Merge branch 'dygraph' of https://github.com/PaddlePaddle/PaddleOCR into dygraph

PPOCRLabel/README.md

@@ -54,7 +54,10 @@ PPOCRLabel can be started in two ways: whl package and Python script. The whl pa
 
 ```bash
 pip install PPOCRLabel  # install
-PPOCRLabel  # run
+
+# Select label mode and run 
+PPOCRLabel  # [Normal mode] for [detection + recognition] labeling
+PPOCRLabel --kie True # [KIE mode] for [detection + recognition + keyword extraction] labeling
 ```
 
 > If you getting this error `OSError: [WinError 126] The specified module could not be found` when you install shapely on windows. Please try to download Shapely whl file using http://www.lfd.uci.edu/~gohlke/pythonlibs/#shapely.
@@ -67,13 +70,18 @@ PPOCRLabel  # run
 ```bash
 pip3 install PPOCRLabel
 pip3 install trash-cli
-PPOCRLabel
+
+# Select label mode and run 
+PPOCRLabel  # [Normal mode] for [detection + recognition] labeling
+PPOCRLabel --kie True # [KIE mode] for [detection + recognition + keyword extraction] labeling
 ```
 
 #### MacOS
 ```bash
 pip3 install PPOCRLabel
 pip3 install opencv-contrib-python-headless==4.2.0.32
+
+# Select label mode and run 
 PPOCRLabel  # [Normal mode] for [detection + recognition] labeling
 PPOCRLabel --kie True # [KIE mode] for [detection + recognition + keyword extraction] labeling
 ```
@@ -90,6 +98,8 @@ pip3 install dist/PPOCRLabel-1.0.2-py2.py3-none-any.whl
 
 ```bash
 cd ./PPOCRLabel  # Switch to the PPOCRLabel directory
+
+# Select label mode and run 
 python PPOCRLabel.py  # [Normal mode] for [detection + recognition] labeling
 python PPOCRLabel.py --kie True # [KIE mode] for [detection + recognition + keyword extraction] labeling
 ```

PPOCRLabel/README_ch.md

@@ -57,7 +57,10 @@ PPOCRLabel可通过whl包与Python脚本两种方式启动，whl包形式启动
 
 ```bash
 pip install PPOCRLabel  # 安装
-PPOCRLabel --lang ch  # 运行
+
+# 选择标签模式来启动
+PPOCRLabel --lang ch  # 启动【普通模式】，用于打【检测+识别】场景的标签
+PPOCRLabel --lang ch --kie True  # 启动 【KIE 模式】，用于打【检测+识别+关键字提取】场景的标签
 ```
 > 注意：通过whl包安装PPOCRLabel会自动下载 `paddleocr` whl包，其中shapely依赖可能会出现 `[winRrror 126] 找不到指定模块的问题。` 的错误，建议从[这里](https://www.lfd.uci.edu/~gohlke/pythonlibs/#shapely)下载并安装
 ##### Ubuntu Linux
@@ -65,13 +68,18 @@ PPOCRLabel --lang ch  # 运行
 ```bash
 pip3 install PPOCRLabel
 pip3 install trash-cli
-PPOCRLabel --lang ch
+
+# 选择标签模式来启动
+PPOCRLabel --lang ch  # 启动【普通模式】，用于打【检测+识别】场景的标签
+PPOCRLabel --lang ch --kie True  # 启动 【KIE 模式】，用于打【检测+识别+关键字提取】场景的标签
 ```
 
 ##### MacOS
 ```bash
 pip3 install PPOCRLabel
 pip3 install opencv-contrib-python-headless==4.2.0.32 # 如果下载过慢请添加"-i https://mirror.baidu.com/pypi/simple"
+
+# 选择标签模式来启动
 PPOCRLabel --lang ch  # 启动【普通模式】，用于打【检测+识别】场景的标签
 PPOCRLabel --lang ch --kie True  # 启动 【KIE 模式】，用于打【检测+识别+关键字提取】场景的标签
 ```
@@ -92,6 +100,8 @@ pip3 install dist/PPOCRLabel-1.0.2-py2.py3-none-any.whl -i https://mirror.baidu.
 
 ```bash
 cd ./PPOCRLabel  # 切换到PPOCRLabel目录
+
+# 选择标签模式来启动
 python PPOCRLabel.py --lang ch  # 启动【普通模式】，用于打【检测+识别】场景的标签
 python PPOCRLabel.py --lang ch --kie True  # 启动 【KIE 模式】，用于打【检测+识别+关键字提取】场景的标签
 ```

deploy/cpp_infer/include/ocr_det.h

@@ -45,8 +45,9 @@ public:
                       const double &det_db_thresh,
                       const double &det_db_box_thresh,
                       const double &det_db_unclip_ratio,
-                      const bool &use_polygon_score, const bool &visualize,
-                      const bool &use_tensorrt, const std::string &precision) {
+                      const bool &use_polygon_score, const bool &use_dilation,
+                      const bool &visualize, const bool &use_tensorrt,
+                      const std::string &precision) {
     this->use_gpu_ = use_gpu;
     this->gpu_id_ = gpu_id;
     this->gpu_mem_ = gpu_mem;
@@ -59,6 +60,7 @@ public:
     this->det_db_box_thresh_ = det_db_box_thresh;
     this->det_db_unclip_ratio_ = det_db_unclip_ratio;
     this->use_polygon_score_ = use_polygon_score;
+    this->use_dilation_ = use_dilation;
 
     this->visualize_ = visualize;
     this->use_tensorrt_ = use_tensorrt;
@@ -71,7 +73,8 @@ public:
   void LoadModel(const std::string &model_dir);
 
   // Run predictor
-  void Run(cv::Mat &img, std::vector<std::vector<std::vector<int>>> &boxes, std::vector<double> *times);
+  void Run(cv::Mat &img, std::vector<std::vector<std::vector<int>>> &boxes,
+           std::vector<double> *times);
 
 private:
   std::shared_ptr<Predictor> predictor_;
@@ -88,6 +91,7 @@ private:
   double det_db_box_thresh_ = 0.5;
   double det_db_unclip_ratio_ = 2.0;
   bool use_polygon_score_ = false;
+  bool use_dilation_ = false;
 
   bool visualize_ = true;
   bool use_tensorrt_ = false;

deploy/cpp_infer/readme.md

@@ -4,16 +4,20 @@
 C++在性能计算上优于python，因此，在大多数CPU、GPU部署场景，多采用C++的部署方式，本节将介绍如何在Linux\Windows (CPU\GPU)环境下配置C++环境并完成
 PaddleOCR模型部署。
 
-* [1. 准备环境](#1)
-  + [1.0 运行准备](#10)
-  + [1.1 编译opencv库](#11)
-  + [1.2 下载或者编译Paddle预测库](#12)
-    - [1.2.1 直接下载安装](#121)
-    - [1.2.2 预测库源码编译](#122)
-* [2 开始运行](#2)
-  + [2.1 将模型导出为inference model](#21)
-  + [2.2 编译PaddleOCR C++预测demo](#22)
-  + [2.3运行demo](#23)
+- [服务器端C++预测](#服务器端c预测)
+  - [1. 准备环境](#1-准备环境)
+    - [1.0 运行准备](#10-运行准备)
+    - [1.1 编译opencv库](#11-编译opencv库)
+    - [1.2 下载或者编译Paddle预测库](#12-下载或者编译paddle预测库)
+      - [1.2.1 直接下载安装](#121-直接下载安装)
+      - [1.2.2 预测库源码编译](#122-预测库源码编译)
+  - [2 开始运行](#2-开始运行)
+    - [2.1 将模型导出为inference model](#21-将模型导出为inference-model)
+    - [2.2 编译PaddleOCR C++预测demo](#22-编译paddleocr-c预测demo)
+    - [2.3 运行demo](#23-运行demo)
+        - [1. 只调用检测：](#1-只调用检测)
+        - [2. 只调用识别：](#2-只调用识别)
+        - [3. 调用串联：](#3-调用串联)
 
 <a name="1"></a>
 
@@ -103,7 +107,7 @@ opencv3/
 
 #### 1.2.1 直接下载安装
 
-* [Paddle预测库官网](https://paddle-inference.readthedocs.io/en/latest/user_guides/download_lib.html) 上提供了不同cuda版本的Linux预测库，可以在官网查看并选择合适的预测库版本（*建议选择paddle版本>=2.0.1版本的预测库* ）。
+* [Paddle预测库官网](https://paddleinference.paddlepaddle.org.cn/user_guides/download_lib.html#linux) 上提供了不同cuda版本的Linux预测库，可以在官网查看并选择合适的预测库版本（*建议选择paddle版本>=2.0.1版本的预测库* ）。
 
 * 下载之后使用下面的方法解压。
 
@@ -249,7 +253,7 @@ CUDNN_LIB_DIR=/your_cudnn_lib_dir
 |gpu_id|int|0|GPU id，使用GPU时有效|
 |gpu_mem|int|4000|申请的GPU内存|
 |cpu_math_library_num_threads|int|10|CPU预测时的线程数，在机器核数充足的情况下，该值越大，预测速度越快|
-|use_mkldnn|bool|true|是否使用mkldnn库|
+|enable_mkldnn|bool|true|是否使用mkldnn库|
 
 - 检测模型相关
 

deploy/cpp_infer/readme_en.md

@@ -78,7 +78,7 @@ opencv3/
 
 #### 1.2.1 Direct download and installation
 
-[Paddle inference library official website](https://paddle-inference.readthedocs.io/en/latest/user_guides/download_lib.html). You can review and select the appropriate version of the inference library on the official website.
+[Paddle inference library official website](https://paddleinference.paddlepaddle.org.cn/user_guides/download_lib.html#linux). You can review and select the appropriate version of the inference library on the official website.
 
 
 * After downloading, use the following command to extract files.
@@ -231,7 +231,7 @@ More parameters are as follows,
 |gpu_id|int|0|GPU id when use_gpu is true|
 |gpu_mem|int|4000|GPU memory requested|
 |cpu_math_library_num_threads|int|10|Number of threads when using CPU inference. When machine cores is enough, the large the value, the faster the inference speed|
-|use_mkldnn|bool|true|Whether to use mkdlnn library|
+|enable_mkldnn|bool|true|Whether to use mkdlnn library|
 
 - Detection related parameters
 

deploy/cpp_infer/src/ocr_det.cpp

@@ -14,7 +14,6 @@
 
 #include <include/ocr_det.h>
 
-
 namespace PaddleOCR {
 
 void DBDetector::LoadModel(const std::string &model_dir) {
@@ -30,13 +29,10 @@ void DBDetector::LoadModel(const std::string &model_dir) {
       if (this->precision_ == "fp16") {
         precision = paddle_infer::Config::Precision::kHalf;
       }
-     if (this->precision_ == "int8") {
+      if (this->precision_ == "int8") {
         precision = paddle_infer::Config::Precision::kInt8;
-      } 
-      config.EnableTensorRtEngine(
-          1 << 20, 10, 3,
-          precision,
-          false, false);
+      }
+      config.EnableTensorRtEngine(1 << 20, 10, 3, precision, false, false);
       std::map<std::string, std::vector<int>> min_input_shape = {
           {"x", {1, 3, 50, 50}},
           {"conv2d_92.tmp_0", {1, 96, 20, 20}},
@@ -105,7 +101,7 @@ void DBDetector::Run(cv::Mat &img,
   cv::Mat srcimg;
   cv::Mat resize_img;
   img.copyTo(srcimg);
-  
+
   auto preprocess_start = std::chrono::steady_clock::now();
   this->resize_op_.Run(img, resize_img, this->max_side_len_, ratio_h, ratio_w,
                        this->use_tensorrt_);
@@ -116,16 +112,16 @@ void DBDetector::Run(cv::Mat &img,
   std::vector<float> input(1 * 3 * resize_img.rows * resize_img.cols, 0.0f);
   this->permute_op_.Run(&resize_img, input.data());
   auto preprocess_end = std::chrono::steady_clock::now();
-    
+
   // Inference.
   auto input_names = this->predictor_->GetInputNames();
   auto input_t = this->predictor_->GetInputHandle(input_names[0]);
   input_t->Reshape({1, 3, resize_img.rows, resize_img.cols});
   auto inference_start = std::chrono::steady_clock::now();
   input_t->CopyFromCpu(input.data());
-  
+
   this->predictor_->Run();
-    
+
   std::vector<float> out_data;
   auto output_names = this->predictor_->GetOutputNames();
   auto output_t = this->predictor_->GetOutputHandle(output_names[0]);
@@ -136,7 +132,7 @@ void DBDetector::Run(cv::Mat &img,
   out_data.resize(out_num);
   output_t->CopyToCpu(out_data.data());
   auto inference_end = std::chrono::steady_clock::now();
-  
+
   auto postprocess_start = std::chrono::steady_clock::now();
   int n2 = output_shape[2];
   int n3 = output_shape[3];
@@ -157,24 +153,29 @@ void DBDetector::Run(cv::Mat &img,
   const double maxvalue = 255;
   cv::Mat bit_map;
   cv::threshold(cbuf_map, bit_map, threshold, maxvalue, cv::THRESH_BINARY);
-  cv::Mat dilation_map;
-  cv::Mat dila_ele = cv::getStructuringElement(cv::MORPH_RECT, cv::Size(2, 2));
-  cv::dilate(bit_map, dilation_map, dila_ele);
+  if (this->use_dilation_) {
+    cv::Mat dila_ele =
+        cv::getStructuringElement(cv::MORPH_RECT, cv::Size(2, 2));
+    cv::dilate(bit_map, bit_map, dila_ele);
+  }
+
   boxes = post_processor_.BoxesFromBitmap(
-      pred_map, dilation_map, this->det_db_box_thresh_,
-      this->det_db_unclip_ratio_, this->use_polygon_score_);
+      pred_map, bit_map, this->det_db_box_thresh_, this->det_db_unclip_ratio_,
+      this->use_polygon_score_);
 
   boxes = post_processor_.FilterTagDetRes(boxes, ratio_h, ratio_w, srcimg);
   auto postprocess_end = std::chrono::steady_clock::now();
   std::cout << "Detected boxes num: " << boxes.size() << endl;
 
-  std::chrono::duration<float> preprocess_diff = preprocess_end - preprocess_start;
+  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;
+  std::chrono::duration<float> postprocess_diff =
+      postprocess_end - postprocess_start;
   times->push_back(double(postprocess_diff.count() * 1000));
-    
+
   //// visualization
   if (this->visualize_) {
     Utility::VisualizeBboxes(srcimg, boxes);

doc/doc_ch/detection.md

@@ -10,6 +10,7 @@
   * [2.1 启动训练](#21-----)
   * [2.2 断点训练](#22-----)
   * [2.3 更换Backbone 训练](#23---backbone---)
+  * [2.4 知识蒸馏训练](#24---distill---)
 - [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-------)
@@ -182,6 +183,15 @@ 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="24---distill---"></a>
+
+## 2.4 知识蒸馏训练
+
+PaddleOCR支持了基于知识蒸馏的检测模型训练过程，更多内容可以参考[知识蒸馏说明文档](./knowledge_distillation.md)。
+
+
+
 <a name="3--------"></a>
 # 3. 模型评估与预测
 

doc/doc_ch/recognition.md

@@ -11,6 +11,7 @@
     - [2.1 数据增强](#数据增强)
     - [2.2 通用模型训练](#通用模型训练)
     - [2.3 多语言模型训练](#多语言模型训练)
+    - [2.4 知识蒸馏训练](#知识蒸馏训练)
 - [3 评估](#评估)
 - [4 预测](#预测)
 - [5 转Inference模型测试](#Inference)
@@ -368,6 +369,13 @@ Eval:
     label_file_list: ["./train_data/french_val.txt"]
     ...
 ```
+
+<a name="知识蒸馏训练"></a>
+
+### 2.4 知识蒸馏训练
+
+PaddleOCR支持了基于知识蒸馏的文本识别模型训练过程，更多内容可以参考[知识蒸馏说明文档](./knowledge_distillation.md)。
+
 <a name="评估"></a>
 ## 3 评估
 

doc/doc_en/detection_en.md

@@ -9,6 +9,7 @@ This section uses the icdar2015 dataset as an example to introduce the training,
   * [2.1 Start Training](#21-start-training)
   * [2.2 Load Trained Model and Continue Training](#22-load-trained-model-and-continue-training)
   * [2.3 Training with New Backbone](#23-training-with-new-backbone)
+  * [2.4 Training with knowledge distillation](#24)
 - [3. Evaluation and Test](#3-evaluation-and-test)
   * [3.1 Evaluation](#31-evaluation)
   * [3.2 Test](#32-test)
@@ -174,6 +175,11 @@ After adding the four-part modules of the network, you only need to configure th
 
 **NOTE**: More details about replace Backbone and other mudule can be found in [doc](add_new_algorithm_en.md).
 
+
+### 2.4 Training with knowledge distillation
+
+Knowledge distillation is supported in PaddleOCR for text detection training process. For more details, please refer to [doc](./knowledge_distillation_en.md).
+
 ## 3. Evaluation and Test
 
 ### 3.1 Evaluation

doc/doc_en/recognition_en.md

@@ -10,6 +10,7 @@
     - [2.1 Data Augmentation](#Data_Augmentation)
     - [2.2 General Training](#Training)
     - [2.3 Multi-language Training](#Multi_language)
+    - [2.4 Training with Knowledge Distillation](#kd)
 
 - [3. Evaluation](#EVALUATION)
 
@@ -361,6 +362,12 @@ Eval:
     ...
 ```
 
+<a name="kd"></a>
+
+### 2.4 Training with Knowledge Distillation
+
+Knowledge distillation is supported in PaddleOCR for text recognition training process. For more details, please refer to [doc](./knowledge_distillation_en.md).
+
 <a name="EVALUATION"></a>
 
 ## 3. Evalution

ppocr/losses/basic_loss.py

@@ -95,9 +95,15 @@ class DMLLoss(nn.Layer):
             self.act = None
 
         self.use_log = use_log
-
         self.jskl_loss = KLJSLoss(mode="js")
 
+    def _kldiv(self, x, target):
+        eps = 1.0e-10
+        loss = target * (paddle.log(target + eps) - x)
+        # batch mean loss
+        loss = paddle.sum(loss) / loss.shape[0]
+        return loss
+
     def forward(self, out1, out2):
         if self.act is not None:
             out1 = self.act(out1)
@@ -106,9 +112,8 @@ class DMLLoss(nn.Layer):
             # for recognition distillation, log is needed for feature map
             log_out1 = paddle.log(out1)
             log_out2 = paddle.log(out2)
-            loss = (F.kl_div(
-                log_out1, out2, reduction='batchmean') + F.kl_div(
-                    log_out2, out1, reduction='batchmean')) / 2.0
+            loss = (
+                self._kldiv(log_out1, out2) + self._kldiv(log_out2, out1)) / 2.0
         else:
             # for detection distillation log is not needed
             loss = self.jskl_loss(out1, out2)

test_tipc/configs/ch_PP-OCRv2_rec/train_infer_python.txt

@@ -49,5 +49,5 @@ inference:tools/infer/predict_rec.py
 null:null
 --benchmark:True
 null:null
-
-
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,32,320]}]

test_tipc/configs/ch_PP-OCRv2_rec_PACT/train_infer_python.txt

@@ -49,5 +49,5 @@ inference:tools/infer/predict_rec.py
 null:null
 --benchmark:True
 null:null
-
-
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,32,320]}]

test_tipc/configs/ch_ppocr_mobile_v2.0_rec/train_infer_python.txt

@@ -49,3 +49,5 @@ inference:tools/infer/predict_rec.py
 --save_log_path:./test/output/
 --benchmark:True
 null:null
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,32,100]}]

test_tipc/configs/ch_ppocr_mobile_v2.0_rec_FPGM/train_infer_python.txt

@@ -48,4 +48,6 @@ inference:tools/infer/predict_rec.py
 --image_dir:./inference/rec_inference
 null:null
 --benchmark:True
-null:null
+null:null
\ No newline at end of file
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,32,320]}]
\ No newline at end of file

test_tipc/configs/ch_ppocr_mobile_v2.0_rec_PACT/train_infer_python.txt

@@ -48,4 +48,6 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ppocr_ke
 --image_dir:./inference/rec_inference
 --save_log_path:./test/output/
 --benchmark:True
-null:null
+null:null
\ No newline at end of file
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,32,320]}]

test_tipc/configs/ch_ppocr_server_v2.0_rec/train_infer_python.txt

@@ -49,3 +49,5 @@ inference:tools/infer/predict_rec.py
 --save_log_path:./test/output/
 --benchmark:True
 null:null
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,32,100]}]

test_tipc/configs/rec_mtb_nrtr/train_infer_python.txt

@@ -49,4 +49,5 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/EN_symbo
 --save_log_path:./test/output/
 --benchmark:True
 null:null
-
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[1,32,100]}]

test_tipc/configs/rec_mv3_none_bilstm_ctc_v2.0/train_infer_python.txt

@@ -49,3 +49,5 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
 --save_log_path:./test/output/
 --benchmark:True
 null:null
+===========================infer_benchmark_params==========================
+random_infer_input:[{float32,[3,32,100]}]
\ No newline at end of file