fix conflict

deploy/cpp_infer/src/preprocess_op.cpp

@@ -40,6 +40,17 @@ void Permute::Run(const cv::Mat *im, float *data) {
   }
 }
 
+void PermuteBatch::Run(const std::vector<cv::Mat> imgs, float *data) {
+    for (int j = 0; j < imgs.size(); j ++){
+        int rh = imgs[j].rows;
+        int rw = imgs[j].cols;
+        int rc = imgs[j].channels();
+        for (int i = 0; i < rc; ++i) {
+            cv::extractChannel(imgs[j], cv::Mat(rh, rw, CV_32FC1, data + (j * rc + i) * rh * rw), i);
+        }
+    }
+}
+    
 void Normalize::Run(cv::Mat *im, const std::vector<float> &mean,
                     const std::vector<float> &scale, const bool is_scale) {
   double e = 1.0;
@@ -90,16 +101,17 @@ void CrnnResizeImg::Run(const cv::Mat &img, cv::Mat &resize_img, float wh_ratio,
   imgC = rec_image_shape[0];
   imgH = rec_image_shape[1];
   imgW = rec_image_shape[2];
-
+    
   imgW = int(32 * wh_ratio);
 
   float ratio = float(img.cols) / float(img.rows);
   int resize_w, resize_h;
+
   if (ceilf(imgH * ratio) > imgW)
     resize_w = imgW;
   else
     resize_w = int(ceilf(imgH * ratio));
-
+    
   cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
              cv::INTER_LINEAR);
   cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0,

deploy/cpp_infer/src/utility.cpp

@@ -147,4 +147,17 @@ cv::Mat Utility::GetRotateCropImage(const cv::Mat &srcimage,
   }
 }
 
+std::vector<int> Utility::argsort(const std::vector<float>& array)
+{
+    const int array_len(array.size());
+    std::vector<int> array_index(array_len, 0);
+    for (int i = 0; i < array_len; ++i)
+        array_index[i] = i;
+
+    std::sort(array_index.begin(), array_index.end(),
+        [&array](int pos1, int pos2) {return (array[pos1] < array[pos2]); });
+
+    return array_index;
+}
+
 } // namespace PaddleOCR
\ No newline at end of file

deploy/cpp_infer/tools/build_opencv.sh

+root_path="/paddle/PaddleOCR/deploy/cpp_infer/opencv-3.4.7"
+install_path=${root_path}/opencv3
+build_dir=${root_path}/build
+
+rm -rf ${build_dir}
+mkdir ${build_dir}
+cd ${build_dir}
+
+cmake .. \
+    -DCMAKE_INSTALL_PREFIX=${install_path} \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DBUILD_SHARED_LIBS=OFF \
+    -DWITH_IPP=OFF \
+    -DBUILD_IPP_IW=OFF \
+    -DWITH_LAPACK=OFF \
+    -DWITH_EIGEN=OFF \
+    -DCMAKE_INSTALL_LIBDIR=lib64 \
+    -DWITH_ZLIB=ON \
+    -DBUILD_ZLIB=ON \
+    -DWITH_JPEG=ON \
+    -DBUILD_JPEG=ON \
+    -DWITH_PNG=ON \
+    -DBUILD_PNG=ON \
+    -DWITH_TIFF=ON \
+    -DBUILD_TIFF=ON
+
+make -j
+make install

deploy/lite/ocr_db_crnn.cc

@@ -12,12 +12,14 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle_api.h" // NOLINT
 #include <chrono>
+#include "paddle_api.h" // NOLINT
+#include "paddle_place.h"
 
 #include "cls_process.h"
 #include "crnn_process.h"
 #include "db_post_process.h"
+#include "AutoLog/auto_log/lite_autolog.h"
 
 using namespace paddle::lite_api; // NOLINT
 using namespace std;
@@ -27,7 +29,7 @@ void NeonMeanScale(const float *din, float *dout, int size,
                    const std::vector<float> mean,
                    const std::vector<float> scale) {
   if (mean.size() != 3 || scale.size() != 3) {
-    std::cerr << "[ERROR] mean or scale size must equal to 3\n";
+    std::cerr << "[ERROR] mean or scale size must equal to 3" << std::endl;
     exit(1);
   }
   float32x4_t vmean0 = vdupq_n_f32(mean[0]);
@@ -159,7 +161,8 @@ void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
                  std::vector<float> &rec_text_score,
                  std::vector<std::string> charactor_dict,
                  std::shared_ptr<PaddlePredictor> predictor_cls,
-                 int use_direction_classify) {
+                 int use_direction_classify,
+                 std::vector<double> *times) {
   std::vector<float> mean = {0.5f, 0.5f, 0.5f};
   std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
 
@@ -169,7 +172,10 @@ void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
   cv::Mat resize_img;
 
   int index = 0;
+
+  std::vector<double> time_info = {0, 0, 0};
   for (int i = boxes.size() - 1; i >= 0; i--) {
+    auto preprocess_start = std::chrono::steady_clock::now();
     crop_img = GetRotateCropImage(srcimg, boxes[i]);
     if (use_direction_classify >= 1) {
       crop_img = RunClsModel(crop_img, predictor_cls);
@@ -188,7 +194,9 @@ void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
     auto *data0 = input_tensor0->mutable_data<float>();
 
     NeonMeanScale(dimg, data0, resize_img.rows * resize_img.cols, mean, scale);
+    auto preprocess_end = std::chrono::steady_clock::now();
     //// Run CRNN predictor
+    auto inference_start = std::chrono::steady_clock::now();
     predictor_crnn->Run();
 
     // Get output and run postprocess
@@ -196,8 +204,10 @@ void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
         std::move(predictor_crnn->GetOutput(0)));
     auto *predict_batch = output_tensor0->data<float>();
     auto predict_shape = output_tensor0->shape();
+    auto inference_end = std::chrono::steady_clock::now();
 
     // ctc decode
+    auto postprocess_start = std::chrono::steady_clock::now();
     std::string str_res;
     int argmax_idx;
     int last_index = 0;
@@ -221,19 +231,33 @@ void RunRecModel(std::vector<std::vector<std::vector<int>>> boxes, cv::Mat img,
     score /= count;
     rec_text.push_back(str_res);
     rec_text_score.push_back(score);
+    auto postprocess_end = std::chrono::steady_clock::now();
+
+    std::chrono::duration<float> preprocess_diff = preprocess_end - preprocess_start;
+    time_info[0] += double(preprocess_diff.count() * 1000);
+    std::chrono::duration<float> inference_diff = inference_end - inference_start;
+    time_info[1] += double(inference_diff.count() * 1000);
+    std::chrono::duration<float> postprocess_diff = postprocess_end - postprocess_start;
+    time_info[2] += double(postprocess_diff.count() * 1000);
+
   }
+
+times->push_back(time_info[0]);
+times->push_back(time_info[1]);
+times->push_back(time_info[2]);
 }
 
 std::vector<std::vector<std::vector<int>>>
 RunDetModel(std::shared_ptr<PaddlePredictor> predictor, cv::Mat img,
-            std::map<std::string, double> Config) {
+            std::map<std::string, double> Config, std::vector<double> *times) {
   // Read img
   int max_side_len = int(Config["max_side_len"]);
   int det_db_use_dilate = int(Config["det_db_use_dilate"]);
 
   cv::Mat srcimg;
   img.copyTo(srcimg);
-
+  
+  auto preprocess_start = std::chrono::steady_clock::now();
   std::vector<float> ratio_hw;
   img = DetResizeImg(img, max_side_len, ratio_hw);
   cv::Mat img_fp;
@@ -248,8 +272,10 @@ RunDetModel(std::shared_ptr<PaddlePredictor> predictor, cv::Mat img,
   std::vector<float> scale = {1 / 0.229f, 1 / 0.224f, 1 / 0.225f};
   const float *dimg = reinterpret_cast<const float *>(img_fp.data);
   NeonMeanScale(dimg, data0, img_fp.rows * img_fp.cols, mean, scale);
+  auto preprocess_end = std::chrono::steady_clock::now();
 
   // Run predictor
+  auto inference_start = std::chrono::steady_clock::now();
   predictor->Run();
 
   // Get output and post process
@@ -257,8 +283,10 @@ RunDetModel(std::shared_ptr<PaddlePredictor> predictor, cv::Mat img,
       std::move(predictor->GetOutput(0)));
   auto *outptr = output_tensor->data<float>();
   auto shape_out = output_tensor->shape();
+  auto inference_end = std::chrono::steady_clock::now();
 
   // Save output
+  auto postprocess_start = std::chrono::steady_clock::now();
   float pred[shape_out[2] * shape_out[3]];
   unsigned char cbuf[shape_out[2] * shape_out[3]];
 
@@ -287,14 +315,23 @@ RunDetModel(std::shared_ptr<PaddlePredictor> predictor, cv::Mat img,
 
   std::vector<std::vector<std::vector<int>>> filter_boxes =
       FilterTagDetRes(boxes, ratio_hw[0], ratio_hw[1], srcimg);
+  auto postprocess_end = std::chrono::steady_clock::now();
+
+  std::chrono::duration<float> preprocess_diff = preprocess_end - preprocess_start;
+  times->push_back(double(preprocess_diff.count() * 1000));
+  std::chrono::duration<float> inference_diff = inference_end - inference_start;
+  times->push_back(double(inference_diff.count() * 1000));
+  std::chrono::duration<float> postprocess_diff = postprocess_end - postprocess_start;
+  times->push_back(double(postprocess_diff.count() * 1000));
 
   return filter_boxes;
 }
 
-std::shared_ptr<PaddlePredictor> loadModel(std::string model_file) {
+std::shared_ptr<PaddlePredictor> loadModel(std::string model_file, int num_threads) {
   MobileConfig config;
   config.set_model_from_file(model_file);
 
+  config.set_threads(num_threads);
   std::shared_ptr<PaddlePredictor> predictor =
       CreatePaddlePredictor<MobileConfig>(config);
   return predictor;
@@ -354,60 +391,285 @@ std::map<std::string, double> LoadConfigTxt(std::string config_path) {
   return dict;
 }
 
-int main(int argc, char **argv) {
-  if (argc < 5) {
-    std::cerr << "[ERROR] usage: " << argv[0]
-              << " det_model_file cls_model_file rec_model_file image_path "
-                 "charactor_dict\n";
+void check_params(int argc, char **argv) {
+  if (argc<=1 || (strcmp(argv[1], "det")!=0 && strcmp(argv[1], "rec")!=0 && strcmp(argv[1], "system")!=0)) {
+    std::cerr << "Please choose one mode of [det, rec, system] !" << std::endl;
     exit(1);
   }
-  std::string det_model_file = argv[1];
-  std::string rec_model_file = argv[2];
-  std::string cls_model_file = argv[3];
-  std::string img_path = argv[4];
-  std::string dict_path = argv[5];
+  if (strcmp(argv[1], "det") == 0) {
+      if (argc < 9){
+        std::cerr << "[ERROR] usage:" << argv[0]
+                  << " det det_model runtime_device num_threads batchsize img_dir det_config lite_benchmark_value" << std::endl;
+        exit(1);
+      }
+  }
+
+  if (strcmp(argv[1], "rec") == 0) {
+      if (argc < 9){
+        std::cerr << "[ERROR] usage:" << argv[0]
+                  << " rec rec_model runtime_device num_threads batchsize img_dir key_txt lite_benchmark_value" << std::endl;
+        exit(1);
+      }
+  }
+
+  if (strcmp(argv[1], "system") == 0) {
+      if (argc < 12){
+        std::cerr << "[ERROR] usage:" << argv[0]
+                  << " system det_model rec_model clas_model runtime_device num_threads batchsize img_dir det_config key_txt lite_benchmark_value" << std::endl;
+        exit(1);
+      }
+  }
+}
+
+void system(char **argv){
+  std::string det_model_file = argv[2];
+  std::string rec_model_file = argv[3];
+  std::string cls_model_file = argv[4];
+  std::string runtime_device = argv[5];
+  std::string precision = argv[6];
+  std::string num_threads = argv[7];
+  std::string batchsize = argv[8];
+  std::string img_dir = argv[9];
+  std::string det_config_path = argv[10];
+  std::string dict_path = argv[11];
+
+  if (strcmp(argv[6], "FP32") != 0 && strcmp(argv[6], "INT8") != 0) {
+      std::cerr << "Only support FP32 or INT8." << std::endl;
+      exit(1);
+  }
+
+  std::vector<cv::String> cv_all_img_names;
+  cv::glob(img_dir, cv_all_img_names);
 
   //// load config from txt file
-  auto Config = LoadConfigTxt("./config.txt");
+  auto Config = LoadConfigTxt(det_config_path);
   int use_direction_classify = int(Config["use_direction_classify"]);
 
-  auto start = std::chrono::system_clock::now();
+  auto charactor_dict = ReadDict(dict_path);
+  charactor_dict.insert(charactor_dict.begin(), "#"); // blank char for ctc
+  charactor_dict.push_back(" ");
+
+  auto det_predictor = loadModel(det_model_file, std::stoi(num_threads));
+  auto rec_predictor = loadModel(rec_model_file, std::stoi(num_threads));
+  auto cls_predictor = loadModel(cls_model_file, std::stoi(num_threads));
+
+  std::vector<double> det_time_info = {0, 0, 0};
+  std::vector<double> rec_time_info = {0, 0, 0};
 
-  auto det_predictor = loadModel(det_model_file);
-  auto rec_predictor = loadModel(rec_model_file);
-  auto cls_predictor = loadModel(cls_model_file);
+  for (int i = 0; i < cv_all_img_names.size(); ++i) {
+    std::cout << "The predict img: " << cv_all_img_names[i] << std::endl;
+    cv::Mat srcimg = cv::imread(cv_all_img_names[i], cv::IMREAD_COLOR);
+
+    if (!srcimg.data) {
+      std::cerr << "[ERROR] image read failed! image path: " << cv_all_img_names[i] << std::endl;
+      exit(1);
+    }
+
+    std::vector<double> det_times;
+    auto boxes = RunDetModel(det_predictor, srcimg, Config, &det_times);
+  
+    std::vector<std::string> rec_text;
+    std::vector<float> rec_text_score;
+  
+    std::vector<double> rec_times;
+    RunRecModel(boxes, srcimg, rec_predictor, rec_text, rec_text_score,
+                charactor_dict, cls_predictor, use_direction_classify, &rec_times);
+  
+    //// visualization
+    auto img_vis = Visualization(srcimg, boxes);
+  
+    //// print recognized text
+    for (int i = 0; i < rec_text.size(); i++) {
+      std::cout << i << "\t" << rec_text[i] << "\t" << rec_text_score[i]
+                <<  std::endl;
+
+    }
+
+    det_time_info[0] += det_times[0];
+    det_time_info[1] += det_times[1];
+    det_time_info[2] += det_times[2];
+    rec_time_info[0] += rec_times[0];
+    rec_time_info[1] += rec_times[1];
+    rec_time_info[2] += rec_times[2];
+  }
+  if (strcmp(argv[12], "True") == 0) {
+    AutoLogger autolog_det(det_model_file, 
+                       runtime_device,
+                       std::stoi(num_threads),
+                       std::stoi(batchsize), 
+                       "dynamic", 
+                       precision, 
+                       det_time_info, 
+                       cv_all_img_names.size());
+    AutoLogger autolog_rec(rec_model_file, 
+                       runtime_device,
+                       std::stoi(num_threads),
+                       std::stoi(batchsize), 
+                       "dynamic", 
+                       precision, 
+                       rec_time_info, 
+                       cv_all_img_names.size());
+
+    autolog_det.report();
+    std::cout << std::endl;
+    autolog_rec.report();
+  }
+}
+
+void det(int argc, char **argv) {
+  std::string det_model_file = argv[2];
+  std::string runtime_device = argv[3];
+  std::string precision = argv[4];
+  std::string num_threads = argv[5];
+  std::string batchsize = argv[6];
+  std::string img_dir = argv[7];
+  std::string det_config_path = argv[8];
+
+  if (strcmp(argv[4], "FP32") != 0 && strcmp(argv[4], "INT8") != 0) {
+      std::cerr << "Only support FP32 or INT8." << std::endl;
+      exit(1);
+  }
+
+  std::vector<cv::String> cv_all_img_names;
+  cv::glob(img_dir, cv_all_img_names);
+
+  //// load config from txt file
+  auto Config = LoadConfigTxt(det_config_path);
+
+  auto det_predictor = loadModel(det_model_file, std::stoi(num_threads));
+
+  std::vector<double> time_info = {0, 0, 0};
+  for (int i = 0; i < cv_all_img_names.size(); ++i) {
+    std::cout << "The predict img: " << cv_all_img_names[i] << std::endl;
+    cv::Mat srcimg = cv::imread(cv_all_img_names[i], cv::IMREAD_COLOR);
+
+    if (!srcimg.data) {
+      std::cerr << "[ERROR] image read failed! image path: " << cv_all_img_names[i] << std::endl;
+      exit(1);
+    }
+
+    std::vector<double> times;
+    auto boxes = RunDetModel(det_predictor, srcimg, Config, &times);
+
+    //// visualization
+    auto img_vis = Visualization(srcimg, boxes);
+    std::cout << boxes.size() << " bboxes have detected:" << std::endl;
+
+    for (int i=0; i<boxes.size(); i++){
+      std::cout << "The " << i << " box:" << std::endl;
+      for (int j=0; j<4; j++){
+        for (int k=0; k<2; k++){
+          std::cout << boxes[i][j][k] << "\t";
+        }
+      }
+      std::cout << std::endl;
+    }
+    time_info[0] += times[0];
+    time_info[1] += times[1];
+    time_info[2] += times[2];
+  }
+
+  if (strcmp(argv[9], "True") == 0) {
+    AutoLogger autolog(det_model_file, 
+                       runtime_device,
+                       std::stoi(num_threads),
+                       std::stoi(batchsize), 
+                       "dynamic", 
+                       precision, 
+                       time_info, 
+                       cv_all_img_names.size());
+    autolog.report();
+  }
+}
+
+void rec(int argc, char **argv) {
+  std::string rec_model_file = argv[2];
+  std::string runtime_device = argv[3];
+  std::string precision = argv[4];
+  std::string num_threads = argv[5];
+  std::string batchsize = argv[6];
+  std::string img_dir = argv[7];
+  std::string dict_path = argv[8];
+
+  if (strcmp(argv[4], "FP32") != 0 && strcmp(argv[4], "INT8") != 0) {
+      std::cerr << "Only support FP32 or INT8." << std::endl;
+      exit(1);
+  }
+
+  std::vector<cv::String> cv_all_img_names;
+  cv::glob(img_dir, cv_all_img_names);
 
   auto charactor_dict = ReadDict(dict_path);
   charactor_dict.insert(charactor_dict.begin(), "#"); // blank char for ctc
   charactor_dict.push_back(" ");
 
-  cv::Mat srcimg = cv::imread(img_path, cv::IMREAD_COLOR);
-  auto boxes = RunDetModel(det_predictor, srcimg, Config);
+  auto rec_predictor = loadModel(rec_model_file, std::stoi(num_threads));
 
-  std::vector<std::string> rec_text;
-  std::vector<float> rec_text_score;
+  std::shared_ptr<PaddlePredictor> cls_predictor;
 
-  RunRecModel(boxes, srcimg, rec_predictor, rec_text, rec_text_score,
-              charactor_dict, cls_predictor, use_direction_classify);
+  std::vector<double> time_info = {0, 0, 0};
+  for (int i = 0; i < cv_all_img_names.size(); ++i) {
+    std::cout << "The predict img: " << cv_all_img_names[i] << std::endl;
+    cv::Mat srcimg = cv::imread(cv_all_img_names[i], cv::IMREAD_COLOR);
 
-  auto end = std::chrono::system_clock::now();
-  auto duration =
-      std::chrono::duration_cast<std::chrono::microseconds>(end - start);
+    if (!srcimg.data) {
+      std::cerr << "[ERROR] image read failed! image path: " << cv_all_img_names[i] << std::endl;
+      exit(1);
+    }
 
-  //// visualization
-  auto img_vis = Visualization(srcimg, boxes);
+    int width = srcimg.cols;
+    int height = srcimg.rows;
+    std::vector<int> upper_left = {0, 0};
+    std::vector<int> upper_right = {width, 0};
+    std::vector<int> lower_right = {width, height};
+    std::vector<int> lower_left  = {0, height};
+    std::vector<std::vector<int>> box = {upper_left, upper_right, lower_right, lower_left};
+    std::vector<std::vector<std::vector<int>>> boxes = {box};
+
+    std::vector<std::string> rec_text;
+    std::vector<float> rec_text_score;
+    std::vector<double> times;
+    RunRecModel(boxes, srcimg, rec_predictor, rec_text, rec_text_score,
+                charactor_dict, cls_predictor, 0, &times);
+  
+    //// print recognized text
+    for (int i = 0; i < rec_text.size(); i++) {
+      std::cout << i << "\t" << rec_text[i] << "\t" << rec_text_score[i]
+                << std::endl;
+    }
+    time_info[0] += times[0];
+    time_info[1] += times[1];
+    time_info[2] += times[2];
+  }
+  // TODO: support autolog
+  if (strcmp(argv[9], "True") == 0) {
+    AutoLogger autolog(rec_model_file, 
+                       runtime_device,
+                       std::stoi(num_threads),
+                       std::stoi(batchsize), 
+                       "dynamic", 
+                       precision, 
+                       time_info, 
+                       cv_all_img_names.size());
+    autolog.report();
+  }
+}
+
+int main(int argc, char **argv) {
+  check_params(argc, argv);
+  std::cout << "mode: " << argv[1] << endl;
 
-  //// print recognized text
-  for (int i = 0; i < rec_text.size(); i++) {
-    std::cout << i << "\t" << rec_text[i] << "\t" << rec_text_score[i]
-              << std::endl;
+  if (strcmp(argv[1], "system") == 0) {
+    system(argv);
   }
 
-  std::cout << "花费了"
-            << double(duration.count()) *
-                   std::chrono::microseconds::period::num /
-                   std::chrono::microseconds::period::den
-            << "秒" << std::endl;
+  if (strcmp(argv[1], "det") == 0) {
+    det(argc, argv);
+  }
+
+  if (strcmp(argv[1], "rec") == 0) {
+    rec(argc, argv);
+  }
 
   return 0;
-}
+}
\ No newline at end of file

deploy/paddle2onnx/readme.md

+# paddle2onnx 模型转化与预测
+
+本章节介绍 PaddleOCR 模型如何转化为 ONNX 模型，并基于 ONNX 引擎预测。
+
+## 1. 环境准备
+
+需要准备 Paddle2ONNX 模型转化环境，和 ONNX 模型预测环境
+
+###  Paddle2ONNX
+
+Paddle2ONNX 支持将 PaddlePaddle 模型格式转化到 ONNX 模型格式，算子目前稳定支持导出 ONNX Opset 9~11，部分Paddle算子支持更低的ONNX Opset转换。
+更多细节可参考 [Paddle2ONNX](https://github.com/PaddlePaddle/Paddle2ONNX/blob/develop/README_zh.md)
+
+- 安装 Paddle2ONNX
+```
+python3.7 -m pip install paddle2onnx
+```
+
+- 安装 ONNX
+```
+# 建议安装 1.4.0 版本，可根据环境更换版本号
+python3.7 -m pip install onnxruntime==1.4.0
+```
+
+## 2. 模型转换
+
+
+- Paddle 模型下载
+
+有两种方式获取Paddle静态图模型：在 [model_list](../../doc/doc_ch/models_list.md) 中下载PaddleOCR提供的预测模型；
+参考[模型导出说明](../../doc/doc_ch/inference.md#训练模型转inference模型)把训练好的权重转为 inference_model。
+
+以 ppocr 检测模型为例：
+
+```
+wget -nc  -P ./inference https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_infer.tar
+cd ./inference && tar xf ch_ppocr_mobile_v2.0_det_infer.tar && cd ..
+```
+
+- 模型转换
+
+使用 Paddle2ONNX 将Paddle静态图模型转换为ONNX模型格式：
+
+```
+paddle2onnx --model_dir=./inference/ch_ppocr_mobile_v2.0_det_infer/ \
+--model_filename=inference.pdmodel \
+--params_filename=inference.pdiparams \
+--save_file=./inference/det_mobile_onnx/model.onnx \
+--opset_version=10 \
+--enable_onnx_checker=True
+```
+
+执行完毕后，ONNX 模型会被保存在 `./inference/det_mobile_onnx/` 路径下
+
+* 注意：以下几个模型暂不支持转换为 ONNX 模型：
+NRTR、SAR、RARE、SRN
+
+## 3. onnx 预测
+
+以检测模型为例，使用 ONNX 预测可执行如下命令：
+
+```
+python3.7 ../../tools/infer/predict_det.py --use_gpu=False --use_onnx=True \
+--det_model_dir=./inference/det_mobile_onnx/model.onnx \
+--image_dir=../../doc/imgs/1.jpg
+```
+
+执行命令后在终端会打印出预测的检测框坐标，并在 `./inference_results/` 下保存可视化结果。
+
+```
+root INFO: 1.jpg  [[[291, 295], [334, 292], [348, 844], [305, 847]], [[344, 296], [379, 294], [387, 669], [353, 671]]]
+The predict time of ../../doc/imgs/1.jpg: 0.06162881851196289
+The visualized image saved in ./inference_results/det_res_1.jpg
+```
+
+* 注意：ONNX暂时不支持变长预测，需要将输入resize到固定输入，预测结果可能与直接使用Paddle预测有细微不同。

deploy/pdserving/README.md

@@ -114,7 +114,7 @@ The recognition model is the same.
     git clone https://github.com/PaddlePaddle/PaddleOCR
 
     # Enter the working directory  
-    cd PaddleOCR/deploy/pdserver/
+    cd PaddleOCR/deploy/pdserving/
     ```
 
     The pdserver directory contains the code to start the pipeline service and send prediction requests, including:

deploy/pdserving/README_CN.md

@@ -112,7 +112,7 @@ python3 -m paddle_serving_client.convert --dirname ./ch_ppocr_mobile_v2.0_rec_in
     git clone https://github.com/PaddlePaddle/PaddleOCR
 
     # 进入到工作目录
-    cd PaddleOCR/deploy/pdserver/
+    cd PaddleOCR/deploy/pdserving/
     ```
     pdserver目录包含启动pipeline服务和发送预测请求的代码，包括：
     ```
@@ -206,7 +206,7 @@ pip3 install paddle-serving-app==0.3.1
 1. 启动服务端程序
 
 ```
-cd win 
+cd win
 python3 ocr_web_server.py gpu(使用gpu方式)
 或者
 python3 ocr_web_server.py cpu(使用cpu方式)

deploy/pdserving/ocr_cpp_client.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.
+# pylint: disable=doc-string-missing
+
+from paddle_serving_client import Client
+import sys
+import numpy as np
+import base64
+import os
+import cv2
+from paddle_serving_app.reader import Sequential, URL2Image, ResizeByFactor
+from paddle_serving_app.reader import Div, Normalize, Transpose
+from ocr_reader import OCRReader
+
+client = Client()
+# TODO:load_client need to load more than one client model.
+# this need to figure out some details.
+client.load_client_config(sys.argv[1:])
+client.connect(["127.0.0.1:9293"])
+
+import paddle
+test_img_dir = "test_img/"
+
+ocr_reader = OCRReader(char_dict_path="../../ppocr/utils/ppocr_keys_v1.txt")
+
+
+def cv2_to_base64(image):
+    return base64.b64encode(image).decode(
+        'utf8')  #data.tostring()).decode('utf8')
+
+
+for img_file in os.listdir(test_img_dir):
+    with open(os.path.join(test_img_dir, img_file), 'rb') as file:
+        image_data = file.read()
+    image = cv2_to_base64(image_data)
+    res_list = []
+    #print(image)
+    fetch_map = client.predict(
+        feed={"x": image}, fetch=["save_infer_model/scale_0.tmp_1"], batch=True)
+    print("fetrch map:", fetch_map)
+    one_batch_res = ocr_reader.postprocess(fetch_map, with_score=True)
+    for res in one_batch_res:
+        res_list.append(res[0])
+    res = {"res": str(res_list)}
+    print(res)

deploy/pdserving/pipeline_http_client.py

@@ -18,13 +18,19 @@ import json
 import base64
 import os
 
+import argparse
+parser = argparse.ArgumentParser(description="args for paddleserving")
+parser.add_argument("--image_dir", type=str, default="../../doc/imgs/")
+args = parser.parse_args()
+
 
 def cv2_to_base64(image):
     return base64.b64encode(image).decode('utf8')
 
 
 url = "http://127.0.0.1:9998/ocr/prediction"
-test_img_dir = "../../doc/imgs/"
+test_img_dir = args.image_dir
+
 for idx, img_file in enumerate(os.listdir(test_img_dir)):
     with open(os.path.join(test_img_dir, img_file), 'rb') as file:
         image_data1 = file.read()
@@ -36,5 +42,4 @@ for idx, img_file in enumerate(os.listdir(test_img_dir)):
         r = requests.post(url=url, data=json.dumps(data))
         print(r.json())
 
-test_img_dir = "../../doc/imgs/"
 print("==> total number of test imgs: ", len(os.listdir(test_img_dir)))

deploy/pdserving/pipeline_rpc_client.py

@@ -30,12 +30,17 @@ def cv2_to_base64(image):
     return base64.b64encode(image).decode('utf8')
 
 
-test_img_dir = "imgs/"
+import argparse
+parser = argparse.ArgumentParser(description="args for paddleserving")
+parser.add_argument("--image_dir", type=str, default="../../doc/imgs/")
+args = parser.parse_args()
+test_img_dir = args.image_dir
+
 for img_file in os.listdir(test_img_dir):
     with open(os.path.join(test_img_dir, img_file), 'rb') as file:
         image_data = file.read()
     image = cv2_to_base64(image_data)
 
-for i in range(1):
-    ret = client.predict(feed_dict={"image": image}, fetch=["res"])
-    print(ret)
+    for i in range(1):
+        ret = client.predict(feed_dict={"image": image}, fetch=["res"])
+        print(ret)

deploy/slim/prune/export_prune_model.py

@@ -30,7 +30,7 @@ from ppocr.modeling.architectures import build_model
 
 from ppocr.postprocess import build_post_process
 from ppocr.metrics import build_metric
-from ppocr.utils.save_load import init_model
+from ppocr.utils.save_load import load_model
 import tools.program as program
 
 
@@ -89,7 +89,7 @@ def main(config, device, logger, vdl_writer):
     logger.info(f"FLOPs after pruning: {flops}")
 
     # load pretrain model
-    pre_best_model_dict = init_model(config, model, logger, None)
+    load_model(config, model)
     metric = program.eval(model, valid_dataloader, post_process_class,
                           eval_class)
     logger.info(f"metric['hmean']: {metric['hmean']}")

deploy/slim/prune/sensitivity_anal.py

@@ -32,7 +32,7 @@ from ppocr.losses import build_loss
 from ppocr.optimizer import build_optimizer
 from ppocr.postprocess import build_post_process
 from ppocr.metrics import build_metric
-from ppocr.utils.save_load import init_model
+from ppocr.utils.save_load import load_model
 import tools.program as program
 
 dist.get_world_size()
@@ -94,7 +94,7 @@ def main(config, device, logger, vdl_writer):
     # build metric
     eval_class = build_metric(config['Metric'])
     # load pretrain model
-    pre_best_model_dict = init_model(config, model, logger, optimizer)
+    pre_best_model_dict = load_model(config, model, optimizer)
 
     logger.info('train dataloader has {} iters, valid dataloader has {} iters'.
                 format(len(train_dataloader), len(valid_dataloader)))

deploy/slim/quantization/export_model.py

@@ -28,7 +28,7 @@ from paddle.jit import to_static
 
 from ppocr.modeling.architectures import build_model
 from ppocr.postprocess import build_post_process
-from ppocr.utils.save_load import init_model
+from ppocr.utils.save_load import load_model
 from ppocr.utils.logging import get_logger
 from tools.program import load_config, merge_config, ArgsParser
 from ppocr.metrics import build_metric
@@ -101,7 +101,7 @@ def main():
     quanter = QAT(config=quant_config)
     quanter.quantize(model)
 
-    init_model(config, model)
+    load_model(config, model)
     model.eval()
 
     # build metric

deploy/slim/quantization/quant.py

@@ -37,7 +37,7 @@ from ppocr.losses import build_loss
 from ppocr.optimizer import build_optimizer
 from ppocr.postprocess import build_post_process
 from ppocr.metrics import build_metric
-from ppocr.utils.save_load import init_model
+from ppocr.utils.save_load import load_model
 import tools.program as program
 from paddleslim.dygraph.quant import QAT
 
@@ -137,7 +137,7 @@ def main(config, device, logger, vdl_writer):
     # build metric
     eval_class = build_metric(config['Metric'])
     # load pretrain model
-    pre_best_model_dict = init_model(config, model, logger, optimizer)
+    pre_best_model_dict = load_model(config, model, optimizer)
 
     logger.info('train dataloader has {} iters, valid dataloader has {} iters'.
                 format(len(train_dataloader), len(valid_dataloader)))

deploy/slim/quantization/quant_kl.py

@@ -37,7 +37,7 @@ from ppocr.losses import build_loss
 from ppocr.optimizer import build_optimizer
 from ppocr.postprocess import build_post_process
 from ppocr.metrics import build_metric
-from ppocr.utils.save_load import init_model
+from ppocr.utils.save_load import load_model
 import tools.program as program
 import paddleslim
 from paddleslim.dygraph.quant import QAT

doc/doc_ch/algorithm_overview.md

-<a name="算法介绍"></a>
-## 算法介绍
+# 两阶段算法
+
+- [两阶段算法](#-----)
+  * [1. 算法介绍](#1)
+    + [1.1 文本检测算法](#11)
+    + [1.2 文本识别算法](#12)
+  * [2. 模型训练](#2)
+  * [3. 模型推理](#3)
+
+<a name="1"></a>
+
+## 1. 算法介绍
 本文给出了PaddleOCR已支持的文本检测算法和文本识别算法列表，以及每个算法在**英文公开数据集**上的模型和指标，主要用于算法简介和算法性能对比，更多包括中文在内的其他数据集上的模型请参考[PP-OCR v2.0 系列模型下载](./models_list.md)。
 
-- [1.文本检测算法](#文本检测算法)
-- [2.文本识别算法](#文本识别算法)
+<a name="11"></a>
 
-<a name="文本检测算法"></a>
-### 1.文本检测算法
+### 1.1 文本检测算法
 
 PaddleOCR开源的文本检测算法列表：
-- [x]  DB([paper]( https://arxiv.org/abs/1911.08947))（ppocr推荐）
-- [x]  EAST([paper](https://arxiv.org/abs/1704.03155))
-- [x]  SAST([paper](https://arxiv.org/abs/1908.05498))
+- [x]  DB([paper]( https://arxiv.org/abs/1911.08947)) [2]（ppocr推荐）
+- [x]  EAST([paper](https://arxiv.org/abs/1704.03155))[1]
+- [x]  SAST([paper](https://arxiv.org/abs/1908.05498))[4]
 - [x]  PSENet([paper](https://arxiv.org/abs/1903.12473v2)）
 
 在ICDAR2015文本检测公开数据集上，算法效果如下：
-
 |模型|骨干网络|precision|recall|Hmean|下载链接|
 | --- | --- | --- | --- | --- | --- |
-|EAST|ResNet50_vd|85.80%|86.71%|86.25%|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_east_v2.0_train.tar)|
-|EAST|MobileNetV3|79.42%|80.64%|80.03%|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_mv3_east_v2.0_train.tar)|
-|DB|ResNet50_vd|86.41%|78.72%|82.38%|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_db_v2.0_train.tar)|
-|DB|MobileNetV3|77.29%|73.08%|75.12%|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_mv3_db_v2.0_train.tar)|
-|SAST|ResNet50_vd|91.39%|83.77%|87.42%|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_icdar15_v2.0_train.tar)|
-|PSE|ResNet50_vd|85.81%|79.53%|82.55%|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.1/en_det/det_r50_vd_pse_v2.0_train.tar)|
-|PSE|MobileNetV3|82.20%|70.48%|75.89%|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.1/en_det/det_mv3_pse_v2.0_train.tar)|
+|EAST|ResNet50_vd|85.80%|86.71%|86.25%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_east_v2.0_train.tar)|
+|EAST|MobileNetV3|79.42%|80.64%|80.03%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_mv3_east_v2.0_train.tar)|
+|DB|ResNet50_vd|86.41%|78.72%|82.38%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_db_v2.0_train.tar)|
+|DB|MobileNetV3|77.29%|73.08%|75.12%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_mv3_db_v2.0_train.tar)|
+|SAST|ResNet50_vd|91.39%|83.77%|87.42%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_icdar15_v2.0_train.tar)|
+|PSE|ResNet50_vd|85.81%|79.53%|82.55%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/en_det/det_r50_vd_pse_v2.0_train.tar)|
+|PSE|MobileNetV3|82.20%|70.48%|75.89%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/en_det/det_mv3_pse_v2.0_train.tar)|
 
 在Total-text文本检测公开数据集上，算法效果如下：
 
 |模型|骨干网络|precision|recall|Hmean|下载链接|
 | --- | --- | --- | --- | --- | --- |
-|SAST|ResNet50_vd|89.63%|78.44%|83.66%|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_totaltext_v2.0_train.tar)|
+|SAST|ResNet50_vd|89.63%|78.44%|83.66%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_totaltext_v2.0_train.tar)|
 
 **说明：** SAST模型训练额外加入了icdar2013、icdar2017、COCO-Text、ArT等公开数据集进行调优。PaddleOCR用到的经过整理格式的英文公开数据集下载：
 * [百度云地址](https://pan.baidu.com/s/12cPnZcVuV1zn5DOd4mqjVw) (提取码: 2bpi)
 * [Google Drive下载地址](https://drive.google.com/drive/folders/1ll2-XEVyCQLpJjawLDiRlvo_i4BqHCJe?usp=sharing)
 
-PaddleOCR文本检测算法的训练和使用请参考文档教程中[模型训练/评估中的文本检测部分](./detection.md)。
-
+<a name="12"></a>
 
-<a name="文本识别算法"></a>
-### 2.文本识别算法
+### 1.2 文本识别算法
 
 PaddleOCR基于动态图开源的文本识别算法列表：
-- [x]  CRNN([paper](https://arxiv.org/abs/1507.05717))（ppocr推荐）
-- [x]  Rosetta([paper](https://arxiv.org/abs/1910.05085))
-- [x]  STAR-Net([paper](http://www.bmva.org/bmvc/2016/papers/paper043/index.html))
-- [x]  RARE([paper](https://arxiv.org/abs/1603.03915v1))
-- [x]  SRN([paper](https://arxiv.org/abs/2003.12294))
-- [x]  NRTR([paper](https://arxiv.org/abs/1806.00926v2))
+- [x]  CRNN([paper](https://arxiv.org/abs/1507.05717))[7]（ppocr推荐）
+- [x]  Rosetta([paper](https://arxiv.org/abs/1910.05085))[10]
+- [x]  STAR-Net([paper](http://www.bmva.org/bmvc/2016/papers/paper043/index.html))[11]
+- [x]  RARE([paper](https://arxiv.org/abs/1603.03915v1))[12]
+- [x]  SRN([paper](https://arxiv.org/abs/2003.12294))[5]
+- [x]  NRTR([paper](https://arxiv.org/abs/1806.00926v2))[13]
 - [x]  SAR([paper](https://arxiv.org/abs/1811.00751v2))
-- [x]  SEED([paper](https://arxiv.org/pdf/2005.10977.pdf))
+- [x] SEED([paper](https://arxiv.org/pdf/2005.10977.pdf))
 
-参考[DTRB](https://arxiv.org/abs/1904.01906) 文字识别训练和评估流程，使用MJSynth和SynthText两个文字识别数据集训练，在IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE数据集上进行评估，算法效果如下：
+参考[DTRB][3](https://arxiv.org/abs/1904.01906)文字识别训练和评估流程，使用MJSynth和SynthText两个文字识别数据集训练，在IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE数据集上进行评估，算法效果如下：
 
 |模型|骨干网络|Avg Accuracy|模型存储命名|下载链接|
 |---|---|---|---|---|
-|Rosetta|Resnet34_vd|80.9%|rec_r34_vd_none_none_ctc|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_none_none_ctc_v2.0_train.tar)|
-|Rosetta|MobileNetV3|78.05%|rec_mv3_none_none_ctc|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_none_none_ctc_v2.0_train.tar)|
-|CRNN|Resnet34_vd|82.76%|rec_r34_vd_none_bilstm_ctc|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_none_bilstm_ctc_v2.0_train.tar)|
-|CRNN|MobileNetV3|79.97%|rec_mv3_none_bilstm_ctc|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_none_bilstm_ctc_v2.0_train.tar)|
-|StarNet|Resnet34_vd|84.44%|rec_r34_vd_tps_bilstm_ctc|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_tps_bilstm_ctc_v2.0_train.tar)|
-|StarNet|MobileNetV3|81.42%|rec_mv3_tps_bilstm_ctc|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_tps_bilstm_ctc_v2.0_train.tar)|
-|RARE|MobileNetV3|82.5%|rec_mv3_tps_bilstm_att |[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_tps_bilstm_att_v2.0_train.tar)|
-|RARE|Resnet34_vd|83.6%|rec_r34_vd_tps_bilstm_att |[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_tps_bilstm_att_v2.0_train.tar)|
-|SRN|Resnet50_vd_fpn| 88.52% | rec_r50fpn_vd_none_srn | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r50_vd_srn_train.tar) |
-|NRTR|NRTR_MTB| 84.3% | rec_mtb_nrtr | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar) |
-|SAR|Resnet31| 87.2% | rec_r31_sar | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_r31_sar_train.tar) |
-|SEED| Aster_Resnet | 85.2% | rec_resnet_stn_bilstm_att | [下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_resnet_stn_bilstm_att.tar)|
-PaddleOCR文本识别算法的训练和使用请参考文档教程中[模型训练/评估中的文本识别部分](./recognition.md)。
+|Rosetta|Resnet34_vd|80.9%|rec_r34_vd_none_none_ctc|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_none_none_ctc_v2.0_train.tar)|
+|Rosetta|MobileNetV3|78.05%|rec_mv3_none_none_ctc|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_none_none_ctc_v2.0_train.tar)|
+|CRNN|Resnet34_vd|82.76%|rec_r34_vd_none_bilstm_ctc|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_none_bilstm_ctc_v2.0_train.tar)|
+|CRNN|MobileNetV3|79.97%|rec_mv3_none_bilstm_ctc|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_none_bilstm_ctc_v2.0_train.tar)|
+|StarNet|Resnet34_vd|84.44%|rec_r34_vd_tps_bilstm_ctc|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_tps_bilstm_ctc_v2.0_train.tar)|
+|StarNet|MobileNetV3|81.42%|rec_mv3_tps_bilstm_ctc|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_tps_bilstm_ctc_v2.0_train.tar)|
+|RARE|MobileNetV3|82.5%|rec_mv3_tps_bilstm_att |[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mv3_tps_bilstm_att_v2.0_train.tar)|
+|RARE|Resnet34_vd|83.6%|rec_r34_vd_tps_bilstm_att |[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_tps_bilstm_att_v2.0_train.tar)|
+|SRN|Resnet50_vd_fpn| 88.52% | rec_r50fpn_vd_none_srn | [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r50_vd_srn_train.tar) |
+|NRTR|NRTR_MTB| 84.3% | rec_mtb_nrtr | [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar) |
+|SAR|Resnet31| 87.2% | rec_r31_sar | [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_r31_sar_train.tar) |
+|SEED|Aster_Resnet| 85.2% | rec_resnet_stn_bilstm_att | [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/rec/rec_resnet_stn_bilstm_att.tar) |
+
+<a name="2"></a>
+
+## 2. 模型训练
+
+PaddleOCR文本检测算法的训练和使用请参考文档教程中[模型训练/评估中的文本检测部分](./detection.md)。文本识别算法的训练和使用请参考文档教程中[模型训练/评估中的文本识别部分](./recognition.md)。
+
+<a name="3"></a>
+
+## 3. 模型推理
+
+上述模型中除PP-OCR系列模型以外，其余模型仅支持基于Python引擎的推理，具体内容可参考[基于Python预测引擎推理](./inference.md)

doc/doc_ch/angle_class.md

@@ -11,7 +11,7 @@
 ## 1. 方法介绍
 文本方向分类器主要用于图片非0度的场景下，在这种场景下需要对图片里检测到的文本行进行一个转正的操作。在PaddleOCR系统内，
 文字检测之后得到的文本行图片经过仿射变换之后送入识别模型，此时只需要对文字进行一个0和180度的角度分类，因此PaddleOCR内置的
-文字角度分类器**只支持了0和180度的分类**。如果想支持更多角度，可以自己修改算法进行支持。
+文本方向分类器**只支持了0和180度的分类**。如果想支持更多角度，可以自己修改算法进行支持。
 
 0和180度数据样本例子：
 
@@ -72,8 +72,6 @@ train/cls/train/word_002.jpg   180
 <a name="启动训练"></a>
 ## 3. 启动训练
 
-### 启动训练
-
 将准备好的txt文件和图片文件夹路径分别写入配置文件的 `Train/Eval.dataset.label_file_list` 和 `Train/Eval.dataset.data_dir` 字段下，`Train/Eval.dataset.data_dir`字段下的路径和文件里记载的图片名构成了图片的绝对路径。
 
 PaddleOCR提供了训练脚本、评估脚本和预测脚本。

doc/doc_ch/config.md

@@ -36,11 +36,10 @@
 |      pretrained_model    |    设置加载预训练模型路径      |  ./pretrain_models/CRNN/best_accuracy  |  \          |
 |      checkpoints         |    加载模型参数路径            |       None        |    用于中断后加载参数继续训练 |
 |      use_visualdl  |    设置是否启用visualdl进行可视化log展示 |          False        |    [教程地址](https://www.paddlepaddle.org.cn/paddle/visualdl) |
-|      infer_img            |    设置预测图像路径或文件夹路径     |       ./infer_img | \|
-|      character_dict_path |    设置字典路径            |  ./ppocr/utils/ppocr_keys_v1.txt  |    \                 |
+|      infer_img            |    设置预测图像路径或文件夹路径     |       ./infer_img | \||
+|      character_dict_path |    设置字典路径            |  ./ppocr/utils/ppocr_keys_v1.txt  | 如果为空，则默认使用小写字母+数字作为字典 |
 |      max_text_length     |    设置文本最大长度        |       25          |                \                 |
-|      character_type      |    设置字符类型            |       ch          |    en/ch, en时将使用默认dict，ch时使用自定义dict|
-|      use_space_char     |    设置是否识别空格             |        True      |          仅在 character_type=ch 时支持空格                 |
+|      use_space_char     |    设置是否识别空格             |        True      |          \|               |
 |      label_list          |    设置方向分类器支持的角度       |    ['0','180']    |     仅在方向分类器中生效 |
 |      save_res_path          |    设置检测模型的结果保存地址       |    ./output/det_db/predicts_db.txt    |     仅在检测模型中生效 |
 
@@ -191,7 +190,6 @@ PaddleOCR目前已支持80种（除中文外）语种识别，`configs/rec/multi
       use_gpu: True
       epoch_num: 500
       ...
-      character_type: it  # 需要识别的语种
       character_dict_path:  {path/of/dict} # 字典文件所在路径
    
    Train:
@@ -212,17 +210,17 @@ PaddleOCR目前已支持80种（除中文外）语种识别，`configs/rec/multi
 
 目前PaddleOCR支持的多语言算法有：
 
-| 配置文件 |  算法名称 |   backbone |   trans   |   seq      |     pred     |  language | character_type |
-| :--------: |  :-------:   | :-------:  |   :-------:   |   :-----:   |  :-----:   | :-----:  | :-----:  |
-| rec_chinese_cht_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 中文繁体  | chinese_cht|
-| rec_en_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 英语（区分大小写）   | EN |
-| rec_french_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 法语 |  french |
-| rec_ger_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 德语   | german |
-| rec_japan_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 日语  | japan |
-| rec_korean_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 韩语  | korean |
-| rec_latin_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 拉丁字母  | latin |
-| rec_arabic_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 阿拉伯字母 |  ar |
-| rec_cyrillic_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 斯拉夫字母  | cyrillic |
-| rec_devanagari_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 梵文字母  | devanagari |
+| 配置文件 |  算法名称 |   backbone |   trans   |   seq      |     pred     |  language |
+| :--------: |  :-------:   | :-------:  |   :-------:   |   :-----:   |  :-----:   | :-----:  |
+| rec_chinese_cht_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 中文繁体  |
+| rec_en_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 英语（区分大小写）   |
+| rec_french_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 法语 |
+| rec_ger_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 德语   |
+| rec_japan_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 日语  |
+| rec_korean_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 韩语  |
+| rec_latin_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 拉丁字母  |
+| rec_arabic_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 阿拉伯字母 |
+| rec_cyrillic_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 斯拉夫字母  |
+| rec_devanagari_lite_train.yml |  CRNN |   Mobilenet_v3 small 0.5 |  None   |  BiLSTM |  ctc  | 梵文字母  |
 
 更多支持语种请参考: [多语言模型](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.1/doc/doc_ch/multi_languages.md#%E8%AF%AD%E7%A7%8D%E7%BC%A9%E5%86%99)

doc/doc_ch/detection.md

 
-# 目录
-- [1. 文字检测](#1-----)
-  * [1.1 数据准备](#11-----)
-  * [1.2 下载预训练模型](#12--------)
-  * [1.3 启动训练](#13-----)
-  * [1.4 断点训练](#14-----)
-  * [1.5 更换Backbone 训练](#15---backbone---)
-  * [1.6 指标评估](#16-----)
-  * [1.7 测试检测效果](#17-------)
-  * [1.8 转inference模型测试](#18--inference----)
-- [2. FAQ](#2-faq)
-
-
-<a name="1-----"></a>
-# 1. 文字检测
+# 文字检测
 
 本节以icdar2015数据集为例，介绍PaddleOCR中检测模型训练、评估、测试的使用方式。
 
+- [1. 准备数据和模型](#1--------)
+  * [1.1 数据准备](#11-----)
+  * [1.2 下载预训练模型](#12--------)
+- [2. 开始训练](#2-----)
+  * [2.1 启动训练](#21-----)
+  * [2.2 断点训练](#22-----)
+  * [2.3 更换Backbone 训练](#23---backbone---)
+- [3. 模型评估与预测](#3--------)
+  * [3.1 指标评估](#31-----)
+  * [3.2 测试检测效果](#32-------)
+- [4. 模型导出与预测](#4--------)
+- [5. FAQ](#5-faq)
+
+<a name="1--------"></a>
+# 1. 准备数据和模型
+
 <a name="11-----"></a>
 ## 1.1 数据准备
 
@@ -83,8 +85,11 @@ wget -P ./pretrain_models/ https://paddle-imagenet-models-name.bj.bcebos.com/dyg
 wget -P ./pretrain_models/ https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/ResNet50_vd_ssld_pretrained.pdparams
 ```
 
-<a name="13-----"></a>
-## 1.3 启动训练
+<a name="2-----"></a>
+# 2. 开始训练
+
+<a name="21-----"></a>
+## 2.1 启动训练
 
 *如果您安装的是cpu版本，请将配置文件中的 `use_gpu` 字段修改为false*
 
@@ -96,6 +101,10 @@ python3 tools/train.py -c configs/det/det_mv3_db.yml \
 # 单机多卡训练，通过 --gpus 参数设置使用的GPU ID
 python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/det/det_mv3_db.yml \
      -o Global.pretrained_model=./pretrain_models/MobileNetV3_large_x0_5_pretrained
+
+# 多机多卡训练，通过 --ips 参数设置使用的机器IP地址，通过 --gpus 参数设置使用的GPU ID
+python3 -m paddle.distributed.launch --ips="xx.xx.xx.xx,xx.xx.xx.xx" --gpus '0,1,2,3' tools/train.py -c configs/det/det_mv3_db.yml \
+     -o Global.pretrained_model=./pretrain_models/MobileNetV3_large_x0_5_pretrained
 ```
 
 上述指令中，通过-c 选择训练使用configs/det/det_db_mv3.yml配置文件。
@@ -106,8 +115,17 @@ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/
 python3 tools/train.py -c configs/det/det_mv3_db.yml -o Optimizer.base_lr=0.0001
 ```
 
-<a name="14-----"></a>
-## 1.4 断点训练
+**注意:** 采用多机多卡训练时，需要替换上面命令中的ips值为您机器的地址，机器之间需要能够相互ping通。另外，训练时需要在多个机器上分别启动命令。查看机器ip地址的命令为`ifconfig`。
+
+如果您想进一步加快训练速度，可以使用[自动混合精度训练](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/01_paddle2.0_introduction/basic_concept/amp_cn.html)， 以单机单卡为例，命令如下：
+```shell
+python3 tools/train.py -c configs/det/det_mv3_db.yml \
+     -o Global.pretrained_model=./pretrain_models/MobileNetV3_large_x0_5_pretrained \
+     Global.use_amp=True Global.scale_loss=1024.0 Global.use_dynamic_loss_scaling=True
+ ```
+
+<a name="22-----"></a>
+## 2.2 断点训练
 
 如果训练程序中断，如果希望加载训练中断的模型从而恢复训练，可以通过指定Global.checkpoints指定要加载的模型路径：
 ```shell
@@ -116,8 +134,8 @@ python3 tools/train.py -c configs/det/det_mv3_db.yml -o Global.checkpoints=./you
 
 **注意**：`Global.checkpoints`的优先级高于`Global.pretrained_model`的优先级，即同时指定两个参数时，优先加载`Global.checkpoints`指定的模型，如果`Global.checkpoints`指定的模型路径有误，会加载`Global.pretrained_model`指定的模型。
 
-<a name="15---backbone---"></a>
-## 1.5 更换Backbone 训练
+<a name="23---backbone---"></a>
+## 2.3 更换Backbone 训练
 
 PaddleOCR将网络划分为四部分，分别在[ppocr/modeling](../../ppocr/modeling)下。 进入网络的数据将按照顺序(transforms->backbones->
 necks->heads)依次通过这四个部分。
@@ -164,8 +182,11 @@ args1: args1
 
 **注意**：如果要更换网络的其他模块，可以参考[文档](./add_new_algorithm.md)。
 
-<a name="16-----"></a>
-## 1.6 指标评估
+<a name="3--------"></a>
+# 3. 模型评估与预测
+
+<a name="31-----"></a>
+## 3.1 指标评估
 
 PaddleOCR计算三个OCR检测相关的指标，分别是：Precision、Recall、Hmean（F-Score）。
 
@@ -177,8 +198,8 @@ python3 tools/eval.py -c configs/det/det_mv3_db.yml  -o Global.checkpoints="{pat
 
 * 注：`box_thresh`、`unclip_ratio`是DB后处理所需要的参数，在评估EAST模型时不需要设置
 
-<a name="17-------"></a>
-## 1.7 测试检测效果
+<a name="32-------"></a>
+## 3.2 测试检测效果
 
 测试单张图像的检测效果
 ```shell
@@ -195,8 +216,8 @@ python3 tools/infer_det.py -c configs/det/det_mv3_db.yml -o Global.infer_img="./
 python3 tools/infer_det.py -c configs/det/det_mv3_db.yml -o Global.infer_img="./doc/imgs_en/" Global.pretrained_model="./output/det_db/best_accuracy"
 ```
 
-<a name="18--inference----"></a>
-## 1.8 转inference模型测试
+<a name="4--------"></a>
+# 4. 模型导出与预测
 
 inference 模型（`paddle.jit.save`保存的模型）
 一般是模型训练，把模型结构和模型参数保存在文件中的固化模型，多用于预测部署场景。
@@ -218,10 +239,11 @@ python3 tools/infer/predict_det.py --det_algorithm="DB" --det_model_dir="./outpu
 python3 tools/infer/predict_det.py --det_algorithm="EAST" --det_model_dir="./output/det_db_inference/" --image_dir="./doc/imgs/" --use_gpu=True
 ```
 
-<a name="2"></a>
-# 2. FAQ
+<a name="5-faq"></a>
+# 5. FAQ
 
 Q1: 训练模型转inference 模型之后预测效果不一致？
+
 **A**：此类问题出现较多，问题多是trained model预测时候的预处理、后处理参数和inference model预测的时候的预处理、后处理参数不一致导致的。以det_mv3_db.yml配置文件训练的模型为例，训练模型、inference模型预测结果不一致问题解决方式如下：
 - 检查[trained model预处理](https://github.com/PaddlePaddle/PaddleOCR/blob/c1ed243fb68d5d466258243092e56cbae32e2c14/configs/det/det_mv3_db.yml#L116)，和[inference model的预测预处理](https://github.com/PaddlePaddle/PaddleOCR/blob/c1ed243fb68d5d466258243092e56cbae32e2c14/tools/infer/predict_det.py#L42)函数是否一致。算法在评估的时候，输入图像大小会影响精度，为了和论文保持一致，训练icdar15配置文件中将图像resize到[736, 1280]，但是在inference model预测的时候只有一套默认参数，会考虑到预测速度问题，默认限制图像最长边为960做resize的。训练模型预处理和inference模型的预处理函数位于[ppocr/data/imaug/operators.py](https://github.com/PaddlePaddle/PaddleOCR/blob/c1ed243fb68d5d466258243092e56cbae32e2c14/ppocr/data/imaug/operators.py#L147)
 - 检查[trained model后处理](https://github.com/PaddlePaddle/PaddleOCR/blob/c1ed243fb68d5d466258243092e56cbae32e2c14/configs/det/det_mv3_db.yml#L51)，和[inference 后处理参数](https://github.com/PaddlePaddle/PaddleOCR/blob/c1ed243fb68d5d466258243092e56cbae32e2c14/tools/infer/utility.py#L50)是否一致。