Skip to content

P
PaddleOCR

weixin_41840029 / PaddleOCR
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleOCR

通知 1
Star 1
Fork 0
P
PaddleOCR
提交 96a8e0f8 编写于 作者: A andyjpaddle
浏览文件
操作

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

上级 29268eaa 6eb5efe0
显示空白变更内容
内联 并排
Showing with 4627 addition and 1948 deletion
README.md
浏览文件 @ 96a8e0f8
...@@ -113,18 +113,19 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel ...@@ -113,18 +113,19 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel
- [Quick Start](./ppstructure/docs/quickstart_en.md) - [Quick Start](./ppstructure/docs/quickstart_en.md)
- [Model Zoo](./ppstructure/docs/models_list_en.md) - [Model Zoo](./ppstructure/docs/models_list_en.md)
- [Model training](./doc/doc_en/training_en.md) - [Model training](./doc/doc_en/training_en.md)
- [Layout Parser](./ppstructure/layout/README.md) - [Layout Analysis](./ppstructure/layout/README.md)
- [Table Recognition](./ppstructure/table/README.md) - [Table Recognition](./ppstructure/table/README.md)
- [DocVQA](./ppstructure/vqa/README.md) - [Key Information Extraction](./ppstructure/kie/README.md)
- [Key Information Extraction](./ppstructure/docs/kie_en.md)
- [Inference and Deployment](./deploy/README.md) - [Inference and Deployment](./deploy/README.md)
- [Python Inference](./ppstructure/docs/inference_en.md) - [Python Inference](./ppstructure/docs/inference_en.md)
- [C++ Inference]() - [C++ Inference](./deploy/cpp_infer/readme.md)
- [Serving](./deploy/pdserving/README.md) - [Serving](./deploy/pdserving/README.md)
- [Academic algorithms](./doc/doc_en/algorithms_en.md) - [Academic Algorithms](./doc/doc_en/algorithm_overview_en.md)
- [Text detection](./doc/doc_en/algorithm_overview_en.md) - [Text detection](./doc/doc_en/algorithm_overview_en.md)
- [Text recognition](./doc/doc_en/algorithm_overview_en.md) - [Text recognition](./doc/doc_en/algorithm_overview_en.md)
- [End-to-end](./doc/doc_en/algorithm_overview_en.md) - [End-to-end OCR](./doc/doc_en/algorithm_overview_en.md)
- [Table Recognition](./doc/doc_en/algorithm_overview_en.md)
- [Key Information Extraction](./doc/doc_en/algorithm_overview_en.md)
- [Add New Algorithms to PaddleOCR](./doc/doc_en/add_new_algorithm_en.md) - [Add New Algorithms to PaddleOCR](./doc/doc_en/add_new_algorithm_en.md)
- Data Annotation and Synthesis - Data Annotation and Synthesis
- [Semi-automatic Annotation Tool: PPOCRLabel](./PPOCRLabel/README.md) - [Semi-automatic Annotation Tool: PPOCRLabel](./PPOCRLabel/README.md)
...@@ -135,9 +136,9 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel ...@@ -135,9 +136,9 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel
- [General OCR Datasets(Chinese/English)](doc/doc_en/dataset/datasets_en.md) - [General OCR Datasets(Chinese/English)](doc/doc_en/dataset/datasets_en.md)
- [HandWritten_OCR_Datasets(Chinese)](doc/doc_en/dataset/handwritten_datasets_en.md) - [HandWritten_OCR_Datasets(Chinese)](doc/doc_en/dataset/handwritten_datasets_en.md)
- [Various OCR Datasets(multilingual)](doc/doc_en/dataset/vertical_and_multilingual_datasets_en.md) - [Various OCR Datasets(multilingual)](doc/doc_en/dataset/vertical_and_multilingual_datasets_en.md)
- [layout analysis](doc/doc_en/dataset/layout_datasets_en.md) - [Layout Analysis](doc/doc_en/dataset/layout_datasets_en.md)
- [table recognition](doc/doc_en/dataset/table_datasets_en.md) - [Table Recognition](doc/doc_en/dataset/table_datasets_en.md)
- [DocVQA](doc/doc_en/dataset/docvqa_datasets_en.md) - [Key Information Extraction](doc/doc_en/dataset/kie_datasets_en.md)
- [Code Structure](./doc/doc_en/tree_en.md) - [Code Structure](./doc/doc_en/tree_en.md)
- [Visualization](#Visualization) - [Visualization](#Visualization)
- [Community](#Community) - [Community](#Community)
...@@ -176,7 +177,7 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel ...@@ -176,7 +177,7 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel
</details> </details>
<details open> <details open>
<summary>PP-Structure</summary> <summary>PP-Structurev2</summary>
- layout analysis + table recognition - layout analysis + table recognition
<div align="center"> <div align="center">
...@@ -185,12 +186,28 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel ...@@ -185,12 +186,28 @@ PaddleOCR support a variety of cutting-edge algorithms related to OCR, and devel
- SER (Semantic entity recognition) - SER (Semantic entity recognition)
<div align="center"> <div align="center">
<img src="./ppstructure/docs/vqa/result_ser/zh_val_0_ser.jpg" width="800"> <img src="https://user-images.githubusercontent.com/25809855/186094456-01a1dd11-1433-4437-9ab2-6480ac94ec0a.png" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185310636-6ce02f7c-790d-479f-b163-ea97a5a04808.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539517-ccf2372a-f026-4a7c-ad28-c741c770f60a.png" width="600">
</div> </div>
- RE (Relation Extraction) - RE (Relation Extraction)
<div align="center"> <div align="center">
<img src="./ppstructure/docs/vqa/result_re/zh_val_21_re.jpg" width="800"> <img src="https://user-images.githubusercontent.com/25809855/186094813-3a8e16cc-42e5-4982-b9f4-0134dfb5688d.png" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185393805-c67ff571-cf7e-4217-a4b0-8b396c4f22bb.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185540080-0431e006-9235-4b6d-b63d-0b3c6e1de48f.jpg" width="600">
</div> </div>
</details> </details>
......
README_ch.md
浏览文件 @ 96a8e0f8
...@@ -27,15 +27,8 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力 ...@@ -27,15 +27,8 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
## 近期更新 ## 近期更新
- **🔥2022.5.11~13 每晚8：30【超强OCR技术详解与产业应用实战】三日直播课** - **🔥2022.7 发布[OCR场景应用集合](./applications)**
- 11日：开源最强OCR系统PP-OCRv3揭秘 - 发布OCR场景应用集合，包含数码管、液晶屏、车牌、高精度SVTR模型等**7个垂类模型**，覆盖通用，制造、金融、交通行业的主要OCR垂类应用。
- 12日：云边端全覆盖的PP-OCRv3训练部署实战
- 13日：OCR产业应用全流程拆解与实战
赶紧扫码报名吧！
<div align="center">
<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "150" height = "150" />
</div>
- **🔥2022.5.9 发布PaddleOCR [release/2.5](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.5)** - **🔥2022.5.9 发布PaddleOCR [release/2.5](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.5)**
- 发布[PP-OCRv3](./doc/doc_ch/ppocr_introduction.md#pp-ocrv3)，速度可比情况下，中文场景效果相比于PP-OCRv2再提升5%，英文场景提升11%，80语种多语言模型平均识别准确率提升5%以上； - 发布[PP-OCRv3](./doc/doc_ch/ppocr_introduction.md#pp-ocrv3)，速度可比情况下，中文场景效果相比于PP-OCRv2再提升5%，英文场景提升11%，80语种多语言模型平均识别准确率提升5%以上；
...@@ -71,24 +64,22 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力 ...@@ -71,24 +64,22 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
## 《动手学OCR》电子书 ## 《动手学OCR》电子书
- [《动手学OCR》电子书📚](./doc/doc_ch/ocr_book.md) - [《动手学OCR》电子书📚](./doc/doc_ch/ocr_book.md)
## 场景应用
- PaddleOCR场景应用覆盖通用，制造、金融、交通行业的主要OCR垂类应用，在PP-OCR、PP-Structure的通用能力基础之上，以notebook的形式展示利用场景数据微调、模型优化方法、数据增广等内容，为开发者快速落地OCR应用提供示范与启发。详情可查看[README](./applications)。
<a name="开源社区"></a> <a name="开源社区"></a>
## 开源社区 ## 开源社区
- **项目合作📑：** 如果您是企业开发者且有明确的OCR垂类应用需求，填写[问卷](https://paddle.wjx.cn/vj/QwF7GKw.aspx)后可免费与官方团队展开不同层次的合作。
- **加入社区👬：** 微信扫描二维码并填写问卷之后，加入交流群领取福利 - **加入社区👬：** 微信扫描二维码并填写问卷之后，加入交流群领取福利
- **获取5月11-13日每晚20:30《OCR超强技术详解与产业应用实战》的直播课链接** - **获取PaddleOCR最新发版解说《OCR超强技术详解与产业应用实战》系列直播课回放链接**
- **10G重磅OCR学习大礼包：**《动手学OCR》电子书，配套讲解视频和notebook项目；66篇OCR相关顶会前沿论文打包放送，包括CVPR、AAAI、IJCAI、ICCV等；PaddleOCR历次发版直播课视频；OCR社区优秀开发者项目分享视频。 - **10G重磅OCR学习大礼包：**《动手学OCR》电子书，配套讲解视频和notebook项目；66篇OCR相关顶会前沿论文打包放送，包括CVPR、AAAI、IJCAI、ICCV等；PaddleOCR历次发版直播课视频；OCR社区优秀开发者项目分享视频。
- **社区项目**🏅️：[社区项目](./doc/doc_ch/thirdparty.md)文档中包含了社区用户**使用PaddleOCR开发的各种工具、应用**以及**为PaddleOCR贡献的功能、优化的文档与代码**等，是官方为社区开发者打造的荣誉墙，也是帮助优质项目宣传的广播站。
- **社区贡献**🏅️：[社区贡献](./doc/doc_ch/thirdparty.md)文档中包含了社区用户**使用PaddleOCR开发的各种工具、应用**以及**为PaddleOCR贡献的功能、优化的文档与代码**等，是官方为社区开发者打造的荣誉墙，也是帮助优质项目宣传的广播站。
- **社区常规赛**🎁：社区常规赛是面向OCR开发者的积分赛事，覆盖文档、代码、模型和应用四大类型，以季度为单位评选并发放奖励，赛题详情与报名方法可参考[链接](https://github.com/PaddlePaddle/PaddleOCR/issues/4982)。 - **社区常规赛**🎁：社区常规赛是面向OCR开发者的积分赛事，覆盖文档、代码、模型和应用四大类型，以季度为单位评选并发放奖励，赛题详情与报名方法可参考[链接](https://github.com/PaddlePaddle/PaddleOCR/issues/4982)。
<div align="center"> <div align="center">
<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "200" height = "200" /> <img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "150" height = "150" />
</div> </div>
<a name="模型下载"></a> <a name="模型下载"></a>
## PP-OCR系列模型列表（更新中） ## PP-OCR系列模型列表（更新中）
...@@ -96,14 +87,21 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力 ...@@ -96,14 +87,21 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
| ------------------------------------- | ----------------------- | --------------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ | | ------------------------------------- | ----------------------- | --------------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ |
| 中英文超轻量PP-OCRv3模型（16.2M） | ch_PP-OCRv3_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar) | | 中英文超轻量PP-OCRv3模型（16.2M） | ch_PP-OCRv3_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_train.tar) |
| 英文超轻量PP-OCRv3模型（13.4M） | en_PP-OCRv3_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_distill_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_train.tar) | | 英文超轻量PP-OCRv3模型（13.4M） | en_PP-OCRv3_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_distill_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_train.tar) |
| 中英文超轻量PP-OCRv2模型（13.0M） | ch_PP-OCRv2_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar) |
| 中英文超轻量PP-OCR mobile模型（9.4M） | ch_ppocr_mobile_v2.0_xx | 移动端&服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_pre.tar) |
| 中英文通用PP-OCR server模型（143.4M） | ch_ppocr_server_v2.0_xx | 服务器端 | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) | [推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar) / [预训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_pre.tar) |
更多模型下载（包括多语言），可以参考[PP-OCR 系列模型下载](./doc/doc_ch/models_list.md)，文档分析相关模型参考[PP-Structure 系列模型下载](./ppstructure/docs/models_list.md) - 超轻量OCR系列更多模型下载（包括多语言），可以参考[PP-OCR系列模型下载](./doc/doc_ch/models_list.md)，文档分析相关模型参考[PP-Structure系列模型下载](./ppstructure/docs/models_list.md)
### PaddleOCR场景应用模型
| 行业 | 类别 | 亮点 | 文档说明 | 模型下载 |
| ---- | ------------ | ---------------------------------- | ------------------------------------------------------------ | --------------------------------------------- |
| 制造 | 数码管识别 | 数码管数据合成、漏识别调优 | [光功率计数码管字符识别](./applications/光功率计数码管字符识别/光功率计数码管字符识别.md) | [下载链接](./applications/README.md#模型下载) |
| 金融 | 通用表单识别 | 多模态通用表单结构化提取 | [多模态表单识别](./applications/多模态表单识别.md) | [下载链接](./applications/README.md#模型下载) |
| 交通 | 车牌识别 | 多角度图像处理、轻量模型、端侧部署 | [轻量级车牌识别](./applications/轻量级车牌识别.md) | [下载链接](./applications/README.md#模型下载) |
- 更多制造、金融、交通行业的主要OCR垂类应用模型（如电表、液晶屏、高精度SVTR模型等），可参考[场景应用模型下载](./applications)
<a name="文档教程"></a> <a name="文档教程"></a>
## 文档教程 ## 文档教程
- [运行环境准备](./doc/doc_ch/environment.md) - [运行环境准备](./doc/doc_ch/environment.md)
...@@ -120,7 +118,7 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力 ...@@ -120,7 +118,7 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
- [知识蒸馏](./doc/doc_ch/knowledge_distillation.md) - [知识蒸馏](./doc/doc_ch/knowledge_distillation.md)
- [推理部署](./deploy/README_ch.md) - [推理部署](./deploy/README_ch.md)
- [基于Python预测引擎推理](./doc/doc_ch/inference_ppocr.md) - [基于Python预测引擎推理](./doc/doc_ch/inference_ppocr.md)
- [基于C++预测引擎推理](./deploy/cpp_infer/readme.md) - [基于C++预测引擎推理](./deploy/cpp_infer/readme_ch.md)
- [服务化部署](./deploy/pdserving/README_CN.md) - [服务化部署](./deploy/pdserving/README_CN.md)
- [端侧部署](./deploy/lite/readme.md) - [端侧部署](./deploy/lite/readme.md)
- [Paddle2ONNX模型转化与预测](./deploy/paddle2onnx/readme.md) - [Paddle2ONNX模型转化与预测](./deploy/paddle2onnx/readme.md)
...@@ -132,16 +130,17 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力 ...@@ -132,16 +130,17 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
- [模型训练](./doc/doc_ch/training.md) - [模型训练](./doc/doc_ch/training.md)
- [版面分析](./ppstructure/layout/README_ch.md) - [版面分析](./ppstructure/layout/README_ch.md)
- [表格识别](./ppstructure/table/README_ch.md) - [表格识别](./ppstructure/table/README_ch.md)
- [关键信息提取](./ppstructure/docs/kie.md) - [关键信息提取](./ppstructure/kie/README_ch.md)
- [DocVQA](./ppstructure/vqa/README_ch.md)
- [推理部署](./deploy/README_ch.md) - [推理部署](./deploy/README_ch.md)
- [基于Python预测引擎推理](./ppstructure/docs/inference.md) - [基于Python预测引擎推理](./ppstructure/docs/inference.md)
- [基于C++预测引擎推理]() - [基于C++预测引擎推理](./deploy/cpp_infer/readme_ch.md)
- [服务化部署](./deploy/pdserving/README_CN.md) - [服务化部署](./deploy/pdserving/README_CN.md)
- [前沿算法与模型🚀](./doc/doc_ch/algorithm.md) - [前沿算法与模型🚀](./doc/doc_ch/algorithm_overview.md)
- [文本检测算法](./doc/doc_ch/algorithm_overview.md#11-%E6%96%87%E6%9C%AC%E6%A3%80%E6%B5%8B%E7%AE%97%E6%B3%95) - [文本检测算法](./doc/doc_ch/algorithm_overview.md)
- [文本识别算法](./doc/doc_ch/algorithm_overview.md#12-%E6%96%87%E6%9C%AC%E8%AF%86%E5%88%AB%E7%AE%97%E6%B3%95) - [文本识别算法](./doc/doc_ch/algorithm_overview.md)
- [端到端算法](./doc/doc_ch/algorithm_overview.md#2-%E6%96%87%E6%9C%AC%E8%AF%86%E5%88%AB%E7%AE%97%E6%B3%95) - [端到端OCR算法](./doc/doc_ch/algorithm_overview.md)
- [表格识别算法](./doc/doc_ch/algorithm_overview.md)
- [关键信息抽取算法](./doc/doc_ch/algorithm_overview.md)
- [使用PaddleOCR架构添加新算法](./doc/doc_ch/add_new_algorithm.md) - [使用PaddleOCR架构添加新算法](./doc/doc_ch/add_new_algorithm.md)
- [场景应用](./applications) - [场景应用](./applications)
- 数据标注与合成 - 数据标注与合成
...@@ -155,7 +154,7 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力 ...@@ -155,7 +154,7 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
- [垂类多语言OCR数据集](doc/doc_ch/dataset/vertical_and_multilingual_datasets.md) - [垂类多语言OCR数据集](doc/doc_ch/dataset/vertical_and_multilingual_datasets.md)
- [版面分析数据集](doc/doc_ch/dataset/layout_datasets.md) - [版面分析数据集](doc/doc_ch/dataset/layout_datasets.md)
- [表格识别数据集](doc/doc_ch/dataset/table_datasets.md) - [表格识别数据集](doc/doc_ch/dataset/table_datasets.md)
- [DocVQA数据集](doc/doc_ch/dataset/docvqa_datasets.md) - [关键信息提取数据集](doc/doc_ch/dataset/kie_datasets.md)
- [代码组织结构](./doc/doc_ch/tree.md) - [代码组织结构](./doc/doc_ch/tree.md)
- [效果展示](#效果展示) - [效果展示](#效果展示)
- [《动手学OCR》电子书📚](./doc/doc_ch/ocr_book.md) - [《动手学OCR》电子书📚](./doc/doc_ch/ocr_book.md)
...@@ -214,12 +213,28 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力 ...@@ -214,12 +213,28 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
- SER（语义实体识别） - SER（语义实体识别）
<div align="center"> <div align="center">
<img src="./ppstructure/docs/vqa/result_ser/zh_val_0_ser.jpg" width="800"> <img src="https://user-images.githubusercontent.com/14270174/185310636-6ce02f7c-790d-479f-b163-ea97a5a04808.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539517-ccf2372a-f026-4a7c-ad28-c741c770f60a.png" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/25809855/186094456-01a1dd11-1433-4437-9ab2-6480ac94ec0a.png" width="600">
</div> </div>
- RE（关系提取） - RE（关系提取）
<div align="center"> <div align="center">
<img src="./ppstructure/docs/vqa/result_re/zh_val_21_re.jpg" width="800"> <img src="https://user-images.githubusercontent.com/14270174/185393805-c67ff571-cf7e-4217-a4b0-8b396c4f22bb.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185540080-0431e006-9235-4b6d-b63d-0b3c6e1de48f.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/25809855/186094813-3a8e16cc-42e5-4982-b9f4-0134dfb5688d.png" width="600">
</div> </div>
</details> </details>
......
applications/发票关键信息抽取.md 0 → 100644
浏览文件 @ 96a8e0f8
# 基于VI-LayoutXLM的发票关键信息抽取
- [1. 项目背景及意义](#1-项目背景及意义)
- [2. 项目内容](#2-项目内容)
- [3. 安装环境](#3-安装环境)
- [4. 关键信息抽取](#4-关键信息抽取)
- [4.1 文本检测](#41-文本检测)
- [4.2 文本识别](#42-文本识别)
- [4.3 语义实体识别](#43-语义实体识别)
- [4.4 关系抽取](#44-关系抽取)
## 1. 项目背景及意义
关键信息抽取在文档场景中被广泛使用，如身份证中的姓名、住址信息抽取，快递单中的姓名、联系方式等关键字段内容的抽取。传统基于模板匹配的方案需要针对不同的场景制定模板并进行适配，较为繁琐，不够鲁棒。基于该问题，我们借助飞桨提供的PaddleOCR套件中的关键信息抽取方案，实现对增值税发票场景的关键信息抽取。
## 2. 项目内容
本项目基于PaddleOCR开源套件，以VI-LayoutXLM多模态关键信息抽取模型为基础，针对增值税发票场景进行适配，提取该场景的关键信息。
## 3. 安装环境
```bash
# 首先git官方的PaddleOCR项目，安装需要的依赖
# 第一次运行打开该注释
git clone https://gitee.com/PaddlePaddle/PaddleOCR.git
cd PaddleOCR
# 安装PaddleOCR的依赖
pip install -r requirements.txt
# 安装关键信息抽取任务的依赖
pip install -r ./ppstructure/vqa/requirements.txt
```
## 4. 关键信息抽取
基于文档图像的关键信息抽取包含3个部分：（1）文本检测（2）文本识别（3）关键信息抽取方法，包括语义实体识别或者关系抽取，下面分别进行介绍。
### 4.1 文本检测
本文重点关注发票的关键信息抽取模型训练与预测过程，因此在关键信息抽取过程中，直接使用标注的文本检测与识别标注信息进行测试，如果你希望自定义该场景的文本检测模型，完成端到端的关键信息抽取部分，请参考[文本检测模型训练教程](../doc/doc_ch/detection.md)，按照训练数据格式准备数据，并完成该场景下垂类文本检测模型的微调过程。
### 4.2 文本识别
本文重点关注发票的关键信息抽取模型训练与预测过程，因此在关键信息抽取过程中，直接使用提供的文本检测与识别标注信息进行测试，如果你希望自定义该场景的文本检测模型，完成端到端的关键信息抽取部分，请参考[文本识别模型训练教程](../doc/doc_ch/recognition.md)，按照训练数据格式准备数据，并完成该场景下垂类文本识别模型的微调过程。
### 4.3 语义实体识别 （Semantic Entity Recognition）
语义实体识别指的是给定一段文本行，确定其类别（如`姓名`、`住址`等类别）。PaddleOCR中提供了基于VI-LayoutXLM的多模态语义实体识别方法，融合文本、位置与版面信息，相比LayoutXLM多模态模型，去除了其中的视觉骨干网络特征提取部分，引入符合阅读顺序的文本行排序方法，同时使用UDML联合互蒸馏方法进行训练，最终在精度与速度方面均超越LayoutXLM。更多关于VI-LayoutXLM的算法介绍与精度指标，请参考：[VI-LayoutXLM算法介绍](../doc/doc_ch/algorithm_kie_vi_layoutxlm.md)。
#### 4.3.1 准备数据
发票场景为例，我们首先需要标注出其中的关键字段，我们将其标注为`问题-答案`的key-value pair，如下，编号No为12270830，则`No`字段标注为question，`12270830`字段标注为answer。如下图所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185381131-76b6e260-04fe-46d9-baca-6bdd7fe0d0ce.jpg" width="800">
</div>
**注意：**
* 如果文本检测模型数据标注过程中，没有标注 **非关键信息内容** 的检测框，那么在标注关键信息抽取任务的时候，也不需要标注该部分，如上图所示；如果标注的过程，如果同时标注了**非关键信息内容** 的检测框，那么我们需要将该部分的label记为other。
* 标注过程中，需要以文本行为单位进行标注，无需标注单个字符的位置信息。
已经处理好的增值税发票数据集从这里下载：[增值税发票数据集下载链接](https://aistudio.baidu.com/aistudio/datasetdetail/165561)。
下载好发票数据集，并解压在train_data目录下，目录结构如下所示。
```
train_data
|--zzsfp
|---class_list.txt
|---imgs/
|---train.json
|---val.json
```
其中`class_list.txt`是包含`other`, `question`, `answer`，3个种类的的类别列表（不区分大小写），`imgs`目录底下，`train.json`与`val.json`分别表示训练与评估集合的标注文件。训练集中包含30张图片，验证集中包含8张图片。部分标注如下所示。
```py
b33.jpg [{"transcription": "No", "label": "question", "points": [[2882, 472], [3026, 472], [3026, 588], [2882, 588]], }, {"transcription": "12269563", "label": "answer", "points": [[3066, 448], [3598, 448], [3598, 576], [3066, 576]], ]}]
```
相比于OCR检测的标注，仅多了`label`字段。
#### 4.3.2 开始训练
VI-LayoutXLM的配置为[ser_vi_layoutxlm_xfund_zh_udml.yml](../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml)，需要修改数据、类别数目以及配置文件。
```yml
Architecture:
model_type: &model_type "vqa"
name: DistillationModel
algorithm: Distillation
Models:
Teacher:
pretrained:
freeze_params: false
return_all_feats: true
model_type: *model_type
algorithm: &algorithm "LayoutXLM"
Transform:
Backbone:
name: LayoutXLMForSer
pretrained: True
# one of base or vi
mode: vi
checkpoints:
# 定义类别数目
num_classes: &num_classes 5
...
PostProcess:
name: DistillationSerPostProcess
model_name: ["Student", "Teacher"]
key: backbone_out
# 定义类别文件
class_path: &class_path train_data/zzsfp/class_list.txt
Train:
dataset:
name: SimpleDataSet
# 定义训练数据目录与标注文件
data_dir: train_data/zzsfp/imgs
label_file_list:
- train_data/zzsfp/train.json
...
Eval:
dataset:
# 定义评估数据目录与标注文件
name: SimpleDataSet
data_dir: train_data/zzsfp/imgs
label_file_list:
- train_data/zzsfp/val.json
...
```
LayoutXLM与VI-LayoutXLM针对该场景的训练结果如下所示。
| 模型 | 迭代轮数 | Hmean |
| :---: | :---: | :---: |
| LayoutXLM | 50 | 100% |
| VI-LayoutXLM | 50 | 100% |
可以看出，由于当前数据量较少，场景比较简单，因此2个模型的Hmean均达到了100%。
#### 4.3.3 模型评估
模型训练过程中，使用的是知识蒸馏的策略，最终保留了学生模型的参数，在评估时，我们需要针对学生模型的配置文件进行修改: [ser_vi_layoutxlm_xfund_zh.yml](../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml)，修改内容与训练配置相同，包括**类别数、类别映射文件、数据目录**。
修改完成后，执行下面的命令完成评估过程。
```bash
# 注意：需要根据你的配置文件地址与保存的模型地址，对评估命令进行修改
python3 tools/eval.py -c ./fapiao/ser_vi_layoutxlm.yml -o Architecture.Backbone.checkpoints=fapiao/models/ser_vi_layoutxlm_fapiao_udml/best_accuracy
```
输出结果如下所示。
```
[2022/08/18 08:49:58] ppocr INFO: metric eval ***************
[2022/08/18 08:49:58] ppocr INFO: precision:1.0
[2022/08/18 08:49:58] ppocr INFO: recall:1.0
[2022/08/18 08:49:58] ppocr INFO: hmean:1.0
[2022/08/18 08:49:58] ppocr INFO: fps:1.9740402401574881
```
#### 4.3.4 模型预测
使用下面的命令进行预测。
```bash
python3 tools/infer_vqa_token_ser.py -c fapiao/ser_vi_layoutxlm.yml -o Architecture.Backbone.checkpoints=fapiao/models/ser_vi_layoutxlm_fapiao_udml/best_accuracy Global.infer_img=./train_data/XFUND/zh_val/val.json Global.infer_mode=False
```
预测结果会保存在配置文件中的`Global.save_res_path`目录中。
部分预测结果如下所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185310636-6ce02f7c-790d-479f-b163-ea97a5a04808.jpg" width="800">
</div>
* 注意：在预测时，使用的文本检测与识别结果为标注的结果，直接从json文件里面进行读取。
如果希望使用OCR引擎结果得到的结果进行推理，则可以使用下面的命令进行推理。
```bash
python3 tools/infer_vqa_token_ser.py -c fapiao/ser_vi_layoutxlm.yml -o Architecture.Backbone.checkpoints=fapiao/models/ser_vi_layoutxlm_fapiao_udml/best_accuracy Global.infer_img=./train_data/zzsfp/imgs/b25.jpg Global.infer_mode=True
```
结果如下所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185384321-61153faa-e407-45c4-8e7c-a39540248189.jpg" width="800">
</div>
它会使用PP-OCRv3的文本检测与识别模型进行获取文本位置与内容信息。
可以看出，由于训练的过程中，没有标注额外的字段为other类别，所以大多数检测出来的字段被预测为question或者answer。
如果希望构建基于你在垂类场景训练得到的OCR检测与识别模型，可以使用下面的方法传入检测与识别的inference 模型路径，即可完成OCR文本检测与识别以及SER的串联过程。
```bash
python3 tools/infer_vqa_token_ser.py -c fapiao/ser_vi_layoutxlm.yml -o Architecture.Backbone.checkpoints=fapiao/models/ser_vi_layoutxlm_fapiao_udml/best_accuracy Global.infer_img=./train_data/zzsfp/imgs/b25.jpg Global.infer_mode=True Global.kie_rec_model_dir="your_rec_model" Global.kie_det_model_dir="your_det_model"
```
### 4.4 关系抽取（Relation Extraction）
使用SER模型，可以获取图像中所有的question与answer的字段，继续这些字段的类别，我们需要进一步获取question与answer之间的连接，因此需要进一步训练关系抽取模型，解决该问题。本文也基于VI-LayoutXLM多模态预训练模型，进行下游RE任务的模型训练。
#### 4.4.1 准备数据
以发票场景为例，相比于SER任务，RE中还需要标记每个文本行的id信息以及链接关系linking，如下所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185387870-dc9125a0-9ceb-4036-abf3-184b6e65dc7d.jpg" width="800">
</div>
标注文件的部分内容如下所示。
```py
b33.jpg [{"transcription": "No", "label": "question", "points": [[2882, 472], [3026, 472], [3026, 588], [2882, 588]], "id": 0, "linking": [[0, 1]]}, {"transcription": "12269563", "label": "answer", "points": [[3066, 448], [3598, 448], [3598, 576], [3066, 576]], "id": 1, "linking": [[0, 1]]}]
```
相比与SER的标注，多了`id`与`linking`的信息，分别表示唯一标识以及连接关系。
已经处理好的增值税发票数据集从这里下载：[增值税发票数据集下载链接](https://aistudio.baidu.com/aistudio/datasetdetail/165561)。
#### 4.4.2 开始训练
基于VI-LayoutXLM的RE任务配置为[re_vi_layoutxlm_xfund_zh_udml.yml](../configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml)，需要修改**数据路径、类别列表文件**。
```yml
Train:
dataset:
name: SimpleDataSet
# 定义训练数据目录与标注文件
data_dir: train_data/zzsfp/imgs
label_file_list:
- train_data/zzsfp/train.json
transforms:
- DecodeImage: # load image
img_mode: RGB
channel_first: False
- VQATokenLabelEncode: # Class handling label
contains_re: True
algorithm: *algorithm
class_path: &class_path train_data/zzsfp/class_list.txt
...
Eval:
dataset:
# 定义评估数据目录与标注文件
name: SimpleDataSet
data_dir: train_data/zzsfp/imgs
label_file_list:
- train_data/zzsfp/val.json
...
```
LayoutXLM与VI-LayoutXLM针对该场景的训练结果如下所示。
| 模型 | 迭代轮数 | Hmean |
| :---: | :---: | :---: |
| LayoutXLM | 50 | 98.0% |
| VI-LayoutXLM | 50 | 99.3% |
可以看出，对于VI-LayoutXLM相比LayoutXLM的Hmean高了1.3%。
#### 4.4.3 模型评估
模型训练过程中，使用的是知识蒸馏的策略，最终保留了学生模型的参数，在评估时，我们需要针对学生模型的配置文件进行修改: [re_vi_layoutxlm_xfund_zh.yml](../configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml)，修改内容与训练配置相同，包括**类别映射文件、数据目录**。
修改完成后，执行下面的命令完成评估过程。
```bash
# 注意：需要根据你的配置文件地址与保存的模型地址，对评估命令进行修改
python3 tools/eval.py -c ./fapiao/re_vi_layoutxlm.yml -o Architecture.Backbone.checkpoints=fapiao/models/re_vi_layoutxlm_fapiao_udml/best_accuracy
```
输出结果如下所示。
```py
[2022/08/18 12:17:14] ppocr INFO: metric eval ***************
[2022/08/18 12:17:14] ppocr INFO: precision:1.0
[2022/08/18 12:17:14] ppocr INFO: recall:0.9873417721518988
[2022/08/18 12:17:14] ppocr INFO: hmean:0.9936305732484078
[2022/08/18 12:17:14] ppocr INFO: fps:2.765963539771157
```
#### 4.4.4 模型预测
使用下面的命令进行预测。
```bash
# -c 后面的是RE任务的配置文件
# -o 后面的字段是RE任务的配置
# -c_ser 后面的是SER任务的配置文件
# -c_ser 后面的字段是SER任务的配置
python3 tools/infer_vqa_token_ser_re.py -c fapiao/re_vi_layoutxlm.yml -o Architecture.Backbone.checkpoints=fapiao/models/re_vi_layoutxlm_fapiao_udml/best_accuracy Global.infer_img=./train_data/zzsfp/val.json Global.infer_mode=False -c_ser fapiao/ser_vi_layoutxlm.yml -o_ser Architecture.Backbone.checkpoints=fapiao/models/ser_vi_layoutxlm_fapiao_udml/best_accuracy
```
预测结果会保存在配置文件中的`Global.save_res_path`目录中。
部分预测结果如下所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185393805-c67ff571-cf7e-4217-a4b0-8b396c4f22bb.jpg" width="800">
</div>
* 注意：在预测时，使用的文本检测与识别结果为标注的结果，直接从json文件里面进行读取。
如果希望使用OCR引擎结果得到的结果进行推理，则可以使用下面的命令进行推理。
```bash
python3 tools/infer_vqa_token_ser_re.py -c fapiao/re_vi_layoutxlm.yml -o Architecture.Backbone.checkpoints=fapiao/models/re_vi_layoutxlm_fapiao_udml/best_accuracy Global.infer_img=./train_data/zzsfp/val.json Global.infer_mode=True -c_ser fapiao/ser_vi_layoutxlm.yml -o_ser Architecture.Backbone.checkpoints=fapiao/models/ser_vi_layoutxlm_fapiao_udml/best_accuracy
```
如果希望构建基于你在垂类场景训练得到的OCR检测与识别模型，可以使用下面的方法传入，即可完成SER + RE的串联过程。
```bash
python3 tools/infer_vqa_token_ser_re.py -c fapiao/re_vi_layoutxlm.yml -o Architecture.Backbone.checkpoints=fapiao/models/re_vi_layoutxlm_fapiao_udml/best_accuracy Global.infer_img=./train_data/zzsfp/val.json Global.infer_mode=True -c_ser fapiao/ser_vi_layoutxlm.yml -o_ser Architecture.Backbone.checkpoints=fapiao/models/ser_vi_layoutxlm_fapiao_udml/best_accuracy Global.kie_rec_model_dir="your_rec_model" Global.kie_det_model_dir="your_det_model"
```
configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_cml.yml
浏览文件 @ 96a8e0f8
...@@ -14,6 +14,9 @@ Global: ...@@ -14,6 +14,9 @@ Global:
use_visualdl: False use_visualdl: False
infer_img: doc/imgs_en/img_10.jpg infer_img: doc/imgs_en/img_10.jpg
save_res_path: ./output/det_db/predicts_db.txt save_res_path: ./output/det_db/predicts_db.txt
use_amp: False
amp_level: O2
amp_custom_black_list: ['exp']
Architecture: Architecture:
name: DistillationModel name: DistillationModel
......
configs/kie/layoutlm_series/re_layoutlmv2_xfund_zh.yml
浏览文件 @ 96a8e0f8
...@@ -11,11 +11,11 @@ Global: ...@@ -11,11 +11,11 @@ Global:
save_inference_dir: save_inference_dir:
use_visualdl: False use_visualdl: False
seed: 2022 seed: 2022
infer_img: ppstructure/docs/vqa/input/zh_val_21.jpg infer_img: ppstructure/docs/kie/input/zh_val_21.jpg
save_res_path: ./output/re_layoutlmv2_xfund_zh/res/ save_res_path: ./output/re_layoutlmv2_xfund_zh/res/
Architecture: Architecture:
model_type: vqa model_type: kie
algorithm: &algorithm "LayoutLMv2" algorithm: &algorithm "LayoutLMv2"
Transform: Transform:
Backbone: Backbone:
......
configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml
浏览文件 @ 96a8e0f8
...@@ -11,11 +11,11 @@ Global: ...@@ -11,11 +11,11 @@ Global:
save_inference_dir: save_inference_dir:
use_visualdl: False use_visualdl: False
seed: 2022 seed: 2022
infer_img: ppstructure/docs/vqa/input/zh_val_21.jpg infer_img: ppstructure/docs/kie/input/zh_val_21.jpg
save_res_path: ./output/re_layoutxlm_xfund_zh/res/ save_res_path: ./output/re_layoutxlm_xfund_zh/res/
Architecture: Architecture:
model_type: vqa model_type: kie
algorithm: &algorithm "LayoutXLM" algorithm: &algorithm "LayoutXLM"
Transform: Transform:
Backbone: Backbone:
......
configs/kie/layoutlm_series/ser_layoutlm_xfund_zh.yml
浏览文件 @ 96a8e0f8
...@@ -11,11 +11,11 @@ Global: ...@@ -11,11 +11,11 @@ Global:
save_inference_dir: save_inference_dir:
use_visualdl: False use_visualdl: False
seed: 2022 seed: 2022
infer_img: ppstructure/docs/vqa/input/zh_val_42.jpg infer_img: ppstructure/docs/kie/input/zh_val_42.jpg
save_res_path: ./output/re_layoutlm_xfund_zh/res save_res_path: ./output/re_layoutlm_xfund_zh/res
Architecture: Architecture:
model_type: vqa model_type: kie
algorithm: &algorithm "LayoutLM" algorithm: &algorithm "LayoutLM"
Transform: Transform:
Backbone: Backbone:
......
configs/kie/layoutlm_series/ser_layoutlmv2_xfund_zh.yml
浏览文件 @ 96a8e0f8
...@@ -11,11 +11,11 @@ Global: ...@@ -11,11 +11,11 @@ Global:
save_inference_dir: save_inference_dir:
use_visualdl: False use_visualdl: False
seed: 2022 seed: 2022
infer_img: ppstructure/docs/vqa/input/zh_val_42.jpg infer_img: ppstructure/docs/kie/input/zh_val_42.jpg
save_res_path: ./output/ser_layoutlmv2_xfund_zh/res/ save_res_path: ./output/ser_layoutlmv2_xfund_zh/res/
Architecture: Architecture:
model_type: vqa model_type: kie
algorithm: &algorithm "LayoutLMv2" algorithm: &algorithm "LayoutLMv2"
Transform: Transform:
Backbone: Backbone:
......
configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml
浏览文件 @ 96a8e0f8
...@@ -11,11 +11,11 @@ Global: ...@@ -11,11 +11,11 @@ Global:
save_inference_dir: save_inference_dir:
use_visualdl: False use_visualdl: False
seed: 2022 seed: 2022
infer_img: ppstructure/docs/vqa/input/zh_val_42.jpg infer_img: ppstructure/docs/kie/input/zh_val_42.jpg
save_res_path: ./output/ser_layoutxlm_xfund_zh/res save_res_path: ./output/ser_layoutxlm_xfund_zh/res
Architecture: Architecture:
model_type: vqa model_type: kie
algorithm: &algorithm "LayoutXLM" algorithm: &algorithm "LayoutXLM"
Transform: Transform:
Backbone: Backbone:
......
configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml
浏览文件 @ 96a8e0f8
...@@ -11,11 +11,13 @@ Global: ...@@ -11,11 +11,13 @@ Global:
save_inference_dir: save_inference_dir:
use_visualdl: False use_visualdl: False
seed: 2022 seed: 2022
infer_img: ppstructure/docs/vqa/input/zh_val_21.jpg infer_img: ppstructure/docs/kie/input/zh_val_21.jpg
save_res_path: ./output/re/xfund_zh/with_gt save_res_path: ./output/re/xfund_zh/with_gt
kie_rec_model_dir:
kie_det_model_dir:
Architecture: Architecture:
model_type: vqa model_type: kie
algorithm: &algorithm "LayoutXLM" algorithm: &algorithm "LayoutXLM"
Transform: Transform:
Backbone: Backbone:
......
configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml
浏览文件 @ 96a8e0f8
...@@ -11,11 +11,11 @@ Global: ...@@ -11,11 +11,11 @@ Global:
save_inference_dir: save_inference_dir:
use_visualdl: False use_visualdl: False
seed: 2022 seed: 2022
infer_img: ppstructure/docs/vqa/input/zh_val_21.jpg infer_img: ppstructure/docs/kie/input/zh_val_21.jpg
save_res_path: ./output/re/xfund_zh/with_gt save_res_path: ./output/re/xfund_zh/with_gt
Architecture: Architecture:
model_type: &model_type "vqa" model_type: &model_type "kie"
name: DistillationModel name: DistillationModel
algorithm: Distillation algorithm: Distillation
Models: Models:
......
configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml
浏览文件 @ 96a8e0f8
...@@ -11,16 +11,18 @@ Global: ...@@ -11,16 +11,18 @@ Global:
save_inference_dir: save_inference_dir:
use_visualdl: False use_visualdl: False
seed: 2022 seed: 2022
infer_img: ppstructure/docs/vqa/input/zh_val_42.jpg infer_img: ppstructure/docs/kie/input/zh_val_42.jpg
# if you want to predict using the groundtruth ocr info, # if you want to predict using the groundtruth ocr info,
# you can use the following config # you can use the following config
# infer_img: train_data/XFUND/zh_val/val.json # infer_img: train_data/XFUND/zh_val/val.json
# infer_mode: False # infer_mode: False
save_res_path: ./output/ser/xfund_zh/res save_res_path: ./output/ser/xfund_zh/res
kie_rec_model_dir:
kie_det_model_dir:
Architecture: Architecture:
model_type: vqa model_type: kie
algorithm: &algorithm "LayoutXLM" algorithm: &algorithm "LayoutXLM"
Transform: Transform:
Backbone: Backbone:
......
configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml
浏览文件 @ 96a8e0f8
...@@ -11,12 +11,12 @@ Global: ...@@ -11,12 +11,12 @@ Global:
save_inference_dir: save_inference_dir:
use_visualdl: False use_visualdl: False
seed: 2022 seed: 2022
infer_img: ppstructure/docs/vqa/input/zh_val_42.jpg infer_img: ppstructure/docs/kie/input/zh_val_42.jpg
save_res_path: ./output/ser_layoutxlm_xfund_zh/res save_res_path: ./output/ser_layoutxlm_xfund_zh/res
Architecture: Architecture:
model_type: &model_type "vqa" model_type: &model_type "kie"
name: DistillationModel name: DistillationModel
algorithm: Distillation algorithm: Distillation
Models: Models:
......
configs/table/SLANet_ch.yml 0 → 100644
浏览文件 @ 96a8e0f8
Global:
use_gpu: True
epoch_num: 400
log_smooth_window: 20
print_batch_step: 20
save_model_dir: ./output/SLANet_ch
save_epoch_step: 400
# evaluation is run every 331 iterations after the 0th iteration
eval_batch_step: [0, 331]
cal_metric_during_train: True
pretrained_model:
checkpoints:
save_inference_dir: ./output/SLANet_ch/infer
use_visualdl: False
infer_img: doc/table/table.jpg
# for data or label process
character_dict_path: ppocr/utils/dict/table_structure_dict_ch.txt
character_type: en
max_text_length: &max_text_length 500
box_format: &box_format xyxyxyxy # 'xywh', 'xyxy', 'xyxyxyxy'
infer_mode: False
use_sync_bn: True
save_res_path: output/infer
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
clip_norm: 5.0
lr:
learning_rate: 0.001
regularizer:
name: 'L2'
factor: 0.00000
Architecture:
model_type: table
algorithm: SLANet
Backbone:
name: PPLCNet
scale: 1.0
pretrained: True
use_ssld: True
Neck:
name: CSPPAN
out_channels: 96
Head:
name: SLAHead
hidden_size: 256
max_text_length: *max_text_length
loc_reg_num: &loc_reg_num 8
Loss:
name: SLALoss
structure_weight: 1.0
loc_weight: 2.0
loc_loss: smooth_l1
PostProcess:
name: TableLabelDecode
merge_no_span_structure: &merge_no_span_structure True
Metric:
name: TableMetric
main_indicator: acc
compute_bbox_metric: False
loc_reg_num: *loc_reg_num
box_format: *box_format
del_thead_tbody: True
Train:
dataset:
name: PubTabDataSet
data_dir: train_data/table/train/
label_file_list: [train_data/table/train.txt]
transforms:
- DecodeImage:
img_mode: BGR
channel_first: False
- TableLabelEncode:
learn_empty_box: False
merge_no_span_structure: *merge_no_span_structure
replace_empty_cell_token: False
loc_reg_num: *loc_reg_num
max_text_length: *max_text_length
- TableBoxEncode:
in_box_format: *box_format
out_box_format: *box_format
- ResizeTableImage:
max_len: 488
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- PaddingTableImage:
size: [488, 488]
- ToCHWImage:
- KeepKeys:
keep_keys: [ 'image', 'structure', 'bboxes', 'bbox_masks', 'shape' ]
loader:
shuffle: True
batch_size_per_card: 48
drop_last: True
num_workers: 1
Eval:
dataset:
name: PubTabDataSet
data_dir: train_data/table/val/
label_file_list: [train_data/table/val.txt]
transforms:
- DecodeImage:
img_mode: BGR
channel_first: False
- TableLabelEncode:
learn_empty_box: False
merge_no_span_structure: *merge_no_span_structure
replace_empty_cell_token: False
loc_reg_num: *loc_reg_num
max_text_length: *max_text_length
- TableBoxEncode:
in_box_format: *box_format
out_box_format: *box_format
- ResizeTableImage:
max_len: 488
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- PaddingTableImage:
size: [488, 488]
- ToCHWImage:
- KeepKeys:
keep_keys: [ 'image', 'structure', 'bboxes', 'bbox_masks', 'shape' ]
loader:
shuffle: False
drop_last: False
batch_size_per_card: 48
num_workers: 1
configs/table/table_mv3.yml
浏览文件 @ 96a8e0f8
...@@ -17,7 +17,7 @@ Global: ...@@ -17,7 +17,7 @@ Global:
# for data or label process # for data or label process
character_dict_path: ppocr/utils/dict/table_structure_dict.txt character_dict_path: ppocr/utils/dict/table_structure_dict.txt
character_type: en character_type: en
max_text_length: &max_text_length 800 max_text_length: &max_text_length 500
box_format: &box_format 'xyxy' # 'xywh', 'xyxy', 'xyxyxyxy' box_format: &box_format 'xyxy' # 'xywh', 'xyxy', 'xyxyxyxy'
infer_mode: False infer_mode: False
...@@ -38,7 +38,8 @@ Architecture: ...@@ -38,7 +38,8 @@ Architecture:
Backbone: Backbone:
name: MobileNetV3 name: MobileNetV3
scale: 1.0 scale: 1.0
model_name: large model_name: small
disable_se: true
Head: Head:
name: TableAttentionHead name: TableAttentionHead
hidden_size: 256 hidden_size: 256
...@@ -89,7 +90,7 @@ Train: ...@@ -89,7 +90,7 @@ Train:
keep_keys: [ 'image', 'structure', 'bboxes', 'bbox_masks', 'shape' ] keep_keys: [ 'image', 'structure', 'bboxes', 'bbox_masks', 'shape' ]
loader: loader:
shuffle: True shuffle: True
batch_size_per_card: 32 batch_size_per_card: 48
drop_last: True drop_last: True
num_workers: 1 num_workers: 1
...@@ -124,5 +125,5 @@ Eval: ...@@ -124,5 +125,5 @@ Eval:
loader: loader:
shuffle: False shuffle: False
drop_last: False drop_last: False
batch_size_per_card: 16 batch_size_per_card: 48
num_workers: 1 num_workers: 1
deploy/android_demo/app/src/main/cpp/native.cpp
浏览文件 @ 96a8e0f8
...@@ -47,7 +47,7 @@ str_to_cpu_mode(const std::string &cpu_mode) { ...@@ -47,7 +47,7 @@ str_to_cpu_mode(const std::string &cpu_mode) {
std::string upper_key; std::string upper_key;
std::transform(cpu_mode.cbegin(), cpu_mode.cend(), upper_key.begin(), std::transform(cpu_mode.cbegin(), cpu_mode.cend(), upper_key.begin(),
::toupper); ::toupper);
auto index = cpu_mode_map.find(upper_key); auto index = cpu_mode_map.find(upper_key.c_str());
if (index == cpu_mode_map.end()) { if (index == cpu_mode_map.end()) {
LOGE("cpu_mode not found %s", upper_key.c_str()); LOGE("cpu_mode not found %s", upper_key.c_str());
return paddle::lite_api::LITE_POWER_HIGH; return paddle::lite_api::LITE_POWER_HIGH;
......
deploy/android_demo/app/src/main/java/com/baidu/paddle/lite/demo/ocr/OCRPredictorNative.java
浏览文件 @ 96a8e0f8
...@@ -54,7 +54,7 @@ public class OCRPredictorNative { ...@@ -54,7 +54,7 @@ public class OCRPredictorNative {
} }
public void destory() { public void destory() {
if (nativePointer > 0) { if (nativePointer != 0) {
release(nativePointer); release(nativePointer);
nativePointer = 0; nativePointer = 0;
} }
......
deploy/cpp_infer/docs/windows_vs2019_build.md
浏览文件 @ 96a8e0f8
...@@ -109,8 +109,10 @@ CUDA_LIB、CUDNN_LIB、TENSORRT_DIR、WITH_GPU、WITH_TENSORRT ...@@ -109,8 +109,10 @@ CUDA_LIB、CUDNN_LIB、TENSORRT_DIR、WITH_GPU、WITH_TENSORRT
运行之前，将下面文件拷贝到`build/Release/`文件夹下 运行之前，将下面文件拷贝到`build/Release/`文件夹下
1. `paddle_inference/paddle/lib/paddle_inference.dll` 1. `paddle_inference/paddle/lib/paddle_inference.dll`
2. `opencv/build/x64/vc15/bin/opencv_world455.dll` 2. `paddle_inference/third_party/install/onnxruntime/lib/onnxruntime.dll`
3. 如果使用openblas版本的预测库还需要拷贝 `paddle_inference/third_party/install/openblas/lib/openblas.dll` 3. `paddle_inference/third_party/install/paddle2onnx/lib/paddle2onnx.dll`
4. `opencv/build/x64/vc15/bin/opencv_world455.dll`
5. 如果使用openblas版本的预测库还需要拷贝 `paddle_inference/third_party/install/openblas/lib/openblas.dll`
### Step4: 预测 ### Step4: 预测
......
deploy/slim/quantization/README_en.md
浏览文件 @ 96a8e0f8
...@@ -73,4 +73,4 @@ python deploy/slim/quantization/export_model.py -c configs/det/ch_ppocr_v2.0/ch_ ...@@ -73,4 +73,4 @@ python deploy/slim/quantization/export_model.py -c configs/det/ch_ppocr_v2.0/ch_
The numerical range of the quantized model parameters derived from the above steps is still FP32, but the numerical range of the parameters is int8. The numerical range of the quantized model parameters derived from the above steps is still FP32, but the numerical range of the parameters is int8.
The derived model can be converted through the `opt tool` of PaddleLite. The derived model can be converted through the `opt tool` of PaddleLite.
For quantitative model deployment, please refer to [Mobile terminal model deployment](../../lite/readme_en.md) For quantitative model deployment, please refer to [Mobile terminal model deployment](../../lite/readme.md)
doc/doc_ch/algorithm.md 已删除 100644 → 0
浏览文件 @ 29268eaa
# 前沿算法与模型
PaddleOCR将**持续新增**支持OCR领域前沿算法与模型，已支持的模型与使用教程可点击下方列表查看：
- [文本检测算法](./algorithm_overview.md#11-%E6%96%87%E6%9C%AC%E6%A3%80%E6%B5%8B%E7%AE%97%E6%B3%95)
- [文本识别算法](./algorithm_overview.md#12-%E6%96%87%E6%9C%AC%E8%AF%86%E5%88%AB%E7%AE%97%E6%B3%95)
- [端到端算法](./algorithm_overview.md#2-%E6%96%87%E6%9C%AC%E8%AF%86%E5%88%AB%E7%AE%97%E6%B3%95)
- [表格识别](./algorithm_overview.md#3-%E8%A1%A8%E6%A0%BC%E8%AF%86%E5%88%AB%E7%AE%97%E6%B3%95)
**欢迎广大开发者合作共建，贡献更多算法，合入有奖🎁！具体可查看[社区常规赛](https://github.com/PaddlePaddle/PaddleOCR/issues/4982)。**
新增算法可参考如下教程：
- [使用PaddleOCR架构添加新算法](./add_new_algorithm.md)
doc/doc_ch/algorithm_kie_layoutxlm.md
浏览文件 @ 96a8e0f8
...@@ -66,10 +66,10 @@ LayoutXLM模型基于SER任务进行推理，可以执行如下命令： ...@@ -66,10 +66,10 @@ LayoutXLM模型基于SER任务进行推理，可以执行如下命令：
```bash ```bash
cd ppstructure cd ppstructure
python3 vqa/predict_vqa_token_ser.py \ python3 kie/predict_kie_token_ser.py \
--vqa_algorithm=LayoutXLM \ --kie_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_layoutxlm_infer \ --ser_model_dir=../inference/ser_layoutxlm_infer \
--image_dir=./docs/vqa/input/zh_val_42.jpg \ --image_dir=./docs/kie/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \ --ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf --vis_font_path=../doc/fonts/simfang.ttf
``` ```
...@@ -77,7 +77,7 @@ python3 vqa/predict_vqa_token_ser.py \ ...@@ -77,7 +77,7 @@ python3 vqa/predict_vqa_token_ser.py \
SER可视化结果默认保存到`./output`文件夹里面，结果示例如下： SER可视化结果默认保存到`./output`文件夹里面，结果示例如下：
<div align="center"> <div align="center">
<img src="../../ppstructure/docs/vqa/result_ser/zh_val_42_ser.jpg" width="800"> <img src="../../ppstructure/docs/kie/result_ser/zh_val_42_ser.jpg" width="800">
</div> </div>
......
doc/doc_ch/algorithm_kie_vi_layoutxlm.md
浏览文件 @ 96a8e0f8
...@@ -59,10 +59,10 @@ VI-LayoutXLM模型基于SER任务进行推理，可以执行如下命令： ...@@ -59,10 +59,10 @@ VI-LayoutXLM模型基于SER任务进行推理，可以执行如下命令：
```bash ```bash
cd ppstructure cd ppstructure
python3 vqa/predict_vqa_token_ser.py \ python3 kie/predict_kie_token_ser.py \
--vqa_algorithm=LayoutXLM \ --kie_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_vi_layoutxlm_infer \ --ser_model_dir=../inference/ser_vi_layoutxlm_infer \
--image_dir=./docs/vqa/input/zh_val_42.jpg \ --image_dir=./docs/kie/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \ --ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf \ --vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx" --ocr_order_method="tb-yx"
...@@ -71,7 +71,7 @@ python3 vqa/predict_vqa_token_ser.py \ ...@@ -71,7 +71,7 @@ python3 vqa/predict_vqa_token_ser.py \
SER可视化结果默认保存到`./output`文件夹里面，结果示例如下： SER可视化结果默认保存到`./output`文件夹里面，结果示例如下：
<div align="center"> <div align="center">
<img src="../../ppstructure/docs/vqa/result_ser/zh_val_42_ser.jpg" width="800"> <img src="../../ppstructure/docs/kie/result_ser/zh_val_42_ser.jpg" width="800">
</div> </div>
......
doc/doc_ch/algorithm_overview.md
浏览文件 @ 96a8e0f8
# 算法汇总 # 前沿算法与模型
- [1. 两阶段OCR算法](#1) - [1. 两阶段OCR算法](#1)
- [1.1 文本检测算法](#11) - [1.1 文本检测算法](#11)
...@@ -7,8 +7,13 @@ ...@@ -7,8 +7,13 @@
- [3. 表格识别算法](#3) - [3. 表格识别算法](#3)
- [4. 关键信息抽取算法](#4) - [4. 关键信息抽取算法](#4)
本文给出了PaddleOCR已支持的OCR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-OCRv3 系列模型下载](./models_list.md)。
>>
PaddleOCR将**持续新增**支持OCR领域前沿算法与模型，**欢迎广大开发者合作共建，贡献更多算法，合入有奖🎁！具体可查看[社区常规赛](https://github.com/PaddlePaddle/PaddleOCR/issues/4982)。**
>>
新增算法可参考教程：[使用PaddleOCR架构添加新算法](./add_new_algorithm.md)
本文给出了PaddleOCR已支持的OCR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)。
<a name="1"></a> <a name="1"></a>
......
doc/doc_ch/algorithm_rec_srn.md
浏览文件 @ 96a8e0f8
...@@ -78,7 +78,7 @@ python3 tools/export_model.py -c configs/rec/rec_r50_fpn_srn.yml -o Global.pretr ...@@ -78,7 +78,7 @@ python3 tools/export_model.py -c configs/rec/rec_r50_fpn_srn.yml -o Global.pretr
SRN文本识别模型推理，可以执行如下命令： SRN文本识别模型推理，可以执行如下命令：
``` ```
python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/en/word_1.png" --rec_model_dir="./inference/rec_srn/" --rec_image_shape="1,64,256" --rec_char_type="ch" --rec_algorithm="SRN" --rec_char_dict_path=./ppocr/utils/ic15_dict.txt --use_space_char=False python3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words/en/word_1.png" --rec_model_dir="./inference/rec_srn/" --rec_image_shape="1,64,256" --rec_algorithm="SRN" --rec_char_dict_path=./ppocr/utils/ic15_dict.txt --use_space_char=False
``` ```
<a name="4-2"></a> <a name="4-2"></a>
......
doc/doc_ch/dataset/kie_datasets.md
浏览文件 @ 96a8e0f8
# 信息抽取数据集 # 关键信息抽取数据集
这里整理了常见的DocVQA数据集，持续更新中，欢迎各位小伙伴贡献数据集～ 这里整理了常见的DocVQA数据集，持续更新中，欢迎各位小伙伴贡献数据集～
- [FUNSD数据集](#funsd) - [FUNSD数据集](#funsd)
- [XFUND数据集](#xfund) - [XFUND数据集](#xfund)
- [wildreceipt数据集](#wildreceipt) - [wildreceipt数据集](#wildreceipt)
......
doc/doc_ch/kie.md
浏览文件 @ 96a8e0f8
...@@ -64,7 +64,7 @@ zh_train_1.jpg [{"transcription": "中国人体器官捐献", "label": "other" ...@@ -64,7 +64,7 @@ zh_train_1.jpg [{"transcription": "中国人体器官捐献", "label": "other"
验证集构建方式与训练集相同。 验证集构建方式与训练集相同。
* 字典文件 **（3）字典文件**
训练集与验证集中的文本行包含标签信息，所有标签的列表存在字典文件中（如`class_list.txt`），字典文件中的每一行表示为一个类别名称。 训练集与验证集中的文本行包含标签信息，所有标签的列表存在字典文件中（如`class_list.txt`），字典文件中的每一行表示为一个类别名称。
...@@ -103,7 +103,7 @@ HEADER ...@@ -103,7 +103,7 @@ HEADER
## 1.3. 数据下载 ## 1.3. 数据下载
如果你没有本地数据集，可以从[XFUND](https://github.com/doc-analysis/XFUND)或者[FUNSD](https://guillaumejaume.github.io/FUNSD/)官网下载数据，然后使用XFUND与FUNSD的处理脚本([XFUND](../../ppstructure/vqa/tools/trans_xfun_data.py), [FUNSD](../../ppstructure/vqa/tools/trans_funsd_label.py))，生成用于PaddleOCR训练的数据格式，并使用公开数据集快速体验关键信息抽取的流程。 如果你没有本地数据集，可以从[XFUND](https://github.com/doc-analysis/XFUND)或者[FUNSD](https://guillaumejaume.github.io/FUNSD/)官网下载数据，然后使用XFUND与FUNSD的处理脚本([XFUND](../../ppstructure/kie/tools/trans_xfun_data.py), [FUNSD](../../ppstructure/kie/tools/trans_funsd_label.py))，生成用于PaddleOCR训练的数据格式，并使用公开数据集快速体验关键信息抽取的流程。
更多关于公开数据集的介绍，请参考[关键信息抽取数据集说明文档](./dataset/kie_datasets.md)。 更多关于公开数据集的介绍，请参考[关键信息抽取数据集说明文档](./dataset/kie_datasets.md)。
...@@ -209,7 +209,7 @@ Architecture: ...@@ -209,7 +209,7 @@ Architecture:
num_classes: &num_classes 7 num_classes: &num_classes 7
PostProcess: PostProcess:
name: VQASerTokenLayoutLMPostProcess name: kieSerTokenLayoutLMPostProcess
# 修改字典文件的路径为你自定义的数据集的字典路径 # 修改字典文件的路径为你自定义的数据集的字典路径
class_path: &class_path train_data/XFUND/class_list_xfun.txt class_path: &class_path train_data/XFUND/class_list_xfun.txt
...@@ -347,25 +347,25 @@ output/ser_vi_layoutxlm_xfund_zh/ ...@@ -347,25 +347,25 @@ output/ser_vi_layoutxlm_xfund_zh/
```bash ```bash
python3 tools/infer_vqa_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.infer_img=./ppstructure/docs/vqa/input/zh_val_42.jpg python3 tools/infer_kie_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.infer_img=./ppstructure/docs/kie/input/zh_val_42.jpg
``` ```
预测图片如下所示，图片会存储在`Global.save_res_path`路径中。 预测图片如下所示，图片会存储在`Global.save_res_path`路径中。
<div align="center"> <div align="center">
<img src="../../ppstructure/docs/vqa/result_ser/zh_val_42_ser.jpg" width="800"> <img src="../../ppstructure/docs/kie/result_ser/zh_val_42_ser.jpg" width="800">
</div> </div>
预测过程中，默认会加载PP-OCRv3的检测识别模型，用于OCR的信息抽取，如果希望加载预先获取的OCR结果，可以使用下面的方式进行预测，指定`Global.infer_img`为标注文件，其中包含图片路径以及OCR信息，同时指定`Global.infer_mode`为False，表示此时不使用OCR预测引擎。 预测过程中，默认会加载PP-OCRv3的检测识别模型，用于OCR的信息抽取，如果希望加载预先获取的OCR结果，可以使用下面的方式进行预测，指定`Global.infer_img`为标注文件，其中包含图片路径以及OCR信息，同时指定`Global.infer_mode`为False，表示此时不使用OCR预测引擎。
```bash ```bash
python3 tools/infer_vqa_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.infer_img=./train_data/XFUND/zh_val/val.json Global.infer_mode=False python3 tools/infer_kie_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.infer_img=./train_data/XFUND/zh_val/val.json Global.infer_mode=False
``` ```
对于上述图片，如果使用标注的OCR结果进行信息抽取，预测结果如下。 对于上述图片，如果使用标注的OCR结果进行信息抽取，预测结果如下。
<div align="center"> <div align="center">
<img src="../../ppstructure/docs/vqa/result_ser_with_gt_ocr/zh_val_42_ser.jpg" width="800"> <img src="../../ppstructure/docs/kie/result_ser_with_gt_ocr/zh_val_42_ser.jpg" width="800">
</div> </div>
可以看出，部分检测框信息更加准确，但是整体信息抽取识别结果基本一致。 可以看出，部分检测框信息更加准确，但是整体信息抽取识别结果基本一致。
...@@ -375,20 +375,26 @@ python3 tools/infer_vqa_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxl ...@@ -375,20 +375,26 @@ python3 tools/infer_vqa_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxl
```bash ```bash
python3 ./tools/infer_vqa_token_ser_re.py -c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_udml_xfund_zh/re_layoutxlm_xfund_zh_v4_udml/best_accuracy/ Global.infer_img=./train_data/XFUND/zh_val/image/ -c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o_ser Architecture.Backbone.checkpoints=pretrain_models/ser_vi_layoutxlm_udml_xfund_zh/best_accuracy/ python3 ./tools/infer_kie_token_ser_re.py \
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_udml_xfund_zh/best_accuracy/ \
Global.infer_img=./train_data/XFUND/zh_val/image/ \
-c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o_ser Architecture.Backbone.checkpoints=pretrain_models/ \
ser_vi_layoutxlm_udml_xfund_zh/best_accuracy/
``` ```
预测结果如下所示。 预测结果如下所示。
<div align="center"> <div align="center">
<img src="../../ppstructure/docs/vqa/result_re/zh_val_42_re.jpg" width="800"> <img src="../../ppstructure/docs/kie/result_re/zh_val_42_re.jpg" width="800">
</div> </div>
如果希望使用标注或者预先获取的OCR信息进行关键信息抽取，同上，可以指定`Global.infer_mode`为False，指定`Global.infer_img`为标注文件。 如果希望使用标注或者预先获取的OCR信息进行关键信息抽取，同上，可以指定`Global.infer_mode`为False，指定`Global.infer_img`为标注文件。
```bash ```bash
python3 ./tools/infer_vqa_token_ser_re.py -c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_udml_xfund_zh/re_layoutxlm_xfund_zh_v4_udml/best_accuracy/ Global.infer_img=./train_data/XFUND/zh_val/val.json Global.infer_mode=False -c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o_ser Architecture.Backbone.checkpoints=pretrain_models/ser_vi_layoutxlm_udml_xfund_zh/best_accuracy/ python3 ./tools/infer_kie_token_ser_re.py -c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_udml_xfund_zh/re_layoutxlm_xfund_zh_v4_udml/best_accuracy/ Global.infer_img=./train_data/XFUND/zh_val/val.json Global.infer_mode=False -c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o_ser Architecture.Backbone.checkpoints=pretrain_models/ser_vi_layoutxlm_udml_xfund_zh/best_accuracy/
``` ```
其中`c_ser`表示SER的配置文件，`o_ser` 后面需要加上待修改的SER模型与配置文件，如预训练权重等。 其中`c_ser`表示SER的配置文件，`o_ser` 后面需要加上待修改的SER模型与配置文件，如预训练权重等。
...@@ -397,7 +403,7 @@ python3 ./tools/infer_vqa_token_ser_re.py -c configs/kie/vi_layoutxlm/re_vi_layo ...@@ -397,7 +403,7 @@ python3 ./tools/infer_vqa_token_ser_re.py -c configs/kie/vi_layoutxlm/re_vi_layo
预测结果如下所示。 预测结果如下所示。
<div align="center"> <div align="center">
<img src="../../ppstructure/docs/vqa/result_re_with_gt_ocr/zh_val_42_re.jpg" width="800"> <img src="../../ppstructure/docs/kie/result_re_with_gt_ocr/zh_val_42_re.jpg" width="800">
</div> </div>
可以看出，直接使用标注的OCR结果的RE预测结果要更加准确一些。 可以看出，直接使用标注的OCR结果的RE预测结果要更加准确一些。
...@@ -417,8 +423,8 @@ inference 模型（`paddle.jit.save`保存的模型） ...@@ -417,8 +423,8 @@ inference 模型（`paddle.jit.save`保存的模型）
```bash ```bash
# -c 后面设置训练算法的yml配置文件 # -c 后面设置训练算法的yml配置文件
# -o 配置可选参数 # -o 配置可选参数
# Global.pretrained_model 参数设置待转换的训练模型地址。 # Architecture.Backbone.checkpoints 参数设置待转换的训练模型地址
# Global.save_inference_dir参数设置转换的模型将保存的地址。 # 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/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.save_inference_dir=./inference/ser_vi_layoutxlm python3 tools/export_model.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.save_inference_dir=./inference/ser_vi_layoutxlm
``` ```
...@@ -440,10 +446,10 @@ VI-LayoutXLM模型基于SER任务进行推理，可以执行如下命令： ...@@ -440,10 +446,10 @@ VI-LayoutXLM模型基于SER任务进行推理，可以执行如下命令：
```bash ```bash
cd ppstructure cd ppstructure
python3 vqa/predict_vqa_token_ser.py \ python3 kie/predict_kie_token_ser.py \
--vqa_algorithm=LayoutXLM \ --kie_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_vi_layoutxlm \ --ser_model_dir=../inference/ser_vi_layoutxlm \
--image_dir=./docs/vqa/input/zh_val_42.jpg \ --image_dir=./docs/kie/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \ --ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf \ --vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx" --ocr_order_method="tb-yx"
...@@ -452,7 +458,7 @@ python3 vqa/predict_vqa_token_ser.py \ ...@@ -452,7 +458,7 @@ python3 vqa/predict_vqa_token_ser.py \
可视化SER结果结果默认保存到`./output`文件夹里面。结果示例如下： 可视化SER结果结果默认保存到`./output`文件夹里面。结果示例如下：
<div align="center"> <div align="center">
<img src="../../ppstructure/docs/vqa/result_ser/zh_val_42_ser.jpg" width="800"> <img src="../../ppstructure/docs/kie/result_ser/zh_val_42_ser.jpg" width="800">
</div> </div>
......
doc/doc_en/algorithm_en.md 已删除 100644 → 0
浏览文件 @ 29268eaa
# Academic Algorithms and Models
PaddleOCR will add cutting-edge OCR algorithms and models continuously. Check out the supported models and tutorials by clicking the following list:
- [text detection algorithms](./algorithm_overview_en.md#11)
- [text recognition algorithms](./algorithm_overview_en.md#12)
- [end-to-end algorithms](./algorithm_overview_en.md#2)
- [table recognition algorithms](./algorithm_overview_en.md#3)
Developers are welcome to contribute more algorithms! Please refer to [add new algorithm](./add_new_algorithm_en.md) guideline.
doc/doc_en/algorithm_kie_layoutxlm_en.md 0 → 100644
浏览文件 @ 96a8e0f8
# KIE Algorithm - LayoutXLM
- [1. Introduction](#1-introduction)
- [2. Environment](#2-environment)
- [3. Model Training / Evaluation / Prediction](#3-model-training--evaluation--prediction)
- [4. Inference and Deployment](#4-inference-and-deployment)
- [4.1 Python Inference](#41-python-inference)
- [4.2 C++ Inference](#42-c-inference)
- [4.3 Serving](#43-serving)
- [4.4 More](#44-more)
- [5. FAQ](#5-faq)
- [Citation](#Citation)
## 1. Introduction
Paper:
> [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836)
>
> Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei
>
> 2021
On XFUND_zh dataset, the algorithm reproduction Hmean is as follows.
|Model|Backbone|Task |Cnnfig|Hmean|Download link|
| --- | --- |--|--- | --- | --- |
|LayoutXLM|LayoutXLM-base|SER |[ser_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml)|90.38%|[trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)/[inference model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh_infer.tar)|
|LayoutXLM|LayoutXLM-base|RE | [re_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml)|74.83%|[trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar)/[inference model(coming soon)]()|
## 2. Environment
Please refer to ["Environment Preparation"](./environment_en.md) to configure the PaddleOCR environment, and refer to ["Project Clone"](./clone_en.md) to clone the project code.
## 3. Model Training / Evaluation / Prediction
Please refer to [KIE tutorial](./kie_en.md)。PaddleOCR has modularized the code structure, so that you only need to **replace the configuration file** to train different models.
## 4. Inference and Deployment
### 4.1 Python Inference
**Note:** Currently, the RE model inference process is still in the process of adaptation. We take SER model as an example to introduce the KIE process based on LayoutXLM model.
First, we need to export the trained model into inference model. Take LayoutXLM model trained on XFUND_zh as an example ([trained model download link](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)). Use the following command to export.
``` bash
wget https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar
tar -xf ser_LayoutXLM_xfun_zh.tar
python3 tools/export_model.py -c configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./ser_LayoutXLM_xfun_zh/best_accuracy Global.save_inference_dir=./inference/ser_layoutxlm
```
Use the following command to infer using LayoutXLM SER model.
```bash
cd ppstructure
python3 kie/predict_kie_token_ser.py \
--kie_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_layoutxlm_infer \
--image_dir=./docs/kie/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf
```
The SER visualization results are saved in the `./output` directory by default. The results are as follows.
<div align="center">
<img src="../../ppstructure/docs/kie/result_ser/zh_val_42_ser.jpg" width="800">
</div>
### 4.2 C++ Inference
Not supported
### 4.3 Serving
Not supported
### 4.4 More
Not supported
## 5. FAQ
## Citation
```bibtex
@article{DBLP:journals/corr/abs-2104-08836,
author = {Yiheng Xu and
Tengchao Lv and
Lei Cui and
Guoxin Wang and
Yijuan Lu and
Dinei Flor{\^{e}}ncio and
Cha Zhang and
Furu Wei},
title = {LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich
Document Understanding},
journal = {CoRR},
volume = {abs/2104.08836},
year = {2021},
url = {https://arxiv.org/abs/2104.08836},
eprinttype = {arXiv},
eprint = {2104.08836},
timestamp = {Thu, 14 Oct 2021 09:17:23 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2104-08836.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@article{DBLP:journals/corr/abs-1912-13318,
author = {Yiheng Xu and
Minghao Li and
Lei Cui and
Shaohan Huang and
Furu Wei and
Ming Zhou},
title = {LayoutLM: Pre-training of Text and Layout for Document Image Understanding},
journal = {CoRR},
volume = {abs/1912.13318},
year = {2019},
url = {http://arxiv.org/abs/1912.13318},
eprinttype = {arXiv},
eprint = {1912.13318},
timestamp = {Mon, 01 Jun 2020 16:20:46 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-1912-13318.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@article{DBLP:journals/corr/abs-2012-14740,
author = {Yang Xu and
Yiheng Xu and
Tengchao Lv and
Lei Cui and
Furu Wei and
Guoxin Wang and
Yijuan Lu and
Dinei A. F. Flor{\^{e}}ncio and
Cha Zhang and
Wanxiang Che and
Min Zhang and
Lidong Zhou},
title = {LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding},
journal = {CoRR},
volume = {abs/2012.14740},
year = {2020},
url = {https://arxiv.org/abs/2012.14740},
eprinttype = {arXiv},
eprint = {2012.14740},
timestamp = {Tue, 27 Jul 2021 09:53:52 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2012-14740.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
```
doc/doc_en/algorithm_kie_sdmgr_en.md 0 → 100644
浏览文件 @ 96a8e0f8
# KIE Algorithm - SDMGR
- [1. Introduction](#1-introduction)
- [2. Environment](#2-environment)
- [3. Model Training / Evaluation / Prediction](#3-model-training--evaluation--prediction)
- [4. Inference and Deployment](#4-inference-and-deployment)
- [4.1 Python Inference](#41-python-inference)
- [4.2 C++ Inference](#42-c-inference)
- [4.3 Serving](#43-serving)
- [4.4 More](#44-more)
- [5. FAQ](#5-faq)
- [Citation](#Citation)
## 1. Introduction
Paper:
> [Spatial Dual-Modality Graph Reasoning for Key Information Extraction](https://arxiv.org/abs/2103.14470)
>
> Hongbin Sun and Zhanghui Kuang and Xiaoyu Yue and Chenhao Lin and Wayne Zhang
>
> 2021
On wildreceipt dataset, the algorithm reproduction Hmean is as follows.
|Model|Backbone |Cnnfig|Hmean|Download link|
| --- | --- | --- | --- | --- |
|SDMGR|VGG6|[configs/kie/sdmgr/kie_unet_sdmgr.yml](../../configs/kie/sdmgr/kie_unet_sdmgr.yml)|86.7%|[trained model]( https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)/[inference model(coming soon)]()|
## 2. 环境配置
Please refer to ["Environment Preparation"](./environment_en.md) to configure the PaddleOCR environment, and refer to ["Project Clone"](./clone_en.md) to clone the project code.
## 3. Model Training / Evaluation / Prediction
SDMGR is a key information extraction algorithm that classifies each detected textline into predefined categories, such as order ID, invoice number, amount, etc.
The training and test data are collected in the wildreceipt dataset, use following command to downloaded the dataset.
```bash
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/wildreceipt.tar && tar xf wildreceipt.tar
```
Create dataset soft link to `PaddleOCR/train_data` directory.
```bash
cd PaddleOCR/ && mkdir train_data && cd train_data
ln -s ../../wildreceipt ./
```
### 3.1 Model training
The config file is `configs/kie/sdmgr/kie_unet_sdmgr.yml`， the default dataset path is `train_data/wildreceipt`.
Use the following command to train the model.
```bash
python3 tools/train.py -c configs/kie/sdmgr/kie_unet_sdmgr.yml -o Global.save_model_dir=./output/kie/
```
### 3.2 Model evaluation
Use the following command to evaluate the model.
```bash
python3 tools/eval.py -c configs/kie/sdmgr/kie_unet_sdmgr.yml -o Global.checkpoints=./output/kie/best_accuracy
```
An example of output information is shown below.
```py
[2022/08/10 05:22:23] ppocr INFO: metric eval ***************
[2022/08/10 05:22:23] ppocr INFO: hmean:0.8670120239257812
[2022/08/10 05:22:23] ppocr INFO: fps:10.18816520530961
```
### 3.3 Model prediction
Use the following command to load the model and predict. During the prediction, the text file storing the image path and OCR information needs to be loaded in advance. Use `Global.infer_img` to assign.
```bash
python3 tools/infer_kie.py -c configs/kie/kie_unet_sdmgr.yml -o Global.checkpoints=kie_vgg16/best_accuracy Global.infer_img=./train_data/wildreceipt/1.txt
```
The visualization results and texts are saved in the `./output/sdmgr_kie/` directory by default. The results are as follows.
<div align="center">
<img src="../../ppstructure/docs/imgs/sdmgr_result.png" width="800">
</div>
## 4. Inference and Deployment
### 4.1 Python Inference
Not supported
### 4.2 C++ Inference
Not supported
### 4.3 Serving
Not supported
### 4.4 More
Not supported
## 5. FAQ
## Citation
```bibtex
@misc{sun2021spatial,
title={Spatial Dual-Modality Graph Reasoning for Key Information Extraction},
author={Hongbin Sun and Zhanghui Kuang and Xiaoyu Yue and Chenhao Lin and Wayne Zhang},
year={2021},
eprint={2103.14470},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
doc/doc_en/algorithm_kie_vi_layoutxlm_en.md 0 → 100644
浏览文件 @ 96a8e0f8
# KIE Algorithm - VI-LayoutXLM
- [1. Introduction](#1-introduction)
- [2. Environment](#2-environment)
- [3. Model Training / Evaluation / Prediction](#3-model-training--evaluation--prediction)
- [4. Inference and Deployment](#4-inference-and-deployment)
- [4.1 Python Inference](#41-python-inference)
- [4.2 C++ Inference](#42-c-inference)
- [4.3 Serving](#43-serving)
- [4.4 More](#44-more)
- [5. FAQ](#5-faq)
- [Citation](#Citation)
## 1. Introduction
VI-LayoutXLM is improved based on LayoutXLM. In the process of downstream finetuning, the visual backbone network module is removed, and the model infernce speed is further improved on the basis of almost lossless accuracy.
On XFUND_zh dataset, the algorithm reproduction Hmean is as follows.
|Model|Backbone|Task |Cnnfig|Hmean|Download link|
| --- | --- |---| --- | --- | --- |
|VI-LayoutXLM |VI-LayoutXLM-base | SER |[ser_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml)|93.19%|[trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)/[inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_infer.tar)|
|VI-LayoutXLM |VI-LayoutXLM-base |RE | [re_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml)|83.92%|[trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar)/[inference model(coming soon)]()|
Please refer to ["Environment Preparation"](./environment_en.md) to configure the PaddleOCR environment, and refer to ["Project Clone"](./clone_en.md) to clone the project code.
## 3. Model Training / Evaluation / Prediction
Please refer to [KIE tutorial](./kie_en.md)。PaddleOCR has modularized the code structure, so that you only need to **replace the configuration file** to train different models.
## 4. Inference and Deployment
### 4.1 Python Inference
**Note:** Currently, the RE model inference process is still in the process of adaptation. We take SER model as an example to introduce the KIE process based on VI-LayoutXLM model.
First, we need to export the trained model into inference model. Take VI-LayoutXLM model trained on XFUND_zh as an example ([trained model download link](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)). Use the following command to export.
``` bash
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar
tar -xf ser_vi_layoutxlm_xfund_pretrained.tar
python3 tools/export_model.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./ser_vi_layoutxlm_xfund_pretrained/best_accuracy Global.save_inference_dir=./inference/ser_vi_layoutxlm_infer
```
Use the following command to infer using VI-LayoutXLM SER model.
```bash
cd ppstructure
python3 kie/predict_kie_token_ser.py \
--kie_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_vi_layoutxlm_infer \
--image_dir=./docs/kie/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx"
```
The SER visualization results are saved in the `./output` folder by default. The results are as follows.
<div align="center">
<img src="../../ppstructure/docs/kie/result_ser/zh_val_42_ser.jpg" width="800">
</div>
### 4.2 C++ Inference
Not supported
### 4.3 Serving
Not supported
### 4.4 More
Not supported
## 5. FAQ
## Citation
```bibtex
@article{DBLP:journals/corr/abs-2104-08836,
author = {Yiheng Xu and
Tengchao Lv and
Lei Cui and
Guoxin Wang and
Yijuan Lu and
Dinei Flor{\^{e}}ncio and
Cha Zhang and
Furu Wei},
title = {LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich
Document Understanding},
journal = {CoRR},
volume = {abs/2104.08836},
year = {2021},
url = {https://arxiv.org/abs/2104.08836},
eprinttype = {arXiv},
eprint = {2104.08836},
timestamp = {Thu, 14 Oct 2021 09:17:23 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2104-08836.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@article{DBLP:journals/corr/abs-1912-13318,
author = {Yiheng Xu and
Minghao Li and
Lei Cui and
Shaohan Huang and
Furu Wei and
Ming Zhou},
title = {LayoutLM: Pre-training of Text and Layout for Document Image Understanding},
journal = {CoRR},
volume = {abs/1912.13318},
year = {2019},
url = {http://arxiv.org/abs/1912.13318},
eprinttype = {arXiv},
eprint = {1912.13318},
timestamp = {Mon, 01 Jun 2020 16:20:46 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-1912-13318.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
@article{DBLP:journals/corr/abs-2012-14740,
author = {Yang Xu and
Yiheng Xu and
Tengchao Lv and
Lei Cui and
Furu Wei and
Guoxin Wang and
Yijuan Lu and
Dinei A. F. Flor{\^{e}}ncio and
Cha Zhang and
Wanxiang Che and
Min Zhang and
Lidong Zhou},
title = {LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding},
journal = {CoRR},
volume = {abs/2012.14740},
year = {2020},
url = {https://arxiv.org/abs/2012.14740},
eprinttype = {arXiv},
eprint = {2012.14740},
timestamp = {Tue, 27 Jul 2021 09:53:52 +0200},
biburl = {https://dblp.org/rec/journals/corr/abs-2012-14740.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}
```
doc/doc_en/algorithm_overview_en.md
浏览文件 @ 96a8e0f8
# OCR Algorithms # Algorithms
- [1. Two-stage Algorithms](#1) - [1. Two-stage OCR Algorithms](#1)
- [1.1 Text Detection Algorithms](#11) - [1.1 Text Detection Algorithms](#11)
- [1.2 Text Recognition Algorithms](#12) - [1.2 Text Recognition Algorithms](#12)
- [2. End-to-end Algorithms](#2) - [2. End-to-end OCR Algorithms](#2)
- [3. Table Recognition Algorithms](#3) - [3. Table Recognition Algorithms](#3)
- [4. Key Information Extraction Algorithms](#4)
This tutorial lists the OCR algorithms supported by PaddleOCR, as well as the models and metrics of each algorithm on **English public datasets**. It is mainly used for algorithm introduction and algorithm performance comparison. For more models on other datasets including Chinese, please refer to [PP-OCRv3 models list](./models_list_en.md).
>>
Developers are welcome to contribute more algorithms! Please refer to [add new algorithm](./add_new_algorithm_en.md) guideline.
This tutorial lists the OCR algorithms supported by PaddleOCR, as well as the models and metrics of each algorithm on **English public datasets**. It is mainly used for algorithm introduction and algorithm performance comparison. For more models on other datasets including Chinese, please refer to [PP-OCR v2.0 models list](./models_list_en.md).
<a name="1"></a> <a name="1"></a>
## 1. Two-stage Algorithms ## 1. Two-stage OCR Algorithms
<a name="11"></a> <a name="11"></a>
...@@ -98,11 +102,12 @@ Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation r ...@@ -98,11 +102,12 @@ Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation r
<a name="2"></a> <a name="2"></a>
## 2. End-to-end Algorithms ## 2. End-to-end OCR Algorithms
Supported end-to-end algorithms (Click the link to get the tutorial): Supported end-to-end algorithms (Click the link to get the tutorial):
- [x] [PGNet](./algorithm_e2e_pgnet_en.md) - [x] [PGNet](./algorithm_e2e_pgnet_en.md)
<a name="3"></a> <a name="3"></a>
## 3. Table Recognition Algorithms ## 3. Table Recognition Algorithms
...@@ -114,3 +119,34 @@ On the PubTabNet dataset, the algorithm result is as follows: ...@@ -114,3 +119,34 @@ On the PubTabNet dataset, the algorithm result is as follows:
|Model|Backbone|Config|Acc|Download link| |Model|Backbone|Config|Acc|Download link|
|---|---|---|---|---| |---|---|---|---|---|
|TableMaster|TableResNetExtra|[configs/table/table_master.yml](../../configs/table/table_master.yml)|77.47%|[trained](https://paddleocr.bj.bcebos.com/ppstructure/models/tablemaster/table_structure_tablemaster_train.tar) / [inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/tablemaster/table_structure_tablemaster_infer.tar)| |TableMaster|TableResNetExtra|[configs/table/table_master.yml](../../configs/table/table_master.yml)|77.47%|[trained](https://paddleocr.bj.bcebos.com/ppstructure/models/tablemaster/table_structure_tablemaster_train.tar) / [inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/tablemaster/table_structure_tablemaster_infer.tar)|
<a name="4"></a>
## 4. Key Information Extraction Algorithms
Supported KIE algorithms (Click the link to get the tutorial):
- [x] [VI-LayoutXLM](./algorithm_kie_vi_laoutxlm_en.md)
- [x] [LayoutLM](./algorithm_kie_laoutxlm_en.md)
- [x] [LayoutLMv2](./algorithm_kie_laoutxlm_en.md)
- [x] [LayoutXLM](./algorithm_kie_laoutxlm_en.md)
- [x] [SDMGR](./algorithm_kie_sdmgr_en.md)
On wildreceipt dataset, the algorithm result is as follows:
|Model|Backbone|Config|Hmean|Download link|
| --- | --- | --- | --- | --- |
|SDMGR|VGG6|[configs/kie/sdmgr/kie_unet_sdmgr.yml](../../configs/kie/sdmgr/kie_unet_sdmgr.yml)|86.7%|[trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)|
On XFUND_zh dataset, the algorithm result is as follows:
|Model|Backbone|Task|Config|Hmean|Download link|
| --- | --- | --- | --- | --- | --- |
|VI-LayoutXLM| VI-LayoutXLM-base | SER | [ser_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml)|**93.19%**|[trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)|
|LayoutXLM| LayoutXLM-base | SER | [ser_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml)|90.38%|[trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)|
|LayoutLM| LayoutLM-base | SER | [ser_layoutlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutlm_xfund_zh.yml)|77.31%|[trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar)|
|LayoutLMv2| LayoutLMv2-base | SER | [ser_layoutlmv2_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutlmv2_xfund_zh.yml)|85.44%|[trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar)|
|VI-LayoutXLM| VI-LayoutXLM-base | RE | [re_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml)|**83.92%**|[trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar)|
|LayoutXLM| LayoutXLM-base | RE | [re_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml)|74.83%|[trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar)|
|LayoutLMv2| LayoutLMv2-base | RE | [re_layoutlmv2_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutlmv2_xfund_zh.yml)|67.77%|[trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar)|
doc/doc_en/dataset/docvqa_datasets_en.md → doc/doc_en/dataset/kie_datasets_en.md
浏览文件 @ 96a8e0f8
## DocVQA dataset ## Key Imnformation Extraction dataset
Here are the common DocVQA datasets, which are being updated continuously. Welcome to contribute datasets~
Here are the common DocVQA datasets, which are being updated continuously. Welcome to contribute datasets.
- [FUNSD dataset](#funsd) - [FUNSD dataset](#funsd)
- [XFUND dataset](#xfund) - [XFUND dataset](#xfund)
- [wildreceipt dataset](#wildreceipt数据集)
<a name="funsd"></a> <a name="funsd"></a>
#### 1. FUNSD dataset #### 1. FUNSD dataset
...@@ -25,3 +27,21 @@ Here are the common DocVQA datasets, which are being updated continuously. Welco ...@@ -25,3 +27,21 @@ Here are the common DocVQA datasets, which are being updated continuously. Welco
</div> </div>
- **Download address**: https://github.com/doc-analysis/XFUND/releases/tag/v1.0 - **Download address**: https://github.com/doc-analysis/XFUND/releases/tag/v1.0
<a name="wildreceipt"></a>
## 3. wildreceipt dataset
- **Data source**: https://arxiv.org/abs/2103.14470
- **Data introduction**: XFUND is an English receipt dataset, which contains 26 different categories. There are 1267 training images and 472 evaluation images, in which 50,000 textlines and boxes are annotated. Part of the image and the annotation box visualization are shown below.
<div align="center">
<img src="../../datasets/wildreceipt_demo/2769.jpeg" width="500">
<img src="../../datasets/wildreceipt_demo/1bbe854b8817dedb8585e0732089fd1f752d2cec.jpeg" width="500">
</div>
**Note：** Boxes with category `Ignore` or `Others` are not visualized here.
- **Download address**：
- Offical dataset: [link](https://download.openmmlab.com/mmocr/data/wildreceipt.tar)
- Dataset converted for PaddleOCR training process: [link](https://paddleocr.bj.bcebos.com/ppstructure/dataset/wildreceipt.tar)
doc/doc_en/kie_en.md 0 → 100644
浏览文件 @ 96a8e0f8
# Key Information Extraction
This tutorial provides a guide to the whole process of key information extraction using PaddleOCR, including data preparation, model training, optimization, evaluation, prediction of semantic entity recognition (SER) and relationship extraction (RE) tasks.
- [1. Data Preparation](#Data-Preparation)
- [1.1. Prepare for dataset](#11-Prepare-for-dataset)
- [1.2. Custom Dataset](#12-Custom-Dataset)
- [1.3. Download data](#13-Download-data)
- [2. Training](#2-Training)
- [2.1. Start Training](#21-start-training)
- [2.2. Resume Training](#22-Resume-Training)
- [2.3. Mixed Precision Training](#23-Mixed-Precision-Training)
- [2.4. Distributed Training](#24-Distributed-Training)
- [2.5. Train using knowledge distillation](#25-Train-using-knowledge-distillation)
- [2.6. Training on other platform](#26-Training-on-other-platform)
- [3. Evaluation and Test](#3-Evaluation-and-Test)
- [3.1. Evaluation](#31-指标评估)
- [3.2. Test](#32-Test)
- [4. Model inference](#4-Model-inference)
- [5. FAQ](#5-faq)
# 1. Data Preparation
## 1.1. Prepare for dataset
PaddleOCR supports the following data format when training KIE models.
- `general data` is used to train a dataset whose annotation is stored in a text file (SimpleDataset).
The default storage path of training data is `PaddleOCR/train_data`. If you already have datasets on your disk, you only need to create a soft link to the dataset directory.
```
# linux and mac os
ln -sf <path/to/dataset> <path/to/paddle_ocr>/train_data/dataset
# windows
mklink /d <path/to/paddle_ocr>/train_data/dataset <path/to/dataset>
```
## 1.2. Custom Dataset
The training process generally includes the training set and the evaluation set. The data formats of the two sets are same.
**(1) Training set**
It is recommended to put the training images into the same folder, record the path and annotation of images in a text file. The contents of the text file are as follows:
```py
" image path annotation information "
zh_train_0.jpg [{"transcription": "汇丰晋信", "label": "other", "points": [[104, 114], [530, 114], [530, 175], [104, 175]], "id": 1, "linking": []}, {"transcription": "受理时间:", "label": "question", "points": [[126, 267], [266, 267], [266, 305], [126, 305]], "id": 7, "linking": [[7, 13]]}, {"transcription": "2020.6.15", "label": "answer", "points": [[321, 239], [537, 239], [537, 285], [321, 285]], "id": 13, "linking": [[7, 13]]}]
zh_train_1.jpg [{"transcription": "中国人体器官捐献", "label": "other", "points": [[544, 459], [954, 459], [954, 517], [544, 517]], "id": 1, "linking": []}, {"transcription": ">编号:MC545715483585", "label": "other", "points": [[1462, 470], [2054, 470], [2054, 543], [1462, 543]], "id": 10, "linking": []}, {"transcription": "CHINAORGANDONATION", "label": "other", "points": [[543, 516], [958, 516], [958, 551], [543, 551]], "id": 14, "linking": []}, {"transcription": "中国人体器官捐献志愿登记表", "label": "header", "points": [[635, 793], [1892, 793], [1892, 904], [635, 904]], "id": 18, "linking": []}]
...
```
**Note:** In the text file, please split the image path and annotation with `\t`. Otherwise, error will happen when training.
The annotation can be parsed by `json` into a list of sub-annotations. Each element in the list is a dict, which stores the required information of each text line. The required fields are as follows.
- transcription: stores the text content of the text line
- label: the category of the text line content
- points: stores the four point position information of the text line
- id: stores the ID information of the text line for RE model training
- linking: stores the connection information between text lines for RE model training
**(2) Evaluation set**
The evaluation set is constructed in the same way as the training set.
**(3) Dictionary file**
The textlines in the training set and the evaluation set contain label information. The list of all labels is stored in the dictionary file (such as `class_list.txt`). Each line in the dictionary file is represented as a label name.
For example, FUND_zh data contains four categories. The contents of the dictionary file are as follows.
```
OTHER
QUESTION
ANSWER
HEADER
```
In the annotation file, the annotation information of the `label` field of the text line content of each annotation needs to belong to the dictionary content.
The final dataset shall have the following file structure.
```
|-train_data
|-data_name
|- train.json
|- train
|- zh_train_0.png
|- zh_train_1.jpg
| ...
|- val.json
|- val
|- zh_val_0.png
|- zh_val_1.jpg
| ...
```
**Note:**
-The category information in the annotation file is not case sensitive. For example, 'HEADER' and 'header' will be seen as the same category ID.
- In the dictionary file, it is recommended to put the `other` category (other textlines that need not be paid attention to can be labeled as `other`) on the first line. When parsing, the category ID of the 'other' category will be resolved to 0, and the textlines predicted as `other` will not be visualized later.
## 1.3. Download data
If you do not have local dataset, you can donwload the source files of [XFUND](https://github.com/doc-analysis/XFUND) or [FUNSD](https://guillaumejaume.github.io/FUNSD) and use the scripts of [XFUND](../../ppstructure/kie/tools/trans_xfun_data.py) or [FUNSD](../../ppstructure/kie/tools/trans_funsd_label.py) for tranform them into PaddleOCR format. Then you can use the public dataset to quick experience KIE.
For more information about public KIE datasets, please refer to [KIE dataset tutorial](./dataset/kie_datasets_en.md).
PaddleOCR also supports the annotation of KIE models. Please refer to [PPOCRLabel tutorial](../../PPOCRLabel/README.md).
# 2. Training
PaddleOCR provides training scripts, evaluation scripts and inference scripts. We will introduce based on VI-LayoutXLM model in this section.
This section will take the VI layoutxlm multimodal pre training model as an example to explain.
> If you want to use the SDMGR based KIE algorithm, please refer to: [SDMGR tutorial](./algorithm_kie_sdmgr_en.md).
## 2.1. Start Training
If you do not use a custom dataset, you can use XFUND_zh that has been processed in PaddleOCR dataset for quick experience.
```bash
mkdir train_data
cd train_data
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar && tar -xf XFUND.tar
cd ..
```
If you don't want to train, and want to directly experience the process of model evaluation, prediction, and inference, you can download the training model provided in PaddleOCR and skip section 2.1.
Use the following command to download the trained model.
```bash
mkdir pretrained_model
cd pretrained_model
# download and uncompress SER model
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar & tar -xf ser_vi_layoutxlm_xfund_pretrained.tar
# download and uncompress RE model
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar & tar -xf re_vi_layoutxlm_xfund_pretrained.tar
```
Start training:
- If your paddlepaddle version is `CPU`, you need to set `Global.use_gpu=False` in your config file.
- During training, PaddleOCR will download the VI-LayoutXLM pretraining model by default. There is no need to download it in advance.
```bash
# GPU training, support single card and multi-cards
# The training log will be save in "{Global.save_model_dir}/train.log"
# train SER model using single card
python3 tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml
# train SER model using multi-cards, you can use --gpus to assign the GPU ids.
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml
# train RE model using single card
python3 tools/train.py -c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml
```
Take the SER model training as an example. After the training is started, you will see the following log output.
```
[2022/08/08 16:28:28] ppocr INFO: epoch: [1/200], global_step: 10, lr: 0.000006, loss: 1.871535, avg_reader_cost: 0.28200 s, avg_batch_cost: 0.82318 s, avg_samples: 8.0, ips: 9.71838 samples/s, eta: 0:51:59
[2022/08/08 16:28:33] ppocr INFO: epoch: [1/200], global_step: 19, lr: 0.000018, loss: 1.461939, avg_reader_cost: 0.00042 s, avg_batch_cost: 0.32037 s, avg_samples: 6.9, ips: 21.53773 samples/s, eta: 0:37:55
[2022/08/08 16:28:39] ppocr INFO: cur metric, precision: 0.11526348939743859, recall: 0.19776657060518732, hmean: 0.14564265817747712, fps: 34.008392345050055
[2022/08/08 16:28:45] ppocr INFO: save best model is to ./output/ser_vi_layoutxlm_xfund_zh/best_accuracy
[2022/08/08 16:28:45] ppocr INFO: best metric, hmean: 0.14564265817747712, precision: 0.11526348939743859, recall: 0.19776657060518732, fps: 34.008392345050055, best_epoch: 1
[2022/08/08 16:28:51] ppocr INFO: save model in ./output/ser_vi_layoutxlm_xfund_zh/latest
```
The following information will be automatically printed.
|Field | meaning|
| :----: | :------: |
|epoch | current iteration round|
|iter | current iteration times|
|lr | current learning rate|
|loss | current loss function|
| reader_cost | current batch data processing time|
| batch_ Cost | total current batch time|
|samples | number of samples in the current batch|
|ips | number of samples processed per second|
PaddleOCR supports evaluation during training. you can modify `eval_batch_step` in the config file `configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml` (default as 19 iters). Trained model with best hmean will be saved as `output/ser_vi_layoutxlm_xfund_zh/best_accuracy/`.
If the evaluation dataset is very large, it's recommended to enlarge the eval interval or evaluate the model after training.
**Note:** for more KIE models training and configuration files, you can go into `configs/kie/` or refer to [Frontier KIE algorithms](./algorithm_overview_en.md).
If you want to train model on your own dataset, you need to modify the data path, dictionary file and category number in the configuration file.
Take `configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml` as an example, contents we need to fix is as follows.
```yaml
Architecture:
# ...
Backbone:
name: LayoutXLMForSer
pretrained: True
mode: vi
# Assuming that n categroies are included in the dictionary file (other is included), the the num_classes is set as 2n-1
num_classes: &num_classes 7
PostProcess:
name: kieSerTokenLayoutLMPostProcess
# Modify the dictionary file path for your custom dataset
class_path: &class_path train_data/XFUND/class_list_xfun.txt
Train:
dataset:
name: SimpleDataSet
# Modify the data path for your training dataset
data_dir: train_data/XFUND/zh_train/image
# Modify the data annotation path for your training dataset
label_file_list:
- train_data/XFUND/zh_train/train.json
...
loader:
# batch size for single card when training
batch_size_per_card: 8
...
Eval:
dataset:
name: SimpleDataSet
# Modify the data path for your evaluation dataset
data_dir: train_data/XFUND/zh_val/image
# Modify the data annotation path for your evaluation dataset
label_file_list:
- train_data/XFUND/zh_val/val.json
...
loader:
# batch size for single card when evaluation
batch_size_per_card: 8
```
**Note that the configuration file for prediction/evaluation must be consistent with the training file.**
## 2.2. Resume Training
If the training process is interrupted and you want to load the saved model to resume training, you can specify the path of the model to be loaded by specifying `Architecture.Backbone.checkpoints`.
```bash
python3 tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy
```
**Note:**
- Priority of `Architecture.Backbone.checkpoints` is higher than` Architecture.Backbone.pretrained`. You need to set `Architecture.Backbone.checkpoints` for model finetuning, resume and evalution. If you want to train with the NLP pretrained model, you need to set `Architecture.Backbone.pretrained` as `True` and set `Architecture.Backbone.checkpoints` as null (`null`).
- PaddleNLP pretrained models are used here for LayoutXLM series models, the model loading and saving logic is same as those in PaddleNLP. Therefore we do not need to set `Global.pretrained_model` or `Global.checkpoints` here.
- If you use knowledge distillation to train the LayoutXLM series models, resuming training is not supported now.
## 2.3. Mixed Precision Training
coming soon!
## 2.4. Distributed Training
During multi-machine multi-gpu training, use the `--ips` parameter to set the used machine IP address, and the `--gpus` parameter to set the used GPU ID:
```bash
python3 -m paddle.distributed.launch --ips="xx.xx.xx.xx,xx.xx.xx.xx" --gpus '0,1,2,3' tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml
```
**Note:** (1) When using multi-machine and multi-gpu training, you need to replace the ips value in the above command with the address of your machine, and the machines need to be able to ping each other. (2) Training needs to be launched separately on multiple machines. The command to view the ip address of the machine is `ifconfig`. (3) For more details about the distributed training speedup ratio, please refer to [Distributed Training Tutorial](./distributed_training_en.md).
## 2.5. Train with Knowledge Distillation
Knowledge distillation is supported in PaddleOCR for KIE model training process. The configuration file is [ser_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml). For more information, please refer to [doc](./knowledge_distillation_en.md).
**Note:** The saving and loading logic of the LayoutXLM series KIE models in PaddleOCR is consistent with PaddleNLP, so only the parameters of the student model are saved in the distillation process. If you want to use the saved model for evaluation, you need to use the configuration of the student model (the student model corresponding to the distillation file above is [ser_vi_layoutxlm_xfund_zh.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml).
## 2.6. Training on other platform
- Windows GPU/CPU
The Windows platform is slightly different from the Linux platform:
Windows platform only supports `single gpu` training and inference, specify GPU for training `set CUDA_VISIBLE_DEVICES=0`
On the Windows platform, DataLoader only supports single-process mode, so you need to set `num_workers` to 0;
- macOS
GPU mode is not supported, you need to set `use_gpu` to False in the configuration file, and the rest of the training evaluation prediction commands are exactly the same as Linux GPU.
- Linux DCU
Running on a DCU device requires setting the environment variable `export HIP_VISIBLE_DEVICES=0,1,2,3`, and the rest of the training and evaluation prediction commands are exactly the same as the Linux GPU.
# 3. Evaluation and Test
## 3.1. Evaluation
The trained model will be saved in `Global.save_model_dir`. When evaluation, you need to set `Architecture.Backbone.checkpoints` as your model directroy. The evaluation dataset can be set by modifying the `Eval.dataset.label_file_list` field in the `configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml` file.
```bash
# GPU evaluation, Global.checkpoints is the weight to be tested
python3 tools/eval.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy
```
The following information will be printed such as precision, recall, hmean and so on.
```py
[2022/08/09 07:59:28] ppocr INFO: metric eval ***************
[2022/08/09 07:59:28] ppocr INFO: precision:0.697476609016161
[2022/08/09 07:59:28] ppocr INFO: recall:0.8861671469740634
[2022/08/09 07:59:28] ppocr INFO: hmean:0.7805806758686339
[2022/08/09 07:59:28] ppocr INFO: fps:17.367364606899105
```
## 3.2. Test
Using the model trained by PaddleOCR, we can quickly get prediction through the following script.
The default prediction image is stored in `Global.infer_img`, and the trained model weight is specified via `-o Global.checkpoints`.
According to the `Global.save_model_dir` and `save_epoch_step` fields set in the configuration file, the following parameters will be saved.
```
output/ser_vi_layoutxlm_xfund_zh/
├── best_accuracy
├── metric.states
├── model_config.json
├── model_state.pdparams
├── best_accuracy.pdopt
├── config.yml
├── train.log
├── latest
├── metric.states
├── model_config.json
├── model_state.pdparams
├── latest.pdopt
```
Among them, best_accuracy.* is the best model on the evaluation set; latest.* is the model of the last epoch.
The configuration file for prediction must be consistent with the training file. If you finish the training process using `python3 tools/train.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml`. You can use the following command for prediction.
```bash
python3 tools/infer_kie_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.infer_img=./ppstructure/docs/kie/input/zh_val_42.jpg
```
The output image is as follows, which is also saved in `Global.save_res_path`.
<div align="center">
<img src="../../ppstructure/docs/kie/result_ser/zh_val_42_ser.jpg" width="800">
</div>
During the prediction process, the detection and recognition model of PP-OCRv3 will be loaded by default for information extraction of OCR. If you want to load the OCR results obtained in advance, you can use the following method to predict, and specify `Global.infer_img` as the annotation file, which contains the image path and OCR information, and specifies `Global.infer_mode` as False, indicating that the OCR inference engine is not used at this time.
```bash
python3 tools/infer_kie_token_ser.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.infer_img=./train_data/XFUND/zh_val/val.json Global.infer_mode=False
```
For the above image, if information extraction is performed using the labeled OCR results, the prediction results are as follows.
<div align="center">
<img src="../../ppstructure/docs/kie/result_ser_with_gt_ocr/zh_val_42_ser.jpg" width="800">
</div>
It can be seen that part of the detection information is more accurate, but the overall information extraction results are basically the same.
In RE model prediction, the SER model result needs to be given first, so the configuration file and model weight of SER need to be loaded at the same time, as shown in the following example.
```bash
python3 ./tools/infer_kie_token_ser_re.py \
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_udml_xfund_zh/best_accuracy/ \
Global.infer_img=./train_data/XFUND/zh_val/image/ \
-c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o_ser Architecture.Backbone.checkpoints=pretrain_models/ \
ser_vi_layoutxlm_udml_xfund_zh/best_accuracy/
```
The result is as follows.
<div align="center">
<img src="../../ppstructure/docs/kie/result_re/zh_val_42_re.jpg" width="800">
</div>
If you want to load the OCR results obtained in advance, you can use the following method to predict, and specify `Global.infer_img` as the annotation file, which contains the image path and OCR information, and specifies `Global.infer_mode` as False, indicating that the OCR inference engine is not used at this time.
```bash
python3 ./tools/infer_kie_token_ser_re.py \
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_udml_xfund_zh/best_accuracy/ \
Global.infer_img=./train_data/XFUND/zh_val/val.json \
Global.infer_mode=False \
-c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o_ser Architecture.Backbone.checkpoints=pretrain_models/ser_vi_layoutxlm_udml_xfund_zh/best_accuracy/
```
`c_ser` denotes SER configurations file, `o_ser` denotes the SER model configurations that will override corresponding content in the file.
The result is as follows.
<div align="center">
<img src="../../ppstructure/docs/kie/result_re_with_gt_ocr/zh_val_42_re.jpg" width="800">
</div>
It can be seen that the re prediction results directly using the annotated OCR results are more accurate.
# 4. Model inference
## 4.1 Export the model
The inference model (the model saved by `paddle.jit.save`) is generally a solidified model saved after the model training is completed, and is mostly used to give prediction in deployment.
The model saved during the training process is the checkpoints model, which saves the parameters of the model and is mostly used to resume training.
Compared with the checkpoints model, the inference model will additionally save the structural information of the model. Therefore, it is easier to deploy because the model structure and model parameters are already solidified in the inference model file, and is suitable for integration with actual systems.
The SER model can be converted to the inference model using the following command.
```bash
# -c Set the training algorithm yml configuration file.
# -o Set optional parameters.
# Architecture.Backbone.checkpoints Set the training model address.
# Global.save_inference_dir Set the address where the converted model will be saved.
python3 tools/export_model.py -c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml -o Architecture.Backbone.checkpoints=./output/ser_vi_layoutxlm_xfund_zh/best_accuracy Global.save_inference_dir=./inference/ser_vi_layoutxlm
```
After the conversion is successful, there are three files in the model save directory:
```
inference/ser_vi_layoutxlm/
├── inference.pdiparams # The parameter file of recognition inference model
├── inference.pdiparams.info # The parameter information of recognition inference model, which can be ignored
└── inference.pdmodel # The program file of recognition
```
Export of RE model is also in adaptation.
## 4.2 Model inference
The VI layoutxlm model performs reasoning based on the ser task, and can execute the following commands:
Using the following command to infer the VI-LayoutXLM model.
```bash
cd ppstructure
python3 kie/predict_kie_token_ser.py \
--kie_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_vi_layoutxlm \
--image_dir=./docs/kie/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx"
```
The visualized result will be saved in `./output`, which is shown as follows.
<div align="center">
<img src="../../ppstructure/docs/kie/result_ser/zh_val_42_ser.jpg" width="800">
</div>
# 5. FAQ
Q1: After the training model is transferred to the inference model, the prediction effect is inconsistent?
**A**：The problems are mostly caused by inconsistent preprocessing and postprocessing parameters when the trained model predicts and the preprocessing and postprocessing parameters when the inference model predicts. You can compare whether there are differences in preprocessing, postprocessing, and prediction in the configuration files used for training.
paddleocr.py
浏览文件 @ 96a8e0f8
...@@ -35,7 +35,7 @@ from tools.infer import predict_system ...@@ -35,7 +35,7 @@ from tools.infer import predict_system
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
logger = get_logger() logger = get_logger()
from ppocr.utils.utility import check_and_read_gif, get_image_file_list from ppocr.utils.utility import check_and_read, get_image_file_list
from ppocr.utils.network import maybe_download, download_with_progressbar, is_link, confirm_model_dir_url from ppocr.utils.network import maybe_download, download_with_progressbar, is_link, confirm_model_dir_url
from tools.infer.utility import draw_ocr, str2bool, check_gpu from tools.infer.utility import draw_ocr, str2bool, check_gpu
from ppstructure.utility import init_args, draw_structure_result from ppstructure.utility import init_args, draw_structure_result
...@@ -289,7 +289,8 @@ MODEL_URLS = { ...@@ -289,7 +289,8 @@ MODEL_URLS = {
'ch': { 'ch': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_layout_infer.tar', 'https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_layout_infer.tar',
'dict_path': 'ppocr/utils/dict/layout_publaynet_dict.txt' 'dict_path':
'ppocr/utils/dict/layout_dict/layout_publaynet_dict.txt'
} }
} }
} }
...@@ -490,7 +491,7 @@ class PaddleOCR(predict_system.TextSystem): ...@@ -490,7 +491,7 @@ class PaddleOCR(predict_system.TextSystem):
download_with_progressbar(img, 'tmp.jpg') download_with_progressbar(img, 'tmp.jpg')
img = 'tmp.jpg' img = 'tmp.jpg'
image_file = img image_file = img
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
with open(image_file, 'rb') as f: with open(image_file, 'rb') as f:
np_arr = np.frombuffer(f.read(), dtype=np.uint8) np_arr = np.frombuffer(f.read(), dtype=np.uint8)
...@@ -584,7 +585,7 @@ class PPStructure(StructureSystem): ...@@ -584,7 +585,7 @@ class PPStructure(StructureSystem):
download_with_progressbar(img, 'tmp.jpg') download_with_progressbar(img, 'tmp.jpg')
img = 'tmp.jpg' img = 'tmp.jpg'
image_file = img image_file = img
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
with open(image_file, 'rb') as f: with open(image_file, 'rb') as f:
np_arr = np.frombuffer(f.read(), dtype=np.uint8) np_arr = np.frombuffer(f.read(), dtype=np.uint8)
...@@ -635,4 +636,6 @@ def main(): ...@@ -635,4 +636,6 @@ def main():
for item in result: for item in result:
item.pop('img') item.pop('img')
item.pop('res')
logger.info(item) logger.info(item)
logger.info('result save to {}'.format(args.output))
ppocr/data/imaug/copy_paste.py
浏览文件 @ 96a8e0f8
...@@ -35,10 +35,12 @@ class CopyPaste(object): ...@@ -35,10 +35,12 @@ class CopyPaste(object):
point_num = data['polys'].shape[1] point_num = data['polys'].shape[1]
src_img = data['image'] src_img = data['image']
src_polys = data['polys'].tolist() src_polys = data['polys'].tolist()
src_texts = data['texts']
src_ignores = data['ignore_tags'].tolist() src_ignores = data['ignore_tags'].tolist()
ext_data = data['ext_data'][0] ext_data = data['ext_data'][0]
ext_image = ext_data['image'] ext_image = ext_data['image']
ext_polys = ext_data['polys'] ext_polys = ext_data['polys']
ext_texts = ext_data['texts']
ext_ignores = ext_data['ignore_tags'] ext_ignores = ext_data['ignore_tags']
indexs = [i for i in range(len(ext_ignores)) if not ext_ignores[i]] indexs = [i for i in range(len(ext_ignores)) if not ext_ignores[i]]
...@@ -53,7 +55,7 @@ class CopyPaste(object): ...@@ -53,7 +55,7 @@ class CopyPaste(object):
src_img = cv2.cvtColor(src_img, cv2.COLOR_BGR2RGB) src_img = cv2.cvtColor(src_img, cv2.COLOR_BGR2RGB)
ext_image = cv2.cvtColor(ext_image, cv2.COLOR_BGR2RGB) ext_image = cv2.cvtColor(ext_image, cv2.COLOR_BGR2RGB)
src_img = Image.fromarray(src_img).convert('RGBA') src_img = Image.fromarray(src_img).convert('RGBA')
for poly, tag in zip(select_polys, select_ignores): for idx, poly, tag in zip(select_idxs, select_polys, select_ignores):
box_img = get_rotate_crop_image(ext_image, poly) box_img = get_rotate_crop_image(ext_image, poly)
src_img, box = self.paste_img(src_img, box_img, src_polys) src_img, box = self.paste_img(src_img, box_img, src_polys)
...@@ -62,6 +64,7 @@ class CopyPaste(object): ...@@ -62,6 +64,7 @@ class CopyPaste(object):
for _ in range(len(box), point_num): for _ in range(len(box), point_num):
box.append(box[-1]) box.append(box[-1])
src_polys.append(box) src_polys.append(box)
src_texts.append(ext_texts[idx])
src_ignores.append(tag) src_ignores.append(tag)
src_img = cv2.cvtColor(np.array(src_img), cv2.COLOR_RGB2BGR) src_img = cv2.cvtColor(np.array(src_img), cv2.COLOR_RGB2BGR)
h, w = src_img.shape[:2] h, w = src_img.shape[:2]
...@@ -70,6 +73,7 @@ class CopyPaste(object): ...@@ -70,6 +73,7 @@ class CopyPaste(object):
src_polys[:, :, 1] = np.clip(src_polys[:, :, 1], 0, h) src_polys[:, :, 1] = np.clip(src_polys[:, :, 1], 0, h)
data['image'] = src_img data['image'] = src_img
data['polys'] = src_polys data['polys'] = src_polys
data['texts'] = src_texts
data['ignore_tags'] = np.array(src_ignores) data['ignore_tags'] = np.array(src_ignores)
return data return data
......
ppocr/metrics/rec_metric.py
浏览文件 @ 96a8e0f8
...@@ -12,7 +12,7 @@ ...@@ -12,7 +12,7 @@
# See the License for the specific language governing permissions and # See the License for the specific language governing permissions and
# limitations under the License. # limitations under the License.
import Levenshtein from rapidfuzz.distance import Levenshtein
import string import string
...@@ -46,8 +46,7 @@ class RecMetric(object): ...@@ -46,8 +46,7 @@ class RecMetric(object):
if self.is_filter: if self.is_filter:
pred = self._normalize_text(pred) pred = self._normalize_text(pred)
target = self._normalize_text(target) target = self._normalize_text(target)
norm_edit_dis += Levenshtein.distance(pred, target) / max( norm_edit_dis += Levenshtein.normalized_distance(pred, target)
len(pred), len(target), 1)
if pred == target: if pred == target:
correct_num += 1 correct_num += 1
all_num += 1 all_num += 1
......
ppocr/modeling/backbones/__init__.py
浏览文件 @ 96a8e0f8
...@@ -52,17 +52,15 @@ def build_backbone(config, model_type): ...@@ -52,17 +52,15 @@ def build_backbone(config, model_type):
support_dict = ['ResNet'] support_dict = ['ResNet']
elif model_type == 'kie': elif model_type == 'kie':
from .kie_unet_sdmgr import Kie_backbone from .kie_unet_sdmgr import Kie_backbone
support_dict = ['Kie_backbone'] from .vqa_layoutlm import LayoutLMForSer, LayoutLMv2ForSer, LayoutLMv2ForRe, LayoutXLMForSer, LayoutXLMForRe
support_dict = [
'Kie_backbone', 'LayoutLMForSer', 'LayoutLMv2ForSer',
'LayoutLMv2ForRe', 'LayoutXLMForSer', 'LayoutXLMForRe'
]
elif model_type == 'table': elif model_type == 'table':
from .table_resnet_vd import ResNet from .table_resnet_vd import ResNet
from .table_mobilenet_v3 import MobileNetV3 from .table_mobilenet_v3 import MobileNetV3
support_dict = ['ResNet', 'MobileNetV3'] support_dict = ['ResNet', 'MobileNetV3']
elif model_type == 'vqa':
from .vqa_layoutlm import LayoutLMForSer, LayoutLMv2ForSer, LayoutLMv2ForRe, LayoutXLMForSer, LayoutXLMForRe
support_dict = [
'LayoutLMForSer', 'LayoutLMv2ForSer', 'LayoutLMv2ForRe',
'LayoutXLMForSer', 'LayoutXLMForRe'
]
else: else:
raise NotImplementedError raise NotImplementedError
......
ppocr/utils/save_load.py
浏览文件 @ 96a8e0f8
...@@ -54,13 +54,15 @@ def load_model(config, model, optimizer=None, model_type='det'): ...@@ -54,13 +54,15 @@ def load_model(config, model, optimizer=None, model_type='det'):
pretrained_model = global_config.get('pretrained_model') pretrained_model = global_config.get('pretrained_model')
best_model_dict = {} best_model_dict = {}
is_float16 = False is_float16 = False
is_nlp_model = model_type == 'kie' and config["Architecture"][
"algorithm"] not in ["SDMGR"]
if model_type == 'vqa': if is_nlp_model is True:
# NOTE: for vqa model dsitillation, resume training is not supported now # NOTE: for kie model dsitillation, resume training is not supported now
if config["Architecture"]["algorithm"] in ["Distillation"]: if config["Architecture"]["algorithm"] in ["Distillation"]:
return best_model_dict return best_model_dict
checkpoints = config['Architecture']['Backbone']['checkpoints'] checkpoints = config['Architecture']['Backbone']['checkpoints']
# load vqa method metric # load kie method metric
if checkpoints: if checkpoints:
if os.path.exists(os.path.join(checkpoints, 'metric.states')): if os.path.exists(os.path.join(checkpoints, 'metric.states')):
with open(os.path.join(checkpoints, 'metric.states'), with open(os.path.join(checkpoints, 'metric.states'),
...@@ -102,8 +104,9 @@ def load_model(config, model, optimizer=None, model_type='det'): ...@@ -102,8 +104,9 @@ def load_model(config, model, optimizer=None, model_type='det'):
continue continue
pre_value = params[key] pre_value = params[key]
if pre_value.dtype == paddle.float16: if pre_value.dtype == paddle.float16:
pre_value = pre_value.astype(paddle.float32)
is_float16 = True is_float16 = True
if pre_value.dtype != value.dtype:
pre_value = pre_value.astype(value.dtype)
if list(value.shape) == list(pre_value.shape): if list(value.shape) == list(pre_value.shape):
new_state_dict[key] = pre_value new_state_dict[key] = pre_value
else: else:
...@@ -160,8 +163,9 @@ def load_pretrained_params(model, path): ...@@ -160,8 +163,9 @@ def load_pretrained_params(model, path):
logger.warning("The pretrained params {} not in model".format(k1)) logger.warning("The pretrained params {} not in model".format(k1))
else: else:
if params[k1].dtype == paddle.float16: if params[k1].dtype == paddle.float16:
params[k1] = params[k1].astype(paddle.float32)
is_float16 = True is_float16 = True
if params[k1].dtype != state_dict[k1].dtype:
params[k1] = params[k1].astype(state_dict[k1].dtype)
if list(state_dict[k1].shape) == list(params[k1].shape): if list(state_dict[k1].shape) == list(params[k1].shape):
new_state_dict[k1] = params[k1] new_state_dict[k1] = params[k1]
else: else:
...@@ -192,10 +196,13 @@ def save_model(model, ...@@ -192,10 +196,13 @@ def save_model(model,
_mkdir_if_not_exist(model_path, logger) _mkdir_if_not_exist(model_path, logger)
model_prefix = os.path.join(model_path, prefix) model_prefix = os.path.join(model_path, prefix)
paddle.save(optimizer.state_dict(), model_prefix + '.pdopt') paddle.save(optimizer.state_dict(), model_prefix + '.pdopt')
if config['Architecture']["model_type"] != 'vqa':
is_nlp_model = config['Architecture']["model_type"] == 'kie' and config[
"Architecture"]["algorithm"] not in ["SDMGR"]
if is_nlp_model is not True:
paddle.save(model.state_dict(), model_prefix + '.pdparams') paddle.save(model.state_dict(), model_prefix + '.pdparams')
metric_prefix = model_prefix metric_prefix = model_prefix
else: # for vqa system, we follow the save/load rules in NLP else: # for kie system, we follow the save/load rules in NLP
if config['Global']['distributed']: if config['Global']['distributed']:
arch = model._layers arch = model._layers
else: else:
......
ppocr/utils/utility.py
浏览文件 @ 96a8e0f8
...@@ -50,7 +50,7 @@ def get_check_global_params(mode): ...@@ -50,7 +50,7 @@ def get_check_global_params(mode):
def _check_image_file(path): def _check_image_file(path):
img_end = {'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif', 'tiff', 'gif'} img_end = {'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif', 'tiff', 'gif', 'pdf'}
return any([path.lower().endswith(e) for e in img_end]) return any([path.lower().endswith(e) for e in img_end])
...@@ -59,7 +59,7 @@ def get_image_file_list(img_file): ...@@ -59,7 +59,7 @@ def get_image_file_list(img_file):
if img_file is None or not os.path.exists(img_file): if img_file is None or not os.path.exists(img_file):
raise Exception("not found any img file in {}".format(img_file)) raise Exception("not found any img file in {}".format(img_file))
img_end = {'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif', 'tiff', 'gif'} img_end = {'jpg', 'bmp', 'png', 'jpeg', 'rgb', 'tif', 'tiff', 'gif', 'pdf'}
if os.path.isfile(img_file) and _check_image_file(img_file): if os.path.isfile(img_file) and _check_image_file(img_file):
imgs_lists.append(img_file) imgs_lists.append(img_file)
elif os.path.isdir(img_file): elif os.path.isdir(img_file):
...@@ -73,7 +73,7 @@ def get_image_file_list(img_file): ...@@ -73,7 +73,7 @@ def get_image_file_list(img_file):
return imgs_lists return imgs_lists
def check_and_read_gif(img_path): def check_and_read(img_path):
if os.path.basename(img_path)[-3:] in ['gif', 'GIF']: if os.path.basename(img_path)[-3:] in ['gif', 'GIF']:
gif = cv2.VideoCapture(img_path) gif = cv2.VideoCapture(img_path)
ret, frame = gif.read() ret, frame = gif.read()
...@@ -84,8 +84,26 @@ def check_and_read_gif(img_path): ...@@ -84,8 +84,26 @@ def check_and_read_gif(img_path):
if len(frame.shape) == 2 or frame.shape[-1] == 1: if len(frame.shape) == 2 or frame.shape[-1] == 1:
frame = cv2.cvtColor(frame, cv2.COLOR_GRAY2RGB) frame = cv2.cvtColor(frame, cv2.COLOR_GRAY2RGB)
imgvalue = frame[:, :, ::-1] imgvalue = frame[:, :, ::-1]
return imgvalue, True return imgvalue, True, False
return None, False elif os.path.basename(img_path)[-3:] in ['pdf']:
import fitz
from PIL import Image
imgs = []
with fitz.open(img_path) as pdf:
for pg in range(0, pdf.pageCount):
page = pdf[pg]
mat = fitz.Matrix(2, 2)
pm = page.getPixmap(matrix=mat, alpha=False)
# if width or height > 2000 pixels, don't enlarge the image
if pm.width > 2000 or pm.height > 2000:
pm = page.getPixmap(matrix=fitz.Matrix(1, 1), alpha=False)
img = Image.frombytes("RGB", [pm.width, pm.height], pm.samples)
img = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
imgs.append(img)
return imgs, False, True
return None, False, False
def load_vqa_bio_label_maps(label_map_path): def load_vqa_bio_label_maps(label_map_path):
......
ppstructure/README.md
浏览文件 @ 96a8e0f8
English | [简体中文](README_ch.md) English | [简体中文](README_ch.md)
- [1. Introduction](#1-introduction) - [1. Introduction](#1-introduction)
- [2. Update log](#2-update-log) - [2. Features](#2-features)
- [3. Features](#3-features) - [3. Results](#3-results)
- [4. Results](#4-results) - [3.1 Layout analysis and table recognition](#31-layout-analysis-and-table-recognition)
- [4.1 Layout analysis and table recognition](#41-layout-analysis-and-table-recognition) - [3.2 Layout Recovery](#32-layout-recovery)
- [4.2 DOC-VQA](#42-doc-vqa) - [3.3 KIE](#33-kie)
- [5. Quick start](#5-quick-start) - [4. Quick start](#4-quick-start)
- [6. PP-Structure System](#6-pp-structure-system) - [5. Model List](#5-model-list)
- [6.1 Layout analysis and table recognition](#61-layout-analysis-and-table-recognition)
- [6.1.1 Layout analysis](#611-layout-analysis)
- [6.1.2 Table recognition](#612-table-recognition)
- [6.2 DOC-VQA](#62-doc-vqa)
- [7. Model List](#7-model-list)
- [7.1 Layout analysis model](#71-layout-analysis-model)
- [7.2 OCR and table recognition model](#72-ocr-and-table-recognition-model)
- [7.3 DOC-VQA model](#73-doc-vqa-model)
## 1. Introduction ## 1. Introduction
PP-Structure is an OCR toolkit that can be used for document analysis and processing with complex structures, designed to help developers better complete document understanding tasks PP-Structure is an intelligent document analysis system developed by the PaddleOCR team, which aims to help developers better complete tasks related to document understanding such as layout analysis and table recognition.
## 2. Update log The pipeline of PP-Structurev2 system is shown below. The document image first passes through the image direction correction module to identify the direction of the entire image and complete the direction correction. Then, two tasks of layout information analysis and key information extraction can be completed.
* 2022.02.12 DOC-VQA add LayoutLMv2 model。
* 2021.12.07 add [DOC-VQA SER and RE tasks](vqa/README.md)。
## 3. Features - In the layout analysis task, the image first goes through the layout analysis model to divide the image into different areas such as text, table, and figure, and then analyze these areas separately. For example, the table area is sent to the form recognition module for structured recognition, and the text area is sent to the OCR engine for text recognition. Finally, the layout recovery module restores it to a word or pdf file with the same layout as the original image;
- In the key information extraction task, the OCR engine is first used to extract the text content, and then the SER(semantic entity recognition) module obtains the semantic entities in the image, and finally the RE(relationship extraction) module obtains the correspondence between the semantic entities, thereby extracting the required key information.
<img src="./docs/ppstructurev2_pipeline.png" width="100%"/>
The main features of PP-Structure are as follows: More technical details: 👉 [PP-Structurev2 Technical Report](docs/PP-Structurev2_introduction.md)
- Support the layout analysis of documents, divide the documents into 5 types of areas **text, title, table, image and list** (conjunction with Layout-Parser) PP-Structurev2 supports independent use or flexible collocation of each module. For example, you can use layout analysis alone or table recognition alone. Click the corresponding link below to get the tutorial for each independent module:
- Support to extract the texts from the text, title, picture and list areas (used in conjunction with PP-OCR)
- Support to extract excel files from the table areas
- Support python whl package and command line usage, easy to use
- Support custom training for layout analysis and table structure tasks
- Support Document Visual Question Answering (DOC-VQA) tasks: Semantic Entity Recognition (SER) and Relation Extraction (RE)
## 4. Results - [Layout Analysis](layout/README.md)
- [Table Recognition](table/README.md)
- [Key Information Extraction](kie/README.md)
- [Layout Recovery](recovery/README.md)
### 4.1 Layout analysis and table recognition ## 2. Features
<img src="docs/table/ppstructure.GIF" width="100%"/> The main features of PP-Structurev2 are as follows:
- Support layout analysis of documents in the form of images/pdfs, which can be divided into areas such as **text, titles, tables, figures, formulas, etc.**;
The figure shows the pipeline of layout analysis + table recognition. The image is first divided into four areas of image, text, title and table by layout analysis, and then OCR detection and recognition is performed on the three areas of image, text and title, and the table is performed table recognition, where the image will also be stored for use. - Support common Chinese and English **table detection** tasks;
- Support structured table recognition, and output the final result to **Excel file**;
### 4.2 DOC-VQA - Support multimodal-based Key Information Extraction (KIE) tasks - **Semantic Entity Recognition** (SER) and **Relation Extraction (RE);
- Support **layout recovery**, that is, restore the document in word or pdf format with the same layout as the original image;
* SER - Support customized training and multiple inference deployment methods such as python whl package quick start;
* - Connect with the semi-automatic data labeling tool PPOCRLabel, which supports the labeling of layout analysis, table recognition, and SER.
![](docs/vqa/result_ser/zh_val_0_ser.jpg) | ![](docs/vqa/result_ser/zh_val_42_ser.jpg)
---|---
Different colored boxes in the figure represent different categories. For xfun dataset, there are three categories: query, answer and header:
* Dark purple: header ## 3. Results
* Light purple: query
* Army green: answer
The corresponding category and OCR recognition results are also marked at the top left of the OCR detection box. PP-Structurev2 supports the independent use or flexible collocation of each module. For example, layout analysis can be used alone, or table recognition can be used alone. Only the visualization effects of several representative usage methods are shown here.
### 3.1 Layout analysis and table recognition
* RE The figure shows the pipeline of layout analysis + table recognition. The image is first divided into four areas of image, text, title and table by layout analysis, and then OCR detection and recognition is performed on the three areas of image, text and title, and the table is performed table recognition, where the image will also be stored for use.
<img src="docs/table/ppstructure.GIF" width="100%"/>
![](docs/vqa/result_re/zh_val_21_re.jpg) | ![](docs/vqa/result_re/zh_val_40_re.jpg)
---|---
### 3.2 Layout recovery
In the figure, the red box represents the question, the blue box represents the answer, and the question and answer are connected by green lines. The corresponding category and OCR recognition results are also marked at the top left of the OCR detection box. The following figure shows the effect of layout recovery based on the results of layout analysis and table recognition in the previous section.
<img src="./docs/recovery/recovery.jpg" width="100%"/>
## 5. Quick start ### 3.3 KIE
Start from [Quick Installation](./docs/quickstart.md) * SER
## 6. PP-Structure System Different colored boxes in the figure represent different categories.
### 6.1 Layout analysis and table recognition <div align="center">
<img src="https://user-images.githubusercontent.com/25809855/186094456-01a1dd11-1433-4437-9ab2-6480ac94ec0a.png" width="600">
</div>
![pipeline](docs/table/pipeline.jpg) <div align="center">
<img src="https://user-images.githubusercontent.com/25809855/186095702-9acef674-12af-4d09-97fc-abf4ab32600e.png" width="600">
</div>
In PP-Structure, the image will be divided into 5 types of areas **text, title, image list and table**. For the first 4 types of areas, directly use PP-OCR system to complete the text detection and recognition. For the table area, after the table structuring process, the table in image is converted into an Excel file with the same table style. <div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539141-68e71c75-5cf7-4529-b2ca-219d29fa5f68.jpg" width="600">
</div>
#### 6.1.1 Layout analysis <div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185310636-6ce02f7c-790d-479f-b163-ea97a5a04808.jpg" width="600">
</div>
Layout analysis classifies image by region, including the use of Python scripts of layout analysis tools, extraction of designated category detection boxes, performance indicators, and custom training layout analysis models. For details, please refer to [document](layout/README.md). <div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539517-ccf2372a-f026-4a7c-ad28-c741c770f60a.png" width="600">
</div>
#### 6.1.2 Table recognition * RE
Table recognition converts table images into excel documents, which include the detection and recognition of table text and the prediction of table structure and cell coordinates. For detailed instructions, please refer to [document](table/README.md) In the figure, the red box represents `Question`, the blue box represents `Answer`, and `Question` and `Answer` are connected by green lines.
### 6.2 DOC-VQA <div align="center">
<img src="https://user-images.githubusercontent.com/25809855/186094813-3a8e16cc-42e5-4982-b9f4-0134dfb5688d.png" width="600">
</div>
Document Visual Question Answering (DOC-VQA) if a type of Visual Question Answering (VQA), which includes Semantic Entity Recognition (SER) and Relation Extraction (RE) tasks. Based on SER task, text recognition and classification in images can be completed. Based on THE RE task, we can extract the relation of the text content in the image, such as judge the problem pair. For details, please refer to [document](vqa/README.md) <div align="center">
<img src="https://user-images.githubusercontent.com/25809855/186095641-5843b4da-34d7-4c1c-943a-b1036a859fe3.png" width="600">
</div>
## 7. Model List <div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185393805-c67ff571-cf7e-4217-a4b0-8b396c4f22bb.jpg" width="600">
</div>
PP-Structure Series Model List (Updating) <div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185540080-0431e006-9235-4b6d-b63d-0b3c6e1de48f.jpg" width="600">
</div>
### 7.1 Layout analysis model ## 4. Quick start
|model name|description|download|label_map| Start from [Quick Start](./docs/quickstart_en.md).
| --- | --- | --- |--- |
| ppyolov2_r50vd_dcn_365e_publaynet | The layout analysis model trained on the PubLayNet dataset can divide image into 5 types of areas **text, title, table, picture, and list** | [PubLayNet](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet.tar) | {0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"}|
### 7.2 OCR and table recognition model ## 5. Model List
|model name|description|model size|download| Some tasks need to use both the structured analysis models and the OCR models. For example, the table recognition task needs to use the table recognition model for structured analysis, and the OCR model to recognize the text in the table. Please select the appropriate models according to your specific needs.
| --- | --- | --- | --- |
|ch_PP-OCRv3_det_slim|[New] slim quantization with distillation lightweight model, supporting Chinese, English, multilingual text detection| 1.1M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_distill_train.tar)|
|ch_PP-OCRv3_rec_slim |[New] Slim qunatization with distillation lightweight model, supporting Chinese, English text recognition| 4.9M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_train.tar) |
|ch_ppstructure_mobile_v2.0_SLANet|Chinese table recognition model trained on PubTabNet dataset based on SLANet|9.3M|[inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_train.tar) |
### 7.3 DOC-VQA model For structural analysis related model downloads, please refer to:
- [PP-Structure Model Zoo](./docs/models_list_en.md)
|model name|description|model size|download| For OCR related model downloads, please refer to:
| --- | --- | --- | --- | - [PP-OCR Model Zoo](../doc/doc_en/models_list_en.md)
|ser_LayoutXLM_xfun_zhd|SER model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|re_LayoutXLM_xfun_zh|RE model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
If you need to use other models, you can download the model in [PPOCR model_list](../doc/doc_en/models_list_en.md) and [PPStructure model_list](./docs/models_list.md)
ppstructure/README_ch.md
浏览文件 @ 96a8e0f8
...@@ -3,135 +3,120 @@ ...@@ -3,135 +3,120 @@
# PP-Structure 文档分析 # PP-Structure 文档分析
- [1. 简介](#1) - [1. 简介](#1)
- [2. 近期更新](#2) - [2. 特性](#2)
- [3. 特性](#3) - [3. 效果展示](#3)
- [4. 效果展示](#4) - [3.1 版面分析和表格识别](#31)
- [4.1 版面分析和表格识别](#41) - [3.2 版面恢复](#32)
- [4.2 DocVQA](#42) - [3.3 关键信息抽取](#33)
- [5. 快速体验](#5) - [4. 快速体验](#4)
- [6. PP-Structure 介绍](#6) - [5. 模型库](#5)
- [6.1 版面分析+表格识别](#61)
- [6.1.1 版面分析](#611)
- [6.1.2 表格识别](#612)
- [6.2 DocVQA](#62)
- [7. 模型库](#7)
- [7.1 版面分析模型](#71)
- [7.2 OCR和表格识别模型](#72)
- [7.3 DocVQA 模型](#73)
<a name="1"></a> <a name="1"></a>
## 1. 简介 ## 1. 简介
PP-Structure是一个可用于复杂文档结构分析和处理的OCR工具包，旨在帮助开发者更好的完成文档理解相关任务。
<a name="2"></a> PP-Structure是PaddleOCR团队自研的智能文档分析系统，旨在帮助开发者更好的完成版面分析、表格识别等文档理解相关任务。
## 2. 近期更新
* 2022.02.12 DocVQA增加LayoutLMv2模型。
* 2021.12.07 新增[DOC-VQA任务SER和RE](vqa/README.md)。
<a name="3"></a>
## 3. 特性
PP-Structure的主要特性如下：
- 支持对图片形式的文档进行版面分析，可以划分**文字、标题、表格、图片以及列表**5类区域（与Layout-Parser联合使用）
- 支持文字、标题、图片以及列表区域提取为文字字段（与PP-OCR联合使用）
- 支持表格区域进行结构化分析，最终结果输出Excel文件
- 支持python whl包和命令行两种方式，简单易用
- 支持版面分析和表格结构化两类任务自定义训练
- 支持文档视觉问答(Document Visual Question Answering，DocVQA)任务-语义实体识别(Semantic Entity Recognition，SER)和关系抽取(Relation Extraction，RE)
<a name="4"></a>
## 4. 效果展示
<a name="41"></a>
### 4.1 版面分析和表格识别
<img src="./docs/table/ppstructure.GIF" width="100%"/>
图中展示了版面分析+表格识别的整体流程，图片先有版面分析划分为图像、文本、标题和表格四种区域，然后对图像、文本和标题三种区域进行OCR的检测识别，对表格进行表格识别，其中图像还会被存储下来以便使用。
<a name="42"></a>
### 4.2 DOC-VQA
* SER
![](./docs/vqa/result_ser/zh_val_0_ser.jpg) | ![](./docs/vqa/result_ser/zh_val_42_ser.jpg) PP-Structurev2系统流程图如下所示，文档图像首先经过图像矫正模块，判断整图方向并完成转正，随后可以完成版面信息分析与关键信息抽取2类任务。
---|--- - 版面分析任务中，图像首先经过版面分析模型，将图像划分为文本、表格、图像等不同区域，随后对这些区域分别进行识别，如，将表格区域送入表格识别模块进行结构化识别，将文本区域送入OCR引擎进行文字识别，最后使用版面恢复模块将其恢复为与原始图像布局一致的word或者pdf格式的文件；
- 关键信息抽取任务中，首先使用OCR引擎提取文本内容，然后由语义实体识别模块获取图像中的语义实体，最后经关系抽取模块获取语义实体之间的对应关系，从而提取需要的关键信息。
<img src="./docs/ppstructurev2_pipeline.png" width="100%"/>
图中不同颜色的框表示不同的类别，对于XFUN数据集，有`QUESTION`, `ANSWER`, `HEADER` 3种类别 更多技术细节：👉 [PP-Structurev2技术报告](docs/PP-Structurev2_introduction.md)
* 深紫色：HEADER PP-Structurev2支持各个模块独立使用或灵活搭配，如，可以单独使用版面分析，或单独使用表格识别，点击下面相应链接获取各个独立模块的使用教程：
* 浅紫色：QUESTION
* 军绿色：ANSWER
在OCR检测框的左上方也标出了对应的类别和OCR识别结果。 - [版面分析](layout/README_ch.md)
- [表格识别](table/README_ch.md)
- [关键信息抽取](kie/README_ch.md)
- [版面复原](recovery/README_ch.md)
* RE <a name="2"></a>
## 2. 特性
![](./docs/vqa/result_re/zh_val_21_re.jpg) | ![](./docs/vqa/result_re/zh_val_40_re.jpg) PP-Structurev2的主要特性如下：
---|--- - 支持对图片/pdf形式的文档进行版面分析，可以划分**文字、标题、表格、图片、公式等**区域；
- 支持通用的中英文**表格检测**任务；
- 支持表格区域进行结构化识别，最终结果输出**Excel文件**；
- 支持基于多模态的关键信息抽取(Key Information Extraction，KIE)任务-**语义实体识别**(Semantic Entity Recognition，SER)和**关系抽取**(Relation Extraction，RE)；
- 支持**版面复原**，即恢复为与原始图像布局一致的word或者pdf格式的文件；
- 支持自定义训练及python whl包调用等多种推理部署方式，简单易用；
- 与半自动数据标注工具PPOCRLabel打通，支持版面分析、表格识别、SER三种任务的标注。
<a name="3"></a>
## 3. 效果展示
PP-Structurev2支持各个模块独立使用或灵活搭配，如，可以单独使用版面分析，或单独使用表格识别，这里仅展示几种代表性使用方式的可视化效果。
图中红色框表示问题，蓝色框表示答案，问题和答案之间使用绿色线连接。在OCR检测框的左上方也标出了对应的类别和OCR识别结果。 <a name="31"></a>
### 3.1 版面分析和表格识别
下图展示了版面分析+表格识别的整体流程，图片先有版面分析划分为图像、文本、标题和表格四种区域，然后对图像、文本和标题三种区域进行OCR的检测识别，对表格进行表格识别，其中图像还会被存储下来以便使用。
<img src="./docs/table/ppstructure.GIF" width="100%"/>
<a name="5"></a> <a name="32"></a>
## 5. 快速体验 ### 3.2 版面恢复
下图展示了基于上一节版面分析和表格识别的结果进行版面恢复的效果。
<img src="./docs/recovery/recovery.jpg" width="100%"/>
请参考[快速使用](./docs/quickstart.md)教程。
<a name="6"></a> <a name="33"></a>
## 6. PP-Structure 介绍 ### 3.3 关键信息抽取
<a name="61"></a> * SER
### 6.1 版面分析+表格识别
![pipeline](./docs/table/pipeline.jpg) 图中不同颜色的框表示不同的类别。
在PP-Structure中，图片会先经由Layout-Parser进行版面分析，在版面分析中，会对图片里的区域进行分类，包括**文字、标题、图片、列表和表格**5类。对于前4类区域，直接使用PP-OCR完成对应区域文字检测与识别。对于表格类区域，经过表格结构化处理后，表格图片转换为相同表格样式的Excel文件。 <div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539141-68e71c75-5cf7-4529-b2ca-219d29fa5f68.jpg" width="600">
</div>
<a name="611"></a> <div align="center">
#### 6.1.1 版面分析 <img src="https://user-images.githubusercontent.com/14270174/185310636-6ce02f7c-790d-479f-b163-ea97a5a04808.jpg" width="600">
</div>
版面分析对文档数据进行区域分类，其中包括版面分析工具的Python脚本使用、提取指定类别检测框、性能指标以及自定义训练版面分析模型，详细内容可以参考[文档](layout/README_ch.md)。 <div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539517-ccf2372a-f026-4a7c-ad28-c741c770f60a.png" width="600">
</div>
<a name="612"></a> <div align="center">
#### 6.1.2 表格识别 <img src="https://user-images.githubusercontent.com/25809855/186094456-01a1dd11-1433-4437-9ab2-6480ac94ec0a.png" width="600">
</div>
表格识别将表格图片转换为excel文档，其中包含对于表格文本的检测和识别以及对于表格结构和单元格坐标的预测，详细说明参考[文档](table/README_ch.md)。 <div align="center">
<img src="https://user-images.githubusercontent.com/25809855/186095702-9acef674-12af-4d09-97fc-abf4ab32600e.png" width="600">
</div>
<a name="62"></a> * RE
### 6.2 DocVQA
DocVQA指文档视觉问答，其中包括语义实体识别 (Semantic Entity Recognition, SER) 和关系抽取 (Relation Extraction, RE) 任务。基于 SER 任务，可以完成对图像中的文本识别与分类；基于 RE 任务，可以完成对图象中的文本内容的关系提取，如判断问题对(pair)，详细说明参考[文档](vqa/README.md)。 图中红色框表示`问题`，蓝色框表示`答案`，`问题`和`答案`之间使用绿色线连接。
<a name="7"></a> <div align="center">
## 7. 模型库 <img src="https://user-images.githubusercontent.com/14270174/185393805-c67ff571-cf7e-4217-a4b0-8b396c4f22bb.jpg" width="600">
</div>
PP-Structure系列模型列表（更新中） <div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185540080-0431e006-9235-4b6d-b63d-0b3c6e1de48f.jpg" width="600">
</div>
<a name="71"></a> <div align="center">
### 7.1 版面分析模型 <img src="https://user-images.githubusercontent.com/25809855/186094813-3a8e16cc-42e5-4982-b9f4-0134dfb5688d.png" width="600">
</div>
|模型名称|模型简介|下载地址| label_map| <div align="center">
| --- | --- | --- | --- | <img src="https://user-images.githubusercontent.com/25809855/186095641-5843b4da-34d7-4c1c-943a-b1036a859fe3.png" width="600">
| ppyolov2_r50vd_dcn_365e_publaynet | PubLayNet 数据集训练的版面分析模型，可以划分**文字、标题、表格、图片以及列表**5类区域 | [PubLayNet](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet.tar) | {0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"}| </div>
<a name="72"></a> <a name="4"></a>
### 7.2 OCR和表格识别模型 ## 4. 快速体验
|模型名称|模型简介|模型大小|下载地址| 请参考[快速使用](./docs/quickstart.md)教程。
| --- | --- | --- | --- |
|ch_PP-OCRv3_det_slim|【最新】slim量化+蒸馏版超轻量模型，支持中英文、多语种文本检测| 1.1M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_distill_train.tar)|
|ch_PP-OCRv3_rec_slim |【最新】slim量化版超轻量模型，支持中英文、数字识别| 4.9M |[推理模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_train.tar) |
|ch_ppstructure_mobile_v2.0_SLANet|基于SLANet在PubTabNet数据集上训练的中文表格识别模型|9.3M|[推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_train.tar) |
<a name="5"></a>
## 5. 模型库
<a name="73"></a> 部分任务需要同时用到结构化分析模型和OCR模型，如表格识别需要使用表格识别模型进行结构化解析，同时也要用到OCR模型对表格内的文字进行识别，请根据具体需求选择合适的模型。
### 7.3 DocVQA 模型
|模型名称|模型简介|模型大小|下载地址| 结构化分析相关模型下载可以参考：
| --- | --- | --- | --- | - [PP-Structure 模型库](./docs/models_list.md)
|ser_LayoutXLM_xfun_zhd|基于LayoutXLM在xfun中文数据集上训练的SER模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|re_LayoutXLM_xfun_zh|基于LayoutXLM在xfun中文数据集上训练的RE模型|1.4G|[推理模型 coming soon]() / [训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
OCR相关模型下载可以参考：
- [PP-OCR 模型库](../doc/doc_ch/models_list.md)
更多模型下载，可以参考 [PP-OCR model_list](../doc/doc_ch/models_list.md) and [PP-Structure model_list](./docs/models_list.md)
ppstructure/docs/PP-Structurev2_introduction.md 0 → 100644
浏览文件 @ 96a8e0f8
# PP-Structurev2
## 目录
- [1. 背景](#1-背景)
- [2. 简介](#3-简介)
- [3. 整图方向矫正](#3-整图方向矫正)
- [4. 版面信息结构化](#4-版面信息结构化)
- [4.1 版面分析](#41-版面分析)
- [4.2 表格识别](#42-表格识别)
- [4.3 版面恢复](#43-版面恢复)
- [5. 关键信息抽取](#5-关键信息抽取)
- [6. Reference](#6-Reference)
## 1. 背景
现实场景中包含大量的文档图像，它们以图片等非结构化形式存储。基于文档图像的结构化分析与信息抽取对于数据的数字化存储以及产业的数字化转型至关重要。基于该考虑，PaddleOCR自研并发布了PP-Structure智能文档分析系统，旨在帮助开发者更好的完成版面分析、表格识别、关键信息抽取等文档理解相关任务。
近期，PaddleOCR团队针对PP-Structurev1的版面分析、表格识别、关键信息抽取模块，进行了共计8个方面的升级，同时新增整图方向矫正、文档复原等功能，打造出一个全新的、效果更优的文档分析系统：PP-Structurev2。
## 2. 简介
PP-Structurev2在PP-Structurev1的基础上进一步改进，主要有以下3个方面升级：
* **系统功能升级** ：新增图像矫正和版面复原模块，图像转word/pdf、关键信息抽取能力全覆盖！
* **系统性能优化** ：
* 版面分析：发布轻量级版面分析模型，速度提升**12倍**，平均CPU耗时仅需**70ms**！
* 表格识别：设计3大优化策略，预测耗时不变情况下，模型精度提升**6%**。
* 关键信息抽取：设计视觉无关模型结构，语义实体识别精度提升**2.8%**，关系抽取精度提升**9.1%**。
* **中文场景适配** ：完成对版面分析与表格识别的中文场景适配，开源**开箱即用**的中文场景版面结构化模型！
PP-Structurev2系统流程图如下所示，文档图像首先经过图像矫正模块，判断整图方向并完成转正，随后可以完成版面信息分析与关键信息抽取2类任务。版面分析任务中，图像首先经过版面分析模型，将图像划分为文本、表格、图像等不同区域，随后对这些区域分别进行识别，如，将表格区域送入表格识别模块进行结构化识别，将文本区域送入OCR引擎进行文字识别，最后使用版面恢复模块将其恢复为与原始图像布局一致的word或者pdf格式的文件；关键信息抽取任务中，首先使用OCR引擎提取文本内容，然后由语义实体识别模块获取图像中的语义实体，最后经关系抽取模块获取语义实体之间的对应关系，从而提取需要的关键信息。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185939247-57e53254-399c-46c4-a610-da4fa79232f5.png" width="1200">
</div>
从算法改进思路来看，对系统中的3个关键子模块，共进行了8个方面的改进。
* 版面分析
* PP-PicoDet：轻量级版面分析模型
* FGD：兼顾全局与局部特征的模型蒸馏算法
* 表格识别
* PP-LCNet: CPU友好型轻量级骨干网络
* CSP-PAN：轻量级高低层特征融合模块
* SLAHead：结构与位置信息对齐的特征解码模块
* 关键信息抽取
* VI-LayoutXLM：视觉特征无关的多模态预训练模型结构
* TB-YX：考虑阅读顺序的文本行排序逻辑
* UDML：联合互学习知识蒸馏策略
最终，与PP-Structurev1相比：
- 版面分析模型参数量减少95.6%，推理速度提升12倍，精度提升0.4%；
- 表格识别预测耗时不变，模型精度提升6%，端到端TEDS提升2%；
- 关键信息抽取模型速度提升2.8倍，语义实体识别模型精度提升2.8%；关系抽取模型精度提升9.1%。
下面对各个模块进行详细介绍。
## 3. 整图方向矫正
由于训练集一般以正方向图像为主，旋转过的文档图像直接输入模型会增加识别难度，影响识别效果。PP-Structurev2引入了整图方向矫正模块来判断含文字图像的方向，并将其进行方向调整。
我们直接调用PaddleClas中提供的文字图像方向分类模型-[PULC_text_image_orientation](https://github.com/PaddlePaddle/PaddleClas/blob/develop/docs/zh_CN/PULC/PULC_text_image_orientation.md)，该模型部分数据集图像如下所示。不同于文本行方向分类器，文字图像方向分类模型针对整图进行方向判别。文字图像方向分类模型在验证集上精度高达99%，单张图像CPU预测耗时仅为`2.16ms`。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185939683-f6465473-3303-4a0c-95be-51f04fb9f387.png" width="600">
</div>
## 4. 版面信息结构化
### 4.1 版面分析
版面分析指的是对图片形式的文档进行区域划分，定位其中的关键区域，如文字、标题、表格、图片等，PP-Structurev1使用了PaddleDetection中开源的高效检测算法PP-YOLOv2完成版面分析的任务。
在PP-Structurev2中，我们发布基于PP-PicoDet的轻量级版面分析模型，并针对版面分析场景定制图像尺度，同时使用FGD知识蒸馏算法，进一步提升模型精度。最终CPU上`41ms`即可完成版面分析过程(仅包含模型推理时间，数据预处理耗时大约50ms左右)。在公开数据集PubLayNet 上，消融实验如下：
| 实验序号 | 策略 | 模型存储(M) | mAP | CPU预测耗时(ms) |
|:------:|:------:|:------:|:------:|:------:|
| 1 | PP-YOLOv2(640*640) | 221 | 93.6% | 512 |
| 2 | PP-PicoDet-LCNet2.5x(640*640) | 29.7 | 92.5% |53.2|
| 3 | PP-PicoDet-LCNet2.5x(800*608) | 29.7 | 94.2% |83.1 |
| 4 | PP-PicoDet-LCNet1.0x(800*608) | 9.7 | 93.5% | 41.2|
| 5 | PP-PicoDet-LCNet1.0x(800*608) + FGD | 9.7 | 94% |41.2|
* 测试条件
* paddle版本：2.3.0
* CPU：Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz，开启mkldnn，线程数为10
在PubLayNet数据集上，与其他方法的性能对比如下表所示。可以看到，和基于Detectron2的版面分析工具layoutparser相比，我们的模型精度高出大约5%，预测速度快约69倍。
| 模型 | mAP | CPU预测耗时 |
|-------------------|-----------|------------|
| layoutparser (Detectron2) | 88.98% | 2.9s |
| PP-Structurev2 (PP-PicoDet) | **94%** | 41.2ms |
[PubLayNet](https://github.com/ibm-aur-nlp/PubLayNet)数据集是一个大型的文档图像数据集，包含Text、Title、Tale、Figure、List，共5个类别。数据集中包含335,703张训练集、11,245张验证集和11,405张测试集。训练数据与标注示例图如下所示：
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185940305-2cd3633b-4c43-4f84-8a6f-5ce6a24e88ce.png" width="600">
</div>
#### 4.1.1 优化策略
**（1）轻量级版面分析模型PP-PicoDet**
`PP-PicoDet`是PaddleDetection中提出的轻量级目标检测模型，通过使用PP-LCNet骨干网络、CSP-PAN特征融合模块、SimOTA标签分配方法等优化策略，最终在CPU与移动端具有卓越的性能。我们将PP-Structurev1中采用的PP-YOLOv2模型替换为`PP-PicoDet`，同时针对版面分析场景优化预测尺度，从针对目标检测设计的`640*640`调整为更适配文档图像的`800*608`，在`1.0x`配置下，模型精度与PP-YOLOv2相当，CPU平均预测速度可提升11倍。
**（1）FGD知识蒸馏**
FGD（Focal and Global Knowledge Distillation for Detectors），是一种兼顾局部全局特征信息的模型蒸馏方法，分为Focal蒸馏和Global蒸馏2个部分。Focal蒸馏分离图像的前景和背景，让学生模型分别关注教师模型的前景和背景部分特征的关键像素；Global蒸馏部分重建不同像素之间的关系并将其从教师转移到学生，以补偿Focal蒸馏中丢失的全局信息。我们基于FGD蒸馏策略，使用教师模型PP-PicoDet-LCNet2.5x（mAP=94.2%）蒸馏学生模型PP-PicoDet-LCNet1.0x（mAP=93.5%），可将学生模型精度提升0.5%，和教师模型仅差0.2%，而预测速度比教师模型快1倍。
#### 4.1.2 场景适配
**（1）中文版面分析**
除了英文公开数据集PubLayNet，我们也在中文场景进行了场景适配与方法验证。[CDLA](https://github.com/buptlihang/CDLA)是一个中文文档版面分析数据集，面向中文文献类（论文）场景，包含正文、标题等10个label。数据集中包含5,000张训练集和1,000张验证集。训练数据与标注示例图如下所示：
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185940445-92b7613f-e431-43c2-9033-3b3618ddae02.png" width="600">
</div>
在CDLA 数据集上，消融实验如下：
| 实验序号 | 策略 | mAP |
|:------:|:------:|:------:|
| 1 | PP-YOLOv2 | 84.7% |
| 2 | PP-PicoDet-LCNet2.5x(800*608) | 87.8% |
| 3 | PP-PicoDet-LCNet1.0x(800*608) | 84.5% |
| 4 | PP-PicoDet-LCNet1.0x(800*608) + FGD | 86.8% |
**（2）表格版面分析**
在实际应用中，很多场景并不关注图像中的图片、文本等版面区域，而仅需要提取文档图像中的表格，此时版面分析任务退化为一个表格检测任务，表格检测往往也是表格识别的前序任务。面向中英文文档场景，我们整理了开源领域含表格的版面分析数据集，包括TableBank、DocBank等。融合后的数据集中包含496,405张训练集与9,495张验证集图像。
在表格数据集上，消融实验如下：
| 实验序号 | 策略 | mAP |
|:------:|:------:|:------:|
| 1 | PP-YOLOv2 |91.3% |
| 2 | PP-PicoDet-LCNet2.5x(800*608) | 95.9% |
| 3 | PP-PicoDet-LCNet1.0x(800*608) | 95.2% |
| 4 | PP-PicoDet-LCNet1.0x(800*608) + FGD | 95.7% |
表格检测效果示意图如下：
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185940654-956ef614-888a-4779-bf63-a6c2b61b97fa.png" width="600">
</div>
### 4.2 表格识别
基于深度学习的表格识别算法种类丰富，PP-Structurev1中，我们基于文本识别算法RARE研发了端到端表格识别算法TableRec-RARE，模型输出为表格结构的HTML表示，进而可以方便地转化为Excel文件。PP-Structurev2中，我们对模型结构和损失函数等5个方面进行升级，提出了 SLANet (Structure Location Alignment Network) ，模型结构如下图所示：
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185940811-089c9265-4be9-4776-b365-6d1125606b4b.png" width="1200">
</div>
在PubTabNet英文表格识别数据集上的消融实验如下：
|策略|Acc|TEDS|推理速度(CPU+MKLDNN)|模型大小|
|---|---|---|---|---|
|TableRec-RARE| 71.73% | 93.88% |779ms |6.8M|
|+PP-LCNet| 74.71% |94.37% |778ms| 8.7M|
|+CSP-PAN| 75.68%| 94.72% |708ms| 9.3M|
|+SLAHead| 77.7%|94.85%| 766ms| 9.2M|
|+MergeToken| 76.31%| 95.89%|766ms| 9.2M|
* 测试环境
* paddle版本：2.3.1
* CPU：Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz，开启mkldnn，线程数为10
在PubtabNet英文表格识别数据集上，和其他方法对比如下：
|策略|Acc|TEDS|推理速度(CPU+MKLDNN)|模型大小|
|---|---|---|---|---|
|TableMaster|77.9%|96.12%|2144ms|253M|
|TableRec-RARE| 71.73% | 93.88% |779ms |6.8M|
|SLANet|76.31%| 95.89%|766ms|9.2M|
#### 4.2.1 优化策略
**（1） CPU友好型轻量级骨干网络PP-LCNet**
PP-LCNet是结合Intel-CPU端侧推理特性而设计的轻量高性能骨干网络，该方案在图像分类任务上取得了比ShuffleNetV2、MobileNetV3、GhostNet等轻量级模型更优的“精度-速度”均衡。PP-Structurev2中，我们采用PP-LCNet作为骨干网络，表格识别模型精度从71.73%提升至72.98%；同时加载通过SSLD知识蒸馏方案训练得到的图像分类模型权重作为表格识别的预训练模型，最终精度进一步提升2.95%至74.71%。
**（2）轻量级高低层特征融合模块CSP-PAN**
对骨干网络提取的特征进行融合，可以有效解决尺度变化较大等复杂场景中的模型预测问题。早期，FPN模块被提出并用于特征融合，但是它的特征融合过程仅包含单向（高->低），融合不够充分。CSP-PAN基于PAN进行改进，在保证特征融合更为充分的同时，使用CSP block、深度可分离卷积等策略减小了计算量。在表格识别场景中，我们进一步将CSP-PAN的通道数从128降低至96以降低模型大小。最终表格识别模型精度提升0.97%至75.68%，预测速度提升10%。
**（3）结构与位置信息对齐的特征解码模块SLAHead**
TableRec-RARE的TableAttentionHead如下图a所示，TableAttentionHead在执行完全部step的计算后拿到最终隐藏层状态表征(hiddens)，随后hiddens经由SDM(Structure Decode Module)和CLDM(Cell Location Decode Module)模块生成全部的表格结构token和单元格坐标。但是这种设计忽略了单元格token和坐标之间一一对应的关系。
PP-Structurev2中，我们设计SLAHead模块，对单元格token和坐标之间做了对齐操作，如下图b所示。在SLAHead中，每一个step的隐藏层状态表征会分别送入SDM和CLDM来得到当前step的token和坐标，每个step的token和坐标输出分别进行concat得到表格的html表达和全部单元格的坐标。此外，考虑到表格识别模型的单元格准确率依赖于表格结构的识别准确，我们将损失函数中表格结构分支与单元格定位分支的权重比从1:1提升到8:1，并使用收敛更稳定的Smoothl1 Loss替换定位分支中的MSE Loss。最终模型精度从75.68%提高至77.7%。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185940968-e3a2fbac-78d7-4b74-af54-a1dab860f470.png" width="1200">
</div>
**（4）其他**
TableRec-RARE算法中，我们使用`<td>`和`</td>`两个单独的token来表示一个非跨行列单元格，这种表示方式限制了网络对于单元格数量较多表格的处理能力。
PP-Structurev2中，我们参考TableMaster中的token处理方法，将`<td>`和`</td>`合并为一个token-`<td></td>`。合并token后，验证集中token长度大于500的图片也参与模型评估，最终模型精度降低为76.31%，但是端到端TEDS提升1.04%。
#### 4.2.2 中文场景适配
除了上述模型策略的升级外，本次升级还开源了中文表格识别模型。在实际应用场景中，表格图像存在着各种各样的倾斜角度（PubTabNet数据集不存在该问题），因此在中文模型中，我们将单元格坐标回归的点数从2个（左上，右下）增加到4个(左上，右上，右下，左下)。在内部测试集上，模型升级前后指标如下：
|模型|acc|
|---|---|
|TableRec-RARE|44.3%|
|SLANet|59.35%|
可视化结果如下，左为输入图像，右为识别的html表格
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185941221-c94e3d45-524c-4073-9644-21ba6a9fd93e.png" width="800">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185941254-31f2b1fa-d594-4037-b1c7-0f24543e5d19.png" width="800">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185941273-2f7131df-3fe7-43b8-9c64-77ad2cf3b947.png" width="800">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185941295-d0672aa8-548d-4e6a-812c-ac5d5fd8a269.png" width="800">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185941324-8036e959-abdc-4dc5-a9f2-730b07b4e3d3.png" width="800">
</div>
### 4.3 版面恢复
版面恢复指的是文档图像经过OCR识别、版面分析、表格识别等方法处理后的内容可以与原始文档保持相同的排版方式，并输出到word等文档中。PP-Structurev2中，我们版面恢复系统，包含版面分析、表格识别、OCR文本检测与识别等子模块。
下图展示了版面恢复的结果：
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185941816-4dabb3e8-a0db-4094-98ea-52e0a0fda8e8.png" width="1200">
</div>
## 5. 关键信息抽取
关键信息抽取指的是针对文档图像的文字内容，提取出用户关注的关键信息，如身份证中的姓名、住址等字段。PP-Structure中支持了基于多模态LayoutLM系列模型的语义实体识别 (Semantic Entity Recognition, SER) 以及关系抽取 (Relation Extraction, RE) 任务。PP-Structurev2中，我们对模型结构以及下游任务训练方法进行升级，提出了VI-LayoutXLM（Visual-feature Independent LayoutXLM），具体流程图如下所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185941978-abec7d4a-5e3a-4141-83f8-088d04ef898e.png" width="1000">
</div>
具体优化策略包括：
* VI-LayoutXLM：视觉特征无关的多模态预训练模型结构
* TB-YX：考虑人类阅读顺序的文本行排序逻辑
* UDML：联合互学习知识蒸馏策略
XFUND-zh数据集上，SER任务的消融实验如下所示。
| 实验序号 | 策略 | 模型大小(G) | 精度 | GPU预测耗时(ms) | CPU预测耗时(ms) |
|:------:|:------:|:------:|:------:|:------:|:------:|
| 1 | LayoutXLM | 1.4 | 89.50% | 59.35 | 766.16 |
| 2 | VI-LayoutXLM | 1.1 | 90.46% | 23.71 | 675.58 |
| 3 | 实验2 + TB-YX文本行排序 | 1.1 | 92.50% | 23.71 | 675.58 |
| 4 | 实验3 + UDML蒸馏 | 1.1 | 93.19% | 23.71 | 675.58 |
| 5 | 实验3 + UDML蒸馏 | 1.1 | **93.19%** | **15.49** | **675.58** |
* 测试条件
* paddle版本：2.3.0
* GPU：V100，实验5的GPU预测耗时使用`trt+fp16`测试得到，环境为cuda10.2+ cudnn8.1.1 + trt7.2.3.4，其他实验的预测耗时统计中没有使用TRT。
* CPU：Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz，开启mkldnn，线程数为10
在XFUND数据集上，与其他方法的效果对比如下所示。
| 模型 | SER Hmean | RE Hmean |
|-------------------|-----------|------------|
| LayoutLMv2-base | 85.44% | 67.77% |
| LayoutXLM-base | 89.24% | 70.73% |
| StrucTexT-large | 92.29% | **86.81%** |
| VI-LayoutXLM-base (ours) | **93.19%** | 83.92% |
### 5.1 优化策略
**(1) VI-LayoutXLM（Visual-feature Independent LayoutXLM）**
LayoutLMv2以及LayoutXLM中引入视觉骨干网络，用于提取视觉特征，并与后续的text embedding进行联合，作为多模态的输入embedding。但是该模块为基于`ResNet_x101_64x4d`的特征提取网络，特征抽取阶段耗时严重，因此我们将其去除，同时仍然保留文本、位置以及布局等信息，最终发现针对LayoutXLM进行改进，下游SER任务精度无损，针对LayoutLMv2进行改进，下游SER任务精度仅降低`2.1%`，而模型大小减小了约`340M`。具体消融实验如下所示。
| 模型 | 模型大小 (G) | F-score | 精度收益 |
|-----------------|----------|---------|--------|
| LayoutLMv2 | 0.76 | 84.20% | - |
| VI-LayoutLMv2 | 0.42 | 82.10% | -2.10% |
| LayoutXLM | 1.4 | 89.50% | - |
| VI-LayouXLM | 1.1 | 90.46% | +0.96% |
同时，基于XFUND数据集，VI-LayoutXLM在RE任务上的精度也进一步提升了`1.06%`。
**(2) TB-YX排序方法（Threshold-Based YX sorting algorithm）**
文本阅读顺序对于信息抽取与文本理解等任务至关重要，传统多模态模型中，没有考虑不同OCR工具可能产生的不正确阅读顺序，而模型输入中包含位置编码，阅读顺序会直接影响预测结果，在预处理中，我们对文本行按照从上到下，从左到右（YX）的顺序进行排序，为防止文本行位置轻微干扰带来的排序结果不稳定问题，在排序的过程中，引入位置偏移阈值Th，对于Y方向距离小于Th的2个文本内容，使用x方向的位置从左到右进行排序。TB-YX排序方法伪代码如下所示。
```py
def order_by_tbyx(ocr_info, th=20):
"""
ocr_info: a list of dict, which contains bbox information([x1, y1, x2, y2])
th: threshold of the position threshold
"""
res = sorted(ocr_info, key=lambda r: (r["bbox"][1], r["bbox"][0])) # sort using y1 first and then x1
for i in range(len(res) - 1):
for j in range(i, 0, -1):
# restore the order using the
if abs(res[j + 1]["bbox"][1] - res[j]["bbox"][1]) < th and \
(res[j + 1]["bbox"][0] < res[j]["bbox"][0]):
tmp = deepcopy(res[j])
res[j] = deepcopy(res[j + 1])
res[j + 1] = deepcopy(tmp)
else:
break
return res
```
不同排序方法的结果对比如下所示，可以看出引入偏离阈值之后，排序结果更加符合人类的阅读顺序。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185942080-9d4bafc9-fa7f-4da4-b139-b2bd703dc76d.png" width="800">
</div>
使用该策略，最终XFUND数据集上，SER任务F1指标提升`2.06%`，RE任务F1指标提升`7.04%`。
**(3) 互学习蒸馏策略**
UDML（Unified-Deep Mutual Learning）联合互学习是PP-OCRv2与PP-OCRv3中采用的对于文本识别非常有效的提升模型效果的策略。在训练时，引入2个完全相同的模型进行互学习，计算2个模型之间的互蒸馏损失函数(DML loss)，同时对transformer中间层的输出结果计算距离损失函数(L2 loss)。使用该策略，最终XFUND数据集上，SER任务F1指标提升`0.6%`，RE任务F1指标提升`5.01%`。
最终优化后模型基于SER任务的可视化结果如下所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185942213-0909135b-3bcd-4d79-9e69-847dfb1c3b82.png" width="800">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185942237-72923b42-8590-42eb-b687-fa819b1c3afd.png" width="800">
</div>
RE任务的可视化结果如下所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185942400-8920dc3c-de7f-46d0-b0bc-baca9536e0e1.png" width="800">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185942416-ca4fd8b0-9227-4c65-b969-0afbda525b85.png" width="800">
</div>
### 5.2 更多场景消融实验
我们在FUNSD数据集上，同时基于RE任务进行对本次升级策略进行验证，具体实验结果如下所示，可以看出该方案针对不同任务，在不同数据集上均有非常明显的精度收益。
#### 5.2.1 XFUND_zh数据集
**RE任务结果**
| 实验序号 | 策略 | 模型大小(G) | F1-score |
|:------:|:------------:|:---------:|:----------:|
| 1 | LayoutXLM | 1.4 | 70.81% |
| 2 | VI-LayoutXLM | 1.1 | 71.87% |
| 3 | 实验2 + PP-OCR排序 | 1.1 | 78.91% |
| 4 | 实验3 + UDML蒸馏 | 1.1 | **83.92%** |
#### 5.2.2 FUNSD数据集
**SER任务结果**
| 实验序号 | 策略 | F1-score |
|:------:|:------:|:------:|
| 1 | LayoutXLM | 82.28% |
| 2 | PP-Structurev2 SER | **87.79%** |
**RE任务结果**
| 实验序号 | 策略 | F1-score |
|:------:|:------:|:------:|
| 1 | LayoutXLM | 53.13% |
| 2 | PP-Structurev2 SER | **74.87%** |
## 6. Reference
* [1] Zhong X, ShafieiBavani E, Jimeno Yepes A. Image-based table recognition: data, model, and evaluation[C]//European Conference on Computer Vision. Springer, Cham, 2020: 564-580.
* [2] Cui C, Gao T, Wei S. Yuning Du, Ruoyu Guo, Shuilong Dong, Bin Lu, Ying Zhou, Xueying Lv, Qiwen Liu, Xiaoguang Hu, Dianhai Yu, and Yanjun Ma* [J]. Pplcnet: A lightweight cpu convolutional neural network, 2021, 3.
* [3] Lin T Y, Dollár P, Girshick R, et al. Feature pyramid networks for object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 2117-2125.
* [4] Yu G, Chang Q, Lv W, et al. PP-PicoDet: A Better Real-Time Object Detector on Mobile Devices[J]. arXiv preprint arXiv:2111.00902, 2021.
* [5] Bochkovskiy A, Wang C Y, Liao H Y M. Yolov4: Optimal speed and accuracy of object detection[J]. arXiv preprint arXiv:2004.10934, 2020.
* [6] Ye J, Qi X, He Y, et al. PingAn-VCGroup's Solution for ICDAR 2021 Competition on Scientific Literature Parsing Task B: Table Recognition to HTML[J]. arXiv preprint arXiv:2105.01848, 2021.
* [7] Zhong X, Tang J, Yepes A J. Publaynet: largest dataset ever for document layout analysis[C]//2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019: 1015-1022.
* [8] CDLA：https://github.com/buptlihang/CDLA
* [9]Gao L, Huang Y, Déjean H, et al. ICDAR 2019 competition on table detection and recognition (cTDaR)[C]//2019 International Conference on Document Analysis and Recognition (ICDAR). IEEE, 2019: 1510-1515.
* [10] Mondal A, Lipps P, Jawahar C V. IIIT-AR-13K: a new dataset for graphical object detection in documents[C]//International Workshop on Document Analysis Systems. Springer, Cham, 2020: 216-230.
* [11] Tal ocr_tabel：https://ai.100tal.com/dataset
* [12] Li M, Cui L, Huang S, et al. Tablebank: A benchmark dataset for table detection and recognition[J]. arXiv preprint arXiv:1903.01949, 2019.
* [13]Li M, Xu Y, Cui L, et al. DocBank: A benchmark dataset for document layout analysis[J]. arXiv preprint arXiv:2006.01038, 2020.
* [14] Xu Y, Li M, Cui L, et al. Layoutlm: Pre-training of text and layout for document image understanding[C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020: 1192-1200.
* [15] Xu Y, Xu Y, Lv T, et al. LayoutLMv2: Multi-modal pre-training for visually-rich document understanding[J]. arXiv preprint arXiv:2012.14740, 2020.
* [16] Xu Y, Lv T, Cui L, et al. Layoutxlm: Multimodal pre-training for multilingual visually-rich document understanding[J]. arXiv preprint arXiv:2104.08836, 2021.
* [17] Xu Y, Lv T, Cui L, et al. XFUND: A Benchmark Dataset for Multilingual Visually Rich Form Understanding[C]//Findings of the Association for Computational Linguistics: ACL 2022. 2022: 3214-3224.
* [18] Jaume G, Ekenel H K, Thiran J P. Funsd: A dataset for form understanding in noisy scanned documents[C]//2019 International Conference on Document Analysis and Recognition Workshops (ICDARW). IEEE, 2019, 2: 1-6.
ppstructure/docs/inference.md
浏览文件 @ 96a8e0f8
# 基于Python预测引擎推理 # 基于Python预测引擎推理
- [1. Structure](#1) - [1. 版面信息抽取](#1)
- [1.1 版面分析+表格识别](#1.1) - [1.1 版面分析+表格识别](#1.1)
- [1.2 版面分析](#1.2) - [1.2 版面分析](#1.2)
- [1.3 表格识别](#1.3) - [1.3 表格识别](#1.3)
- [2. DocVQA](#2) - [2. 关键信息抽取](#2)
<a name="1"></a> <a name="1"></a>
## 1. Structure ## 1. 版面信息抽取
进入`ppstructure`目录 进入`ppstructure`目录
```bash ```bash
cd ppstructure cd ppstructure
```` ```
下载模型 下载模型
```bash ```bash
mkdir inference && cd inference mkdir inference && cd inference
# 下载PP-OCRv2文本检测模型并解压 # 下载PP-Structurev2版面分析模型并解压
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar && tar xf ch_PP-OCRv2_det_slim_quant_infer.tar wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_layout_infer.tar && tar xf picodet_lcnet_x1_0_layout_infer.tar
# 下载PP-OCRv2文本识别模型并解压 # 下载PP-OCRv3文本检测模型并解压
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar && tar xf ch_PP-OCRv2_rec_slim_quant_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar && tar xf ch_PP-OCRv3_det_infer.tar
# 下载超轻量级英文表格预测模型并解压 # 下载PP-OCRv3文本识别模型并解压
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar && tar xf en_ppocr_mobile_v2.0_table_structure_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar && tar xf ch_PP-OCRv3_rec_infer.tar
# 下载PP-Structurev2表格识别模型并解压
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf ch_ppstructure_mobile_v2.0_SLANet_infer.tar
cd .. cd ..
``` ```
<a name="1.1"></a> <a name="1.1"></a>
### 1.1 版面分析+表格识别 ### 1.1 版面分析+表格识别
```bash ```bash
python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_infer \ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv3_det_infer \
--rec_model_dir=inference/ch_PP-OCRv2_rec_slim_quant_infer \ --rec_model_dir=inference/ch_PP-OCRv3_rec_infer \
--table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer \ --table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \
--layout_model_dir=inference/picodet_lcnet_x1_0_layout_infer \
--image_dir=./docs/table/1.png \ --image_dir=./docs/table/1.png \
--rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \ --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \ --table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \
--output=../output \ --output=../output \
--vis_font_path=../doc/fonts/simfang.ttf --vis_font_path=../doc/fonts/simfang.ttf
``` ```
...@@ -41,19 +44,23 @@ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_i ...@@ -41,19 +44,23 @@ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_i
<a name="1.2"></a> <a name="1.2"></a>
### 1.2 版面分析 ### 1.2 版面分析
```bash ```bash
python3 predict_system.py --image_dir=./docs/table/1.png --table=false --ocr=false --output=../output/ python3 predict_system.py --layout_model_dir=inference/picodet_lcnet_x1_0_layout_infer \
--image_dir=./docs/table/1.png \
--output=../output \
--table=false \
--ocr=false
``` ```
运行完成后，每张图片会在`output`字段指定的目录下的`structure`目录下有一个同名目录，图片区域会被裁剪之后保存下来，图片名为表格在图片里的坐标。版面分析结果会存储在`res.txt`文件中。 运行完成后，每张图片会在`output`字段指定的目录下的`structure`目录下有一个同名目录，图片区域会被裁剪之后保存下来，图片名为表格在图片里的坐标。版面分析结果会存储在`res.txt`文件中。
<a name="1.3"></a> <a name="1.3"></a>
### 1.3 表格识别 ### 1.3 表格识别
```bash ```bash
python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_infer \ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv3_det_infer \
--rec_model_dir=inference/ch_PP-OCRv2_rec_slim_quant_infer \ --rec_model_dir=inference/ch_PP-OCRv3_rec_infer \
--table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer \ --table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \
--image_dir=./docs/table/table.jpg \ --image_dir=./docs/table/table.jpg \
--rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \ --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \ --table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \
--output=../output \ --output=../output \
--vis_font_path=../doc/fonts/simfang.ttf \ --vis_font_path=../doc/fonts/simfang.ttf \
--layout=false --layout=false
...@@ -61,20 +68,22 @@ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_i ...@@ -61,20 +68,22 @@ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_i
运行完成后，每张图片会在`output`字段指定的目录下的`structure`目录下有一个同名目录，表格会存储为一个excel，excel文件名为`[0,0,img_h,img_w]`。 运行完成后，每张图片会在`output`字段指定的目录下的`structure`目录下有一个同名目录，表格会存储为一个excel，excel文件名为`[0,0,img_h,img_w]`。
<a name="2"></a> <a name="2"></a>
## 2. DocVQA ## 2. 关键信息抽取
```bash ```bash
cd ppstructure cd ppstructure
# 下载模型
mkdir inference && cd inference mkdir inference && cd inference
# 下载SER xfun 模型并解压 # 下载SER XFUND 模型并解压
wget https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_ser_pretrained.tar && tar xf PP-Layout_v1.0_ser_pretrained.tar wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_infer.tar && tar -xf ser_vi_layoutxlm_xfund_infer.tar
cd .. cd ..
python3 kie/predict_kie_token_ser.py \
python3 predict_system.py --model_name_or_path=vqa/PP-Layout_v1.0_ser_pretrained/ \ --kie_algorithm=LayoutXLM \
--mode=vqa \ --ser_model_dir=../inference/ser_vi_layoutxlm_xfund_infer \
--image_dir=vqa/images/input/zh_val_0.jpg \ --image_dir=./docs/kie/input/zh_val_42.jpg \
--vis_font_path=../doc/fonts/simfang.ttf --ser_dict_path=../ppocr/utils/dict/kie_dict/xfund_class_list.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx"
``` ```
运行完成后，每张图片会在`output`字段指定的目录下的`vqa`目录下存放可视化之后的图片，图片名和输入图片名一致。
运行完成后，每张图片会在`output`字段指定的目录下的`kie`目录下存放可视化之后的图片，图片名和输入图片名一致。
ppstructure/docs/inference_en.md
浏览文件 @ 96a8e0f8
# Python Inference # Python Inference
- [1. Structure](#1) - [1. Layout Structured Analysis](#1)
- [1.1 layout analysis + table recognition](#1.1) - [1.1 layout analysis + table recognition](#1.1)
- [1.2 layout analysis](#1.2) - [1.2 layout analysis](#1.2)
- [1.3 table recognition](#1.3) - [1.3 table recognition](#1.3)
- [2. DocVQA](#2) - [2. Key Information Extraction](#2)
<a name="1"></a> <a name="1"></a>
## 1. Structure ## 1. Layout Structured Analysis
Go to the `ppstructure` directory Go to the `ppstructure` directory
```bash ```bash
...@@ -18,23 +18,26 @@ download model ...@@ -18,23 +18,26 @@ download model
```bash ```bash
mkdir inference && cd inference mkdir inference && cd inference
# Download the PP-OCRv2 text detection model and unzip it # Download the PP-Structurev2 layout analysis model and unzip it
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar && tar xf ch_PP-OCRv2_det_slim_quant_infer.tar wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_layout_infer.tar && tar xf picodet_lcnet_x1_0_layout_infer.tar
# Download the PP-OCRv2 text recognition model and unzip it # Download the PP-OCRv3 text detection model and unzip it
wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar && tar xf ch_PP-OCRv2_rec_slim_quant_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar && tar xf ch_PP-OCRv3_det_infer.tar
# Download the ultra-lightweight English table structure model and unzip it # Download the PP-OCRv3 text recognition model and unzip it
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar && tar xf en_ppocr_mobile_v2.0_table_structure_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar && tar xf ch_PP-OCRv3_rec_infer.tar
# Download the PP-Structurev2 form recognition model and unzip it
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf ch_ppstructure_mobile_v2.0_SLANet_infer.tar
cd .. cd ..
``` ```
<a name="1.1"></a> <a name="1.1"></a>
### 1.1 layout analysis + table recognition ### 1.1 layout analysis + table recognition
```bash ```bash
python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_infer \ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv3_det_infer \
--rec_model_dir=inference/ch_PP-OCRv2_rec_slim_quant_infer \ --rec_model_dir=inference/ch_PP-OCRv3_rec_infer \
--table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer \ --table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \
--layout_model_dir=inference/picodet_lcnet_x1_0_layout_infer \
--image_dir=./docs/table/1.png \ --image_dir=./docs/table/1.png \
--rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \ --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \ --table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \
--output=../output \ --output=../output \
--vis_font_path=../doc/fonts/simfang.ttf --vis_font_path=../doc/fonts/simfang.ttf
``` ```
...@@ -43,19 +46,23 @@ After the operation is completed, each image will have a directory with the same ...@@ -43,19 +46,23 @@ After the operation is completed, each image will have a directory with the same
<a name="1.2"></a> <a name="1.2"></a>
### 1.2 layout analysis ### 1.2 layout analysis
```bash ```bash
python3 predict_system.py --image_dir=./docs/table/1.png --table=false --ocr=false --output=../output/ python3 predict_system.py --layout_model_dir=inference/picodet_lcnet_x1_0_layout_infer \
--image_dir=./docs/table/1.png \
--output=../output \
--table=false \
--ocr=false
``` ```
After the operation is completed, each image will have a directory with the same name in the `structure` directory under the directory specified by the `output` field. Each picture in image will be cropped and saved. The filename of picture area is their coordinates in the image. Layout analysis results will be stored in the `res.txt` file After the operation is completed, each image will have a directory with the same name in the `structure` directory under the directory specified by the `output` field. Each picture in image will be cropped and saved. The filename of picture area is their coordinates in the image. Layout analysis results will be stored in the `res.txt` file
<a name="1.3"></a> <a name="1.3"></a>
### 1.3 table recognition ### 1.3 table recognition
```bash ```bash
python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_infer \ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv3_det_infer \
--rec_model_dir=inference/ch_PP-OCRv2_rec_slim_quant_infer \ --rec_model_dir=inference/ch_PP-OCRv3_rec_infer \
--table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer \ --table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \
--image_dir=./docs/table/table.jpg \ --image_dir=./docs/table/table.jpg \
--rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \ --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \ --table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \
--output=../output \ --output=../output \
--vis_font_path=../doc/fonts/simfang.ttf \ --vis_font_path=../doc/fonts/simfang.ttf \
--layout=false --layout=false
...@@ -63,19 +70,22 @@ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_i ...@@ -63,19 +70,22 @@ python3 predict_system.py --det_model_dir=inference/ch_PP-OCRv2_det_slim_quant_i
After the operation is completed, each image will have a directory with the same name in the `structure` directory under the directory specified by the `output` field. Each table in the image will be stored as an excel. The filename of excel is their coordinates in the image. After the operation is completed, each image will have a directory with the same name in the `structure` directory under the directory specified by the `output` field. Each table in the image will be stored as an excel. The filename of excel is their coordinates in the image.
<a name="2"></a> <a name="2"></a>
## 2. DocVQA ## 2. Key Information Extraction
```bash ```bash
cd ppstructure cd ppstructure
# download model
mkdir inference && cd inference mkdir inference && cd inference
wget https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_ser_pretrained.tar && tar xf PP-Layout_v1.0_ser_pretrained.tar # download model
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_infer.tar && tar -xf ser_vi_layoutxlm_xfund_infer.tar
cd .. cd ..
python3 kie/predict_kie_token_ser.py \
python3 predict_system.py --model_name_or_path=vqa/PP-Layout_v1.0_ser_pretrained/ \ --kie_algorithm=LayoutXLM \
--mode=vqa \ --ser_model_dir=../inference/ser_vi_layoutxlm_xfund_infer \
--image_dir=vqa/images/input/zh_val_0.jpg \ --image_dir=./docs/kie/input/zh_val_42.jpg \
--vis_font_path=../doc/fonts/simfang.ttf --ser_dict_path=../ppocr/utils/dict/kie_dict/xfund_class_list.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx"
``` ```
After the operation is completed, each image will store the visualized image in the `vqa` directory under the directory specified by the `output` field, and the image name is the same as the input image name.
After the operation is completed, each image will store the visualized image in the `kie` directory under the directory specified by the `output` field, and the image name is the same as the input image name.
ppstructure/docs/installation.md 已删除 100644 → 0
浏览文件 @ 29268eaa
- [快速安装](#快速安装)
- [1. PaddlePaddle 和 PaddleOCR](#1-paddlepaddle-和-paddleocr)
- [2. 安装其他依赖](#2-安装其他依赖)
- [2.1 VQA所需依赖](#21--vqa所需依赖)
# 快速安装
## 1. PaddlePaddle 和 PaddleOCR
可参考[PaddleOCR安装文档](../../doc/doc_ch/installation.md)
## 2. 安装其他依赖
### 2.1 VQA所需依赖
* paddleocr
```bash
pip3 install paddleocr
```
* PaddleNLP
```bash
git clone https://github.com/PaddlePaddle/PaddleNLP -b develop
cd PaddleNLP
pip3 install -e .
```
ppstructure/docs/installation_en.md 已删除 100644 → 0
浏览文件 @ 29268eaa
# Quick installation
- [1. PaddlePaddle 和 PaddleOCR](#1)
- [2. Install other dependencies](#2)
- [2.1 VQA](#21)
<a name="1"></a>
## 1. PaddlePaddle and PaddleOCR
Please refer to [PaddleOCR installation documentation](../../doc/doc_en/installation_en.md)
<a name="2"></a>
## 2. Install other dependencies
<a name="21"></a>
### 2.1 VQA
* paddleocr
```bash
pip3 install paddleocr
```
* PaddleNLP
```bash
git clone https://github.com/PaddlePaddle/PaddleNLP -b develop
cd PaddleNLP
pip3 install -e .
```
ppstructure/docs/models_list.md
浏览文件 @ 96a8e0f8
...@@ -10,13 +10,17 @@ ...@@ -10,13 +10,17 @@
<a name="1"></a> <a name="1"></a>
## 1. 版面分析模型 ## 1. 版面分析模型
|模型名称|模型简介|下载地址|label_map| |模型名称|模型简介|推理模型大小|下载地址|dict path|
| --- | --- | --- | --- | | --- | --- | --- | --- | --- |
| ppyolov2_r50vd_dcn_365e_publaynet | PubLayNet 数据集训练的版面分析模型，可以划分**文字、标题、表格、图片以及列表**5类区域 | [推理模型](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet.tar) / [训练模型](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet_pretrained.pdparams) |{0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"}| | picodet_lcnet_x1_0_fgd_layout | 基于PicoDet LCNet_x1_0和FGD蒸馏在PubLayNet 数据集训练的英文版面分析模型，可以划分**文字、标题、表格、图片以及列表**5类区域 | 9.7M | [推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout.pdparams) | [PubLayNet dict](../../ppocr/utils/dict/layout_dict/layout_publaynet_dict.txt) |
| ppyolov2_r50vd_dcn_365e_tableBank_word | TableBank Word 数据集训练的版面分析模型，只能检测表格 | [推理模型](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_tableBank_word.tar) | {0:"Table"}| | ppyolov2_r50vd_dcn_365e_publaynet | 基于PP-YOLOv2在PubLayNet数据集上训练的英文版面分析模型 | 221M | [推理模型](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet.tar) / [训练模型](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet_pretrained.pdparams) | 同上 |
| ppyolov2_r50vd_dcn_365e_tableBank_latex | TableBank Latex 数据集训练的版面分析模型，只能检测表格 | [推理模型](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_tableBank_latex.tar) | {0:"Table"}| | picodet_lcnet_x1_0_fgd_layout_cdla | CDLA数据集训练的中文版面分析模型，可以划分为**表格、图片、图片标题、表格、表格标题、页眉、脚本、引用、公式**10类区域 | 9.7M | [推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_cdla_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_cdla.pdparams) | [CDLA dict](../../ppocr/utils/dict/layout_dict/layout_cdla_dict.txt) |
| picodet_lcnet_x1_0_fgd_layout_table | 表格数据集训练的版面分析模型，支持中英文文档表格区域的检测 | 9.7M | [推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_table_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_table.pdparams) | [Table dict](../../ppocr/utils/dict/layout_dict/layout_table_dict.txt) |
| ppyolov2_r50vd_dcn_365e_tableBank_word | 基于PP-YOLOv2在TableBank Word 数据集训练的版面分析模型，支持英文文档表格区域的检测 | 221M | [推理模型](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_tableBank_word.tar) | 同上 |
| ppyolov2_r50vd_dcn_365e_tableBank_latex | 基于PP-YOLOv2在TableBank Latex数据集训练的版面分析模型，支持英文文档表格区域的检测 | 221M | [推理模型](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_tableBank_latex.tar) | 同上 |
<a name="2"></a> <a name="2"></a>
## 2. OCR和表格识别模型 ## 2. OCR和表格识别模型
<a name="21"></a> <a name="21"></a>
...@@ -24,8 +28,8 @@ ...@@ -24,8 +28,8 @@
|模型名称|模型简介|推理模型大小|下载地址| |模型名称|模型简介|推理模型大小|下载地址|
| --- | --- | --- | --- | | --- | --- | --- | --- |
|en_ppocr_mobile_v2.0_table_det|PubLayNet数据集训练的英文表格场景的文字检测|4.7M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_det_train.tar) | |en_ppocr_mobile_v2.0_table_det|PubTabNet数据集训练的英文表格场景的文字检测|4.7M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_det_train.tar) |
|en_ppocr_mobile_v2.0_table_rec|PubLayNet数据集训练的英文表格场景的文字识别|6.9M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_rec_train.tar) | |en_ppocr_mobile_v2.0_table_rec|PubTabNet数据集训练的英文表格场景的文字识别|6.9M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_rec_train.tar) |
如需要使用其他OCR模型，可以在 [PP-OCR model_list](../../doc/doc_ch/models_list.md) 下载模型或者使用自己训练好的模型配置到 `det_model_dir`, `rec_model_dir`两个字段即可。 如需要使用其他OCR模型，可以在 [PP-OCR model_list](../../doc/doc_ch/models_list.md) 下载模型或者使用自己训练好的模型配置到 `det_model_dir`, `rec_model_dir`两个字段即可。
...@@ -34,9 +38,9 @@ ...@@ -34,9 +38,9 @@
|模型名称|模型简介|推理模型大小|下载地址| |模型名称|模型简介|推理模型大小|下载地址|
| --- | --- | --- | --- | | --- | --- | --- | --- |
|en_ppocr_mobile_v2.0_table_structure|基于TableRec-RARE在PubTabNet数据集上训练的英文表格识别模型|18.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) | |en_ppocr_mobile_v2.0_table_structure|基于TableRec-RARE在PubTabNet数据集上训练的英文表格识别模型|6.8M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
|en_ppstructure_mobile_v2.0_SLANet|基于SLANet在PubTabNet数据集上训练的英文表格识别模型|9M|[推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_train.tar) | |en_ppstructure_mobile_v2.0_SLANet|基于SLANet在PubTabNet数据集上训练的英文表格识别模型|9.2M|[推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_train.tar) |
|ch_ppstructure_mobile_v2.0_SLANet|基于SLANet在PubTabNet数据集上训练的中文表格识别模型|9.3M|[推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_train.tar) | |ch_ppstructure_mobile_v2.0_SLANet|基于SLANet的中文表格识别模型|9.3M|[推理模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_train.tar) |
<a name="3"></a> <a name="3"></a>
......
ppstructure/docs/models_list_en.md
浏览文件 @ 96a8e0f8
...@@ -4,18 +4,20 @@ ...@@ -4,18 +4,20 @@
- [2. OCR and Table Recognition](#2-ocr-and-table-recognition) - [2. OCR and Table Recognition](#2-ocr-and-table-recognition)
- [2.1 OCR](#21-ocr) - [2.1 OCR](#21-ocr)
- [2.2 Table Recognition](#22-table-recognition) - [2.2 Table Recognition](#22-table-recognition)
- [3. VQA](#3-vqa) - [3. KIE](#3-kie)
- [4. KIE](#4-kie)
<a name="1"></a> <a name="1"></a>
## 1. Layout Analysis ## 1. Layout Analysis
|model name| description |download|label_map| |model name| description | inference model size |download|dict path|
| --- |---------------------------------------------------------------------------------------------------------------------------------------------------------| --- | --- | | --- |---------------------------------------------------------------------------------------------------------------------------------------------------------| --- | --- | --- |
| ppyolov2_r50vd_dcn_365e_publaynet | The layout analysis model trained on the PubLayNet dataset, the model can recognition 5 types of areas such as **text, title, table, picture and list** | [inference model](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet.tar) / [trained model](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet_pretrained.pdparams) |{0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"}| | picodet_lcnet_x1_0_fgd_layout | The layout analysis English model trained on the PubLayNet dataset based on PicoDet LCNet_x1_0 and FGD . the model can recognition 5 types of areas such as **Text, Title, Table, Picture and List** | 9.7M | [inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout.pdparams) | [PubLayNet dict](../../ppocr/utils/dict/layout_dict/layout_publaynet_dict.txt) |
| ppyolov2_r50vd_dcn_365e_tableBank_word | The layout analysis model trained on the TableBank Word dataset, the model can only detect tables | [inference model](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_tableBank_word.tar) | {0:"Table"}| | ppyolov2_r50vd_dcn_365e_publaynet | The layout analysis English model trained on the PubLayNet dataset based on PP-YOLOv2 | 221M | [inference_moel]](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet.tar) / [trained model](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet_pretrained.pdparams) | sme as above |
| ppyolov2_r50vd_dcn_365e_tableBank_latex | The layout analysis model trained on the TableBank Latex dataset, the model can only detect tables | [inference model](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_tableBank_latex.tar) | {0:"Table"}| | picodet_lcnet_x1_0_fgd_layout_cdla | The layout analysis Chinese model trained on the CDLA dataset, the model can recognition 10 types of areas such as **Table、Figure、Figure caption、Table、Table caption、Header、Footer、Reference、Equation** | 9.7M | [inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_cdla_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_cdla.pdparams) | [CDLA dict](../../ppocr/utils/dict/layout_dict/layout_cdla_dict.txt) |
| picodet_lcnet_x1_0_fgd_layout_table | The layout analysis model trained on the table dataset, the model can detect tables in Chinese and English documents | 9.7M | [inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_table_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_table.pdparams) | [Table dict](../../ppocr/utils/dict/layout_dict/layout_table_dict.txt) |
| ppyolov2_r50vd_dcn_365e_tableBank_word | The layout analysis model trained on the TableBank Word dataset based on PP-YOLOv2, the model can detect tables in English documents | 221M | [inference model](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_tableBank_word.tar) | same as above |
| ppyolov2_r50vd_dcn_365e_tableBank_latex | The layout analysis model trained on the TableBank Latex dataset based on PP-YOLOv2, the model can detect tables in English documents | 221M | [inference model](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_tableBank_latex.tar) | same as above |
<a name="2"></a> <a name="2"></a>
## 2. OCR and Table Recognition ## 2. OCR and Table Recognition
...@@ -35,24 +37,30 @@ If you need to use other OCR models, you can download the model in [PP-OCR model ...@@ -35,24 +37,30 @@ If you need to use other OCR models, you can download the model in [PP-OCR model
|model| description |inference model size|download| |model| description |inference model size|download|
| --- |-----------------------------------------------------------------------------| --- | --- | | --- |-----------------------------------------------------------------------------| --- | --- |
|en_ppocr_mobile_v2.0_table_structure| English table recognition model trained on PubTabNet dataset based on TableRec-RARE |18.6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) | |en_ppocr_mobile_v2.0_table_structure| English table recognition model trained on PubTabNet dataset based on TableRec-RARE |6.8M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
|en_ppstructure_mobile_v2.0_SLANet|English table recognition model trained on PubTabNet dataset based on SLANet|9M|[inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_train.tar) | |en_ppstructure_mobile_v2.0_SLANet|English table recognition model trained on PubTabNet dataset based on SLANet|9.2M|[inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_train.tar) |
|ch_ppstructure_mobile_v2.0_SLANet|Chinese table recognition model trained on PubTabNet dataset based on SLANet|9.3M|[inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_train.tar) | |ch_ppstructure_mobile_v2.0_SLANet|Chinese table recognition model based on SLANet|9.3M|[inference model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_train.tar) |
<a name="3"></a> <a name="3"></a>
## 3. VQA ## 3. KIE
|model| description |inference model size|download| On XFUND_zh dataset, Accuracy and time cost of different models on V100 GPU are as follows.
| --- |----------------------------------------------------------------| --- | --- |
|ser_LayoutXLM_xfun_zh| SER model trained on xfun Chinese dataset based on LayoutXLM |1.4G|[inference model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) | |Model|Backbone|Task|Config|Hmean|Time cost(ms)|Download link|
|re_LayoutXLM_xfun_zh| Re model trained on xfun Chinese dataset based on LayoutXLM |1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) | | --- | --- | --- | --- | --- | --- |--- |
|ser_LayoutLMv2_xfun_zh| SER model trained on xfun Chinese dataset based on LayoutXLMv2 |778M|[inference model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar) | |VI-LayoutXLM| VI-LayoutXLM-base | SER | [ser_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml)|**93.19%**| 15.49| [trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)|
|re_LayoutLMv2_xfun_zh| Re model trained on xfun Chinese dataset based on LayoutXLMv2 |765M|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar) | |LayoutXLM| LayoutXLM-base | SER | [ser_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml)|90.38%| 19.49 |[trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)|
|ser_LayoutLM_xfun_zh| SER model trained on xfun Chinese dataset based on LayoutLM |430M|[inference model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar) | |LayoutLM| LayoutLM-base | SER | [ser_layoutlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutlm_xfund_zh.yml)|77.31%|-|[trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar)|
|LayoutLMv2| LayoutLMv2-base | SER | [ser_layoutlmv2_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutlmv2_xfund_zh.yml)|85.44%|31.46|[trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar)|
<a name="4"></a> |VI-LayoutXLM| VI-LayoutXLM-base | RE | [re_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml)|**83.92%**|15.49|[trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar)|
## 4. KIE |LayoutXLM| LayoutXLM-base | RE | [re_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml)|74.83%|19.49|[trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar)|
|LayoutLMv2| LayoutLMv2-base | RE | [re_layoutlmv2_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutlmv2_xfund_zh.yml)|67.77%|31.46|[trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar)|
|model|description|model size|download|
| --- | --- | --- | --- | * Note: The above time cost information just considers inference time without preprocess or postprocess, test environment: `V100 GPU + CUDA 10.2 + CUDNN 8.1.1 + TRT 7.2.3.4`
|SDMGR|Key Information Extraction Model|78M|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)|
On wildreceipt dataset, the algorithm result is as follows:
|Model|Backbone|Config|Hmean|Download link|
| --- | --- | --- | --- | --- |
|SDMGR|VGG6|[configs/kie/sdmgr/kie_unet_sdmgr.yml](../../configs/kie/sdmgr/kie_unet_sdmgr.yml)|86.7%|[trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/kie/kie_vgg16.tar)|
ppstructure/docs/quickstart.md
浏览文件 @ 96a8e0f8
# PP-Structure 快速开始 # PP-Structure 快速开始
- [1. 安装依赖包](#1-安装依赖包) - [1. 准备环境](#1-准备环境)
- [2. 便捷使用](#2-便捷使用) - [2. 便捷使用](#2-便捷使用)
- [2.1 命令行使用](#21-命令行使用) - [2.1 命令行使用](#21-命令行使用)
- [2.1.1 图像方向分类+版面分析+表格识别](#211-图像方向分类版面分析表格识别) - [2.1.1 图像方向分类+版面分析+表格识别](#211-图像方向分类版面分析表格识别)
- [2.1.2 版面分析+表格识别](#212-版面分析表格识别) - [2.1.2 版面分析+表格识别](#212-版面分析表格识别)
- [2.1.3 版面分析](#213-版面分析) - [2.1.3 版面分析](#213-版面分析)
- [2.1.4 表格识别](#214-表格识别) - [2.1.4 表格识别](#214-表格识别)
- [2.1.5 DocVQA](#215-docvqa) - [2.1.5 关键信息抽取](#215-关键信息抽取)
- [2.2 代码使用](#22-代码使用) - [2.1.6 版面恢复](#216-版面恢复)
- [2.2.1 图像方向分类版面分析表格识别](#221-图像方向分类版面分析表格识别) - [2.2 Python脚本使用](#22-Python脚本使用)
- [2.2.1 图像方向分类+版面分析+表格识别](#221-图像方向分类版面分析表格识别)
- [2.2.2 版面分析+表格识别](#222-版面分析表格识别) - [2.2.2 版面分析+表格识别](#222-版面分析表格识别)
- [2.2.3 版面分析](#223-版面分析) - [2.2.3 版面分析](#223-版面分析)
- [2.2.4 表格识别](#224-表格识别) - [2.2.4 表格识别](#224-表格识别)
- [2.2.5 DocVQA](#225-docvqa) - [2.2.5 关键信息抽取](#225-关键信息抽取)
- [2.2.6 版面恢复](#226-版面恢复)
- [2.3 返回结果说明](#23-返回结果说明) - [2.3 返回结果说明](#23-返回结果说明)
- [2.3.1 版面分析+表格识别](#231-版面分析表格识别) - [2.3.1 版面分析+表格识别](#231-版面分析表格识别)
- [2.3.2 DocVQA](#232-docvqa) - [2.3.2 关键信息抽取](#232-关键信息抽取)
- [2.4 参数说明](#24-参数说明) - [2.4 参数说明](#24-参数说明)
- [3. 小结](#3-小结)
<a name="1"></a> <a name="1"></a>
## 1. 安装依赖包 ## 1. 准备环境
### 1.1 安装PaddlePaddle
> 如果您没有基础的Python运行环境，请参考[运行环境准备](../../doc/doc_ch/environment.md)。
- 您的机器安装的是CUDA9或CUDA10，请运行以下命令安装
```bash
python3 -m pip install paddlepaddle-gpu -i https://mirror.baidu.com/pypi/simple
```
- 您的机器是CPU，请运行以下命令安装
```bash
python3 -m pip install paddlepaddle -i https://mirror.baidu.com/pypi/simple
```
更多的版本需求，请参照[飞桨官网安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
### 1.2 安装PaddleOCR whl包
```bash ```bash
# 安装 paddleocr，推荐使用2.5+版本 # 安装 paddleocr，推荐使用2.6版本
pip3 install "paddleocr>=2.5" pip3 install "paddleocr>=2.6"
# 安装 DocVQA依赖包paddlenlp（如不需要DocVQA功能，可跳过）
pip install paddlenlp # 安装 图像方向分类依赖包paddleclas（如不需要图像方向分类功能，可跳过）
pip3 install paddleclas
# 安装 关键信息抽取 依赖包（如不需要KIE功能，可跳过）
pip3 install -r kie/requirements.txt
``` ```
<a name="2"></a> <a name="2"></a>
...@@ -40,37 +63,46 @@ pip install paddlenlp ...@@ -40,37 +63,46 @@ pip install paddlenlp
<a name="211"></a> <a name="211"></a>
#### 2.1.1 图像方向分类+版面分析+表格识别 #### 2.1.1 图像方向分类+版面分析+表格识别
```bash ```bash
paddleocr --image_dir=PaddleOCR/ppstructure/docs/table/1.png --type=structure --image_orientation=true paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure --image_orientation=true
``` ```
<a name="212"></a> <a name="212"></a>
#### 2.1.2 版面分析+表格识别 #### 2.1.2 版面分析+表格识别
```bash ```bash
paddleocr --image_dir=PaddleOCR/ppstructure/docs/table/1.png --type=structure paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure
``` ```
<a name="213"></a> <a name="213"></a>
#### 2.1.3 版面分析 #### 2.1.3 版面分析
```bash ```bash
paddleocr --image_dir=PaddleOCR/ppstructure/docs/table/1.png --type=structure --table=false --ocr=false paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure --table=false --ocr=false
``` ```
<a name="214"></a> <a name="214"></a>
#### 2.1.4 表格识别 #### 2.1.4 表格识别
```bash ```bash
paddleocr --image_dir=PaddleOCR/ppstructure/docs/table/table.jpg --type=structure --layout=false paddleocr --image_dir=ppstructure/docs/table/table.jpg --type=structure --layout=false
``` ```
<a name="215"></a> <a name="215"></a>
#### 2.1.5 DocVQA
请参考：[文档视觉问答](../vqa/README.md)。 #### 2.1.5 关键信息抽取
关键信息抽取暂不支持通过whl包调用，详细使用教程请参考：[关键信息抽取教程](../kie/README_ch.md)。
<a name="216"></a>
#### 2.1.6 版面恢复
```bash
paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure --recovery=true
```
<a name="22"></a> <a name="22"></a>
### 2.2 代码使用
### 2.2 Python脚本使用
<a name="221"></a> <a name="221"></a>
#### 2.2.1 图像方向分类版面分析表格识别 #### 2.2.1 图像方向分类+版面分析+表格识别
```python ```python
import os import os
...@@ -80,7 +112,7 @@ from paddleocr import PPStructure,draw_structure_result,save_structure_res ...@@ -80,7 +112,7 @@ from paddleocr import PPStructure,draw_structure_result,save_structure_res
table_engine = PPStructure(show_log=True, image_orientation=True) table_engine = PPStructure(show_log=True, image_orientation=True)
save_folder = './output' save_folder = './output'
img_path = 'PaddleOCR/ppstructure/docs/table/1.png' img_path = 'ppstructure/docs/table/1.png'
img = cv2.imread(img_path) img = cv2.imread(img_path)
result = table_engine(img) result = table_engine(img)
save_structure_res(result, save_folder,os.path.basename(img_path).split('.')[0]) save_structure_res(result, save_folder,os.path.basename(img_path).split('.')[0])
...@@ -91,7 +123,7 @@ for line in result: ...@@ -91,7 +123,7 @@ for line in result:
from PIL import Image from PIL import Image
font_path = 'PaddleOCR/doc/fonts/simfang.ttf' # PaddleOCR下提供字体包 font_path = 'doc/fonts/simfang.ttf' # PaddleOCR下提供字体包
image = Image.open(img_path).convert('RGB') image = Image.open(img_path).convert('RGB')
im_show = draw_structure_result(image, result,font_path=font_path) im_show = draw_structure_result(image, result,font_path=font_path)
im_show = Image.fromarray(im_show) im_show = Image.fromarray(im_show)
...@@ -109,7 +141,7 @@ from paddleocr import PPStructure,draw_structure_result,save_structure_res ...@@ -109,7 +141,7 @@ from paddleocr import PPStructure,draw_structure_result,save_structure_res
table_engine = PPStructure(show_log=True) table_engine = PPStructure(show_log=True)
save_folder = './output' save_folder = './output'
img_path = 'PaddleOCR/ppstructure/docs/table/1.png' img_path = 'ppstructure/docs/table/1.png'
img = cv2.imread(img_path) img = cv2.imread(img_path)
result = table_engine(img) result = table_engine(img)
save_structure_res(result, save_folder,os.path.basename(img_path).split('.')[0]) save_structure_res(result, save_folder,os.path.basename(img_path).split('.')[0])
...@@ -120,7 +152,7 @@ for line in result: ...@@ -120,7 +152,7 @@ for line in result:
from PIL import Image from PIL import Image
font_path = 'PaddleOCR/doc/fonts/simfang.ttf' # PaddleOCR下提供字体包 font_path = 'doc/fonts/simfang.ttf' # PaddleOCR下提供字体包
image = Image.open(img_path).convert('RGB') image = Image.open(img_path).convert('RGB')
im_show = draw_structure_result(image, result,font_path=font_path) im_show = draw_structure_result(image, result,font_path=font_path)
im_show = Image.fromarray(im_show) im_show = Image.fromarray(im_show)
...@@ -138,7 +170,7 @@ from paddleocr import PPStructure,save_structure_res ...@@ -138,7 +170,7 @@ from paddleocr import PPStructure,save_structure_res
table_engine = PPStructure(table=False, ocr=False, show_log=True) table_engine = PPStructure(table=False, ocr=False, show_log=True)
save_folder = './output' save_folder = './output'
img_path = 'PaddleOCR/ppstructure/docs/table/1.png' img_path = 'ppstructure/docs/table/1.png'
img = cv2.imread(img_path) img = cv2.imread(img_path)
result = table_engine(img) result = table_engine(img)
save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0]) save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0])
...@@ -149,6 +181,7 @@ for line in result: ...@@ -149,6 +181,7 @@ for line in result:
``` ```
<a name="224"></a> <a name="224"></a>
#### 2.2.4 表格识别 #### 2.2.4 表格识别
```python ```python
...@@ -159,7 +192,7 @@ from paddleocr import PPStructure,save_structure_res ...@@ -159,7 +192,7 @@ from paddleocr import PPStructure,save_structure_res
table_engine = PPStructure(layout=False, show_log=True) table_engine = PPStructure(layout=False, show_log=True)
save_folder = './output' save_folder = './output'
img_path = 'PaddleOCR/ppstructure/docs/table/table.jpg' img_path = 'ppstructure/docs/table/table.jpg'
img = cv2.imread(img_path) img = cv2.imread(img_path)
result = table_engine(img) result = table_engine(img)
save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0]) save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0])
...@@ -170,13 +203,40 @@ for line in result: ...@@ -170,13 +203,40 @@ for line in result:
``` ```
<a name="225"></a> <a name="225"></a>
#### 2.2.5 DocVQA #### 2.2.5 关键信息抽取
关键信息抽取暂不支持通过whl包调用，详细使用教程请参考：[关键信息抽取教程](../kie/README_ch.md)。
请参考：[文档视觉问答](../vqa/README.md)。 <a name="226"></a>
#### 2.2.6 版面恢复
```python
import os
import cv2
from paddleocr import PPStructure,save_structure_res
from paddelocr.ppstructure.recovery.recovery_to_doc import sorted_layout_boxes, convert_info_docx
table_engine = PPStructure(layout=False, show_log=True)
save_folder = './output'
img_path = 'ppstructure/docs/table/1.png'
img = cv2.imread(img_path)
result = table_engine(img)
save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0])
for line in result:
line.pop('img')
print(line)
h, w, _ = img.shape
res = sorted_layout_boxes(res, w)
convert_info_docx(img, result, save_folder, os.path.basename(img_path).split('.')[0])
```
<a name="23"></a> <a name="23"></a>
### 2.3 返回结果说明 ### 2.3 返回结果说明
PP-Structure的返回结果为一个dict组成的list，示例如下 PP-Structure的返回结果为一个dict组成的list，示例如下：
<a name="231"></a> <a name="231"></a>
#### 2.3.1 版面分析+表格识别 #### 2.3.1 版面分析+表格识别
...@@ -189,7 +249,7 @@ PP-Structure的返回结果为一个dict组成的list，示例如下 ...@@ -189,7 +249,7 @@ PP-Structure的返回结果为一个dict组成的list，示例如下
} }
] ]
``` ```
dict 里各个字段说明如下 dict 里各个字段说明如下：
| 字段 | 说明| | 字段 | 说明|
| --- |---| | --- |---|
...@@ -208,9 +268,9 @@ dict 里各个字段说明如下 ...@@ -208,9 +268,9 @@ dict 里各个字段说明如下
``` ```
<a name="232"></a> <a name="232"></a>
#### 2.3.2 DocVQA #### 2.3.2 关键信息抽取
请参考：[文档视觉问答](../vqa/README.md)。 请参考：[关键信息抽取教程](../kie/README_ch.md)。
<a name="24"></a> <a name="24"></a>
### 2.4 参数说明 ### 2.4 参数说明
...@@ -226,15 +286,21 @@ dict 里各个字段说明如下 ...@@ -226,15 +286,21 @@ dict 里各个字段说明如下
| layout_dict_path | 版面分析模型字典| ../ppocr/utils/dict/layout_publaynet_dict.txt | | layout_dict_path | 版面分析模型字典| ../ppocr/utils/dict/layout_publaynet_dict.txt |
| layout_score_threshold | 版面分析模型检测框阈值| 0.5| | layout_score_threshold | 版面分析模型检测框阈值| 0.5|
| layout_nms_threshold | 版面分析模型nms阈值| 0.5| | layout_nms_threshold | 版面分析模型nms阈值| 0.5|
| vqa_algorithm | vqa模型算法| LayoutXLM| | kie_algorithm | kie模型算法| LayoutXLM|
| ser_model_dir | ser模型 inference 模型地址| None| | ser_model_dir | ser模型 inference 模型地址| None|
| ser_dict_path | ser模型字典| ../train_data/XFUND/class_list_xfun.txt| | ser_dict_path | ser模型字典| ../train_data/XFUND/class_list_xfun.txt|
| mode | structure or vqa | structure | | mode | structure or kie | structure |
| image_orientation | 前向中是否执行图像方向分类 | False | | image_orientation | 前向中是否执行图像方向分类 | False |
| layout | 前向中是否执行版面分析 | True | | layout | 前向中是否执行版面分析 | True |
| table | 前向中是否执行表格识别 | True | | table | 前向中是否执行表格识别 | True |
| ocr | 对于版面分析中的非表格区域，是否执行ocr。当layout为False时会被自动设置为False| True | | ocr | 对于版面分析中的非表格区域，是否执行ocr。当layout为False时会被自动设置为False| True |
| recovery | 前向中是否执行版面恢复| False | | recovery | 前向中是否执行版面恢复| False |
| save_pdf | 版面恢复导出docx文件的同时，是否导出pdf文件 | False |
| structure_version | 模型版本，可选 PP-structure和PP-structurev2 | PP-structure | | structure_version | 模型版本，可选 PP-structure和PP-structurev2 | PP-structure |
大部分参数和PaddleOCR whl包保持一致，见 [whl包文档](../../doc/doc_ch/whl.md) 大部分参数和PaddleOCR whl包保持一致，见 [whl包文档](../../doc/doc_ch/whl.md)
<a name="3"></a>
## 3. 小结
通过本节内容，相信您已经熟练掌握通过PaddleOCR whl包调用PP-Structure相关功能的使用方法，您可以参考[文档教程](../../README_ch.md#文档教程)，获取包括模型训练、推理部署等更详细的使用教程。
\ No newline at end of file
ppstructure/docs/quickstart_en.md
浏览文件 @ 96a8e0f8
# PP-Structure Quick Start # PP-Structure Quick Start
- [1. Install package](#1-install-package) - [1. Environment Preparation](#1-environment-preparation)
- [2. Use](#2-use) - [2. Quick Use](#2-quick-use)
- [2.1 Use by command line](#21-use-by-command-line) - [2.1 Use by command line](#21-use-by-command-line)
- [2.1.1 image orientation + layout analysis + table recognition](#211-image-orientation--layout-analysis--table-recognition) - [2.1.1 image orientation + layout analysis + table recognition](#211-image-orientation--layout-analysis--table-recognition)
- [2.1.2 layout analysis + table recognition](#212-layout-analysis--table-recognition) - [2.1.2 layout analysis + table recognition](#212-layout-analysis--table-recognition)
- [2.1.3 layout analysis](#213-layout-analysis) - [2.1.3 layout analysis](#213-layout-analysis)
- [2.1.4 table recognition](#214-table-recognition) - [2.1.4 table recognition](#214-table-recognition)
- [2.1.5 DocVQA](#215-docvqa) - [2.1.5 Key Information Extraction](#215-Key-Information-Extraction)
- [2.2 Use by code](#22-use-by-code) - [2.1.6 layout recovery](#216-layout-recovery)
- [2.2 Use by python script](#22-use-by-python-script)
- [2.2.1 image orientation + layout analysis + table recognition](#221-image-orientation--layout-analysis--table-recognition) - [2.2.1 image orientation + layout analysis + table recognition](#221-image-orientation--layout-analysis--table-recognition)
- [2.2.2 layout analysis + table recognition](#222-layout-analysis--table-recognition) - [2.2.2 layout analysis + table recognition](#222-layout-analysis--table-recognition)
- [2.2.3 layout analysis](#223-layout-analysis) - [2.2.3 layout analysis](#223-layout-analysis)
- [2.2.4 table recognition](#224-table-recognition) - [2.2.4 table recognition](#224-table-recognition)
- [2.2.5 DocVQA](#225-docvqa) - [2.2.5 Key Information Extraction](#225-Key-Information-Extraction)
- [2.2.6 layout recovery](#226-layout-recovery)
- [2.3 Result description](#23-result-description) - [2.3 Result description](#23-result-description)
- [2.3.1 layout analysis + table recognition](#231-layout-analysis--table-recognition) - [2.3.1 layout analysis + table recognition](#231-layout-analysis--table-recognition)
- [2.3.2 DocVQA](#232-docvqa) - [2.3.2 Key Information Extraction](#232-Key-Information-Extraction)
- [2.4 Parameter Description](#24-parameter-description) - [2.4 Parameter Description](#24-parameter-description)
- [3. Summary](#3-summary)
<a name="1"></a> <a name="1"></a>
## 1. Install package ## 1. Environment Preparation
### 1.1 Install PaddlePaddle
> If you do not have a Python environment, please refer to [Environment Preparation](./environment_en.md).
- If you have CUDA 9 or CUDA 10 installed on your machine, please run the following command to install
```bash
python3 -m pip install paddlepaddle-gpu -i https://mirror.baidu.com/pypi/simple
```
- If you have no available GPU on your machine, please run the following command to install the CPU version
```bash
python3 -m pip install paddlepaddle -i https://mirror.baidu.com/pypi/simple
```
For more software version requirements, please refer to the instructions in [Installation Document](https://www.paddlepaddle.org.cn/install/quick) for operation.
### 1.2 Install PaddleOCR Whl Package
```bash ```bash
# Install paddleocr, version 2.5+ is recommended # Install paddleocr, version 2.6 is recommended
pip3 install "paddleocr>=2.5" pip3 install "paddleocr>=2.6"
# Install the DocVQA dependency package paddlenlp (if you do not use the DocVQA, you can skip it)
pip install paddlenlp
# Install the image direction classification dependency package paddleclas (if you do not use the image direction classification, you can skip it)
pip3 install paddleclas
# Install the KIE dependency packages (if you do not use the KIE, you can skip it)
pip3 install -r kie/requirements.txt
``` ```
<a name="2"></a> <a name="2"></a>
## 2. Use
## 2. Quick Use
<a name="21"></a> <a name="21"></a>
### 2.1 Use by command line ### 2.1 Use by command line
...@@ -40,34 +66,40 @@ pip install paddlenlp ...@@ -40,34 +66,40 @@ pip install paddlenlp
<a name="211"></a> <a name="211"></a>
#### 2.1.1 image orientation + layout analysis + table recognition #### 2.1.1 image orientation + layout analysis + table recognition
```bash ```bash
paddleocr --image_dir=PaddleOCR/ppstructure/docs/table/1.png --type=structure --image_orientation=true paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure --image_orientation=true
``` ```
<a name="212"></a> <a name="212"></a>
#### 2.1.2 layout analysis + table recognition #### 2.1.2 layout analysis + table recognition
```bash ```bash
paddleocr --image_dir=PaddleOCR/ppstructure/docs/table/1.png --type=structure paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure
``` ```
<a name="213"></a> <a name="213"></a>
#### 2.1.3 layout analysis #### 2.1.3 layout analysis
```bash ```bash
paddleocr --image_dir=PaddleOCR/ppstructure/docs/table/1.png --type=structure --table=false --ocr=false paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure --table=false --ocr=false
``` ```
<a name="214"></a> <a name="214"></a>
#### 2.1.4 table recognition #### 2.1.4 table recognition
```bash ```bash
paddleocr --image_dir=PaddleOCR/ppstructure/docs/table/table.jpg --type=structure --layout=false paddleocr --image_dir=ppstructure/docs/table/table.jpg --type=structure --layout=false
``` ```
<a name="215"></a> <a name="215"></a>
#### 2.1.5 DocVQA #### 2.1.5 Key Information Extraction
Key information extraction does not currently support use by the whl package. For detailed usage tutorials, please refer to: [Key Information Extraction](../kie/README.md).
Please refer to: [Documentation Visual Q&A](../vqa/README.md) . <a name="216"></a>
#### 2.1.6 layout recovery
```bash
paddleocr --image_dir=ppstructure/docs/table/1.png --type=structure --recovery=true
```
<a name="22"></a> <a name="22"></a>
### 2.2 Use by code ### 2.2 Use by python script
<a name="221"></a> <a name="221"></a>
#### 2.2.1 image orientation + layout analysis + table recognition #### 2.2.1 image orientation + layout analysis + table recognition
...@@ -80,7 +112,7 @@ from paddleocr import PPStructure,draw_structure_result,save_structure_res ...@@ -80,7 +112,7 @@ from paddleocr import PPStructure,draw_structure_result,save_structure_res
table_engine = PPStructure(show_log=True, image_orientation=True) table_engine = PPStructure(show_log=True, image_orientation=True)
save_folder = './output' save_folder = './output'
img_path = 'PaddleOCR/ppstructure/docs/table/1.png' img_path = 'ppstructure/docs/table/1.png'
img = cv2.imread(img_path) img = cv2.imread(img_path)
result = table_engine(img) result = table_engine(img)
save_structure_res(result, save_folder,os.path.basename(img_path).split('.')[0]) save_structure_res(result, save_folder,os.path.basename(img_path).split('.')[0])
...@@ -91,7 +123,7 @@ for line in result: ...@@ -91,7 +123,7 @@ for line in result:
from PIL import Image from PIL import Image
font_path = 'PaddleOCR/doc/fonts/simfang.ttf' # PaddleOCR下提供字体包 font_path = 'doc/fonts/simfang.ttf' # PaddleOCR下提供字体包
image = Image.open(img_path).convert('RGB') image = Image.open(img_path).convert('RGB')
im_show = draw_structure_result(image, result,font_path=font_path) im_show = draw_structure_result(image, result,font_path=font_path)
im_show = Image.fromarray(im_show) im_show = Image.fromarray(im_show)
...@@ -109,7 +141,7 @@ from paddleocr import PPStructure,draw_structure_result,save_structure_res ...@@ -109,7 +141,7 @@ from paddleocr import PPStructure,draw_structure_result,save_structure_res
table_engine = PPStructure(show_log=True) table_engine = PPStructure(show_log=True)
save_folder = './output' save_folder = './output'
img_path = 'PaddleOCR/ppstructure/docs/table/1.png' img_path = 'ppstructure/docs/table/1.png'
img = cv2.imread(img_path) img = cv2.imread(img_path)
result = table_engine(img) result = table_engine(img)
save_structure_res(result, save_folder,os.path.basename(img_path).split('.')[0]) save_structure_res(result, save_folder,os.path.basename(img_path).split('.')[0])
...@@ -120,7 +152,7 @@ for line in result: ...@@ -120,7 +152,7 @@ for line in result:
from PIL import Image from PIL import Image
font_path = 'PaddleOCR/doc/fonts/simfang.ttf' # PaddleOCR下提供字体包 font_path = 'doc/fonts/simfang.ttf' # font provieded in PaddleOCR
image = Image.open(img_path).convert('RGB') image = Image.open(img_path).convert('RGB')
im_show = draw_structure_result(image, result,font_path=font_path) im_show = draw_structure_result(image, result,font_path=font_path)
im_show = Image.fromarray(im_show) im_show = Image.fromarray(im_show)
...@@ -138,7 +170,7 @@ from paddleocr import PPStructure,save_structure_res ...@@ -138,7 +170,7 @@ from paddleocr import PPStructure,save_structure_res
table_engine = PPStructure(table=False, ocr=False, show_log=True) table_engine = PPStructure(table=False, ocr=False, show_log=True)
save_folder = './output' save_folder = './output'
img_path = 'PaddleOCR/ppstructure/docs/table/1.png' img_path = 'ppstructure/docs/table/1.png'
img = cv2.imread(img_path) img = cv2.imread(img_path)
result = table_engine(img) result = table_engine(img)
save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0]) save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0])
...@@ -159,7 +191,7 @@ from paddleocr import PPStructure,save_structure_res ...@@ -159,7 +191,7 @@ from paddleocr import PPStructure,save_structure_res
table_engine = PPStructure(layout=False, show_log=True) table_engine = PPStructure(layout=False, show_log=True)
save_folder = './output' save_folder = './output'
img_path = 'PaddleOCR/ppstructure/docs/table/table.jpg' img_path = 'ppstructure/docs/table/table.jpg'
img = cv2.imread(img_path) img = cv2.imread(img_path)
result = table_engine(img) result = table_engine(img)
save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0]) save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0])
...@@ -170,9 +202,35 @@ for line in result: ...@@ -170,9 +202,35 @@ for line in result:
``` ```
<a name="225"></a> <a name="225"></a>
#### 2.2.5 DocVQA #### 2.2.5 Key Information Extraction
Please refer to: [Documentation Visual Q&A](../vqa/README.md) . Key information extraction does not currently support use by the whl package. For detailed usage tutorials, please refer to: [Key Information Extraction](../kie/README.md).
<a name="226"></a>
#### 2.2.6 layout recovery
```python
import os
import cv2
from paddleocr import PPStructure,save_structure_res
from paddelocr.ppstructure.recovery.recovery_to_doc import sorted_layout_boxes, convert_info_docx
table_engine = PPStructure(layout=False, show_log=True)
save_folder = './output'
img_path = 'ppstructure/docs/table/1.png'
img = cv2.imread(img_path)
result = table_engine(img)
save_structure_res(result, save_folder, os.path.basename(img_path).split('.')[0])
for line in result:
line.pop('img')
print(line)
h, w, _ = img.shape
res = sorted_layout_boxes(res, w)
convert_info_docx(img, result, save_folder, os.path.basename(img_path).split('.')[0])
```
<a name="23"></a> <a name="23"></a>
### 2.3 Result description ### 2.3 Result description
...@@ -208,9 +266,9 @@ After the recognition is completed, each image will have a directory with the sa ...@@ -208,9 +266,9 @@ After the recognition is completed, each image will have a directory with the sa
``` ```
<a name="232"></a> <a name="232"></a>
#### 2.3.2 DocVQA #### 2.3.2 Key Information Extraction
Please refer to: [Documentation Visual Q&A](../vqa/README.md) . Please refer to: [Key Information Extraction](../kie/README.md) .
<a name="24"></a> <a name="24"></a>
### 2.4 Parameter Description ### 2.4 Parameter Description
...@@ -226,15 +284,21 @@ Please refer to: [Documentation Visual Q&A](../vqa/README.md) . ...@@ -226,15 +284,21 @@ Please refer to: [Documentation Visual Q&A](../vqa/README.md) .
| layout_dict_path | The dictionary path of layout analysis model| ../ppocr/utils/dict/layout_publaynet_dict.txt | | layout_dict_path | The dictionary path of layout analysis model| ../ppocr/utils/dict/layout_publaynet_dict.txt |
| layout_score_threshold | The box threshold path of layout analysis model| 0.5| | layout_score_threshold | The box threshold path of layout analysis model| 0.5|
| layout_nms_threshold | The nms threshold path of layout analysis model| 0.5| | layout_nms_threshold | The nms threshold path of layout analysis model| 0.5|
| vqa_algorithm | vqa model algorithm| LayoutXLM| | kie_algorithm | kie model algorithm| LayoutXLM|
| ser_model_dir | Ser model inference model path| None| | ser_model_dir | Ser model inference model path| None|
| ser_dict_path | The dictionary path of Ser model| ../train_data/XFUND/class_list_xfun.txt| | ser_dict_path | The dictionary path of Ser model| ../train_data/XFUND/class_list_xfun.txt|
| mode | structure or vqa | structure | | mode | structure or kie | structure |
| image_orientation | Whether to perform image orientation classification in forward | False | | image_orientation | Whether to perform image orientation classification in forward | False |
| layout | Whether to perform layout analysis in forward | True | | layout | Whether to perform layout analysis in forward | True |
| table | Whether to perform table recognition in forward | True | | table | Whether to perform table recognition in forward | True |
| ocr | Whether to perform ocr for non-table areas in layout analysis. When layout is False, it will be automatically set to False| True | | ocr | Whether to perform ocr for non-table areas in layout analysis. When layout is False, it will be automatically set to False| True |
| recovery | Whether to perform layout recovery in forward| False | | recovery | Whether to perform layout recovery in forward| False |
| save_pdf | Whether to convert docx to pdf when recovery| False |
| structure_version | Structure version, optional PP-structure and PP-structurev2 | PP-structure | | structure_version | Structure version, optional PP-structure and PP-structurev2 | PP-structure |
Most of the parameters are consistent with the PaddleOCR whl package, see [whl package documentation](../../doc/doc_en/whl.md) Most of the parameters are consistent with the PaddleOCR whl package, see [whl package documentation](../../doc/doc_en/whl.md)
<a name="3"></a>
## 3. Summary
Through the content in this section, you can master the use of PP-Structure related functions through PaddleOCR whl package. Please refer to [documentation tutorial](../../README.md) for more detailed usage tutorials including model training, inference and deployment, etc.
\ No newline at end of file
ppstructure/docs/table/layout.jpg
查看替换文件 @ 29268eaa 浏览文件 @ 96a8e0f8
ppstructure/docs/table/layout.jpg

671.5 KB | W: | H:

ppstructure/docs/table/layout.jpg

671.5 KB | W: | H:

ppstructure/docs/table/layout.jpg
ppstructure/docs/table/layout.jpg
ppstructure/docs/table/layout.jpg
ppstructure/docs/table/layout.jpg
  • 2-up
  • Swipe
  • Onion skin
ppstructure/docs/table/paper-image.jpg
查看替换文件 @ 29268eaa 浏览文件 @ 96a8e0f8
ppstructure/docs/table/paper-image.jpg

671.5 KB | W: | H:

ppstructure/docs/table/paper-image.jpg

671.5 KB | W: | H:

ppstructure/docs/table/paper-image.jpg
ppstructure/docs/table/paper-image.jpg
ppstructure/docs/table/paper-image.jpg
ppstructure/docs/table/paper-image.jpg
  • 2-up
  • Swipe
  • Onion skin
ppstructure/kie/README.md 0 → 100644
浏览文件 @ 96a8e0f8
English | [简体中文](README_ch.md)
- [1. Introduction](#1-introduction)
- [2. Accuracy and performance](#2-Accuracy-and-performance)
- [3. Visualization](#3-Visualization)
- [3.1 SER](#31-ser)
- [3.2 RE](#32-re)
- [4. Usage](#4-usage)
- [4.1 Prepare for the environment](#41-Prepare-for-the-environment)
- [4.2 Quick start](#42-Quick-start)
- [4.3 More](#43-More)
- [5. Reference](#5-Reference)
- [6. License](#6-License)
## 1. Introduction
Key information extraction (KIE) refers to extracting key information from text or images. As downstream task of OCR, the key information extraction task of document image has many practical application scenarios, such as form recognition, ticket information extraction, ID card information extraction, etc.
PP-Structure conducts research based on the LayoutXLM multi-modal, and proposes the VI-LayoutXLM, which gets rid of visual features when finetuning the downstream tasks. An textline sorting method is also utilized to fit in reading order. What's more, UDML knowledge distillation is used for higher accuracy. Finally, the accuracy and inference speed of VI-LayoutXLM surpass those of LayoutXLM.
The main features of the key information extraction module in PP-Structure are as follows.
- Integrate multi-modal methods such as [LayoutXLM](https://arxiv.org/pdf/2104.08836.pdf), VI-LayoutXLM, and PP-OCR inference engine.
- Supports Semantic Entity Recognition (SER) and Relation Extraction (RE) tasks based on multimodal methods. Based on the SER task, the text recognition and classification in the image can be completed; based on the RE task, the relationship extraction of the text content in the image can be completed, such as judging the problem pair (pair).
- Supports custom training for SER tasks and RE tasks.
- Supports end-to-end system prediction and evaluation of OCR+SER.
- Supports end-to-end system prediction of OCR+SER+RE.
- Support SER model export and inference using PaddleInference.
## 2. Accuracy and performance
We evaluate the methods on the Chinese dataset of [XFUND](https://github.com/doc-analysis/XFUND), and the performance is as follows
|Model | Backbone | Task | Config file | Hmean | Inference time (ms) | Download link|
| --- | --- | --- | --- | --- | --- | --- |
|VI-LayoutXLM| VI-LayoutXLM-base | SER | [ser_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml)|**93.19%**| 15.49|[trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)|
|LayoutXLM| LayoutXLM-base | SER | [ser_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml)|90.38%| 19.49 | [trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)|
|VI-LayoutXLM| VI-LayoutXLM-base | RE | [re_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml)|**83.92%**| 15.49|[trained model](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar)|
|LayoutXLM| LayoutXLM-base | RE | [re_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml)|74.83%| 19.49|[trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar)|
* Note：Inference environment：V100 GPU + cuda10.2 + cudnn8.1.1 + TensorRT 7.2.3.4，tested using fp16.
For more KIE models in PaddleOCR, please refer to [KIE model zoo](../../doc/doc_en/algorithm_overview_en.md).
## 3. Visualization
There are two main solutions to the key information extraction task based on VI-LayoutXLM series model.
(1) Text detection + text recognition + semantic entity recognition (SER)
(2) Text detection + text recognition + semantic entity recognition (SER) + relationship extraction (RE)
The following images are demo results of the SER and RE models. For more detailed introduction to the above solutions, please refer to [KIE Guide](./how_to_do_kie.md).
### 3.1 SER
Demo results for SER task are as follows.
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539141-68e71c75-5cf7-4529-b2ca-219d29fa5f68.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185310636-6ce02f7c-790d-479f-b163-ea97a5a04808.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539517-ccf2372a-f026-4a7c-ad28-c741c770f60a.png" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539735-37b5c2ef-629d-43fe-9abb-44bb717ef7ee.jpg" width="600">
</div>
**Note:** test pictures are from [xfund dataset](https://github.com/doc-analysis/XFUND), [invoice dataset](https://aistudio.baidu.com/aistudio/datasetdetail/165561) and a composite ID card dataset.
Boxes of different colors in the image represent different categories.
The invoice and application form images have three categories: `request`, `answer` and `header`. The `question` and 'answer' can be used to extract the relationship.
For the ID card image, the mdoel can be directly identify the key information such as `name`, `gender`, `nationality`, so that the subsequent relationship extraction process is not required, and the key information extraction task can be completed using only on model.
### 3.2 RE
Demo results for RE task are as follows.
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185393805-c67ff571-cf7e-4217-a4b0-8b396c4f22bb.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185540080-0431e006-9235-4b6d-b63d-0b3c6e1de48f.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185540291-f64e5daf-6d42-4e7c-bbbb-471e3fac4fcc.png" width="600">
</div>
Red boxes are questions, blue boxes are answers. The green lines means the two conected objects are a pair.
## 4. Usage
### 4.1 Prepare for the environment
Use the following command to install KIE dependencies.
```bash
git clone https://github.com/PaddlePaddle/PaddleOCR.git
cd PaddleOCR
pip install -r requirements.txt
pip install -r ppstructure/kie/requirements.txt
# 安装PaddleOCR引擎用于预测
pip install paddleocr -U
```
The visualized results of SER are saved in the `./output` folder by default. Examples of results are as follows.
<div align="center">
<img src="../../ppstructure/docs/kie/result_ser/zh_val_42_ser.jpg" width="800">
</div>
### 4.2 Quick start
Here we use XFUND dataset to quickly experience the SER model and RE model.
#### 4.2.1 Prepare for the dataset
```bash
mkdir train_data
cd train_data
# download and uncompress the dataset
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar && tar -xf XFUND.tar
cd ..
```
#### 4.2.2 Predict images using the trained model
Use the following command to download the models.
```bash
mkdir pretrained_model
cd pretrained_model
# download and uncompress the SER trained model
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar && tar -xf ser_vi_layoutxlm_xfund_pretrained.tar
# download and uncompress the RE trained model
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar && tar -xf re_vi_layoutxlm_xfund_pretrained.tar
```
If you want to use OCR engine to obtain end-to-end prediction results, you can use the following command to predict.
```bash
# just predict using SER trained model
python3 tools/infer_kie_token_ser.py \
-c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/ser_vi_layoutxlm_xfund_pretrained/best_accuracy \
Global.infer_img=./ppstructure/docs/kie/input/zh_val_42.jpg
# predict using SER and RE trained model at the same time
python3 ./tools/infer_kie_token_ser_re.py \
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_xfund_pretrained/best_accuracy \
Global.infer_img=./train_data/XFUND/zh_val/image/zh_val_42.jpg \
-c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o_ser Architecture.Backbone.checkpoints=./pretrain_models/ser_vi_layoutxlm_xfund_pretrained/best_accuracy
```
The visual result images and the predicted text file will be saved in the `Global.save_res_path` directory.
If you want to load the text detection and recognition results collected before, you can use the following command to predict.
```bash
# just predict using SER trained model
python3 tools/infer_kie_token_ser.py \
-c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/ser_vi_layoutxlm_xfund_pretrained/best_accuracy \
Global.infer_img=./train_data/XFUND/zh_val/val.json \
Global.infer_mode=False
# predict using SER and RE trained model at the same time
python3 ./tools/infer_kie_token_ser_re.py \
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_xfund_pretrained/best_accuracy \
Global.infer_img=./train_data/XFUND/zh_val/val.json \
Global.infer_mode=False \
-c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o_ser Architecture.Backbone.checkpoints=./pretrain_models/ser_vi_layoutxlm_xfund_pretrained/best_accuracy
```
#### 4.2.3 Inference using PaddleInference
At present, only SER model supports inference using PaddleInference.
Firstly, download the inference SER inference model.
```bash
mkdir inference
cd inference
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_infer.tar && tar -xf ser_vi_layoutxlm_xfund_infer.tar
```
Use the following command for inference.
```bash
cd ppstructure
python3 kie/predict_kie_token_ser.py \
--kie_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_vi_layoutxlm_xfund_infer \
--image_dir=./docs/kie/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx"
```
The visual results and text file will be saved in directory `output`.
### 4.3 More
For training, evaluation and inference tutorial for KIE models, please refer to [KIE doc](../../doc/doc_en/kie_en.md).
For training, evaluation and inference tutorial for text detection models, please refer to [text detection doc](../../doc/doc_en/detection_en.md).
For training, evaluation and inference tutorial for text recognition models, please refer to [text recognition doc](../../doc/doc_en/recognition.md).
If you want to finish the KIE tasks in your scene, and don't know what to prepare, please refer to [End cdoc](../../doc/doc_en/recognition.md).
To complete the key information extraction task in your own scenario from data preparation to model selection, please refer to: [Guide to End-to-end KIE](./how_to_do_kie_en.md)。
## 5. Reference
- LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding, https://arxiv.org/pdf/2104.08836.pdf
- microsoft/unilm/layoutxlm, https://github.com/microsoft/unilm/tree/master/layoutxlm
- XFUND dataset, https://github.com/doc-analysis/XFUND
## 6. License
The content of this project itself is licensed under the [Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/)
ppstructure/kie/README_ch.md 0 → 100644
浏览文件 @ 96a8e0f8
[English](README.md) | 简体中文
# 关键信息抽取
- [1. 简介](#1-简介)
- [2. 精度与性能](#2-精度与性能)
- [3. 效果演示](#3-效果演示)
- [3.1 SER](#31-ser)
- [3.2 RE](#32-re)
- [4. 使用](#4-使用)
- [4.1 准备环境](#41-准备环境)
- [4.2 快速开始](#42-快速开始)
- [4.3 更多](#43-更多)
- [5. 参考链接](#5-参考链接)
- [6. License](#6-License)
## 1. 简介
关键信息抽取 (Key Information Extraction, KIE)指的是是从文本或者图像中，抽取出关键的信息。针对文档图像的关键信息抽取任务作为OCR的下游任务，存在非常多的实际应用场景，如表单识别、车票信息抽取、身份证信息抽取等。
PP-Structure 基于 LayoutXLM 文档多模态系列方法进行研究与优化，设计了视觉特征无关的多模态模型结构VI-LayoutXLM，同时引入符合阅读顺序的文本行排序方法以及UDML联合互学习蒸馏方法，最终在精度与速度均超越LayoutXLM。
PP-Structure中关键信息抽取模块的主要特性如下：
- 集成[LayoutXLM](https://arxiv.org/pdf/2104.08836.pdf)、VI-LayoutXLM等多模态模型以及PP-OCR预测引擎。
- 支持基于多模态方法的语义实体识别 (Semantic Entity Recognition, SER) 以及关系抽取 (Relation Extraction, RE) 任务。基于 SER 任务，可以完成对图像中的文本识别与分类；基于 RE 任务，可以完成对图象中的文本内容的关系提取，如判断问题对(pair)。
- 支持SER任务和RE任务的自定义训练。
- 支持OCR+SER的端到端系统预测与评估。
- 支持OCR+SER+RE的端到端系统预测。
- 支持SER模型的动转静导出与基于PaddleInfernece的模型推理。
## 2. 精度与性能
我们在 [XFUND](https://github.com/doc-analysis/XFUND) 的中文数据集上对算法进行了评估，SER与RE上的任务性能如下
|模型|骨干网络|任务|配置文件|hmean|预测耗时(ms)|下载链接|
| --- | --- | --- | --- | --- | --- | --- |
|VI-LayoutXLM| VI-LayoutXLM-base | SER | [ser_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh_udml.yml)|**93.19%**| 15.49|[训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar)|
|LayoutXLM| LayoutXLM-base | SER | [ser_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/ser_layoutxlm_xfund_zh.yml)|90.38%| 19.49 | [训练模型](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar)|
|VI-LayoutXLM| VI-LayoutXLM-base | RE | [re_vi_layoutxlm_xfund_zh_udml.yml](../../configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh_udml.yml)|**83.92%**| 15.49|[训练模型](https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar)|
|LayoutXLM| LayoutXLM-base | RE | [re_layoutxlm_xfund_zh.yml](../../configs/kie/layoutlm_series/re_layoutxlm_xfund_zh.yml)|74.83%| 19.49|[训练模型](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar)|
* 注：预测耗时测试条件：V100 GPU + cuda10.2 + cudnn8.1.1 + TensorRT 7.2.3.4，使用FP16进行测试。
更多关于PaddleOCR中关键信息抽取模型的介绍，请参考[关键信息抽取模型库](../../doc/doc_ch/algorithm_overview.md)。
## 3. 效果演示
基于多模态模型的关键信息抽取任务有2种主要的解决方案。
（1）文本检测 + 文本识别 + 语义实体识别(SER)
（2）文本检测 + 文本识别 + 语义实体识别(SER) + 关系抽取(RE)
下面给出SER与RE任务的示例效果，关于上述解决方案的详细介绍，请参考[关键信息抽取全流程指南](./how_to_do_kie.md)。
### 3.1 SER
对于SER任务，效果如下所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539141-68e71c75-5cf7-4529-b2ca-219d29fa5f68.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185310636-6ce02f7c-790d-479f-b163-ea97a5a04808.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539517-ccf2372a-f026-4a7c-ad28-c741c770f60a.png" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185539735-37b5c2ef-629d-43fe-9abb-44bb717ef7ee.jpg" width="600">
</div>
**注意：** 测试图片来源于[XFUND数据集](https://github.com/doc-analysis/XFUND)、[发票数据集](https://aistudio.baidu.com/aistudio/datasetdetail/165561)以及合成的身份证数据集。
图中不同颜色的框表示不同的类别。
图中的发票以及申请表图像，有`QUESTION`, `ANSWER`, `HEADER` 3种类别，识别的`QUESTION`, `ANSWER`可以用于后续的问题与答案的关系抽取。
图中的身份证图像，则直接识别出其中的`姓名`、`性别`、`民族`等关键信息，这样就无需后续的关系抽取过程，一个模型即可完成关键信息抽取。
### 3.2 RE
对于RE任务，效果如下所示。
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185393805-c67ff571-cf7e-4217-a4b0-8b396c4f22bb.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185540080-0431e006-9235-4b6d-b63d-0b3c6e1de48f.jpg" width="600">
</div>
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185540291-f64e5daf-6d42-4e7c-bbbb-471e3fac4fcc.png" width="600">
</div>
红色框是问题，蓝色框是答案。绿色线条表示连接的两端为一个key-value的pair。
## 4. 使用
### 4.1 准备环境
使用下面的命令安装运行SER与RE关键信息抽取的依赖。
```bash
git clone https://github.com/PaddlePaddle/PaddleOCR.git
cd PaddleOCR
pip install -r requirements.txt
pip install -r ppstructure/kie/requirements.txt
# 安装PaddleOCR引擎用于预测
pip install paddleocr -U
```
### 4.2 快速开始
下面XFUND数据集，快速体验SER模型与RE模型。
#### 4.2.1 准备数据
```bash
mkdir train_data
cd train_data
# 下载与解压数据
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar && tar -xf XFUND.tar
cd ..
```
#### 4.2.2 基于动态图的预测
首先下载模型。
```bash
mkdir pretrained_model
cd pretrained_model
# 下载并解压SER预训练模型
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_pretrained.tar && tar -xf ser_vi_layoutxlm_xfund_pretrained.tar
# 下载并解压RE预训练模型
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/re_vi_layoutxlm_xfund_pretrained.tar && tar -xf re_vi_layoutxlm_xfund_pretrained.tar
```
如果希望使用OCR引擎，获取端到端的预测结果，可以使用下面的命令进行预测。
```bash
# 仅预测SER模型
python3 tools/infer_kie_token_ser.py \
-c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/ser_vi_layoutxlm_xfund_pretrained/best_accuracy \
Global.infer_img=./ppstructure/docs/kie/input/zh_val_42.jpg
# SER + RE模型串联
python3 ./tools/infer_kie_token_ser_re.py \
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_xfund_pretrained/best_accuracy \
Global.infer_img=./train_data/XFUND/zh_val/image/zh_val_42.jpg \
-c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o_ser Architecture.Backbone.checkpoints=./pretrain_models/ser_vi_layoutxlm_xfund_pretrained/best_accuracy
```
`Global.save_res_path`目录中会保存可视化的结果图像以及预测的文本文件。
如果希望加载标注好的文本检测与识别结果，仅预测可以使用下面的命令进行预测。
```bash
# 仅预测SER模型
python3 tools/infer_kie_token_ser.py \
-c configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/ser_vi_layoutxlm_xfund_pretrained/best_accuracy \
Global.infer_img=./train_data/XFUND/zh_val/val.json \
Global.infer_mode=False
# SER + RE模型串联
python3 ./tools/infer_kie_token_ser_re.py \
-c configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yml \
-o Architecture.Backbone.checkpoints=./pretrain_models/re_vi_layoutxlm_xfund_pretrained/best_accuracy \
Global.infer_img=./train_data/XFUND/zh_val/val.json \
Global.infer_mode=False \
-c_ser configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yml \
-o_ser Architecture.Backbone.checkpoints=./pretrain_models/ser_vi_layoutxlm_xfund_pretrained/best_accuracy
```
#### 4.2.3 基于PaddleInference的预测
目前仅SER模型支持PaddleInference推理。
首先下载SER的推理模型。
```bash
mkdir inference
cd inference
wget https://paddleocr.bj.bcebos.com/ppstructure/models/vi_layoutxlm/ser_vi_layoutxlm_xfund_infer.tar && tar -xf ser_vi_layoutxlm_xfund_infer.tar
```
执行下面的命令进行预测。
```bash
cd ppstructure
python3 kie/predict_kie_token_ser.py \
--kie_algorithm=LayoutXLM \
--ser_model_dir=../inference/ser_vi_layoutxlm_xfund_infer \
--image_dir=./docs/kie/input/zh_val_42.jpg \
--ser_dict_path=../train_data/XFUND/class_list_xfun.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--ocr_order_method="tb-yx"
```
可视化结果保存在`output`目录下。
### 4.3 更多
关于KIE模型的训练评估与推理，请参考：[关键信息抽取教程](../../doc/doc_ch/kie.md)。
关于文本检测模型的训练评估与推理，请参考：[文本检测教程](../../doc/doc_ch/detection.md)。
关于文本识别模型的训练评估与推理，请参考：[文本识别教程](../../doc/doc_ch/recognition.md)。
关于怎样在自己的场景中完成关键信息抽取任务，请参考：[关键信息抽取全流程指南](./how_to_do_kie.md)。
## 5. 参考链接
- LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding, https://arxiv.org/pdf/2104.08836.pdf
- microsoft/unilm/layoutxlm, https://github.com/microsoft/unilm/tree/master/layoutxlm
- XFUND dataset, https://github.com/doc-analysis/XFUND
## 6. License
The content of this project itself is licensed under the [Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/)
ppstructure/kie/how_to_do_kie_en.md 0 → 100644
浏览文件 @ 96a8e0f8
# Key Information Extraction Pipeline
- [1. Introduction](#1-Introduction)
- [1.1 Background](#11-Background)
- [1.2 Mainstream Deep-learning Solutions](#12-Mainstream-Deep-learning-Solutions)
- [2. KIE Pipeline](#2-KIE-Pipeline)
- [2.1 Train OCR Models](#21-Train-OCR-Models)
- [2.2 Train KIE Models](#22-Train-KIE-Models)
- [3. Reference](#3-Reference)
## 1. Introduction
### 1.1 Background
Key information extraction (KIE) refers to extracting key information from text or images. As the downstream task of OCR, KIE of document image has many practical application scenarios, such as form recognition, ticket information extraction, ID card information extraction, etc. However, it is time-consuming and laborious to extract key information from these document images by manpower. It's challengable but also valuable to combine multi-modal features (visual, layout, text, etc) together and complete KIE tasks.
For the document images in a specific scene, the position and layout of the key information are relatively fixed. Therefore, in the early stage of the research, there are many methods based on template matching to extract the key information. This method is still widely used in many simple scenarios at present. However, it takes long time to adjut the template for different scenarios.
The KIE in the document image generally contains 2 subtasks, which is as shown follows.
* (1) SER: semantic entity recognition, which classifies each detected textline, such as dividing it into name and ID card. As shown in the red boxes in the following figure.
* (2) RE: relationship extraction, which matches the question and answer based on SER results. As shown in the figure below, the yellow arrows match the question and answer.
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185726510-faba470d-2c79-4784-b8da-6c1aa5af9572.png" width="800">
</div>
### 1.2 Mainstream Deep-learning Solutions
General KIE methods are based on Named Entity Recognition (NER), but such methods only use text information and ignore location and visual feature information, which leads to limited accuracy. In recent years, most scholars have started to combine mutil-modal features to improve the accuracy of KIE model. The main methods are as follows:
* (1) Grid based methods. These methods mainly focus on the fusion of multi-modal information at the image level. Most texts are of character granularity. The text and structure information embedding method is simple, such as the algorithm of chargrid [1].
* (2) Token based methods. These methods refer to the NLP methods such as Bert, which encode the position, vision and other feature information into the multi-modal model, and conduct pre-training on large-scale datasets, so that in downstream tasks, only a small amount of annotation data is required to obtain excellent results. The representative algorithms are layoutlm [2], layoutlmv2 [3], layoutxlm [4], structext [5], etc.
* (3) GCN based methods. These methods try to learn the structural information between images and characters, so as to solve the problem of extracting open set information (templates not seen in the training set), such as GCN [6], SDMGR [7] and other algorithms.
* (4) End to end based methods: these methods put the existing OCR character recognition and KIE information extraction tasks into a unified network for common learning, and strengthen each other in the learning process. Such as TRIE [8].
For more detailed introduction of the algorithms, please refer to Chapter 6 of [Diving into OCR](https://aistudio.baidu.com/aistudio/education/group/info/25207).
## 2. KIE Pipeline
Token based methods such as LayoutXLM are implemented in PaddleOCR. What's more, in PP-Structurev2, we simplify the LayoutXLM model and proposed VI-LayoutXLM, in which the visual feature extraction module is removed for speed-up. The textline sorting strategy conforming to the human reading order and UDML knowledge distillation strategy are utilized for higher model accuracy.
In the non end-to-end KIE method, KIE needs at least ** 2 steps**. Firstly, the OCR model is used to extract the text and its position. Secondly, the KIE model is used to extract the key information according to the image, text position and text content.
### 2.1 Train OCR Models
#### 2.1.1 Text Detection
**(1) Data**
Most of the models provided in PaddleOCR are general models. In the process of text detection, the detection of adjacent text lines is generally based on the distance of the position. As shown in the figure above, when using PP-OCRv3 general English detection model for text detection, it is easy to detect the two fields representing different propoerties as one. Therefore, it is suggested to finetune a detection model according to your scenario firstly during the KIE task.
During data annotation, the different key information needs to be separated. Otherwise, it will increase the difficulty of subsequent KIE tasks.
For downstream tasks, generally speaking, `200~300` training images can guarantee the basic training effect. If there is not too much prior knowledge, **`200~300`** images can be labeled firstly for subsequent text detection model training.
**(2) Model**
In terms of model selection, PP-OCRv3 detection model is recommended. For more information about the training methods of the detection model, please refer to: [Text detection tutorial](../../doc/doc_en/detection_en.md) and [PP-OCRv3 detection model tutorial](../../doc/doc_ch/PPOCRv3_det_train.md).
#### 2.1.2 Text recognition
Compared with the natural scene, the text recognition in the document image is generally relatively easier (the background is not too complex), so **it is suggested to** try the PP-OCRv3 general text recognition model provided in PaddleOCR ([PP-OCRv3 model list](../../doc/doc_en/models_list_en.md))
**(1) Data**
However, there are also some challenges in some document scenarios, such as rare words in ID card scenarios and special fonts in invoice and other scenarios. These problems will increase the difficulty of text recognition. At this time, if you want to ensure or further improve the model accuracy, it is recommended to load PP-OCRv3 model based on the text recognition dataset of specific document scenarios for finetuning.
In the process of model finetuning, it is recommended to prepare at least `5000` vertical scene text recognition images to ensure the basic model fine-tuning effect. If you want to improve the accuracy and generalization ability of the model, you can synthesize more text recognition images similar to the scene, collect general real text recognition data from the public data set, and add them to the text recognition training process. In the training process, it is suggested that the ratio of real data, synthetic data and general data of each epoch should be around `1:1:1`, which can be controlled by setting the sampling ratio of different data sources. If there are 3 training text files, including 10k, 20k and 50k pieces of data respectively, the data can be set in the configuration file as follows:
```yml
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/
label_file_list:
- ./train_data/train_list_10k.txt
- ./train_data/train_list_10k.txt
- ./train_data/train_list_50k.txt
ratio_list: [1.0, 0.5, 0.2]
...
```
**(2) Model**
In terms of model selection, PP-OCRv3 recognition model is recommended. For more information about the training methods of the recognition model, please refer to: [Text recognition tutorial](../../doc/doc_en/recognition_en.md) and [PP-OCRv3 model list](../../doc/doc_en/models_list_en.md).
### 2.2 Train KIE Models
There are two main methods to extract the key information from the recognized texts.
(1) Directly use SER model to obtain the key information category. For example, in the ID card scenario, we mark "name" and "Geoff Sample" as "name_key" and "name_value", respectively. The **text field** corresponding to the category "name_value" finally identified is the key information we need.
(2) Joint use SER and RE models. For this case, we firstly use SER model to obtain all questions (keys) and questions (values) for the image text, and then use RE model to match all keys and values to find the relationship, so as to complete the extraction of key information.
#### 2.2.1 SER
Take the ID card scenario as an example. The key information generally includes `name`, `DOB`, etc. We can directly mark the corresponding fields as specific categories, as shown in the following figure.
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185728456-dc396f47-0880-4279-9c7c-c99601bf16a7.png" width="500">
</div>
**Note:**
- In the labeling process, text content without key information about KIE shall be labeled as`other`, which is equivalent to background information. For example, in the ID card scenario, if we do not pay attention to `DOB` information, we can mark the categories of `DOB` and `Area manager` as `other`.
- In the annotation process of, it is required to annotate the **textline** position rather than the character.
In terms of data, generally speaking, for relatively fixed scenes, **50** training images can achieve acceptable effects. You can refer to [PPOCRLabel](../../PPOCRLabel/README.md) for finish the labeling process.
In terms of model, it is recommended to use the VI-layoutXLM model proposed in PP-Structurev2. It is improved based on the LayoutXLM model, removing the visual feature extraction module, and further improving the model inference speed without the significant reduction on model accuracy. For more tutorials, please refer to [VI-LayoutXLM introduction](../../doc/doc_en/algorithm_kie_vi_layoutxlm_en.md) and [KIE tutorial](../../doc/doc_en/kie_en.md).
#### 2.2.2 SER + RE
The SER model is mainly used to identify all keys and values in the document image, and the RE model is mainly used to match all keys and values.
Taking the ID card scenario as an example, the key information generally includes key information such as `name`, `DOB`, etc. in the SER stage, we need to identify all questions (keys) and answers (values). The demo annotation is as follows. All keys can be annotated as `question`, and all keys can be annotated as `answer`.
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185728881-b6055e01-034c-4584-aaa6-97c9c25fb61b.png" width="500">
</div>
In the RE stage, the ID and connection information of each field need to be marked, as shown in the following figure.
<div align="center">
<img src="https://user-images.githubusercontent.com/14270174/185728948-a4208013-5038-4025-9a93-0c6d51447488.png" width="500">
</div>
For each textline, you need to add 'ID' and 'linking' field information. The 'ID' records the unique identifier of the textline. Different text contents in the same images cannot be repeated. The 'linking' is a list that records the connection information between different texts. If the ID of the field "name" is 0 and the ID of the field "Geoff Sample" is 1, then they all have [[0, 1]] 'linking' marks, indicating that the fields with `id=0` and `id=1` form a key value relationship (the fields such as DOB and Expires are similar, and will not be repeated here).
**Note:**
-During annotation, if value is multiple textines, a key value pair can be added in linking, such as `[[0, 1], [0, 2]]`.
In terms of data, generally speaking, for relatively fixed scenes, about **50** training images can achieve acceptable effects.
In terms of model, it is recommended to use the VI-layoutXLM model proposed in PP-Structurev2. It is improved based on the LayoutXLM model, removing the visual feature extraction module, and further improving the model inference speed without the significant reduction on model accuracy. For more tutorials, please refer to [VI-LayoutXLM introduction](../../doc/doc_en/algorithm_kie_vi_layoutxlm_en.md) and [KIE tutorial](../../doc/doc_en/kie_en.md).
## 3. Reference
[1] Katti A R, Reisswig C, Guder C, et al. Chargrid: Towards understanding 2d documents[J]. arXiv preprint arXiv:1809.08799, 2018.
[2] Xu Y, Li M, Cui L, et al. Layoutlm: Pre-training of text and layout for document image understanding[C]//Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2020: 1192-1200.
[3] Xu Y, Xu Y, Lv T, et al. LayoutLMv2: Multi-modal pre-training for visually-rich document understanding[J]. arXiv preprint arXiv:2012.14740, 2020.
[4]: Xu Y, Lv T, Cui L, et al. Layoutxlm: Multimodal pre-training for multilingual visually-rich document understanding[J]. arXiv preprint arXiv:2104.08836, 2021.
[5] Li Y, Qian Y, Yu Y, et al. StrucTexT: Structured Text Understanding with Multi-Modal Transformers[C]//Proceedings of the 29th ACM International Conference on Multimedia. 2021: 1912-1920.
[6] Liu X, Gao F, Zhang Q, et al. Graph convolution for multimodal information extraction from visually rich documents[J]. arXiv preprint arXiv:1903.11279, 2019.
[7] Sun H, Kuang Z, Yue X, et al. Spatial Dual-Modality Graph Reasoning for Key Information Extraction[J]. arXiv preprint arXiv:2103.14470, 2021.
[8] Zhang P, Xu Y, Cheng Z, et al. Trie: End-to-end text reading and information extraction for document understanding[C]//Proceedings of the 28th ACM International Conference on Multimedia. 2020: 1413-1422.
ppstructure/vqa/predict_vqa_token_ser.py → ppstructure/kie/predict_kie_token_ser.py
浏览文件 @ 96a8e0f8
...@@ -30,7 +30,7 @@ from ppocr.data import create_operators, transform ...@@ -30,7 +30,7 @@ from ppocr.data import create_operators, transform
from ppocr.postprocess import build_post_process from ppocr.postprocess import build_post_process
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppocr.utils.visual import draw_ser_results from ppocr.utils.visual import draw_ser_results
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
from ppstructure.utility import parse_args from ppstructure.utility import parse_args
from paddleocr import PaddleOCR from paddleocr import PaddleOCR
...@@ -49,7 +49,7 @@ class SerPredictor(object): ...@@ -49,7 +49,7 @@ class SerPredictor(object):
pre_process_list = [{ pre_process_list = [{
'VQATokenLabelEncode': { 'VQATokenLabelEncode': {
'algorithm': args.vqa_algorithm, 'algorithm': args.kie_algorithm,
'class_path': args.ser_dict_path, 'class_path': args.ser_dict_path,
'contains_re': False, 'contains_re': False,
'ocr_engine': self.ocr_engine, 'ocr_engine': self.ocr_engine,
...@@ -138,7 +138,7 @@ def main(args): ...@@ -138,7 +138,7 @@ def main(args):
os.path.join(args.output, 'infer.txt'), mode='w', os.path.join(args.output, 'infer.txt'), mode='w',
encoding='utf-8') as f_w: encoding='utf-8') as f_w:
for image_file in image_file_list: for image_file in image_file_list:
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
img = cv2.imread(image_file) img = cv2.imread(image_file)
img = img[:, :, ::-1] img = img[:, :, ::-1]
......
ppstructure/vqa/requirements.txt → ppstructure/kie/requirements.txt
浏览文件 @ 96a8e0f8
sentencepiece sentencepiece
yacs yacs
seqeval seqeval
paddlenlp>=2.2.1 git+https://github.com/PaddlePaddle/PaddleNLP
pypandoc pypandoc
attrdict attrdict
python_docx python_docx
ppstructure/vqa/tools/eval_with_label_end2end.py → ppstructure/kie/tools/eval_with_label_end2end.py
浏览文件 @ 96a8e0f8
...@@ -20,7 +20,7 @@ from shapely.geometry import Polygon ...@@ -20,7 +20,7 @@ from shapely.geometry import Polygon
import numpy as np import numpy as np
from collections import defaultdict from collections import defaultdict
import operator import operator
import Levenshtein from rapidfuzz.distance import Levenshtein
import argparse import argparse
import json import json
import copy import copy
......
ppstructure/layout/README.md → ppstructure/layout/README_ch.md
浏览文件 @ 96a8e0f8
- [1. 简介](#1-简介) - [1. 简介](#1-简介)
- [2. 安装](#2-安装) - [2. 安装](#2-安装)
- [2.1 安装PaddlePaddle](#21-安装paddlepaddle) - [2.1 安装PaddlePaddle](#21-安装paddlepaddle)
- [2.2 安装PaddleDetection](#22-安装paddledetection) - [2.2 安装PaddleDetection](#22-安装paddledetection)
- [3. 数据准备](#3-数据准备) - [3. 数据准备](#3-数据准备)
- [3.1 英文数据集](#31-英文数据集) - [3.1 英文数据集](#31-英文数据集)
- [3.2 更多数据集](#32-更多数据集) - [3.2 更多数据集](#32-更多数据集)
- [4. 开始训练](#4-开始训练) - [4. 开始训练](#4-开始训练)
- [4.1 启动训练](#41-启动训练) - [4.1 启动训练](#41-启动训练)
- [4.2 FGD蒸馏训练](#42-FGD蒸馏训练) - [4.2 FGD蒸馏训练](#42-FGD蒸馏训练)
- [5. 模型评估与预测](#5-模型评估与预测) - [5. 模型评估与预测](#5-模型评估与预测)
- [5.1 指标评估](#51-指标评估) - [5.1 指标评估](#51-指标评估)
- [5.2 测试版面分析结果](#52-测试版面分析结果) - [5.2 测试版面分析结果](#52-测试版面分析结果)
- [6 模型导出与预测](#6-模型导出与预测) - [6 模型导出与预测](#6-模型导出与预测)
- [6.1 模型导出](#61-模型导出) - [6.1 模型导出](#61-模型导出)
- [6.2 模型推理](#62-模型推理) - [6.2 模型推理](#62-模型推理)
# 版面分析 # 版面分析
...@@ -63,7 +52,7 @@ python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simp ...@@ -63,7 +52,7 @@ python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simp
git clone https://github.com/PaddlePaddle/PaddleDetection.git git clone https://github.com/PaddlePaddle/PaddleDetection.git
``` ```
- **（2）安装其他依赖 ** - **（2）安装其他依赖**
```bash ```bash
cd PaddleDetection cd PaddleDetection
...@@ -166,15 +155,17 @@ json文件包含所有图像的标注，数据以字典嵌套的方式存放， ...@@ -166,15 +155,17 @@ json文件包含所有图像的标注，数据以字典嵌套的方式存放，
提供了训练脚本、评估脚本和预测脚本，本节将以PubLayNet预训练模型为例进行讲解。 提供了训练脚本、评估脚本和预测脚本，本节将以PubLayNet预训练模型为例进行讲解。
如果不希望训练，直接体验后面的模型评估、预测、动转静、推理的流程，可以下载提供的预训练模型，并跳过本部分。 如果不希望训练，直接体验后面的模型评估、预测、动转静、推理的流程，可以下载提供的预训练模型(PubLayNet数据集)，并跳过本部分。
``` ```
mkdir pretrained_model mkdir pretrained_model
cd pretrained_model cd pretrained_model
# 下载并解压PubLayNet预训练模型 # 下载PubLayNet预训练模型
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_layout.pdparams wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_layout.pdparams
``` ```
下载更多[版面分析模型](../docs/models_list.md)（中文CDLA数据集预训练模型、表格预训练模型）
### 4.1. 启动训练 ### 4.1. 启动训练
开始训练: 开始训练:
...@@ -184,7 +175,7 @@ wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_ ...@@ -184,7 +175,7 @@ wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_
如果你希望训练自己的数据集，需要修改配置文件中的数据配置、类别数。 如果你希望训练自己的数据集，需要修改配置文件中的数据配置、类别数。
以`configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml` 为例，修改的内容如下所示。 以`configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml` 为例，修改的内容如下所示。
```yaml ```yaml
metric: COCO metric: COCO
...@@ -223,16 +214,20 @@ TestDataset: ...@@ -223,16 +214,20 @@ TestDataset:
# 训练日志会自动保存到 log 目录中 # 训练日志会自动保存到 log 目录中
# 单卡训练 # 单卡训练
export CUDA_VISIBLE_DEVICES=0
python3 tools/train.py \ python3 tools/train.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \ -c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
--eval --eval
# 多卡训练，通过--gpus参数指定卡号 # 多卡训练，通过--gpus参数指定卡号
export CUDA_VISIBLE_DEVICES=0,1,2,3
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \ -c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
--eval --eval
``` ```
**注意：**如果训练时显存out memory，将TrainReader中batch_size调小，同时LearningRate中base_lr等比例减小。发布的config均由8卡训练得到，如果改变GPU卡数为1，那么base_lr需要减小8倍。
正常启动训练后，会看到以下log输出： 正常启动训练后，会看到以下log输出：
``` ```
...@@ -254,9 +249,11 @@ PaddleDetection支持了基于FGD([Focal and Global Knowledge Distillation for D ...@@ -254,9 +249,11 @@ PaddleDetection支持了基于FGD([Focal and Global Knowledge Distillation for D
更换数据集，修改【TODO】配置中的数据配置、类别数，具体可以参考4.1。启动训练： 更换数据集，修改【TODO】配置中的数据配置、类别数，具体可以参考4.1。启动训练：
```bash ```bash
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \ # 单卡训练
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \ export CUDA_VISIBLE_DEVICES=0
--slim_config configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x2_5_layout.yml \ python3 tools/train.py \
-c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
--slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
--eval --eval
``` ```
...@@ -267,13 +264,13 @@ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \ ...@@ -267,13 +264,13 @@ python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \
### 5.1. 指标评估 ### 5.1. 指标评估
训练中模型参数默认保存在`output/picodet_lcnet_x1_0_layout`目录下。在评估指标时，需要设置`weights`指向保存的参数文件。评估数据集可以通过 `configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml` 修改`EvalDataset`中的 `image_dir`、`anno_path`和`dataset_dir` 设置。 训练中模型参数默认保存在`output/picodet_lcnet_x1_0_layout`目录下。在评估指标时，需要设置`weights`指向保存的参数文件。评估数据集可以通过 `configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml` 修改`EvalDataset`中的 `image_dir`、`anno_path`和`dataset_dir` 设置。
```bash ```bash
# GPU 评估， weights 为待测权重 # GPU 评估， weights 为待测权重
python3 tools/eval.py \ python3 tools/eval.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \ -c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
-o weigths=./output/picodet_lcnet_x1_0_layout/best_model -o weights=./output/picodet_lcnet_x1_0_layout/best_model
``` ```
会输出以下信息，打印出mAP、AP0.5等信息。 会输出以下信息，打印出mAP、AP0.5等信息。
...@@ -299,8 +296,8 @@ python3 tools/eval.py \ ...@@ -299,8 +296,8 @@ python3 tools/eval.py \
``` ```
python3 tools/eval.py \ python3 tools/eval.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \ -c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
--slim_config configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x2_5_layout.yml \ --slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
-o weights=output/picodet_lcnet_x2_5_layout/best_model -o weights=output/picodet_lcnet_x2_5_layout/best_model
``` ```
...@@ -311,18 +308,17 @@ python3 tools/eval.py \ ...@@ -311,18 +308,17 @@ python3 tools/eval.py \
### 5.2. 测试版面分析结果 ### 5.2. 测试版面分析结果
预测使用的配置文件必须与训练一致，如您通过 `python3 tools/train.py -c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml` 完成了模型的训练过程。 预测使用的配置文件必须与训练一致，如您通过 `python3 tools/train.py -c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml` 完成了模型的训练过程。
使用 PaddleDetection 训练好的模型，您可以使用如下命令进行中文模型预测。
使用 PaddleDetection 训练好的模型，您可以使用如下命令进行模型预测。
```bash ```bash
python3 tools/infer.py \ python3 tools/infer.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \ -c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
-o weights='output/picodet_lcnet_x1_0_layout/best_model.pdparams' \ -o weights='output/picodet_lcnet_x1_0_layout/best_model.pdparams' \
--infer_img='docs/images/layout.jpg' \ --infer_img='docs/images/layout.jpg' \
--output_dir=output_dir/ \ --output_dir=output_dir/ \
--draw_threshold=0.4 --draw_threshold=0.5
``` ```
- `--infer_img`: 推理单张图片，也可以通过`--infer_dir`推理文件中的所有图片。 - `--infer_img`: 推理单张图片，也可以通过`--infer_dir`推理文件中的所有图片。
...@@ -335,16 +331,15 @@ python3 tools/infer.py \ ...@@ -335,16 +331,15 @@ python3 tools/infer.py \
``` ```
python3 tools/infer.py \ python3 tools/infer.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \ -c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
--slim_config configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x2_5_layout.yml \ --slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
-o weights='output/picodet_lcnet_x2_5_layout/best_model.pdparams' \ -o weights='output/picodet_lcnet_x2_5_layout/best_model.pdparams' \
--infer_img='docs/images/layout.jpg' \ --infer_img='docs/images/layout.jpg' \
--output_dir=output_dir/ \ --output_dir=output_dir/ \
--draw_threshold=0.4 --draw_threshold=0.5
``` ```
## 6. 模型导出与预测 ## 6. 模型导出与预测
...@@ -356,7 +351,7 @@ inference 模型（`paddle.jit.save`保存的模型） 一般是模型训练， ...@@ -356,7 +351,7 @@ inference 模型（`paddle.jit.save`保存的模型） 一般是模型训练，
```bash ```bash
python3 tools/export_model.py \ python3 tools/export_model.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \ -c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
-o weights=output/picodet_lcnet_x1_0_layout/best_model \ -o weights=output/picodet_lcnet_x1_0_layout/best_model \
--output_dir=output_inference/ --output_dir=output_inference/
``` ```
...@@ -377,8 +372,8 @@ FGD蒸馏模型转inference模型步骤如下： ...@@ -377,8 +372,8 @@ FGD蒸馏模型转inference模型步骤如下：
```bash ```bash
python3 tools/export_model.py \ python3 tools/export_model.py \
-c configs/picodet/legacy_model/application/publayernet_lcnet_x1_5/picodet_student.yml \ -c configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x1_0_layout.yml \
--slim_config configs/picodet/legacy_model/application/publayernet_lcnet_x1_5/picodet_teacher.yml \ --slim_config configs/picodet/legacy_model/application/layout_analysis/picodet_lcnet_x2_5_layout.yml \
-o weights=./output/picodet_lcnet_x2_5_layout/best_model \ -o weights=./output/picodet_lcnet_x2_5_layout/best_model \
--output_dir=output_inference/ --output_dir=output_inference/
``` ```
...@@ -466,4 +461,3 @@ preprocess_time(ms): 2172.50, inference_time(ms): 11.90, postprocess_time(ms): 1 ...@@ -466,4 +461,3 @@ preprocess_time(ms): 2172.50, inference_time(ms): 11.90, postprocess_time(ms): 1
year={2022} year={2022}
} }
``` ```
ppstructure/layout/__init__.py 0 → 100644
浏览文件 @ 96a8e0f8
# copyright (c) 2020 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.
ppstructure/layout/layout_in_ocr.md 已删除 100644 → 0
浏览文件 @ 29268eaa
- [1. 简介](#1-简介)
- [2. 安装](#2-安装)
- [2.1 安装PaddlePaddle](#21-安装paddlepaddle)
- [2.2 安装PaddleDetection](#22-安装paddledetection)
- [3. 数据准备](#3-数据准备)
- [3.1 英文数据集](#31-英文数据集)
- [3.2 更多数据集](#32-更多数据集)
- [4. 开始训练](#4-开始训练)
- [4.1 启动训练](#41-启动训练)
- [4.2 FGD蒸馏训练](#42-FGD蒸馏训练)
- [5. 模型评估与预测](#5-模型评估与预测)
- [5.1 指标评估](#51-指标评估)
- [5.2 测试版面分析结果](#52-测试版面分析结果)
- [6 模型导出与预测](#6-模型导出与预测)
- [6.1 模型导出](#61-模型导出)
- [6.2 模型推理](#62-模型推理)
# 版面分析
## 1. 简介
版面分析指的是对图片形式的文档进行区域划分，定位其中的关键区域，如文字、标题、表格、图片等。版面分析算法基于[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection)的轻量模型PP-PicoDet进行开发。
<div align="center">
<img src="../docs/layout/layout.png" width="800">
</div>
## 2. 安装依赖
### 2.1. 安装PaddlePaddle
- **（1) 安装PaddlePaddle**
```bash
python3 -m pip install --upgrade pip
# GPU安装
python3 -m pip install "paddlepaddle-gpu>=2.2" -i https://mirror.baidu.com/pypi/simple
# CPU安装
python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simple
```
更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
### 2.2. 安装PaddleDetection
- **（1）下载PaddleDetection源码**
```bash
git clone https://github.com/PaddlePaddle/PaddleDetection.git
```
- **（2）安装其他依赖 **
```bash
cd PaddleDetection
python3 -m pip install -r requirements.txt
```
## 3. 数据准备
如果希望直接体验预测过程，可以跳过数据准备，下载我们提供的预训练模型。
### 3.1. 英文数据集
下载文档分析数据集[PubLayNet](https://developer.ibm.com/exchanges/data/all/publaynet/)（数据集96G），包含5个类：`{0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"}`
```
# 下载数据
wget https://dax-cdn.cdn.appdomain.cloud/dax-publaynet/1.0.0/publaynet.tar.gz
# 解压数据
tar -xvf publaynet.tar.gz
```
解压之后的**目录结构：**
```
|-publaynet
|- test
|- PMC1277013_00004.jpg
|- PMC1291385_00002.jpg
| ...
|- train.json
|- train
|- PMC1291385_00002.jpg
|- PMC1277013_00004.jpg
| ...
|- val.json
|- val
|- PMC538274_00004.jpg
|- PMC539300_00004.jpg
| ...
```
**数据分布：**
| File or Folder | Description | num |
| :------------- | :------------- | ------- |
| `train/` | 训练集图片 | 335,703 |
| `val/` | 验证集图片 | 11,245 |
| `test/` | 测试集图片 | 11,405 |
| `train.json` | 训练集标注文件 | - |
| `val.json` | 验证集标注文件 | - |
**标注格式：**
json文件包含所有图像的标注，数据以字典嵌套的方式存放，包含以下key：
- info，表示标注文件info。
- licenses，表示标注文件licenses。
- images，表示标注文件中图像信息列表，每个元素是一张图像的信息。如下为其中一张图像的信息：
```
{
'file_name': 'PMC4055390_00006.jpg', # file_name
'height': 601, # image height
'width': 792, # image width
'id': 341427 # image id
}
```
- annotations，表示标注文件中目标物体的标注信息列表，每个元素是一个目标物体的标注信息。如下为其中一个目标物体的标注信息：
```
{
'segmentation': # 物体的分割标注
'area': 60518.099043117836, # 物体的区域面积
'iscrowd': 0, # iscrowd
'image_id': 341427, # image id
'bbox': [50.58, 490.86, 240.15, 252.16], # bbox [x1,y1,w,h]
'category_id': 1, # category_id
'id': 3322348 # image id
}
```
### 3.2. 更多数据集
我们提供了CDLA(中文版面分析)、TableBank(表格版面分析)等数据集的下连接，处理为上述标注文件json格式，即可以按相同方式进行训练。
| dataset | 简介 |
| ------------------------------------------------------------ | ------------------------------------------------------------ |
| [cTDaR2019_cTDaR](https://cndplab-founder.github.io/cTDaR2019/) | 用于表格检测(TRACKA)和表格识别(TRACKB)。图片类型包含历史数据集(以cTDaR_t0开头，如cTDaR_t00872.jpg)和现代数据集(以cTDaR_t1开头，cTDaR_t10482.jpg)。 |
| [IIIT-AR-13K](http://cvit.iiit.ac.in/usodi/iiitar13k.php) | 手动注释公开的年度报告中的图形或页面而构建的数据集，包含5类：table, figure, natural image, logo, and signature |
| [CDLA](https://github.com/buptlihang/CDLA) | 中文文档版面分析数据集，面向中文文献类（论文）场景，包含10类：Table、Figure、Figure caption、Table、Table caption、Header、Footer、Reference、Equation |
| [TableBank](https://github.com/doc-analysis/TableBank) | 用于表格检测和识别大型数据集，包含Word和Latex2种文档格式 |
| [DocBank](https://github.com/doc-analysis/DocBank) | 使用弱监督方法构建的大规模数据集(500K文档页面)，用于文档布局分析，包含12类：Author、Caption、Date、Equation、Figure、Footer、List、Paragraph、Reference、Section、Table、Title |
## 4. 开始训练
提供了训练脚本、评估脚本和预测脚本，本节将以PubLayNet预训练模型为例进行讲解。
如果不希望训练，直接体验后面的模型评估、预测、动转静、推理的流程，可以下载提供的预训练模型，并跳过本部分。
```
mkdir pretrained_model
cd pretrained_model
# 下载并解压PubLayNet预训练模型
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_layout.pdparams
```
### 4.1. 启动训练
开始训练:
* 修改配置文件
如果你希望训练自己的数据集，需要修改配置文件中的数据配置、类别数。
以`configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml` 为例，修改的内容如下所示。
```yaml
metric: COCO
# 类别数
num_classes: 5
TrainDataset:
!COCODataSet
# 修改为你自己的训练数据目录
image_dir: train
# 修改为你自己的训练数据标签文件
anno_path: train.json
# 修改为你自己的训练数据根目录
dataset_dir: /root/publaynet/
data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']
EvalDataset:
!COCODataSet
# 修改为你自己的验证数据目录
image_dir: val
# 修改为你自己的验证数据标签文件
anno_path: val.json
# 修改为你自己的验证数据根目录
dataset_dir: /root/publaynet/
TestDataset:
!ImageFolder
# 修改为你自己的测试数据标签文件
anno_path: /root/publaynet/val.json
```
* 开始训练，在训练时，会默认下载PP-PicoDet预训练模型，这里无需预先下载。
```bash
# GPU训练 支持单卡，多卡训练
# 训练日志会自动保存到 log 目录中
# 单卡训练
python3 tools/train.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \
--eval
# 多卡训练，通过--gpus参数指定卡号
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \
--eval
```
正常启动训练后，会看到以下log输出：
```
[08/15 04:02:30] ppdet.utils.checkpoint INFO: Finish loading model weights: /root/.cache/paddle/weights/LCNet_x1_0_pretrained.pdparams
[08/15 04:02:46] ppdet.engine INFO: Epoch: [0] [ 0/1929] learning_rate: 0.040000 loss_vfl: 1.216707 loss_bbox: 1.142163 loss_dfl: 0.544196 loss: 2.903065 eta: 17 days, 13:50:26 batch_cost: 15.7452 data_cost: 2.9112 ips: 1.5243 images/s
[08/15 04:03:19] ppdet.engine INFO: Epoch: [0] [ 20/1929] learning_rate: 0.064000 loss_vfl: 1.180627 loss_bbox: 0.939552 loss_dfl: 0.442436 loss: 2.628206 eta: 2 days, 12:18:53 batch_cost: 1.5770 data_cost: 0.0008 ips: 15.2184 images/s
[08/15 04:03:47] ppdet.engine INFO: Epoch: [0] [ 40/1929] learning_rate: 0.088000 loss_vfl: 0.543321 loss_bbox: 1.071401 loss_dfl: 0.457817 loss: 2.057003 eta: 2 days, 0:07:03 batch_cost: 1.3190 data_cost: 0.0007 ips: 18.1954 images/s
[08/15 04:04:12] ppdet.engine INFO: Epoch: [0] [ 60/1929] learning_rate: 0.112000 loss_vfl: 0.630989 loss_bbox: 0.859183 loss_dfl: 0.384702 loss: 1.883143 eta: 1 day, 19:01:29 batch_cost: 1.2177 data_cost: 0.0006 ips: 19.7087 images/s
```
- `--eval`表示训练的同时，进行评估， 评估过程中默认将最佳模型，保存为 `output/picodet_lcnet_x1_0_layout/best_accuracy` 。
**注意，预测/评估时的配置文件请务必与训练一致。**
### 4.2. FGD蒸馏训练
PaddleDetection支持了基于FGD([Focal and Global Knowledge Distillation for Detectors](https://arxiv.org/abs/2111.11837v1))蒸馏的目标检测模型训练过程，FGD蒸馏分为两个部分`Focal`和`Global`。`Focal`蒸馏分离图像的前景和背景，让学生模型分别关注教师模型的前景和背景部分特征的关键像素；`Global`蒸馏部分重建不同像素之间的关系并将其从教师转移到学生，以补偿`Focal`蒸馏中丢失的全局信息。
更换数据集，修改【TODO】配置中的数据配置、类别数，具体可以参考4.1。启动训练：
```bash
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \
--slim_config configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x2_5_layout.yml \
--eval
```
- `-c`: 指定模型配置文件。
- `--slim_config`: 指定压缩策略配置文件。
## 5. 模型评估与预测
### 5.1. 指标评估
训练中模型参数默认保存在`output/picodet_lcnet_x1_0_layout`目录下。在评估指标时，需要设置`weights`指向保存的参数文件。评估数据集可以通过 `configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml` 修改`EvalDataset`中的 `image_dir`、`anno_path`和`dataset_dir` 设置。
```bash
# GPU 评估， weights 为待测权重
python3 tools/eval.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \
-o weigths=./output/picodet_lcnet_x1_0_layout/best_model
```
会输出以下信息，打印出mAP、AP0.5等信息。
```py
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.935
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.979
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.956
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.404
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.782
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.969
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.539
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.938
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.949
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.495
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.818
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.978
[08/15 07:07:09] ppdet.engine INFO: Total sample number: 11245, averge FPS: 24.405059207157436
[08/15 07:07:09] ppdet.engine INFO: Best test bbox ap is 0.935.
```
使用FGD蒸馏模型进行评估：
```
python3 tools/eval.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \
--slim_config configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x2_5_layout.yml \
-o weights=output/picodet_lcnet_x2_5_layout/best_model
```
- `-c`: 指定模型配置文件。
- `--slim_config`: 指定蒸馏策略配置文件。
- `-o weights`: 指定蒸馏算法训好的模型路径。
### 5.2. 测试版面分析结果
预测使用的配置文件必须与训练一致，如您通过 `python3 tools/train.py -c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml` 完成了模型的训练过程。
使用 PaddleDetection 训练好的模型，您可以使用如下命令进行中文模型预测。
```bash
python3 tools/infer.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \
-o weights='output/picodet_lcnet_x1_0_layout/best_model.pdparams' \
--infer_img='docs/images/layout.jpg' \
--output_dir=output_dir/ \
--draw_threshold=0.4
```
- `--infer_img`: 推理单张图片，也可以通过`--infer_dir`推理文件中的所有图片。
- `--output_dir`: 指定可视化结果保存路径。
- `--draw_threshold`:指定绘制结果框的NMS阈值。
预测图片如下所示，图片会存储在`output_dir`路径中。
使用FGD蒸馏模型进行测试：
```
python3 tools/infer.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \
--slim_config configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x2_5_layout.yml \
-o weights='output/picodet_lcnet_x2_5_layout/best_model.pdparams' \
--infer_img='docs/images/layout.jpg' \
--output_dir=output_dir/ \
--draw_threshold=0.4
```
## 6. 模型导出与预测
### 6.1 模型导出
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。 训练过程中保存的模型是checkpoints模型，保存的只有模型的参数，多用于恢复训练等。 与checkpoints模型相比，inference 模型会额外保存模型的结构信息，在预测部署、加速推理上性能优越，灵活方便，适合于实际系统集成。
版面分析模型转inference模型步骤如下：
```bash
python3 tools/export_model.py \
-c configs/picodet/legacy_model/application/layout_detection/picodet_lcnet_x1_0_layout.yml \
-o weights=output/picodet_lcnet_x1_0_layout/best_model \
--output_dir=output_inference/
```
* 如无需导出后处理，请指定：`-o export.benchmark=True`（如果-o已出现过，此处删掉-o）
* 如无需导出NMS，请指定：`-o export.nms=False`
转换成功后，在目录下有三个文件：
```
output_inference/picodet_lcnet_x1_0_layout/
├── model.pdiparams # inference模型的参数文件
├── model.pdiparams.info # inference模型的参数信息，可忽略
└── model.pdmodel # inference模型的模型结构文件
```
FGD蒸馏模型转inference模型步骤如下：
```bash
python3 tools/export_model.py \
-c configs/picodet/legacy_model/application/publayernet_lcnet_x1_5/picodet_student.yml \
--slim_config configs/picodet/legacy_model/application/publayernet_lcnet_x1_5/picodet_teacher.yml \
-o weights=./output/picodet_lcnet_x2_5_layout/best_model \
--output_dir=output_inference/
```
### 6.2 模型推理
版面恢复任务进行推理，可以执行如下命令：
```bash
python3 deploy/python/infer.py \
--model_dir=output_inference/picodet_lcnet_x1_0_layout/ \
--image_file=docs/images/layout.jpg \
--device=CPU
```
- --device：指定GPU、CPU设备
模型推理完成，会看到以下log输出
```
------------------------------------------
----------- Model Configuration -----------
Model Arch: PicoDet
Transform Order:
--transform op: Resize
--transform op: NormalizeImage
--transform op: Permute
--transform op: PadStride
--------------------------------------------
class_id:0, confidence:0.9921, left_top:[20.18,35.66],right_bottom:[341.58,600.99]
class_id:0, confidence:0.9914, left_top:[19.77,611.42],right_bottom:[341.48,901.82]
class_id:0, confidence:0.9904, left_top:[369.36,375.10],right_bottom:[691.29,600.59]
class_id:0, confidence:0.9835, left_top:[369.60,608.60],right_bottom:[691.38,736.72]
class_id:0, confidence:0.9830, left_top:[369.58,805.38],right_bottom:[690.97,901.80]
class_id:0, confidence:0.9716, left_top:[383.68,271.44],right_bottom:[688.93,335.39]
class_id:0, confidence:0.9452, left_top:[370.82,34.48],right_bottom:[688.10,63.54]
class_id:1, confidence:0.8712, left_top:[370.84,771.03],right_bottom:[519.30,789.13]
class_id:3, confidence:0.9856, left_top:[371.28,67.85],right_bottom:[685.73,267.72]
save result to: output/layout.jpg
Test iter 0
------------------ Inference Time Info ----------------------
total_time(ms): 2196.0, img_num: 1
average latency time(ms): 2196.00, QPS: 0.455373
preprocess_time(ms): 2172.50, inference_time(ms): 11.90, postprocess_time(ms): 11.60
```
- Model：模型结构
- Transform Order：预处理操作
- class_id、confidence、left_top、right_bottom：分别表示类别id、置信度、左上角坐标、右下角坐标
- save result to：可视化版面分析结果保存路径，默认保存到`./output`文件夹
- Inference Time Info：推理时间，其中preprocess_time表示预处理耗时，inference_time表示模型预测耗时，postprocess_time表示后处理耗时
可视化版面结果如下图所示
<div align="center">
<img src="../docs/layout/layout_res.jpg" width="800">
</div>
## Citations
```
@inproceedings{zhong2019publaynet,
title={PubLayNet: largest dataset ever for document layout analysis},
author={Zhong, Xu and Tang, Jianbin and Yepes, Antonio Jimeno},
booktitle={2019 International Conference on Document Analysis and Recognition (ICDAR)},
year={2019},
volume={},
number={},
pages={1015-1022},
doi={10.1109/ICDAR.2019.00166},
ISSN={1520-5363},
month={Sep.},
organization={IEEE}
}
@inproceedings{yang2022focal,
title={Focal and global knowledge distillation for detectors},
author={Yang, Zhendong and Li, Zhe and Jiang, Xiaohu and Gong, Yuan and Yuan, Zehuan and Zhao, Danpei and Yuan, Chun},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={4643--4652},
year={2022}
}
```
ppstructure/layout/predict_layout.py
浏览文件 @ 96a8e0f8
...@@ -28,12 +28,13 @@ import tools.infer.utility as utility ...@@ -28,12 +28,13 @@ import tools.infer.utility as utility
from ppocr.data import create_operators, transform from ppocr.data import create_operators, transform
from ppocr.postprocess import build_post_process from ppocr.postprocess import build_post_process
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
from ppstructure.utility import parse_args from ppstructure.utility import parse_args
from picodet_postprocess import PicoDetPostProcess from picodet_postprocess import PicoDetPostProcess
logger = get_logger() logger = get_logger()
class LayoutPredictor(object): class LayoutPredictor(object):
def __init__(self, args): def __init__(self, args):
pre_process_list = [{ pre_process_list = [{
...@@ -109,7 +110,7 @@ def main(args): ...@@ -109,7 +110,7 @@ def main(args):
repeats = 50 repeats = 50
for image_file in image_file_list: for image_file in image_file_list:
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
img = cv2.imread(image_file) img = cv2.imread(image_file)
if img is None: if img is None:
......
ppstructure/predict_system.py
浏览文件 @ 96a8e0f8
...@@ -28,13 +28,12 @@ import time ...@@ -28,13 +28,12 @@ import time
import logging import logging
from copy import deepcopy from copy import deepcopy
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from tools.infer.predict_system import TextSystem from tools.infer.predict_system import TextSystem
from ppstructure.layout.predict_layout import LayoutPredictor from ppstructure.layout.predict_layout import LayoutPredictor
from ppstructure.table.predict_table import TableSystem, to_excel from ppstructure.table.predict_table import TableSystem, to_excel
from ppstructure.utility import parse_args, draw_structure_result from ppstructure.utility import parse_args, draw_structure_result
from ppstructure.recovery.recovery_to_doc import convert_info_docx
logger = get_logger() logger = get_logger()
...@@ -75,10 +74,10 @@ class StructureSystem(object): ...@@ -75,10 +74,10 @@ class StructureSystem(object):
else: else:
self.table_system = TableSystem(args) self.table_system = TableSystem(args)
elif self.mode == 'vqa': elif self.mode == 'kie':
raise NotImplementedError raise NotImplementedError
def __call__(self, img, return_ocr_result_in_table=False): def __call__(self, img, img_idx=0, return_ocr_result_in_table=False):
time_dict = { time_dict = {
'image_orientation': 0, 'image_orientation': 0,
'layout': 0, 'layout': 0,
...@@ -86,7 +85,7 @@ class StructureSystem(object): ...@@ -86,7 +85,7 @@ class StructureSystem(object):
'table_match': 0, 'table_match': 0,
'det': 0, 'det': 0,
'rec': 0, 'rec': 0,
'vqa': 0, 'kie': 0,
'all': 0 'all': 0
} }
start = time.time() start = time.time()
...@@ -169,36 +168,40 @@ class StructureSystem(object): ...@@ -169,36 +168,40 @@ class StructureSystem(object):
'type': region['label'].lower(), 'type': region['label'].lower(),
'bbox': [x1, y1, x2, y2], 'bbox': [x1, y1, x2, y2],
'img': roi_img, 'img': roi_img,
'res': res 'res': res,
'img_idx': img_idx
}) })
end = time.time() end = time.time()
time_dict['all'] = end - start time_dict['all'] = end - start
return res_list, time_dict return res_list, time_dict
elif self.mode == 'vqa': elif self.mode == 'kie':
raise NotImplementedError raise NotImplementedError
return None, None return None, None
def save_structure_res(res, save_folder, img_name): def save_structure_res(res, save_folder, img_name, img_idx=0):
excel_save_folder = os.path.join(save_folder, img_name) excel_save_folder = os.path.join(save_folder, img_name)
os.makedirs(excel_save_folder, exist_ok=True) os.makedirs(excel_save_folder, exist_ok=True)
res_cp = deepcopy(res) res_cp = deepcopy(res)
# save res # save res
with open( with open(
os.path.join(excel_save_folder, 'res.txt'), 'w', os.path.join(excel_save_folder, 'res_{}.txt'.format(img_idx)),
'w',
encoding='utf8') as f: encoding='utf8') as f:
for region in res_cp: for region in res_cp:
roi_img = region.pop('img') roi_img = region.pop('img')
f.write('{}\n'.format(json.dumps(region))) f.write('{}\n'.format(json.dumps(region)))
if region['type'] == 'table' and len(region[ if region['type'].lower() == 'table' and len(region[
'res']) > 0 and 'html' in region['res']: 'res']) > 0 and 'html' in region['res']:
excel_path = os.path.join(excel_save_folder, excel_path = os.path.join(
'{}.xlsx'.format(region['bbox'])) excel_save_folder,
'{}_{}.xlsx'.format(region['bbox'], img_idx))
to_excel(region['res']['html'], excel_path) to_excel(region['res']['html'], excel_path)
elif region['type'] == 'figure': elif region['type'].lower() == 'figure':
img_path = os.path.join(excel_save_folder, img_path = os.path.join(
'{}.jpg'.format(region['bbox'])) excel_save_folder,
'{}_{}.jpg'.format(region['bbox'], img_idx))
cv2.imwrite(img_path, roi_img) cv2.imwrite(img_path, roi_img)
...@@ -214,11 +217,13 @@ def main(args): ...@@ -214,11 +217,13 @@ def main(args):
for i, image_file in enumerate(image_file_list): for i, image_file in enumerate(image_file_list):
logger.info("[{}/{}] {}".format(i, img_num, image_file)) logger.info("[{}/{}] {}".format(i, img_num, image_file))
img, flag = check_and_read_gif(image_file) img, flag_gif, flag_pdf = check_and_read(image_file)
img_name = os.path.basename(image_file).split('.')[0] img_name = os.path.basename(image_file).split('.')[0]
if not flag: if not flag_gif and not flag_pdf:
img = cv2.imread(image_file) img = cv2.imread(image_file)
if not flag_pdf:
if img is None: if img is None:
logger.error("error in loading image:{}".format(image_file)) logger.error("error in loading image:{}".format(image_file))
continue continue
...@@ -228,14 +233,59 @@ def main(args): ...@@ -228,14 +233,59 @@ def main(args):
save_structure_res(res, save_folder, img_name) save_structure_res(res, save_folder, img_name)
draw_img = draw_structure_result(img, res, args.vis_font_path) draw_img = draw_structure_result(img, res, args.vis_font_path)
img_save_path = os.path.join(save_folder, img_name, 'show.jpg') img_save_path = os.path.join(save_folder, img_name, 'show.jpg')
elif structure_sys.mode == 'vqa': elif structure_sys.mode == 'kie':
raise NotImplementedError raise NotImplementedError
# draw_img = draw_ser_results(img, res, args.vis_font_path) # draw_img = draw_ser_results(img, res, args.vis_font_path)
# img_save_path = os.path.join(save_folder, img_name + '.jpg') # img_save_path = os.path.join(save_folder, img_name + '.jpg')
cv2.imwrite(img_save_path, draw_img) cv2.imwrite(img_save_path, draw_img)
logger.info('result save to {}'.format(img_save_path)) logger.info('result save to {}'.format(img_save_path))
if args.recovery: if args.recovery:
convert_info_docx(img, res, save_folder, img_name) try:
from ppstructure.recovery.recovery_to_doc import sorted_layout_boxes, convert_info_docx
h, w, _ = img.shape
res = sorted_layout_boxes(res, w)
convert_info_docx(img, res, save_folder, img_name,
args.save_pdf)
except Exception as ex:
logger.error(
"error in layout recovery image:{}, err msg: {}".format(
image_file, ex))
continue
else:
pdf_imgs = img
all_res = []
for index, img in enumerate(pdf_imgs):
res, time_dict = structure_sys(img, index)
if structure_sys.mode == 'structure' and res != []:
save_structure_res(res, save_folder, img_name, index)
draw_img = draw_structure_result(img, res,
args.vis_font_path)
img_save_path = os.path.join(save_folder, img_name,
'show_{}.jpg'.format(index))
elif structure_sys.mode == 'kie':
raise NotImplementedError
# draw_img = draw_ser_results(img, res, args.vis_font_path)
# img_save_path = os.path.join(save_folder, img_name + '.jpg')
if res != []:
cv2.imwrite(img_save_path, draw_img)
logger.info('result save to {}'.format(img_save_path))
if args.recovery and res != []:
from ppstructure.recovery.recovery_to_doc import sorted_layout_boxes, convert_info_docx
h, w, _ = img.shape
res = sorted_layout_boxes(res, w)
all_res += res
if args.recovery and all_res != []:
try:
convert_info_docx(img, all_res, save_folder, img_name,
args.save_pdf)
except Exception as ex:
logger.error(
"error in layout recovery image:{}, err msg: {}".format(
image_file, ex))
continue
logger.info("Predict time : {:.3f}s".format(time_dict['all'])) logger.info("Predict time : {:.3f}s".format(time_dict['all']))
......
ppstructure/recovery/README.md
浏览文件 @ 96a8e0f8
...@@ -6,6 +6,8 @@ English | [简体中文](README_ch.md) ...@@ -6,6 +6,8 @@ English | [简体中文](README_ch.md)
- [2.1 Installation dependencies](#2.1) - [2.1 Installation dependencies](#2.1)
- [2.2 Install PaddleOCR](#2.2) - [2.2 Install PaddleOCR](#2.2)
- [3. Quick Start](#3) - [3. Quick Start](#3)
- [3.1 Download models](#3.1)
- [3.2 Layout recovery](#3.2)
<a name="1"></a> <a name="1"></a>
...@@ -17,8 +19,9 @@ Layout recovery combines [layout analysis](../layout/README.md)、[table recogni ...@@ -17,8 +19,9 @@ Layout recovery combines [layout analysis](../layout/README.md)、[table recogni
The following figure shows the result： The following figure shows the result：
<div align="center"> <div align="center">
<img src="../docs/table/recovery.jpg" width = "700" /> <img src="../docs/recovery/recovery.jpg" width = "700" />
</div> </div>
<a name="2"></a> <a name="2"></a>
## 2. Install ## 2. Install
...@@ -33,14 +36,14 @@ The following figure shows the result： ...@@ -33,14 +36,14 @@ The following figure shows the result：
python3 -m pip install --upgrade pip python3 -m pip install --upgrade pip
# GPU installation # GPU installation
python3 -m pip install "paddlepaddle-gpu>=2.2" -i https://mirror.baidu.com/pypi/simple python3 -m pip install "paddlepaddle-gpu" -i https://mirror.baidu.com/pypi/simple
# CPU installation # CPU installation
python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simple python3 -m pip install "paddlepaddle" -i https://mirror.baidu.com/pypi/simple
```` ````
For more requirements, please refer to the instructions in [Installation Documentation](https://www.paddlepaddle.org.cn/install/quick). For more requirements, please refer to the instructions in [Installation Documentation](https://www.paddlepaddle.org.cn/en/install/quick?docurl=/documentation/docs/en/install/pip/macos-pip_en.html).
<a name="2.2"></a> <a name="2.2"></a>
...@@ -67,20 +70,61 @@ python3 -m pip install -r ppstructure/recovery/requirements.txt ...@@ -67,20 +70,61 @@ python3 -m pip install -r ppstructure/recovery/requirements.txt
## 3. Quick Start ## 3. Quick Start
```python <a name="3.1"></a>
### 3.1 Download models
If input is English document, download English models:
```bash
cd PaddleOCR/ppstructure cd PaddleOCR/ppstructure
# download model # download model
mkdir inference && cd inference mkdir inference && cd inference
# Download the detection model of the ultra-lightweight English PP-OCRv3 model and unzip it # Download the detection model of the ultra-lightweight English PP-OCRv3 model and unzip it
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar && tar xf ch_PP-OCRv3_det_infer.tar https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar && tar xf en_PP-OCRv3_det_infer.tar
# Download the recognition model of the ultra-lightweight English PP-OCRv3 model and unzip it # Download the recognition model of the ultra-lightweight English PP-OCRv3 model and unzip it
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar && tar xf ch_PP-OCRv3_rec_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_infer.tar && tar xf en_PP-OCRv3_rec_infer.tar
# Download the ultra-lightweight English table inch model and unzip it # Download the ultra-lightweight English table inch model and unzip it
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar && tar xf en_ppocr_mobile_v2.0_table_structure_infer.tar wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf en_ppstructure_mobile_v2.0_SLANet_infer.tar
# Download the layout model of publaynet dataset and unzip it
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_infer.tar && tar xf picodet_lcnet_x1_0_fgd_layout_infer.tar
cd .. cd ..
# run ```
python3 predict_system.py --det_model_dir=inference/en_PP-OCRv3_det_infer --rec_model_dir=inference/en_PP-OCRv3_rec_infer --table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer --rec_char_dict_path=../ppocr/utils/en_dict.txt --table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt --output ./output/table --rec_image_shape=3,48,320 --vis_font_path=../doc/fonts/simfang.ttf --recovery=True --image_dir=./docs/table/1.png If input is Chinese document，download Chinese models:
[Chinese and English ultra-lightweight PP-OCRv3 model](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/README.md#pp-ocr-series-model-listupdate-on-september-8th)、[表格识别模型](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/ppstructure/docs/models_list.md#22-表格识别模型)、[版面分析模型](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/ppstructure/docs/models_list.md#1-版面分析模型)
<a name="3.2"></a>
### 3.2 Layout recovery
```bash
python3 predict_system.py \
--image_dir=./docs/table/1.png \
--det_model_dir=inference/en_PP-OCRv3_det_infer \
--rec_model_dir=inference/en_PP-OCRv3_rec_infer \
--rec_char_dict_path=../ppocr/utils/en_dict.txt \
--table_model_dir=inference/en_ppstructure_mobile_v2.0_SLANet_infer \
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \
--layout_model_dir=inference/picodet_lcnet_x1_0_fgd_layout_infer \
--layout_dict_path=../ppocr/utils/dict/layout_dict/layout_publaynet_dict.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--recovery=True \
--save_pdf=False \
--output=../output/
``` ```
After running, the docx of each picture will be saved in the directory specified by the output field After running, the docx of each picture will be saved in the directory specified by the output field
Field：
- image_dir：test file测试文件， can be picture, picture directory, pdf file, pdf file directory
- det_model_dir：OCR detection model path
- rec_model_dir：OCR recognition model path
- rec_char_dict_path：OCR recognition dict path. If the Chinese model is used, change to "../ppocr/utils/ppocr_keys_v1.txt". And if you trained the model on your own dataset, change to the trained dictionary
- table_model_dir：tabel recognition model path
- table_char_dict_path：tabel recognition dict path. If the Chinese model is used, no need to change
- layout_model_dir：layout analysis model path
- layout_dict_path：layout analysis dict path. If the Chinese model is used, change to "../ppocr/utils/dict/layout_dict/layout_cdla_dict.txt"
- recovery：whether to enable layout of recovery, default False
- save_pdf：when recovery file, whether to save pdf file, default False
- output：save the recovery result path
ppstructure/recovery/README_ch.md
浏览文件 @ 96a8e0f8
...@@ -8,6 +8,8 @@ ...@@ -8,6 +8,8 @@
- [2.2 安装PaddleOCR](#2.2) - [2.2 安装PaddleOCR](#2.2)
- [3. 使用](#3) - [3. 使用](#3)
- [3.1 下载模型](#3.1)
- [3.2 版面恢复](#3.2)
<a name="1"></a> <a name="1"></a>
...@@ -16,11 +18,12 @@ ...@@ -16,11 +18,12 @@
版面恢复就是在OCR识别后，内容仍然像原文档图片那样排列着，段落不变、顺序不变的输出到word文档中等。 版面恢复就是在OCR识别后，内容仍然像原文档图片那样排列着，段落不变、顺序不变的输出到word文档中等。
版面恢复结合了[版面分析](../layout/README_ch.md)、[表格识别](../table/README_ch.md)技术，从而更好地恢复图片、表格、标题等内容，下图展示了版面恢复的结果： 版面恢复结合了[版面分析](../layout/README_ch.md)、[表格识别](../table/README_ch.md)技术，从而更好地恢复图片、表格、标题等内容，支持pdf文档、文档图片格式的输入文件，下图展示了版面恢复的结果：
<div align="center"> <div align="center">
<img src="../docs/table/recovery.jpg" width = "700" /> <img src="../docs/recovery/recovery.jpg" width = "700" />
</div> </div>
<a name="2"></a> <a name="2"></a>
## 2. 安装 ## 2. 安装
...@@ -35,21 +38,15 @@ ...@@ -35,21 +38,15 @@
python3 -m pip install --upgrade pip python3 -m pip install --upgrade pip
# GPU安装 # GPU安装
python3 -m pip install "paddlepaddle-gpu>=2.2" -i https://mirror.baidu.com/pypi/simple python3 -m pip install "paddlepaddle-gpu" -i https://mirror.baidu.com/pypi/simple
# CPU安装 # CPU安装
python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simple python3 -m pip install "paddlepaddle" -i https://mirror.baidu.com/pypi/simple
``` ```
更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。 更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
* **(2)安装依赖**
```bash
python3 -m pip install -r ppstructure/recovery/requirements.txt
```
<a name="2.2"></a> <a name="2.2"></a>
### 2.2 安装PaddleOCR ### 2.2 安装PaddleOCR
...@@ -75,23 +72,66 @@ python3 -m pip install -r ppstructure/recovery/requirements.txt ...@@ -75,23 +72,66 @@ python3 -m pip install -r ppstructure/recovery/requirements.txt
## 3. 使用 ## 3. 使用
恢复给定文档的版面： <a name="3.1"></a>
### 3.1 下载模型
如果输入为英文文档类型，下载英文模型
```python ```bash
cd PaddleOCR/ppstructure cd PaddleOCR/ppstructure
# 下载模型 # 下载模型
mkdir inference && cd inference mkdir inference && cd inference
# 下载超英文轻量级PP-OCRv3模型的检测模型并解压 # 下载英文超轻量PP-OCRv3检测模型并解压
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar && tar xf ch_PP-OCRv3_det_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar && tar xf en_PP-OCRv3_det_infer.tar
# 下载英文轻量级PP-OCRv3模型的识别模型并解压 # 下载英文超轻量PP-OCRv3识别模型并解压
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar && tar xf ch_PP-OCRv3_rec_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_infer.tar && tar xf en_PP-OCRv3_rec_infer.tar
# 下载超轻量级英文表格英寸模型并解压 # 下载英文表格识别模型并解压
wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar && tar xf en_ppocr_mobile_v2.0_table_structure_infer.tar wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/en_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf en_ppstructure_mobile_v2.0_SLANet_infer.tar
# 下载英文版面分析模型
wget https://paddleocr.bj.bcebos.com/ppstructure/models/layout/picodet_lcnet_x1_0_fgd_layout_infer.tar && tar xf picodet_lcnet_x1_0_fgd_layout_infer.tar
cd .. cd ..
# 执行预测
python3 predict_system.py --det_model_dir=inference/en_PP-OCRv3_det_infer --rec_model_dir=inference/en_PP-OCRv3_rec_infer --table_model_dir=inference/en_ppocr_mobile_v2.0_table_structure_infer --rec_char_dict_path=../ppocr/utils/en_dict.txt --table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt --output ./output/table --rec_image_shape=3,48,320 --vis_font_path=../doc/fonts/simfang.ttf --recovery=True --image_dir=./docs/table/1.png
``` ```
运行完成后，每张图片的docx文档会保存到output字段指定的目录下 如果输入为中文文档类型，在下述链接中下载中文模型即可：
[PP-OCRv3中英文超轻量文本检测和识别模型](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/README_ch.md#pp-ocr%E7%B3%BB%E5%88%97%E6%A8%A1%E5%9E%8B%E5%88%97%E8%A1%A8%E6%9B%B4%E6%96%B0%E4%B8%AD)、[表格识别模型](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/ppstructure/docs/models_list.md#22-表格识别模型)、[版面分析模型](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/ppstructure/docs/models_list.md#1-版面分析模型)
<a name="3.2"></a>
### 3.2 版面恢复
使用下载的模型恢复给定文档的版面，以英文模型为例，执行如下命令：
```bash
python3 predict_system.py \
--image_dir=./docs/table/1.png \
--det_model_dir=inference/en_PP-OCRv3_det_infer \
--rec_model_dir=inference/en_PP-OCRv3_rec_infer \
--rec_char_dict_path=../ppocr/utils/en_dict.txt \
--table_model_dir=inference/en_ppstructure_mobile_v2.0_SLANet_infer \
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict.txt \
--layout_model_dir=inference/picodet_lcnet_x1_0_fgd_layout_infer \
--layout_dict_path=../ppocr/utils/dict/layout_dict/layout_publaynet_dict.txt \
--vis_font_path=../doc/fonts/simfang.ttf \
--recovery=True \
--save_pdf=False \
--output=../output/
```
运行完成后，恢复版面的docx文档会保存到`output`字段指定的目录下
字段含义：
- image_dir：测试文件，可以是图片、图片目录、pdf文件、pdf文件目录
- det_model_dir：OCR检测模型路径
- rec_model_dir：OCR识别模型路径
- rec_char_dict_path：OCR识别字典，如果更换为中文模型，需要更改为"../ppocr/utils/ppocr_keys_v1.txt"，如果您在自己的数据集上训练的模型，则更改为训练的字典的文件
- table_model_dir：表格识别模型路径
- table_char_dict_path：表格识别字典，如果更换为中文模型，不需要更换字典
- layout_model_dir：版面分析模型路径
- layout_dict_path：版面分析字典，如果更换为中文模型，需要更改为"../ppocr/utils/dict/layout_dict/layout_cdla_dict.txt"
- recovery：是否进行版面恢复，默认False
- save_pdf：进行版面恢复导出docx文档的同时，是否保存为pdf文件，默认为False
- output：版面恢复结果保存路径
ppstructure/recovery/__init__.py 0 → 100644
浏览文件 @ 96a8e0f8
# copyright (c) 2020 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.
ppstructure/recovery/recovery_to_doc.py
浏览文件 @ 96a8e0f8
...@@ -22,21 +22,23 @@ from docx import shared ...@@ -22,21 +22,23 @@ from docx import shared
from docx.enum.text import WD_ALIGN_PARAGRAPH from docx.enum.text import WD_ALIGN_PARAGRAPH
from docx.enum.section import WD_SECTION from docx.enum.section import WD_SECTION
from docx.oxml.ns import qn from docx.oxml.ns import qn
from docx.enum.table import WD_TABLE_ALIGNMENT
from ppstructure.recovery.table_process import HtmlToDocx
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
logger = get_logger() logger = get_logger()
def convert_info_docx(img, res, save_folder, img_name): def convert_info_docx(img, res, save_folder, img_name, save_pdf):
doc = Document() doc = Document()
doc.styles['Normal'].font.name = 'Times New Roman' doc.styles['Normal'].font.name = 'Times New Roman'
doc.styles['Normal']._element.rPr.rFonts.set(qn('w:eastAsia'), u'宋体') doc.styles['Normal']._element.rPr.rFonts.set(qn('w:eastAsia'), u'宋体')
doc.styles['Normal'].font.size = shared.Pt(6.5) doc.styles['Normal'].font.size = shared.Pt(6.5)
h, w, _ = img.shape
res = sorted_layout_boxes(res, w)
flag = 1 flag = 1
for i, region in enumerate(res): for i, region in enumerate(res):
img_idx = region['img_idx']
if flag == 2 and region['layout'] == 'single': if flag == 2 and region['layout'] == 'single':
section = doc.add_section(WD_SECTION.CONTINUOUS) section = doc.add_section(WD_SECTION.CONTINUOUS)
section._sectPr.xpath('./w:cols')[0].set(qn('w:num'), '1') section._sectPr.xpath('./w:cols')[0].set(qn('w:num'), '1')
...@@ -46,10 +48,10 @@ def convert_info_docx(img, res, save_folder, img_name): ...@@ -46,10 +48,10 @@ def convert_info_docx(img, res, save_folder, img_name):
section._sectPr.xpath('./w:cols')[0].set(qn('w:num'), '2') section._sectPr.xpath('./w:cols')[0].set(qn('w:num'), '2')
flag = 2 flag = 2
if region['type'] == 'Figure': if region['type'].lower() == 'figure':
excel_save_folder = os.path.join(save_folder, img_name) excel_save_folder = os.path.join(save_folder, img_name)
img_path = os.path.join(excel_save_folder, img_path = os.path.join(excel_save_folder,
'{}.jpg'.format(region['bbox'])) '{}_{}.jpg'.format(region['bbox'], img_idx))
paragraph_pic = doc.add_paragraph() paragraph_pic = doc.add_paragraph()
paragraph_pic.alignment = WD_ALIGN_PARAGRAPH.CENTER paragraph_pic.alignment = WD_ALIGN_PARAGRAPH.CENTER
run = paragraph_pic.add_run("") run = paragraph_pic.add_run("")
...@@ -57,40 +59,38 @@ def convert_info_docx(img, res, save_folder, img_name): ...@@ -57,40 +59,38 @@ def convert_info_docx(img, res, save_folder, img_name):
run.add_picture(img_path, width=shared.Inches(5)) run.add_picture(img_path, width=shared.Inches(5))
elif flag == 2: elif flag == 2:
run.add_picture(img_path, width=shared.Inches(2)) run.add_picture(img_path, width=shared.Inches(2))
elif region['type'] == 'Title': elif region['type'].lower() == 'title':
doc.add_heading(region['res'][0]['text']) doc.add_heading(region['res'][0]['text'])
elif region['type'] == 'Text': elif region['type'].lower() == 'table':
paragraph = doc.add_paragraph()
new_parser = HtmlToDocx()
new_parser.table_style = 'TableGrid'
table = new_parser.handle_table(html=region['res']['html'])
new_table = deepcopy(table)
new_table.alignment = WD_TABLE_ALIGNMENT.CENTER
paragraph.add_run().element.addnext(new_table._tbl)
else:
paragraph = doc.add_paragraph() paragraph = doc.add_paragraph()
paragraph_format = paragraph.paragraph_format paragraph_format = paragraph.paragraph_format
for i, line in enumerate(region['res']): for i, line in enumerate(region['res']):
if i == 0: if i == 0:
paragraph_format.first_line_indent = shared.Inches(0.25) paragraph_format.first_line_indent = shared.Inches(0.25)
text_run = paragraph.add_run(line['text'] + ' ') text_run = paragraph.add_run(line['text'] + ' ')
text_run.font.size = shared.Pt(9) text_run.font.size = shared.Pt(10)
elif region['type'] == 'Table':
pypandoc.convert(
source=region['res']['html'],
format='html',
to='docx',
outputfile='tmp.docx')
tmp_doc = Document('tmp.docx')
paragraph = doc.add_paragraph()
table = tmp_doc.tables[0]
new_table = deepcopy(table)
new_table.style = doc.styles['Table Grid']
from docx.enum.table import WD_TABLE_ALIGNMENT
new_table.alignment = WD_TABLE_ALIGNMENT.CENTER
paragraph.add_run().element.addnext(new_table._tbl)
os.remove('tmp.docx')
else:
continue
# save to docx # save to docx
docx_path = os.path.join(save_folder, '{}.docx'.format(img_name)) docx_path = os.path.join(save_folder, '{}.docx'.format(img_name))
doc.save(docx_path) doc.save(docx_path)
logger.info('docx save to {}'.format(docx_path)) logger.info('docx save to {}'.format(docx_path))
# save to pdf
if save_pdf:
pdf_path = os.path.join(save_folder, '{}.pdf'.format(img_name))
from docx2pdf import convert
convert(docx_path, pdf_path)
logger.info('pdf save to {}'.format(pdf_path))
def sorted_layout_boxes(res, w): def sorted_layout_boxes(res, w):
""" """
...@@ -137,7 +137,7 @@ def sorted_layout_boxes(res, w): ...@@ -137,7 +137,7 @@ def sorted_layout_boxes(res, w):
res_left = [] res_left = []
res_right = [] res_right = []
break break
elif _boxes[i]['bbox'][0] < w / 4 and _boxes[i]['bbox'][2] < 3*w / 4: elif _boxes[i]['bbox'][0] < w / 4 and _boxes[i]['bbox'][2] < 3 * w / 4:
_boxes[i]['layout'] = 'double' _boxes[i]['layout'] = 'double'
res_left.append(_boxes[i]) res_left.append(_boxes[i])
i += 1 i += 1
......
ppstructure/recovery/requirements.txt
浏览文件 @ 96a8e0f8
opencv-contrib-python==4.4.0.46
pypandoc pypandoc
python-docx python-docx
docx2pdf
fitz
PyMuPDF
\ No newline at end of file
ppstructure/recovery/table_process.py 0 → 100644
浏览文件 @ 96a8e0f8
# Copyright (c) 2022 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.
"""
This code is refer from:https://github.com/pqzx/html2docx/blob/8f6695a778c68befb302e48ac0ed5201ddbd4524/htmldocx/h2d.py
"""
import re, argparse
import io, os
import urllib.request
from urllib.parse import urlparse
from html.parser import HTMLParser
import docx, docx.table
from docx import Document
from docx.shared import RGBColor, Pt, Inches
from docx.enum.text import WD_COLOR, WD_ALIGN_PARAGRAPH
from docx.oxml import OxmlElement
from docx.oxml.ns import qn
from bs4 import BeautifulSoup
# values in inches
INDENT = 0.25
LIST_INDENT = 0.5
MAX_INDENT = 5.5 # To stop indents going off the page
# Style to use with tables. By default no style is used.
DEFAULT_TABLE_STYLE = None
# Style to use with paragraphs. By default no style is used.
DEFAULT_PARAGRAPH_STYLE = None
def get_filename_from_url(url):
return os.path.basename(urlparse(url).path)
def is_url(url):
"""
Not to be used for actually validating a url, but in our use case we only
care if it's a url or a file path, and they're pretty distinguishable
"""
parts = urlparse(url)
return all([parts.scheme, parts.netloc, parts.path])
def fetch_image(url):
"""
Attempts to fetch an image from a url.
If successful returns a bytes object, else returns None
:return:
"""
try:
with urllib.request.urlopen(url) as response:
# security flaw?
return io.BytesIO(response.read())
except urllib.error.URLError:
return None
def remove_last_occurence(ls, x):
ls.pop(len(ls) - ls[::-1].index(x) - 1)
def remove_whitespace(string, leading=False, trailing=False):
"""Remove white space from a string.
Args:
string(str): The string to remove white space from.
leading(bool, optional): Remove leading new lines when True.
trailing(bool, optional): Remove trailing new lines when False.
Returns:
str: The input string with new line characters removed and white space squashed.
Examples:
Single or multiple new line characters are replaced with space.
>>> remove_whitespace("abc\\ndef")
'abc def'
>>> remove_whitespace("abc\\n\\n\\ndef")
'abc def'
New line characters surrounded by white space are replaced with a single space.
>>> remove_whitespace("abc \\n \\n \\n def")
'abc def'
>>> remove_whitespace("abc \\n \\n \\n def")
'abc def'
Leading and trailing new lines are replaced with a single space.
>>> remove_whitespace("\\nabc")
' abc'
>>> remove_whitespace(" \\n abc")
' abc'
>>> remove_whitespace("abc\\n")
'abc '
>>> remove_whitespace("abc \\n ")
'abc '
Use ``leading=True`` to remove leading new line characters, including any surrounding
white space:
>>> remove_whitespace("\\nabc", leading=True)
'abc'
>>> remove_whitespace(" \\n abc", leading=True)
'abc'
Use ``trailing=True`` to remove trailing new line characters, including any surrounding
white space:
>>> remove_whitespace("abc \\n ", trailing=True)
'abc'
"""
# Remove any leading new line characters along with any surrounding white space
if leading:
string = re.sub(r'^\s*\n+\s*', '', string)
# Remove any trailing new line characters along with any surrounding white space
if trailing:
string = re.sub(r'\s*\n+\s*$', '', string)
# Replace new line characters and absorb any surrounding space.
string = re.sub(r'\s*\n\s*', ' ', string)
# TODO need some way to get rid of extra spaces in e.g. text <span> </span> text
return re.sub(r'\s+', ' ', string)
def delete_paragraph(paragraph):
# https://github.com/python-openxml/python-docx/issues/33#issuecomment-77661907
p = paragraph._element
p.getparent().remove(p)
p._p = p._element = None
font_styles = {
'b': 'bold',
'strong': 'bold',
'em': 'italic',
'i': 'italic',
'u': 'underline',
's': 'strike',
'sup': 'superscript',
'sub': 'subscript',
'th': 'bold',
}
font_names = {
'code': 'Courier',
'pre': 'Courier',
}
styles = {
'LIST_BULLET': 'List Bullet',
'LIST_NUMBER': 'List Number',
}
class HtmlToDocx(HTMLParser):
def __init__(self):
super().__init__()
self.options = {
'fix-html': True,
'images': True,
'tables': True,
'styles': True,
}
self.table_row_selectors = [
'table > tr',
'table > thead > tr',
'table > tbody > tr',
'table > tfoot > tr'
]
self.table_style = DEFAULT_TABLE_STYLE
self.paragraph_style = DEFAULT_PARAGRAPH_STYLE
def set_initial_attrs(self, document=None):
self.tags = {
'span': [],
'list': [],
}
if document:
self.doc = document
else:
self.doc = Document()
self.bs = self.options['fix-html'] # whether or not to clean with BeautifulSoup
self.document = self.doc
self.include_tables = True #TODO add this option back in?
self.include_images = self.options['images']
self.include_styles = self.options['styles']
self.paragraph = None
self.skip = False
self.skip_tag = None
self.instances_to_skip = 0
def copy_settings_from(self, other):
"""Copy settings from another instance of HtmlToDocx"""
self.table_style = other.table_style
self.paragraph_style = other.paragraph_style
def get_cell_html(self, soup):
# Returns string of td element with opening and closing <td> tags removed
# Cannot use find_all as it only finds element tags and does not find text which
# is not inside an element
return ' '.join([str(i) for i in soup.contents])
def add_styles_to_paragraph(self, style):
if 'text-align' in style:
align = style['text-align']
if align == 'center':
self.paragraph.paragraph_format.alignment = WD_ALIGN_PARAGRAPH.CENTER
elif align == 'right':
self.paragraph.paragraph_format.alignment = WD_ALIGN_PARAGRAPH.RIGHT
elif align == 'justify':
self.paragraph.paragraph_format.alignment = WD_ALIGN_PARAGRAPH.JUSTIFY
if 'margin-left' in style:
margin = style['margin-left']
units = re.sub(r'[0-9]+', '', margin)
margin = int(float(re.sub(r'[a-z]+', '', margin)))
if units == 'px':
self.paragraph.paragraph_format.left_indent = Inches(min(margin // 10 * INDENT, MAX_INDENT))
# TODO handle non px units
def add_styles_to_run(self, style):
if 'color' in style:
if 'rgb' in style['color']:
color = re.sub(r'[a-z()]+', '', style['color'])
colors = [int(x) for x in color.split(',')]
elif '#' in style['color']:
color = style['color'].lstrip('#')
colors = tuple(int(color[i:i+2], 16) for i in (0, 2, 4))
else:
colors = [0, 0, 0]
# TODO map colors to named colors (and extended colors...)
# For now set color to black to prevent crashing
self.run.font.color.rgb = RGBColor(*colors)
if 'background-color' in style:
if 'rgb' in style['background-color']:
color = color = re.sub(r'[a-z()]+', '', style['background-color'])
colors = [int(x) for x in color.split(',')]
elif '#' in style['background-color']:
color = style['background-color'].lstrip('#')
colors = tuple(int(color[i:i+2], 16) for i in (0, 2, 4))
else:
colors = [0, 0, 0]
# TODO map colors to named colors (and extended colors...)
# For now set color to black to prevent crashing
self.run.font.highlight_color = WD_COLOR.GRAY_25 #TODO: map colors
def apply_paragraph_style(self, style=None):
try:
if style:
self.paragraph.style = style
elif self.paragraph_style:
self.paragraph.style = self.paragraph_style
except KeyError as e:
raise ValueError(f"Unable to apply style {self.paragraph_style}.") from e
def parse_dict_string(self, string, separator=';'):
new_string = string.replace(" ", '').split(separator)
string_dict = dict([x.split(':') for x in new_string if ':' in x])
return string_dict
def handle_li(self):
# check list stack to determine style and depth
list_depth = len(self.tags['list'])
if list_depth:
list_type = self.tags['list'][-1]
else:
list_type = 'ul' # assign unordered if no tag
if list_type == 'ol':
list_style = styles['LIST_NUMBER']
else:
list_style = styles['LIST_BULLET']
self.paragraph = self.doc.add_paragraph(style=list_style)
self.paragraph.paragraph_format.left_indent = Inches(min(list_depth * LIST_INDENT, MAX_INDENT))
self.paragraph.paragraph_format.line_spacing = 1
def add_image_to_cell(self, cell, image):
# python-docx doesn't have method yet for adding images to table cells. For now we use this
paragraph = cell.add_paragraph()
run = paragraph.add_run()
run.add_picture(image)
def handle_img(self, current_attrs):
if not self.include_images:
self.skip = True
self.skip_tag = 'img'
return
src = current_attrs['src']
# fetch image
src_is_url = is_url(src)
if src_is_url:
try:
image = fetch_image(src)
except urllib.error.URLError:
image = None
else:
image = src
# add image to doc
if image:
try:
if isinstance(self.doc, docx.document.Document):
self.doc.add_picture(image)
else:
self.add_image_to_cell(self.doc, image)
except FileNotFoundError:
image = None
if not image:
if src_is_url:
self.doc.add_paragraph("<image: %s>" % src)
else:
# avoid exposing filepaths in document
self.doc.add_paragraph("<image: %s>" % get_filename_from_url(src))
def handle_table(self, html):
"""
To handle nested tables, we will parse tables manually as follows:
Get table soup
Create docx table
Iterate over soup and fill docx table with new instances of this parser
Tell HTMLParser to ignore any tags until the corresponding closing table tag
"""
doc = Document()
table_soup = BeautifulSoup(html, 'html.parser')
rows, cols_len = self.get_table_dimensions(table_soup)
table = doc.add_table(len(rows), cols_len)
table.style = doc.styles['Table Grid']
cell_row = 0
for index, row in enumerate(rows):
cols = self.get_table_columns(row)
cell_col = 0
for col in cols:
colspan = int(col.attrs.get('colspan', 1))
rowspan = int(col.attrs.get('rowspan', 1))
cell_html = self.get_cell_html(col)
if col.name == 'th':
cell_html = "<b>%s</b>" % cell_html
docx_cell = table.cell(cell_row, cell_col)
while docx_cell.text != '': # Skip the merged cell
cell_col += 1
docx_cell = table.cell(cell_row, cell_col)
cell_to_merge = table.cell(cell_row + rowspan - 1, cell_col + colspan - 1)
if docx_cell != cell_to_merge:
docx_cell.merge(cell_to_merge)
child_parser = HtmlToDocx()
child_parser.copy_settings_from(self)
child_parser.add_html_to_cell(cell_html or ' ', docx_cell) # occupy the position
cell_col += colspan
cell_row += 1
# skip all tags until corresponding closing tag
self.instances_to_skip = len(table_soup.find_all('table'))
self.skip_tag = 'table'
self.skip = True
self.table = None
return table
def handle_link(self, href, text):
# Link requires a relationship
is_external = href.startswith('http')
rel_id = self.paragraph.part.relate_to(
href,
docx.opc.constants.RELATIONSHIP_TYPE.HYPERLINK,
is_external=True # don't support anchor links for this library yet
)
# Create the w:hyperlink tag and add needed values
hyperlink = docx.oxml.shared.OxmlElement('w:hyperlink')
hyperlink.set(docx.oxml.shared.qn('r:id'), rel_id)
# Create sub-run
subrun = self.paragraph.add_run()
rPr = docx.oxml.shared.OxmlElement('w:rPr')
# add default color
c = docx.oxml.shared.OxmlElement('w:color')
c.set(docx.oxml.shared.qn('w:val'), "0000EE")
rPr.append(c)
# add underline
u = docx.oxml.shared.OxmlElement('w:u')
u.set(docx.oxml.shared.qn('w:val'), 'single')
rPr.append(u)
subrun._r.append(rPr)
subrun._r.text = text
# Add subrun to hyperlink
hyperlink.append(subrun._r)
# Add hyperlink to run
self.paragraph._p.append(hyperlink)
def handle_starttag(self, tag, attrs):
if self.skip:
return
if tag == 'head':
self.skip = True
self.skip_tag = tag
self.instances_to_skip = 0
return
elif tag == 'body':
return
current_attrs = dict(attrs)
if tag == 'span':
self.tags['span'].append(current_attrs)
return
elif tag == 'ol' or tag == 'ul':
self.tags['list'].append(tag)
return # don't apply styles for now
elif tag == 'br':
self.run.add_break()
return
self.tags[tag] = current_attrs
if tag in ['p', 'pre']:
self.paragraph = self.doc.add_paragraph()
self.apply_paragraph_style()
elif tag == 'li':
self.handle_li()
elif tag == "hr":
# This implementation was taken from:
# https://github.com/python-openxml/python-docx/issues/105#issuecomment-62806373
self.paragraph = self.doc.add_paragraph()
pPr = self.paragraph._p.get_or_add_pPr()
pBdr = OxmlElement('w:pBdr')
pPr.insert_element_before(pBdr,
'w:shd', 'w:tabs', 'w:suppressAutoHyphens', 'w:kinsoku', 'w:wordWrap',
'w:overflowPunct', 'w:topLinePunct', 'w:autoSpaceDE', 'w:autoSpaceDN',
'w:bidi', 'w:adjustRightInd', 'w:snapToGrid', 'w:spacing', 'w:ind',
'w:contextualSpacing', 'w:mirrorIndents', 'w:suppressOverlap', 'w:jc',
'w:textDirection', 'w:textAlignment', 'w:textboxTightWrap',
'w:outlineLvl', 'w:divId', 'w:cnfStyle', 'w:rPr', 'w:sectPr',
'w:pPrChange'
)
bottom = OxmlElement('w:bottom')
bottom.set(qn('w:val'), 'single')
bottom.set(qn('w:sz'), '6')
bottom.set(qn('w:space'), '1')
bottom.set(qn('w:color'), 'auto')
pBdr.append(bottom)
elif re.match('h[1-9]', tag):
if isinstance(self.doc, docx.document.Document):
h_size = int(tag[1])
self.paragraph = self.doc.add_heading(level=min(h_size, 9))
else:
self.paragraph = self.doc.add_paragraph()
elif tag == 'img':
self.handle_img(current_attrs)
return
elif tag == 'table':
self.handle_table()
return
# set new run reference point in case of leading line breaks
if tag in ['p', 'li', 'pre']:
self.run = self.paragraph.add_run()
# add style
if not self.include_styles:
return
if 'style' in current_attrs and self.paragraph:
style = self.parse_dict_string(current_attrs['style'])
self.add_styles_to_paragraph(style)
def handle_endtag(self, tag):
if self.skip:
if not tag == self.skip_tag:
return
if self.instances_to_skip > 0:
self.instances_to_skip -= 1
return
self.skip = False
self.skip_tag = None
self.paragraph = None
if tag == 'span':
if self.tags['span']:
self.tags['span'].pop()
return
elif tag == 'ol' or tag == 'ul':
remove_last_occurence(self.tags['list'], tag)
return
elif tag == 'table':
self.table_no += 1
self.table = None
self.doc = self.document
self.paragraph = None
if tag in self.tags:
self.tags.pop(tag)
# maybe set relevant reference to None?
def handle_data(self, data):
if self.skip:
return
# Only remove white space if we're not in a pre block.
if 'pre' not in self.tags:
# remove leading and trailing whitespace in all instances
data = remove_whitespace(data, True, True)
if not self.paragraph:
self.paragraph = self.doc.add_paragraph()
self.apply_paragraph_style()
# There can only be one nested link in a valid html document
# You cannot have interactive content in an A tag, this includes links
# https://html.spec.whatwg.org/#interactive-content
link = self.tags.get('a')
if link:
self.handle_link(link['href'], data)
else:
# If there's a link, dont put the data directly in the run
self.run = self.paragraph.add_run(data)
spans = self.tags['span']
for span in spans:
if 'style' in span:
style = self.parse_dict_string(span['style'])
self.add_styles_to_run(style)
# add font style and name
for tag in self.tags:
if tag in font_styles:
font_style = font_styles[tag]
setattr(self.run.font, font_style, True)
if tag in font_names:
font_name = font_names[tag]
self.run.font.name = font_name
def ignore_nested_tables(self, tables_soup):
"""
Returns array containing only the highest level tables
Operates on the assumption that bs4 returns child elements immediately after
the parent element in `find_all`. If this changes in the future, this method will need to be updated
:return:
"""
new_tables = []
nest = 0
for table in tables_soup:
if nest:
nest -= 1
continue
new_tables.append(table)
nest = len(table.find_all('table'))
return new_tables
def get_table_rows(self, table_soup):
# If there's a header, body, footer or direct child tr tags, add row dimensions from there
return table_soup.select(', '.join(self.table_row_selectors), recursive=False)
def get_table_columns(self, row):
# Get all columns for the specified row tag.
return row.find_all(['th', 'td'], recursive=False) if row else []
def get_table_dimensions(self, table_soup):
# Get rows for the table
rows = self.get_table_rows(table_soup)
# Table is either empty or has non-direct children between table and tr tags
# Thus the row dimensions and column dimensions are assumed to be 0
cols = self.get_table_columns(rows[0]) if rows else []
# Add colspan calculation column number
col_count = 0
for col in cols:
colspan = col.attrs.get('colspan', 1)
col_count += int(colspan)
# return len(rows), col_count
return rows, col_count
def get_tables(self):
if not hasattr(self, 'soup'):
self.include_tables = False
return
# find other way to do it, or require this dependency?
self.tables = self.ignore_nested_tables(self.soup.find_all('table'))
self.table_no = 0
def run_process(self, html):
if self.bs and BeautifulSoup:
self.soup = BeautifulSoup(html, 'html.parser')
html = str(self.soup)
if self.include_tables:
self.get_tables()
self.feed(html)
def add_html_to_document(self, html, document):
if not isinstance(html, str):
raise ValueError('First argument needs to be a %s' % str)
elif not isinstance(document, docx.document.Document) and not isinstance(document, docx.table._Cell):
raise ValueError('Second argument needs to be a %s' % docx.document.Document)
self.set_initial_attrs(document)
self.run_process(html)
def add_html_to_cell(self, html, cell):
self.set_initial_attrs(cell)
self.run_process(html)
def parse_html_file(self, filename_html, filename_docx=None):
with open(filename_html, 'r') as infile:
html = infile.read()
self.set_initial_attrs()
self.run_process(html)
if not filename_docx:
path, filename = os.path.split(filename_html)
filename_docx = '%s/new_docx_file_%s' % (path, filename)
self.doc.save('%s.docx' % filename_docx)
def parse_html_string(self, html):
self.set_initial_attrs()
self.run_process(html)
return self.doc
\ No newline at end of file
ppstructure/table/README.md
浏览文件 @ 96a8e0f8
...@@ -33,8 +33,8 @@ We evaluated the algorithm on the PubTabNet<sup>[1]</sup> eval dataset, and the ...@@ -33,8 +33,8 @@ We evaluated the algorithm on the PubTabNet<sup>[1]</sup> eval dataset, and the
|Method|Acc|[TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src)|Speed| |Method|Acc|[TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src)|Speed|
| --- | --- | --- | ---| | --- | --- | --- | ---|
| EDD<sup>[2]</sup> |x| 88.3 |x| | EDD<sup>[2]</sup> |x| 88.3 |x|
| TableRec-RARE(ours) |73.8%| 95.3% |1550ms| | TableRec-RARE(ours) | 71.73%| 93.88% |779ms|
| SLANet(ours) | 76.2%| 95.85% |766ms| | SLANet(ours) | 76.31%| 95.89%|766ms|
The performance indicators are explained as follows: The performance indicators are explained as follows:
- Acc: The accuracy of the table structure in each image, a wrong token is considered an error. - Acc: The accuracy of the table structure in each image, a wrong token is considered an error.
...@@ -59,16 +59,16 @@ cd PaddleOCR/ppstructure ...@@ -59,16 +59,16 @@ cd PaddleOCR/ppstructure
# download model # download model
mkdir inference && cd inference mkdir inference && cd inference
# Download the PP-OCRv3 text detection model and unzip it # Download the PP-OCRv3 text detection model and unzip it
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar && tar xf ch_PP-OCRv3_det_slim_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar && tar xf ch_PP-OCRv3_det_infer.tar
# Download the PP-OCRv3 text recognition model and unzip it # Download the PP-OCRv3 text recognition model and unzip it
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_infer.tar && tar xf ch_PP-OCRv3_rec_slim_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar && tar xf ch_PP-OCRv3_rec_infer.tar
# Download the PP-Structurev2 form recognition model and unzip it # Download the PP-Structurev2 form recognition model and unzip it
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf ch_ppstructure_mobile_v2.0_SLANet_infer.tar wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf ch_ppstructure_mobile_v2.0_SLANet_infer.tar
cd .. cd ..
# run # run
python3.7 table/predict_table.py \ python3.7 table/predict_table.py \
--det_model_dir=inference/ch_PP-OCRv3_det_slim_infer \ --det_model_dir=inference/ch_PP-OCRv3_det_infer \
--rec_model_dir=inference/ch_PP-OCRv3_rec_slim_infer \ --rec_model_dir=inference/ch_PP-OCRv3_rec_infer \
--table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \ --table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \
--rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \ --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \ --table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \
......
ppstructure/table/README_ch.md
浏览文件 @ 96a8e0f8
...@@ -39,8 +39,8 @@ ...@@ -39,8 +39,8 @@
|算法|Acc|[TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src)|Speed| |算法|Acc|[TEDS(Tree-Edit-Distance-based Similarity)](https://github.com/ibm-aur-nlp/PubTabNet/tree/master/src)|Speed|
| --- | --- | --- | ---| | --- | --- | --- | ---|
| EDD<sup>[2]</sup> |x| 88.3% |x| | EDD<sup>[2]</sup> |x| 88.3% |x|
| TableRec-RARE(ours) |73.8%| 95.3% |1550ms| | TableRec-RARE(ours) | 71.73%| 93.88% |779ms|
| SLANet(ours) | 76.2%| 95.85% |766ms| | SLANet(ours) |76.31%| 95.89%|766ms|
性能指标解释如下： 性能指标解释如下：
- Acc: 模型对每张图像里表格结构的识别准确率，错一个token就算错误。 - Acc: 模型对每张图像里表格结构的识别准确率，错一个token就算错误。
...@@ -64,16 +64,16 @@ cd PaddleOCR/ppstructure ...@@ -64,16 +64,16 @@ cd PaddleOCR/ppstructure
# 下载模型 # 下载模型
mkdir inference && cd inference mkdir inference && cd inference
# 下载PP-OCRv3文本检测模型并解压 # 下载PP-OCRv3文本检测模型并解压
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar && tar xf ch_PP-OCRv3_det_slim_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar && tar xf ch_PP-OCRv3_det_infer.tar
# 下载PP-OCRv3文本识别模型并解压 # 下载PP-OCRv3文本识别模型并解压
wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_infer.tar && tar xf ch_PP-OCRv3_rec_slim_infer.tar wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar && tar xf ch_PP-OCRv3_rec_infer.tar
# 下载PP-Structurev2表格识别模型并解压 # 下载PP-Structurev2表格识别模型并解压
wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf ch_ppstructure_mobile_v2.0_SLANet_infer.tar wget https://paddleocr.bj.bcebos.com/ppstructure/models/slanet/ch_ppstructure_mobile_v2.0_SLANet_infer.tar && tar xf ch_ppstructure_mobile_v2.0_SLANet_infer.tar
cd .. cd ..
# 执行表格识别 # 执行表格识别
python table/predict_table.py \ python table/predict_table.py \
--det_model_dir=inference/ch_PP-OCRv3_det_slim_infer \ --det_model_dir=inference/ch_PP-OCRv3_det_infer \
--rec_model_dir=inference/ch_PP-OCRv3_rec_slim_infer \ --rec_model_dir=inference/ch_PP-OCRv3_rec_infer \
--table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \ --table_model_dir=inference/ch_ppstructure_mobile_v2.0_SLANet_infer \
--rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \ --rec_char_dict_path=../ppocr/utils/ppocr_keys_v1.txt \
--table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \ --table_char_dict_path=../ppocr/utils/dict/table_structure_dict_ch.txt \
......
ppstructure/table/predict_structure.py
浏览文件 @ 96a8e0f8
...@@ -29,7 +29,7 @@ import tools.infer.utility as utility ...@@ -29,7 +29,7 @@ import tools.infer.utility as utility
from ppocr.data import create_operators, transform from ppocr.data import create_operators, transform
from ppocr.postprocess import build_post_process from ppocr.postprocess import build_post_process
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
from ppocr.utils.visual import draw_rectangle from ppocr.utils.visual import draw_rectangle
from ppstructure.utility import parse_args from ppstructure.utility import parse_args
...@@ -133,7 +133,7 @@ def main(args): ...@@ -133,7 +133,7 @@ def main(args):
os.path.join(args.output, 'infer.txt'), mode='w', os.path.join(args.output, 'infer.txt'), mode='w',
encoding='utf-8') as f_w: encoding='utf-8') as f_w:
for image_file in image_file_list: for image_file in image_file_list:
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
img = cv2.imread(image_file) img = cv2.imread(image_file)
if img is None: if img is None:
......
ppstructure/table/predict_table.py
浏览文件 @ 96a8e0f8
...@@ -31,7 +31,7 @@ import tools.infer.predict_rec as predict_rec ...@@ -31,7 +31,7 @@ import tools.infer.predict_rec as predict_rec
import tools.infer.predict_det as predict_det import tools.infer.predict_det as predict_det
import tools.infer.utility as utility import tools.infer.utility as utility
from tools.infer.predict_system import sorted_boxes from tools.infer.predict_system import sorted_boxes
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppstructure.table.matcher import TableMatch from ppstructure.table.matcher import TableMatch
from ppstructure.table.table_master_match import TableMasterMatcher from ppstructure.table.table_master_match import TableMasterMatcher
...@@ -194,7 +194,7 @@ def main(args): ...@@ -194,7 +194,7 @@ def main(args):
for i, image_file in enumerate(image_file_list): for i, image_file in enumerate(image_file_list):
logger.info("[{}/{}] {}".format(i, img_num, image_file)) logger.info("[{}/{}] {}".format(i, img_num, image_file))
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
excel_path = os.path.join( excel_path = os.path.join(
args.output, os.path.basename(image_file).split('.')[0] + '.xlsx') args.output, os.path.basename(image_file).split('.')[0] + '.xlsx')
if not flag: if not flag:
......
ppstructure/table/table_metric/table_metric.py
浏览文件 @ 96a8e0f8
...@@ -9,7 +9,7 @@ ...@@ -9,7 +9,7 @@
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# Apache 2.0 License for more details. # Apache 2.0 License for more details.
import distance from rapidfuzz.distance import Levenshtein
from apted import APTED, Config from apted import APTED, Config
from apted.helpers import Tree from apted.helpers import Tree
from lxml import etree, html from lxml import etree, html
...@@ -39,17 +39,6 @@ class TableTree(Tree): ...@@ -39,17 +39,6 @@ class TableTree(Tree):
class CustomConfig(Config): class CustomConfig(Config):
@staticmethod
def maximum(*sequences):
"""Get maximum possible value
"""
return max(map(len, sequences))
def normalized_distance(self, *sequences):
"""Get distance from 0 to 1
"""
return float(distance.levenshtein(*sequences)) / self.maximum(*sequences)
def rename(self, node1, node2): def rename(self, node1, node2):
"""Compares attributes of trees""" """Compares attributes of trees"""
#print(node1.tag) #print(node1.tag)
...@@ -58,23 +47,12 @@ class CustomConfig(Config): ...@@ -58,23 +47,12 @@ class CustomConfig(Config):
if node1.tag == 'td': if node1.tag == 'td':
if node1.content or node2.content: if node1.content or node2.content:
#print(node1.content, ) #print(node1.content, )
return self.normalized_distance(node1.content, node2.content) return Levenshtein.normalized_distance(node1.content, node2.content)
return 0. return 0.
class CustomConfig_del_short(Config): class CustomConfig_del_short(Config):
@staticmethod
def maximum(*sequences):
"""Get maximum possible value
"""
return max(map(len, sequences))
def normalized_distance(self, *sequences):
"""Get distance from 0 to 1
"""
return float(distance.levenshtein(*sequences)) / self.maximum(*sequences)
def rename(self, node1, node2): def rename(self, node1, node2):
"""Compares attributes of trees""" """Compares attributes of trees"""
if (node1.tag != node2.tag) or (node1.colspan != node2.colspan) or (node1.rowspan != node2.rowspan): if (node1.tag != node2.tag) or (node1.colspan != node2.colspan) or (node1.rowspan != node2.rowspan):
...@@ -90,21 +68,10 @@ class CustomConfig_del_short(Config): ...@@ -90,21 +68,10 @@ class CustomConfig_del_short(Config):
node1_content = ['####'] node1_content = ['####']
if len(node2_content) < 3: if len(node2_content) < 3:
node2_content = ['####'] node2_content = ['####']
return self.normalized_distance(node1_content, node2_content) return Levenshtein.normalized_distance(node1_content, node2_content)
return 0. return 0.
class CustomConfig_del_block(Config): class CustomConfig_del_block(Config):
@staticmethod
def maximum(*sequences):
"""Get maximum possible value
"""
return max(map(len, sequences))
def normalized_distance(self, *sequences):
"""Get distance from 0 to 1
"""
return float(distance.levenshtein(*sequences)) / self.maximum(*sequences)
def rename(self, node1, node2): def rename(self, node1, node2):
"""Compares attributes of trees""" """Compares attributes of trees"""
if (node1.tag != node2.tag) or (node1.colspan != node2.colspan) or (node1.rowspan != node2.rowspan): if (node1.tag != node2.tag) or (node1.colspan != node2.colspan) or (node1.rowspan != node2.rowspan):
...@@ -120,7 +87,7 @@ class CustomConfig_del_block(Config): ...@@ -120,7 +87,7 @@ class CustomConfig_del_block(Config):
while ' ' in node2_content: while ' ' in node2_content:
print(node2_content.index(' ')) print(node2_content.index(' '))
node2_content.pop(node2_content.index(' ')) node2_content.pop(node2_content.index(' '))
return self.normalized_distance(node1_content, node2_content) return Levenshtein.normalized_distance(node1_content, node2_content)
return 0. return 0.
class TEDS(object): class TEDS(object):
......
ppstructure/utility.py
浏览文件 @ 96a8e0f8
...@@ -38,7 +38,7 @@ def init_args(): ...@@ -38,7 +38,7 @@ def init_args():
parser.add_argument( parser.add_argument(
"--layout_dict_path", "--layout_dict_path",
type=str, type=str,
default="../ppocr/utils/dict/layout_dict/layout_pubalynet_dict.txt") default="../ppocr/utils/dict/layout_dict/layout_publaynet_dict.txt")
parser.add_argument( parser.add_argument(
"--layout_score_threshold", "--layout_score_threshold",
type=float, type=float,
...@@ -49,8 +49,8 @@ def init_args(): ...@@ -49,8 +49,8 @@ def init_args():
type=float, type=float,
default=0.5, default=0.5,
help="Threshold of nms.") help="Threshold of nms.")
# params for vqa # params for kie
parser.add_argument("--vqa_algorithm", type=str, default='LayoutXLM') parser.add_argument("--kie_algorithm", type=str, default='LayoutXLM')
parser.add_argument("--ser_model_dir", type=str) parser.add_argument("--ser_model_dir", type=str)
parser.add_argument( parser.add_argument(
"--ser_dict_path", "--ser_dict_path",
...@@ -63,7 +63,7 @@ def init_args(): ...@@ -63,7 +63,7 @@ def init_args():
"--mode", "--mode",
type=str, type=str,
default='structure', default='structure',
help='structure and vqa is supported') help='structure and kie is supported')
parser.add_argument( parser.add_argument(
"--image_orientation", "--image_orientation",
type=bool, type=bool,
...@@ -84,11 +84,18 @@ def init_args(): ...@@ -84,11 +84,18 @@ def init_args():
type=str2bool, type=str2bool,
default=True, default=True,
help='In the forward, whether the non-table area is recognition by ocr') help='In the forward, whether the non-table area is recognition by ocr')
# param for recovery
parser.add_argument( parser.add_argument(
"--recovery", "--recovery",
type=bool, type=str2bool,
default=False, default=False,
help='Whether to enable layout of recovery') help='Whether to enable layout of recovery')
parser.add_argument(
"--save_pdf",
type=str2bool,
default=False,
help='Whether to save pdf file')
return parser return parser
......
ppstructure/vqa/README.md 已删除 100644 → 0
浏览文件 @ 29268eaa
English | [简体中文](README_ch.md)
- [1 Introduction](#1-introduction)
- [2. Performance](#2-performance)
- [3. Effect demo](#3-effect-demo)
- [3.1 SER](#31-ser)
- [3.2 RE](#32-re)
- [4. Install](#4-install)
- [4.1 Install dependencies](#41-install-dependencies)
- [5.3 RE](#53-re)
- [6. Reference Links](#6-reference-links)
- [License](#license)
# Document Visual Question Answering
## 1 Introduction
VQA refers to visual question answering, which mainly asks and answers image content. DOC-VQA is one of the VQA tasks. DOC-VQA mainly asks questions about the text content of text images.
The DOC-VQA algorithm in PP-Structure is developed based on the PaddleNLP natural language processing algorithm library.
The main features are as follows:
- Integrate [LayoutXLM](https://arxiv.org/pdf/2104.08836.pdf) model and PP-OCR prediction engine.
- Supports Semantic Entity Recognition (SER) and Relation Extraction (RE) tasks based on multimodal methods. Based on the SER task, the text recognition and classification in the image can be completed; based on the RE task, the relationship extraction of the text content in the image can be completed, such as judging the problem pair (pair).
- Supports custom training for SER tasks and RE tasks.
- Supports end-to-end system prediction and evaluation of OCR+SER.
- Supports end-to-end system prediction of OCR+SER+RE.
This project is an open source implementation of [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/pdf/2104.08836.pdf) on Paddle 2.2,
Included fine-tuning code on [XFUND dataset](https://github.com/doc-analysis/XFUND).
## 2. Performance
We evaluate the algorithm on the Chinese dataset of [XFUND](https://github.com/doc-analysis/XFUND), and the performance is as follows
| Model | Task | hmean | Model download address |
|:---:|:---:|:---:| :---:|
| LayoutXLM | SER | 0.9038 | [link](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
| LayoutXLM | RE | 0.7483 | [link](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
| LayoutLMv2 | SER | 0.8544 | [link](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar)
| LayoutLMv2 | RE | 0.6777 | [link](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar) |
| LayoutLM | SER | 0.7731 | [link](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar) |
## 3. Effect demo
**Note:** The test images are from the XFUND dataset.
<a name="31"></a>
### 3.1 SER
![](../docs/vqa/result_ser/zh_val_0_ser.jpg) | ![](../docs/vqa/result_ser/zh_val_42_ser.jpg)
---|---
Boxes with different colors in the figure represent different categories. For the XFUND dataset, there are 3 categories: `QUESTION`, `ANSWER`, `HEADER`
* Dark purple: HEADER
* Light purple: QUESTION
* Army Green: ANSWER
The corresponding categories and OCR recognition results are also marked on the upper left of the OCR detection frame.
<a name="32"></a>
### 3.2 RE
![](../docs/vqa/result_re/zh_val_21_re.jpg) | ![](../docs/vqa/result_re/zh_val_40_re.jpg)
---|---
The red box in the figure represents the question, the blue box represents the answer, and the question and the answer are connected by a green line. The corresponding categories and OCR recognition results are also marked on the upper left of the OCR detection frame.
## 4. Install
### 4.1 Install dependencies
- **(1) Install PaddlePaddle**
```bash
python3 -m pip install --upgrade pip
# GPU installation
python3 -m pip install "paddlepaddle-gpu>=2.2" -i https://mirror.baidu.com/pypi/simple
# CPU installation
python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simple
````
For more requirements, please refer to the instructions in [Installation Documentation](https://www.paddlepaddle.org.cn/install/quick).
### 4.2 Install PaddleOCR
- **(1) pip install PaddleOCR whl package quickly (prediction only)**
```bash
python3 -m pip install paddleocr
````
- **(2) Download VQA source code (prediction + training)**
```bash
[Recommended] git clone https://github.com/PaddlePaddle/PaddleOCR
# If the pull cannot be successful due to network problems, you can also choose to use the hosting on the code cloud:
git clone https://gitee.com/paddlepaddle/PaddleOCR
# Note: Code cloud hosting code may not be able to synchronize the update of this github project in real time, there is a delay of 3 to 5 days, please use the recommended method first.
````
- **(3) Install VQA's `requirements`**
```bash
python3 -m pip install -r ppstructure/vqa/requirements.txt
````
## 5. Usage
### 5.1 Data and Model Preparation
If you want to experience the prediction process directly, you can download the pre-training model provided by us, skip the training process, and just predict directly.
* Download the processed dataset
The download address of the processed XFUND Chinese dataset: [link](https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar).
Download and unzip the dataset, and place the dataset in the current directory after unzipping.
```shell
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar
````
* Convert the dataset
If you need to train other XFUND datasets, you can use the following commands to convert the datasets
```bash
python3 ppstructure/vqa/tools/trans_xfun_data.py --ori_gt_path=path/to/json_path --output_path=path/to/save_path
````
* Download the pretrained models
```bash
mkdir pretrain && cd pretrain
#download the SER model
wget https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar && tar -xvf ser_LayoutXLM_xfun_zh.tar
#download the RE model
wget https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar && tar -xvf re_LayoutXLM_xfun_zh.tar
cd ../
````
<a name="52"></a>
### 5.2 SER
Before starting training, you need to modify the following four fields
1. `Train.dataset.data_dir`: point to the directory where the training set images are stored
2. `Train.dataset.label_file_list`: point to the training set label file
3. `Eval.dataset.data_dir`: refers to the directory where the validation set images are stored
4. `Eval.dataset.label_file_list`: point to the validation set label file
* start training
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/ser/layoutxlm.yml
````
Finally, `precision`, `recall`, `hmean` and other indicators will be printed.
In the `./output/ser_layoutxlm/` folder will save the training log, the optimal model and the model for the latest epoch.
* resume training
To resume training, assign the folder path of the previously trained model to the `Architecture.Backbone.checkpoints` field.
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
````
* evaluate
Evaluation requires assigning the folder path of the model to be evaluated to the `Architecture.Backbone.checkpoints` field.
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
````
Finally, `precision`, `recall`, `hmean` and other indicators will be printed
* `OCR + SER` tandem prediction based on training engine
Use the following command to complete the series prediction of `OCR engine + SER`, taking the SER model based on LayoutXLM as an example::
```shell
python3.7 tools/export_model.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/ Global.save_inference_dir=output/ser/infer
````
Finally, the prediction result visualization image and the prediction result text file will be saved in the directory configured by the `config.Global.save_res_path` field. The prediction result text file is named `infer_results.txt`.
* End-to-end evaluation of `OCR + SER` prediction system
First use the `tools/infer_vqa_token_ser.py` script to complete the prediction of the dataset, then use the following command to evaluate.
```shell
export CUDA_VISIBLE_DEVICES=0
python3 tools/eval_with_label_end2end.py --gt_json_path XFUND/zh_val/xfun_normalize_val.json --pred_json_path output_res/infer_results.txt
````
* export model
Use the following command to complete the model export of the SER model, taking the SER model based on LayoutXLM as an example:
```shell
python3.7 tools/export_model.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/ Global.save_inference_dir=output/ser/infer
```
The converted model will be stored in the directory specified by the `Global.save_inference_dir` field.
* `OCR + SER` tandem prediction based on prediction engine
Use the following command to complete the tandem prediction of `OCR + SER` based on the prediction engine, taking the SER model based on LayoutXLM as an example:
```shell
cd ppstructure
CUDA_VISIBLE_DEVICES=0 python3.7 vqa/predict_vqa_token_ser.py --vqa_algorithm=LayoutXLM --ser_model_dir=../output/ser/infer --ser_dict_path=../train_data/XFUND/class_list_xfun.txt --vis_font_path=../doc/fonts/simfang.ttf --image_dir=docs/vqa/input/zh_val_42.jpg --output=output
```
After the prediction is successful, the visualization images and results will be saved in the directory specified by the `output` field
<a name="53"></a>
### 5.3 RE
* start training
Before starting training, you need to modify the following four fields
1. `Train.dataset.data_dir`: point to the directory where the training set images are stored
2. `Train.dataset.label_file_list`: point to the training set label file
3. `Eval.dataset.data_dir`: refers to the directory where the validation set images are stored
4. `Eval.dataset.label_file_list`: point to the validation set label file
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml
````
Finally, `precision`, `recall`, `hmean` and other indicators will be printed.
In the `./output/re_layoutxlm/` folder will save the training log, the optimal model and the model for the latest epoch.
* resume training
To resume training, assign the folder path of the previously trained model to the `Architecture.Backbone.checkpoints` field.
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
````
* evaluate
Evaluation requires assigning the folder path of the model to be evaluated to the `Architecture.Backbone.checkpoints` field.
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
````
Finally, `precision`, `recall`, `hmean` and other indicators will be printed
* Use `OCR engine + SER + RE` tandem prediction
Use the following command to complete the series prediction of `OCR engine + SER + RE`, taking the pretrained SER and RE models as an example:
```shell
export CUDA_VISIBLE_DEVICES=0
python3 tools/infer_vqa_token_ser_re.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=pretrain/re_LayoutXLM_xfun_zh/Global.infer_img=ppstructure/docs/vqa/input/zh_val_21.jpg -c_ser configs/vqa/ser/layoutxlm. yml -o_ser Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/
````
Finally, the prediction result visualization image and the prediction result text file will be saved in the directory configured by the `config.Global.save_res_path` field. The prediction result text file is named `infer_results.txt`.
* export model
cooming soon
* `OCR + SER + RE` tandem prediction based on prediction engine
cooming soon
## 6. Reference Links
- LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding, https://arxiv.org/pdf/2104.08836.pdf
- microsoft/unilm/layoutxlm, https://github.com/microsoft/unilm/tree/master/layoutxlm
- XFUND dataset, https://github.com/doc-analysis/XFUND
## License
The content of this project itself is licensed under the [Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/)
ppstructure/vqa/README_ch.md 已删除 100644 → 0
浏览文件 @ 29268eaa
[English](README.md) | 简体中文
- [1. 简介](#1-简介)
- [2. 性能](#2-性能)
- [3. 效果演示](#3-效果演示)
- [3.1 SER](#31-ser)
- [3.2 RE](#32-re)
- [4. 安装](#4-安装)
- [4.1 安装依赖](#41-安装依赖)
- [4.2 安装PaddleOCR（包含 PP-OCR 和 VQA）](#42-安装paddleocr包含-pp-ocr-和-vqa)
- [5. 使用](#5-使用)
- [5.1 数据和预训练模型准备](#51-数据和预训练模型准备)
- [5.2 SER](#52-ser)
- [5.3 RE](#53-re)
- [6. 参考链接](#6-参考链接)
- [License](#license)
# 文档视觉问答（DOC-VQA）
## 1. 简介
VQA指视觉问答，主要针对图像内容进行提问和回答,DOC-VQA是VQA任务中的一种，DOC-VQA主要针对文本图像的文字内容提出问题。
PP-Structure 里的 DOC-VQA算法基于PaddleNLP自然语言处理算法库进行开发。
主要特性如下：
- 集成[LayoutXLM](https://arxiv.org/pdf/2104.08836.pdf)模型以及PP-OCR预测引擎。
- 支持基于多模态方法的语义实体识别 (Semantic Entity Recognition, SER) 以及关系抽取 (Relation Extraction, RE) 任务。基于 SER 任务，可以完成对图像中的文本识别与分类；基于 RE 任务，可以完成对图象中的文本内容的关系提取，如判断问题对(pair)。
- 支持SER任务和RE任务的自定义训练。
- 支持OCR+SER的端到端系统预测与评估。
- 支持OCR+SER+RE的端到端系统预测。
本项目是 [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/pdf/2104.08836.pdf) 在 Paddle 2.2上的开源实现，
包含了在 [XFUND数据集](https://github.com/doc-analysis/XFUND) 上的微调代码。
## 2. 性能
我们在 [XFUND](https://github.com/doc-analysis/XFUND) 的中文数据集上对算法进行了评估，性能如下
| 模型 | 任务 | hmean | 模型下载地址 |
|:---:|:---:|:---:| :---:|
| LayoutXLM | SER | 0.9038 | [链接](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
| LayoutXLM | RE | 0.7483 | [链接](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
| LayoutLMv2 | SER | 0.8544 | [链接](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLMv2_xfun_zh.tar)
| LayoutLMv2 | RE | 0.6777 | [链接](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutLMv2_xfun_zh.tar) |
| LayoutLM | SER | 0.7731 | [链接](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutLM_xfun_zh.tar) |
## 3. 效果演示
**注意：** 测试图片来源于XFUND数据集。
### 3.1 SER
![](../docs/vqa/result_ser/zh_val_0_ser.jpg) | ![](../docs/vqa/result_ser/zh_val_42_ser.jpg)
---|---
图中不同颜色的框表示不同的类别，对于XFUND数据集，有`QUESTION`, `ANSWER`, `HEADER` 3种类别
* 深紫色：HEADER
* 浅紫色：QUESTION
* 军绿色：ANSWER
在OCR检测框的左上方也标出了对应的类别和OCR识别结果。
### 3.2 RE
![](../docs/vqa/result_re/zh_val_21_re.jpg) | ![](../docs/vqa/result_re/zh_val_40_re.jpg)
---|---
图中红色框表示问题，蓝色框表示答案，问题和答案之间使用绿色线连接。在OCR检测框的左上方也标出了对应的类别和OCR识别结果。
## 4. 安装
### 4.1 安装依赖
- **（1) 安装PaddlePaddle**
```bash
python3 -m pip install --upgrade pip
# GPU安装
python3 -m pip install "paddlepaddle-gpu>=2.2" -i https://mirror.baidu.com/pypi/simple
# CPU安装
python3 -m pip install "paddlepaddle>=2.2" -i https://mirror.baidu.com/pypi/simple
```
更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
### 4.2 安装PaddleOCR（包含 PP-OCR 和 VQA）
- **（1）pip快速安装PaddleOCR whl包（仅预测）**
```bash
python3 -m pip install paddleocr
```
- **（2）下载VQA源码（预测+训练）**
```bash
【推荐】git clone https://github.com/PaddlePaddle/PaddleOCR
# 如果因为网络问题无法pull成功，也可选择使用码云上的托管：
git clone https://gitee.com/paddlepaddle/PaddleOCR
# 注：码云托管代码可能无法实时同步本github项目更新，存在3~5天延时，请优先使用推荐方式。
```
- **（3）安装VQA的`requirements`**
```bash
python3 -m pip install -r ppstructure/vqa/requirements.txt
```
## 5. 使用
### 5.1 数据和预训练模型准备
如果希望直接体验预测过程，可以下载我们提供的预训练模型，跳过训练过程，直接预测即可。
* 下载处理好的数据集
处理好的XFUND中文数据集下载地址：[链接](https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar)。
下载并解压该数据集，解压后将数据集放置在当前目录下。
```shell
wget https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar
```
* 转换数据集
若需进行其他XFUND数据集的训练，可使用下面的命令进行数据集的转换
```bash
python3 ppstructure/vqa/tools/trans_xfun_data.py --ori_gt_path=path/to/json_path --output_path=path/to/save_path
```
* 下载预训练模型
```bash
mkdir pretrain && cd pretrain
#下载SER模型
wget https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar && tar -xvf ser_LayoutXLM_xfun_zh.tar
#下载RE模型
wget https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar && tar -xvf re_LayoutXLM_xfun_zh.tar
cd ../
```
### 5.2 SER
启动训练之前，需要修改下面的四个字段
1. `Train.dataset.data_dir`：指向训练集图片存放目录
2. `Train.dataset.label_file_list`：指向训练集标注文件
3. `Eval.dataset.data_dir`：指指向验证集图片存放目录
4. `Eval.dataset.label_file_list`：指向验证集标注文件
* 启动训练
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/ser/layoutxlm.yml
```
最终会打印出`precision`, `recall`, `hmean`等指标。
在`./output/ser_layoutxlm/`文件夹中会保存训练日志，最优的模型和最新epoch的模型。
* 恢复训练
恢复训练需要将之前训练好的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段。
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
```
* 评估
评估需要将待评估的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段。
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
```
最终会打印出`precision`, `recall`, `hmean`等指标
* 基于训练引擎的`OCR + SER`串联预测
使用如下命令即可完成基于训练引擎的`OCR + SER`的串联预测, 以基于LayoutXLM的SER模型为例:
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/infer_vqa_token_ser.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/ Global.infer_img=doc/vqa/input/zh_val_42.jpg
```
最终会在`config.Global.save_res_path`字段所配置的目录下保存预测结果可视化图像以及预测结果文本文件，预测结果文本文件名为`infer_results.txt`。
* 对`OCR + SER`预测系统进行端到端评估
首先使用 `tools/infer_vqa_token_ser.py` 脚本完成数据集的预测，然后使用下面的命令进行评估。
```shell
export CUDA_VISIBLE_DEVICES=0
python3 tools/eval_with_label_end2end.py --gt_json_path XFUND/zh_val/xfun_normalize_val.json --pred_json_path output_res/infer_results.txt
```
* 模型导出
使用如下命令即可完成SER模型的模型导出, 以基于LayoutXLM的SER模型为例:
```shell
python3.7 tools/export_model.py -c configs/vqa/ser/layoutxlm.yml -o Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/ Global.save_inference_dir=output/ser/infer
```
转换后的模型会存放在`Global.save_inference_dir`字段指定的目录下。
* 基于预测引擎的`OCR + SER`串联预测
使用如下命令即可完成基于预测引擎的`OCR + SER`的串联预测, 以基于LayoutXLM的SER模型为例:
```shell
cd ppstructure
CUDA_VISIBLE_DEVICES=0 python3.7 vqa/predict_vqa_token_ser.py --vqa_algorithm=LayoutXLM --ser_model_dir=../output/ser/infer --ser_dict_path=../train_data/XFUND/class_list_xfun.txt --vis_font_path=../doc/fonts/simfang.ttf --image_dir=docs/vqa/input/zh_val_42.jpg --output=output
```
预测成功后，可视化图片和结果会保存在`output`字段指定的目录下
### 5.3 RE
* 启动训练
启动训练之前，需要修改下面的四个字段
1. `Train.dataset.data_dir`：指向训练集图片存放目录
2. `Train.dataset.label_file_list`：指向训练集标注文件
3. `Eval.dataset.data_dir`：指指向验证集图片存放目录
4. `Eval.dataset.label_file_list`：指向验证集标注文件
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml
```
最终会打印出`precision`, `recall`, `hmean`等指标。
在`./output/re_layoutxlm/`文件夹中会保存训练日志，最优的模型和最新epoch的模型。
* 恢复训练
恢复训练需要将之前训练好的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段。
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
```
* 评估
评估需要将待评估的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段。
```shell
CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=path/to/model_dir
```
最终会打印出`precision`, `recall`, `hmean`等指标
* 基于训练引擎的`OCR + SER + RE`串联预测
使用如下命令即可完成基于训练引擎的`OCR + SER + RE`串联预测, 以基于LayoutXLMSER和RE模型为例：
```shell
export CUDA_VISIBLE_DEVICES=0
python3 tools/infer_vqa_token_ser_re.py -c configs/vqa/re/layoutxlm.yml -o Architecture.Backbone.checkpoints=pretrain/re_LayoutXLM_xfun_zh/ Global.infer_img=ppstructure/docs/vqa/input/zh_val_21.jpg -c_ser configs/vqa/ser/layoutxlm.yml -o_ser Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/
```
最终会在`config.Global.save_res_path`字段所配置的目录下保存预测结果可视化图像以及预测结果文本文件，预测结果文本文件名为`infer_results.txt`。
* 模型导出
cooming soon
* 基于预测引擎的`OCR + SER + RE`串联预测
cooming soon
## 6. 参考链接
- LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding, https://arxiv.org/pdf/2104.08836.pdf
- microsoft/unilm/layoutxlm, https://github.com/microsoft/unilm/tree/master/layoutxlm
- XFUND dataset, https://github.com/doc-analysis/XFUND
## License
The content of this project itself is licensed under the [Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/)
test_tipc/configs/ch_PP-OCRv2_rec/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/ch_PP-OCRv3_rec/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_image_shape="3,48,320" ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_image_shape="3,48,320"
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/ch_ppocr_mobile_v2_0_rec/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/ch_ppocr_server_v2_0_rec/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/en_table_structure/table_mv3.yml
浏览文件 @ 96a8e0f8
...@@ -4,7 +4,7 @@ Global: ...@@ -4,7 +4,7 @@ Global:
log_smooth_window: 20 log_smooth_window: 20
print_batch_step: 5 print_batch_step: 5
save_model_dir: ./output/table_mv3/ save_model_dir: ./output/table_mv3/
save_epoch_step: 3 save_epoch_step: 400
# evaluation is run every 400 iterations after the 0th iteration # evaluation is run every 400 iterations after the 0th iteration
eval_batch_step: [0, 40000] eval_batch_step: [0, 40000]
cal_metric_during_train: True cal_metric_during_train: True
...@@ -17,7 +17,8 @@ Global: ...@@ -17,7 +17,8 @@ Global:
# for data or label process # for data or label process
character_dict_path: ppocr/utils/dict/table_structure_dict.txt character_dict_path: ppocr/utils/dict/table_structure_dict.txt
character_type: en character_type: en
max_text_length: 800 max_text_length: &max_text_length 500
box_format: &box_format 'xyxy' # 'xywh', 'xyxy', 'xyxyxyxy'
infer_mode: False infer_mode: False
Optimizer: Optimizer:
...@@ -37,12 +38,14 @@ Architecture: ...@@ -37,12 +38,14 @@ Architecture:
Backbone: Backbone:
name: MobileNetV3 name: MobileNetV3
scale: 1.0 scale: 1.0
model_name: large model_name: small
disable_se: true
Head: Head:
name: TableAttentionHead name: TableAttentionHead
hidden_size: 256 hidden_size: 256
loc_type: 2 loc_type: 2
max_text_length: 800 max_text_length: *max_text_length
loc_reg_num: &loc_reg_num 4
Loss: Loss:
name: TableAttentionLoss name: TableAttentionLoss
...@@ -70,6 +73,8 @@ Train: ...@@ -70,6 +73,8 @@ Train:
learn_empty_box: False learn_empty_box: False
merge_no_span_structure: False merge_no_span_structure: False
replace_empty_cell_token: False replace_empty_cell_token: False
loc_reg_num: *loc_reg_num
max_text_length: *max_text_length
- TableBoxEncode: - TableBoxEncode:
- ResizeTableImage: - ResizeTableImage:
max_len: 488 max_len: 488
...@@ -102,6 +107,8 @@ Eval: ...@@ -102,6 +107,8 @@ Eval:
learn_empty_box: False learn_empty_box: False
merge_no_span_structure: False merge_no_span_structure: False
replace_empty_cell_token: False replace_empty_cell_token: False
loc_reg_num: *loc_reg_num
max_text_length: *max_text_length
- TableBoxEncode: - TableBoxEncode:
- ResizeTableImage: - ResizeTableImage:
max_len: 488 max_len: 488
......
test_tipc/configs/en_table_structure/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -54,6 +54,6 @@ random_infer_input:[{float32,[3,488,488]}] ...@@ -54,6 +54,6 @@ random_infer_input:[{float32,[3,488,488]}]
===========================train_benchmark_params========================== ===========================train_benchmark_params==========================
batch_size:32 batch_size:32
fp_items:fp32|fp16 fp_items:fp32|fp16
epoch:1 epoch:2
--profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile --profiler_options:batch_range=[10,20];state=GPU;tracer_option=Default;profile_path=model.profile
flags:FLAGS_eager_delete_tensor_gb=0.0;FLAGS_fraction_of_gpu_memory_to_use=0.98;FLAGS_conv_workspace_size_limit=4096 flags:FLAGS_eager_delete_tensor_gb=0.0;FLAGS_fraction_of_gpu_memory_to_use=0.98;FLAGS_conv_workspace_size_limit=4096
test_tipc/configs/layoutxlm_ser/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -9,7 +9,7 @@ Global.save_model_dir:./output/ ...@@ -9,7 +9,7 @@ Global.save_model_dir:./output/
Train.loader.batch_size_per_card:lite_train_lite_infer=4|whole_train_whole_infer=8 Train.loader.batch_size_per_card:lite_train_lite_infer=4|whole_train_whole_infer=8
Architecture.Backbone.checkpoints:null Architecture.Backbone.checkpoints:null
train_model_name:latest train_model_name:latest
train_infer_img_dir:ppstructure/docs/vqa/input/zh_val_42.jpg train_infer_img_dir:ppstructure/docs/kie/input/zh_val_42.jpg
null:null null:null
## ##
trainer:norm_train trainer:norm_train
...@@ -37,7 +37,7 @@ export2:null ...@@ -37,7 +37,7 @@ export2:null
infer_model:null infer_model:null
infer_export:null infer_export:null
infer_quant:False infer_quant:False
inference:ppstructure/vqa/predict_vqa_token_ser.py --vqa_algorithm=LayoutXLM --ser_dict_path=train_data/XFUND/class_list_xfun.txt --output=output inference:ppstructure/kie/predict_kie_token_ser.py --kie_algorithm=LayoutXLM --ser_dict_path=train_data/XFUND/class_list_xfun.txt --output=output
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
...@@ -45,7 +45,7 @@ inference:ppstructure/vqa/predict_vqa_token_ser.py --vqa_algorithm=LayoutXLM - ...@@ -45,7 +45,7 @@ inference:ppstructure/vqa/predict_vqa_token_ser.py --vqa_algorithm=LayoutXLM -
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--ser_model_dir: --ser_model_dir:
--image_dir:./ppstructure/docs/vqa/input/zh_val_42.jpg --image_dir:./ppstructure/docs/kie/input/zh_val_42.jpg
null:null null:null
--benchmark:False --benchmark:False
null:null null:null
......
test_tipc/configs/rec_mtb_nrtr/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/EN_symbo ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/EN_symbo
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_mv3_none_bilstm_ctc_v2_0/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_mv3_none_none_ctc_v2_0/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_mv3_tps_bilstm_att_v2_0/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_mv3_tps_bilstm_ctc_v2_0/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_r31_robustscanner/train_infer_python.txt
浏览文件 @ 96a8e0f8
===========================train_params=========================== ===========================train_params===========================
model_name:rec_r31_robustscanner model_name:rec_r31_robustscanner
python:python python:python3.7
gpu_list:0|0,1 gpu_list:0|0,1
Global.use_gpu:True|True Global.use_gpu:True|True
Global.auto_cast:null Global.auto_cast:null
...@@ -39,11 +39,11 @@ infer_export:tools/export_model.py -c test_tipc/configs/rec_r31_robustscanner/re ...@@ -39,11 +39,11 @@ infer_export:tools/export_model.py -c test_tipc/configs/rec_r31_robustscanner/re
infer_quant:False infer_quant:False
inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/dict90.txt --rec_image_shape="3,48,48,160" --use_space_char=False --rec_algorithm="RobustScanner" inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/dict90.txt --rec_image_shape="3,48,48,160" --use_space_char=False --rec_algorithm="RobustScanner"
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:True|False --enable_mkldnn:False
--cpu_threads:1|6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False|False --use_tensorrt:False
--precision:fp32|int8 --precision:fp32
--rec_model_dir: --rec_model_dir:
--image_dir:./inference/rec_inference --image_dir:./inference/rec_inference
--save_log_path:./test/output/ --save_log_path:./test/output/
......
test_tipc/configs/rec_r31_sar/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/dict90.t ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/dict90.t
--use_gpu:True --use_gpu:True
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_r32_gaspin_bilstm_att/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/dict/spi ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/dict/spi
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_r34_vd_none_bilstm_ctc_v2_0/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_r34_vd_none_none_ctc_v2_0/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_r34_vd_tps_bilstm_att_v2_0/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_r34_vd_tps_bilstm_ctc_v2_0/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_r45_abinet/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_r45_visionlan/rec_r45_visionlan.yml 0 → 100644
浏览文件 @ 96a8e0f8
Global:
use_gpu: true
epoch_num: 8
log_smooth_window: 200
print_batch_step: 200
save_model_dir: ./output/rec/r45_visionlan
save_epoch_step: 1
# evaluation is run every 2000 iterations
eval_batch_step: [0, 2000]
cal_metric_during_train: True
pretrained_model:
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img: doc/imgs_words/en/word_2.png
# for data or label process
character_dict_path:
max_text_length: &max_text_length 25
training_step: &training_step LA
infer_mode: False
use_space_char: False
save_res_path: ./output/rec/predicts_visionlan.txt
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
clip_norm: 20.0
group_lr: true
training_step: *training_step
lr:
name: Piecewise
decay_epochs: [6]
values: [0.0001, 0.00001]
regularizer:
name: 'L2'
factor: 0
Architecture:
model_type: rec
algorithm: VisionLAN
Transform:
Backbone:
name: ResNet45
strides: [2, 2, 2, 1, 1]
Head:
name: VLHead
n_layers: 3
n_position: 256
n_dim: 512
max_text_length: *max_text_length
training_step: *training_step
Loss:
name: VLLoss
mode: *training_step
weight_res: 0.5
weight_mas: 0.5
PostProcess:
name: VLLabelDecode
Metric:
name: RecMetric
is_filter: true
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/ic15_data/
label_file_list: ["./train_data/ic15_data/rec_gt_train.txt"]
transforms:
- DecodeImage: # load image
img_mode: RGB
channel_first: False
- ABINetRecAug:
- VLLabelEncode: # Class handling label
- VLRecResizeImg:
image_shape: [3, 64, 256]
- KeepKeys:
keep_keys: ['image', 'label', 'label_res', 'label_sub', 'label_id', 'length'] # dataloader will return list in this order
loader:
shuffle: True
batch_size_per_card: 220
drop_last: True
num_workers: 4
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data/ic15_data
label_file_list: ["./train_data/ic15_data/rec_gt_test.txt"]
transforms:
- DecodeImage: # load image
img_mode: RGB
channel_first: False
- VLLabelEncode: # Class handling label
- VLRecResizeImg:
image_shape: [3, 64, 256]
- KeepKeys:
keep_keys: ['image', 'label', 'label_res', 'label_sub', 'label_id', 'length'] # dataloader will return list in this order
loader:
shuffle: False
drop_last: False
batch_size_per_card: 64
num_workers: 4
test_tipc/configs/rec_r45_visionlan/train_infer_python.txt 0 → 100644
浏览文件 @ 96a8e0f8
===========================train_params===========================
model_name:rec_r45_visionlan
python:python3.7
gpu_list:0|0,1
Global.use_gpu:True|True
Global.auto_cast:null
Global.epoch_num:lite_train_lite_infer=2|whole_train_whole_infer=300
Global.save_model_dir:./output/
Train.loader.batch_size_per_card:lite_train_lite_infer=32|whole_train_whole_infer=64
Global.pretrained_model:null
train_model_name:latest
train_infer_img_dir:./inference/rec_inference
null:null
##
trainer:norm_train
norm_train:tools/train.py -c test_tipc/configs/rec_r45_visionlan/rec_r45_visionlan.yml -o
pact_train:null
fpgm_train:null
distill_train:null
null:null
null:null
##
===========================eval_params===========================
eval:tools/eval.py -c test_tipc/configs/rec_r45_visionlan/rec_r45_visionlan.yml -o
null:null
##
===========================infer_params===========================
Global.save_inference_dir:./output/
Global.checkpoints:
norm_export:tools/export_model.py -c test_tipc/configs/rec_r45_visionlan/rec_r45_visionlan.yml -o
quant_export:null
fpgm_export:null
distill_export:null
export1:null
export2:null
##
train_model:./inference/rec_r45_visionlan_train/best_accuracy
infer_export:tools/export_model.py -c test_tipc/configs/rec_r45_visionlan/rec_r45_visionlan.yml -o
infer_quant:False
inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dict.txt --rec_image_shape="3,64,256" --rec_algorithm="VisionLAN" --use_space_char=False
--use_gpu:True|False
--enable_mkldnn:False
--cpu_threads:6
--rec_batch_num:1
--use_tensorrt:False
--precision:fp32
--rec_model_dir:
--image_dir:./inference/rec_inference
--save_log_path:./test/output/
--benchmark:True
null:null
===========================infer_benchmark_params==========================
random_infer_input:[{float32,[3,64,256]}]
test_tipc/configs/rec_r50_fpn_vd_none_srn/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_svtrnet/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/ic15_dic
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/rec_vitstr_none_ce/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/EN_symbo ...@@ -41,7 +41,7 @@ inference:tools/infer/predict_rec.py --rec_char_dict_path=./ppocr/utils/EN_symbo
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
--rec_batch_num:1|6 --rec_batch_num:1
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--rec_model_dir: --rec_model_dir:
......
test_tipc/configs/vi_layoutxlm_ser/train_infer_python.txt
浏览文件 @ 96a8e0f8
...@@ -9,7 +9,7 @@ Global.save_model_dir:./output/ ...@@ -9,7 +9,7 @@ Global.save_model_dir:./output/
Train.loader.batch_size_per_card:lite_train_lite_infer=4|whole_train_whole_infer=8 Train.loader.batch_size_per_card:lite_train_lite_infer=4|whole_train_whole_infer=8
Architecture.Backbone.checkpoints:null Architecture.Backbone.checkpoints:null
train_model_name:latest train_model_name:latest
train_infer_img_dir:ppstructure/docs/vqa/input/zh_val_42.jpg train_infer_img_dir:ppstructure/docs/kie/input/zh_val_42.jpg
null:null null:null
## ##
trainer:norm_train trainer:norm_train
...@@ -37,7 +37,7 @@ export2:null ...@@ -37,7 +37,7 @@ export2:null
infer_model:null infer_model:null
infer_export:null infer_export:null
infer_quant:False infer_quant:False
inference:ppstructure/vqa/predict_vqa_token_ser.py --vqa_algorithm=LayoutXLM --ser_dict_path=train_data/XFUND/class_list_xfun.txt --output=output --ocr_order_method=tb-yx inference:ppstructure/kie/predict_kie_token_ser.py --kie_algorithm=LayoutXLM --ser_dict_path=train_data/XFUND/class_list_xfun.txt --output=output --ocr_order_method=tb-yx
--use_gpu:True|False --use_gpu:True|False
--enable_mkldnn:False --enable_mkldnn:False
--cpu_threads:6 --cpu_threads:6
...@@ -45,7 +45,7 @@ inference:ppstructure/vqa/predict_vqa_token_ser.py --vqa_algorithm=LayoutXLM - ...@@ -45,7 +45,7 @@ inference:ppstructure/vqa/predict_vqa_token_ser.py --vqa_algorithm=LayoutXLM -
--use_tensorrt:False --use_tensorrt:False
--precision:fp32 --precision:fp32
--ser_model_dir: --ser_model_dir:
--image_dir:./ppstructure/docs/vqa/input/zh_val_42.jpg --image_dir:./ppstructure/docs/kie/input/zh_val_42.jpg
null:null null:null
--benchmark:False --benchmark:False
null:null null:null
......
test_tipc/prepare.sh
浏览文件 @ 96a8e0f8
...@@ -107,8 +107,7 @@ if [ ${MODE} = "benchmark_train" ];then ...@@ -107,8 +107,7 @@ if [ ${MODE} = "benchmark_train" ];then
cd ../ cd ../
fi fi
if [ ${model_name} == "layoutxlm_ser" ] || [ ${model_name} == "vi_layoutxlm_ser" ]; then if [ ${model_name} == "layoutxlm_ser" ] || [ ${model_name} == "vi_layoutxlm_ser" ]; then
pip install -r ppstructure/vqa/requirements.txt pip install -r ppstructure/kie/requirements.txt
pip install paddlenlp\>=2.3.5 --force-reinstall -i https://mirrors.aliyun.com/pypi/simple/
wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar --no-check-certificate wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar --no-check-certificate
cd ./train_data/ && tar xf XFUND.tar cd ./train_data/ && tar xf XFUND.tar
# expand gt.txt 10 times # expand gt.txt 10 times
...@@ -161,6 +160,8 @@ if [ ${MODE} = "lite_train_lite_infer" ];then ...@@ -161,6 +160,8 @@ if [ ${MODE} = "lite_train_lite_infer" ];then
ln -s ./icdar2015_lite ./icdar2015 ln -s ./icdar2015_lite ./icdar2015
wget -nc -P ./ic15_data/ https://paddleocr.bj.bcebos.com/dataset/rec_gt_train_lite.txt --no-check-certificate wget -nc -P ./ic15_data/ https://paddleocr.bj.bcebos.com/dataset/rec_gt_train_lite.txt --no-check-certificate
wget -nc -P ./ic15_data/ https://paddleocr.bj.bcebos.com/dataset/rec_gt_test_lite.txt --no-check-certificate wget -nc -P ./ic15_data/ https://paddleocr.bj.bcebos.com/dataset/rec_gt_test_lite.txt --no-check-certificate
mv ic15_data/rec_gt_train_lite.txt ic15_data/rec_gt_train.txt
mv ic15_data/rec_gt_test_lite.txt ic15_data/rec_gt_test.txt
cd ../ cd ../
cd ./inference && tar xf rec_inference.tar && cd ../ cd ./inference && tar xf rec_inference.tar && cd ../
if [ ${model_name} == "ch_PP-OCRv2_det" ] || [ ${model_name} == "ch_PP-OCRv2_det_PACT" ]; then if [ ${model_name} == "ch_PP-OCRv2_det" ] || [ ${model_name} == "ch_PP-OCRv2_det_PACT" ]; then
...@@ -221,8 +222,7 @@ if [ ${MODE} = "lite_train_lite_infer" ];then ...@@ -221,8 +222,7 @@ if [ ${MODE} = "lite_train_lite_infer" ];then
cd ./pretrain_models/ && tar xf rec_r32_gaspin_bilstm_att_train.tar && cd ../ cd ./pretrain_models/ && tar xf rec_r32_gaspin_bilstm_att_train.tar && cd ../
fi fi
if [ ${model_name} == "layoutxlm_ser" ] || [ ${model_name} == "vi_layoutxlm_ser" ]; then if [ ${model_name} == "layoutxlm_ser" ] || [ ${model_name} == "vi_layoutxlm_ser" ]; then
pip install -r ppstructure/vqa/requirements.txt pip install -r ppstructure/kie/requirements.txt
pip install paddlenlp\>=2.3.5 --force-reinstall -i https://mirrors.aliyun.com/pypi/simple/
wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar --no-check-certificate wget -nc -P ./train_data/ https://paddleocr.bj.bcebos.com/ppstructure/dataset/XFUND.tar --no-check-certificate
cd ./train_data/ && tar xf XFUND.tar cd ./train_data/ && tar xf XFUND.tar
cd ../ cd ../
......
tools/eval.py
浏览文件 @ 96a8e0f8
...@@ -23,6 +23,7 @@ __dir__ = os.path.dirname(os.path.abspath(__file__)) ...@@ -23,6 +23,7 @@ __dir__ = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, __dir__) sys.path.insert(0, __dir__)
sys.path.insert(0, os.path.abspath(os.path.join(__dir__, '..'))) sys.path.insert(0, os.path.abspath(os.path.join(__dir__, '..')))
import paddle
from ppocr.data import build_dataloader from ppocr.data import build_dataloader
from ppocr.modeling.architectures import build_model from ppocr.modeling.architectures import build_model
from ppocr.postprocess import build_post_process from ppocr.postprocess import build_post_process
...@@ -86,6 +87,30 @@ def main(): ...@@ -86,6 +87,30 @@ def main():
else: else:
model_type = None model_type = None
# build metric
eval_class = build_metric(config['Metric'])
# amp
use_amp = config["Global"].get("use_amp", False)
amp_level = config["Global"].get("amp_level", 'O2')
amp_custom_black_list = config['Global'].get('amp_custom_black_list',[])
if use_amp:
AMP_RELATED_FLAGS_SETTING = {
'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
'FLAGS_max_inplace_grad_add': 8,
}
paddle.fluid.set_flags(AMP_RELATED_FLAGS_SETTING)
scale_loss = config["Global"].get("scale_loss", 1.0)
use_dynamic_loss_scaling = config["Global"].get(
"use_dynamic_loss_scaling", False)
scaler = paddle.amp.GradScaler(
init_loss_scaling=scale_loss,
use_dynamic_loss_scaling=use_dynamic_loss_scaling)
if amp_level == "O2":
model = paddle.amp.decorate(
models=model, level=amp_level, master_weight=True)
else:
scaler = None
best_model_dict = load_model( best_model_dict = load_model(
config, model, model_type=config['Architecture']["model_type"]) config, model, model_type=config['Architecture']["model_type"])
if len(best_model_dict): if len(best_model_dict):
...@@ -93,11 +118,9 @@ def main(): ...@@ -93,11 +118,9 @@ def main():
for k, v in best_model_dict.items(): for k, v in best_model_dict.items():
logger.info('{}:{}'.format(k, v)) logger.info('{}:{}'.format(k, v))
# build metric
eval_class = build_metric(config['Metric'])
# start eval # start eval
metric = program.eval(model, valid_dataloader, post_process_class, metric = program.eval(model, valid_dataloader, post_process_class,
eval_class, model_type, extra_input) eval_class, model_type, extra_input, scaler, amp_level, amp_custom_black_list)
logger.info('metric eval ***************') logger.info('metric eval ***************')
for k, v in metric.items(): for k, v in metric.items():
logger.info('{}:{}'.format(k, v)) logger.info('{}:{}'.format(k, v))
......
tools/infer/predict_cls.py
浏览文件 @ 96a8e0f8
...@@ -30,7 +30,7 @@ import traceback ...@@ -30,7 +30,7 @@ import traceback
import tools.infer.utility as utility import tools.infer.utility as utility
from ppocr.postprocess import build_post_process from ppocr.postprocess import build_post_process
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
logger = get_logger() logger = get_logger()
...@@ -128,7 +128,7 @@ def main(args): ...@@ -128,7 +128,7 @@ def main(args):
valid_image_file_list = [] valid_image_file_list = []
img_list = [] img_list = []
for image_file in image_file_list: for image_file in image_file_list:
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
img = cv2.imread(image_file) img = cv2.imread(image_file)
if img is None: if img is None:
......
tools/infer/predict_det.py
浏览文件 @ 96a8e0f8
...@@ -27,7 +27,7 @@ import sys ...@@ -27,7 +27,7 @@ import sys
import tools.infer.utility as utility import tools.infer.utility as utility
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
from ppocr.data import create_operators, transform from ppocr.data import create_operators, transform
from ppocr.postprocess import build_post_process from ppocr.postprocess import build_post_process
import json import json
...@@ -289,7 +289,7 @@ if __name__ == "__main__": ...@@ -289,7 +289,7 @@ if __name__ == "__main__":
os.makedirs(draw_img_save) os.makedirs(draw_img_save)
save_results = [] save_results = []
for image_file in image_file_list: for image_file in image_file_list:
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
img = cv2.imread(image_file) img = cv2.imread(image_file)
if img is None: if img is None:
......
tools/infer/predict_e2e.py
浏览文件 @ 96a8e0f8
...@@ -27,7 +27,7 @@ import sys ...@@ -27,7 +27,7 @@ import sys
import tools.infer.utility as utility import tools.infer.utility as utility
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
from ppocr.data import create_operators, transform from ppocr.data import create_operators, transform
from ppocr.postprocess import build_post_process from ppocr.postprocess import build_post_process
...@@ -148,7 +148,7 @@ if __name__ == "__main__": ...@@ -148,7 +148,7 @@ if __name__ == "__main__":
if not os.path.exists(draw_img_save): if not os.path.exists(draw_img_save):
os.makedirs(draw_img_save) os.makedirs(draw_img_save)
for image_file in image_file_list: for image_file in image_file_list:
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
img = cv2.imread(image_file) img = cv2.imread(image_file)
if img is None: if img is None:
......
tools/infer/predict_rec.py
浏览文件 @ 96a8e0f8
...@@ -30,7 +30,7 @@ import paddle ...@@ -30,7 +30,7 @@ import paddle
import tools.infer.utility as utility import tools.infer.utility as utility
from ppocr.postprocess import build_post_process from ppocr.postprocess import build_post_process
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
logger = get_logger() logger = get_logger()
...@@ -349,6 +349,13 @@ class TextRecognizer(object): ...@@ -349,6 +349,13 @@ class TextRecognizer(object):
for beg_img_no in range(0, img_num, batch_num): for beg_img_no in range(0, img_num, batch_num):
end_img_no = min(img_num, beg_img_no + batch_num) end_img_no = min(img_num, beg_img_no + batch_num)
norm_img_batch = [] norm_img_batch = []
if self.rec_algorithm == "SRN":
encoder_word_pos_list = []
gsrm_word_pos_list = []
gsrm_slf_attn_bias1_list = []
gsrm_slf_attn_bias2_list = []
if self.rec_algorithm == "SAR":
valid_ratios = []
imgC, imgH, imgW = self.rec_image_shape[:3] imgC, imgH, imgW = self.rec_image_shape[:3]
max_wh_ratio = imgW / imgH max_wh_ratio = imgW / imgH
# max_wh_ratio = 0 # max_wh_ratio = 0
...@@ -357,22 +364,16 @@ class TextRecognizer(object): ...@@ -357,22 +364,16 @@ class TextRecognizer(object):
wh_ratio = w * 1.0 / h wh_ratio = w * 1.0 / h
max_wh_ratio = max(max_wh_ratio, wh_ratio) max_wh_ratio = max(max_wh_ratio, wh_ratio)
for ino in range(beg_img_no, end_img_no): for ino in range(beg_img_no, end_img_no):
if self.rec_algorithm == "SAR": if self.rec_algorithm == "SAR":
norm_img, _, _, valid_ratio = self.resize_norm_img_sar( norm_img, _, _, valid_ratio = self.resize_norm_img_sar(
img_list[indices[ino]], self.rec_image_shape) img_list[indices[ino]], self.rec_image_shape)
norm_img = norm_img[np.newaxis, :] norm_img = norm_img[np.newaxis, :]
valid_ratio = np.expand_dims(valid_ratio, axis=0) valid_ratio = np.expand_dims(valid_ratio, axis=0)
valid_ratios = []
valid_ratios.append(valid_ratio) valid_ratios.append(valid_ratio)
norm_img_batch.append(norm_img) norm_img_batch.append(norm_img)
elif self.rec_algorithm == "SRN": elif self.rec_algorithm == "SRN":
norm_img = self.process_image_srn( norm_img = self.process_image_srn(
img_list[indices[ino]], self.rec_image_shape, 8, 25) img_list[indices[ino]], self.rec_image_shape, 8, 25)
encoder_word_pos_list = []
gsrm_word_pos_list = []
gsrm_slf_attn_bias1_list = []
gsrm_slf_attn_bias2_list = []
encoder_word_pos_list.append(norm_img[1]) encoder_word_pos_list.append(norm_img[1])
gsrm_word_pos_list.append(norm_img[2]) gsrm_word_pos_list.append(norm_img[2])
gsrm_slf_attn_bias1_list.append(norm_img[3]) gsrm_slf_attn_bias1_list.append(norm_img[3])
...@@ -399,7 +400,9 @@ class TextRecognizer(object): ...@@ -399,7 +400,9 @@ class TextRecognizer(object):
norm_img_batch.append(norm_img) norm_img_batch.append(norm_img)
elif self.rec_algorithm == "RobustScanner": elif self.rec_algorithm == "RobustScanner":
norm_img, _, _, valid_ratio = self.resize_norm_img_sar( norm_img, _, _, valid_ratio = self.resize_norm_img_sar(
img_list[indices[ino]], self.rec_image_shape, width_downsample_ratio=0.25) img_list[indices[ino]],
self.rec_image_shape,
width_downsample_ratio=0.25)
norm_img = norm_img[np.newaxis, :] norm_img = norm_img[np.newaxis, :]
valid_ratio = np.expand_dims(valid_ratio, axis=0) valid_ratio = np.expand_dims(valid_ratio, axis=0)
valid_ratios = [] valid_ratios = []
...@@ -484,11 +487,7 @@ class TextRecognizer(object): ...@@ -484,11 +487,7 @@ class TextRecognizer(object):
elif self.rec_algorithm == "RobustScanner": elif self.rec_algorithm == "RobustScanner":
valid_ratios = np.concatenate(valid_ratios) valid_ratios = np.concatenate(valid_ratios)
word_positions_list = np.concatenate(word_positions_list) word_positions_list = np.concatenate(word_positions_list)
inputs = [ inputs = [norm_img_batch, valid_ratios, word_positions_list]
norm_img_batch,
valid_ratios,
word_positions_list
]
if self.use_onnx: if self.use_onnx:
input_dict = {} input_dict = {}
...@@ -555,7 +554,7 @@ def main(args): ...@@ -555,7 +554,7 @@ def main(args):
res = text_recognizer([img] * int(args.rec_batch_num)) res = text_recognizer([img] * int(args.rec_batch_num))
for image_file in image_file_list: for image_file in image_file_list:
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
img = cv2.imread(image_file) img = cv2.imread(image_file)
if img is None: if img is None:
......
tools/infer/predict_sr.py
浏览文件 @ 96a8e0f8
...@@ -30,7 +30,7 @@ import paddle ...@@ -30,7 +30,7 @@ import paddle
import tools.infer.utility as utility import tools.infer.utility as utility
from ppocr.postprocess import build_post_process from ppocr.postprocess import build_post_process
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
logger = get_logger() logger = get_logger()
...@@ -120,7 +120,7 @@ def main(args): ...@@ -120,7 +120,7 @@ def main(args):
res = text_recognizer([img] * int(args.sr_batch_num)) res = text_recognizer([img] * int(args.sr_batch_num))
for image_file in image_file_list: for image_file in image_file_list:
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
img = Image.open(image_file).convert("RGB") img = Image.open(image_file).convert("RGB")
if img is None: if img is None:
......
tools/infer/predict_system.py
浏览文件 @ 96a8e0f8
...@@ -32,7 +32,7 @@ import tools.infer.utility as utility ...@@ -32,7 +32,7 @@ import tools.infer.utility as utility
import tools.infer.predict_rec as predict_rec import tools.infer.predict_rec as predict_rec
import tools.infer.predict_det as predict_det import tools.infer.predict_det as predict_det
import tools.infer.predict_cls as predict_cls import tools.infer.predict_cls as predict_cls
from ppocr.utils.utility import get_image_file_list, check_and_read_gif from ppocr.utils.utility import get_image_file_list, check_and_read
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from tools.infer.utility import draw_ocr_box_txt, get_rotate_crop_image from tools.infer.utility import draw_ocr_box_txt, get_rotate_crop_image
logger = get_logger() logger = get_logger()
...@@ -120,11 +120,14 @@ def sorted_boxes(dt_boxes): ...@@ -120,11 +120,14 @@ def sorted_boxes(dt_boxes):
_boxes = list(sorted_boxes) _boxes = list(sorted_boxes)
for i in range(num_boxes - 1): for i in range(num_boxes - 1):
if abs(_boxes[i + 1][0][1] - _boxes[i][0][1]) < 10 and \ for j in range(i, 0, -1):
(_boxes[i + 1][0][0] < _boxes[i][0][0]): if abs(_boxes[j + 1][0][1] - _boxes[j][0][1]) < 10 and \
tmp = _boxes[i] (_boxes[j + 1][0][0] < _boxes[j][0][0]):
_boxes[i] = _boxes[i + 1] tmp = _boxes[j]
_boxes[i + 1] = tmp _boxes[j] = _boxes[j + 1]
_boxes[j + 1] = tmp
else:
break
return _boxes return _boxes
...@@ -156,7 +159,7 @@ def main(args): ...@@ -156,7 +159,7 @@ def main(args):
count = 0 count = 0
for idx, image_file in enumerate(image_file_list): for idx, image_file in enumerate(image_file_list):
img, flag = check_and_read_gif(image_file) img, flag, _ = check_and_read(image_file)
if not flag: if not flag:
img = cv2.imread(image_file) img = cv2.imread(image_file)
if img is None: if img is None:
......
tools/infer/utility.py
浏览文件 @ 96a8e0f8
...@@ -181,14 +181,21 @@ def create_predictor(args, mode, logger): ...@@ -181,14 +181,21 @@ def create_predictor(args, mode, logger):
return sess, sess.get_inputs()[0], None, None return sess, sess.get_inputs()[0], None, None
else: else:
model_file_path = model_dir + "/inference.pdmodel" file_names = ['model', 'inference']
params_file_path = model_dir + "/inference.pdiparams" for file_name in file_names:
model_file_path = '{}/{}.pdmodel'.format(model_dir, file_name)
params_file_path = '{}/{}.pdiparams'.format(model_dir, file_name)
if os.path.exists(model_file_path) and os.path.exists(
params_file_path):
break
if not os.path.exists(model_file_path): if not os.path.exists(model_file_path):
raise ValueError("not find model file path {}".format( raise ValueError(
model_file_path)) "not find model.pdmodel or inference.pdmodel in {}".format(
model_dir))
if not os.path.exists(params_file_path): if not os.path.exists(params_file_path):
raise ValueError("not find params file path {}".format( raise ValueError(
params_file_path)) "not find model.pdiparams or inference.pdiparams in {}".format(
model_dir))
config = inference.Config(model_file_path, params_file_path) config = inference.Config(model_file_path, params_file_path)
...@@ -221,7 +228,8 @@ def create_predictor(args, mode, logger): ...@@ -221,7 +228,8 @@ def create_predictor(args, mode, logger):
# collect shape # collect shape
if args.shape_info_filename is not None: if args.shape_info_filename is not None:
if not os.path.exists(args.shape_info_filename): if not os.path.exists(args.shape_info_filename):
config.collect_shape_range_info(args.shape_info_filename) config.collect_shape_range_info(
args.shape_info_filename)
logger.info( logger.info(
f"collect dynamic shape info into : {args.shape_info_filename}" f"collect dynamic shape info into : {args.shape_info_filename}"
) )
...@@ -542,7 +550,7 @@ def text_visual(texts, ...@@ -542,7 +550,7 @@ def text_visual(texts,
def base64_to_cv2(b64str): def base64_to_cv2(b64str):
import base64 import base64
data = base64.b64decode(b64str.encode('utf8')) data = base64.b64decode(b64str.encode('utf8'))
data = np.fromstring(data, np.uint8) data = np.frombuffer(data, np.uint8)
data = cv2.imdecode(data, cv2.IMREAD_COLOR) data = cv2.imdecode(data, cv2.IMREAD_COLOR)
return data return data
......
tools/infer_kie.py
浏览文件 @ 96a8e0f8
...@@ -88,6 +88,29 @@ def draw_kie_result(batch, node, idx_to_cls, count): ...@@ -88,6 +88,29 @@ def draw_kie_result(batch, node, idx_to_cls, count):
cv2.imwrite(save_path, vis_img) cv2.imwrite(save_path, vis_img)
logger.info("The Kie Image saved in {}".format(save_path)) logger.info("The Kie Image saved in {}".format(save_path))
def write_kie_result(fout, node, data):
"""
Write infer result to output file, sorted by the predict label of each line.
The format keeps the same as the input with additional score attribute.
"""
import json
label = data['label']
annotations = json.loads(label)
max_value, max_idx = paddle.max(node, -1), paddle.argmax(node, -1)
node_pred_label = max_idx.numpy().tolist()
node_pred_score = max_value.numpy().tolist()
res = []
for i, label in enumerate(node_pred_label):
pred_score = '{:.2f}'.format(node_pred_score[i])
pred_res = {
'label': label,
'transcription': annotations[i]['transcription'],
'score': pred_score,
'points': annotations[i]['points'],
}
res.append(pred_res)
res.sort(key=lambda x: x['label'])
fout.writelines([json.dumps(res, ensure_ascii=False) + '\n'])
def main(): def main():
global_config = config['Global'] global_config = config['Global']
...@@ -114,7 +137,7 @@ def main(): ...@@ -114,7 +137,7 @@ def main():
warmup_times = 0 warmup_times = 0
count_t = [] count_t = []
with open(save_res_path, "wb") as fout: with open(save_res_path, "w") as fout:
with open(config['Global']['infer_img'], "rb") as f: with open(config['Global']['infer_img'], "rb") as f:
lines = f.readlines() lines = f.readlines()
for index, data_line in enumerate(lines): for index, data_line in enumerate(lines):
...@@ -139,6 +162,8 @@ def main(): ...@@ -139,6 +162,8 @@ def main():
node = F.softmax(node, -1) node = F.softmax(node, -1)
count_t.append(time.time() - st) count_t.append(time.time() - st)
draw_kie_result(batch, node, idx_to_cls, index) draw_kie_result(batch, node, idx_to_cls, index)
write_kie_result(fout, node, data)
fout.close()
logger.info("success!") logger.info("success!")
logger.info("It took {} s for predict {} images.".format( logger.info("It took {} s for predict {} images.".format(
np.sum(count_t), len(count_t))) np.sum(count_t), len(count_t)))
......
tools/infer_vqa_token_ser.py → tools/infer_kie_token_ser.py
浏览文件 @ 96a8e0f8
...@@ -75,6 +75,8 @@ class SerPredictor(object): ...@@ -75,6 +75,8 @@ class SerPredictor(object):
self.ocr_engine = PaddleOCR( self.ocr_engine = PaddleOCR(
use_angle_cls=False, use_angle_cls=False,
show_log=False, show_log=False,
rec_model_dir=global_config.get("kie_rec_model_dir", None),
det_model_dir=global_config.get("kie_det_model_dir", None),
use_gpu=global_config['use_gpu']) use_gpu=global_config['use_gpu'])
# create data ops # create data ops
......
tools/infer_vqa_token_ser_re.py → tools/infer_kie_token_ser_re.py
浏览文件 @ 96a8e0f8
...@@ -39,7 +39,7 @@ from ppocr.utils.visual import draw_re_results ...@@ -39,7 +39,7 @@ from ppocr.utils.visual import draw_re_results
from ppocr.utils.logging import get_logger from ppocr.utils.logging import get_logger
from ppocr.utils.utility import get_image_file_list, load_vqa_bio_label_maps, print_dict from ppocr.utils.utility import get_image_file_list, load_vqa_bio_label_maps, print_dict
from tools.program import ArgsParser, load_config, merge_config from tools.program import ArgsParser, load_config, merge_config
from tools.infer_vqa_token_ser import SerPredictor from tools.infer_kie_token_ser import SerPredictor
class ReArgsParser(ArgsParser): class ReArgsParser(ArgsParser):
...@@ -205,9 +205,7 @@ if __name__ == '__main__': ...@@ -205,9 +205,7 @@ if __name__ == '__main__':
result = ser_re_engine(data) result = ser_re_engine(data)
result = result[0] result = result[0]
fout.write(img_path + "\t" + json.dumps( fout.write(img_path + "\t" + json.dumps(
{ result, ensure_ascii=False) + "\n")
"ser_result": result,
}, ensure_ascii=False) + "\n")
img_res = draw_re_results(img_path, result) img_res = draw_re_results(img_path, result)
cv2.imwrite(save_img_path, img_res) cv2.imwrite(save_img_path, img_res)
......
tools/infer_sr.py
浏览文件 @ 96a8e0f8
...@@ -63,14 +63,14 @@ def main(): ...@@ -63,14 +63,14 @@ def main():
elif op_name in ['SRResize']: elif op_name in ['SRResize']:
op[op_name]['infer_mode'] = True op[op_name]['infer_mode'] = True
elif op_name == 'KeepKeys': elif op_name == 'KeepKeys':
op[op_name]['keep_keys'] = ['imge_lr'] op[op_name]['keep_keys'] = ['img_lr']
transforms.append(op) transforms.append(op)
global_config['infer_mode'] = True global_config['infer_mode'] = True
ops = create_operators(transforms, global_config) ops = create_operators(transforms, global_config)
save_res_path = config['Global'].get('save_res_path', "./infer_result") save_visual_path = config['Global'].get('save_visual', "infer_result/")
if not os.path.exists(os.path.dirname(save_res_path)): if not os.path.exists(os.path.dirname(save_visual_path)):
os.makedirs(os.path.dirname(save_res_path)) os.makedirs(os.path.dirname(save_visual_path))
model.eval() model.eval()
for file in get_image_file_list(config['Global']['infer_img']): for file in get_image_file_list(config['Global']['infer_img']):
...@@ -87,7 +87,7 @@ def main(): ...@@ -87,7 +87,7 @@ def main():
fm_sr = (sr_img.numpy() * 255).transpose(1, 2, 0).astype(np.uint8) fm_sr = (sr_img.numpy() * 255).transpose(1, 2, 0).astype(np.uint8)
fm_lr = (lr_img.numpy() * 255).transpose(1, 2, 0).astype(np.uint8) fm_lr = (lr_img.numpy() * 255).transpose(1, 2, 0).astype(np.uint8)
img_name_pure = os.path.split(file)[-1] img_name_pure = os.path.split(file)[-1]
cv2.imwrite("infer_result/sr_{}".format(img_name_pure), cv2.imwrite("{}/sr_{}".format(save_visual_path, img_name_pure),
fm_sr[:, :, ::-1]) fm_sr[:, :, ::-1])
logger.info("The visualized image saved in infer_result/sr_{}".format( logger.info("The visualized image saved in infer_result/sr_{}".format(
img_name_pure)) img_name_pure))
......
tools/program.py
浏览文件 @ 96a8e0f8
...@@ -162,18 +162,18 @@ def to_float32(preds): ...@@ -162,18 +162,18 @@ def to_float32(preds):
for k in preds: for k in preds:
if isinstance(preds[k], dict) or isinstance(preds[k], list): if isinstance(preds[k], dict) or isinstance(preds[k], list):
preds[k] = to_float32(preds[k]) preds[k] = to_float32(preds[k])
else: elif isinstance(preds[k], paddle.Tensor):
preds[k] = paddle.to_tensor(preds[k], dtype='float32') preds[k] = preds[k].astype(paddle.float32)
elif isinstance(preds, list): elif isinstance(preds, list):
for k in range(len(preds)): for k in range(len(preds)):
if isinstance(preds[k], dict): if isinstance(preds[k], dict):
preds[k] = to_float32(preds[k]) preds[k] = to_float32(preds[k])
elif isinstance(preds[k], list): elif isinstance(preds[k], list):
preds[k] = to_float32(preds[k]) preds[k] = to_float32(preds[k])
else: elif isinstance(preds[k], paddle.Tensor):
preds[k] = paddle.to_tensor(preds[k], dtype='float32') preds[k] = preds[k].astype(paddle.float32)
else: elif isinstance(preds, paddle.Tensor):
preds = paddle.to_tensor(preds, dtype='float32') preds = preds.astype(paddle.float32)
return preds return preds
...@@ -190,7 +190,9 @@ def train(config, ...@@ -190,7 +190,9 @@ def train(config,
pre_best_model_dict, pre_best_model_dict,
logger, logger,
log_writer=None, log_writer=None,
scaler=None): scaler=None,
amp_level='O2',
amp_custom_black_list=[]):
cal_metric_during_train = config['Global'].get('cal_metric_during_train', cal_metric_during_train = config['Global'].get('cal_metric_during_train',
False) False)
calc_epoch_interval = config['Global'].get('calc_epoch_interval', 1) calc_epoch_interval = config['Global'].get('calc_epoch_interval', 1)
...@@ -230,7 +232,8 @@ def train(config, ...@@ -230,7 +232,8 @@ def train(config,
use_srn = config['Architecture']['algorithm'] == "SRN" use_srn = config['Architecture']['algorithm'] == "SRN"
extra_input_models = [ extra_input_models = [
"SRN", "NRTR", "SAR", "SEED", "SVTR", "SPIN", "VisionLAN", "RobustScanner" "SRN", "NRTR", "SAR", "SEED", "SVTR", "SPIN", "VisionLAN",
"RobustScanner"
] ]
extra_input = False extra_input = False
if config['Architecture']['algorithm'] == 'Distillation': if config['Architecture']['algorithm'] == 'Distillation':
...@@ -276,10 +279,10 @@ def train(config, ...@@ -276,10 +279,10 @@ def train(config,
model_average = True model_average = True
# use amp # use amp
if scaler: if scaler:
with paddle.amp.auto_cast(level='O2'): with paddle.amp.auto_cast(level=amp_level, custom_black_list=amp_custom_black_list):
if model_type == 'table' or extra_input: if model_type == 'table' or extra_input:
preds = model(images, data=batch[1:]) preds = model(images, data=batch[1:])
elif model_type in ["kie", 'vqa']: elif model_type in ["kie"]:
preds = model(batch) preds = model(batch)
else: else:
preds = model(images) preds = model(images)
...@@ -292,7 +295,7 @@ def train(config, ...@@ -292,7 +295,7 @@ def train(config,
else: else:
if model_type == 'table' or extra_input: if model_type == 'table' or extra_input:
preds = model(images, data=batch[1:]) preds = model(images, data=batch[1:])
elif model_type in ["kie", 'vqa', 'sr']: elif model_type in ["kie", 'sr']:
preds = model(batch) preds = model(batch)
else: else:
preds = model(images) preds = model(images)
...@@ -381,7 +384,9 @@ def train(config, ...@@ -381,7 +384,9 @@ def train(config,
eval_class, eval_class,
model_type, model_type,
extra_input=extra_input, extra_input=extra_input,
scaler=scaler) scaler=scaler,
amp_level=amp_level,
amp_custom_black_list=amp_custom_black_list)
cur_metric_str = 'cur metric, {}'.format(', '.join( cur_metric_str = 'cur metric, {}'.format(', '.join(
['{}: {}'.format(k, v) for k, v in cur_metric.items()])) ['{}: {}'.format(k, v) for k, v in cur_metric.items()]))
logger.info(cur_metric_str) logger.info(cur_metric_str)
...@@ -472,7 +477,9 @@ def eval(model, ...@@ -472,7 +477,9 @@ def eval(model,
eval_class, eval_class,
model_type=None, model_type=None,
extra_input=False, extra_input=False,
scaler=None): scaler=None,
amp_level='O2',
amp_custom_black_list = []):
model.eval() model.eval()
with paddle.no_grad(): with paddle.no_grad():
total_frame = 0.0 total_frame = 0.0
...@@ -493,46 +500,27 @@ def eval(model, ...@@ -493,46 +500,27 @@ def eval(model,
# use amp # use amp
if scaler: if scaler:
with paddle.amp.auto_cast(level='O2'): with paddle.amp.auto_cast(level=amp_level, custom_black_list=amp_custom_black_list):
if model_type == 'table' or extra_input: if model_type == 'table' or extra_input:
preds = model(images, data=batch[1:]) preds = model(images, data=batch[1:])
elif model_type in ["kie", 'vqa']: elif model_type in ["kie"]:
preds = model(batch) preds = model(batch)
elif model_type in ['sr']: elif model_type in ['sr']:
preds = model(batch) preds = model(batch)
sr_img = preds["sr_img"] sr_img = preds["sr_img"]
lr_img = preds["lr_img"] lr_img = preds["lr_img"]
for i in (range(sr_img.shape[0])):
fm_sr = (sr_img[i].numpy() * 255).transpose(
1, 2, 0).astype(np.uint8)
fm_lr = (lr_img[i].numpy() * 255).transpose(
1, 2, 0).astype(np.uint8)
cv2.imwrite("output/images/{}_{}_sr.jpg".format(
sum_images, i), fm_sr)
cv2.imwrite("output/images/{}_{}_lr.jpg".format(
sum_images, i), fm_lr)
else: else:
preds = model(images) preds = model(images)
preds = to_float32(preds)
else: else:
if model_type == 'table' or extra_input: if model_type == 'table' or extra_input:
preds = model(images, data=batch[1:]) preds = model(images, data=batch[1:])
elif model_type in ["kie", 'vqa']: elif model_type in ["kie"]:
preds = model(batch) preds = model(batch)
elif model_type in ['sr']: elif model_type in ['sr']:
preds = model(batch) preds = model(batch)
sr_img = preds["sr_img"] sr_img = preds["sr_img"]
lr_img = preds["lr_img"] lr_img = preds["lr_img"]
for i in (range(sr_img.shape[0])):
fm_sr = (sr_img[i].numpy() * 255).transpose(
1, 2, 0).astype(np.uint8)
fm_lr = (lr_img[i].numpy() * 255).transpose(
1, 2, 0).astype(np.uint8)
cv2.imwrite("output/images/{}_{}_sr.jpg".format(
sum_images, i), fm_sr)
cv2.imwrite("output/images/{}_{}_lr.jpg".format(
sum_images, i), fm_lr)
else: else:
preds = model(images) preds = model(images)
...@@ -545,9 +533,10 @@ def eval(model, ...@@ -545,9 +533,10 @@ def eval(model,
# Obtain usable results from post-processing methods # Obtain usable results from post-processing methods
total_time += time.time() - start total_time += time.time() - start
# Evaluate the results of the current batch # Evaluate the results of the current batch
if model_type in ['kie']: if model_type in ['table', 'kie']:
if post_process_class is None:
eval_class(preds, batch_numpy) eval_class(preds, batch_numpy)
elif model_type in ['table', 'vqa']: else:
post_result = post_process_class(preds, batch_numpy) post_result = post_process_class(preds, batch_numpy)
eval_class(post_result, batch_numpy) eval_class(post_result, batch_numpy)
elif model_type in ['sr']: elif model_type in ['sr']:
......
tools/train.py
浏览文件 @ 96a8e0f8
...@@ -138,15 +138,15 @@ def main(config, device, logger, vdl_writer): ...@@ -138,15 +138,15 @@ def main(config, device, logger, vdl_writer):
# build metric # build metric
eval_class = build_metric(config['Metric']) eval_class = build_metric(config['Metric'])
# load pretrain model
pre_best_model_dict = load_model(config, model, optimizer,
config['Architecture']["model_type"])
logger.info('train dataloader has {} iters'.format(len(train_dataloader))) logger.info('train dataloader has {} iters'.format(len(train_dataloader)))
if valid_dataloader is not None: if valid_dataloader is not None:
logger.info('valid dataloader has {} iters'.format( logger.info('valid dataloader has {} iters'.format(
len(valid_dataloader))) len(valid_dataloader)))
use_amp = config["Global"].get("use_amp", False) use_amp = config["Global"].get("use_amp", False)
amp_level = config["Global"].get("amp_level", 'O2')
amp_custom_black_list = config['Global'].get('amp_custom_black_list',[])
if use_amp: if use_amp:
AMP_RELATED_FLAGS_SETTING = { AMP_RELATED_FLAGS_SETTING = {
'FLAGS_cudnn_batchnorm_spatial_persistent': 1, 'FLAGS_cudnn_batchnorm_spatial_persistent': 1,
...@@ -159,17 +159,22 @@ def main(config, device, logger, vdl_writer): ...@@ -159,17 +159,22 @@ def main(config, device, logger, vdl_writer):
scaler = paddle.amp.GradScaler( scaler = paddle.amp.GradScaler(
init_loss_scaling=scale_loss, init_loss_scaling=scale_loss,
use_dynamic_loss_scaling=use_dynamic_loss_scaling) use_dynamic_loss_scaling=use_dynamic_loss_scaling)
if amp_level == "O2":
model, optimizer = paddle.amp.decorate( model, optimizer = paddle.amp.decorate(
models=model, optimizers=optimizer, level='O2', master_weight=True) models=model, optimizers=optimizer, level=amp_level, master_weight=True)
else: else:
scaler = None scaler = None
# load pretrain model
pre_best_model_dict = load_model(config, model, optimizer,
config['Architecture']["model_type"])
if config['Global']['distributed']: if config['Global']['distributed']:
model = paddle.DataParallel(model) model = paddle.DataParallel(model)
# start train # start train
program.train(config, train_dataloader, valid_dataloader, device, model, program.train(config, train_dataloader, valid_dataloader, device, model,
loss_class, optimizer, lr_scheduler, post_process_class, loss_class, optimizer, lr_scheduler, post_process_class,
eval_class, pre_best_model_dict, logger, vdl_writer, scaler) eval_class, pre_best_model_dict, logger, vdl_writer, scaler,amp_level, amp_custom_black_list)
def test_reader(config, device, logger): def test_reader(config, device, logger):
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或