Merge remote-tracking branch 'upstream/dygraph' into dy1

README_ch.md

@@ -8,7 +8,7 @@ PaddleOCR同时支持动态图与静态图两种编程范式
 - 静态图版本：develop分支
 
 **近期更新**
-- 2021.1.4 [FAQ](./doc/doc_ch/FAQ.md)新增5个高频问题，总数142个，每周一都会更新，欢迎大家持续关注。
+- 2021.1.11 [FAQ](./doc/doc_ch/FAQ.md)新增5个高频问题，总数147个，每周一都会更新，欢迎大家持续关注。
 - 2020.12.15 更新数据合成工具[Style-Text](./StyleText/README_ch.md)，可以批量合成大量与目标场景类似的图像，在多个场景验证，效果明显提升。
 - 2020.11.25 更新半自动标注工具[PPOCRLabel](./PPOCRLabel/README_ch.md)，辅助开发者高效完成标注任务，输出格式与PP-OCR训练任务完美衔接。
 - 2020.9.22 更新PP-OCR技术文章，https://arxiv.org/abs/2009.09941
@@ -102,7 +102,7 @@ PaddleOCR同时支持动态图与静态图两种编程范式
 - FAQ
     - [【精选】OCR精选10个问题](./doc/doc_ch/FAQ.md)
     - [【理论篇】OCR通用31个问题](./doc/doc_ch/FAQ.md)
-    - [【实战篇】PaddleOCR实战101个问题](./doc/doc_ch/FAQ.md)
+    - [【实战篇】PaddleOCR实战106个问题](./doc/doc_ch/FAQ.md)
 - [技术交流群](#欢迎加入PaddleOCR技术交流群)
 - [参考文献](./doc/doc_ch/reference.md)
 - [许可证书](#许可证书)

StyleText/README.md

@@ -52,7 +52,7 @@ If you save the model in another location, please modify the address of the mode
 
 ```
 bg_generator:
-  pretrain: style_text_rec/bg_generator
+  pretrain: style_text_models/bg_generator
 ...
 text_generator:
   pretrain: style_text_models/text_generator

configs/det/det_mv3_db.yml

@@ -102,6 +102,7 @@ Train:
     drop_last: False
     batch_size_per_card: 16
     num_workers: 8
+    use_shared_memory: False
 
 Eval:
   dataset:
@@ -128,4 +129,5 @@ Eval:
     shuffle: False
     drop_last: False
     batch_size_per_card: 1 # must be 1
-    num_workers: 8
\ No newline at end of file
+    num_workers: 8
+    use_shared_memory: False
\ No newline at end of file

configs/rec/rec_icdar15_train.yml

@@ -76,6 +76,7 @@ Train:
     batch_size_per_card: 256
     drop_last: True
     num_workers: 8
+    use_shared_memory: False
 
 Eval:
   dataset:
@@ -96,3 +97,4 @@ Eval:
     drop_last: False
     batch_size_per_card: 256
     num_workers: 4
+    use_shared_memory: False

deploy/cpp_infer/include/preprocess_op.h

@@ -47,18 +47,20 @@ public:
 class ResizeImgType0 {
 public:
   virtual void Run(const cv::Mat &img, cv::Mat &resize_img, int max_size_len,
-                   float &ratio_h, float &ratio_w);
+                   float &ratio_h, float &ratio_w, bool use_tensorrt);
 };
 
 class CrnnResizeImg {
 public:
   virtual void Run(const cv::Mat &img, cv::Mat &resize_img, float wh_ratio,
+                   bool use_tensorrt = false,
                    const std::vector<int> &rec_image_shape = {3, 32, 320});
 };
 
 class ClsResizeImg {
 public:
   virtual void Run(const cv::Mat &img, cv::Mat &resize_img,
+                   bool use_tensorrt = false,
                    const std::vector<int> &rec_image_shape = {3, 48, 192});
 };
 

deploy/cpp_infer/src/main.cpp

@@ -77,7 +77,7 @@ int main(int argc, char **argv) {
   auto end = std::chrono::system_clock::now();
   auto duration =
       std::chrono::duration_cast<std::chrono::microseconds>(end - start);
-  std::cout << "Cost"
+  std::cout << "Cost  "
             << double(duration.count()) *
                    std::chrono::microseconds::period::num /
                    std::chrono::microseconds::period::den

deploy/cpp_infer/src/ocr_cls.cpp

@@ -25,7 +25,7 @@ cv::Mat Classifier::Run(cv::Mat &img) {
   int index = 0;
   float wh_ratio = float(img.cols) / float(img.rows);
 
-  this->resize_op_.Run(img, resize_img, cls_image_shape);
+  this->resize_op_.Run(img, resize_img, this->use_tensorrt_, cls_image_shape);
 
   this->normalize_op_.Run(&resize_img, this->mean_, this->scale_,
                           this->is_scale_);

deploy/cpp_infer/src/ocr_det.cpp

@@ -61,7 +61,8 @@ void DBDetector::Run(cv::Mat &img,
   cv::Mat srcimg;
   cv::Mat resize_img;
   img.copyTo(srcimg);
-  this->resize_op_.Run(img, resize_img, this->max_side_len_, ratio_h, ratio_w);
+  this->resize_op_.Run(img, resize_img, this->max_side_len_, ratio_h, ratio_w,
+                       this->use_tensorrt_);
 
   this->normalize_op_.Run(&resize_img, this->mean_, this->scale_,
                           this->is_scale_);

deploy/cpp_infer/src/ocr_rec.cpp

@@ -33,7 +33,7 @@ void CRNNRecognizer::Run(std::vector<std::vector<std::vector<int>>> boxes,
 
     float wh_ratio = float(crop_img.cols) / float(crop_img.rows);
 
-    this->resize_op_.Run(crop_img, resize_img, wh_ratio);
+    this->resize_op_.Run(crop_img, resize_img, wh_ratio, this->use_tensorrt_);
 
     this->normalize_op_.Run(&resize_img, this->mean_, this->scale_,
                             this->is_scale_);

deploy/cpp_infer/src/preprocess_op.cpp

@@ -60,7 +60,8 @@ void Normalize::Run(cv::Mat *im, const std::vector<float> &mean,
 }
 
 void ResizeImgType0::Run(const cv::Mat &img, cv::Mat &resize_img,
-                         int max_size_len, float &ratio_h, float &ratio_w) {
+                         int max_size_len, float &ratio_h, float &ratio_w,
+                         bool use_tensorrt) {
   int w = img.cols;
   int h = img.rows;
 
@@ -89,14 +90,19 @@ void ResizeImgType0::Run(const cv::Mat &img, cv::Mat &resize_img,
     resize_w = 32;
   else
     resize_w = (resize_w / 32) * 32;
-
-  cv::resize(img, resize_img, cv::Size(resize_w, resize_h));
-
-  ratio_h = float(resize_h) / float(h);
-  ratio_w = float(resize_w) / float(w);
+  if (!use_tensorrt) {
+    cv::resize(img, resize_img, cv::Size(resize_w, resize_h));
+    ratio_h = float(resize_h) / float(h);
+    ratio_w = float(resize_w) / float(w);
+  } else {
+    cv::resize(img, resize_img, cv::Size(640, 640));
+    ratio_h = float(640) / float(h);
+    ratio_w = float(640) / float(w);
+  }
 }
 
 void CrnnResizeImg::Run(const cv::Mat &img, cv::Mat &resize_img, float wh_ratio,
+                        bool use_tensorrt,
                         const std::vector<int> &rec_image_shape) {
   int imgC, imgH, imgW;
   imgC = rec_image_shape[0];
@@ -111,12 +117,27 @@ void CrnnResizeImg::Run(const cv::Mat &img, cv::Mat &resize_img, float wh_ratio,
     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);
+  if (!use_tensorrt) {
+    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,
+                       int(imgW - resize_img.cols), cv::BORDER_CONSTANT,
+                       {127, 127, 127});
+  } else {
+    int k = int(img.cols * 32 / img.rows);
+    if (k >= 100) {
+      cv::resize(img, resize_img, cv::Size(100, 32), 0.f, 0.f,
+                 cv::INTER_LINEAR);
+    } else {
+      cv::resize(img, resize_img, cv::Size(k, 32), 0.f, 0.f, cv::INTER_LINEAR);
+      cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0, int(100 - k),
+                         cv::BORDER_CONSTANT, {127, 127, 127});
+    }
+  }
 }
 
 void ClsResizeImg::Run(const cv::Mat &img, cv::Mat &resize_img,
+                       bool use_tensorrt,
                        const std::vector<int> &rec_image_shape) {
   int imgC, imgH, imgW;
   imgC = rec_image_shape[0];
@@ -130,11 +151,15 @@ void ClsResizeImg::Run(const cv::Mat &img, cv::Mat &resize_img,
   else
     resize_w = int(ceilf(imgH * ratio));
 
-  cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
-             cv::INTER_LINEAR);
-  if (resize_w < imgW) {
-    cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0, imgW - resize_w,
-                       cv::BORDER_CONSTANT, cv::Scalar(0, 0, 0));
+  if (!use_tensorrt) {
+    cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
+               cv::INTER_LINEAR);
+    if (resize_w < imgW) {
+      cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0, imgW - resize_w,
+                         cv::BORDER_CONSTANT, cv::Scalar(0, 0, 0));
+    }
+  } else {
+    cv::resize(img, resize_img, cv::Size(100, 32), 0.f, 0.f, cv::INTER_LINEAR);
   }
 }
 

doc/doc_ch/algorithm_overview.md

@@ -21,13 +21,13 @@ PaddleOCR开源的文本检测算法列表：
 |EAST|MobileNetV3|78.24%|79.15%|78.69%|[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|91.83%|81.80%|86.52%|[下载链接](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_icdar15_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)|
 
 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|89.05%|76.80%|82.47%|[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%|[下载链接](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)
 

doc/doc_en/algorithm_overview_en.md

@@ -23,13 +23,13 @@ On the ICDAR2015 dataset, the text detection result is as follows:
 |EAST|MobileNetV3|78.24%|79.15%|78.69%|[Download link](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%|[Download link](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%|[Download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_mv3_db_v2.0_train.tar)|
-|SAST|ResNet50_vd|91.83%|81.80%|86.52%|[Download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_icdar15_v2.0_train.tar)|
+|SAST|ResNet50_vd|91.39%|83.77%|87.42%|[Download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_icdar15_v2.0_train.tar)|
 
 On Total-Text dataset, the text detection result is as follows:
 
 |Model|Backbone|precision|recall|Hmean|Download link|
 | --- | --- | --- | --- | --- | --- |
-|SAST|ResNet50_vd|89.05%|76.80%|82.47%|[Download link](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%|[Download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_totaltext_v2.0_train.tar)|
 
 **Note：** Additional data, like icdar2013, icdar2017, COCO-Text, ArT, was added to the model training of SAST. Download English public dataset in organized format used by PaddleOCR from [Baidu Drive](https://pan.baidu.com/s/12cPnZcVuV1zn5DOd4mqjVw) (download code: 2bpi).
 

ppocr/data/__init__.py

@@ -66,8 +66,10 @@ def build_dataloader(config, mode, device, logger):
     batch_size = loader_config['batch_size_per_card']
     drop_last = loader_config['drop_last']
     num_workers = loader_config['num_workers']
-
-    use_shared_memory = False
+    if 'use_shared_memory' in loader_config.keys():
+        use_shared_memory = loader_config['use_shared_memory']
+    else:
+        use_shared_memory = True
     if mode == "Train":
         #Distribute data to multiple cards
         batch_sampler = DistributedBatchSampler(
@@ -75,7 +77,6 @@ def build_dataloader(config, mode, device, logger):
             batch_size=batch_size,
             shuffle=False,
             drop_last=drop_last)
-        use_shared_memory = True
     else:
         #Distribute data to single card
         batch_sampler = BatchSampler(