None:
+                label = vertical_flip(label)
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class Resize(SegTransform):
+    """调整图像大小（resize），当存在标注图像时，则同步进行处理。
+
+    - 当目标大小（target_size）类型为int时，根据插值方式，
+      将图像resize为[target_size, target_size]。
+    - 当目标大小（target_size）类型为list或tuple时，根据插值方式，
+      将图像resize为target_size, target_size的输入应为[w, h]或（w, h）。
+
+    Args:
+        target_size (int|list|tuple): 目标大小。
+        interp (str): resize的插值方式，与opencv的插值方式对应，
+            可选的值为['NEAREST', 'LINEAR', 'CUBIC', 'AREA', 'LANCZOS4']，默认为"LINEAR"。
+
+    Raises:
+        TypeError: target_size不是int/list/tuple。
+        ValueError:  target_size为list/tuple时元素个数不等于2。
+        AssertionError: interp的取值不在['NEAREST', 'LINEAR', 'CUBIC', 'AREA', 'LANCZOS4']之内。
+    """
+
+    # The interpolation mode
+    interp_dict = {
+        'NEAREST': cv2.INTER_NEAREST,
+        'LINEAR': cv2.INTER_LINEAR,
+        'CUBIC': cv2.INTER_CUBIC,
+        'AREA': cv2.INTER_AREA,
+        'LANCZOS4': cv2.INTER_LANCZOS4
+    }
+
+    def __init__(self, target_size, interp='LINEAR'):
+        self.interp = interp
+        assert interp in self.interp_dict, "interp should be one of {}".format(
+            interp_dict.keys())
+        if isinstance(target_size, list) or isinstance(target_size, tuple):
+            if len(target_size) != 2:
+                raise ValueError(
+                    'when target is list or tuple, it should include 2 elements, but it is {}'
+                    .format(target_size))
+        elif not isinstance(target_size, int):
+            raise TypeError(
+                "Type of target_size is invalid. Must be Integer or List or tuple, now is {}"
+                .format(type(target_size)))
+
+        self.target_size = target_size
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+                其中，im_info跟新字段为：
+                    -shape_before_resize (tuple): 保存resize之前图像的形状(h, w）。
+
+        Raises:
+            ZeroDivisionError: im的短边为0。
+            TypeError: im不是np.ndarray数据。
+            ValueError: im不是3维nd.ndarray。
+        """
+        if im_info is None:
+            im_info = OrderedDict()
+        im_info.append(('resize', im.shape[:2]))
+
+        if not isinstance(im, np.ndarray):
+            raise TypeError("ResizeImage: image type is not np.ndarray.")
+        if len(im.shape) != 3:
+            raise ValueError('ResizeImage: image is not 3-dimensional.')
+        im_shape = im.shape
+        im_size_min = np.min(im_shape[0:2])
+        im_size_max = np.max(im_shape[0:2])
+        if float(im_size_min) == 0:
+            raise ZeroDivisionError('ResizeImage: min size of image is 0')
+
+        if isinstance(self.target_size, int):
+            resize_w = self.target_size
+            resize_h = self.target_size
+        else:
+            resize_w = self.target_size[0]
+            resize_h = self.target_size[1]
+        im_scale_x = float(resize_w) / float(im_shape[1])
+        im_scale_y = float(resize_h) / float(im_shape[0])
+
+        im = cv2.resize(
+            im,
+            None,
+            None,
+            fx=im_scale_x,
+            fy=im_scale_y,
+            interpolation=self.interp_dict[self.interp])
+        if label is not None:
+            label = cv2.resize(
+                label,
+                None,
+                None,
+                fx=im_scale_x,
+                fy=im_scale_y,
+                interpolation=self.interp_dict['NEAREST'])
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class ResizeByLong(SegTransform):
+    """对图像长边resize到固定值，短边按比例进行缩放。当存在标注图像时，则同步进行处理。
+
+    Args:
+        long_size (int): resize后图像的长边大小。
+    """
+
+    def __init__(self, long_size):
+        self.long_size = long_size
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+                其中，im_info新增字段为：
+                    -shape_before_resize (tuple): 保存resize之前图像的形状(h, w)。
+        """
+        if im_info is None:
+            im_info = OrderedDict()
+
+        im_info.append(('resize', im.shape[:2]))
+        im = resize_long(im, self.long_size)
+        if label is not None:
+            label = resize_long(label, self.long_size, cv2.INTER_NEAREST)
+
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class ResizeByShort(SegTransform):
+    """根据图像的短边调整图像大小（resize）。
+
+    1. 获取图像的长边和短边长度。
+    2. 根据短边与short_size的比例，计算长边的目标长度，
+       此时高、宽的resize比例为short_size/原图短边长度。
+    3. 如果max_size>0，调整resize比例：
+       如果长边的目标长度>max_size，则高、宽的resize比例为max_size/原图长边长度。
+    4. 根据调整大小的比例对图像进行resize。
+
+    Args:
+        target_size (int): 短边目标长度。默认为800。
+        max_size (int): 长边目标长度的最大限制。默认为1333。
+
+     Raises:
+        TypeError: 形参数据类型不满足需求。
+    """
+
+    def __init__(self, short_size=800, max_size=1333):
+        self.max_size = int(max_size)
+        if not isinstance(short_size, int):
+            raise TypeError(
+                "Type of short_size is invalid. Must be Integer, now is {}".
+                format(type(short_size)))
+        self.short_size = short_size
+        if not (isinstance(self.max_size, int)):
+            raise TypeError("max_size: input type is invalid.")
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (numnp.ndarraypy): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                   当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                   存储与图像相关信息的字典和标注图像np.ndarray数据。
+                   其中，im_info更新字段为：
+                       -shape_before_resize (tuple): 保存resize之前图像的形状(h, w)。
+
+        Raises:
+            TypeError: 形参数据类型不满足需求。
+            ValueError: 数据长度不匹配。
+        """
+        if im_info is None:
+            im_info = OrderedDict()
+        if not isinstance(im, np.ndarray):
+            raise TypeError("ResizeByShort: image type is not numpy.")
+        if len(im.shape) != 3:
+            raise ValueError('ResizeByShort: image is not 3-dimensional.')
+        im_info.append(('resize', im.shape[:2]))
+        im_short_size = min(im.shape[0], im.shape[1])
+        im_long_size = max(im.shape[0], im.shape[1])
+        scale = float(self.short_size) / im_short_size
+        if self.max_size > 0 and np.round(scale *
+                                          im_long_size) > self.max_size:
+            scale = float(self.max_size) / float(im_long_size)
+        resized_width = int(round(im.shape[1] * scale))
+        resized_height = int(round(im.shape[0] * scale))
+        im = cv2.resize(
+            im, (resized_width, resized_height),
+            interpolation=cv2.INTER_NEAREST)
+        if label is not None:
+            im = cv2.resize(
+                label, (resized_width, resized_height),
+                interpolation=cv2.INTER_NEAREST)
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class ResizeRangeScaling(SegTransform):
+    """对图像长边随机resize到指定范围内，短边按比例进行缩放。当存在标注图像时，则同步进行处理。
+
+    Args:
+        min_value (int): 图像长边resize后的最小值。默认值400。
+        max_value (int): 图像长边resize后的最大值。默认值600。
+
+    Raises:
+        ValueError: min_value大于max_value
+    """
+
+    def __init__(self, min_value=400, max_value=600):
+        if min_value > max_value:
+            raise ValueError('min_value must be less than max_value, '
+                             'but they are {} and {}.'.format(min_value,
+                                                              max_value))
+        self.min_value = min_value
+        self.max_value = max_value
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+        if self.min_value == self.max_value:
+            random_size = self.max_value
+        else:
+            random_size = int(
+                np.random.uniform(self.min_value, self.max_value) + 0.5)
+        im = resize_long(im, random_size, cv2.INTER_LINEAR)
+        if label is not None:
+            label = resize_long(label, random_size, cv2.INTER_NEAREST)
+
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class ResizeStepScaling(SegTransform):
+    """对图像按照某一个比例resize，这个比例以scale_step_size为步长
+    在[min_scale_factor, max_scale_factor]随机变动。当存在标注图像时，则同步进行处理。
+
+    Args:
+        min_scale_factor（float), resize最小尺度。默认值0.75。
+        max_scale_factor (float), resize最大尺度。默认值1.25。
+        scale_step_size (float), resize尺度范围间隔。默认值0.25。
+
+    Raises:
+        ValueError: min_scale_factor大于max_scale_factor
+    """
+
+    def __init__(self,
+                 min_scale_factor=0.75,
+                 max_scale_factor=1.25,
+                 scale_step_size=0.25):
+        if min_scale_factor > max_scale_factor:
+            raise ValueError(
+                'min_scale_factor must be less than max_scale_factor, '
+                'but they are {} and {}.'.format(min_scale_factor,
+                                                 max_scale_factor))
+        self.min_scale_factor = min_scale_factor
+        self.max_scale_factor = max_scale_factor
+        self.scale_step_size = scale_step_size
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+        if self.min_scale_factor == self.max_scale_factor:
+            scale_factor = self.min_scale_factor
+
+        elif self.scale_step_size == 0:
+            scale_factor = np.random.uniform(self.min_scale_factor,
+                                             self.max_scale_factor)
+
+        else:
+            num_steps = int((self.max_scale_factor - self.min_scale_factor) /
+                            self.scale_step_size + 1)
+            scale_factors = np.linspace(self.min_scale_factor,
+                                        self.max_scale_factor,
+                                        num_steps).tolist()
+            np.random.shuffle(scale_factors)
+            scale_factor = scale_factors[0]
+
+        im = cv2.resize(
+            im, (0, 0),
+            fx=scale_factor,
+            fy=scale_factor,
+            interpolation=cv2.INTER_LINEAR)
+        if label is not None:
+            label = cv2.resize(
+                label, (0, 0),
+                fx=scale_factor,
+                fy=scale_factor,
+                interpolation=cv2.INTER_NEAREST)
+
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class Normalize(SegTransform):
+    """对图像进行标准化。
+    1.尺度缩放到 [0,1]。
+    2.对图像进行减均值除以标准差操作。
+
+    Args:
+        mean (list): 图像数据集的均值。默认值[0.5, 0.5, 0.5]。
+        std (list): 图像数据集的标准差。默认值[0.5, 0.5, 0.5]。
+
+    Raises:
+        ValueError: mean或std不是list对象。std包含0。
+    """
+
+    def __init__(self, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]):
+        self.mean = mean
+        self.std = std
+        if not (isinstance(self.mean, list) and isinstance(self.std, list)):
+            raise ValueError("{}: input type is invalid.".format(self))
+        from functools import reduce
+        if reduce(lambda x, y: x * y, self.std) == 0:
+            raise ValueError('{}: std is invalid!'.format(self))
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+         Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+
+        mean = np.array(self.mean)[np.newaxis, np.newaxis, :]
+        std = np.array(self.std)[np.newaxis, np.newaxis, :]
+        im = normalize(im, mean, std)
+
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class Padding(SegTransform):
+    """对图像或标注图像进行padding，padding方向为右和下。
+    根据提供的值对图像或标注图像进行padding操作。
+
+    Args:
+        target_size (int|list|tuple): padding后图像的大小。
+        im_padding_value (list): 图像padding的值。默认为[127.5, 127.5, 127.5]。
+        label_padding_value (int): 标注图像padding的值。默认值为255。
+
+    Raises:
+        TypeError: target_size不是int|list|tuple。
+        ValueError:  target_size为list|tuple时元素个数不等于2。
+    """
+
+    def __init__(self,
+                 target_size,
+                 im_padding_value=[127.5, 127.5, 127.5],
+                 label_padding_value=255):
+        if isinstance(target_size, list) or isinstance(target_size, tuple):
+            if len(target_size) != 2:
+                raise ValueError(
+                    'when target is list or tuple, it should include 2 elements, but it is {}'
+                    .format(target_size))
+        elif not isinstance(target_size, int):
+            raise TypeError(
+                "Type of target_size is invalid. Must be Integer or List or tuple, now is {}"
+                .format(type(target_size)))
+        self.target_size = target_size
+        self.im_padding_value = im_padding_value
+        self.label_padding_value = label_padding_value
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+                其中，im_info新增字段为：
+                    -shape_before_padding (tuple): 保存padding之前图像的形状(h, w）。
+
+        Raises:
+            ValueError: 输入图像im或label的形状大于目标值
+        """
+        if im_info is None:
+            im_info = OrderedDict()
+        im_info.append(('padding', im.shape[:2]))
+
+        im_height, im_width = im.shape[0], im.shape[1]
+        if isinstance(self.target_size, int):
+            target_height = self.target_size
+            target_width = self.target_size
+        else:
+            target_height = self.target_size[1]
+            target_width = self.target_size[0]
+        pad_height = target_height - im_height
+        pad_width = target_width - im_width
+        if pad_height < 0 or pad_width < 0:
+            raise ValueError(
+                'the size of image should be less than target_size, but the size of image ({}, {}), is larger than target_size ({}, {})'
+                .format(im_width, im_height, target_width, target_height))
+        else:
+            im = cv2.copyMakeBorder(
+                im,
+                0,
+                pad_height,
+                0,
+                pad_width,
+                cv2.BORDER_CONSTANT,
+                value=self.im_padding_value)
+            if label is not None:
+                label = cv2.copyMakeBorder(
+                    label,
+                    0,
+                    pad_height,
+                    0,
+                    pad_width,
+                    cv2.BORDER_CONSTANT,
+                    value=self.label_padding_value)
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class RandomPaddingCrop(SegTransform):
+    """对图像和标注图进行随机裁剪，当所需要的裁剪尺寸大于原图时，则进行padding操作。
+
+    Args:
+        crop_size (int|list|tuple): 裁剪图像大小。默认为512。
+        im_padding_value (list): 图像padding的值。默认为[127.5, 127.5, 127.5]。
+        label_padding_value (int): 标注图像padding的值。默认值为255。
+
+    Raises:
+        TypeError: crop_size不是int/list/tuple。
+        ValueError:  target_size为list/tuple时元素个数不等于2。
+    """
+
+    def __init__(self,
+                 crop_size=512,
+                 im_padding_value=[127.5, 127.5, 127.5],
+                 label_padding_value=255):
+        if isinstance(crop_size, list) or isinstance(crop_size, tuple):
+            if len(crop_size) != 2:
+                raise ValueError(
+                    'when crop_size is list or tuple, it should include 2 elements, but it is {}'
+                    .format(crop_size))
+        elif not isinstance(crop_size, int):
+            raise TypeError(
+                "Type of crop_size is invalid. Must be Integer or List or tuple, now is {}"
+                .format(type(crop_size)))
+        self.crop_size = crop_size
+        self.im_padding_value = im_padding_value
+        self.label_padding_value = label_padding_value
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+         Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+        if isinstance(self.crop_size, int):
+            crop_width = self.crop_size
+            crop_height = self.crop_size
+        else:
+            crop_width = self.crop_size[0]
+            crop_height = self.crop_size[1]
+
+        img_height = im.shape[0]
+        img_width = im.shape[1]
+
+        if img_height == crop_height and img_width == crop_width:
+            if label is None:
+                return (im, im_info)
+            else:
+                return (im, im_info, label)
+        else:
+            pad_height = max(crop_height - img_height, 0)
+            pad_width = max(crop_width - img_width, 0)
+            if (pad_height > 0 or pad_width > 0):
+                im = cv2.copyMakeBorder(
+                    im,
+                    0,
+                    pad_height,
+                    0,
+                    pad_width,
+                    cv2.BORDER_CONSTANT,
+                    value=self.im_padding_value)
+                if label is not None:
+                    label = cv2.copyMakeBorder(
+                        label,
+                        0,
+                        pad_height,
+                        0,
+                        pad_width,
+                        cv2.BORDER_CONSTANT,
+                        value=self.label_padding_value)
+                img_height = im.shape[0]
+                img_width = im.shape[1]
+
+            if crop_height > 0 and crop_width > 0:
+                h_off = np.random.randint(img_height - crop_height + 1)
+                w_off = np.random.randint(img_width - crop_width + 1)
+
+                im = im[h_off:(crop_height + h_off), w_off:(w_off + crop_width
+                                                            ), :]
+                if label is not None:
+                    label = label[h_off:(crop_height + h_off), w_off:(
+                        w_off + crop_width)]
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class RandomBlur(SegTransform):
+    """以一定的概率对图像进行高斯模糊。
+
+    Args：
+        prob (float): 图像模糊概率。默认为0.1。
+    """
+
+    def __init__(self, prob=0.1):
+        self.prob = prob
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+        if self.prob <= 0:
+            n = 0
+        elif self.prob >= 1:
+            n = 1
+        else:
+            n = int(1.0 / self.prob)
+        if n > 0:
+            if np.random.randint(0, n) == 0:
+                radius = np.random.randint(3, 10)
+                if radius % 2 != 1:
+                    radius = radius + 1
+                if radius > 9:
+                    radius = 9
+                im = cv2.GaussianBlur(im, (radius, radius), 0, 0)
+
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class RandomRotate(SegTransform):
+    """对图像进行随机旋转, 模型训练时的数据增强操作。
+    在旋转区间[-rotate_range, rotate_range]内，对图像进行随机旋转，当存在标注图像时，同步进行，
+    并对旋转后的图像和标注图像进行相应的padding。
+
+    Args:
+        rotate_range (float): 最大旋转角度。默认为15度。
+        im_padding_value (list): 图像padding的值。默认为[127.5, 127.5, 127.5]。
+        label_padding_value (int): 标注图像padding的值。默认为255。
+
+    """
+
+    def __init__(self,
+                 rotate_range=15,
+                 im_padding_value=[127.5, 127.5, 127.5],
+                 label_padding_value=255):
+        self.rotate_range = rotate_range
+        self.im_padding_value = im_padding_value
+        self.label_padding_value = label_padding_value
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+        if self.rotate_range > 0:
+            (h, w) = im.shape[:2]
+            do_rotation = np.random.uniform(-self.rotate_range,
+                                            self.rotate_range)
+            pc = (w // 2, h // 2)
+            r = cv2.getRotationMatrix2D(pc, do_rotation, 1.0)
+            cos = np.abs(r[0, 0])
+            sin = np.abs(r[0, 1])
+
+            nw = int((h * sin) + (w * cos))
+            nh = int((h * cos) + (w * sin))
+
+            (cx, cy) = pc
+            r[0, 2] += (nw / 2) - cx
+            r[1, 2] += (nh / 2) - cy
+            dsize = (nw, nh)
+            im = cv2.warpAffine(
+                im,
+                r,
+                dsize=dsize,
+                flags=cv2.INTER_LINEAR,
+                borderMode=cv2.BORDER_CONSTANT,
+                borderValue=self.im_padding_value)
+            label = cv2.warpAffine(
+                label,
+                r,
+                dsize=dsize,
+                flags=cv2.INTER_NEAREST,
+                borderMode=cv2.BORDER_CONSTANT,
+                borderValue=self.label_padding_value)
+
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class RandomScaleAspect(SegTransform):
+    """裁剪并resize回原始尺寸的图像和标注图像。
+    按照一定的面积比和宽高比对图像进行裁剪，并reszie回原始图像的图像，当存在标注图时，同步进行。
+
+    Args：
+        min_scale (float)：裁取图像占原始图像的面积比，取值[0，1]，为0时则返回原图。默认为0.5。
+        aspect_ratio (float): 裁取图像的宽高比范围，非负值，为0时返回原图。默认为0.33。
+    """
+
+    def __init__(self, min_scale=0.5, aspect_ratio=0.33):
+        self.min_scale = min_scale
+        self.aspect_ratio = aspect_ratio
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+        if self.min_scale != 0 and self.aspect_ratio != 0:
+            img_height = im.shape[0]
+            img_width = im.shape[1]
+            for i in range(0, 10):
+                area = img_height * img_width
+                target_area = area * np.random.uniform(self.min_scale, 1.0)
+                aspectRatio = np.random.uniform(self.aspect_ratio,
+                                                1.0 / self.aspect_ratio)
+
+                dw = int(np.sqrt(target_area * 1.0 * aspectRatio))
+                dh = int(np.sqrt(target_area * 1.0 / aspectRatio))
+                if (np.random.randint(10) < 5):
+                    tmp = dw
+                    dw = dh
+                    dh = tmp
+
+                if (dh < img_height and dw < img_width):
+                    h1 = np.random.randint(0, img_height - dh)
+                    w1 = np.random.randint(0, img_width - dw)
+
+                    im = im[h1:(h1 + dh), w1:(w1 + dw), :]
+                    label = label[h1:(h1 + dh), w1:(w1 + dw)]
+                    im = cv2.resize(
+                        im, (img_width, img_height),
+                        interpolation=cv2.INTER_LINEAR)
+                    label = cv2.resize(
+                        label, (img_width, img_height),
+                        interpolation=cv2.INTER_NEAREST)
+                    break
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class RandomDistort(SegTransform):
+    """对图像进行随机失真。
+
+    1. 对变换的操作顺序进行随机化操作。
+    2. 按照1中的顺序以一定的概率对图像进行随机像素内容变换。
+
+    Args:
+        brightness_range (float): 明亮度因子的范围。默认为0.5。
+        brightness_prob (float): 随机调整明亮度的概率。默认为0.5。
+        contrast_range (float): 对比度因子的范围。默认为0.5。
+        contrast_prob (float): 随机调整对比度的概率。默认为0.5。
+        saturation_range (float): 饱和度因子的范围。默认为0.5。
+        saturation_prob (float): 随机调整饱和度的概率。默认为0.5。
+        hue_range (int): 色调因子的范围。默认为18。
+        hue_prob (float): 随机调整色调的概率。默认为0.5。
+    """
+
+    def __init__(self,
+                 brightness_range=0.5,
+                 brightness_prob=0.5,
+                 contrast_range=0.5,
+                 contrast_prob=0.5,
+                 saturation_range=0.5,
+                 saturation_prob=0.5,
+                 hue_range=18,
+                 hue_prob=0.5):
+        self.brightness_range = brightness_range
+        self.brightness_prob = brightness_prob
+        self.contrast_range = contrast_range
+        self.contrast_prob = contrast_prob
+        self.saturation_range = saturation_range
+        self.saturation_prob = saturation_prob
+        self.hue_range = hue_range
+        self.hue_prob = hue_prob
+
+    def __call__(self, im, im_info=None, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当label为空时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当label不为空时，返回的tuple为(im, im_info, label)，分别对应图像np.ndarray数据、
+                存储与图像相关信息的字典和标注图像np.ndarray数据。
+        """
+        brightness_lower = 1 - self.brightness_range
+        brightness_upper = 1 + self.brightness_range
+        contrast_lower = 1 - self.contrast_range
+        contrast_upper = 1 + self.contrast_range
+        saturation_lower = 1 - self.saturation_range
+        saturation_upper = 1 + self.saturation_range
+        hue_lower = -self.hue_range
+        hue_upper = self.hue_range
+        ops = [brightness, contrast, saturation, hue]
+        random.shuffle(ops)
+        params_dict = {
+            'brightness': {
+                'brightness_lower': brightness_lower,
+                'brightness_upper': brightness_upper
+            },
+            'contrast': {
+                'contrast_lower': contrast_lower,
+                'contrast_upper': contrast_upper
+            },
+            'saturation': {
+                'saturation_lower': saturation_lower,
+                'saturation_upper': saturation_upper
+            },
+            'hue': {
+                'hue_lower': hue_lower,
+                'hue_upper': hue_upper
+            }
+        }
+        prob_dict = {
+            'brightness': self.brightness_prob,
+            'contrast': self.contrast_prob,
+            'saturation': self.saturation_prob,
+            'hue': self.hue_prob
+        }
+        for id in range(4):
+            params = params_dict[ops[id].__name__]
+            prob = prob_dict[ops[id].__name__]
+            params['im'] = im
+            if np.random.uniform(0, 1) < prob:
+                im = ops[id](**params)
+        if label is None:
+            return (im, im_info)
+        else:
+            return (im, im_info, label)
+
+
+class ArrangeSegmenter(SegTransform):
+    """获取训练/验证/预测所需的信息。
+
+    Args:
+        mode (str): 指定数据用于何种用途，取值范围为['train', 'eval', 'test', 'quant']。
+
+    Raises:
+        ValueError: mode的取值不在['train', 'eval', 'test', 'quant']之内
+    """
+
+    def __init__(self, mode):
+        if mode not in ['train', 'eval', 'test', 'quant']:
+            raise ValueError(
+                "mode should be defined as one of ['train', 'eval', 'test', 'quant']!"
+            )
+        self.mode = mode
+
+    def __call__(self, im, im_info, label=None):
+        """
+        Args:
+            im (np.ndarray): 图像np.ndarray数据。
+            im_info (list): 存储图像reisze或padding前的shape信息，如
+                [('resize', [200, 300]), ('padding', [400, 600])]表示
+                图像在过resize前shape为(200, 300)， 过padding前shape为
+                (400, 600)
+            label (np.ndarray): 标注图像np.ndarray数据。
+
+        Returns:
+            tuple: 当mode为'train'或'eval'时，返回的tuple为(im, label)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；
+                当mode为'test'时，返回的tuple为(im, im_info)，分别对应图像np.ndarray数据、存储与图像相关信息的字典；当mode为
+                'quant'时，返回的tuple为(im,)，为图像np.ndarray数据。
+        """
+        im = permute(im, False)
+        if self.mode == 'train' or self.mode == 'eval':
+            label = label[np.newaxis, :, :]
+            return (im, label)
+        elif self.mode == 'test':
+            return (im, im_info)
+        else:
+            return (im, )
+
+
+class ComposedSegTransforms(Compose):
+    """ 语义分割模型(UNet/DeepLabv3p)的图像处理流程，具体如下
+        训练阶段：
+        1. 随机对图像以0.5的概率水平翻转
+        2. 按不同的比例随机Resize原图
+        3. 从原图中随机crop出大小为train_crop_size大小的子图，如若crop出来的图小于train_crop_size，则会将图padding到对应大小
+        4. 图像归一化
+        预测阶段：
+        1. 图像归一化
+
+        Args:
+            mode(str): 图像处理所处阶段，训练/验证/预测，分别对应'train', 'eval', 'test'
+            train_crop_size(list): 模型训练阶段，随机从原图crop的大小
+            mean(list): 图像均值
+            std(list): 图像方差
+    """
+
+    def __init__(self,
+                 mode,
+                 train_crop_size=[769, 769],
+                 mean=[0.5, 0.5, 0.5],
+                 std=[0.5, 0.5, 0.5]):
+        if mode == 'train':
+            # 训练时的transforms，包含数据增强
+            transforms = [
+                RandomHorizontalFlip(prob=0.5), ResizeStepScaling(),
+                RandomPaddingCrop(crop_size=train_crop_size), Normalize(
+                    mean=mean, std=std)
+            ]
+        else:
+            # 验证/预测时的transforms
+            transforms = [Normalize(mean=mean, std=std)]
+
+        super(ComposedSegTransforms, self).__init__(transforms)

deploy/openvino/python/utils/__init__.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+

deploy/openvino/python/utils/logging.py

+# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+import os
+import sys
+import colorama
+from colorama import init
+
+init(autoreset=True)
+levels = {0: 'ERROR', 1: 'WARNING', 2: 'INFO', 3: 'DEBUG'}
+log_level = 2
+
+def log(level=2, message="", use_color=False):
+    current_time = time.time()
+    time_array = time.localtime(current_time)
+    current_time = time.strftime("%Y-%m-%d %H:%M:%S", time_array)
+    if log_level >= level:
+        if use_color:
+            print("\033[1;31;40m{} [{}]\t{}\033[0m".format(current_time, levels[
+                level], message).encode("utf-8").decode("latin1"))
+        else:
+            print("{} [{}]\t{}".format(current_time, levels[level], message)
+                  .encode("utf-8").decode("latin1"))
+        sys.stdout.flush()
+
+
+def debug(message="", use_color=False):
+    log(level=3, message=message, use_color=use_color)
+
+
+def info(message="", use_color=False):
+    log(level=2, message=message, use_color=use_color)
+
+
+def warning(message="", use_color=True):
+    log(level=1, message=message, use_color=use_color)
+
+
+def error(message="", use_color=True, exit=True):
+    log(level=0, message=message, use_color=use_color)
+    if exit:
+        sys.exit(-1)

deploy/openvino/scripts/bootstrap.sh

-# 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/openvino/scripts/build.sh

-# openvino预编译库的路径
-OPENVINO_DIR=/path/to/inference_engine/
-# gflags预编译库的路径
-GFLAGS_DIR=/path/to/gflags
+# OpenVINO预编译库的路径
+OPENVINO_DIR=$INTEL_OPENVINO_DIR/inference_engine
+
 # ngraph lib的路径，编译openvino时通常会生成
-NGRAPH_LIB=/path/to/ngraph/lib/
+NGRAPH_LIB=$INTEL_OPENVINO_DIR/deployment_tools/ngraph/lib
+
+# gflags预编译库的路径
+GFLAGS_DIR=$(pwd)/deps/gflags
+# glog预编译库的路径
+GLOG_DIR=$(pwd)/deps/glog
+
+# opencv使用自带预编译版本
+OPENCV_DIR=$(pwd)/deps/opencv/
+
+#cpu架构
+ARCH=x86
+export ARCH
 
-# opencv预编译库的路径, 如果使用自带预编译版本可不修改
-OPENCV_DIR=$(pwd)/deps/opencv3gcc4.8/
-# 下载自带预编译版本
-sh $(pwd)/scripts/bootstrap.sh
+#下载并编译third-part lib
+sh $(pwd)/scripts/install_third-party.sh
 
 rm -rf build
 mkdir -p build
@@ -16,6 +25,8 @@ cd build
 cmake .. \
     -DOPENCV_DIR=${OPENCV_DIR} \
     -DGFLAGS_DIR=${GFLAGS_DIR} \
+    -DGLOG_DIR=${GLOG_DIR} \
     -DOPENVINO_DIR=${OPENVINO_DIR} \
-    -DNGRAPH_LIB=${NGRAPH_LIB} 
+    -DNGRAPH_LIB=${NGRAPH_LIB} \
+    -DARCH=${ARCH}
 make

deploy/openvino/scripts/install_third-party.sh

+# download third-part lib
+if [ ! -d "./deps" ]; then
+    mkdir deps
+fi
+if [ ! -d "./deps/gflag" ]; then
+    cd deps
+    git clone https://github.com/gflags/gflags
+    cd gflags
+    cmake .
+    make -j 8
+    cd ..
+    cd ..
+fi
+if [ ! -d "./deps/glog" ]; then
+    cd deps
+    git clone https://github.com/google/glog
+    sudo apt-get install autoconf automake libtool
+    cd glog
+    ./autogen.sh
+    ./configure
+    make -j 8
+    cd ..
+    cd ..
+fi
+
+if [ "$ARCH" = "x86" ]; then
+    OPENCV_URL=https://bj.bcebos.com/paddlex/deploy/x86opencv/opencv.tar.bz2 
+else
+    OPENCV_URL=https://bj.bcebos.com/paddlex/deploy/armopencv/opencv.tar.bz2
+fi
+if [ ! -d "./deps/opencv" ]; then
+    cd deps
+    wget -c ${OPENCV_URL}
+    tar xvfj opencv.tar.bz2
+    rm -rf opencv.tar.bz2
+    cd ..
+fi

deploy/openvino/src/paddlex.cpp

@@ -13,6 +13,8 @@
 // limitations under the License.
 
 #include "include/paddlex/paddlex.h"
+#include <iostream>
+#include <fstream>
 
 using namespace InferenceEngine;
 
@@ -24,12 +26,28 @@ void Model::create_predictor(const std::string& model_dir,
     Core ie;
     network_ = ie.ReadNetwork(model_dir, model_dir.substr(0, model_dir.size() - 4) + ".bin");
     network_.setBatchSize(1);
-    InputInfo::Ptr input_info = network_.getInputsInfo().begin()->second;
 
-    input_info->getPreProcess().setResizeAlgorithm(RESIZE_BILINEAR);
-    input_info->setLayout(Layout::NCHW);
-    input_info->setPrecision(Precision::FP32);
+    InputsDataMap inputInfo(network_.getInputsInfo());
+    std::string imageInputName;
+    for (const auto & inputInfoItem : inputInfo) {
+      if (inputInfoItem.second->getTensorDesc().getDims().size() == 4){
+        imageInputName = inputInfoItem.first;
+        inputInfoItem.second->setPrecision(Precision::FP32);
+        inputInfoItem.second->getPreProcess().setResizeAlgorithm(RESIZE_BILINEAR);
+        inputInfoItem.second->setLayout(Layout::NCHW);
+      }
+      if (inputInfoItem.second->getTensorDesc().getDims().size() == 2){
+        imageInputName = inputInfoItem.first;
+        inputInfoItem.second->setPrecision(Precision::FP32);
+      }
+    }
+    if(device == "MYRIAD"){
+      std::map<std::string, std::string> networkConfig;
+      networkConfig["VPU_HW_STAGES_OPTIMIZATION"] = "ON";
+      executable_network_ = ie.LoadNetwork(network_, device, networkConfig);
+    }else{
       executable_network_ = ie.LoadNetwork(network_, device);
+    }
     load_config(cfg_dir);
 }
 
@@ -49,21 +67,25 @@ bool Model::load_config(const std::string& cfg_dir) {
     }
   }
   // 构建数据处理流
-  transforms_.Init(config["Transforms"], to_rgb);
-  // 读入label list
-  labels.clear();
-  labels = config["_Attributes"]["labels"].as<std::vector<std::string>>();
+  transforms_.Init(config["Transforms"], type, to_rgb);
+  // 读入label lis
+  for (const auto& item : config["_Attributes"]["labels"]) {
+    int index = labels.size();
+    labels[index] = item.as<std::string>();
+  }
+
   return true;
 }
 
-bool Model::preprocess(cv::Mat* input_im) {
-  if (!transforms_.Run(input_im, inputs_)) {
+bool Model::preprocess(cv::Mat* input_im, ImageBlob* inputs) {
+  if (!transforms_.Run(input_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()!"
@@ -78,16 +100,17 @@ bool Model::predict(const cv::Mat& im, ClsResult* result) {
   // 处理输入图像
   InferRequest infer_request = executable_network_.CreateInferRequest();
   std::string input_name = network_.getInputsInfo().begin()->first;
-  inputs_ = infer_request.GetBlob(input_name);
-
-  auto im_clone = im.clone();
-  if (!preprocess(&im_clone)) {
+  inputs_.blob = infer_request.GetBlob(input_name);
+  cv::Mat im_clone = im.clone();
+  if (!preprocess(&im_clone, &inputs_)) {
     std::cerr << "Preprocess failed!" << std::endl;
     return false;
   }
   
+
   infer_request.Infer();
 
+
   std::string output_name = network_.getOutputsInfo().begin()->first;
   output_ = infer_request.GetBlob(output_name);
   MemoryBlob::CPtr moutput = as<MemoryBlob>(output_);
@@ -99,10 +122,198 @@ bool Model::predict(const cv::Mat& im, ClsResult* result) {
   result->category_id = std::distance(outputs_data, ptr);
   result->score = *ptr;
   result->category = labels[result->category_id];
-  //for (int i=0;i<sizeof(outputs_data);i++){
-  //    std::cout <<  labels[i] << std::endl;
-  //    std::cout <<  outputs_[i] << std::endl;
-  //    }
+
+}
+
+bool Model::predict(const cv::Mat& im, DetResult* result) {
+  inputs_.clear();
+  result->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;
+  }
+  InferRequest infer_request = executable_network_.CreateInferRequest();
+  InputsDataMap input_maps = network_.getInputsInfo();
+  std::string inputName;
+  for (const auto & input_map : input_maps) {
+    if (input_map.second->getTensorDesc().getDims().size() == 4){
+      inputName = input_map.first;
+      inputs_.blob = infer_request.GetBlob(inputName);
+    }
+    if (input_map.second->getTensorDesc().getDims().size() == 2){
+      inputName = input_map.first;
+      inputs_.ori_im_size_ = infer_request.GetBlob(inputName);
+    }
+  } 
+  cv::Mat im_clone = im.clone();
+  if (!preprocess(&im_clone, &inputs_)) {
+    std::cerr << "Preprocess failed!" << std::endl;
+    return false;
+  }
+  
+
+  infer_request.Infer();
+
+
+  OutputsDataMap out_map = network_.getOutputsInfo(); 
+  auto iter = out_map.begin();
+  std::string outputName = iter->first;
+  std::cout << "output: " << outputName << std::endl;
+  Blob::Ptr output = infer_request.GetBlob(outputName);
+  MemoryBlob::CPtr moutput = as<MemoryBlob>(output);
+  TensorDesc blob_output = moutput->getTensorDesc();
+  std::vector<size_t> output_shape = blob_output.getDims();
+  auto moutputHolder = moutput->rmap();
+  float* data = moutputHolder.as<float *>();
+  int size = 1;
+  for (auto& i : output_shape){
+    size *= static_cast<int>(i);
+  }
+  int num_boxes = size / 6;
+  for (int i = 0; i < num_boxes; ++i){
+    if(data[i * 6] > 0){
+      Box box;
+      box.category_id = static_cast<int>(data[i * 6]);
+      box.category = labels[box.category_id];
+      box.score = data[i * 6 + 1];
+      float xmin = data[i * 6 + 2];
+      float ymin = data[i * 6 + 3];
+      float xmax = data[i * 6 + 4];
+      float ymax = data[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));
+    }
+  }
+}
+
+
+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;
+  }
+  //
+  InferRequest infer_request = executable_network_.CreateInferRequest();
+  std::string input_name = network_.getInputsInfo().begin()->first;
+  inputs_.blob = infer_request.GetBlob(input_name);
+  
+  //
+  cv::Mat im_clone = im.clone();
+  if (!preprocess(&im_clone, &inputs_)) {
+    std::cerr << "Preprocess failed!" << std::endl;
+    return false;
+  }
+
+  //
+  infer_request.Infer();
+  if (count_num_ >= 20){
+    total_time_ = total_time_ + time_used.count();
+  }
+ 
+ 
+  OutputsDataMap out_map = network_.getOutputsInfo();
+  auto iter = out_map.begin();
+  iter++;
+  std::string output_name_score = iter->first;
+  std::cout << iter->first  << std::endl;
+  Blob::Ptr output_score = infer_request.GetBlob(output_name_score);
+  MemoryBlob::CPtr moutput_score = as<MemoryBlob>(output_score);
+  TensorDesc blob_score = moutput_score->getTensorDesc();
+  std::vector<size_t> output_score_shape = blob_score.getDims();
+  int size = 1;
+  for (auto& i : output_score_shape) {
+    size *= static_cast<int>(i);
+    result->score_map.shape.push_back(static_cast<int>(i));
+  }
+  result->score_map.data.resize(size);
+  auto moutputHolder_score = moutput_score->rmap();
+  float* score_data = moutputHolder_score.as<float *>();
+  memcpy(result->score_map.data.data(),score_data,moutput_score->byteSize());
+
+  iter++;
+  std::string output_name_label = iter->first;
+  std::cout << iter->first  << std::endl;
+  Blob::Ptr output_label = infer_request.GetBlob(output_name_label);
+  MemoryBlob::CPtr moutput_label = as<MemoryBlob>(output_label);
+  TensorDesc blob_label = moutput_label->getTensorDesc();
+  std::vector<size_t> output_label_shape = blob_label.getDims();
+  size = 1;
+  for (auto& i : output_label_shape) {
+    size *= static_cast<int>(i);
+    result->label_map.shape.push_back(static_cast<int>(i));
+  }
+  result->label_map.data.resize(size);
+  auto moutputHolder_label = moutput_label->rmap();
+  int* label_data = moutputHolder_label.as<int *>();
+  memcpy(result->label_map.data.data(),label_data,moutput_label->byteSize());
+
+
+
+  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());
+  int idx = 1;
+  int len_postprocess = inputs_.im_size_before_resize_.size();
+  for (std::vector<std::string>::reverse_iterator iter =
+           inputs_.reshape_order_.rbegin();
+       iter != inputs_.reshape_order_.rend();
+       ++iter) {
+    if (*iter == "padding") {
+      auto before_shape = inputs_.im_size_before_resize_[len_postprocess - idx];
+      inputs_.im_size_before_resize_.pop_back();
+      auto padding_w = before_shape[0];
+      auto padding_h = before_shape[1];
+      mask_label = mask_label(cv::Rect(0, 0, padding_h, padding_w));
+      mask_score = mask_score(cv::Rect(0, 0, padding_h, padding_w));
+    } else if (*iter == "resize") {
+      auto before_shape = inputs_.im_size_before_resize_[len_postprocess - idx];
+      inputs_.im_size_before_resize_.pop_back();
+      auto resize_w = before_shape[0];
+      auto resize_h = before_shape[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_LINEAR);
+    }
+    ++idx;
+  }
+  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};
+  return true;
 }
   
 }  // namespce of PaddleX

deploy/openvino/src/transforms.cpp

@@ -13,8 +13,10 @@
 // limitations under the License.
 
 #include <iostream>
+#include <fstream>
 #include <string>
 #include <vector>
+#include <math.h>
 
 #include "include/paddlex/transforms.h"
 
@@ -26,7 +28,7 @@ std::map<std::string, int> interpolations = {{"LINEAR", cv::INTER_LINEAR},
                                              {"CUBIC", cv::INTER_CUBIC},
                                              {"LANCZOS4", cv::INTER_LANCZOS4}};
 
-bool Normalize::Run(cv::Mat* im){
+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] =
@@ -40,19 +42,6 @@ bool Normalize::Run(cv::Mat* im){
   return true;
 }
 
-bool CenterCrop::Run(cv::Mat* im) {
-  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);
-  return true;
-}
 
 
 float ResizeByShort::GenerateScale(const cv::Mat& im) {
@@ -70,17 +59,117 @@ float ResizeByShort::GenerateScale(const cv::Mat& im) {
   return scale;
 }
 
-bool ResizeByShort::Run(cv::Mat* im) {
+bool ResizeByShort::Run(cv::Mat* im, ImageBlob* data) {
+  data->im_size_before_resize_.push_back({im->rows, 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);
+  int width = static_cast<int>(round(scale * im->cols));
+  int height = static_cast<int>(round(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;
 }
 
-void Transforms::Init(const YAML::Node& transforms_node, bool to_rgb) {
+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_resize_.push_back({im->rows, im->cols});
+  data->reshape_order_.push_back("padding");
+
+  int padding_w = 0;
+  int padding_h = 0;
+  if (width_ > 1 & height_ > 1) {
+    padding_w = width_ - im->cols;
+    padding_h = height_ - im->rows;
+  } else if (coarsest_stride_ >= 1) {
+    int h = im->rows;
+    int w = im->cols;
+    padding_h =
+        ceil(h * 1.0 / coarsest_stride_) * coarsest_stride_ - im->rows;
+    padding_w =
+        ceil(w * 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::Scalar value = cv::Scalar(im_value_[0], im_value_[1], im_value_[2]);
+  cv::copyMakeBorder(
+      *im, *im, 0, padding_h, 0, padding_w, cv::BORDER_CONSTANT, value);
+  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_.push_back({im->rows, 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_.push_back({im->rows, 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, std::string type, bool to_rgb) {
   transforms_.clear();
   to_rgb_ = to_rgb;
+  type_ = type;
   for (const auto& item : transforms_node) {
     std::string name = item.begin()->first.as<std::string>();
     std::cout << "trans name: " << name << std::endl;
@@ -94,10 +183,16 @@ std::shared_ptr<Transform> Transforms::CreateTransform(
     const std::string& transform_name) {
   if (transform_name == "Normalize") {
     return std::make_shared<Normalize>();
-  } else if (transform_name == "CenterCrop") {
-    return std::make_shared<CenterCrop>();
   } 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;
@@ -105,15 +200,26 @@ std::shared_ptr<Transform> Transforms::CreateTransform(
   }
 }
 
-bool Transforms::Run(cv::Mat* im, Blob::Ptr blob) {
+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);
   
+  if(type_ == "detector" ){
+    LockedMemory<void> input2Mapped = as<MemoryBlob>(data->ori_im_size_)->wmap();
+    float *p = input2Mapped.as<float*>();
+    p[0] = im->rows;
+    p[1] = im->cols;
+  }
+  //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)) {
+    if (!transforms_[i]->Run(im,data)) {
       std::cerr << "Apply transforms to image failed!" << std::endl;
       return false;
     }
@@ -121,13 +227,15 @@ bool Transforms::Run(cv::Mat* im, Blob::Ptr blob) {
 
   // 将图像由NHWC转为NCHW格式
   // 同时转为连续的内存块存储到Blob
-  SizeVector blobSize = blob->getTensorDesc().getDims();
+  
+  SizeVector blobSize = data->blob->getTensorDesc().getDims();
   const size_t width = blobSize[3];
   const size_t height = blobSize[2];
   const size_t channels = blobSize[1];
-  MemoryBlob::Ptr mblob = InferenceEngine::as<MemoryBlob>(blob);
+  MemoryBlob::Ptr mblob = InferenceEngine::as<MemoryBlob>(data->blob);
   auto mblobHolder = mblob->wmap();
   float *blob_data = mblobHolder.as<float *>();
+    
   for (size_t c = 0; c < channels; c++) {
       for (size_t  h = 0; h < height; h++) {
           for (size_t w = 0; w < width; w++) {

deploy/openvino/src/visualize.cpp

+//   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 Visualize(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 Visualize(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 PaddleX