Merge branch 'develop' into develop_draw

deploy/README.md

@@ -14,3 +14,5 @@
     - [模型量化](../docs/deploy/paddlelite/slim/quant.md)
     - [模型裁剪](../docs/deploy/paddlelite/slim/prune.md)
   - [Android平台](../docs/deploy/paddlelite/android.md)
+- [OpenVINO部署](../docs/deploy/openvino/introduction.md)
+- [树莓派部署](../docs/deploy/raspberry/Raspberry.md)
\ No newline at end of file

deploy/cpp/demo/classifier.cpp

@@ -29,6 +29,10 @@ using namespace std::chrono;  // NOLINT
 DEFINE_string(model_dir, "", "Path of inference model");
 DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
 DEFINE_bool(use_trt, false, "Infering with TensorRT");
+DEFINE_bool(use_mkl, true, "Infering with MKL");
+DEFINE_int32(mkl_thread_num,
+             omp_get_num_procs(),
+             "Number of mkl threads");
 DEFINE_int32(gpu_id, 0, "GPU card id");
 DEFINE_string(key, "", "key of encryption");
 DEFINE_string(image, "", "Path of test image file");
@@ -56,6 +60,8 @@ int main(int argc, char** argv) {
   model.Init(FLAGS_model_dir,
              FLAGS_use_gpu,
              FLAGS_use_trt,
+             FLAGS_use_mkl,
+             FLAGS_mkl_thread_num,
              FLAGS_gpu_id,
              FLAGS_key);
 

deploy/cpp/demo/detector.cpp

@@ -31,6 +31,10 @@ using namespace std::chrono;  // NOLINT
 DEFINE_string(model_dir, "", "Path of inference model");
 DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
 DEFINE_bool(use_trt, false, "Infering with TensorRT");
+DEFINE_bool(use_mkl, true, "Infering with MKL");
+DEFINE_int32(mkl_thread_num,
+             omp_get_num_procs(),
+             "Number of mkl threads");
 DEFINE_int32(gpu_id, 0, "GPU card id");
 DEFINE_string(key, "", "key of encryption");
 DEFINE_string(image, "", "Path of test image file");
@@ -61,6 +65,8 @@ int main(int argc, char** argv) {
   model.Init(FLAGS_model_dir,
              FLAGS_use_gpu,
              FLAGS_use_trt,
+             FLAGS_use_mkl,
+             FLAGS_mkl_thread_num,
              FLAGS_gpu_id,
              FLAGS_key);
   int imgs = 1;

deploy/cpp/demo/segmenter.cpp

@@ -30,6 +30,10 @@ using namespace std::chrono;  // NOLINT
 DEFINE_string(model_dir, "", "Path of inference model");
 DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
 DEFINE_bool(use_trt, false, "Infering with TensorRT");
+DEFINE_bool(use_mkl, true, "Infering with MKL");
+DEFINE_int32(mkl_thread_num,
+             omp_get_num_procs(),
+             "Number of mkl threads");
 DEFINE_int32(gpu_id, 0, "GPU card id");
 DEFINE_string(key, "", "key of encryption");
 DEFINE_string(image, "", "Path of test image file");
@@ -58,6 +62,8 @@ int main(int argc, char** argv) {
   model.Init(FLAGS_model_dir,
              FLAGS_use_gpu,
              FLAGS_use_trt,
+             FLAGS_use_mkl,
+             FLAGS_mkl_thread_num,
              FLAGS_gpu_id,
              FLAGS_key);
   int imgs = 1;

deploy/cpp/demo/video_classifier.cpp

@@ -35,8 +35,12 @@ using namespace std::chrono;  // NOLINT
 DEFINE_string(model_dir, "", "Path of inference model");
 DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
 DEFINE_bool(use_trt, false, "Infering with TensorRT");
+DEFINE_bool(use_mkl, true, "Infering with MKL");
 DEFINE_int32(gpu_id, 0, "GPU card id");
 DEFINE_string(key, "", "key of encryption");
+DEFINE_int32(mkl_thread_num,
+             omp_get_num_procs(),
+             "Number of mkl threads");
 DEFINE_bool(use_camera, false, "Infering with Camera");
 DEFINE_int32(camera_id, 0, "Camera id");
 DEFINE_string(video_path, "", "Path of input video");
@@ -62,6 +66,8 @@ int main(int argc, char** argv) {
   model.Init(FLAGS_model_dir,
              FLAGS_use_gpu,
              FLAGS_use_trt,
+             FLAGS_use_mkl,
+             FLAGS_mkl_thread_num,
              FLAGS_gpu_id,
              FLAGS_key);
 

deploy/cpp/demo/video_detector.cpp

@@ -35,6 +35,7 @@ using namespace std::chrono;  // NOLINT
 DEFINE_string(model_dir, "", "Path of inference model");
 DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
 DEFINE_bool(use_trt, false, "Infering with TensorRT");
+DEFINE_bool(use_mkl, true, "Infering with MKL");
 DEFINE_int32(gpu_id, 0, "GPU card id");
 DEFINE_bool(use_camera, false, "Infering with Camera");
 DEFINE_int32(camera_id, 0, "Camera id");
@@ -42,6 +43,9 @@ DEFINE_string(video_path, "", "Path of input video");
 DEFINE_bool(show_result, false, "show the result of each frame with a window");
 DEFINE_bool(save_result, true, "save the result of each frame to a video");
 DEFINE_string(key, "", "key of encryption");
+DEFINE_int32(mkl_thread_num,
+             omp_get_num_procs(),
+             "Number of mkl threads");
 DEFINE_string(save_dir, "output", "Path to save visualized image");
 DEFINE_double(threshold,
               0.5,
@@ -64,6 +68,8 @@ int main(int argc, char** argv) {
   model.Init(FLAGS_model_dir,
              FLAGS_use_gpu,
              FLAGS_use_trt,
+             FLAGS_use_mkl,
+             FLAGS_mkl_thread_num,
              FLAGS_gpu_id,
              FLAGS_key);
   // Open video

deploy/cpp/demo/video_segmenter.cpp

@@ -35,8 +35,12 @@ using namespace std::chrono;  // NOLINT
 DEFINE_string(model_dir, "", "Path of inference model");
 DEFINE_bool(use_gpu, false, "Infering with GPU or CPU");
 DEFINE_bool(use_trt, false, "Infering with TensorRT");
+DEFINE_bool(use_mkl, true, "Infering with MKL");
 DEFINE_int32(gpu_id, 0, "GPU card id");
 DEFINE_string(key, "", "key of encryption");
+DEFINE_int32(mkl_thread_num,
+             omp_get_num_procs(),
+             "Number of mkl threads");
 DEFINE_bool(use_camera, false, "Infering with Camera");
 DEFINE_int32(camera_id, 0, "Camera id");
 DEFINE_string(video_path, "", "Path of input video");
@@ -62,6 +66,8 @@ int main(int argc, char** argv) {
   model.Init(FLAGS_model_dir,
              FLAGS_use_gpu,
              FLAGS_use_trt,
+             FLAGS_use_mkl,
+             FLAGS_mkl_thread_num,
              FLAGS_gpu_id,
              FLAGS_key);
   // Open video

deploy/cpp/include/paddlex/paddlex.h

@@ -70,6 +70,8 @@ class Model {
    * @param model_dir: the directory which contains model.yml
    * @param use_gpu: use gpu or not when infering
    * @param use_trt: use Tensor RT or not when infering
+   * @param use_mkl: use mkl or not when infering
+   * @param mkl_thread_num: number of threads for mkldnn when infering
    * @param gpu_id: the id of gpu when infering with using gpu
    * @param key: the key of encryption when using encrypted model
    * @param use_ir_optim: use ir optimization when infering
@@ -77,15 +79,26 @@ class Model {
   void Init(const std::string& model_dir,
             bool use_gpu = false,
             bool use_trt = false,
+            bool use_mkl = true,
+            int mkl_thread_num = 4,
             int gpu_id = 0,
             std::string key = "",
             bool use_ir_optim = true) {
-    create_predictor(model_dir, use_gpu, use_trt, gpu_id, key, use_ir_optim);
+    create_predictor(
+                     model_dir,
+                     use_gpu,
+                     use_trt,
+                     use_mkl,
+                     mkl_thread_num,
+                     gpu_id,
+                     key,
+                     use_ir_optim);
   }
-
   void create_predictor(const std::string& model_dir,
                         bool use_gpu = false,
                         bool use_trt = false,
+                        bool use_mkl = true,
+                        int mkl_thread_num = 4,
                         int gpu_id = 0,
                         std::string key = "",
                         bool use_ir_optim = true);

deploy/cpp/include/paddlex/results.h

@@ -47,7 +47,7 @@ struct Box {
   // confidence score
   float score;
   std::vector<float> coordinate;
-  Mask<float> mask;
+  Mask<int> mask;
 };
 
 /*

deploy/cpp/include/paddlex/transforms.h

@@ -21,6 +21,7 @@
 #include <unordered_map>
 #include <utility>
 #include <vector>
+#include <iostream>
 
 #include <opencv2/core/core.hpp>
 #include <opencv2/highgui/highgui.hpp>
@@ -216,8 +217,7 @@ class Padding : public Transform {
     }
     if (item["im_padding_value"].IsDefined()) {
       im_value_ = item["im_padding_value"].as<std::vector<float>>();
-    }
-    else {
+    } else {
       im_value_ = {0, 0, 0};
     }
   }

deploy/cpp/src/paddlex.cpp

@@ -11,16 +11,25 @@
 // 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 <math.h>
 #include <omp.h>
 #include <algorithm>
 #include <fstream>
 #include <cstring>
 #include "include/paddlex/paddlex.h"
+
+#include <opencv2/core/core.hpp>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/imgproc/imgproc.hpp>
+
 namespace PaddleX {
 
 void Model::create_predictor(const std::string& model_dir,
                              bool use_gpu,
                              bool use_trt,
+                             bool use_mkl,
+                             int mkl_thread_num,
                              int gpu_id,
                              std::string key,
                              bool use_ir_optim) {
@@ -40,7 +49,7 @@ void Model::create_predictor(const std::string& model_dir,
   }
 #endif
   if (yaml_input == "") {
-    // 读取配置文件
+    // read yaml file
     std::ifstream yaml_fin(yaml_file);
     yaml_fin.seekg(0, std::ios::end);
     size_t yaml_file_size = yaml_fin.tellg();
@@ -48,7 +57,7 @@ void Model::create_predictor(const std::string& model_dir,
     yaml_fin.seekg(0);
     yaml_fin.read(&yaml_input[0], yaml_file_size);
   }
-  // 读取配置文件内容
+  // load yaml file
   if (!load_config(yaml_input)) {
     std::cerr << "Parse file 'model.yml' failed!" << std::endl;
     exit(-1);
@@ -57,6 +66,10 @@ void Model::create_predictor(const std::string& model_dir,
   if (key == "") {
     config.SetModel(model_file, params_file);
   }
+  if (use_mkl && name != "HRNet" && name != "DeepLabv3p") {
+    config.EnableMKLDNN();
+    config.SetCpuMathLibraryNumThreads(mkl_thread_num);
+  }
   if (use_gpu) {
     config.EnableUseGpu(100, gpu_id);
   } else {
@@ -64,13 +77,13 @@ void Model::create_predictor(const std::string& model_dir,
   }
   config.SwitchUseFeedFetchOps(false);
   config.SwitchSpecifyInputNames(true);
-  // 开启图优化
+  // enable graph Optim
 #if defined(__arm__) || defined(__aarch64__)
   config.SwitchIrOptim(false);
 #else
   config.SwitchIrOptim(use_ir_optim);
 #endif
-  // 开启内存优化
+  // enable Memory Optim
   config.EnableMemoryOptim();
   if (use_trt) {
     config.EnableTensorRtEngine(
@@ -108,9 +121,9 @@ bool Model::load_config(const std::string& yaml_input) {
       return false;
     }
   }
-  // 构建数据处理流
+  // build data preprocess stream
   transforms_.Init(config["Transforms"], to_rgb);
-  // 读入label list
+  // read label list
   labels.clear();
   for (const auto& item : config["_Attributes"]["labels"]) {
     int index = labels.size();
@@ -152,19 +165,19 @@ bool Model::predict(const cv::Mat& im, ClsResult* result) {
                  "to function predict()!" << std::endl;
     return false;
   }
-  // 处理输入图像
+  // im preprocess
   if (!preprocess(im, &inputs_)) {
     std::cerr << "Preprocess failed!" << std::endl;
     return false;
   }
-  // 使用加载的模型进行预测
+  // predict
   auto in_tensor = predictor_->GetInputTensor("image");
   int h = inputs_.new_im_size_[0];
   int w = inputs_.new_im_size_[1];
   in_tensor->Reshape({1, 3, h, w});
   in_tensor->copy_from_cpu(inputs_.im_data_.data());
   predictor_->ZeroCopyRun();
-  // 取出模型的输出结果
+  // get result
   auto output_names = predictor_->GetOutputNames();
   auto output_tensor = predictor_->GetOutputTensor(output_names[0]);
   std::vector<int> output_shape = output_tensor->shape();
@@ -174,7 +187,7 @@ bool Model::predict(const cv::Mat& im, ClsResult* result) {
   }
   outputs_.resize(size);
   output_tensor->copy_to_cpu(outputs_.data());
-  // 对模型输出结果进行后处理
+  // postprocess
   auto ptr = std::max_element(std::begin(outputs_), std::end(outputs_));
   result->category_id = std::distance(std::begin(outputs_), ptr);
   result->score = *ptr;
@@ -198,12 +211,12 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
     return false;
   }
   inputs_batch_.assign(im_batch.size(), ImageBlob());
-  // 处理输入图像
+  // preprocess
   if (!preprocess(im_batch, &inputs_batch_, thread_num)) {
     std::cerr << "Preprocess failed!" << std::endl;
     return false;
   }
-  // 使用加载的模型进行预测
+  // predict
   int batch_size = im_batch.size();
   auto in_tensor = predictor_->GetInputTensor("image");
   int h = inputs_batch_[0].new_im_size_[0];
@@ -218,7 +231,7 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
   in_tensor->copy_from_cpu(inputs_data.data());
   // in_tensor->copy_from_cpu(inputs_.im_data_.data());
   predictor_->ZeroCopyRun();
-  // 取出模型的输出结果
+  // get result
   auto output_names = predictor_->GetOutputNames();
   auto output_tensor = predictor_->GetOutputTensor(output_names[0]);
   std::vector<int> output_shape = output_tensor->shape();
@@ -228,7 +241,7 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
   }
   outputs_.resize(size);
   output_tensor->copy_to_cpu(outputs_.data());
-  // 对模型输出结果进行后处理
+  // postprocess
   (*results).clear();
   (*results).resize(batch_size);
   int single_batch_size = size / batch_size;
@@ -258,7 +271,7 @@ bool Model::predict(const cv::Mat& im, DetResult* result) {
     return false;
   }
 
-  // 处理输入图像
+  // preprocess
   if (!preprocess(im, &inputs_)) {
     std::cerr << "Preprocess failed!" << std::endl;
     return false;
@@ -288,7 +301,7 @@ bool Model::predict(const cv::Mat& im, DetResult* result) {
     im_info_tensor->copy_from_cpu(im_info);
     im_shape_tensor->copy_from_cpu(im_shape);
   }
-  // 使用加载的模型进行预测
+  // predict
   predictor_->ZeroCopyRun();
 
   std::vector<float> output_box;
@@ -306,7 +319,7 @@ bool Model::predict(const cv::Mat& im, DetResult* result) {
     return true;
   }
   int num_boxes = size / 6;
-  // 解析预测框box
+  // box postprocess
   for (int i = 0; i < num_boxes; ++i) {
     Box box;
     box.category_id = static_cast<int>(round(output_box[i * 6]));
@@ -321,7 +334,7 @@ bool Model::predict(const cv::Mat& im, DetResult* result) {
     box.coordinate = {xmin, ymin, w, h};
     result->boxes.push_back(std::move(box));
   }
-  // 实例分割需解析mask
+  // mask postprocess
   if (name == "MaskRCNN") {
     std::vector<float> output_mask;
     auto output_mask_tensor = predictor_->GetOutputTensor(output_names[1]);
@@ -337,12 +350,22 @@ bool Model::predict(const cv::Mat& im, DetResult* result) {
     result->mask_resolution = output_mask_shape[2];
     for (int i = 0; i < result->boxes.size(); ++i) {
       Box* box = &result->boxes[i];
-      auto begin_mask =
-          output_mask.begin() + (i * classes + box->category_id) * mask_pixels;
-      auto end_mask = begin_mask + mask_pixels;
-      box->mask.data.assign(begin_mask, end_mask);
       box->mask.shape = {static_cast<int>(box->coordinate[2]),
                          static_cast<int>(box->coordinate[3])};
+      auto begin_mask =
+          output_mask.data() + (i * classes + box->category_id) * mask_pixels;
+      cv::Mat bin_mask(result->mask_resolution,
+                     result->mask_resolution,
+                     CV_32FC1,
+                     begin_mask);
+      cv::resize(bin_mask,
+               bin_mask,
+               cv::Size(box->mask.shape[0], box->mask.shape[1]));
+      cv::threshold(bin_mask, bin_mask, 0.5, 1, cv::THRESH_BINARY);
+      auto mask_int_begin = reinterpret_cast<float*>(bin_mask.data);
+      auto mask_int_end =
+        mask_int_begin + box->mask.shape[0] * box->mask.shape[1];
+      box->mask.data.assign(mask_int_begin, mask_int_end);
     }
   }
   return true;
@@ -366,12 +389,12 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
 
   inputs_batch_.assign(im_batch.size(), ImageBlob());
   int batch_size = im_batch.size();
-  // 处理输入图像
+  // preprocess
   if (!preprocess(im_batch, &inputs_batch_, thread_num)) {
     std::cerr << "Preprocess failed!" << std::endl;
     return false;
   }
-  // 对RCNN类模型做批量padding
+  // RCNN model padding
   if (batch_size > 1) {
     if (name == "FasterRCNN" || name == "MaskRCNN") {
       int max_h = -1;
@@ -452,10 +475,10 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
     im_info_tensor->copy_from_cpu(im_info.data());
     im_shape_tensor->copy_from_cpu(im_shape.data());
   }
-  // 使用加载的模型进行预测
+  // predict
   predictor_->ZeroCopyRun();
 
-  // 读取所有box
+  // get all box
   std::vector<float> output_box;
   auto output_names = predictor_->GetOutputNames();
   auto output_box_tensor = predictor_->GetOutputTensor(output_names[0]);
@@ -472,7 +495,7 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
   }
   auto lod_vector = output_box_tensor->lod();
   int num_boxes = size / 6;
-  // 解析预测框box
+  // box postprocess
   (*results).clear();
   (*results).resize(batch_size);
   for (int i = 0; i < lod_vector[0].size() - 1; ++i) {
@@ -492,7 +515,7 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
     }
   }
 
-  // 实例分割需解析mask
+  // mask postprocess
   if (name == "MaskRCNN") {
     std::vector<float> output_mask;
     auto output_mask_tensor = predictor_->GetOutputTensor(output_names[1]);
@@ -509,14 +532,24 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
     for (int i = 0; i < lod_vector[0].size() - 1; ++i) {
       (*results)[i].mask_resolution = output_mask_shape[2];
       for (int j = 0; j < (*results)[i].boxes.size(); ++j) {
-        Box* box = &(*results)[i].boxes[j];
+        Box* box = &(*results)[i].boxes[i];
         int category_id = box->category_id;
-        auto begin_mask = output_mask.begin() +
-                          (mask_idx * classes + category_id) * mask_pixels;
-        auto end_mask = begin_mask + mask_pixels;
-        box->mask.data.assign(begin_mask, end_mask);
         box->mask.shape = {static_cast<int>(box->coordinate[2]),
                           static_cast<int>(box->coordinate[3])};
+        auto begin_mask =
+          output_mask.data() + (i * classes + box->category_id) * mask_pixels;
+        cv::Mat bin_mask(output_mask_shape[2],
+                      output_mask_shape[2],
+                      CV_32FC1,
+                      begin_mask);
+        cv::resize(bin_mask,
+                bin_mask,
+                cv::Size(box->mask.shape[0], box->mask.shape[1]));
+        cv::threshold(bin_mask, bin_mask, 0.5, 1, cv::THRESH_BINARY);
+        auto mask_int_begin = reinterpret_cast<float*>(bin_mask.data);
+        auto mask_int_end =
+          mask_int_begin + box->mask.shape[0] * box->mask.shape[1];
+        box->mask.data.assign(mask_int_begin, mask_int_end);
         mask_idx++;
       }
     }
@@ -537,7 +570,7 @@ bool Model::predict(const cv::Mat& im, SegResult* result) {
     return false;
   }
 
-  // 处理输入图像
+  // preprocess
   if (!preprocess(im, &inputs_)) {
     std::cerr << "Preprocess failed!" << std::endl;
     return false;
@@ -549,10 +582,10 @@ bool Model::predict(const cv::Mat& im, SegResult* result) {
   im_tensor->Reshape({1, 3, h, w});
   im_tensor->copy_from_cpu(inputs_.im_data_.data());
 
-  // 使用加载的模型进行预测
+  // predict
   predictor_->ZeroCopyRun();
 
-  // 获取预测置信度，经过argmax后的labelmap
+  // get labelmap
   auto output_names = predictor_->GetOutputNames();
   auto output_label_tensor = predictor_->GetOutputTensor(output_names[0]);
   std::vector<int> output_label_shape = output_label_tensor->shape();
@@ -565,7 +598,7 @@ bool Model::predict(const cv::Mat& im, SegResult* result) {
   result->label_map.data.resize(size);
   output_label_tensor->copy_to_cpu(result->label_map.data.data());
 
-  // 获取预测置信度scoremap
+  // get scoremap
   auto output_score_tensor = predictor_->GetOutputTensor(output_names[1]);
   std::vector<int> output_score_shape = output_score_tensor->shape();
   size = 1;
@@ -577,7 +610,7 @@ bool Model::predict(const cv::Mat& im, SegResult* result) {
   result->score_map.data.resize(size);
   output_score_tensor->copy_to_cpu(result->score_map.data.data());
 
-  // 解析输出结果到原图大小
+  // get origin image result
   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],
@@ -647,7 +680,7 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
     return false;
   }
 
-  // 处理输入图像
+  // preprocess
   inputs_batch_.assign(im_batch.size(), ImageBlob());
   if (!preprocess(im_batch, &inputs_batch_, thread_num)) {
     std::cerr << "Preprocess failed!" << std::endl;
@@ -670,10 +703,10 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
   im_tensor->copy_from_cpu(inputs_data.data());
   // im_tensor->copy_from_cpu(inputs_.im_data_.data());
 
-  // 使用加载的模型进行预测
+  // predict
   predictor_->ZeroCopyRun();
 
-  // 获取预测置信度，经过argmax后的labelmap
+  // get labelmap
   auto output_names = predictor_->GetOutputNames();
   auto output_label_tensor = predictor_->GetOutputTensor(output_names[0]);
   std::vector<int> output_label_shape = output_label_tensor->shape();
@@ -698,7 +731,7 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
               (*results)[i].label_map.data.data());
   }
 
-  // 获取预测置信度scoremap
+  // get scoremap
   auto output_score_tensor = predictor_->GetOutputTensor(output_names[1]);
   std::vector<int> output_score_shape = output_score_tensor->shape();
   size = 1;
@@ -722,7 +755,7 @@ bool Model::predict(const std::vector<cv::Mat>& im_batch,
               (*results)[i].score_map.data.data());
   }
 
-  // 解析输出结果到原图大小
+  // get origin image result
   for (int i = 0; i < batch_size; ++i) {
     std::vector<uint8_t> label_map((*results)[i].label_map.data.begin(),
                                    (*results)[i].label_map.data.end());

deploy/cpp/src/transforms.cpp

@@ -12,12 +12,14 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include "include/paddlex/transforms.h"
+
+#include <math.h>
+
 #include <iostream>
 #include <string>
 #include <vector>
-#include <math.h>
 
-#include "include/paddlex/transforms.h"
 
 namespace PaddleX {
 
@@ -195,7 +197,7 @@ std::shared_ptr<Transform> Transforms::CreateTransform(
 }
 
 bool Transforms::Run(cv::Mat* im, ImageBlob* data) {
-  // 按照transforms中预处理算子顺序处理图像
+  // do all preprocess ops by order
   if (to_rgb_) {
     cv::cvtColor(*im, *im, cv::COLOR_BGR2RGB);
   }
@@ -211,8 +213,8 @@ bool Transforms::Run(cv::Mat* im, ImageBlob* data) {
     }
   }
 
-  // 将图像由NHWC转为NCHW格式
-  // 同时转为连续的内存块存储到ImageBlob
+  // data format NHWC to NCHW
+  // img data save to ImageBlob
   int h = im->rows;
   int w = im->cols;
   int c = im->channels();

deploy/cpp/src/visualize.cpp

@@ -47,7 +47,7 @@ cv::Mat Visualize(const cv::Mat& img,
                             boxes[i].coordinate[2],
                             boxes[i].coordinate[3]);
 
-    // 生成预测框和标题
+    // draw box and title
     std::string text = boxes[i].category;
     int c1 = colormap[3 * boxes[i].category_id + 0];
     int c2 = colormap[3 * boxes[i].category_id + 1];
@@ -63,13 +63,13 @@ cv::Mat Visualize(const cv::Mat& img,
     origin.x = roi.x;
     origin.y = roi.y;
 
-    // 生成预测框标题的背景
+    // background
     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
+    // draw
     cv::rectangle(vis_img, roi, roi_color, 2);
     cv::rectangle(vis_img, text_back, roi_color, -1);
     cv::putText(vis_img,
@@ -80,18 +80,16 @@ cv::Mat Visualize(const cv::Mat& img,
                 cv::Scalar(255, 255, 255),
                 thickness);
 
-    // 生成实例分割mask
+    // mask
     if (boxes[i].mask.data.size() == 0) {
       continue;
     }
-    cv::Mat bin_mask(result.mask_resolution,
-                     result.mask_resolution,
+    std::vector<float> mask_data;
+    mask_data.assign(boxes[i].mask.data.begin(), boxes[i].mask.data.end());
+    cv::Mat bin_mask(boxes[i].mask.shape[1],
+                     boxes[i].mask.shape[0],
                      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];

deploy/lite/android/sdk/src/main/java/com/baidu/paddlex/preprocess/Transforms.java

@@ -23,6 +23,7 @@ import org.opencv.core.Scalar;
 import org.opencv.core.Size;
 import org.opencv.imgproc.Imgproc;
 import java.util.ArrayList;
+import java.util.Date;
 import java.util.HashMap;
 import java.util.List;
 
@@ -101,6 +102,15 @@ public class Transforms {
                 if (info.containsKey("coarsest_stride")) {
                     padding.coarsest_stride = (int) info.get("coarsest_stride");
                 }
+                if (info.containsKey("im_padding_value")) {
+                    List<Double> im_padding_value = (List<Double>) info.get("im_padding_value");
+                    if (im_padding_value.size()!=3){
+                        Log.e(TAG, "len of im_padding_value in padding must == 3.");
+                    }
+                    for (int k =0; i<im_padding_value.size(); i++){
+                        padding.paddding_value[k] = im_padding_value.get(k);
+                    }
+                }
                 if (info.containsKey("target_size")) {
                     if (info.get("target_size") instanceof Integer) {
                         padding.width = (int) info.get("target_size");
@@ -124,7 +134,7 @@ public class Transforms {
         if(transformsMode.equalsIgnoreCase("RGB")){
             Imgproc.cvtColor(inputMat, inputMat, Imgproc.COLOR_BGR2RGB);
         }else if(!transformsMode.equalsIgnoreCase("BGR")){
-            Log.e(TAG, "transformsMode only support RGB or BGR");
+            Log.e(TAG, "transformsMode only support RGB or BGR.");
         }
         inputMat.convertTo(inputMat, CvType.CV_32FC(3));
 
@@ -136,16 +146,15 @@ public class Transforms {
         int h = inputMat.height();
         int c = inputMat.channels();
         imageBlob.setImageData(new float[w * h * c]);
-        int[] channelStride = new int[]{w * h, w * h * 2};
-        for (int y = 0; y < h; y++) {
-            for (int x = 0;
-                 x < w; x++) {
-                double[] color = inputMat.get(y, x);
-                imageBlob.getImageData()[y * w + x]  =  (float) (color[0]);
-                imageBlob.getImageData()[y * w + x +  channelStride[0]] = (float) (color[1]);
-                imageBlob.getImageData()[y * w + x +  channelStride[1]] = (float) (color[2]);
-            }
+
+        Mat singleChannelMat = new Mat(h, w, CvType.CV_32FC(1));
+        float[] singleChannelImageData = new float[w * h];
+        for (int i = 0; i < c; i++) {
+            Core.extractChannel(inputMat, singleChannelMat, i);
+            singleChannelMat.get(0, 0, singleChannelImageData);
+            System.arraycopy(singleChannelImageData ,0, imageBlob.getImageData(),i*w*h, w*h);
         }
+
         return imageBlob;
     }
 
@@ -248,6 +257,7 @@ public class Transforms {
         private double width;
         private double height;
         private double coarsest_stride;
+        private double[] paddding_value = {0.0, 0.0, 0.0};
 
         public Mat run(Mat inputMat, ImageBlob imageBlob) {
             int origin_w = inputMat.width();
@@ -264,7 +274,7 @@ public class Transforms {
             }
             imageBlob.setNewImageSize(inputMat.height(),2);
             imageBlob.setNewImageSize(inputMat.width(),3);
-            Core.copyMakeBorder(inputMat, inputMat, 0, (int)padding_h, 0, (int)padding_w, Core.BORDER_CONSTANT, new Scalar(0));
+            Core.copyMakeBorder(inputMat, inputMat, 0, (int)padding_h, 0, (int)padding_w, Core.BORDER_CONSTANT, new Scalar(paddding_value));
             return inputMat;
         }
     }

deploy/lite/android/sdk/src/main/java/com/baidu/paddlex/visual/Visualize.java

@@ -31,8 +31,11 @@ import org.opencv.core.Scalar;
 import org.opencv.core.Size;
 import org.opencv.imgproc.Imgproc;
 
+import java.nio.ByteBuffer;
+import java.nio.FloatBuffer;
 import java.util.ArrayList;
 import java.util.Arrays;
+import java.util.Date;
 import java.util.List;
 import java.util.ListIterator;
 import java.util.Map;
@@ -120,13 +123,11 @@ public class Visualize {
         int new_w = (int)imageBlob.getNewImageSize()[3];
         Mat mask = new Mat(new_h, new_w, CvType.CV_32FC(1));
         float[] scoreData = new float[new_h*new_w];
-        for  (int h = 0; h < new_h; h++) {
-            for  (int w = 0; w < new_w; w++){
-                scoreData[new_h * h + w] =  (1-result.getMask().getScoreData()[cutoutClass + h * new_h + w]) * 255;
-            }
-        }
+        System.arraycopy(result.getMask().getScoreData() ,cutoutClass*new_h*new_w, scoreData ,0, new_h*new_w);
         mask.put(0,0, scoreData);
+        Core.multiply(mask, new Scalar(255), mask);
         mask.convertTo(mask,CvType.CV_8UC(1));
+
         ListIterator<Map.Entry<String, int[]>> reverseReshapeInfo = new ArrayList<Map.Entry<String, int[]>>(imageBlob.getReshapeInfo().entrySet()).listIterator(imageBlob.getReshapeInfo().size());
         while (reverseReshapeInfo.hasPrevious()) {
             Map.Entry<String, int[]> entry = reverseReshapeInfo.previous();

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_file, "", "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");
@@ -35,8 +35,8 @@ int main(int argc, char** argv) {
     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;
+  if (FLAGS_cfg_file == "") {
+    std::cerr << "--cfg_file need to be defined" << std::endl;
     return -1;
   }
   if (FLAGS_image == "" & FLAGS_image_list == "") {
@@ -44,11 +44,11 @@ int main(int argc, char** argv) {
     return -1;
   }
 
-  // 加载模型
+  // load model
   PaddleX::Model model;
-  model.Init(FLAGS_model_dir, FLAGS_cfg_dir, FLAGS_device);
+  model.Init(FLAGS_model_dir, FLAGS_cfg_file, FLAGS_device);
 
-  // 进行预测
+  // predict
   if (FLAGS_image_list != "") {
     std::ifstream inf(FLAGS_image_list);
     if (!inf) {

deploy/openvino/demo/detector.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 <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_file, "", "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_file == "") {
+    std::cerr << "--cfg_file 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;
+  }
+
+  // load model
+  PaddleX::Model model;
+  model.Init(FLAGS_model_dir, FLAGS_cfg_file, FLAGS_device);
+
+  int imgs = 1;
+  auto colormap = PaddleX::GenerateColorMap(model.labels.size());
+  // predict
+  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 != "") {
+    // visualize
+      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_file, "", "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_file == "") {
+    std::cerr << "--cfg_file 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;
+  }
+
+  // load model
+  PaddleX::Model model;
+  model.Init(FLAGS_model_dir, FLAGS_cfg_file, FLAGS_device);
+  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;
+    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);
+    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, 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/config_parser.h

@@ -54,4 +54,4 @@ class ConfigPaser {
   YAML::Node Transforms_;
 };
 
-}  // namespace PaddleDetection
+}  // namespace PaddleX

deploy/openvino/include/paddlex/paddlex.h

@@ -17,6 +17,8 @@
 #include <functional>
 #include <iostream>
 #include <numeric>
+#include <map>
+#include <string>
 
 #include "yaml-cpp/yaml.h"
 
@@ -30,35 +32,40 @@
 #include "include/paddlex/config_parser.h"
 #include "include/paddlex/results.h"
 #include "include/paddlex/transforms.h"
-using namespace InferenceEngine;
+
 
 namespace PaddleX {
 
 class Model {
  public:
   void Init(const std::string& model_dir,
-            const std::string& cfg_dir,
+            const std::string& cfg_file,
             std::string device) {
-    create_predictor(model_dir, cfg_dir,  device);
+    create_predictor(model_dir, cfg_file,  device);
   }
 
   void create_predictor(const std::string& model_dir,
-                        const std::string& cfg_dir,
+                        const std::string& cfg_file,
                         std::string device);
 
   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_;
-  Blob::Ptr output_;
-  CNNNetwork network_;
-  ExecutableNetwork executable_network_;
+  ImageBlob inputs_;
+  InferenceEngine::Blob::Ptr output_;
+  InferenceEngine::CNNNetwork network_;
+  InferenceEngine::ExecutableNetwork executable_network_;
 };
-}  // namespce of PaddleX
+}  // namespace PaddleX

deploy/openvino/include/paddlex/results.h

@@ -61,11 +61,11 @@ 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();
     score_map.clear();
   }
 };
-}  // namespce of PaddleX
+}  // namespace PaddleX

deploy/openvino/include/paddlex/transforms.h

@@ -16,26 +16,54 @@
 
 #include <yaml-cpp/yaml.h>
 
-#include <memory>
-#include <string>
 #include <unordered_map>
 #include <utility>
+#include <memory>
+#include <string>
 #include <vector>
+#include <iostream>
 
 #include <opencv2/core/core.hpp>
 #include <opencv2/highgui/highgui.hpp>
 #include <opencv2/imgproc/imgproc.hpp>
-
 #include <inference_engine.hpp>
-using namespace InferenceEngine;
+
 
 namespace PaddleX {
 
+/*
+ * @brief
+ * This class represents object for storing all preprocessed data
+ * */
+class ImageBlob {
+ public:
+  // Original image height and width
+  InferenceEngine::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
+  InferenceEngine::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 +73,7 @@ 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_;
@@ -61,8 +89,8 @@ class ResizeByShort : public Transform {
     } else {
       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 +98,55 @@ 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 +160,65 @@ 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/__init__.py

+# 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.

deploy/openvino/python/convertor.py

+# 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.
+
+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:[w,h]")
+    parser.add_argument(
+        "--data_type",
+        "-dp",
+        default="FP32",
+        help="option, FP32 or FP16, the data_type of openvino IR")
+    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")
+        
+    import x2paddle.convert as x2pc
+    x2pc.paddle2onnx(args.model_dir, 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:]
+    if model.__class__.__name__ == "YOLOV3":
+        shape = shape + ",[1,2]"
+        inputs = "image,im_size"
+        onnx_parser.set_defaults(input = inputs)
+    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 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.
+
+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_file",
+        "-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_file
+
+    #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 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.
+
+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
+
+
+class Predictor:
+    def __init__(self, model_xml, model_yaml, device="CPU"):
+        self.device = device
+        if not osp.exists(model_xml):
+            print("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):
+            print("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
+        print("Creating Inference Engine")
+        ie = IECore()
+        print("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, config=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
+        print("Starting inference in synchronous mode")
+        res = self.predictor.infer(inputs=feed_dict)
+
+        #Processing output blob
+        print("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()
+            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)
+        score_map = np.squeeze(preds[score_name])
+        score_map = np.transpose(score_map, (1, 2, 0))
+        label_name = next(it)
+        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 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.
+
+from . import cls_transforms
+from . import det_transforms
+from . import seg_transforms

deploy/openvino/python/transforms/cls_transforms.py

+# 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.
+
+from .ops import *
+import random
+import os.path as osp
+import numpy as np
+from PIL import Image, ImageEnhance
+
+
+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
+
+    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:
+            outputs = op(im, label)
+            im = outputs[0]
+            if len(outputs) == 2:
+                label = outputs[1]
+        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:
+                print(
+                    "{} is already in ComposedTransforms, need to remove it from add_augmenters().".
+                    format(type(aug).__name__))
+        self.transforms = augmenters + self.transforms
+
+
+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 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':
+            pass
+        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 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.
+
+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 .ops import *
+
+
+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
+
+    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
+            outputs = op(im, im_info, label_info)
+            im = outputs[0]
+        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:
+                print(
+                    "{} 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 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 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':
+            pass
+        elif self.mode == 'eval':
+            pass
+        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 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
+            pass
+        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/ops.py

+# 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.
+
+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 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.
+
+from .ops import *
+import random
+import os.path as osp
+import numpy as np
+from PIL import Image
+import cv2
+from collections import OrderedDict
+
+
+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
+
+
+    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:
+                print("{} 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 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
+            pass
+        else:
+            # 验证/预测时的transforms
+            transforms = [Normalize(mean=mean, std=std)]
+
+        super(ComposedSegTransforms, self).__init__(transforms)

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

@@ -13,28 +13,47 @@
 // limitations under the License.
 
 #include "include/paddlex/paddlex.h"
+#include <iostream>
+#include <fstream>
 
-using namespace InferenceEngine;
 
 namespace PaddleX {
 
 void Model::create_predictor(const std::string& model_dir,
-                            const std::string& cfg_dir,
+                            const std::string& cfg_file,
                             std::string device) {
-    Core ie;
-    network_ = ie.ReadNetwork(model_dir, model_dir.substr(0, model_dir.size() - 4) + ".bin");
+    InferenceEngine::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);
+    InferenceEngine::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(InferenceEngine::Precision::FP32);
+        inputInfoItem.second->getPreProcess().setResizeAlgorithm(
+          InferenceEngine::RESIZE_BILINEAR);
+        inputInfoItem.second->setLayout(InferenceEngine::Layout::NCHW);
+      }
+      if (inputInfoItem.second->getTensorDesc().getDims().size() == 2) {
+        imageInputName = inputInfoItem.first;
+        inputInfoItem.second->setPrecision(InferenceEngine::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);
+    }
+    load_config(cfg_file);
 }
 
-bool Model::load_config(const std::string& cfg_dir) {
-  YAML::Node config = YAML::LoadFile(cfg_dir);
+bool Model::load_config(const std::string& cfg_file) {
+  YAML::Node config = YAML::LoadFile(cfg_file);
   type = config["_Attributes"]["model_type"].as<std::string>();
   name = config["Model"].as<std::string>();
   bool to_rgb = true;
@@ -48,22 +67,26 @@ bool Model::load_config(const std::string& cfg_dir) {
       return false;
     }
   }
-  // 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>>();
+  // init preprocess ops
+  transforms_.Init(config["Transforms"], type, to_rgb);
+  // read label list
+  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()!"
@@ -75,34 +98,221 @@ bool Model::predict(const cv::Mat& im, ClsResult* result) {
               << std::endl;
     return false;
   }
-  // 处理输入图像
-  InferRequest infer_request = executable_network_.CreateInferRequest();
+  // preprocess
+  InferenceEngine::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;
   }
 
+  // predict
   infer_request.Infer();
 
   std::string output_name = network_.getOutputsInfo().begin()->first;
   output_ = infer_request.GetBlob(output_name);
-  MemoryBlob::CPtr moutput = as<MemoryBlob>(output_);
+  InferenceEngine::MemoryBlob::CPtr moutput =
+    InferenceEngine::as<InferenceEngine::MemoryBlob>(output_);
   auto moutputHolder = moutput->rmap();
   float* outputs_data = moutputHolder.as<float *>();
 
-  // 对模型输出结果进行后处理
+  // post process
   auto ptr = std::max_element(outputs_data, outputs_data+sizeof(outputs_data));
   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;
-  //    }
+  return true;
+}
+
+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;
+  }
+  InferenceEngine::InferRequest infer_request =
+    executable_network_.CreateInferRequest();
+  InferenceEngine::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();
+
+  InferenceEngine::OutputsDataMap out_map = network_.getOutputsInfo();
+  auto iter = out_map.begin();
+  std::string outputName = iter->first;
+  InferenceEngine::Blob::Ptr output = infer_request.GetBlob(outputName);
+  InferenceEngine::MemoryBlob::CPtr moutput =
+    InferenceEngine::as<InferenceEngine::MemoryBlob>(output);
+  InferenceEngine::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));
+    }
+  }
 }
 
-}  // namespce of PaddleX
+
+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;
+  }
+  // init infer
+  InferenceEngine::InferRequest infer_request =
+    executable_network_.CreateInferRequest();
+  std::string input_name = network_.getInputsInfo().begin()->first;
+  inputs_.blob = infer_request.GetBlob(input_name);
+
+  // preprocess
+  cv::Mat im_clone = im.clone();
+  if (!preprocess(&im_clone, &inputs_)) {
+    std::cerr << "Preprocess failed!" << std::endl;
+    return false;
+  }
+
+  // predict
+  infer_request.Infer();
+
+  InferenceEngine::OutputsDataMap out_map = network_.getOutputsInfo();
+  auto iter = out_map.begin();
+  iter++;
+  std::string output_name_score = iter->first;
+  InferenceEngine::Blob::Ptr output_score =
+    infer_request.GetBlob(output_name_score);
+  InferenceEngine::MemoryBlob::CPtr moutput_score =
+    InferenceEngine::as<InferenceEngine::MemoryBlob>(output_score);
+  InferenceEngine::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;
+  InferenceEngine::Blob::Ptr output_label =
+    infer_request.GetBlob(output_name_label);
+  InferenceEngine::MemoryBlob::CPtr moutput_label =
+    InferenceEngine::as<InferenceEngine::MemoryBlob>(output_label);
+  InferenceEngine::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;
+}
+}  // namespace PaddleX

deploy/openvino/src/transforms.cpp

@@ -12,11 +12,15 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include "include/paddlex/transforms.h"
+
+#include <math.h>
+
 #include <iostream>
+#include <fstream>
 #include <string>
 #include <vector>
 
-#include "include/paddlex/transforms.h"
 
 namespace PaddleX {
 
@@ -26,7 +30,7 @@ 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 +44,6 @@ 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 +61,115 @@ 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;
@@ -105,27 +200,38 @@ std::shared_ptr<Transform> Transforms::CreateTransform(
   }
 }
 
-bool Transforms::Run(cv::Mat* im, Blob::Ptr blob) {
-  // 按照transforms中预处理算子顺序处理图像
+bool Transforms::Run(cv::Mat* im, ImageBlob* data) {
+  // preprocess by order
   if (to_rgb_) {
     cv::cvtColor(*im, *im, cv::COLOR_BGR2RGB);
   }
   (*im).convertTo(*im, CV_32FC3);
+  if (type_ == "detector") {
+    InferenceEngine::LockedMemory<void> input2Mapped =
+      InferenceEngine::as<InferenceEngine::MemoryBlob>(
+        data->ori_im_size_)->wmap();
+    float *p = input2Mapped.as<float*>();
+    p[0] = im->rows;
+    p[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;
     }
   }
 
-  // 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();
+  // image format NHWC to NCHW
+  // img data save to ImageBlob
+  InferenceEngine::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);
+  InferenceEngine::MemoryBlob::Ptr mblob =
+    InferenceEngine::as<InferenceEngine::MemoryBlob>(data->blob);
   auto mblobHolder = mblob->wmap();
   float *blob_data = mblobHolder.as<float *>();
   for (size_t c = 0; c < channels; c++) {

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]);
+
+    // draw box and title
+    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;
+
+    // background
+    cv::Rect text_back = cv::Rect(boxes[i].coordinate[0],
+                                  boxes[i].coordinate[1] - text_size.height,
+                                  text_size.width,
+                                  text_size.height);
+
+    // draw
+    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);
+
+    // 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

deploy/raspberry/CMakeLists.txt

+cmake_minimum_required(VERSION 3.0)
+project(PaddleX CXX C)
+
+
+option(WITH_STATIC_LIB "Compile demo with static/shared library, default use static."   OFF)
+
+SET(CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake" ${CMAKE_MODULE_PATH})
+SET(LITE_DIR "" CACHE PATH "Location of libraries")
+SET(OPENCV_DIR "" CACHE PATH "Location of libraries")
+SET(NGRAPH_LIB "" CACHE PATH "Location of libraries")
+
+
+include(cmake/yaml-cpp.cmake)
+
+include_directories("${CMAKE_SOURCE_DIR}/")
+link_directories("${CMAKE_CURRENT_BINARY_DIR}")
+include_directories("${CMAKE_CURRENT_BINARY_DIR}/ext/yaml-cpp/src/ext-yaml-cpp/include")
+link_directories("${CMAKE_CURRENT_BINARY_DIR}/ext/yaml-cpp/lib")
+
+macro(safe_set_static_flag)
+    foreach(flag_var
+        CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE
+        CMAKE_CXX_FLAGS_MINSIZEREL CMAKE_CXX_FLAGS_RELWITHDEBINFO)
+      if(${flag_var} MATCHES "/MD")
+        string(REGEX REPLACE "/MD" "/MT" ${flag_var} "${${flag_var}}")
+      endif(${flag_var} MATCHES "/MD")
+    endforeach(flag_var)
+endmacro()
+
+if (NOT DEFINED LITE_DIR OR ${LITE_DIR} STREQUAL "")
+	message(FATAL_ERROR "please set LITE_DIR with -LITE_DIR=/path/influence_engine")
+endif()
+
+if (NOT DEFINED OPENCV_DIR OR ${OPENCV_DIR} STREQUAL "")
+    message(FATAL_ERROR "please set OPENCV_DIR with -DOPENCV_DIR=/path/opencv")
+endif()
+
+if (NOT DEFINED GFLAGS_DIR OR ${GFLAGS_DIR} STREQUAL "")
+    message(FATAL_ERROR "please set GFLAGS_DIR with -DGFLAGS_DIR=/path/gflags")
+endif()
+
+
+
+
+
+link_directories("${LITE_DIR}/lib")
+include_directories("${LITE_DIR}/include")
+
+
+
+link_directories("${GFLAGS_DIR}/lib")
+include_directories("${GFLAGS_DIR}/include")
+
+
+
+
+if (WIN32)
+  find_package(OpenCV REQUIRED PATHS ${OPENCV_DIR}/build/ NO_DEFAULT_PATH)
+  unset(OpenCV_DIR CACHE)
+else ()
+  find_package(OpenCV REQUIRED PATHS ${OPENCV_DIR}/cmake NO_DEFAULT_PATH)
+endif ()
+
+include_directories(${OpenCV_INCLUDE_DIRS})
+
+if (WIN32)
+    add_definitions("/DGOOGLE_GLOG_DLL_DECL=")
+    set(CMAKE_C_FLAGS_DEBUG   "${CMAKE_C_FLAGS_DEBUG} /bigobj /MTd")
+    set(CMAKE_C_FLAGS_RELEASE  "${CMAKE_C_FLAGS_RELEASE} /bigobj /MT")
+    set(CMAKE_CXX_FLAGS_DEBUG  "${CMAKE_CXX_FLAGS_DEBUG} /bigobj /MTd")
+    set(CMAKE_CXX_FLAGS_RELEASE   "${CMAKE_CXX_FLAGS_RELEASE} /bigobj /MT")
+    if (WITH_STATIC_LIB)
+        safe_set_static_flag()
+        add_definitions(-DSTATIC_LIB)
+    endif()
+else()
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mfloat-abi=hard -mfpu=neon-vfpv4 -g -o2 -fopenmp -std=c++11")
+    set(CMAKE_STATIC_LIBRARY_PREFIX "")
+endif()
+
+
+if(WITH_STATIC_LIB)
+	set(DEPS ${LITE_DIR}/lib/libpaddle_full_api_shared${CMAKE_STATIC_LIBRARY_SUFFIX})
+else()
+	set(DEPS ${LITE_DIR}/lib/libpaddle_full_api_shared${CMAKE_SHARED_LIBRARY_SUFFIX})
+endif()
+
+if (NOT WIN32)
+    set(DEPS ${DEPS}
+        glog gflags  z  yaml-cpp
+        )
+else()
+    set(DEPS ${DEPS}
+        glog gflags_static libprotobuf zlibstatic xxhash libyaml-cppmt)
+    set(DEPS ${DEPS} libcmt shlwapi)
+endif(NOT WIN32)
+
+
+if (NOT WIN32)
+    set(EXTERNAL_LIB "-ldl -lrt -lgomp -lz -lm -lpthread")
+    set(DEPS ${DEPS} ${EXTERNAL_LIB})
+endif()
+
+set(DEPS ${DEPS} ${OpenCV_LIBS})
+add_executable(classifier 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/raspberry/cmake/yaml-cpp.cmake

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

deploy/raspberry/demo/classifier.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 <fstream>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "include/paddlex/paddlex.h"
+
+DEFINE_string(model_dir, "", "Path of inference model");
+DEFINE_string(cfg_file, "", "Path of PaddelX model yml file");
+DEFINE_string(image, "", "Path of test image file");
+DEFINE_string(image_list, "", "Path of test image list file");
+DEFINE_int32(thread_num, 1, "num of thread to infer");
+
+int main(int argc, char** argv) {
+  // Parsing command-line
+  google::ParseCommandLineFlags(&argc, &argv, true);
+
+  if (FLAGS_model_dir == "") {
+    std::cerr << "--model_dir need to be defined" << std::endl;
+    return -1;
+  }
+  if (FLAGS_cfg_file == "") {
+    std::cerr << "--cfg_flie 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;
+  }
+
+  // load model
+  PaddleX::Model model;
+  model.Init(FLAGS_model_dir, FLAGS_cfg_file, FLAGS_thread_num);
+  std::cout << "init is done" << std::endl;
+  // predict
+  if (FLAGS_image_list != "") {
+    std::ifstream inf(FLAGS_image_list);
+    if (!inf) {
+      std::cerr << "Fail to open file " << FLAGS_image_list << std::endl;
+      return -1;
+    }
+    std::string image_path;
+
+    while (getline(inf, image_path)) {
+      PaddleX::ClsResult result;
+      cv::Mat im = cv::imread(image_path, 1);
+      model.predict(im, &result);
+      std::cout << "Predict label: " << result.category
+                << ", label_id:" << result.category_id
+                << ", score: " << result.score << std::endl;
+    }
+  } else {
+    PaddleX::ClsResult result;
+    cv::Mat im = cv::imread(FLAGS_image, 1);
+    model.predict(im, &result);
+    std::cout << "Predict label: " << result.category
+              << ", label_id:" << result.category_id
+              << ", score: " << result.score << std::endl;
+  }
+
+  return 0;
+}

deploy/raspberry/demo/detector.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 <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_file, "", "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_int32(thread_num, 1, "num of thread to infer");
+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_file == "") {
+    std::cerr << "--cfg_file 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;
+  }
+
+  // load model
+  PaddleX::Model model;
+  model.Init(FLAGS_model_dir, FLAGS_cfg_file, FLAGS_thread_num);
+
+  int imgs = 1;
+  auto colormap = PaddleX::GenerateColorMap(model.labels.size());
+  // predict
+  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 != "") {
+    // visualize
+      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/raspberry/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_file, "", "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(save_dir, "", "Path to save visualized image");
+DEFINE_int32(batch_size, 1, "Batch size of infering");
+DEFINE_int32(thread_num, 1, "num of thread to infer");
+
+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_file == "") {
+    std::cerr << "--cfg_file 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;
+  }
+
+  // load model
+  std::cout << "init start" << std::endl;
+  PaddleX::Model model;
+  model.Init(FLAGS_model_dir, FLAGS_cfg_file, FLAGS_thread_num);
+  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;
+
+    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);
+    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, FLAGS_image);
+      cv::imwrite(save_path, vis_img);
+      std::cout << "Visualized` output saved as " << save_path << std::endl;
+    }
+    result.clear();
+  }
+  return 0;
+}

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

deploy/raspberry/include/paddlex/paddlex.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 <arm_neon.h>
+#include <paddle_api.h>
+
+#include <functional>
+#include <iostream>
+#include <numeric>
+#include <map>
+#include <string>
+#include <memory>
+
+#include "include/paddlex/config_parser.h"
+#include "include/paddlex/results.h"
+#include "include/paddlex/transforms.h"
+
+
+
+#include "yaml-cpp/yaml.h"
+
+
+
+
+#ifdef _WIN32
+#define OS_PATH_SEP "\\"
+#else
+#define OS_PATH_SEP "/"
+#endif
+
+
+
+
+namespace PaddleX {
+
+class Model {
+ public:
+  void Init(const std::string& model_dir,
+            const std::string& cfg_file,
+            int thread_num) {
+    create_predictor(model_dir, cfg_file, thread_num);
+  }
+
+  void create_predictor(const std::string& model_dir,
+                        const std::string& cfg_file,
+                        int thread_num);
+
+  bool load_config(const std::string& model_dir);
+
+  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::map<int, std::string> labels;
+  Transforms transforms_;
+  ImageBlob inputs_;
+  std::shared_ptr<paddle::lite_api::PaddlePredictor> predictor_;
+};
+}  // namespace PaddleX

deploy/raspberry/include/paddlex/results.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 <string>
+#include <vector>
+
+namespace PaddleX {
+
+template <class T>
+struct Mask {
+  std::vector<T> data;
+  std::vector<int> shape;
+  void clear() {
+    data.clear();
+    shape.clear();
+  }
+};
+
+struct Box {
+  int category_id;
+  std::string category;
+  float score;
+  std::vector<float> coordinate;
+  Mask<float> mask;
+};
+
+class BaseResult {
+ public:
+  std::string type = "base";
+};
+
+class ClsResult : public BaseResult {
+ public:
+  int category_id;
+  std::string category;
+  float score;
+  std::string type = "cls";
+};
+
+class DetResult : public BaseResult {
+ public:
+  std::vector<Box> boxes;
+  int mask_resolution;
+  std::string type = "det";
+  void clear() { boxes.clear(); }
+};
+
+class SegResult : public BaseResult {
+ public:
+  Mask<int64_t> label_map;
+  Mask<float> score_map;
+  void clear() {
+    label_map.clear();
+    score_map.clear();
+  }
+};
+}  // namespace PaddleX

deploy/raspberry/include/paddlex/transforms.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 <yaml-cpp/yaml.h>
+#include <paddle_api.h>
+
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+#include <iostream>
+
+#include <opencv2/core/core.hpp>
+#include <opencv2/highgui/highgui.hpp>
+#include <opencv2/imgproc/imgproc.hpp>
+
+
+
+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);
+
+  // 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
+  std::unique_ptr<paddle::lite_api::Tensor> input_tensor_;
+
+  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, ImageBlob* data) = 0;
+};
+
+class Normalize : public Transform {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    mean_ = item["mean"].as<std::vector<float>>();
+    std_ = item["std"].as<std::vector<float>>();
+  }
+
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  std::vector<float> mean_;
+  std::vector<float> std_;
+};
+
+class ResizeByShort : public Transform {
+ public:
+  virtual void Init(const YAML::Node& item) {
+    short_size_ = item["short_size"].as<int>();
+    if (item["max_size"].IsDefined()) {
+      max_size_ = item["max_size"].as<int>();
+    } else {
+      max_size_ = -1;
+    }
+  }
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  float GenerateScale(const cv::Mat& im);
+  int short_size_;
+  int max_size_;
+};
+
+/*
+ * @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:
+  virtual void Init(const YAML::Node& item) {
+    if (item["crop_size"].IsScalar()) {
+      height_ = item["crop_size"].as<int>();
+      width_ = item["crop_size"].as<int>();
+    } else if (item["crop_size"].IsSequence()) {
+      std::vector<int> crop_size = item["crop_size"].as<std::vector<int>>();
+      width_ = crop_size[0];
+      height_ = crop_size[1];
+    }
+  }
+  virtual bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  int height_;
+  int width_;
+};
+
+
+/*
+ * @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);
+  std::shared_ptr<Transform> CreateTransform(const std::string& name);
+  bool Run(cv::Mat* im, ImageBlob* data);
+
+ private:
+  std::vector<std::shared_ptr<Transform>> transforms_;
+  bool to_rgb_ = true;
+};
+
+}  // namespace PaddleX

deploy/raspberry/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/raspberry/python/__init__.py

+# 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.
\ No newline at end of file

deploy/raspberry/python/demo.py

+# 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.
+
+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(
+        "--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_file",
+        "-c",
+        type=str,
+        default=None,
+        help="Path to PaddelX model yml file")
+
+    parser.add_argument(
+        "--thread_num",
+        "-t",
+        type=int,
+        default=1,
+        help="Path to PaddelX model yml file")
+
+    parser.add_argument(
+        "--input_shape",
+        "-ip",
+        type=str,
+        default=None,
+        help=" image input shape of model [NCHW] like [1,3,224,244] ")
+
+    return parser
+
+
+def main():
+    parser = arg_parser()
+    args = parser.parse_args()
+    model_nb = args.model_dir
+    model_yaml = args.cfg_file
+    thread_num = args.thread_num
+    input_shape = args.input_shape
+    input_shape = input_shape[1:-1].split(",", 3)
+    shape = list(map(int, input_shape))
+    #model init
+    predictor = deploy.Predictor(model_nb, model_yaml, thread_num, shape)
+
+    #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/raspberry/python/transforms/__init__.py

+# 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.
+
+from . import cls_transforms
+from . import det_transforms
+from . import seg_transforms

deploy/raspberry/python/transforms/cls_transforms.py

+# 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.
+
+from .ops import *
+import random
+import os.path as osp
+import numpy as np
+from PIL import Image, ImageEnhance
+
+
+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
+
+    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:
+            outputs = op(im, label)
+            im = outputs[0]
+            if len(outputs) == 2:
+                label = outputs[1]
+        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:
+                print(
+                    "{} is already in ComposedTransforms, need to remove it from add_augmenters().".
+                    format(type(aug).__name__))
+        self.transforms = augmenters + self.transforms
+
+
+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 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':
+            pass
+        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/raspberry/python/transforms/det_transforms.py

+# 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.
+
+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 .ops import *
+
+
+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
+
+    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
+            outputs = op(im, im_info, label_info)
+            im = outputs[0]
+        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:
+                print(
+                    "{} 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 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 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':
+            pass
+        elif self.mode == 'eval':
+            pass
+        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 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
+            pass
+        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/raspberry/python/transforms/ops.py

+# 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.
+
+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/raspberry/python/transforms/seg_transforms.py

+# coding: utf8
+# 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.
+
+from .ops import *
+import random
+import os.path as osp
+import numpy as np
+from PIL import Image
+import cv2
+from collections import OrderedDict
+
+
+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
+
+
+    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:
+                print("{} 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 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
+            pass
+        else:
+            # 验证/预测时的transforms
+            transforms = [Normalize(mean=mean, std=std)]
+
+        super(ComposedSegTransforms, self).__init__(transforms)

deploy/raspberry/scripts/build.sh

+# Paddle-Lite预编译库的路径
+LITE_DIR=/path/to/Paddle-Lite/inference/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/opencv
+# 下载自带预编译版本
+exec $(pwd)/scripts/install_third-party.sh
+
+rm -rf build
+mkdir -p build
+cd build
+cmake .. \
+    -DOPENCV_DIR=${OPENCV_DIR} \
+    -DGFLAGS_DIR=${GFLAGS_DIR} \
+    -DLITE_DIR=${LITE_DIR} \
+    -DCMAKE_CXX_FLAGS="-march=armv7-a"  
+make

deploy/raspberry/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 4
+    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 4
+    cd ..
+    cd ..
+fi
+OPENCV_URL=https://bj.bcebos.com/paddlex/deploy/armopencv/opencv.tar.bz2
+if [ ! -d "./deps/opencv" ]; then
+    cd deps
+    wget -c ${OPENCV_URL}
+    tar xvfj opencv.tar.bz2
+    rm -rf opencv.tar.bz2
+    cd ..
+fi

deploy/raspberry/src/paddlex.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/paddlex.h"
+#include <iostream>
+#include <fstream>
+
+
+namespace PaddleX {
+
+void Model::create_predictor(const std::string& model_dir,
+                            const std::string& cfg_file,
+                            int thread_num) {
+  paddle::lite_api::MobileConfig config;
+  config.set_model_from_file(model_dir);
+  config.set_threads(thread_num);
+  load_config(cfg_file);
+  predictor_ =
+    paddle::lite_api::CreatePaddlePredictor<paddle::lite_api::MobileConfig>(
+      config);
+}
+
+bool Model::load_config(const std::string& cfg_file) {
+  YAML::Node config = YAML::LoadFile(cfg_file);
+  type = config["_Attributes"]["model_type"].as<std::string>();
+  name = config["Model"].as<std::string>();
+  bool to_rgb = true;
+  if (config["TransformsMode"].IsDefined()) {
+    std::string mode = config["TransformsMode"].as<std::string>();
+    if (mode == "BGR") {
+      to_rgb = false;
+    } else if (mode != "RGB") {
+      std::cerr << "[Init] Only 'RGB' or 'BGR' is supported for TransformsMode"
+                << std::endl;
+      return false;
+    }
+  }
+  // init preprocess ops
+  transforms_.Init(config["Transforms"], to_rgb);
+  // read label list
+  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, 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()!"
+              << 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;
+  }
+  // preprocess
+  inputs_.input_tensor_ = std::move(predictor_->GetInput(0));
+  cv::Mat im_clone = im.clone();
+  if (!preprocess(&im_clone, &inputs_)) {
+    std::cerr << "Preprocess failed!" << std::endl;
+    return false;
+  }
+  // predict
+  predictor_->Run();
+
+  std::unique_ptr<const paddle::lite_api::Tensor> output_tensor(
+    std::move(predictor_->GetOutput(0)));
+  const float *outputs_data = output_tensor->mutable_data<float>();
+
+
+  // postprocess
+  auto ptr = std::max_element(outputs_data, outputs_data+sizeof(outputs_data));
+  result->category_id = std::distance(outputs_data, ptr);
+  result->score = *ptr;
+  result->category = labels[result->category_id];
+}
+
+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;
+  }
+  inputs_.input_tensor_ = std::move(predictor_->GetInput(0));
+
+  cv::Mat im_clone = im.clone();
+  if (!preprocess(&im_clone, &inputs_)) {
+    std::cerr << "Preprocess failed!" << std::endl;
+    return false;
+  }
+  int h = inputs_.new_im_size_[0];
+  int w = inputs_.new_im_size_[1];
+  if (name == "YOLOv3") {
+    std::unique_ptr<paddle::lite_api::Tensor> im_size_tensor(
+      std::move(predictor_->GetInput(1)));
+    const std::vector<int64_t> IM_SIZE_SHAPE = {1, 2};
+    im_size_tensor->Resize(IM_SIZE_SHAPE);
+    auto *im_size_data = im_size_tensor->mutable_data<int>();
+    memcpy(im_size_data, inputs_.ori_im_size_.data(), 1*2*sizeof(int));
+  }
+  predictor_->Run();
+  auto output_names = predictor_->GetOutputNames();
+  auto output_box_tensor = predictor_->GetTensor(output_names[0]);
+  const float *output_box = output_box_tensor->mutable_data<float>();
+  std::vector<int64_t> output_box_shape = output_box_tensor->shape();
+  int size = 1;
+  for (const auto& i : output_box_shape) {
+    size *= i;
+  }
+  int num_boxes = size / 6;
+  for (int i = 0; i < num_boxes; ++i) {
+    Box box;
+    box.category_id = static_cast<int>(round(output_box[i * 6]));
+    box.category = labels[box.category_id];
+    box.score = output_box[i * 6 + 1];
+    float xmin = output_box[i * 6 + 2];
+    float ymin = output_box[i * 6 + 3];
+    float xmax = output_box[i * 6 + 4];
+    float ymax = output_box[i * 6 + 5];
+    float w = xmax - xmin + 1;
+    float h = ymax - ymin + 1;
+    box.coordinate = {xmin, ymin, w, h};
+    result->boxes.push_back(std::move(box));
+  }
+  return true;
+}
+
+
+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;
+  }
+  inputs_.input_tensor_ = std::move(predictor_->GetInput(0));
+  cv::Mat im_clone = im.clone();
+  if (!preprocess(&im_clone, &inputs_)) {
+    std::cerr << "Preprocess failed!" << std::endl;
+    return false;
+  }
+  std::cout << "Preprocess is done" << std::endl;
+  predictor_->Run();
+  auto output_names = predictor_->GetOutputNames();
+
+  auto output_label_tensor = predictor_->GetTensor(output_names[0]);
+  const int64_t *label_data = output_label_tensor->mutable_data<int64_t>();
+  std::vector<int64_t> output_label_shape = output_label_tensor->shape();
+  int size = 1;
+  for (const auto& i : output_label_shape) {
+    size *= i;
+    result->label_map.shape.push_back(i);
+  }
+  result->label_map.data.resize(size);
+  memcpy(result->label_map.data.data(), label_data, size*sizeof(int64_t));
+
+  auto output_score_tensor = predictor_->GetTensor(output_names[1]);
+  const float *score_data = output_score_tensor->mutable_data<float>();
+  std::vector<int64_t> output_score_shape = output_score_tensor->shape();
+  size = 1;
+  for (const auto& i : output_score_shape) {
+    size *= i;
+    result->score_map.shape.push_back(i);
+  }
+  result->score_map.data.resize(size);
+  memcpy(result->score_map.data.data(), score_data, size*sizeof(float));
+
+
+  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;
+}
+}  // namespace PaddleX

deploy/raspberry/src/transforms.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/transforms.h"
+
+#include <math.h>
+
+#include <iostream>
+#include <string>
+#include <vector>
+
+
+
+namespace PaddleX {
+
+std::map<std::string, int> interpolations = {{"LINEAR", cv::INTER_LINEAR},
+                                             {"NEAREST", cv::INTER_NEAREST},
+                                             {"AREA", cv::INTER_AREA},
+                                             {"CUBIC", cv::INTER_CUBIC},
+                                             {"LANCZOS4", cv::INTER_LANCZOS4}};
+
+bool Normalize::Run(cv::Mat* im, ImageBlob* data) {
+  for (int h = 0; h < im->rows; h++) {
+    for (int w = 0; w < im->cols; w++) {
+      im->at<cv::Vec3f>(h, w)[0] =
+          (im->at<cv::Vec3f>(h, w)[0] / 255.0 - mean_[0]) / std_[0];
+      im->at<cv::Vec3f>(h, w)[1] =
+          (im->at<cv::Vec3f>(h, w)[1] / 255.0 - mean_[1]) / std_[1];
+      im->at<cv::Vec3f>(h, w)[2] =
+          (im->at<cv::Vec3f>(h, w)[2] / 255.0 - mean_[2]) / std_[2];
+    }
+  }
+  return true;
+}
+
+
+
+float ResizeByShort::GenerateScale(const cv::Mat& im) {
+  int origin_w = im.cols;
+  int origin_h = im.rows;
+  int im_size_max = std::max(origin_w, origin_h);
+  int im_size_min = std::min(origin_w, origin_h);
+  float scale =
+      static_cast<float>(short_size_) / static_cast<float>(im_size_min);
+  if (max_size_ > 0) {
+    if (round(scale * im_size_max) > max_size_) {
+      scale = static_cast<float>(max_size_) / static_cast<float>(im_size_max);
+    }
+  }
+  return scale;
+}
+
+bool ResizeByShort::Run(cv::Mat* im, ImageBlob* data) {
+  data->im_size_before_resize_.push_back({im->rows, im->cols});
+  data->reshape_order_.push_back("resize");
+
+  float scale = GenerateScale(*im);
+  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;
+}
+
+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, bool to_rgb) {
+  transforms_.clear();
+  to_rgb_ = to_rgb;
+  for (const auto& item : transforms_node) {
+    std::string name = item.begin()->first.as<std::string>();
+    std::cout << "trans name: " << name << std::endl;
+    std::shared_ptr<Transform> transform = CreateTransform(name);
+    transform->Init(item.begin()->second);
+    transforms_.push_back(transform);
+  }
+}
+
+std::shared_ptr<Transform> Transforms::CreateTransform(
+    const std::string& transform_name) {
+  if (transform_name == "Normalize") {
+    return std::make_shared<Normalize>();
+  } else if (transform_name == "ResizeByShort") {
+    return std::make_shared<ResizeByShort>();
+  } else if (transform_name == "CenterCrop") {
+    return std::make_shared<CenterCrop>();
+  } else if (transform_name == "Resize") {
+    return std::make_shared<Resize>();
+  } else if (transform_name == "Padding") {
+    return std::make_shared<Padding>();
+  } else if (transform_name == "ResizeByLong") {
+    return std::make_shared<ResizeByLong>();
+  } else {
+    std::cerr << "There's unexpected transform(name='" << transform_name
+              << "')." << std::endl;
+    exit(-1);
+  }
+}
+
+bool Transforms::Run(cv::Mat* im, ImageBlob* data) {
+  // preprocess by order
+  if (to_rgb_) {
+    cv::cvtColor(*im, *im, cv::COLOR_BGR2RGB);
+  }
+  (*im).convertTo(*im, CV_32FC3);
+  data->ori_im_size_[0] = im->rows;
+  data->ori_im_size_[1] = im->cols;
+  data->new_im_size_[0] = im->rows;
+  data->new_im_size_[1] = im->cols;
+
+  for (int i = 0; i < transforms_.size(); ++i) {
+    if (!transforms_[i]->Run(im, data)) {
+      std::cerr << "Apply transforms to image failed!" << std::endl;
+      return false;
+    }
+  }
+
+  // image format NHWC to NCHW
+  // img data save to ImageBlob
+  int height = im->rows;
+  int width = im->cols;
+  int channels = im->channels();
+  const std::vector<int64_t> INPUT_SHAPE = {1, channels, height, width};
+  data->input_tensor_->Resize(INPUT_SHAPE);
+  auto *input_data = data->input_tensor_->mutable_data<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++) {
+              input_data[c * width * height + h * width + w] =
+                      im->at<cv::Vec3f>(h, w)[c];
+          }
+      }
+  }
+  return true;
+}
+}  // namespace PaddleX

deploy/raspberry/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

docs/apis/deploy.md

@@ -45,7 +45,7 @@ predict(image, topk=1)
 
 ### batch_predict 接口
 ```
-batch_predict(image_list, topk=1, thread_num=2)
+batch_predict(image_list, topk=1)
 ```
 批量图片预测接口。
 
@@ -53,4 +53,3 @@ batch_predict(image_list, topk=1, thread_num=2)
 >
 > > * **image_list** (list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是图像路径或numpy数组(HWC排列，BGR格式)。
 > > * **topk** (int): 图像分类时使用的参数，表示预测前topk个可能的分类。
-> > * **thread_num** (int): 并发执行各图像预处理时的线程数。

docs/apis/models/classification.md

@@ -62,7 +62,7 @@ evaluate(self, eval_dataset, batch_size=1, epoch_id=None, return_details=False)
 ### predict
 
 ```python
-predict(self, img_file, transforms=None, topk=5)
+predict(self, img_file, transforms=None, topk=1)
 ```
 
 > 分类模型预测接口。需要注意的是，只有在训练过程中定义了eval_dataset，模型在保存时才会将预测时的图像处理流程保存在`ResNet50.test_transforms`和`ResNet50.eval_transforms`中。如未在训练时定义eval_dataset，那在调用预测`predict`接口时，用户需要再重新定义test_transforms传入给`predict`接口。
@@ -81,7 +81,7 @@ predict(self, img_file, transforms=None, topk=5)
 ### batch_predict
 
 ```python
-batch_predict(self, img_file_list, transforms=None, topk=5, thread_num=2)
+batch_predict(self, img_file_list, transforms=None, topk=1)
 ```
 
 > 分类模型批量预测接口。需要注意的是，只有在训练过程中定义了eval_dataset，模型在保存时才会将预测时的图像处理流程保存在`ResNet50.test_transforms`和`ResNet50.eval_transforms`中。如未在训练时定义eval_dataset，那在调用预测`batch_predict`接口时，用户需要再重新定义test_transforms传入给`batch_predict`接口。
@@ -91,7 +91,6 @@ batch_predict(self, img_file_list, transforms=None, topk=5, thread_num=2)
 > > - **img_file_list** (list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是图像路径或numpy数组(HWC排列，BGR格式)。
 > > - **transforms** (paddlex.cls.transforms): 数据预处理操作。
 > > - **topk** (int): 预测时前k个最大值。
-> > - **thread_num** (int): 并发执行各图像预处理时的线程数。
 
 > **返回值**
 >

docs/apis/models/detection.md

@@ -108,7 +108,7 @@ predict(self, img_file, transforms=None)
 ### batch_predict
 
 ```python
-batch_predict(self, img_file_list, transforms=None, thread_num=2)
+batch_predict(self, img_file_list, transforms=None)
 ```
 
 > PPYOLO模型批量预测接口。需要注意的是，只有在训练过程中定义了eval_dataset，模型在保存时才会将预测时的图像处理流程保存在`YOLOv3.test_transforms`和`YOLOv3.eval_transforms`中。如未在训练时定义eval_dataset，那在调用预测`batch_predict`接口时，用户需要再重新定义`test_transforms`传入给`batch_predict`接口
@@ -117,7 +117,6 @@ batch_predict(self, img_file_list, transforms=None, thread_num=2)
 >
 > > - **img_file_list** (str|np.ndarray): 对列表（或元组）中的图像同时进行预测，列表中的元素是预测图像路径或numpy数组(HWC排列，BGR格式)。
 > > - **transforms** (paddlex.det.transforms): 数据预处理操作。
-> > - **thread_num** (int): 并发执行各图像预处理时的线程数。
 >
 > **返回值**
 >
@@ -222,7 +221,7 @@ predict(self, img_file, transforms=None)
 ### batch_predict
 
 ```python
-batch_predict(self, img_file_list, transforms=None, thread_num=2)
+batch_predict(self, img_file_list, transforms=None)
 ```
 
 > YOLOv3模型批量预测接口。需要注意的是，只有在训练过程中定义了eval_dataset，模型在保存时才会将预测时的图像处理流程保存在`YOLOv3.test_transforms`和`YOLOv3.eval_transforms`中。如未在训练时定义eval_dataset，那在调用预测`batch_predict`接口时，用户需要再重新定义`test_transforms`传入给`batch_predict`接口
@@ -231,7 +230,6 @@ batch_predict(self, img_file_list, transforms=None, thread_num=2)
 >
 > > - **img_file_list** (str|np.ndarray): 对列表（或元组）中的图像同时进行预测，列表中的元素是预测图像路径或numpy数组(HWC排列，BGR格式)。
 > > - **transforms** (paddlex.det.transforms): 数据预处理操作。
-> > - **thread_num** (int): 并发执行各图像预处理时的线程数。
 >
 > **返回值**
 >
@@ -327,7 +325,7 @@ predict(self, img_file, transforms=None)
 ### batch_predict
 
 ```python
-batch_predict(self, img_file_list, transforms=None, thread_num=2)
+batch_predict(self, img_file_list, transforms=None)
 ```
 
 > FasterRCNN模型批量预测接口。需要注意的是，只有在训练过程中定义了eval_dataset，模型在保存时才会将预测时的图像处理流程保存在`FasterRCNN.test_transforms`和`FasterRCNN.eval_transforms`中。如未在训练时定义eval_dataset，那在调用预测`batch_predict`接口时，用户需要再重新定义test_transforms传入给`batch_predict`接口。
@@ -336,7 +334,6 @@ batch_predict(self, img_file_list, transforms=None, thread_num=2)
 >
 > > - **img_file_list** (list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素是预测图像路径或numpy数组(HWC排列，BGR格式)。
 > > - **transforms** (paddlex.det.transforms): 数据预处理操作。
-> > - **thread_num** (int): 并发执行各图像预处理时的线程数。
 >
 > **返回值**
 >

docs/apis/models/instance_segmentation.md

@@ -88,7 +88,7 @@ predict(self, img_file, transforms=None)
 #### batch_predict
 
 ```python
-batch_predict(self, img_file_list, transforms=None, thread_num=2)
+batch_predict(self, img_file_list, transforms=None)
 ```
 
 > MaskRCNN模型批量预测接口。需要注意的是，只有在训练过程中定义了eval_dataset，模型在保存时才会将预测时的图像处理流程保存在`FasterRCNN.test_transforms`和`FasterRCNN.eval_transforms`中。如未在训练时定义eval_dataset，那在调用预测`batch_predict`接口时，用户需要再重新定义test_transforms传入给`batch_predict`接口。
@@ -97,7 +97,6 @@ batch_predict(self, img_file_list, transforms=None, thread_num=2)
 >
 > > - **img_file_list** (list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是预测图像路径或numpy数组(HWC排列，BGR格式)。
 > > - **transforms** (paddlex.det.transforms): 数据预处理操作。
-> > - **thread_num** (int): 并发执行各图像预处理时的线程数。
 >
 > **返回值**
 >

docs/apis/models/semantic_segmentation.md

@@ -95,7 +95,7 @@ predict(self, img_file, transforms=None):
 ### batch_predict
 
 ```
-batch_predict(self, img_file_list, transforms=None, thread_num=2):
+batch_predict(self, img_file_list, transforms=None):
 ```
 
 > DeepLabv3p模型批量预测接口。需要注意的是，只有在训练过程中定义了eval_dataset，模型在保存时才会将预测时的图像处理流程保存在`DeepLabv3p.test_transforms`和`DeepLabv3p.eval_transforms`中。如未在训练时定义eval_dataset，那在调用预测`batch_predict`接口时，用户需要再重新定义test_transforms传入给`batch_predict`接口。
@@ -104,7 +104,6 @@ batch_predict(self, img_file_list, transforms=None, thread_num=2):
 > >
 > > - **img_file_list** (list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是预测图像路径或numpy数组(HWC排列，BGR格式)。
 > > - **transforms** (paddlex.seg.transforms): 数据预处理操作。
-> > - **thread_num** (int): 并发执行各图像预处理时的线程数。
 
 > **返回值**
 > >

docs/deploy/openvino/export_openvino_model.md

+# OpenVINO模型转换
+将Paddle模型转换为OpenVINO的Inference Engine  
+
+## 环境依赖
+
+* ONNX 1.5.0+
+* PaddleX 1.0+
+* OpenVINO 2020.4
+
+**说明**：PaddleX安装请参考[PaddleX](https://paddlex.readthedocs.io/zh_CN/develop/install.html) ， OpenVINO安装请参考[OpenVINO](https://docs.openvinotoolkit.org/latest/index.html)，ONNX请安装1.5.0以上版本否则会出现转模型错误。
+
+请确保系统已经安装好上述基本软件，**下面所有示例以工作目录 `/root/projects/`演示**。
+
+## 导出inference模型
+paddle模型转openvino之前需要先把paddle模型导出为inference格式模型，导出的模型将包括__model__、__params__和model.yml三个文件名，导出命令如下
+```
+paddlex --export_inference --model_dir=/path/to/paddle_model --save_dir=./inference_model --fixed_input_shape=[w,h]
+```
+
+## 导出OpenVINO模型
+
+```
+cd /root/projects/python
+
+python convertor.py --model_dir /path/to/inference_model --save_dir /path/to/openvino_model --fixed_input_shape [w,h]
+```
+**转换成功后会在save_dir下出现后缀名为.xml、.bin、.mapping三个文件**  
+转换参数说明如下：
+
+|  参数   | 说明  |
+|  ----  | ----  |
+| --model_dir  | Paddle模型路径，请确保__model__, \_\_params__model.yml在同一个目录|
+| --save_dir  | OpenVINO模型保存路径 |
+| --fixed_input_shape  | 模型输入的[W,H] |
+| --data type(option)  | FP32、FP16，默认为FP32，VPU下的IR需要为FP16 |  
+**注意**：
+- 由于OpenVINO不支持ONNX的resize-11 OP的原因，目前还不支持Paddle的分割模型
+- YOLOv3在通过OpenVINO部署时，由于OpenVINO对ONNX OP的支持限制，我们在将YOLOv3的Paddle模型导出时，对最后一层multiclass_nms进行了特殊处理，导出的ONNX模型，最终输出的Box结果包括背景类别（而Paddle模型不包含），此处在OpenVINO的部署代码中，我们通过后处理过滤了背景类别。

docs/deploy/openvino/index.rst

@@ -6,6 +6,8 @@ OpenVINO部署
    :maxdepth: 2
    :caption: 文档目录:
 
+   introduction.md
    windows.md
    linux.md
-   intel_movidius.md
+   python.md
+   export_openvino_model.md 

docs/deploy/openvino/intel_movidius.md

-# Intel计算棒

docs/deploy/openvino/introduction.md

+# OpenVINO部署简介
+PaddleX支持将训练好的Paddle模型通过OpenVINO实现模型的预测加速，OpenVINO详细资料与安装流程请参考[OpenVINO](https://docs.openvinotoolkit.org/latest/index.html)
+
+## 部署支持情况
+下表提供了PaddleX在不同环境下对使用OpenVINO加速的支持情况  
+
+|硬件平台|Linux|Windows|Raspbian OS|c++|python |分类|检测|分割|
+| ----|  ---- | ---- | ----|  ---- | ---- |---- | ---- |---- |
+|CPU|支持|支持|不支持|支持|支持|支持|支持|不支持|
+|VPU|支持|支持|支持|支持|支持|支持|不支持|不支持|  
+
+
+**注意**：其中Raspbian OS为树莓派操作系统。检测模型仅支持YOLOV3，由于OpenVINO不支持ONNX的resize-11 OP的原因，目前还不支持Paddle的分割模型
+
+## 部署流程
+**PaddleX到OpenVINO的部署流程可以分为如下两步**： 
+
+  * **模型转换**:将Paddle的模型转换为OpenVINO的Inference Engine
+  * **预测部署**:使用Inference Engine进行预测
+
+## 模型转换 
+**模型转换请参考文档[模型转换](./export_openvino_model.md)**  
+**说明**：由于不同软硬件平台下OpenVINO模型转换方法一致，故如何转换模型后续文档中不再赘述。
+
+## 预测部署
+由于不同软硬下部署OpenVINO实现预测的方式不完全一致，具体请参考：  
+**[Linux](./linux.md)**:介绍了PaddleX在操作系统为Linux或者Raspbian OS，编程语言为C++，硬件平台为
+CPU或者VPU的情况下使用OpenVINO进行预测加速  
+
+**[Windows](./windows.md)**:介绍了PaddleX在操作系统为Window，编程语言为C++，硬件平台为CPU或者VPU的情况下使用OpenVINO进行预测加速  
+
+**[Python](./python.md)**:介绍了PaddleX在python下使用OpenVINO进行预测加速
\ No newline at end of file

docs/deploy/openvino/linux.md

 # Linux平台
+
+
+## 前置条件
+
+* OS: Ubuntu、Raspbian OS
+* GCC* 5.4.0
+* CMake 3.0+
+* PaddleX 1.0+
+* OpenVINO 2020.4
+* 硬件平台：CPU、VPU
+
+**说明**：PaddleX安装请参考[PaddleX](https://paddlex.readthedocs.io/zh_CN/develop/install.html) ， OpenVINO安装请根据相应的系统参考[OpenVINO-Linux](https://docs.openvinotoolkit.org/latest/_docs_install_guides_installing_openvino_linux.html)或者[OpenVINO-Raspbian](https://docs.openvinotoolkit.org/latest/openvino_docs_install_guides_installing_openvino_raspbian.html)
+
+请确保系统已经安装好上述基本软件，并配置好相应环境，**下面所有示例以工作目录 `/root/projects/`演示**。
+
+
+
+## 预测部署  
+
+文档提供了c++下预测部署的方法，如果需要在python下预测部署请参考[python预测部署](./python.md)
+
+### Step1 下载PaddleX预测代码
+```
+mkdir -p /root/projects
+cd /root/projects
+git clone https://github.com/PaddlePaddle/PaddleX.git
+```
+**说明**：其中C++预测代码在PaddleX/deploy/openvino 目录，该目录不依赖任何PaddleX下其他目录。
+
+### Step2 软件依赖
+提供了依赖软件预编包或者一键编译，用户不需要单独下载或编译第三方依赖软件。若需要自行编译第三方依赖软件请参考：
+
+- gflags：编译请参考 [编译文档](https://gflags.github.io/gflags/#download)  
+
+- glog：编译请参考[编译文档](https://github.com/google/glog)
+
+- opencv: 编译请参考
+[编译文档](https://docs.opencv.org/master/d7/d9f/tutorial_linux_install.html)
+
+
+
+### Step3: 编译
+编译`cmake`的命令在`scripts/build.sh`中，若在树莓派(Raspbian OS)上编译请修改ARCH参数x86为armv7，若自行编译第三方依赖软件请根据Step1中编译软件的实际情况修改主要参数，其主要内容说明如下：
+```
+# openvino预编译库的路径
+OPENVINO_DIR=$INTEL_OPENVINO_DIR/inference_engine
+# gflags预编译库的路径
+GFLAGS_DIR=$(pwd)/deps/gflags
+# glog预编译库的路径
+GLOG_DIR=$(pwd)/deps/glog
+# ngraph lib预编译库的路径
+NGRAPH_LIB=$INTEL_OPENVINO_DIR/deployment_tools/ngraph/lib
+# opencv预编译库的路径
+OPENCV_DIR=$(pwd)/deps/opencv/
+#cpu架构（x86或armv7）
+ARCH=x86
+```
+执行`build`脚本：
+ ```shell
+ sh ./scripts/build.sh
+ ```  
+
+### Step4: 预测
+
+编译成功后，分类任务的预测可执行程序为`classifier`，检测任务的预测可执行程序为`detector`，其主要命令参数说明如下：
+
+|  参数   | 说明  |
+|  ----  | ----  |
+| --model_dir  | 模型转换生成的.xml文件路径，请保证模型转换生成的三个文件在同一路径下|
+| --image  | 要预测的图片文件路径 |
+| --image_list  | 按行存储图片路径的.txt文件 |
+| --device  | 运行的平台，可选项{"CPU"，"MYRIAD"}，默认值为"CPU"，如在VPU上请使用"MYRIAD"|
+| --cfg_file | PaddleX model 的.yml配置文件 |
+| --save_dir | 可视化结果图片保存地址，仅适用于检测任务，默认值为" "既不保存可视化结果 |
+
+### 样例
+`样例一`：
+linux系统在CPU下做单张图片的分类任务预测  
+测试图片 `/path/to/test_img.jpeg`  
+
+```shell
+./build/classifier --model_dir=/path/to/openvino_model --image=/path/to/test_img.jpeg --cfg_file=/path/to/PadlleX_model.yml
+```
+
+
+`样例二`:
+linux系统在CPU下做多张图片的检测任务预测，并保存预测可视化结果
+预测的多个图片`/path/to/image_list.txt`，image_list.txt内容的格式如下：
+```
+/path/to/images/test_img1.jpeg
+/path/to/images/test_img2.jpeg
+...
+/path/to/images/test_imgn.jpeg
+```
+
+```shell
+./build/detector --model_dir=/path/to/models/openvino_model --image_list=/root/projects/images_list.txt --cfg_file=/path/to/PadlleX_model.yml --save_dir ./output
+```
+
+`样例三`:  
+树莓派(Raspbian OS)在VPU下做单张图片分类任务预测
+测试图片 `/path/to/test_img.jpeg`  
+
+```shell
+./build/classifier --model_dir=/path/to/openvino_model --image=/path/to/test_img.jpeg --cfg_file=/path/to/PadlleX_model.yml --device=MYRIAD
+```
+
+## 性能测试
+`测试一`：  
+在服务器CPU下测试了OpenVINO对PaddleX部署的加速性能：
+- CPU：Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz
+- OpenVINO： 2020.4
+- PaddleX：采用Paddle预测库(1.8)，打开mkldnn加速，打开多线程。
+- 模型来自PaddleX tutorials，Batch Size均为1，耗时单位为ms/image，只计算模型运行时间，不包括数据的预处理和后处理，20张图片warmup，100张图片测试性能。
+
+|模型| PaddleX| OpenVINO |  图片输入大小|
+|---|---|---|---|
+|resnet-50 | 20.56 | 16.12 | 224*224 |
+|mobilenet-V2 | 5.16 | 2.31 |224*224|
+|yolov3-mobilnetv1 |76.63| 46.26|608*608 |  
+
+`测试二`:
+在PC机上插入VPU架构的神经计算棒(NCS2)，通过Openvino加速。
+- CPU：Intel(R) Core(TM) i5-4300U 1.90GHz
+- VPU：Movidius Neural Compute Stick2
+- OpenVINO： 2020.4
+- 模型来自PaddleX tutorials，Batch Size均为1，耗时单位为ms/image，只计算模型运行时间，不包括数据的预处理和后处理，20张图片warmup，100张图片测试性能。  
+
+|模型|OpenVINO|输入图片|
+|---|---|---|
+|mobilenetV2|24.00|224*224|
+|resnet50_vd_ssld|58.53|224*224|  
+
+`测试三`:
+在树莓派3B上插入VPU架构的神经计算棒(NCS2)，通过Openvino加速。
+- CPU ：ARM Cortex-A72 1.2GHz 64bit
+- VPU：Movidius Neural Compute Stick2
+- OpenVINO 2020.4
+- 模型来自paddleX tutorials，Batch Size均为1，耗时单位为ms/image，只计算模型运行时间，不包括数据的预处理和后处理，20张图片warmup，100张图片测试性能。  
+
+|模型|OpenVINO|输入图片大小|
+|---|---|---|
+|mobilenetV2|43.15|224*224|
+|resnet50|82.66|224*224|  

docs/deploy/openvino/python.md

+# Python预测部署
+文档说明了在python下基于OpenVINO的预测部署，部署前需要先将paddle模型转换为OpenVINO的Inference Engine，请参考[模型转换](docs/deploy/openvino/export_openvino_model.md)。目前CPU硬件上支持PadlleX的分类、检测、分割模型；VPU上支持PaddleX的分类模型。
+
+## 前置条件
+* Python 3.6+
+* OpenVINO 2020.4
+
+**说明**：OpenVINO安装请参考[OpenVINO](https://docs.openvinotoolkit.org/latest/index.html)  
+
+
+请确保系统已经安装好上述基本软件，**下面所有示例以工作目录 `/root/projects/`演示**。
+
+## 预测部署
+运行/root/projects/PaddleX/deploy/openvino/python目录下demo.py文件可以进行预测，其命令参数说明如下：
+
+|  参数   | 说明  |
+|  ----  | ----  |
+| --model_dir  | 模型转换生成的.xml文件路径，请保证模型转换生成的三个文件在同一路径下|
+| --img  | 要预测的图片文件路径 |
+| --image_list  | 按行存储图片路径的.txt文件 |
+| --device  | 运行的平台, 默认值为"CPU" |
+| --cfg_file | PaddleX model 的.yml配置文件 |
+
+### 样例
+`样例一`：  
+测试图片 `/path/to/test_img.jpeg`  
+
+```
+cd /root/projects/python  
+
+python demo.py --model_dir /path/to/openvino_model --img /path/to/test_img.jpeg --cfg_file /path/to/PadlleX_model.yml
+```  
+
+样例二`:
+
+预测多个图片`/path/to/image_list.txt`，image_list.txt内容的格式如下：
+
+```
+/path/to/images/test_img1.jpeg
+/path/to/images/test_img2.jpeg
+...
+/path/to/images/test_imgn.jpeg
+```
+
+```
+cd /root/projects/python  
+
+python demo.py --model_dir /path/to/models/openvino_model --image_list /root/projects/images_list.txt --cfg_file=/path/to/PadlleX_model.yml
+```

docs/deploy/openvino/windows.md

 # Windows平台
+
+## 说明
+Windows 平台下，我们使用`Visual Studio 2019 Community` 进行了测试。微软从`Visual Studio 2017`开始即支持直接管理`CMake`跨平台编译项目，但是直到`2019`才提供了稳定和完全的支持，所以如果你想使用CMake管理项目编译构建，我们推荐你使用`Visual Studio 2019`环境下构建。
+
+## 前置条件
+* Visual Studio 2019
+* OpenVINO 2020.4
+* CMake 3.0+
+
+**说明**：PaddleX安装请参考[PaddleX](https://paddlex.readthedocs.io/zh_CN/develop/install.html) ， OpenVINO安装请参考[OpenVINO-Windows](https://docs.openvinotoolkit.org/latest/openvino_docs_install_guides_installing_openvino_windows.html)  
+
+**注意**：安装完OpenVINO后需要手动添加OpenVINO目录到系统环境变量，否则在运行程序时会出现找不到dll的情况。以安装OpenVINO时不改变OpenVINO安装目录情况下为示例，流程如下
+- 我的电脑->属性->高级系统设置->环境变量
+    - 在系统变量中找到Path（如没有，自行创建），并双击编辑
+    - 新建，分别将OpenVINO以下路径填入并保存:  
+      `C:\Program File (x86)\IntelSWTools\openvino\inference_engine\bin\intel64\Release`  
+      `C:\Program File (x86)\IntelSWTools\openvino\inference_engine\external\tbb\bin`  
+      `C:\Program File (x86)\IntelSWTools\openvino\deployment_tools\ngraph\lib`  
+
+请确保系统已经安装好上述基本软件，并配置好相应环境，**下面所有示例以工作目录为 `D:\projects`演示。**
+
+## 预测部署  
+
+文档提供了c++下预测部署的方法，如果需要在python下预测部署请参考[python预测部署](./python.md)
+
+### Step1: 下载PaddleX预测代码
+
+```shell
+d:
+mkdir projects
+cd projects
+git clone https://github.com/PaddlePaddle/PaddleX.git
+```
+
+**说明**：其中`C++`预测代码在`PaddleX\deploy\openvino` 目录，该目录不依赖任何`PaddleX`下其他目录。
+
+### Step2 软件依赖
+提供了依赖软件预编译库:
+- [gflas-glog](https://bj.bcebos.com/paddlex/deploy/windows/third-parts.zip)  
+- [opencv](https://bj.bcebos.com/paddleseg/deploy/opencv-3.4.6-vc14_vc15.exe)  
+请下载上面两个连接的预编译库。若需要自行下载请参考：
+- gflags:[下载地址](https://docs.microsoft.com/en-us/windows-hardware/drivers/debugger/gflags)
+- glog：[编译文档](https://github.com/google/glog)
+- opencv:[下载地址](https://opencv.org/releases/)  
+下载完opencv后需要配置环境变量，如下流程所示  
+    - 我的电脑->属性->高级系统设置->环境变量
+    - 在系统变量中找到Path（如没有，自行创建），并双击编辑
+    - 新建，将opencv路径填入并保存，如`D:\projects\opencv\build\x64\vc14\bin`
+
+### Step3: 使用Visual Studio 2019直接编译CMake
+1. 打开Visual Studio 2019 Community，点击`继续但无需代码`
+2. 点击： `文件`->`打开`->`CMake` 选择C++预测代码所在路径（例如`D:\projects\PaddleX\deploy\openvino`），并打开`CMakeList.txt`：
+3. 点击：`项目`->`CMake设置`
+4. 点击`浏览`，分别设置编译选项指定`OpenVINO`、`Gflags`、`GLOG`、`NGRAPH`、`OPENCV`的路径  
+
+|  参数名   | 含义  |
+|  ----  | ----  |
+| OPENCV_DIR  | opencv库路径 |
+| OPENVINO_DIR | OpenVINO推理库路径，在OpenVINO安装目录下的deployment/inference_engine目录，若未修改OpenVINO默认安装目录可以不用修改 |
+| NGRAPH_LIB | OpenVINO的ngraph库路径，在OpenVINO安装目录下的deployment/ngraph/lib目录，若未修改OpenVINO默认安装目录可以不用修改 |
+| GFLAGS_DIR | gflags库路径 |
+| GLOG_DIR  | glog库路径 |
+| WITH_STATIC_LIB | 是否静态编译，默认为True |  
+
+**设置完成后**, 点击`保存并生成CMake缓存以加载变量`。
+5. 点击`生成`->`全部生成`
+### Step5: 预测
+上述`Visual Studio 2019`编译产出的可执行文件在`out\build\x64-Release`目录下，打开`cmd`，并切换到该目录：
+
+```
+D:
+cd D:\projects\PaddleX\deploy\openvino\out\build\x64-Release
+```
+
+* 编译成功后，图片预测demo的入口程序为`detector.exe`，`classifier.exe`，用户可根据自己的模型类型选择，其主要命令参数说明如下：
+
+|  参数   | 说明  |
+|  ----  | ----  |
+| --model_dir  | 模型转换生成的.xml文件路径，请保证模型转换生成的三个文件在同一路径下|
+| --image  | 要预测的图片文件路径 |
+| --image_list  | 按行存储图片路径的.txt文件 |
+| --device  | 运行的平台，可选项{"CPU"，"MYRIAD"}，默认值为"CPU"，如在VPU上请使用"MYRIAD"|
+| --cfg_file | PaddleX model 的.yml配置文件 |
+| --save_dir | 可视化结果图片保存地址，仅适用于检测任务，默认值为" "既不保存可视化结果 |
+
+### 样例
+`样例一`：
+在CPU下做单张图片的分类任务预测  
+测试图片 `/path/to/test_img.jpeg`  
+
+```shell
+./classifier.exe --model_dir=/path/to/openvino_model --image=/path/to/test_img.jpeg --cfg_file=/path/to/PadlleX_model.yml
+```
+
+`样例二`:
+在CPU下做多张图片的检测任务预测，并保存预测可视化结果
+预测多个图片`/path/to/image_list.txt`，image_list.txt内容的格式如下：
+```
+/path/to/images/test_img1.jpeg
+/path/to/images/test_img2.jpeg
+...
+/path/to/images/test_imgn.jpeg
+```
+
+```shell
+./detector.exe --model_dir=/path/to/models/openvino_model --image_list=/root/projects/images_list.txt --cfg_file=/path/to/PadlleX_model.yml --save_dir ./output
+```
+
+`样例三`:  
+在VPU下做单张图片分类任务预测
+测试图片 `/path/to/test_img.jpeg`  
+
+```shell
+.classifier.exe --model_dir=/path/to/openvino_model --image=/path/to/test_img.jpeg --cfg_file=/path/to/PadlleX_model.yml --device=MYRIAD
+```

docs/deploy/raspberry/Raspberry.md

+# 树莓派
+PaddleX支持通过Paddle-Lite和基于OpenVINO的神经计算棒(NCS2)这两种方式在树莓派上完成预测部署。
+
+
+## 硬件环境配置  
+
+对于尚未安装系统的树莓派首先需要进行系统安装、环境配置等步骤来初始化硬件环境，过程中需要的软硬件如下：
+
+- 硬件：micro SD，显示器，键盘，鼠标
+- 软件：Raspbian OS
+### Step1：系统安装
+- 格式化micro SD卡为FAT格式，Windows和Mac下建议使用[SD Memory Card Formatter](https://www.sdcard.org/downloads/formatter/)工具，Linux下请参考[NOOBS For Raspberry Pi](http://qdosmsq.dunbar-it.co.uk/blog/2013/06/noobs-for-raspberry-pi/)  
+- 下载NOOBS版本的Raspbian OS [下载地址](https://www.raspberrypi.org/downloads/)并将解压后的文件复制到SD中，插入SD后给树莓派通电，然后将自动安装系统
+### Step2：环境配置
+- 启用VNC和SSH服务：打开LX终端输入，输入如下命令，选择Interfacing Option然后选择P2 SSH 和 P3 VNC分别打开SSH与VNC。打开后就可以通过SSH或者VNC的方式连接树莓派
+```
+sudo raspi-config
+```
+- 更换源：由于树莓派官方源速度很慢，建议在官网查询国内源 [树莓派软件源](https://www.jianshu.com/p/67b9e6ebf8a0)。更换后执行
+```
+sudo apt-get update
+sudo apt-get upgrade
+```
+
+## Paddle-Lite部署
+基于Paddle-Lite的部署目前可以支持PaddleX的分类、分割与检测模型，其实检测模型仅支持YOLOV3  
+部署的流程包括：PaddleX模型转换与转换后的模型部署  
+
+**说明**：PaddleX安装请参考[PaddleX](https://paddlex.readthedocs.io/zh_CN/develop/install.html)，Paddle-Lite详细资料请参考[Paddle-Lite](https://paddle-lite.readthedocs.io/zh/latest/index.html)
+
+请确保系统已经安装好上述基本软件，并配置好相应环境，**下面所有示例以工作目录 `/root/projects/`演示**。
+
+## Paddle-Lite模型转换
+将PaddleX模型转换为Paddle-Lite模型，具体请参考[Paddle-Lite模型转换](./export_nb_model.md)
+
+## Paddle-Lite 预测
+### Step1 下载PaddleX预测代码
+```
+mkdir -p /root/projects
+cd /root/projects
+git clone https://github.com/PaddlePaddle/PaddleX.git
+```
+**说明**：其中C++预测代码在PaddleX/deploy/raspberry 目录，该目录不依赖任何PaddleX下其他目录，如果需要在python下预测部署请参考[Python预测部署](./python.md)。  
+
+### Step2：Paddle-Lite预编译库下载
+提供了下载的opt工具对应的Paddle-Lite在ArmLinux下面的预编译库:[Paddle-Lite(ArmLinux)预编译库](https://bj.bcebos.com/paddlex/deploy/lite/inference_lite_2.6.1_armlinux.tar.bz2)。  
+建议用户使用预编译库，若需要自行编译，在树莓派上LX终端输入
+```
+git clone https://github.com/PaddlePaddle/Paddle-Lite.git
+cd Paddle-Lite
+sudo ./lite/tools/build.sh  --arm_os=armlinux --arm_abi=armv7hf --arm_lang=gcc  --build_extra=ON full_publish
+```  
+
+预编库位置：`./build.lite.armlinux.armv7hf.gcc/inference_lite_lib.armlinux.armv7hf/cxx`  
+
+**注意**：预测库版本需要跟opt版本一致，更多Paddle-Lite编译内容请参考[Paddle-Lite编译](https://paddle-lite.readthedocs.io/zh/latest/user_guides/source_compile.html)；更多预编译Paddle-Lite预测库请参考[Paddle-Lite Release Note](https://github.com/PaddlePaddle/Paddle-Lite/releases)
+
+### Step3 软件依赖
+提供了依赖软件的预编包或者一键编译，用户不需要单独下载或编译第三方依赖软件。若需要自行编译第三方依赖软件请参考：
+
+- gflags：编译请参考 [编译文档](https://gflags.github.io/gflags/#download)  
+
+- glog：编译请参考[编译文档](https://github.com/google/glog)
+
+- opencv: 编译请参考
+[编译文档](https://docs.opencv.org/master/d7/d9f/tutorial_linux_install.html)
+### Step4: 编译
+编译`cmake`的命令在`scripts/build.sh`中，修改LITE_DIR为Paddle-Lite预测库目录，若自行编译第三方依赖软件请根据Step1中编译软件的实际情况修改主要参数，其主要内容说明如下：
+```
+# Paddle-Lite预编译库的路径
+LITE_DIR=/path/to/Paddle-Lite/inference/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预编译库的路径
+OPENCV_DIR=$(pwd)/deps/opencv/
+```
+执行`build`脚本：
+ ```shell
+ sh ./scripts/build.sh
+ ```  
+
+
+### Step3: 预测
+
+编译成功后，分类任务的预测可执行程序为`classifier`,分割任务的预测可执行程序为`segmenter`，检测任务的预测可执行程序为`detector`，其主要命令参数说明如下：  
+|  参数   | 说明  |
+|  ----  | ----  |
+| --model_dir  | 模型转换生成的.xml文件路径，请保证模型转换生成的三个文件在同一路径下|
+| --image  | 要预测的图片文件路径 |
+| --image_list  | 按行存储图片路径的.txt文件 |
+| --thread_num | 预测的线程数，默认值为1 |
+| --cfg_file | PaddleX model 的.yml配置文件 |
+| --save_dir | 可视化结果图片保存地址，仅适用于检测和分割任务，默认值为" "既不保存可视化结果 |
+
+### 样例
+`样例一`：
+单张图片分类任务  
+测试图片 `/path/to/test_img.jpeg`  
+
+```shell
+./build/classifier --model_dir=/path/to/nb_model
+--image=/path/to/test_img.jpeg --cfg_file=/path/to/PadlleX_model.yml  --thread_num=4
+```
+
+
+`样例二`:
+多张图片分割任务
+预测多个图片`/path/to/image_list.txt`，image_list.txt内容的格式如下：
+```
+/path/to/images/test_img1.jpeg
+/path/to/images/test_img2.jpeg
+...
+/path/to/images/test_imgn.jpeg
+```
+
+```shell
+./build/segmenter --model_dir=/path/to/models/nb_model --image_list=/root/projects/images_list.txt --cfg_file=/path/to/PadlleX_model.yml --save_dir ./output --thread_num=4  
+```  
+
+## 性能测试
+### 测试环境：
+硬件：Raspberry Pi 3 Model B
+系统：raspbian OS
+软件：paddle-lite 2.6.1
+### 测试结果
+单位ms，num表示paddle-lite下使用的线程数
+|模型|lite(num=4)|输入图片大小|
+| ----|  ---- | ----|
+|mobilenet-v2|136.19|224*224|
+|resnet-50|1131.42|224*224|
+|deeplabv3|2162.03|512*512|
+|hrnet|6118.23|512*512|
+|yolov3-darknet53|4741.15|320*320|
+|yolov3-mobilenet|1424.01|320*320|
+|densenet121|1144.92|224*224|
+|densenet161|2751.57|224*224|
+|densenet201|1847.06|224*224|
+|HRNet_W18|1753.06|224*224|
+|MobileNetV1|177.63|224*224|
+|MobileNetV3_large_ssld|133.99|224*224|
+|MobileNetV3_small_ssld|53.99|224*224|
+|ResNet101|2290.56|224*224|
+|ResNet101_vd|2337.51|224*224|
+|ResNet101_vd_ssld|3124.49|224*224|
+|ShuffleNetV2|115.97|224*224|
+|Xception41|1418.29|224*224|
+|Xception65|2094.7|224*224|  
+
+从测试结果看建议用户在树莓派上使用MobileNetV1-V3,ShuffleNetV2这类型的小型网络
+
+## NCS2部署
+树莓派支持通过OpenVINO在NCS2上跑PaddleX模型预测，目前仅支持PaddleX的分类网络，基于NCS2的方式包含Paddle模型转OpenVINO IR以及部署IR在NCS2上进行预测两个步骤。
+- 模型转换请参考：[PaddleX模型转换为OpenVINO IR]('./openvino/export_openvino_model.md')，raspbian OS上的OpenVINO不支持模型转换，需要先在host侧转换FP16的IR。
+- 预测部署请参考[OpenVINO部署](./openvino/linux.md)中VPU在raspbian OS部署的部分

docs/deploy/raspberry/export_nb_model.md

+# Paddle-Lite模型转换
+将PaddleX模型转换为Paddle-Lite的nb模型，模型转换主要包括PaddleX转inference model和inference model转Paddle-Lite nb模型
+### Step1：导出inference模型
+PaddleX模型转Paddle-Lite模型之前需要先把PaddleX模型导出为inference格式模型，导出的模型将包括__model__、__params__和model.yml三个文件名。具体方法请参考[Inference模型导出](../export_model.md)。
+### Step2：导出Paddle-Lite模型
+Paddle-Lite模型需要通过Paddle-Lite的opt工具转出模型，下载并解压: [模型优化工具opt（2.6.1-linux）](https://bj.bcebos.com/paddlex/deploy/Rasoberry/opt.zip)，在Linux系统下运行：
+``` bash
+./opt --model_file=<model_path> \
+      --param_file=<param_path> \
+      --valid_targets=arm \
+      --optimize_out_type=naive_buffer \
+      --optimize_out=model_output_name
+```
+|  参数   | 说明  |
+|  ----  | ----  |
+| --model_file  | 导出inference模型中包含的网络结构文件：`__model__`所在的路径|
+| --param_file  | 导出inference模型中包含的参数文件：`__params__`所在的路径|
+| --valid_targets  |指定模型可执行的backend，这里请指定为`arm`|
+| --optimize_out_type | 输出模型类型，目前支持两种类型：protobuf和naive_buffer，其中naive_buffer是一种更轻量级的序列化/反序列化，这里请指定为`naive_buffer`|  
+
+
+若安装了python版本的Paddle-Lite也可以通过如下方式转换
+``` 
+./paddle_lite_opt --model_file=<model_path> \
+      --param_file=<param_path> \
+      --valid_targets=arm \
+      --optimize_out_type=naive_buffer \
+      --optimize_out=model_output_name
+```
+
+更多详细的使用方法和参数含义请参考: [使用opt转化模型](https://paddle-lite.readthedocs.io/zh/latest/user_guides/opt/opt_bin.html)，更多opt预编译版本请参考[Paddle-Lite Release Note](https://github.com/PaddlePaddle/Paddle-Lite/releases)
+
+**注意**：opt版本需要跟预测库版本保持一致，如使2.6.0版本预测库，请从上面Release Note中下载2.6.0版本的opt转换模型
\ No newline at end of file

docs/deploy/raspberry/index.rst

+树莓派部署
+=======================================
+
+
+.. toctree::
+   :maxdepth: 2
+   :caption: 文档目录:
+
+   Raspberry.md
+   python.md
+   export_nb_model.md 
\ No newline at end of file

docs/deploy/raspberry/python.md

+# Python预测部署
+文档说明了在树莓派上使用Python版本的Paddle-Lite进行PaddleX模型好的预测部署，根据下面的命令安装Python版本的Paddle-Lite预测库，若安装不成功用户也可以下载whl文件进行安装[Paddle-Lite_2.6.0_python](https://github.com/PaddlePaddle/Paddle-Lite/releases/download/v2.6.0/armlinux_python_installer.zip)，更多版本请参考[Paddle-Lite Release Note](https://github.com/PaddlePaddle/Paddle-Lite/releases)
+```
+python -m pip install paddlelite
+```
+部署前需要先将PaddleX模型转换为Paddle-Lite的nb模型，具体请参考[Paddle-Lite模型转换](./export_nb_model.md)
+**注意**：若用户使用2.6.0的Python预测库，请下载2.6.0版本的opt转换工具转换模型
+
+
+
+## 前置条件
+* Python 3.6+
+* Paddle-Lite_python 2.6.0+
+
+请确保系统已经安装好上述基本软件，**下面所有示例以工作目录 `/root/projects/`演示**。
+
+## 预测部署
+运行/root/projects/PaddleX/deploy/raspberry/python目录下demo.py文件可以进行预测，其命令参数说明如下：
+
+|  参数   | 说明  |
+|  ----  | ----  |
+| --model_dir  | 模型转换生成的.xml文件路径，请保证模型转换生成的三个文件在同一路径下|
+| --img  | 要预测的图片文件路径 |
+| --image_list  | 按行存储图片路径的.txt文件 |
+| --cfg_file | PaddleX model 的.yml配置文件 |
+| --thread_num  | 预测的线程数, 默认值为1 |
+| --input_shape  | 模型输入中图片输入的大小[N,C,H.W] |
+
+### 样例
+`样例一`：  
+测试图片 `/path/to/test_img.jpeg`  
+
+```
+cd /root/projects/python  
+
+python demo.py --model_dir /path/to/openvino_model --img /path/to/test_img.jpeg --cfg_file /path/to/PadlleX_model.yml --thread_num 4 --input_shape [1,3,224,224]
+```  
+
+样例二`:
+
+预测多个图片`/path/to/image_list.txt`，image_list.txt内容的格式如下：
+
+```
+/path/to/images/test_img1.jpeg
+/path/to/images/test_img2.jpeg
+...
+/path/to/images/test_imgn.jpeg
+```
+
+```
+cd /root/projects/python  
+
+python demo.py --model_dir /path/to/models/openvino_model --image_list /root/projects/images_list.txt --cfg_file=/path/to/PadlleX_model.yml --thread_num 4 --input_shape [1,3,224,224]
+```

docs/deploy/server/cpp/linux.md

@@ -125,6 +125,8 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
 | image_list  | 按行存储图片路径的.txt文件 |
 | use_gpu  | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
 | use_trt  | 是否使用 TensorRT 预测, 支持值为0或1(默认值为0) |
+| use_mkl  | 是否使用 MKL加速CPU预测, 支持值为0或1(默认值为1) |
+| mkl_thread_num | MKL推理的线程数，默认为cpu处理器个数 |
 | gpu_id  | GPU 设备ID, 默认值为0 |
 | save_dir | 保存可视化结果的路径, 默认值为"output"，**classfier无该参数** |
 | key | 加密过程中产生的密钥信息，默认值为""表示加载的是未加密的模型 |
@@ -141,6 +143,8 @@ yaml-cpp.zip文件下载后无需解压，在cmake/yaml.cmake中将`URL https://
 | video_path | 视频文件的路径 |
 | use_gpu  | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
 | use_trt  | 是否使用 TensorRT 预测, 支持值为0或1(默认值为0) |
+| use_mkl  | 是否使用 MKL加速CPU预测, 支持值为0或1(默认值为1) |
+| mkl_thread_num | MKL推理的线程数，默认为cpu处理器个数 |
 | gpu_id  | GPU 设备ID, 默认值为0 |
 | show_result | 对视频文件做预测时，是否在屏幕上实时显示预测可视化结果(因加入了延迟处理，故显示结果不能反映真实的帧率)，支持值为0或1(默认值为0) |
 | save_result | 是否将每帧的预测可视结果保存为视频文件，支持值为0或1(默认值为1) |

docs/deploy/server/cpp/windows.md

@@ -109,6 +109,8 @@ cd D:\projects\PaddleX\deploy\cpp\out\build\x64-Release
 | image  | 要预测的图片文件路径 |
 | image_list  | 按行存储图片路径的.txt文件 |
 | use_gpu  | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
+| use_mkl  | 是否使用 MKL加速CPU预测, 支持值为0或1(默认值为1) |
+| mkl_thread_num | MKL推理的线程数，默认为cpu处理器个数 |
 | gpu_id  | GPU 设备ID, 默认值为0 |
 | save_dir | 保存可视化结果的路径, 默认值为"output"，classifier无该参数 |
 | key | 加密过程中产生的密钥信息，默认值为""表示加载的是未加密的模型 |
@@ -124,6 +126,8 @@ cd D:\projects\PaddleX\deploy\cpp\out\build\x64-Release
 | camera_id | 摄像头设备ID，默认值为0 |
 | video_path | 视频文件的路径 |
 | use_gpu  | 是否使用 GPU 预测, 支持值为0或1(默认值为0) |
+| use_mkl  | 是否使用 MKL加速CPU预测, 支持值为0或1(默认值为1) |
+| mkl_thread_num | MKL推理的线程数，默认为cpu处理器个数 |
 | gpu_id  | GPU 设备ID, 默认值为0 |
 | show_result | 对视频文件做预测时，是否在屏幕上实时显示预测可视化结果(因加入了延迟处理，故显示结果不能反映真实的帧率)，支持值为0或1(默认值为0) |
 | save_result | 是否将每帧的预测可视结果保存为视频文件，支持值为0或1(默认值为1) |

docs/examples/meter_reader.md

@@ -70,7 +70,6 @@ cd PaddleX/examples/meter_reader/
 | save_dir	| 保存可视化结果的路径, 默认值为"output"|
 | score_threshold | 检测模型输出结果中，预测得分低于该阈值的框将被滤除，默认值为0.5|
 | seg_batch_size | 分割的批量大小，默认为2 |
-| seg_thread_num	| 分割预测的线程数，默认为cpu处理器个数 |
 | use_camera | 是否使用摄像头采集图片，默认为False |
 | camera_id | 摄像头设备ID，默认值为0 |
 | use_erode | 是否使用图像腐蚀对分割预测图进行细分，默认为False |

examples/meter_reader/README.md

@@ -79,7 +79,6 @@ cd PaddleX/examples/meter_reader/
 | save_dir	| 保存可视化结果的路径, 默认值为"output"|
 | score_threshold | 检测模型输出结果中，预测得分低于该阈值的框将被滤除，默认值为0.5|
 | seg_batch_size | 分割的批量大小，默认为2 |
-| seg_thread_num	| 分割预测的线程数，默认为cpu处理器个数 |
 | use_camera | 是否使用摄像头采集图片，默认为False |
 | camera_id | 摄像头设备ID，默认值为0 |
 | use_erode | 是否使用图像腐蚀对分割预测图进行细分，默认为False |

examples/meter_reader/deploy/python/reader_deploy.py

@@ -105,12 +105,6 @@ def parse_args():
         help="Segmentation batch size",
         type=int,
         default=2)
-    parser.add_argument(
-        '--seg_thread_num',
-        dest='seg_thread_num',
-        help="Thread number of segmentation preprocess",
-        type=int,
-        default=2)
 
     return parser.parse_args()
 
@@ -143,8 +137,7 @@ class MeterReader:
                 use_erode=True,
                 erode_kernel=4,
                 score_threshold=0.5,
-                seg_batch_size=2,
-                seg_thread_num=2):
+                seg_batch_size=2):
         if isinstance(im_file, str):
             im = cv2.imread(im_file).astype('float32')
         else:
@@ -190,8 +183,7 @@ class MeterReader:
                 meter_images.append(resized_meters[j - i])
             result = self.segmenter.batch_predict(
                 transforms=self.seg_transforms,
-                img_file_list=meter_images,
-                thread_num=seg_thread_num)
+                img_file_list=meter_images)
             if use_erode:
                 kernel = np.ones((erode_kernel, erode_kernel), np.uint8)
                 for i in range(len(result)):
@@ -334,7 +326,7 @@ def infer(args):
         for im_file in image_lists:
             meter_reader.predict(im_file, args.save_dir, args.use_erode,
                                  args.erode_kernel, args.score_threshold,
-                                 args.seg_batch_size, args.seg_thread_num)
+                                 args.seg_batch_size)
     elif args.use_camera:
         cap_video = cv2.VideoCapture(args.camera_id)
         if not cap_video.isOpened():
@@ -347,7 +339,7 @@ def infer(args):
             if ret:
                 meter_reader.predict(frame, args.save_dir, args.use_erode,
                                      args.erode_kernel, args.score_threshold,
-                                     args.seg_batch_size, args.seg_thread_num)
+                                     args.seg_batch_size)
                 if cv2.waitKey(1) & 0xFF == ord('q'):
                     break
             else:

examples/meter_reader/reader_infer.py

@@ -105,12 +105,6 @@ def parse_args():
         help="Segmentation batch size",
         type=int,
         default=2)
-    parser.add_argument(
-        '--seg_thread_num',
-        dest='seg_thread_num',
-        help="Thread number of segmentation preprocess",
-        type=int,
-        default=2)
 
     return parser.parse_args()
 
@@ -143,8 +137,7 @@ class MeterReader:
                 use_erode=True,
                 erode_kernel=4,
                 score_threshold=0.5,
-                seg_batch_size=2,
-                seg_thread_num=2):
+                seg_batch_size=2):
         if isinstance(im_file, str):
             im = cv2.imread(im_file).astype('float32')
         else:
@@ -190,8 +183,7 @@ class MeterReader:
                 meter_images.append(resized_meters[j - i])
             result = self.segmenter.batch_predict(
                 transforms=self.seg_transforms,
-                img_file_list=meter_images,
-                thread_num=seg_thread_num)
+                img_file_list=meter_images)
             if use_erode:
                 kernel = np.ones((erode_kernel, erode_kernel), np.uint8)
                 for i in range(len(result)):
@@ -334,7 +326,7 @@ def infer(args):
         for im_file in image_lists:
             meter_reader.predict(im_file, args.save_dir, args.use_erode,
                                  args.erode_kernel, args.score_threshold,
-                                 args.seg_batch_size, args.seg_thread_num)
+                                 args.seg_batch_size)
     elif args.use_camera:
         cap_video = cv2.VideoCapture(args.camera_id)
         if not cap_video.isOpened():
@@ -347,7 +339,7 @@ def infer(args):
             if ret:
                 meter_reader.predict(frame, args.save_dir, args.use_erode,
                                      args.erode_kernel, args.score_threshold,
-                                     args.seg_batch_size, args.seg_thread_num)
+                                     args.seg_batch_size)
                 if cv2.waitKey(1) & 0xFF == ord('q'):
                     break
             else:

paddlex/__init__.py

@@ -57,4 +57,4 @@ log_level = 2
 
 from . import interpret
 
-__version__ = '1.1.1'
+__version__ = '1.1.4'

paddlex/command.py

@@ -51,6 +51,12 @@ def arg_parser():
         action="store_true",
         default=False,
         help="export onnx model for deployment")
+    parser.add_argument(
+        "--onnx_opset",
+        "-oo",
+        type=int,
+        default=10,
+        help="when use paddle2onnx, set onnx opset version to export")
     parser.add_argument(
         "--data_conversion",
         "-dc",
@@ -162,7 +168,7 @@ def main():
             logging.error(
                 "paddlex --export_inference --model_dir model_path --save_dir infer_model"
             )
-        pdx.convertor.export_onnx_model(model, args.save_dir)
+        pdx.convertor.export_onnx_model(model, args.save_dir, args.onnx_opset)
 
     if args.data_conversion:
         assert args.source is not None, "--source should be defined while converting dataset"
@@ -183,7 +189,7 @@ def main():
 
     if args.split_dataset:
         assert args.dataset_dir is not None, "--dataset_dir should be defined while spliting dataset"
-        assert args.format is not None, "--form should be defined while spliting dataset"
+        assert args.format is not None, "--format should be defined while spliting dataset"
         assert args.val_value is not None, "--val_value should be defined while spliting dataset"
 
         dataset_dir = args.dataset_dir

paddlex/convertor.py

@@ -29,10 +29,12 @@ def export_onnx(model_dir, save_dir, fixed_input_shape):
     export_onnx_model(model, save_dir)
 
 
-def export_onnx_model(model, save_dir):
-    if model.model_type == "detector" or model.__class__.__name__ == "FastSCNN":
+def export_onnx_model(model, save_dir, opset_version=10):
+    if model.__class__.__name__ == "FastSCNN" or (
+            model.model_type == "detector" and
+            model.__class__.__name__ != "YOLOv3"):
         logging.error(
-            "Only image classifier models and semantic segmentation models(except FastSCNN) are supported to export to ONNX"
+            "Only image classifier models, detection models(YOLOv3) and semantic segmentation models(except FastSCNN) are supported to export to ONNX"
         )
     try:
         import x2paddle
@@ -41,6 +43,406 @@ def export_onnx_model(model, save_dir):
     except:
         logging.error(
             "You need to install x2paddle first, pip install x2paddle>=0.7.4")
-    from x2paddle.op_mapper.paddle_op_mapper import PaddleOpMapper
+    if opset_version == 10 and model.__class__.__name__ == "YOLOv3":
+        logging.warning(
+            "Export for openVINO by default, the output of multiclass_nms exported to onnx will contains background. If you need onnx completely consistent with paddle, please use X2Paddle to export"
+        )
+        x2paddle.op_mapper.paddle2onnx.opset10.paddle_custom_layer.multiclass_nms.multiclass_nms = multiclass_nms_for_openvino
+    from x2paddle.op_mapper.paddle2onnx.paddle_op_mapper import PaddleOpMapper
     mapper = PaddleOpMapper()
-    mapper.convert(model.test_prog, save_dir)
+    mapper.convert(
+        model.test_prog,
+        save_dir,
+        scope=model.scope,
+        opset_version=opset_version)
+
+
+def multiclass_nms_for_openvino(op, block):
+    """
+    Convert the paddle multiclass_nms to onnx op.
+    This op is get the select boxes from origin boxes.
+    This op is for OpenVINO, which donn't support dynamic shape).
+    """
+    import math
+    import sys
+    import numpy as np
+    import paddle.fluid.core as core
+    import paddle.fluid as fluid
+    import onnx
+    import warnings
+    from onnx import helper, onnx_pb
+    inputs = dict()
+    outputs = dict()
+    attrs = dict()
+    for name in op.input_names:
+        inputs[name] = op.input(name)
+    for name in op.output_names:
+        outputs[name] = op.output(name)
+    for name in op.attr_names:
+        attrs[name] = op.attr(name)
+
+    result_name = outputs['Out'][0]
+    background = attrs['background_label']
+    normalized = attrs['normalized']
+    if normalized == False:
+        warnings.warn(
+            'The parameter normalized of multiclass_nms OP of Paddle is False, which has diff with ONNX. \
+                         Please set normalized=True in multiclass_nms of Paddle'
+        )
+
+    #convert the paddle attribute to onnx tensor
+    name_score_threshold = [outputs['Out'][0] + "@score_threshold"]
+    name_iou_threshold = [outputs['Out'][0] + "@iou_threshold"]
+    name_keep_top_k = [outputs['Out'][0] + '@keep_top_k']
+    name_keep_top_k_2D = [outputs['Out'][0] + '@keep_top_k_1D']
+
+    node_score_threshold = onnx.helper.make_node(
+        'Constant',
+        inputs=[],
+        outputs=name_score_threshold,
+        value=onnx.helper.make_tensor(
+            name=name_score_threshold[0] + "@const",
+            data_type=onnx.TensorProto.FLOAT,
+            dims=(),
+            vals=[float(attrs['score_threshold'])]))
+
+    node_iou_threshold = onnx.helper.make_node(
+        'Constant',
+        inputs=[],
+        outputs=name_iou_threshold,
+        value=onnx.helper.make_tensor(
+            name=name_iou_threshold[0] + "@const",
+            data_type=onnx.TensorProto.FLOAT,
+            dims=(),
+            vals=[float(attrs['nms_threshold'])]))
+
+    node_keep_top_k = onnx.helper.make_node(
+        'Constant',
+        inputs=[],
+        outputs=name_keep_top_k,
+        value=onnx.helper.make_tensor(
+            name=name_keep_top_k[0] + "@const",
+            data_type=onnx.TensorProto.INT64,
+            dims=(),
+            vals=[np.int64(attrs['keep_top_k'])]))
+
+    node_keep_top_k_2D = onnx.helper.make_node(
+        'Constant',
+        inputs=[],
+        outputs=name_keep_top_k_2D,
+        value=onnx.helper.make_tensor(
+            name=name_keep_top_k_2D[0] + "@const",
+            data_type=onnx.TensorProto.INT64,
+            dims=[1, 1],
+            vals=[np.int64(attrs['keep_top_k'])]))
+
+    # the paddle data format is x1,y1,x2,y2
+    kwargs = {'center_point_box': 0}
+
+    name_select_nms = [outputs['Out'][0] + "@select_index"]
+    node_select_nms= onnx.helper.make_node(
+        'NonMaxSuppression',
+        inputs=inputs['BBoxes'] + inputs['Scores'] + name_keep_top_k +\
+            name_iou_threshold + name_score_threshold,
+        outputs=name_select_nms)
+    # step 1 nodes select the nms class
+    node_list = [
+        node_score_threshold, node_iou_threshold, node_keep_top_k,
+        node_keep_top_k_2D, node_select_nms
+    ]
+
+    # create some const value to use
+    name_const_value = [result_name+"@const_0",
+        result_name+"@const_1",\
+        result_name+"@const_2",\
+        result_name+"@const_-1"]
+    value_const_value = [0, 1, 2, -1]
+    for name, value in zip(name_const_value, value_const_value):
+        node = onnx.helper.make_node(
+            'Constant',
+            inputs=[],
+            outputs=[name],
+            value=onnx.helper.make_tensor(
+                name=name + "@const",
+                data_type=onnx.TensorProto.INT64,
+                dims=[1],
+                vals=[value]))
+        node_list.append(node)
+
+    # In this code block, we will deocde the raw score data, reshape N * C * M to 1 * N*C*M
+    # and the same time, decode the select indices to 1 * D, gather the select_indices
+    outputs_gather_1_ = [result_name + "@gather_1_"]
+    node_gather_1_ = onnx.helper.make_node(
+        'Gather',
+        inputs=name_select_nms + [result_name + "@const_1"],
+        outputs=outputs_gather_1_,
+        axis=1)
+    node_list.append(node_gather_1_)
+    outputs_gather_1 = [result_name + "@gather_1"]
+    node_gather_1 = onnx.helper.make_node(
+        'Unsqueeze',
+        inputs=outputs_gather_1_,
+        outputs=outputs_gather_1,
+        axes=[0])
+    node_list.append(node_gather_1)
+
+    outputs_gather_2_ = [result_name + "@gather_2_"]
+    node_gather_2_ = onnx.helper.make_node(
+        'Gather',
+        inputs=name_select_nms + [result_name + "@const_2"],
+        outputs=outputs_gather_2_,
+        axis=1)
+    node_list.append(node_gather_2_)
+
+    outputs_gather_2 = [result_name + "@gather_2"]
+    node_gather_2 = onnx.helper.make_node(
+        'Unsqueeze',
+        inputs=outputs_gather_2_,
+        outputs=outputs_gather_2,
+        axes=[0])
+    node_list.append(node_gather_2)
+
+    # reshape scores N * C * M to (N*C*M) * 1
+    outputs_reshape_scores_rank1 = [result_name + "@reshape_scores_rank1"]
+    node_reshape_scores_rank1 = onnx.helper.make_node(
+        "Reshape",
+        inputs=inputs['Scores'] + [result_name + "@const_-1"],
+        outputs=outputs_reshape_scores_rank1)
+    node_list.append(node_reshape_scores_rank1)
+
+    # get the shape of scores
+    outputs_shape_scores = [result_name + "@shape_scores"]
+    node_shape_scores = onnx.helper.make_node(
+        'Shape', inputs=inputs['Scores'], outputs=outputs_shape_scores)
+    node_list.append(node_shape_scores)
+
+    # gather the index: 2 shape of scores
+    outputs_gather_scores_dim1 = [result_name + "@gather_scores_dim1"]
+    node_gather_scores_dim1 = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_shape_scores + [result_name + "@const_2"],
+        outputs=outputs_gather_scores_dim1,
+        axis=0)
+    node_list.append(node_gather_scores_dim1)
+
+    # mul class * M
+    outputs_mul_classnum_boxnum = [result_name + "@mul_classnum_boxnum"]
+    node_mul_classnum_boxnum = onnx.helper.make_node(
+        'Mul',
+        inputs=outputs_gather_1 + outputs_gather_scores_dim1,
+        outputs=outputs_mul_classnum_boxnum)
+    node_list.append(node_mul_classnum_boxnum)
+
+    # add class * M * index
+    outputs_add_class_M_index = [result_name + "@add_class_M_index"]
+    node_add_class_M_index = onnx.helper.make_node(
+        'Add',
+        inputs=outputs_mul_classnum_boxnum + outputs_gather_2,
+        outputs=outputs_add_class_M_index)
+    node_list.append(node_add_class_M_index)
+
+    # Squeeze the indices to 1 dim
+    outputs_squeeze_select_index = [result_name + "@squeeze_select_index"]
+    node_squeeze_select_index = onnx.helper.make_node(
+        'Squeeze',
+        inputs=outputs_add_class_M_index,
+        outputs=outputs_squeeze_select_index,
+        axes=[0, 2])
+    node_list.append(node_squeeze_select_index)
+
+    # gather the data from flatten scores
+    outputs_gather_select_scores = [result_name + "@gather_select_scores"]
+    node_gather_select_scores = onnx.helper.make_node('Gather',
+        inputs=outputs_reshape_scores_rank1 + \
+            outputs_squeeze_select_index,
+        outputs=outputs_gather_select_scores,
+        axis=0)
+    node_list.append(node_gather_select_scores)
+
+    # get nums to input TopK
+    outputs_shape_select_num = [result_name + "@shape_select_num"]
+    node_shape_select_num = onnx.helper.make_node(
+        'Shape',
+        inputs=outputs_gather_select_scores,
+        outputs=outputs_shape_select_num)
+    node_list.append(node_shape_select_num)
+
+    outputs_gather_select_num = [result_name + "@gather_select_num"]
+    node_gather_select_num = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_shape_select_num + [result_name + "@const_0"],
+        outputs=outputs_gather_select_num,
+        axis=0)
+    node_list.append(node_gather_select_num)
+
+    outputs_unsqueeze_select_num = [result_name + "@unsqueeze_select_num"]
+    node_unsqueeze_select_num = onnx.helper.make_node(
+        'Unsqueeze',
+        inputs=outputs_gather_select_num,
+        outputs=outputs_unsqueeze_select_num,
+        axes=[0])
+    node_list.append(node_unsqueeze_select_num)
+
+    outputs_concat_topK_select_num = [result_name + "@conat_topK_select_num"]
+    node_conat_topK_select_num = onnx.helper.make_node(
+        'Concat',
+        inputs=outputs_unsqueeze_select_num + name_keep_top_k_2D,
+        outputs=outputs_concat_topK_select_num,
+        axis=0)
+    node_list.append(node_conat_topK_select_num)
+
+    outputs_cast_concat_topK_select_num = [
+        result_name + "@concat_topK_select_num"
+    ]
+    node_outputs_cast_concat_topK_select_num = onnx.helper.make_node(
+        'Cast',
+        inputs=outputs_concat_topK_select_num,
+        outputs=outputs_cast_concat_topK_select_num,
+        to=6)
+    node_list.append(node_outputs_cast_concat_topK_select_num)
+    # get min(topK, num_select)
+    outputs_compare_topk_num_select = [
+        result_name + "@compare_topk_num_select"
+    ]
+    node_compare_topk_num_select = onnx.helper.make_node(
+        'ReduceMin',
+        inputs=outputs_cast_concat_topK_select_num,
+        outputs=outputs_compare_topk_num_select,
+        keepdims=0)
+    node_list.append(node_compare_topk_num_select)
+
+    # unsqueeze the indices to 1D tensor
+    outputs_unsqueeze_topk_select_indices = [
+        result_name + "@unsqueeze_topk_select_indices"
+    ]
+    node_unsqueeze_topk_select_indices = onnx.helper.make_node(
+        'Unsqueeze',
+        inputs=outputs_compare_topk_num_select,
+        outputs=outputs_unsqueeze_topk_select_indices,
+        axes=[0])
+    node_list.append(node_unsqueeze_topk_select_indices)
+
+    # cast the indices to INT64
+    outputs_cast_topk_indices = [result_name + "@cast_topk_indices"]
+    node_cast_topk_indices = onnx.helper.make_node(
+        'Cast',
+        inputs=outputs_unsqueeze_topk_select_indices,
+        outputs=outputs_cast_topk_indices,
+        to=7)
+    node_list.append(node_cast_topk_indices)
+
+    # select topk scores  indices
+    outputs_topk_select_topk_indices = [result_name + "@topk_select_topk_values",\
+        result_name + "@topk_select_topk_indices"]
+    node_topk_select_topk_indices = onnx.helper.make_node(
+        'TopK',
+        inputs=outputs_gather_select_scores + outputs_cast_topk_indices,
+        outputs=outputs_topk_select_topk_indices)
+    node_list.append(node_topk_select_topk_indices)
+
+    # gather topk label, scores, boxes
+    outputs_gather_topk_scores = [result_name + "@gather_topk_scores"]
+    node_gather_topk_scores = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_gather_select_scores +
+        [outputs_topk_select_topk_indices[1]],
+        outputs=outputs_gather_topk_scores,
+        axis=0)
+    node_list.append(node_gather_topk_scores)
+
+    outputs_gather_topk_class = [result_name + "@gather_topk_class"]
+    node_gather_topk_class = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_gather_1 + [outputs_topk_select_topk_indices[1]],
+        outputs=outputs_gather_topk_class,
+        axis=1)
+    node_list.append(node_gather_topk_class)
+
+    # gather the boxes need to gather the boxes id, then get boxes
+    outputs_gather_topk_boxes_id = [result_name + "@gather_topk_boxes_id"]
+    node_gather_topk_boxes_id = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_gather_2 + [outputs_topk_select_topk_indices[1]],
+        outputs=outputs_gather_topk_boxes_id,
+        axis=1)
+    node_list.append(node_gather_topk_boxes_id)
+
+    # squeeze the gather_topk_boxes_id to 1 dim
+    outputs_squeeze_topk_boxes_id = [result_name + "@squeeze_topk_boxes_id"]
+    node_squeeze_topk_boxes_id = onnx.helper.make_node(
+        'Squeeze',
+        inputs=outputs_gather_topk_boxes_id,
+        outputs=outputs_squeeze_topk_boxes_id,
+        axes=[0, 2])
+    node_list.append(node_squeeze_topk_boxes_id)
+
+    outputs_gather_select_boxes = [result_name + "@gather_select_boxes"]
+    node_gather_select_boxes = onnx.helper.make_node(
+        'Gather',
+        inputs=inputs['BBoxes'] + outputs_squeeze_topk_boxes_id,
+        outputs=outputs_gather_select_boxes,
+        axis=1)
+    node_list.append(node_gather_select_boxes)
+
+    # concat the final result
+    # before concat need to cast the class to float
+    outputs_cast_topk_class = [result_name + "@cast_topk_class"]
+    node_cast_topk_class = onnx.helper.make_node(
+        'Cast',
+        inputs=outputs_gather_topk_class,
+        outputs=outputs_cast_topk_class,
+        to=1)
+    node_list.append(node_cast_topk_class)
+
+    outputs_unsqueeze_topk_scores = [result_name + "@unsqueeze_topk_scores"]
+    node_unsqueeze_topk_scores = onnx.helper.make_node(
+        'Unsqueeze',
+        inputs=outputs_gather_topk_scores,
+        outputs=outputs_unsqueeze_topk_scores,
+        axes=[0, 2])
+    node_list.append(node_unsqueeze_topk_scores)
+
+    inputs_concat_final_results = outputs_cast_topk_class + outputs_unsqueeze_topk_scores +\
+        outputs_gather_select_boxes
+    outputs_sort_by_socre_results = [result_name + "@concat_topk_scores"]
+    node_sort_by_socre_results = onnx.helper.make_node(
+        'Concat',
+        inputs=inputs_concat_final_results,
+        outputs=outputs_sort_by_socre_results,
+        axis=2)
+    node_list.append(node_sort_by_socre_results)
+
+    # select topk classes indices
+    outputs_squeeze_cast_topk_class = [
+        result_name + "@squeeze_cast_topk_class"
+    ]
+    node_squeeze_cast_topk_class = onnx.helper.make_node(
+        'Squeeze',
+        inputs=outputs_cast_topk_class,
+        outputs=outputs_squeeze_cast_topk_class,
+        axes=[0, 2])
+    node_list.append(node_squeeze_cast_topk_class)
+    outputs_neg_squeeze_cast_topk_class = [
+        result_name + "@neg_squeeze_cast_topk_class"
+    ]
+    node_neg_squeeze_cast_topk_class = onnx.helper.make_node(
+        'Neg',
+        inputs=outputs_squeeze_cast_topk_class,
+        outputs=outputs_neg_squeeze_cast_topk_class)
+    node_list.append(node_neg_squeeze_cast_topk_class)
+    outputs_topk_select_classes_indices = [result_name + "@topk_select_topk_classes_scores",\
+        result_name + "@topk_select_topk_classes_indices"]
+    node_topk_select_topk_indices = onnx.helper.make_node(
+        'TopK',
+        inputs=outputs_neg_squeeze_cast_topk_class + outputs_cast_topk_indices,
+        outputs=outputs_topk_select_classes_indices)
+    node_list.append(node_topk_select_topk_indices)
+    outputs_concat_final_results = outputs['Out']
+    node_concat_final_results = onnx.helper.make_node(
+        'Gather',
+        inputs=outputs_sort_by_socre_results +
+        [outputs_topk_select_classes_indices[1]],
+        outputs=outputs_concat_final_results,
+        axis=1)
+    node_list.append(node_concat_final_results)
+    return node_list

paddlex/cv/models/base.py

@@ -23,6 +23,7 @@ import yaml
 import copy
 import json
 import functools
+import multiprocessing as mp
 import paddlex.utils.logging as logging
 from paddlex.utils import seconds_to_hms
 from paddlex.utils.utils import EarlyStop
@@ -76,6 +77,16 @@ class BaseAPI:
         self.completed_epochs = 0
         self.scope = fluid.global_scope()
 
+        # 线程池，在模型在预测时用于对输入数据以图片为单位进行并行处理
+        # 主要用于batch_predict接口
+        thread_num = mp.cpu_count() if mp.cpu_count() < 8 else 8
+        self.thread_pool = mp.pool.ThreadPool(thread_num)
+
+    def reset_thread_pool(self, thread_num):
+        self.thread_pool.close()
+        self.thread_pool.join()
+        self.thread_pool = mp.pool.ThreadPool(thread_num)
+
     def _get_single_card_bs(self, batch_size):
         if batch_size % len(self.places) == 0:
             return int(batch_size // len(self.places))
@@ -356,16 +367,6 @@ class BaseAPI:
         ]
         test_outputs = list(self.test_outputs.values())
         with fluid.scope_guard(self.scope):
-            if self.__class__.__name__ == 'MaskRCNN':
-                from paddlex.utils.save import save_mask_inference_model
-                save_mask_inference_model(
-                    dirname=save_dir,
-                    executor=self.exe,
-                    params_filename='__params__',
-                    feeded_var_names=test_input_names,
-                    target_vars=test_outputs,
-                    main_program=self.test_prog)
-            else:
             fluid.io.save_inference_model(
                 dirname=save_dir,
                 executor=self.exe,

paddlex/cv/models/classifier.py

@@ -279,16 +279,18 @@ class BaseClassifier(BaseAPI):
         return metrics
 
     @staticmethod
-    def _preprocess(images, transforms, model_type, class_name, thread_num=1):
+    def _preprocess(images, transforms, model_type, class_name, thread_pool=None):
         arrange_transforms(
             model_type=model_type,
             class_name=class_name,
             transforms=transforms,
             mode='test')
-        pool = ThreadPool(thread_num)
-        batch_data = pool.map(transforms, images)
-        pool.close()
-        pool.join()
+        if thread_pool is not None:
+            batch_data = thread_pool.map(transforms, images)
+        else:
+            batch_data = list()
+            for image in images:
+                batch_data.append(transforms(image))
         padding_batch = generate_minibatch(batch_data)
         im = np.array([data[0] for data in padding_batch])
 
@@ -344,15 +346,13 @@ class BaseClassifier(BaseAPI):
     def batch_predict(self,
                       img_file_list,
                       transforms=None,
-                      topk=1,
-                      thread_num=2):
+                      topk=1):
         """预测。
         Args:
             img_file_list(list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是图像路径
                 也可以是解码后的排列格式为（H，W，C）且类型为float32且为BGR格式的数组。
             transforms (paddlex.cls.transforms): 数据预处理操作。
             topk (int): 预测时前k个最大值。
-            thread_num (int): 并发执行各图像预处理时的线程数。
         Returns:
             list: 每个元素都为列表，表示各图像的预测结果。在各图像的预测列表中，其中元素均为字典。字典的关键字为'category_id'、'category'、'score'，
             分别对应预测类别id、预测类别标签、预测得分。
@@ -367,7 +367,7 @@ class BaseClassifier(BaseAPI):
             transforms = self.test_transforms
         im = BaseClassifier._preprocess(img_file_list, transforms,
                                         self.model_type,
-                                        self.__class__.__name__, thread_num)
+                                        self.__class__.__name__, self.thread_pool)
 
         with fluid.scope_guard(self.scope):
             result = self.exe.run(self.test_prog,

paddlex/cv/models/deeplabv3p.py

@@ -448,16 +448,18 @@ class DeepLabv3p(BaseAPI):
         return metrics
 
     @staticmethod
-    def _preprocess(images, transforms, model_type, class_name, thread_num=1):
+    def _preprocess(images, transforms, model_type, class_name, thread_pool=None):
         arrange_transforms(
             model_type=model_type,
             class_name=class_name,
             transforms=transforms,
             mode='test')
-        pool = ThreadPool(thread_num)
-        batch_data = pool.map(transforms, images)
-        pool.close()
-        pool.join()
+        if thread_pool is not None:
+            batch_data = thread_pool.map(transforms, images)
+        else:
+            batch_data = list()
+            for image in images:
+                batch_data.append(transforms(image))
         padding_batch = generate_minibatch(batch_data)
         im = np.array(
             [data[0] for data in padding_batch],
@@ -522,13 +524,12 @@ class DeepLabv3p(BaseAPI):
         preds = DeepLabv3p._postprocess(result, im_info)
         return preds[0]
 
-    def batch_predict(self, img_file_list, transforms=None, thread_num=2):
+    def batch_predict(self, img_file_list, transforms=None):
         """预测。
         Args:
             img_file_list(list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是图像路径
                 也可以是解码后的排列格式为（H，W，C）且类型为float32且为BGR格式的数组。
             transforms(paddlex.cv.transforms): 数据预处理操作。
-            thread_num (int): 并发执行各图像预处理时的线程数。
 
         Returns:
             list: 每个元素都为列表，表示各图像的预测结果。各图像的预测结果用字典表示，包含关键字'label_map'和'score_map', 'label_map'存储预测结果灰度图，
@@ -543,7 +544,7 @@ class DeepLabv3p(BaseAPI):
             transforms = self.test_transforms
         im, im_info = DeepLabv3p._preprocess(
             img_file_list, transforms, self.model_type,
-            self.__class__.__name__, thread_num)
+            self.__class__.__name__, self.thread_pool)
 
         with fluid.scope_guard(self.scope):
             result = self.exe.run(self.test_prog,

paddlex/cv/models/faster_rcnn.py

@@ -376,16 +376,18 @@ class FasterRCNN(BaseAPI):
         return metrics
 
     @staticmethod
-    def _preprocess(images, transforms, model_type, class_name, thread_num=1):
+    def _preprocess(images, transforms, model_type, class_name, thread_pool=None):
         arrange_transforms(
             model_type=model_type,
             class_name=class_name,
             transforms=transforms,
             mode='test')
-        pool = ThreadPool(thread_num)
-        batch_data = pool.map(transforms, images)
-        pool.close()
-        pool.join()
+        if thread_pool is not None:
+            batch_data = thread_pool.map(transforms, images)
+        else:
+            batch_data = list()
+            for image in images:
+                batch_data.append(transforms(image))
         padding_batch = generate_minibatch(batch_data)
         im = np.array([data[0] for data in padding_batch])
         im_resize_info = np.array([data[1] for data in padding_batch])
@@ -453,14 +455,13 @@ class FasterRCNN(BaseAPI):
 
         return preds[0]
 
-    def batch_predict(self, img_file_list, transforms=None, thread_num=2):
+    def batch_predict(self, img_file_list, transforms=None):
         """预测。
 
         Args:
             img_file_list(list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是图像路径
                 也可以是解码后的排列格式为（H，W，C）且类型为float32且为BGR格式的数组。
             transforms (paddlex.det.transforms): 数据预处理操作。
-            thread_num (int): 并发执行各图像预处理时的线程数。
 
         Returns:
             list: 每个元素都为列表，表示各图像的预测结果。在各图像的预测结果列表中，每个预测结果由预测框类别标签、
@@ -477,7 +478,7 @@ class FasterRCNN(BaseAPI):
             transforms = self.test_transforms
         im, im_resize_info, im_shape = FasterRCNN._preprocess(
             img_file_list, transforms, self.model_type,
-            self.__class__.__name__, thread_num)
+            self.__class__.__name__, self.thread_pool)
 
         with fluid.scope_guard(self.scope):
             result = self.exe.run(self.test_prog,

paddlex/cv/models/mask_rcnn.py

@@ -408,14 +408,13 @@ class MaskRCNN(FasterRCNN):
 
         return preds[0]
 
-    def batch_predict(self, img_file_list, transforms=None, thread_num=2):
+    def batch_predict(self, img_file_list, transforms=None):
         """预测。
 
         Args:
             img_file_list(list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是图像路径
                 也可以是解码后的排列格式为（H，W，C）且类型为float32且为BGR格式的数组。
             transforms (paddlex.det.transforms): 数据预处理操作。
-            thread_num (int): 并发执行各图像预处理时的线程数。
         Returns:
             dict: 每个元素都为列表，表示各图像的预测结果。在各图像的预测结果列表中，每个预测结果由预测框类别标签、预测框类别名称、
                   预测框坐标(坐标格式为[xmin, ymin, w, h]）、
@@ -432,7 +431,7 @@ class MaskRCNN(FasterRCNN):
             transforms = self.test_transforms
         im, im_resize_info, im_shape = FasterRCNN._preprocess(
             img_file_list, transforms, self.model_type,
-            self.__class__.__name__, thread_num)
+            self.__class__.__name__, self.thread_pool)
 
         with fluid.scope_guard(self.scope):
             result = self.exe.run(self.test_prog,

paddlex/cv/models/ppyolo.py

@@ -447,16 +447,18 @@ class PPYOLO(BaseAPI):
         return evaluate_metrics
 
     @staticmethod
-    def _preprocess(images, transforms, model_type, class_name, thread_num=1):
+    def _preprocess(images, transforms, model_type, class_name, thread_pool=None):
         arrange_transforms(
             model_type=model_type,
             class_name=class_name,
             transforms=transforms,
             mode='test')
-        pool = ThreadPool(thread_num)
-        batch_data = pool.map(transforms, images)
-        pool.close()
-        pool.join()
+        if thread_pool is not None:
+            batch_data = thread_pool.map(transforms, images)
+        else:
+            batch_data = list()
+            for image in images:
+                batch_data.append(transforms(image))
         padding_batch = generate_minibatch(batch_data)
         im = np.array(
             [data[0] for data in padding_batch],
@@ -520,14 +522,13 @@ class PPYOLO(BaseAPI):
                                     len(images), self.num_classes, self.labels)
         return preds[0]
 
-    def batch_predict(self, img_file_list, transforms=None, thread_num=2):
+    def batch_predict(self, img_file_list, transforms=None):
         """预测。
 
         Args:
             img_file_list (list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是图像路径，也可以是解码后的排列格式为（H，W，C）
                 且类型为float32且为BGR格式的数组。
             transforms (paddlex.det.transforms): 数据预处理操作。
-            thread_num (int): 并发执行各图像预处理时的线程数。
         Returns:
             list: 每个元素都为列表，表示各图像的预测结果。在各图像的预测结果列表中，每个预测结果由预测框类别标签、
               预测框类别名称、预测框坐标(坐标格式为[xmin, ymin, w, h]）、
@@ -543,7 +544,7 @@ class PPYOLO(BaseAPI):
             transforms = self.test_transforms
         im, im_size = PPYOLO._preprocess(img_file_list, transforms,
                                          self.model_type,
-                                         self.__class__.__name__, thread_num)
+                                         self.__class__.__name__, self.thread_pool)
 
         with fluid.scope_guard(self.scope):
             result = self.exe.run(self.test_prog,

paddlex/cv/models/slim/prune_config.py

@@ -91,7 +91,23 @@ sensitivities_data = {
     'DeepLabv3p_Xception65_aspp_decoder':
     'https://bj.bcebos.com/paddlex/slim_prune/deeplab_xception65_with_aspp_decoder.sensitivities',
     'DeepLabv3p_Xception41_aspp_decoder':
-    'https://bj.bcebos.com/paddlex/slim_prune/deeplab_xception41_with_aspp_decoder.sensitivities'
+    'https://bj.bcebos.com/paddlex/slim_prune/deeplab_xception41_with_aspp_decoder.sensitivities',
+    'HRNet_W18_Seg':
+    'https://bj.bcebos.com/paddlex/slim_prune/hrnet_w18.sensitivities',
+    'HRNet_W30_Seg':
+    'https://bj.bcebos.com/paddlex/slim_prune/hrnet_w30.sensitivities',
+    'HRNet_W32_Seg':
+    'https://bj.bcebos.com/paddlex/slim_prune/hrnet_w32.sensitivities',
+    'HRNet_W40_Seg':
+    'https://bj.bcebos.com/paddlex/slim_prune/hrnet_w40.sensitivities',
+    'HRNet_W44_Seg':
+    'https://bj.bcebos.com/paddlex/slim_prune/hrnet_w44.sensitivities',
+    'HRNet_W48_Seg':
+    'https://bj.bcebos.com/paddlex/slim_prune/hrnet_w48.sensitivities',
+    'HRNet_W64_Seg':
+    'https://bj.bcebos.com/paddlex/slim_prune/hrnet_w64.sensitivities',
+    'FastSCNN':
+    'https://bj.bcebos.com/paddlex/slim_prune/fast_scnn.sensitivities'
 }
 
 
@@ -105,6 +121,8 @@ def get_sensitivities(flag, model, save_dir):
     elif hasattr(model, 'encoder_with_aspp') or hasattr(model,
                                                         'enable_decoder'):
         model_type = model_type + '_' + 'aspp' + '_' + 'decoder'
+    if model_type.startswith('HRNet') and model.model_type == 'segmenter':
+        model_type = '{}_W{}_Seg'.format(model_type, model.width)
     if osp.isfile(flag):
         return flag
     elif flag == 'DEFAULT':
@@ -244,6 +262,28 @@ def get_prune_params(model):
             if i in prune_names:
                 prune_names.remove(i)
 
+    elif model_type.startswith('HRNet') and model.model_type == 'segmenter':
+        for param in program.global_block().all_parameters():
+            if 'weight' not in param.name:
+                continue
+            prune_names.append(param.name)
+        params_not_prune = ['conv-1_weights']
+        for i in params_not_prune:
+            if i in prune_names:
+                prune_names.remove(i)
+
+    elif model_type.startswith('FastSCNN'):
+        for param in program.global_block().all_parameters():
+            if 'weight' not in param.name:
+                continue
+            if 'dwise' in param.name or 'depthwise' in param.name or 'logit' in param.name:
+                continue
+            prune_names.append(param.name)
+        params_not_prune = ['classifier/weights']
+        for i in params_not_prune:
+            if i in prune_names:
+                prune_names.remove(i)
+
     elif model_type.startswith('DeepLabv3p'):
         for param in program.global_block().all_parameters():
             if 'weight' not in param.name:

paddlex/cv/nets/detection/yolo_v3.py

@@ -311,7 +311,7 @@ class YOLOv3:
 
     def _upsample(self, input, scale=2, name=None):
         out = fluid.layers.resize_nearest(
-            input=input, scale=float(scale), name=name)
+            input=input, scale=float(scale), name=name, align_corners=False)
         return out
 
     def _detection_block(self,

paddlex/cv/nets/hrnet.py

@@ -235,10 +235,13 @@ class HRNet(object):
                         name=name + '_layer_' + str(i + 1) + '_' + str(j + 1))
                     if self.feature_maps == "stage4":
                         y = fluid.layers.resize_bilinear(
-                            input=y, out_shape=[height, width])
+                            input=y,
+                            out_shape=[height, width],
+                            align_corners=False,
+                            align_mode=1)
                     else:
                         y = fluid.layers.resize_nearest(
-                            input=y, scale=2**(j - i))
+                            input=y, scale=2**(j - i), align_corners=False)
                     residual = fluid.layers.elementwise_add(
                         x=residual, y=y, act=None)
                 elif j < i:

paddlex/deploy.py

@@ -16,6 +16,7 @@ import os.path as osp
 import cv2
 import numpy as np
 import yaml
+import multiprocessing as mp
 import paddlex
 import paddle.fluid as fluid
 from paddlex.cv.transforms import build_transforms
@@ -79,12 +80,21 @@ class Predictor:
         self.predictor = self.create_predictor(use_gpu, gpu_id, use_mkl,
                                                mkl_thread_num, use_trt,
                                                use_glog, memory_optimize)
+        # 线程池，在模型在预测时用于对输入数据以图片为单位进行并行处理
+        # 主要用于batch_predict接口
+        thread_num = mp.cpu_count() if mp.cpu_count() < 8 else 8
+        self.thread_pool = mp.pool.ThreadPool(thread_num)
+
+    def reset_thread_pool(self, thread_num):
+        self.thread_pool.close()
+        self.thread_pool.join()
+        self.thread_pool = mp.pool.ThreadPool(thread_num)
 
     def create_predictor(self,
                          use_gpu=True,
                          gpu_id=0,
                          use_mkl=False,
-                         mkl_thread_num=4,
+                         mkl_thread_num=psutil.cpu_count(),
                          use_trt=False,
                          use_glog=False,
                          memory_optimize=True):
@@ -98,6 +108,7 @@ class Predictor:
         else:
             config.disable_gpu()
         if use_mkl:
+            if self.model_name not in ["HRNet", "DeepLabv3p"]:
                 config.enable_mkldnn()
                 config.set_cpu_math_library_num_threads(mkl_thread_num)
         if use_glog:
@@ -114,7 +125,7 @@ class Predictor:
         predictor = fluid.core.create_paddle_predictor(config)
         return predictor
 
-    def preprocess(self, image, thread_num=1):
+    def preprocess(self, image, thread_pool=None):
         """ 对图像做预处理
 
             Args:
@@ -128,7 +139,7 @@ class Predictor:
                 self.transforms,
                 self.model_type,
                 self.model_name,
-                thread_num=thread_num)
+                thread_pool=thread_pool)
             res['image'] = im
         elif self.model_type == "detector":
             if self.model_name in ["PPYOLO", "YOLOv3"]:
@@ -137,7 +148,7 @@ class Predictor:
                     self.transforms,
                     self.model_type,
                     self.model_name,
-                    thread_num=thread_num)
+                    thread_pool=thread_pool)
                 res['image'] = im
                 res['im_size'] = im_size
             if self.model_name.count('RCNN') > 0:
@@ -146,7 +157,7 @@ class Predictor:
                     self.transforms,
                     self.model_type,
                     self.model_name,
-                    thread_num=thread_num)
+                    thread_pool=thread_pool)
                 res['image'] = im
                 res['im_info'] = im_resize_info
                 res['im_shape'] = im_shape
@@ -156,7 +167,7 @@ class Predictor:
                 self.transforms,
                 self.model_type,
                 self.model_name,
-                thread_num=thread_num)
+                thread_pool=thread_pool)
             res['image'] = im
             res['im_info'] = im_info
         return res
@@ -253,17 +264,16 @@ class Predictor:
 
         return results[0]
 
-    def batch_predict(self, image_list, topk=1, thread_num=2):
+    def batch_predict(self, image_list, topk=1):
         """ 图片预测
 
             Args:
                 image_list(list|tuple): 对列表（或元组）中的图像同时进行预测，列表中的元素可以是图像路径
                     也可以是解码后的排列格式为（H，W，C）且类型为float32且为BGR格式的数组。
-                thread_num (int): 并发执行各图像预处理时的线程数。
 
                 topk(int): 分类预测时使用，表示预测前topk的结果
         """
-        preprocessed_input = self.preprocess(image_list)
+        preprocessed_input = self.preprocess(image_list, self.thread_pool)
         model_pred = self.raw_predict(preprocessed_input)
         im_shape = None if 'im_shape' not in preprocessed_input else preprocessed_input[
             'im_shape']

setup.py

@@ -19,7 +19,7 @@ long_description = "PaddlePaddle Entire Process Development Toolkit"
 
 setuptools.setup(
     name="paddlex",
-    version='1.1.1',
+    version='1.1.4',
     author="paddlex",
     author_email="paddlex@baidu.com",
     description=long_description,