Merge branch 'dygraph' into add_multi_doc

configs/det/det_r50_vd_sast_icdar15.yml

@@ -20,6 +20,7 @@ Global:
   infer_img:
   save_res_path: ./output/sast_r50_vd_ic15/predicts_sast.txt
 
+
 Architecture:
   model_type: det
   algorithm: SAST

configs/e2e/e2e_r50_vd_pg.yml

+Global:
+  use_gpu: True
+  epoch_num: 600
+  log_smooth_window: 20
+  print_batch_step: 10
+  save_model_dir: ./output/pgnet_r50_vd_totaltext/
+  save_epoch_step: 10
+  # evaluation is run every 0 iterationss after the 1000th iteration
+  eval_batch_step: [ 0, 1000 ]
+  # 1. If pretrained_model is saved in static mode, such as classification pretrained model
+  #    from static branch, load_static_weights must be set as True.
+  # 2. If you want to finetune the pretrained models we provide in the docs,
+  #    you should set load_static_weights as False.
+  load_static_weights: False
+  cal_metric_during_train: False
+  pretrained_model:
+  checkpoints:
+  save_inference_dir:
+  use_visualdl: False
+  infer_img:
+  valid_set: totaltext # two mode: totaltext valid curved words, partvgg valid non-curved words
+  save_res_path: ./output/pgnet_r50_vd_totaltext/predicts_pgnet.txt
+  character_dict_path: ppocr/utils/ic15_dict.txt
+  character_type: EN
+  max_text_length: 50 # the max length in seq
+  max_text_nums: 30 # the max seq nums in a pic
+  tcl_len: 64
+
+Architecture:
+  model_type: e2e
+  algorithm: PGNet
+  Transform:
+  Backbone:
+    name: ResNet
+    layers: 50
+  Neck:
+    name: PGFPN
+  Head:
+    name: PGHead
+
+Loss:
+  name: PGLoss
+  tcl_bs: 64
+  max_text_length: 50 # the same as Global: max_text_length
+  max_text_nums: 30 # the same as Global：max_text_nums
+  pad_num: 36 # the length of dict for pad
+
+Optimizer:
+  name: Adam
+  beta1: 0.9
+  beta2: 0.999
+  lr:
+    learning_rate: 0.001
+  regularizer:
+    name: 'L2'
+    factor: 0
+
+
+PostProcess:
+  name: PGPostProcess
+  score_thresh: 0.5
+Metric:
+  name: E2EMetric
+  character_dict_path: ppocr/utils/ic15_dict.txt
+  main_indicator: f_score_e2e
+
+Train:
+  dataset:
+    name: PGDataSet
+    label_file_list: [.././train_data/total_text/train/]
+    ratio_list: [1.0]
+    data_format: icdar #two data format: icdar/textnet
+    transforms:
+      - DecodeImage: # load image
+          img_mode: BGR
+          channel_first: False
+      - PGProcessTrain:
+          batch_size: 14  # same as loader: batch_size_per_card
+          min_crop_size: 24
+          min_text_size: 4
+          max_text_size: 512
+      - KeepKeys:
+          keep_keys: [ 'images', 'tcl_maps', 'tcl_label_maps', 'border_maps','direction_maps', 'training_masks', 'label_list', 'pos_list', 'pos_mask' ] # dataloader will return list in this order
+  loader:
+    shuffle: True
+    drop_last: True
+    batch_size_per_card: 14
+    num_workers: 16
+
+Eval:
+  dataset:
+    name: PGDataSet
+    data_dir: ./train_data/
+    label_file_list: [./train_data/total_text/test/]
+    transforms:
+      - DecodeImage: # load image
+          img_mode: RGB
+          channel_first: False
+      - E2ELabelEncode:
+      - E2EResizeForTest:
+          max_side_len: 768
+      - NormalizeImage:
+          scale: 1./255.
+          mean: [ 0.485, 0.456, 0.406 ]
+          std: [ 0.229, 0.224, 0.225 ]
+          order: 'hwc'
+      - ToCHWImage:
+      - KeepKeys:
+          keep_keys: [ 'image', 'shape', 'polys', 'strs', 'tags' ]
+  loader:
+    shuffle: False
+    drop_last: False
+    batch_size_per_card: 1 # must be 1
+    num_workers: 2
\ No newline at end of file

doc/doc_ch/inference.md

@@ -13,6 +13,7 @@ inference 模型（`paddle.jit.save`保存的模型）
     - [检测模型转inference模型](#检测模型转inference模型)
     - [识别模型转inference模型](#识别模型转inference模型)  
     - [方向分类模型转inference模型](#方向分类模型转inference模型)
+    - [端到端模型转inference模型](#端到端模型转inference模型)
 
 - [二、文本检测模型推理](#文本检测模型推理)
     - [1. 超轻量中文检测模型推理](#超轻量中文检测模型推理)
@@ -27,10 +28,13 @@ inference 模型（`paddle.jit.save`保存的模型）
     - [4. 自定义文本识别字典的推理](#自定义文本识别字典的推理)
     - [5. 多语言模型的推理](#多语言模型的推理)
 
-- [四、方向分类模型推理](#方向识别模型推理)
+- [四、端到端模型推理](#端到端模型推理)
+    - [1. PGNet端到端模型推理](#PGNet端到端模型推理)
+
+- [五、方向分类模型推理](#方向识别模型推理)
     - [1. 方向分类模型推理](#方向分类模型推理)
 
-- [五、文本检测、方向分类和文字识别串联推理](#文本检测、方向分类和文字识别串联推理)
+- [六、文本检测、方向分类和文字识别串联推理](#文本检测、方向分类和文字识别串联推理)
     - [1. 超轻量中文OCR模型推理](#超轻量中文OCR模型推理)
     - [2. 其他模型推理](#其他模型推理)
 
@@ -118,6 +122,32 @@ python3 tools/export_model.py -c configs/cls/cls_mv3.yml -o Global.pretrained_mo
     ├── inference.pdiparams.info    # 分类inference模型的参数信息，可忽略
     └── inference.pdmodel           # 分类inference模型的program文件
 ```
+<a name="端到端模型转inference模型"></a>
+### 端到端模型转inference模型
+
+下载端到端模型：
+```
+wget -P ./ch_lite/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar && tar xf ./ch_lite/ch_ppocr_mobile_v2.0_cls_train.tar -C ./ch_lite/
+```
+
+端到端模型转inference模型与检测的方式相同，如下：
+```
+# -c 后面设置训练算法的yml配置文件
+# -o 配置可选参数
+# Global.pretrained_model 参数设置待转换的训练模型地址，不用添加文件后缀 .pdmodel，.pdopt或.pdparams。
+# Global.load_static_weights 参数需要设置为 False。
+# Global.save_inference_dir参数设置转换的模型将保存的地址。
+
+python3 tools/export_model.py -c configs/e2e/e2e_r50_vd_pg.yml -o Global.pretrained_model=./ch_lite/ch_ppocr_mobile_v2.0_cls_train/best_accuracy Global.load_static_weights=False Global.save_inference_dir=./inference/e2e/
+```
+
+转换成功后，在目录下有三个文件：
+```
+/inference/e2e/
+    ├── inference.pdiparams         # 分类inference模型的参数文件
+    ├── inference.pdiparams.info    # 分类inference模型的参数信息，可忽略
+    └── inference.pdmodel           # 分类inference模型的program文件
+```
 
 <a name="文本检测模型推理"></a>
 ## 二、文本检测模型推理
@@ -332,8 +362,38 @@ python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/korean/1.jpg" -
 Predicts of ./doc/imgs_words/korean/1.jpg:('바탕으로', 0.9948904)
 ```
 
+<a name="端到端模型推理"></a>
+## 四、端到端模型推理
+
+端到端模型推理，默认使用PGNet模型的配置参数。当不使用PGNet模型时，在推理时，需要通过传入相应的参数进行算法适配，细节参考下文。
+<a name="PGNet端到端模型推理"></a>
+### 1. PGNet端到端模型推理
+#### (1). 四边形文本检测模型（ICDAR2015）  
+首先将PGNet端到端训练过程中保存的模型，转换成inference model。以基于Resnet50_vd骨干网络，在ICDAR2015英文数据集训练的模型为例([模型下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/pgnet/en_server_pgnetA.tar))，可以使用如下命令进行转换：
+```
+python3 tools/export_model.py -c configs/e2e/e2e_r50_vd_pg.yml -o Global.pretrained_model=./en_server_pgnetA/iter_epoch_450 Global.load_static_weights=False Global.save_inference_dir=./inference/e2e
+```
+**PGNet端到端模型推理，需要设置参数`--e2e_algorithm="PGNet"`**，可以执行如下命令：
+```
+python3 tools/infer/predict_e2e.py --e2e_algorithm="PGNet" --image_dir="./doc/imgs_en/img_10.jpg" --e2e_model_dir="./inference/e2e/"  --e2e_pgnet_polygon=False
+```
+可视化文本检测结果默认保存到`./inference_results`文件夹里面，结果文件的名称前缀为'e2e_res'。结果示例如下：
+
+![](../imgs_results/e2e_res_img_10_pgnet.jpg)
+
+#### (2). 弯曲文本检测模型（Total-Text）  
+和四边形文本检测模型共用一个推理模型
+**PGNet端到端模型推理，需要设置参数`--e2e_algorithm="PGNet"`，同时，还需要增加参数`--e2e_pgnet_polygon=True`，**可以执行如下命令：
+```
+python3.7 tools/infer/predict_e2e.py --e2e_algorithm="PGNet" --image_dir="./doc/imgs_en/img623.jpg" --e2e_model_dir="./inference/e2e/" --e2e_pgnet_polygon=True
+```
+可视化文本端到端结果默认保存到`./inference_results`文件夹里面，结果文件的名称前缀为'e2e_res'。结果示例如下：
+
+![](../imgs_results/e2e_res_img623_pgnet.jpg)
+
+
 <a name="方向分类模型推理"></a>
-## 四、方向分类模型推理
+## 五、方向分类模型推理
 
 下面将介绍方向分类模型推理。
 
@@ -358,7 +418,7 @@ Predicts of ./doc/imgs_words/ch/word_4.jpg:['0', 0.9999982]
 ```
 
 <a name="文本检测、方向分类和文字识别串联推理"></a>
-## 五、文本检测、方向分类和文字识别串联推理
+## 六、文本检测、方向分类和文字识别串联推理
 <a name="超轻量中文OCR模型推理"></a>
 ### 1. 超轻量中文OCR模型推理
 

doc/doc_ch/pgnet.md

+# 端对端OCR算法-PGNet
+- [一、简介](#简介)
+- [二、环境配置](#环境配置)
+- [三、快速使用](#快速使用)
+- [四、快速训练](#开始训练)
+- [五、预测推理](#预测推理)
+
+
+<a name="简介"></a>
+##简介
+OCR算法可以分为两阶段算法和端对端的算法。二阶段OCR算法一般分为两个部分，文本检测和文本识别算法，文件检测算法从图像中得到文本行的检测框，然后识别算法去识别文本框中的内容。而端对端OCR算法可以在一个算法中完成文字检测和文字识别，其基本思想是设计一个同时具有检测单元和识别模块的模型，共享其中两者的CNN特征，并联合训练。由于一个算法即可完成文字识别，端对端模型更小，速度更快。
+
+### PGNet算法介绍
+近些年来，端对端OCR算法得到了良好的发展，包括MaskTextSpotter系列、TextSnake、TextDragon、PGNet系列等算法。在这些算法中，PGNet算法具备其他算法不具备的优势，包括：
+- 设计PGNet loss指导训练，不需要字符级别的标注
+- 不需要NMS和ROI相关操作，加速预测
+- 提出预测文本行内的阅读顺序模块；
+- 提出基于图的修正模块（GRM）来进一步提高模型识别性能
+- 精度更高，预测速度更快
+
+PGNet算法细节详见[论文](https://www.aaai.org/AAAI21Papers/AAAI-2885.WangP.pdf)， 算法原理图如下所示：
+![](../pgnet_framework.png)
+输入图像经过特征提取送入四个分支，分别是：文本边缘偏移量预测TBO模块，文本中心线预测TCL模块，文本方向偏移量预测TDO模块，以及文本字符分类图预测TCC模块。
+其中TBO以及TCL的输出经过后处理后可以得到文本的检测结果，TCL、TDO、TCC负责文本识别。
+其检测识别效果图如下：
+![](../imgs_results/e2e_res_img293_pgnet.png)
+![](../imgs_results/e2e_res_img295_pgnet.png)
+
+<a name="环境配置"></a>
+##环境配置
+请先参考[快速安装](./installation.md)配置PaddleOCR运行环境。
+
+*注意：也可以通过 whl 包安装使用PaddleOCR，具体参考[Paddleocr Package使用说明](./whl.md)。*
+
+<a name="快速使用"></a>
+##快速使用
+### inference模型下载
+本节以训练好的端到端模型为例，快速使用模型预测，首先下载训练好的端到端inference模型[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/pgnet/e2e_server_pgnetA_infer.tar)
+```
+mkdir inference && cd inference
+# 下载英文端到端模型并解压
+wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/pgnet/e2e_server_pgnetA_infer.tar && tar xf e2e_server_pgnetA_infer.tar
+```
+* windows 环境下如果没有安装wget,下载模型时可将链接复制到浏览器中下载，并解压放置在相应目录下
+
+解压完毕后应有如下文件结构：
+```
+├── e2e_server_pgnetA_infer
+│   ├── inference.pdiparams
+│   ├── inference.pdiparams.info
+│   └── inference.pdmodel
+```
+### 单张图像或者图像集合预测
+```bash
+# 预测image_dir指定的单张图像
+python3 tools/infer/predict_e2e.py --e2e_algorithm="PGNet" --image_dir="./doc/imgs_en/img623.jpg" --e2e_model_dir="./inference/e2e/" --e2e_pgnet_polygon=True
+
+# 预测image_dir指定的图像集合
+python3 tools/infer/predict_e2e.py --e2e_algorithm="PGNet" --image_dir="./doc/imgs_en/" --e2e_model_dir="./inference/e2e/" --e2e_pgnet_polygon=True
+
+# 如果想使用CPU进行预测，需设置use_gpu参数为False
+python3 tools/infer/predict_e2e.py --e2e_algorithm="PGNet" --image_dir="./doc/imgs_en/img623.jpg" --e2e_model_dir="./inference/e2e/" --e2e_pgnet_polygon=True --use_gpu=False
+```
+<a name="开始训练"></a>
+##开始训练
+本节以totaltext数据集为例，介绍PaddleOCR中端到端模型的训练、评估与测试。
+###数据形式为icdar, 十六点标注数据
+解压数据集和下载标注文件后，PaddleOCR/train_data/total_text/train/ 有两个文件夹，分别是：
+```
+/PaddleOCR/train_data/total_text/train/
+  |- rgb/            total_text数据集的训练数据
+      |- gt_0.png
+      | ...  
+  |- total_text.txt  total_text数据集的训练标注
+```
+
+提供的标注文件格式如下，中间用"\t"分隔：
+```
+" 图像文件名                    json.dumps编码的图像标注信息"
+rgb/gt_0.png    [{"transcription": "EST", "points": [[1004.0,689.0],[1019.0,698.0],[1034.0,708.0],[1049.0,718.0],[1064.0,728.0],[1079.0,738.0],[1095.0,748.0],[1094.0,774.0],[1079.0,765.0],[1065.0,756.0],[1050.0,747.0],[1036.0,738.0],[1021.0,729.0],[1007.0,721.0]]}, {...}]
+```
+json.dumps编码前的图像标注信息是包含多个字典的list，字典中的 `points` 表示文本框的四个点的坐标(x, y)，从左上角的点开始顺时针排列。
+`transcription` 表示当前文本框的文字，**当其内容为“###”时，表示该文本框无效，在训练时会跳过。**
+如果您想在其他数据集上训练，可以按照上述形式构建标注文件。
+
+### 快速启动训练
+
+模型训练一般分两步骤进行，第一步可以选择用合成数据训练，第二步加载第一步训练好的模型训练，这边我们提供了第一步训练好的模型，可以直接加载，从第二步开始训练
+[下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/pgnet/train_step1.tar)
+```shell
+cd PaddleOCR/
+下载ResNet50_vd的动态图预训练模型
+wget -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/dygraph_v2.0/pgnet/train_step1.tar
+可以得到以下的文件格式
+./pretrain_models/train_step1/
+  └─ best_accuracy.pdopt
+  └─ best_accuracy.states
+  └─ best_accuracy.pdparams
+
+```
+
+*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*
+
+```shell
+# 单机单卡训练 e2e 模型
+python3 tools/train.py -c configs/e2e/e2e_r50_vd_pg.yml -o Global.pretrained_model=./pretrain_models/train_step1/best_accuracy Global.load_static_weights=False
+# 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/e2e/e2e_r50_vd_pg.yml -o Global.pretrained_model=./pretrain_models/train_step1/best_accuracy  Global.load_static_weights=False
+```
+
+上述指令中，通过-c 选择训练使用configs/e2e/e2e_r50_vd_pg.yml配置文件。
+有关配置文件的详细解释，请参考[链接](./config.md)。
+
+您也可以通过-o参数在不需要修改yml文件的情况下，改变训练的参数，比如，调整训练的学习率为0.0001
+```shell
+python3 tools/train.py -c configs/e2e/e2e_r50_vd_pg.yml -o Optimizer.base_lr=0.0001
+```
+
+#### 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
+python3 tools/train.py -c configs/e2e/e2e_r50_vd_pg.yml -o Global.checkpoints=./your/trained/model
+```
+
+**注意**：`Global.checkpoints`的优先级高于`Global.pretrain_weights`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.pretrain_weights`指定的模型。
+
+<a name="预测推理"></a>
+## 预测推理
+
+PaddleOCR计算三个OCR端到端相关的指标，分别是：Precision、Recall、Hmean。
+
+运行如下代码，根据配置文件`e2e_r50_vd_pg.yml`中`save_res_path`指定的测试集检测结果文件，计算评估指标。
+
+评估时设置后处理参数`max_side_len=768`，使用不同数据集、不同模型训练，可调整参数进行优化
+训练中模型参数默认保存在`Global.save_model_dir`目录下。在评估指标时，需要设置`Global.checkpoints`指向保存的参数文件。
+```shell
+python3 tools/eval.py -c configs/e2e/e2e_r50_vd_pg.yml  -o Global.checkpoints="{path/to/weights}/best_accuracy"
+```
+
+### 测试端到端效果
+测试单张图像的端到端识别效果
+```shell
+python3 tools/infer_e2e.py -c configs/e2e/e2e_r50_vd_pg.yml -o Global.infer_img="./doc/imgs_en/img_10.jpg" Global.pretrained_model="./output/det_db/best_accuracy" Global.load_static_weights=false
+```
+
+测试文件夹下所有图像的端到端识别效果
+```shell
+python3 tools/infer_e2e.py -c configs/e2e/e2e_r50_vd_pg.yml -o Global.infer_img="./doc/imgs_en/" Global.pretrained_model="./output/det_db/best_accuracy" Global.load_static_weights=false
+```
+
+###转为推理模型
+### (1). 四边形文本检测模型（ICDAR2015）  
+首先将PGNet端到端训练过程中保存的模型，转换成inference model。以基于Resnet50_vd骨干网络，以英文数据集训练的模型为例[模型下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/pgnet/en_server_pgnetA.tar) ，可以使用如下命令进行转换：
+```
+wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/pgnet/en_server_pgnetA.tar && tar xf en_server_pgnetA.tar
+python3 tools/export_model.py -c configs/e2e/e2e_r50_vd_pg.yml -o Global.pretrained_model=./en_server_pgnetA/iter_epoch_450 Global.load_static_weights=False Global.save_inference_dir=./inference/e2e
+```
+**PGNet端到端模型推理，需要设置参数`--e2e_algorithm="PGNet"`**，可以执行如下命令：
+```
+python3 tools/infer/predict_e2e.py --e2e_algorithm="PGNet" --image_dir="./doc/imgs_en/img_10.jpg" --e2e_model_dir="./inference/e2e/"  --e2e_pgnet_polygon=False
+```
+可视化文本检测结果默认保存到`./inference_results`文件夹里面，结果文件的名称前缀为'e2e_res'。结果示例如下：
+
+![](../imgs_results/e2e_res_img_10_pgnet.jpg)
+
+### (2). 弯曲文本检测模型（Total-Text）
+对于弯曲文本样例
+
+**PGNet端到端模型推理，需要设置参数`--e2e_algorithm="PGNet"`，同时，还需要增加参数`--e2e_pgnet_polygon=True`，**可以执行如下命令：
+```
+python3 tools/infer/predict_e2e.py --e2e_algorithm="PGNet" --image_dir="./doc/imgs_en/img623.jpg" --e2e_model_dir="./inference/e2e/" --e2e_pgnet_polygon=True
+```
+可视化文本端到端结果默认保存到`./inference_results`文件夹里面，结果文件的名称前缀为'e2e_res'。结果示例如下：
+
+![](../imgs_results/e2e_res_img623_pgnet.jpg)

ppocr/data/__init__.py

@@ -34,6 +34,7 @@ import paddle.distributed as dist
 from ppocr.data.imaug import transform, create_operators
 from ppocr.data.simple_dataset import SimpleDataSet
 from ppocr.data.lmdb_dataset import LMDBDataSet
+from ppocr.data.pgnet_dataset import PGDataSet
 
 __all__ = ['build_dataloader', 'transform', 'create_operators']
 
@@ -54,7 +55,7 @@ signal.signal(signal.SIGTERM, term_mp)
 def build_dataloader(config, mode, device, logger, seed=None):
     config = copy.deepcopy(config)
 
-    support_dict = ['SimpleDataSet', 'LMDBDataSet']
+    support_dict = ['SimpleDataSet', 'LMDBDataSet', 'PGDataSet']
     module_name = config[mode]['dataset']['name']
     assert module_name in support_dict, Exception(
         'DataSet only support {}'.format(support_dict))
@@ -72,14 +73,14 @@ def build_dataloader(config, mode, device, logger, seed=None):
     else:
         use_shared_memory = True
     if mode == "Train":
-        #Distribute data to multiple cards
+        # Distribute data to multiple cards
         batch_sampler = DistributedBatchSampler(
             dataset=dataset,
             batch_size=batch_size,
             shuffle=shuffle,
             drop_last=drop_last)
     else:
-        #Distribute data to single card
+        # Distribute data to single card
         batch_sampler = BatchSampler(
             dataset=dataset,
             batch_size=batch_size,

ppocr/data/imaug/__init__.py

@@ -28,6 +28,7 @@ from .label_ops import *
 
 from .east_process import *
 from .sast_process import *
+from .pg_process import *
 
 
 def transform(data, ops=None):

ppocr/data/imaug/label_ops.py

@@ -187,6 +187,34 @@ class CTCLabelEncode(BaseRecLabelEncode):
         return dict_character
 
 
+class E2ELabelEncode(BaseRecLabelEncode):
+    def __init__(self,
+                 max_text_length,
+                 character_dict_path=None,
+                 character_type='EN',
+                 use_space_char=False,
+                 **kwargs):
+        super(E2ELabelEncode,
+              self).__init__(max_text_length, character_dict_path,
+                             character_type, use_space_char)
+        self.pad_num = len(self.dict)  # the length to pad
+
+    def __call__(self, data):
+        text_label_index_list, temp_text = [], []
+        texts = data['strs']
+        for text in texts:
+            text = text.lower()
+            temp_text = []
+            for c_ in text:
+                if c_ in self.dict:
+                    temp_text.append(self.dict[c_])
+            temp_text = temp_text + [self.pad_num] * (self.max_text_len -
+                                                      len(temp_text))
+            text_label_index_list.append(temp_text)
+        data['strs'] = np.array(text_label_index_list)
+        return data
+
+
 class AttnLabelEncode(BaseRecLabelEncode):
     """ Convert between text-label and text-index """
 

ppocr/data/imaug/operators.py

@@ -197,7 +197,6 @@ class DetResizeForTest(object):
             sys.exit(0)
         ratio_h = resize_h / float(h)
         ratio_w = resize_w / float(w)
-        # return img, np.array([h, w])
         return img, [ratio_h, ratio_w]
 
     def resize_image_type2(self, img):
@@ -206,7 +205,6 @@ class DetResizeForTest(object):
         resize_w = w
         resize_h = h
 
-        # Fix the longer side
         if resize_h > resize_w:
             ratio = float(self.resize_long) / resize_h
         else:
@@ -223,3 +221,72 @@ class DetResizeForTest(object):
         ratio_w = resize_w / float(w)
 
         return img, [ratio_h, ratio_w]
+
+
+class E2EResizeForTest(object):
+    def __init__(self, **kwargs):
+        super(E2EResizeForTest, self).__init__()
+        self.max_side_len = kwargs['max_side_len']
+        self.valid_set = kwargs['valid_set']
+
+    def __call__(self, data):
+        img = data['image']
+        src_h, src_w, _ = img.shape
+        if self.valid_set == 'totaltext':
+            im_resized, [ratio_h, ratio_w] = self.resize_image_for_totaltext(
+                img, max_side_len=self.max_side_len)
+        else:
+            im_resized, (ratio_h, ratio_w) = self.resize_image(
+                img, max_side_len=self.max_side_len)
+        data['image'] = im_resized
+        data['shape'] = np.array([src_h, src_w, ratio_h, ratio_w])
+        return data
+
+    def resize_image_for_totaltext(self, im, max_side_len=512):
+
+        h, w, _ = im.shape
+        resize_w = w
+        resize_h = h
+        ratio = 1.25
+        if h * ratio > max_side_len:
+            ratio = float(max_side_len) / resize_h
+        resize_h = int(resize_h * ratio)
+        resize_w = int(resize_w * ratio)
+
+        max_stride = 128
+        resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+        resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+        im = cv2.resize(im, (int(resize_w), int(resize_h)))
+        ratio_h = resize_h / float(h)
+        ratio_w = resize_w / float(w)
+        return im, (ratio_h, ratio_w)
+
+    def resize_image(self, im, max_side_len=512):
+        """
+        resize image to a size multiple of max_stride 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
+
+        # Fix the longer side
+        if resize_h > resize_w:
+            ratio = float(max_side_len) / resize_h
+        else:
+            ratio = float(max_side_len) / resize_w
+
+        resize_h = int(resize_h * ratio)
+        resize_w = int(resize_w * ratio)
+
+        max_stride = 128
+        resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
+        resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
+        im = cv2.resize(im, (int(resize_w), int(resize_h)))
+        ratio_h = resize_h / float(h)
+        ratio_w = resize_w / float(w)
+
+        return im, (ratio_h, ratio_w)

ppocr/data/pgnet_dataset.py

+# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import numpy as np
+import os
+from paddle.io import Dataset
+from .imaug import transform, create_operators
+import random
+
+
+class PGDataSet(Dataset):
+    def __init__(self, config, mode, logger, seed=None):
+        super(PGDataSet, self).__init__()
+
+        self.logger = logger
+        self.seed = seed
+        self.mode = mode
+        global_config = config['Global']
+        dataset_config = config[mode]['dataset']
+        loader_config = config[mode]['loader']
+
+        label_file_list = dataset_config.pop('label_file_list')
+        data_source_num = len(label_file_list)
+        ratio_list = dataset_config.get("ratio_list", [1.0])
+        if isinstance(ratio_list, (float, int)):
+            ratio_list = [float(ratio_list)] * int(data_source_num)
+        self.data_format = dataset_config.get('data_format', 'icdar')
+        assert len(
+            ratio_list
+        ) == data_source_num, "The length of ratio_list should be the same as the file_list."
+        self.do_shuffle = loader_config['shuffle']
+
+        logger.info("Initialize indexs of datasets:%s" % label_file_list)
+        self.data_lines = self.get_image_info_list(label_file_list, ratio_list,
+                                                   self.data_format)
+        self.data_idx_order_list = list(range(len(self.data_lines)))
+        if mode.lower() == "train":
+            self.shuffle_data_random()
+
+        self.ops = create_operators(dataset_config['transforms'], global_config)
+
+    def shuffle_data_random(self):
+        if self.do_shuffle:
+            random.seed(self.seed)
+            random.shuffle(self.data_lines)
+        return
+
+    def extract_polys(self, poly_txt_path):
+        """
+        Read text_polys, txt_tags, txts from give txt file.
+        """
+        text_polys, txt_tags, txts = [], [], []
+        with open(poly_txt_path) as f:
+            for line in f.readlines():
+                poly_str, txt = line.strip().split('\t')
+                poly = list(map(float, poly_str.split(',')))
+                if self.mode.lower() == "eval":
+                    while len(poly) < 100:
+                        poly.append(-1)
+                text_polys.append(
+                    np.array(
+                        poly, dtype=np.float32).reshape(-1, 2))
+                txts.append(txt)
+                txt_tags.append(txt == '###')
+
+        return np.array(list(map(np.array, text_polys))), \
+               np.array(txt_tags, dtype=np.bool), txts
+
+    def extract_info_textnet(self, im_fn, img_dir=''):
+        """
+        Extract information from line in textnet format.
+        """
+        info_list = im_fn.split('\t')
+        img_path = ''
+        for ext in [
+                'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif', 'tiff', 'gif', 'JPG'
+        ]:
+            if os.path.exists(os.path.join(img_dir, info_list[0] + "." + ext)):
+                img_path = os.path.join(img_dir, info_list[0] + "." + ext)
+                break
+
+        if img_path == '':
+            print('Image {0} NOT found in {1}, and it will be ignored.'.format(
+                info_list[0], img_dir))
+
+        nBox = (len(info_list) - 1) // 9
+        wordBBs, txts, txt_tags = [], [], []
+        for n in range(0, nBox):
+            wordBB = list(map(float, info_list[n * 9 + 1:(n + 1) * 9]))
+            txt = info_list[(n + 1) * 9]
+            wordBBs.append([[wordBB[0], wordBB[1]], [wordBB[2], wordBB[3]],
+                            [wordBB[4], wordBB[5]], [wordBB[6], wordBB[7]]])
+            txts.append(txt)
+            if txt == '###':
+                txt_tags.append(True)
+            else:
+                txt_tags.append(False)
+        return img_path, np.array(wordBBs, dtype=np.float32), txt_tags, txts
+
+    def get_image_info_list(self, file_list, ratio_list, data_format='textnet'):
+        if isinstance(file_list, str):
+            file_list = [file_list]
+        data_lines = []
+        for idx, data_source in enumerate(file_list):
+            image_files = []
+            if data_format == 'icdar':
+                image_files = [(data_source, x) for x in
+                               os.listdir(os.path.join(data_source, 'rgb'))
+                               if x.split('.')[-1] in [
+                                   'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif',
+                                   'tiff', 'gif', 'JPG'
+                               ]]
+            elif data_format == 'textnet':
+                with open(data_source) as f:
+                    image_files = [(data_source, x.strip())
+                                   for x in f.readlines()]
+            else:
+                print("Unrecognized data format...")
+                exit(-1)
+            random.seed(self.seed)
+            image_files = random.sample(
+                image_files, round(len(image_files) * ratio_list[idx]))
+            data_lines.extend(image_files)
+        return data_lines
+
+    def __getitem__(self, idx):
+        file_idx = self.data_idx_order_list[idx]
+        data_path, data_line = self.data_lines[file_idx]
+        try:
+            if self.data_format == 'icdar':
+                im_path = os.path.join(data_path, 'rgb', data_line)
+                if self.mode.lower() == "eval":
+                    poly_path = os.path.join(data_path, 'poly_gt',
+                                             data_line.split('.')[0] + '.txt')
+                else:
+                    poly_path = os.path.join(data_path, 'poly',
+                                             data_line.split('.')[0] + '.txt')
+                text_polys, text_tags, text_strs = self.extract_polys(poly_path)
+            else:
+                image_dir = os.path.join(os.path.dirname(data_path), 'image')
+                im_path, text_polys, text_tags, text_strs = self.extract_info_textnet(
+                    data_line, image_dir)
+
+            data = {
+                'img_path': im_path,
+                'polys': text_polys,
+                'tags': text_tags,
+                'strs': text_strs
+            }
+            with open(data['img_path'], 'rb') as f:
+                img = f.read()
+                data['image'] = img
+            outs = transform(data, self.ops)
+
+        except Exception as e:
+            self.logger.error(
+                "When parsing line {}, error happened with msg: {}".format(
+                    self.data_idx_order_list[idx], e))
+            outs = None
+        if outs is None:
+            return self.__getitem__(np.random.randint(self.__len__()))
+        return outs
+
+    def __len__(self):
+        return len(self.data_idx_order_list)

ppocr/losses/__init__.py

@@ -29,10 +29,11 @@ def build_loss(config):
     # cls loss
     from .cls_loss import ClsLoss
 
+    # e2e loss
+    from .e2e_pg_loss import PGLoss
     support_dict = [
         'DBLoss', 'EASTLoss', 'SASTLoss', 'CTCLoss', 'ClsLoss', 'AttentionLoss',
-        'SRNLoss'
-    ]
+        'SRNLoss', 'PGLoss']
 
     config = copy.deepcopy(config)
     module_name = config.pop('name')

ppocr/losses/e2e_pg_loss.py

+# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from paddle import nn
+import paddle
+
+from .det_basic_loss import DiceLoss
+from ppocr.utils.e2e_utils.extract_batchsize import pre_process
+
+
+class PGLoss(nn.Layer):
+    def __init__(self,
+                 tcl_bs,
+                 max_text_length,
+                 max_text_nums,
+                 pad_num,
+                 eps=1e-6,
+                 **kwargs):
+        super(PGLoss, self).__init__()
+        self.tcl_bs = tcl_bs
+        self.max_text_nums = max_text_nums
+        self.max_text_length = max_text_length
+        self.pad_num = pad_num
+        self.dice_loss = DiceLoss(eps=eps)
+
+    def border_loss(self, f_border, l_border, l_score, l_mask):
+        l_border_split, l_border_norm = paddle.tensor.split(
+            l_border, num_or_sections=[4, 1], axis=1)
+        f_border_split = f_border
+        b, c, h, w = l_border_norm.shape
+        l_border_norm_split = paddle.expand(
+            x=l_border_norm, shape=[b, 4 * c, h, w])
+        b, c, h, w = l_score.shape
+        l_border_score = paddle.expand(x=l_score, shape=[b, 4 * c, h, w])
+        b, c, h, w = l_mask.shape
+        l_border_mask = paddle.expand(x=l_mask, shape=[b, 4 * c, h, w])
+        border_diff = l_border_split - f_border_split
+        abs_border_diff = paddle.abs(border_diff)
+        border_sign = abs_border_diff < 1.0
+        border_sign = paddle.cast(border_sign, dtype='float32')
+        border_sign.stop_gradient = True
+        border_in_loss = 0.5 * abs_border_diff * abs_border_diff * border_sign + \
+                         (abs_border_diff - 0.5) * (1.0 - border_sign)
+        border_out_loss = l_border_norm_split * border_in_loss
+        border_loss = paddle.sum(border_out_loss * l_border_score * l_border_mask) / \
+                      (paddle.sum(l_border_score * l_border_mask) + 1e-5)
+        return border_loss
+
+    def direction_loss(self, f_direction, l_direction, l_score, l_mask):
+        l_direction_split, l_direction_norm = paddle.tensor.split(
+            l_direction, num_or_sections=[2, 1], axis=1)
+        f_direction_split = f_direction
+        b, c, h, w = l_direction_norm.shape
+        l_direction_norm_split = paddle.expand(
+            x=l_direction_norm, shape=[b, 2 * c, h, w])
+        b, c, h, w = l_score.shape
+        l_direction_score = paddle.expand(x=l_score, shape=[b, 2 * c, h, w])
+        b, c, h, w = l_mask.shape
+        l_direction_mask = paddle.expand(x=l_mask, shape=[b, 2 * c, h, w])
+        direction_diff = l_direction_split - f_direction_split
+        abs_direction_diff = paddle.abs(direction_diff)
+        direction_sign = abs_direction_diff < 1.0
+        direction_sign = paddle.cast(direction_sign, dtype='float32')
+        direction_sign.stop_gradient = True
+        direction_in_loss = 0.5 * abs_direction_diff * abs_direction_diff * direction_sign + \
+                            (abs_direction_diff - 0.5) * (1.0 - direction_sign)
+        direction_out_loss = l_direction_norm_split * direction_in_loss
+        direction_loss = paddle.sum(direction_out_loss * l_direction_score * l_direction_mask) / \
+                         (paddle.sum(l_direction_score * l_direction_mask) + 1e-5)
+        return direction_loss
+
+    def ctcloss(self, f_char, tcl_pos, tcl_mask, tcl_label, label_t):
+        f_char = paddle.transpose(f_char, [0, 2, 3, 1])
+        tcl_pos = paddle.reshape(tcl_pos, [-1, 3])
+        tcl_pos = paddle.cast(tcl_pos, dtype=int)
+        f_tcl_char = paddle.gather_nd(f_char, tcl_pos)
+        f_tcl_char = paddle.reshape(f_tcl_char,
+                                    [-1, 64, 37])  # len(Lexicon_Table)+1
+        f_tcl_char_fg, f_tcl_char_bg = paddle.split(f_tcl_char, [36, 1], axis=2)
+        f_tcl_char_bg = f_tcl_char_bg * tcl_mask + (1.0 - tcl_mask) * 20.0
+        b, c, l = tcl_mask.shape
+        tcl_mask_fg = paddle.expand(x=tcl_mask, shape=[b, c, 36 * l])
+        tcl_mask_fg.stop_gradient = True
+        f_tcl_char_fg = f_tcl_char_fg * tcl_mask_fg + (1.0 - tcl_mask_fg) * (
+            -20.0)
+        f_tcl_char_mask = paddle.concat([f_tcl_char_fg, f_tcl_char_bg], axis=2)
+        f_tcl_char_ld = paddle.transpose(f_tcl_char_mask, (1, 0, 2))
+        N, B, _ = f_tcl_char_ld.shape
+        input_lengths = paddle.to_tensor([N] * B, dtype='int64')
+        cost = paddle.nn.functional.ctc_loss(
+            log_probs=f_tcl_char_ld,
+            labels=tcl_label,
+            input_lengths=input_lengths,
+            label_lengths=label_t,
+            blank=self.pad_num,
+            reduction='none')
+        cost = cost.mean()
+        return cost
+
+    def forward(self, predicts, labels):
+        images, tcl_maps, tcl_label_maps, border_maps \
+            , direction_maps, training_masks, label_list, pos_list, pos_mask = labels
+        # for all the batch_size
+        pos_list, pos_mask, label_list, label_t = pre_process(
+            label_list, pos_list, pos_mask, self.max_text_length,
+            self.max_text_nums, self.pad_num, self.tcl_bs)
+
+        f_score, f_border, f_direction, f_char = predicts['f_score'], predicts['f_border'], predicts['f_direction'], \
+                                                 predicts['f_char']
+        score_loss = self.dice_loss(f_score, tcl_maps, training_masks)
+        border_loss = self.border_loss(f_border, border_maps, tcl_maps,
+                                       training_masks)
+        direction_loss = self.direction_loss(f_direction, direction_maps,
+                                             tcl_maps, training_masks)
+        ctc_loss = self.ctcloss(f_char, pos_list, pos_mask, label_list, label_t)
+        loss_all = score_loss + border_loss + direction_loss + 5 * ctc_loss
+
+        losses = {
+            'loss': loss_all,
+            "score_loss": score_loss,
+            "border_loss": border_loss,
+            "direction_loss": direction_loss,
+            "ctc_loss": ctc_loss
+        }
+        return losses

ppocr/metrics/__init__.py

@@ -26,8 +26,9 @@ def build_metric(config):
     from .det_metric import DetMetric
     from .rec_metric import RecMetric
     from .cls_metric import ClsMetric
+    from .e2e_metric import E2EMetric
 
-    support_dict = ['DetMetric', 'RecMetric', 'ClsMetric']
+    support_dict = ['DetMetric', 'RecMetric', 'ClsMetric', 'E2EMetric']
 
     config = copy.deepcopy(config)
     module_name = config.pop('name')

ppocr/metrics/e2e_metric.py

+# Copyright (c) 2021 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 __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+__all__ = ['E2EMetric']
+
+from ppocr.utils.e2e_metric.Deteval import get_socre, combine_results
+from ppocr.utils.e2e_utils.extract_textpoint import get_dict
+
+
+class E2EMetric(object):
+    def __init__(self,
+                 character_dict_path,
+                 main_indicator='f_score_e2e',
+                 **kwargs):
+        self.label_list = get_dict(character_dict_path)
+        self.max_index = len(self.label_list)
+        self.main_indicator = main_indicator
+        self.reset()
+
+    def __call__(self, preds, batch, **kwargs):
+        temp_gt_polyons_batch = batch[2]
+        temp_gt_strs_batch = batch[3]
+        ignore_tags_batch = batch[4]
+        gt_polyons_batch = []
+        gt_strs_batch = []
+
+        temp_gt_polyons_batch = temp_gt_polyons_batch[0].tolist()
+        for temp_list in temp_gt_polyons_batch:
+            t = []
+            for index in temp_list:
+                if index[0] != -1 and index[1] != -1:
+                    t.append(index)
+            gt_polyons_batch.append(t)
+
+        temp_gt_strs_batch = temp_gt_strs_batch[0].tolist()
+        for temp_list in temp_gt_strs_batch:
+            t = ""
+            for index in temp_list:
+                if index < self.max_index:
+                    t += self.label_list[index]
+            gt_strs_batch.append(t)
+
+        for pred, gt_polyons, gt_strs, ignore_tags in zip(
+            [preds], [gt_polyons_batch], [gt_strs_batch], ignore_tags_batch):
+            # prepare gt
+            gt_info_list = [{
+                'points': gt_polyon,
+                'text': gt_str,
+                'ignore': ignore_tag
+            } for gt_polyon, gt_str, ignore_tag in
+                            zip(gt_polyons, gt_strs, ignore_tags)]
+            # prepare det
+            e2e_info_list = [{
+                'points': det_polyon,
+                'text': pred_str
+            } for det_polyon, pred_str in zip(pred['points'], pred['strs'])]
+            result = get_socre(gt_info_list, e2e_info_list)
+            self.results.append(result)
+
+    def get_metric(self):
+        metircs = combine_results(self.results)
+        self.reset()
+        return metircs
+
+    def reset(self):
+        self.results = []  # clear results

ppocr/metrics/eval_det_iou.py

@@ -200,7 +200,8 @@ class DetectionIoUEvaluator(object):
         methodPrecision = 0 if numGlobalCareDet == 0 else float(
             matchedSum) / numGlobalCareDet
         methodHmean = 0 if methodRecall + methodPrecision == 0 else 2 * \
-            methodRecall * methodPrecision / (methodRecall + methodPrecision)
+                                                                    methodRecall * methodPrecision / (
+                                                                            methodRecall + methodPrecision)
         # print(methodRecall, methodPrecision, methodHmean)
         # sys.exit(-1)
         methodMetrics = {