Skip to content

GitLab

Commit aa59fca5 authored by Leif's avatar Leif
Browse files

Merge remote-tracking branch 'origin/dygraph' into dygraph

parents 12d15752 f01f24c7
Hide whitespace changes
Inline Side-by-side
Showing with 2546 additions and 294 deletions
README.md
View file @ aa59fca5
......@@ -3,10 +3,6 @@ English | [简体中文](README_ch.md)
<p align="center">
<img src="./doc/PaddleOCR_log.png" align="middle" width = "600"/>
<p align="center">
------------------------------------------------------------------------------------------
<p align="left">
<a href="./LICENSE"><img src="https://img.shields.io/badge/license-Apache%202-dfd.svg"></a>
<a href="https://github.com/PaddlePaddle/PaddleOCR/releases"><img src="https://img.shields.io/github/v/release/PaddlePaddle/PaddleOCR?color=ffa"></a>
......@@ -32,60 +28,42 @@ PaddleOCR aims to create multilingual, awesome, leading, and practical OCR tools
- [more](./doc/doc_en/update_en.md)
## Features
- PP-OCR - A series of high-quality pre-trained models, comparable to commercial products
- Ultra lightweight PP-OCRv2 series models: detection (3.1M) + direction classifier (1.4M) + recognition 8.5M) = 13.0M
- Ultra lightweight PP-OCR mobile series models: detection (3.0M) + direction classifier (1.4M) + recognition (5.0M) = 9.4M
- General PP-OCR server series models: detection (47.1M) + direction classifier (1.4M) + recognition (94.9M) = 143.4M
- Support Chinese, English, and digit recognition, vertical text recognition, and long text recognition
- Support multi-lingual recognition: about 80 languages like Korean, Japanese, German, French, etc
- PP-Structure: a document structurize system
- Support layout analysis and table recognition (support export to Excel)
- Support key information extraction
- Support DocVQA
- Rich OCR toolkit
- Semi-automatic data annotation tool, i.e., PPOCRLabel: support fast and efficient data annotation
- Data synthesis tool, i.e., Style-Text: easy to synthesize a large number of images which are similar to the target scene image
- Support user-defined training, provides rich predictive inference deployment solutions
- Support PIP installation, easy to use
- Support Linux, Windows, MacOS and other systems
## Visualization
<div align="center">
<img src="doc/imgs_results/ch_ppocr_mobile_v2.0/test_add_91.jpg" width="800">
<img src="doc/imgs_results/multi_lang/img_01.jpg" width="800">
<img src="doc/imgs_results/multi_lang/img_02.jpg" width="800">
</div>
## Features
The above pictures are the visualizations of the general ppocr_server model. For more effect pictures, please see [More visualizations](./doc/doc_en/visualization_en.md).
PaddleOCR support a variety of cutting-edge algorithms related to OCR, and developed industrial featured models/solution [PP-OCR](./doc/doc_en/ppocr_introduction_en.md) and [PP-Structure](./ppstructure/README.md) on this basis, and get through the whole process of data production, model training, compression, inference and deployment.
<a name="Community"></a>
## Community
- Scan the QR code below with your Wechat, you can join the official technical discussion group. Looking forward to your participation.
![](./doc/features_en.png)
<div align="center">
<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "200" height = "200" />
</div>
> It is recommended to start with the “quick experience” in the document tutorial
## Quick Experience
You can also quickly experience the ultra-lightweight OCR : [Online Experience](https://www.paddlepaddle.org.cn/hub/scene/ocr)
- Web online experience for the ultra-lightweight OCR: [Online Experience](https://www.paddlepaddle.org.cn/hub/scene/ocr)
- Mobile DEMO experience (based on EasyEdge and Paddle-Lite, supports iOS and Android systems): [Sign in to the website to obtain the QR code for installing the App](https://ai.baidu.com/easyedge/app/openSource?from=paddlelite)
- One line of code quick use: [Quick Start](./doc/doc_en/quickstart_en.md)
<a name="book"></a>
## E-book: *Dive Into OCR*
- [Dive Into OCR 📚](./doc/doc_en/ocr_book_en.md)
<a name="Community"></a>
## Community
Mobile DEMO experience (based on EasyEdge and Paddle-Lite, supports iOS and Android systems): [Sign in to the website to obtain the QR code for installing the App](https://ai.baidu.com/easyedge/app/openSource?from=paddlelite)
- **Join us**👬: Scan the QR code below with your Wechat, you can join the official technical discussion group. Looking forward to your participation.
- **Contribution**🏅️: [Contribution page](./doc/doc_en/thirdparty.md) contains various tools and applications developed by community developers using PaddleOCR, as well as the functions, optimized documents and codes contributed to PaddleOCR. It is an official honor wall for community developers and a broadcasting station to help publicize high-quality projects.
- **Regular Season**🎁: The community regular season is a point competition for OCR developers, covering four types: documents, codes, models and applications. Awards are selected and awarded on a quarterly basis. Please refer to the [link](https://github.com/PaddlePaddle/PaddleOCR/issues/4982) for more details.
Also, you can scan the QR code below to install the App (**Android support only**)
<div align="center">
<img src="./doc/ocr-android-easyedge.png" width = "200" height = "200" />
<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "200" height = "200" />
</div>
- [**OCR Quick Start**](./doc/doc_en/quickstart_en.md)
<a name="Supported-Chinese-model-list"></a>
## PP-OCR Series Model List（Update on September 8th）
| Model introduction | Model name | Recommended scene | Detection model | Direction classifier | Recognition model |
......@@ -95,72 +73,67 @@ Mobile DEMO experience (based on EasyEdge and Paddle-Lite, supports iOS and Andr
| Chinese and English general PP-OCR model (143.4M) | ch_ppocr_server_v2.0_xx | Server |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_det_train.tar) |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_train.tar) |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_server_v2.0_rec_train.tar) |
For more model downloads (including multiple languages), please refer to [PP-OCR series model downloads](./doc/doc_en/models_list_en.md).
For a new language request, please refer to [Guideline for new language_requests](#language_requests).
- For more model downloads (including multiple languages), please refer to [PP-OCR series model downloads](./doc/doc_en/models_list_en.md).
- For a new language request, please refer to [Guideline for new language_requests](#language_requests).
- For structural document analysis models, please refer to [PP-Structure models](./ppstructure/docs/models_list_en.md).
## Tutorials
- [Environment Preparation](./doc/doc_en/environment_en.md)
- [Quick Start](./doc/doc_en/quickstart_en.md)
- [PaddleOCR Overview and Project Clone](./doc/doc_en/paddleOCR_overview_en.md)
- PP-OCR Industry Landing: from Training to Deployment
- [PP-OCR Model Zoo](./doc/doc_en/models_en.md)
- [PP-OCR Model Download](./doc/doc_en/models_list_en.md)
- [Python Inference for PP-OCR Model Zoo](./doc/doc_en/inference_ppocr_en.md)
- [PP-OCR Training](./doc/doc_en/training_en.md)
- [PP-OCR 🔥](./doc/doc_en/ppocr_introduction_en.md)
- [Quick Start](./doc/doc_en/quickstart_en.md)
- [Model Zoo](./doc/doc_en/models_en.md)
- [Model training](./doc/doc_en/training_en.md)
- [Text Detection](./doc/doc_en/detection_en.md)
- [Text Recognition](./doc/doc_en/recognition_en.md)
- [Text Direction Classification](./doc/doc_en/angle_class_en.md)
- [Yml Configuration](./doc/doc_en/config_en.md)
- PP-OCR Models Compression
- [Knowledge Distillation](./doc/doc_en/knowledge_distillation_en.md)
- Model Compression
- [Model Quantization](./deploy/slim/quantization/README_en.md)
- [Model Pruning](./deploy/slim/prune/README_en.md)
- Inference and Deployment
- [C++ Inference](./deploy/cpp_infer/readme_en.md)
- [Knowledge Distillation](./doc/doc_en/knowledge_distillation_en.md)
- [Inference and Deployment](./deploy/README.md)
- [Python Inference](./doc/doc_en/inference_ppocr_en.md)
- [C++ Inference](./deploy/cpp_infer/readme.md)
- [Serving](./deploy/pdserving/README.md)
- [Mobile](./deploy/lite/readme_en.md)
- [Mobile](./deploy/lite/readme.md)
- [Paddle2ONNX](./deploy/paddle2onnx/readme.md)
- [Benchmark](./doc/doc_en/benchmark_en.md)
- [PP-Structure: Information Extraction](./ppstructure/README.md)
- [Layout Parser](./ppstructure/layout/README.md)
- [Table Recognition](./ppstructure/table/README.md)
- [DocVQA](./ppstructure/vqa/README.md)
- [Key Information Extraction](./ppstructure/docs/kie.md)
- Academic Circles
- [Two-stage Algorithm](./doc/doc_en/algorithm_overview_en.md)
- [PGNet Algorithm](./doc/doc_en/pgnet_en.md)
- [Python Inference](./doc/doc_en/inference_en.md)
- [PP-Structure 🔥](./ppstructure/README.md)
- [Quick Start](./ppstructure/docs/quickstart_en.md)
- [Model Zoo](./ppstructure/docs/models_list_en.md)
- [Model training](./doc/doc_en/training_en.md)
- [Layout Parser](./ppstructure/layout/README.md)
- [Table Recognition](./ppstructure/table/README.md)
- [DocVQA](./ppstructure/vqa/README.md)
- [Key Information Extraction](./ppstructure/docs/kie_en.md)
- [Inference and Deployment](./deploy/README.md)
- [Python Inference](./ppstructure/docs/inference_en.md)
- [C++ Inference]()
- [Serving](./deploy/pdserving/README.md)
- [Academic algorithms](./doc/doc_en/algorithms_en.md)
- [Text detection](./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)
- [Add New Algorithms to PaddleOCR](./doc/doc_en/add_new_algorithm_en.md)
- Data Annotation and Synthesis
- [Semi-automatic Annotation Tool: PPOCRLabel](./PPOCRLabel/README.md)
- [Data Synthesis Tool: Style-Text](./StyleText/README.md)
- [Other Data Annotation Tools](./doc/doc_en/data_annotation_en.md)
- [Other Data Synthesis Tools](./doc/doc_en/data_synthesis_en.md)
- Datasets
- [General OCR Datasets(Chinese/English)](./doc/doc_en/datasets_en.md)
- [HandWritten_OCR_Datasets(Chinese)](./doc/doc_en/handwritten_datasets_en.md)
- [Various OCR Datasets(multilingual)](./doc/doc_en/vertical_and_multilingual_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)
- [Various OCR Datasets(multilingual)](doc/doc_en/dataset/vertical_and_multilingual_datasets_en.md)
- [Code Structure](./doc/doc_en/tree_en.md)
- [Visualization](#Visualization)
- [Community](#Community)
- [New language requests](#language_requests)
- [FAQ](./doc/doc_en/FAQ_en.md)
- [Community](#Community)
- [References](./doc/doc_en/reference_en.md)
- [License](#LICENSE)
- [Contribution](#CONTRIBUTION)
<a name="PP-OCRv2"></a>
## PP-OCRv2 Pipeline
<div align="center">
<img src="./doc/ppocrv2_framework.jpg" width="800">
</div>
[1] PP-OCR is a practical ultra-lightweight OCR system. It is mainly composed of three parts: DB text detection, detection frame correction and CRNN text recognition. The system adopts 19 effective strategies from 8 aspects including backbone network selection and adjustment, prediction head design, data augmentation, learning rate transformation strategy, regularization parameter selection, pre-training model use, and automatic model tailoring and quantization to optimize and slim down the models of each module (as shown in the green box above). The final results are an ultra-lightweight Chinese and English OCR model with an overall size of 3.5M and a 2.8M English digital OCR model. For more details, please refer to the PP-OCR technical article (https://arxiv.org/abs/2009.09941).
[2] On the basis of PP-OCR, PP-OCRv2 is further optimized in five aspects. The detection model adopts CML(Collaborative Mutual Learning) knowledge distillation strategy and CopyPaste data expansion strategy. The recognition model adopts LCNet lightweight backbone network, U-DML knowledge distillation strategy and enhanced CTC loss function improvement (as shown in the red box above), which further improves the inference speed and prediction effect. For more details, please refer to the technical report of PP-OCRv2 (https://arxiv.org/abs/2109.03144).
<a name="Visualization"></a>
## Visualization [more](./doc/doc_en/visualization_en.md)
- Chinese OCR model
<div align="center">
......@@ -198,19 +171,3 @@ More details, please refer to [Multilingual OCR Development Plan](https://github
<a name="LICENSE"></a>
## License
This project is released under <a href="https://github.com/PaddlePaddle/PaddleOCR/blob/master/LICENSE">Apache 2.0 license</a>
<a name="CONTRIBUTION"></a>
## Contribution
We welcome all the contributions to PaddleOCR and appreciate for your feedback very much.
- Many thanks to [Khanh Tran](https://github.com/xxxpsyduck) and [Karl Horky](https://github.com/karlhorky) for contributing and revising the English documentation.
- Many thanks to [zhangxin](https://github.com/ZhangXinNan) for contributing the new visualize function、add .gitignore and discard set PYTHONPATH manually.
- Many thanks to [lyl120117](https://github.com/lyl120117) for contributing the code for printing the network structure.
- Thanks [xiangyubo](https://github.com/xiangyubo) for contributing the handwritten Chinese OCR datasets.
- Thanks [authorfu](https://github.com/authorfu) for contributing Android demo and [xiadeye](https://github.com/xiadeye) contributing iOS demo, respectively.
- Thanks [BeyondYourself](https://github.com/BeyondYourself) for contributing many great suggestions and simplifying part of the code style.
- Thanks [tangmq](https://gitee.com/tangmq) for contributing Dockerized deployment services to PaddleOCR and supporting the rapid release of callable Restful API services.
- Thanks [lijinhan](https://github.com/lijinhan) for contributing a new way, i.e., java SpringBoot, to achieve the request for the Hubserving deployment.
- Thanks [Mejans](https://github.com/Mejans) for contributing the Occitan corpus and character set.
- Thanks [LKKlein](https://github.com/LKKlein) for contributing a new deploying package with the Golang program language.
- Thanks [Evezerest](https://github.com/Evezerest), [ninetailskim](https://github.com/ninetailskim), [edencfc](https://github.com/edencfc), [BeyondYourself](https://github.com/BeyondYourself) and [1084667371](https://github.com/1084667371) for contributing a new data annotation tool, i.e., PPOCRLabel。
README_ch.md
View file @ aa59fca5
......@@ -17,11 +17,18 @@
PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力开发者训练出更好的模型，并应用落地。
<div align="center">
<img src="./doc/imgs_results/ch_ppocr_mobile_v2.0/test_add_91.jpg" width="800">
</div>
<div align="center">
<img src="./doc/imgs_results/ch_ppocr_mobile_v2.0/00018069.jpg" width="800">
</div>
## 近期更新
- 2021.12.21《动手学OCR · 十讲》课程开讲，12月21日起每晚八点半线上授课！[免费报名地址](https://aistudio.baidu.com/aistudio/course/introduce/25207)。
- 2021.12.21 发布PaddleOCR v2.4。OCR算法新增1种文本检测算法（PSENet），3种文本识别算法（NRTR、SEED、SAR）；文档结构化算法新增1种关键信息提取算法（SDMGR，[文档](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.4/ppstructure/docs/kie.md)），3种DocVQA算法（LayoutLM、LayoutLMv2，LayoutXLM，[文档](https://github.com/PaddlePaddle/PaddleOCR/tree/release/2.4/ppstructure/vqa)）。
- PaddleOCR研发团队对最新发版内容技术深入解读，9月8日晚上20:15，[课程回放](https://aistudio.baidu.com/aistudio/education/group/info/6758)。
- 2021.9.7 发布PaddleOCR v2.3与[PP-OCRv2](#PP-OCRv2)，CPU推理速度相比于PP-OCR server提升220%；效果相比于PP-OCR mobile 提升7%。
- 2021.8.3 发布PaddleOCR v2.2，新增文档结构分析[PP-Structure](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.2/ppstructure/README_ch.md)工具包，支持版面分析与表格识别（含Excel导出）。
......@@ -29,41 +36,35 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
## 特性
- PP-OCR系列高质量预训练模型，准确的识别效果
- 超轻量PP-OCRv2系列：检测（3.1M）+ 方向分类器（1.4M）+ 识别（8.5M）= 13.0M
- 超轻量PP-OCR mobile移动端系列：检测（3.0M）+方向分类器（1.4M）+ 识别（5.0M）= 9.4M
- 通用PP-OCR server系列：检测（47.1M）+方向分类器（1.4M）+ 识别（94.9M）= 143.4M
- 支持中英文数字组合识别、竖排文本识别、长文本识别
- 支持多语言识别：韩语、日语、德语、法语等约80种语言
- PP-Structure文档结构化系统
- 支持版面分析与表格识别（含Excel导出）
- 支持关键信息提取任务
- 支持DocVQA任务
- 丰富易用的OCR相关工具组件
- 半自动数据标注工具PPOCRLabel：支持快速高效的数据标注
- 数据合成工具Style-Text：批量合成大量与目标场景类似的图像
- 支持用户自定义训练，提供丰富的预测推理部署方案
- 支持PIP快速安装使用
- 可运行于Linux、Windows、MacOS等多种系统
支持多种OCR相关前沿算法，在此基础上打造产业级特色模型[PP-OCR](./doc/doc_ch/ppocr_introduction.md)和[PP-Structure](./ppstructure/README_ch.md)，并打通数据生产、模型训练、压缩、预测部署全流程。
![](./doc/features.png)
> 上述内容的使用方法建议从文档教程中的快速开始体验
<a name="贡献代码"></a>
## 社区、社区贡献与社区常规赛
## 快速开始
- 在线网站体验：超轻量PP-OCR mobile模型体验地址：https://www.paddlepaddle.org.cn/hub/scene/ocr
- 移动端demo体验：[安装包DEMO下载地址](https://ai.baidu.com/easyedge/app/openSource?from=paddlelite)(基于EasyEdge和Paddle-Lite, 支持iOS和Android系统)
- 一行命令快速使用：[快速开始（中英文/多语言/文档分析）](./doc/doc_ch/quickstart.md)
<a name="电子书"></a>
## 《动手学OCR》电子书
- [《动手学OCR》电子书📚](./doc/doc_ch/ocr_book.md)
- 加入社区：微信扫描下方二维码加入官方交流群，与各行各业开发者充分交流，期待您的加入。
- 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。如果您的OCR项目未被收集在文档中，可根据文档说明与我们联系。
- 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)。
<a name="开源社区"></a>
## 开源社区
- **加入社区**👬：微信扫描下方二维码加入官方交流群，与各行各业开发者充分交流，期待您的加入。
- **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)。
<div align="center">
<img src="https://raw.githubusercontent.com/PaddlePaddle/PaddleOCR/dygraph/doc/joinus.PNG" width = "200" height = "200" />
</div>
## 零代码体验
- 在线网站体验：超轻量PP-OCR mobile模型体验地址：https://www.paddlepaddle.org.cn/hub/scene/ocr
- 移动端：[安装包DEMO下载地址](https://ai.baidu.com/easyedge/app/openSource?from=paddlelite)(基于EasyEdge和Paddle-Lite, 支持iOS和Android系统)
<a name="模型下载"></a>
## PP-OCR系列模型列表（更新中）
......@@ -74,73 +75,82 @@ 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-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)
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)
## 文档教程
- [运行环境准备](./doc/doc_ch/environment.md)
- [快速开始（中英文/多语言/版面分析）](./doc/doc_ch/quickstart.md)
- [PaddleOCR全景图与项目克隆](./doc/doc_ch/paddleOCR_overview.md)
- PP-OCR产业落地：从训练到部署
- [PP-OCR模型库](./doc/doc_ch/models.md)
- [PP-OCR模型下载](./doc/doc_ch/models_list.md)
- [Python引擎的PP-OCR模型库推理](./doc/doc_ch/inference_ppocr.md)
- [PP-OCR模型训练](./doc/doc_ch/training.md)
- [快速开始（中英文/多语言/文档分析）](./doc/doc_ch/quickstart.md)
- [PP-OCR文本检测识别🔥](./doc/doc_ch/ppocr_introduction.md)
- [快速开始](./doc/doc_ch/quickstart.md)
- [模型库](./doc/doc_ch/models_list.md)
- [模型训练](./doc/doc_ch/training.md)
- [文本检测](./doc/doc_ch/detection.md)
- [文本识别](./doc/doc_ch/recognition.md)
- [文本方向分类器](./doc/doc_ch/angle_class.md)
- [配置文件内容与生成](./doc/doc_ch/config.md)
- PP-OCR模型压缩
- [知识蒸馏](./doc/doc_ch/knowledge_distillation.md)
- 模型压缩
- [模型量化](./deploy/slim/quantization/README.md)
- [模型裁剪](./deploy/slim/prune/README.md)
- PP-OCR模型推理部署
- [知识蒸馏](./doc/doc_ch/knowledge_distillation.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C++预测引擎推理](./deploy/cpp_infer/readme.md)
- [服务化部署](./deploy/pdserving/README_CN.md)
- [端侧部署](./deploy/lite/readme.md)
- [Paddle2ONNX模型转化与预测](./deploy/paddle2onnx/readme.md)
- [Benchmark](./doc/doc_ch/benchmark.md)
- [PP-Structure信息提取](./ppstructure/README_ch.md)
- [版面分析](./ppstructure/layout/README_ch.md)
- [表格识别](./ppstructure/table/README_ch.md)
- [DocVQA](./ppstructure/vqa/README.md)
- [关键信息提取](./ppstructure/docs/kie.md)
- OCR学术圈
- [两阶段模型介绍与下载](./doc/doc_ch/algorithm_overview.md)
- [端到端PGNet算法](./doc/doc_ch/pgnet.md)
- [基于Python脚本预测引擎推理](./doc/doc_ch/inference.md)
- [PP-Structure文档分析🔥](./ppstructure/README_ch.md)
- [快速开始](./ppstructure/docs/quickstart.md)
- [模型库](./ppstructure/docs/models_list.md)
- [模型训练](./doc/doc_ch/training.md)
- [版面分析](./ppstructure/layout/README_ch.md)
- [表格识别](./ppstructure/table/README_ch.md)
- [关键信息提取](./ppstructure/docs/kie.md)
- [DocVQA](./ppstructure/vqa/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C++预测引擎推理]()
- [服务化部署](./deploy/pdserving/README_CN.md)
- [前沿算法与模型🚀](./doc/doc_ch/algorithm.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)
- [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)
- [使用PaddleOCR架构添加新算法](./doc/doc_ch/add_new_algorithm.md)
- [场景应用](./doc/doc_ch/application.md)
- [金融场景（表单/票据等）]()
- [工业场景（电表度数/车牌等）]()
- [教育场景（手写体/公式等）]()
- [医疗场景（化验单等）]()
- 数据标注与合成
- [半自动标注工具PPOCRLabel](./PPOCRLabel/README_ch.md)
- [数据合成工具Style-Text](./StyleText/README_ch.md)
- [其它数据标注工具](./doc/doc_ch/data_annotation.md)
- [其它数据合成工具](./doc/doc_ch/data_synthesis.md)
- 数据集
- [通用中英文OCR数据集](./doc/doc_ch/datasets.md)
- [手写中文OCR数据集](./doc/doc_ch/handwritten_datasets.md)
- [垂类多语言OCR数据集](./doc/doc_ch/vertical_and_multilingual_datasets.md)
- [通用中英文OCR数据集](doc/doc_ch/dataset/datasets.md)
- [手写中文OCR数据集](doc/doc_ch/dataset/handwritten_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)
- [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)
- [代码组织结构](./doc/doc_ch/tree.md)
- [效果展示](#效果展示)
- [《动手学OCR》电子书📚](./doc/doc_ch/ocr_book.md)
- [开源社区](#开源社区)
- FAQ
- [通用问题](./doc/doc_ch/FAQ.md)
- [PaddleOCR实战问题](./doc/doc_ch/FAQ.md)
- [参考文献](./doc/doc_ch/reference.md)
- [许可证书](#许可证书)
- [代码组织结构](./doc/doc_ch/tree.md)
<a name="PP-OCRv2"></a>
## PP-OCRv2 Pipeline
<div align="center">
<img src="./doc/ppocrv2_framework.jpg" width="800">
</div>
[1] PP-OCR是一个实用的超轻量OCR系统。主要由DB文本检测、检测框矫正和CRNN文本识别三部分组成。该系统从骨干网络选择和调整、预测头部的设计、数据增强、学习率变换策略、正则化参数选择、预训练模型使用以及模型自动裁剪量化8个方面，采用19个有效策略，对各个模块的模型进行效果调优和瘦身(如绿框所示)，最终得到整体大小为3.5M的超轻量中英文OCR和2.8M的英文数字OCR。更多细节请参考PP-OCR技术方案 https://arxiv.org/abs/2009.09941
[2] PP-OCRv2在PP-OCR的基础上，进一步在5个方面重点优化，检测模型采用CML协同互学习知识蒸馏策略和CopyPaste数据增广策略；识别模型采用LCNet轻量级骨干网络、UDML 改进知识蒸馏策略和[Enhanced CTC loss](./doc/doc_ch/enhanced_ctc_loss.md)损失函数改进（如上图红框所示），进一步在推理速度和预测效果上取得明显提升。更多细节请参考PP-OCRv2[技术报告](https://arxiv.org/abs/2109.03144)。
<a name="效果展示"></a>
## 效果展示 [more](./doc/doc_ch/visualization.md)
- 中文模型
<details open>
<summary>PP-OCRv2 中文模型</summary>
<div align="center">
<img src="doc/imgs_results/ch_ppocr_mobile_v2.0/test_add_91.jpg" width="800">
......@@ -151,16 +161,49 @@ PaddleOCR旨在打造一套丰富、领先、且实用的OCR工具库，助力
<img src="./doc/imgs_results/ch_ppocr_mobile_v2.0/rotate_00052204.jpg" width="800">
</div>
- 英文模型
</details>
<details open>
<summary>PP-OCRv2 英文模型</summary>
<div align="center">
<img src="./doc/imgs_results/ch_ppocr_mobile_v2.0/img_12.jpg" width="800">
</div>
- 其他语言模型
</details>
<details open>
<summary>PP-OCRv2 其他语言模型</summary>
<div align="center">
<img src="./doc/imgs_results/french_0.jpg" width="800">
<img src="./doc/imgs_results/korean.jpg" width="800">
</div>
</details>
<details open>
<summary>PP-Structure 文档分析</summary>
- 版面分析+表格识别
<div align="center">
<img src="./ppstructure/docs/table/ppstructure.GIF" width="800">
</div>
- SER（语义实体识别）
<div align="center">
<img src="./ppstructure/docs/vqa/result_ser/zh_val_0_ser.jpg" width="800">
</div>
- RE（关系提取）
<div align="center">
<img src="./ppstructure/docs/vqa/result_re/zh_val_21_re.jpg" width="800">
</div>
</details>
<a name="许可证书"></a>
## 许可证书
......
applications/多模态表单识别.md 0 → 100644
View file @ aa59fca5
# 1 项目说明
计算机视觉在金融领域的应用覆盖文字识别、图像识别、视频识别等，其中文字识别（OCR）是金融领域中的核心AI能力，其应用覆盖客户服务、风险防控、运营管理等各项业务，针对的对象包括通用卡证票据识别（银行卡、身份证、营业执照等）、通用文本表格识别（印刷体、多语言、手写体等）以及一些金融特色票据凭证。通过因此如果能够在结构化信息提取时同时利用文字、页面布局等信息，便可增强不同版式下的泛化性。
表单识别旨在识别各种具有表格性质的证件、房产证、营业执照、个人信息表、发票等关键键值对(如姓名-张三)，其广泛应用于银行、证券、公司财务等领域，具有很高的商业价值。本次范例项目开源了全流程表单识别方案，能够在多个场景快速实现迁移能力。表单识别通常存在以下难点：
- 人工摘录工作效率低；
- 国内常见表单版式多；
- 传统技术方案泛化效果不满足。
表单识别包含两大阶段：OCR阶段和文档视觉问答阶段。
其中，OCR阶段选取了PaddleOCR的PP-OCRv2模型，主要由文本检测和文本识别两个模块组成。DOC-VQA文档视觉问答阶段基于PaddleNLP自然语言处理算法库实现的LayoutXLM模型，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。本案例流程如 **图1** 所示：
<center><img src='https://ai-studio-static-online.cdn.bcebos.com/9bd844b970f94e5ba0bc0c5799bd819ea9b1861bb306471fabc2d628864d418e'></center>
<center>图1 多模态表单识别流程图</center>
注：欢迎再AIStudio领取免费算力体验线上实训，项目链接: [多模态表单识别](https://aistudio.baidu.com/aistudio/projectdetail/3815918)(配备Tesla V100、A100等高级算力资源)
# 2 安装说明
下载PaddleOCR源码，本项目中已经帮大家打包好的PaddleOCR(已经修改好配置文件)，无需下载解压即可，只需安装依赖环境~
```python
! unzip -q PaddleOCR.zip
```
```python
# 如仍需安装or安装更新，可以执行以下步骤
! git clone https://github.com/PaddlePaddle/PaddleOCR.git -b dygraph
# ! git clone https://gitee.com/PaddlePaddle/PaddleOCR
```
```python
# 安装依赖包
! pip install -U pip
! pip install -r /home/aistudio/PaddleOCR/requirements.txt
! pip install paddleocr
! pip install yacs gnureadline paddlenlp==2.2.1
! pip install xlsxwriter
```
# 3 数据准备
这里使用[XFUN数据集](https://github.com/doc-analysis/XFUND)做为实验数据集。 XFUN数据集是微软提出的一个用于KIE任务的多语言数据集，共包含七个数据集，每个数据集包含149张训练集和50张验证集
分别为：ZH(中文)、JA(日语)、ES(西班牙)、FR(法语)、IT(意大利)、DE(德语)、PT(葡萄牙)
本次实验选取中文数据集作为我们的演示数据集。法语数据集作为实践课程的数据集，数据集样例图如 **图2** 所示。
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/0f84137778cd4ab6899c64109d452290e9c678ccf01744978bc9c0647adbba45" width="1000" ></center>
<center>图2 数据集样例，左中文，右法语</center>
## 3.1 下载处理好的数据集
处理好的XFUND中文数据集下载地址：[https://paddleocr.bj.bcebos.com/dataset/XFUND.tar](https://paddleocr.bj.bcebos.com/dataset/XFUND.tar) ,可以运行如下指令完成中文数据集下载和解压。
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/31e3dbee31d441d2a36d45b5af660e832dfa2f437f4d49a1914312a15b6a29a7"></center>
<center>图3 下载数据集</center>
```python
! wget https://paddleocr.bj.bcebos.com/dataset/XFUND.tar
! tar -xf XFUND.tar
# XFUN其他数据集使用下面的代码进行转换
# 代码链接：https://github.com/PaddlePaddle/PaddleOCR/blob/release%2F2.4/ppstructure/vqa/helper/trans_xfun_data.py
# %cd PaddleOCR
# !python3 ppstructure/vqa/tools/trans_xfun_data.py --ori_gt_path=path/to/json_path --output_path=path/to/save_path
# %cd ../
```
运行上述指令后在 /home/aistudio/PaddleOCR/ppstructure/vqa/XFUND 目录下有2个文件夹，目录结构如下所示：
```bash
/home/aistudio/PaddleOCR/ppstructure/vqa/XFUND
└─ zh_train/ 训练集
├── image/ 图片存放文件夹
├── xfun_normalize_train.json 标注信息
└─ zh_val/ 验证集
├── image/ 图片存放文件夹
├── xfun_normalize_val.json 标注信息
```
该数据集的标注格式为
```bash
{
"height": 3508, # 图像高度
"width": 2480, # 图像宽度
"ocr_info": [
{
"text": "邮政地址:", # 单个文本内容
"label": "question", # 文本所属类别
"bbox": [261, 802, 483, 859], # 单个文本框
"id": 54, # 文本索引
"linking": [[54, 60]], # 当前文本和其他文本的关系 [question, answer]
"words": []
},
{
"text": "湖南省怀化市市辖区",
"label": "answer",
"bbox": [487, 810, 862, 859],
"id": 60,
"linking": [[54, 60]],
"words": []
}
]
}
```
## 3.2 转换为PaddleOCR检测和识别格式
使用XFUND训练PaddleOCR检测和识别模型，需要将数据集格式改为训练需求的格式。
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/9a709f19e7174725a8cfb09fd922ade74f8e9eb73ae1438596cbb2facef9c24a"></center>
<center>图4 转换为OCR格式</center>
- **文本检测** 标注文件格式如下，中间用'\t'分隔：
" 图像文件名 json.dumps编码的图像标注信息"
ch4_test_images/img_61.jpg [{"transcription": "MASA", "points": [[310, 104], [416, 141], [418, 216], [312, 179]]}, {...}]
json.dumps编码前的图像标注信息是包含多个字典的list，字典中的 `points` 表示文本框的四个点的坐标(x, y)，从左上角的点开始顺时针排列。 `transcription` 表示当前文本框的文字，***当其内容为“###”时，表示该文本框无效，在训练时会跳过。***
- **文本识别** 标注文件的格式如下， txt文件中默认请将图片路径和图片标签用'\t'分割，如用其他方式分割将造成训练报错。
```
" 图像文件名 图像标注信息 "
train_data/rec/train/word_001.jpg 简单可依赖
train_data/rec/train/word_002.jpg 用科技让复杂的世界更简单
...
```
```python
! unzip -q /home/aistudio/data/data140302/XFUND_ori.zip -d /home/aistudio/data/data140302/
```
已经提供转换脚本，执行如下代码即可转换成功：
```python
%cd /home/aistudio/
! python trans_xfund_data.py
```
# 4 OCR
选用飞桨OCR开发套件[PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/README_ch.md)中的PP-OCRv2模型进行文本检测和识别。PP-OCRv2在PP-OCR的基础上，进一步在5个方面重点优化，检测模型采用CML协同互学习知识蒸馏策略和CopyPaste数据增广策略；识别模型采用LCNet轻量级骨干网络、UDML 改进知识蒸馏策略和[Enhanced CTC loss](https://github.com/PaddlePaddle/PaddleOCR/blob/dygraph/doc/doc_ch/enhanced_ctc_loss.md)损失函数改进，进一步在推理速度和预测效果上取得明显提升。更多细节请参考PP-OCRv2[技术报告](https://arxiv.org/abs/2109.03144)。
## 4.1 文本检测
我们使用2种方案进行训练、评估：
- **PP-OCRv2中英文超轻量检测预训练模型**
- **XFUND数据集+fine-tune**
### **4.1.1 方案1：预训练模型**
**1）下载预训练模型**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/2aff41ee8fce4e9bac8295cc00720217bde2aeee7ee7473689848bed0b6fde05"></center>
<center>图5 文本检测方案1-下载预训练模型</center>
PaddleOCR已经提供了PP-OCR系列模型，部分模型展示如下表所示：
| 模型简介 | 模型名称 | 推荐场景 | 检测模型 | 方向分类器 | 识别模型 |
| ------------------------------------- | ----------------------- | --------------- | ------------------------------------------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------ |
| 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)
这里我们使用PP-OCRv2中英文超轻量检测模型，下载并解压预训练模型：
```python
%cd /home/aistudio/PaddleOCR/pretrain/
! wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_distill_train.tar
! tar -xf ch_PP-OCRv2_det_distill_train.tar && rm -rf ch_PP-OCRv2_det_distill_train.tar
% cd ..
```
**2）模型评估**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/75b0e977dfb74a83851f8828460759f337b1b7a0c33c47a08a30f3570e1e2e74"></center>
<center>图6 文本检测方案1-模型评估</center>
接着使用下载的超轻量检测模型在XFUND验证集上进行评估，由于蒸馏需要包含多个网络，甚至多个Student网络，在计算指标的时候只需要计算一个Student网络的指标即可，key字段设置为Student则表示只计算Student网络的精度。
```
Metric:
name: DistillationMetric
base_metric_name: DetMetric
main_indicator: hmean
key: "Student"
```
首先修改配置文件`configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_distill.yml`中的以下字段：
```
Eval.dataset.data_dir：指向验证集图片存放目录
Eval.dataset.label_file_list：指向验证集标注文件
```
然后在XFUND验证集上进行评估，具体代码如下：
```python
%cd /home/aistudio/PaddleOCR
! python tools/eval.py \
-c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_distill.yml \
-o Global.checkpoints="./pretrain_models/ch_PP-OCRv2_det_distill_train/best_accuracy"
```
使用预训练模型进行评估，指标如下所示：
| 方案 | hmeans |
| -------- | -------- |
| PP-OCRv2中英文超轻量检测预训练模型 | 77.26% |
使用文本检测预训练模型在XFUND验证集上评估，达到77%左右，充分说明ppocr提供的预训练模型有一定的泛化能力。
### **4.1.2 方案2：XFUND数据集+fine-tune**
PaddleOCR提供的蒸馏预训练模型包含了多个模型的参数，我们提取Student模型的参数，在XFUND数据集上进行finetune，可以参考如下代码：
```python
import paddle
# 加载预训练模型
all_params = paddle.load("pretrain/ch_PP-OCRv2_det_distill_train/best_accuracy.pdparams")
# 查看权重参数的keys
# print(all_params.keys())
# 学生模型的权重提取
s_params = {key[len("student_model."):]: all_params[key] for key in all_params if "student_model." in key}
# 查看学生模型权重参数的keys
print(s_params.keys())
# 保存
paddle.save(s_params, "pretrain/ch_PP-OCRv2_det_distill_train/student.pdparams")
```
**1)模型训练**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/560c44b8dd604da7987bd25da0a882156ffcfb7f6bcb44108fe9bde77512e572"></center>
<center>图7 文本检测方案2-模型训练</center>
修改配置文件`configs/det/ch_PP-OCRv2_det_student.yml`中的以下字段：
```
Global.pretrained_model：指向预训练模型路径
Train.dataset.data_dir：指向训练集图片存放目录
Train.dataset.label_file_list：指向训练集标注文件
Eval.dataset.data_dir：指向验证集图片存放目录
Eval.dataset.label_file_list：指向验证集标注文件
Optimizer.lr.learning_rate：调整学习率，本实验设置为0.005
Train.dataset.transforms.EastRandomCropData.size：训练尺寸改为[1600, 1600]
Eval.dataset.transforms.DetResizeForTest：评估尺寸，添加如下参数
limit_side_len: 1600
limit_type: 'min'
```
执行下面命令启动训练：
```python
! CUDA_VISIBLE_DEVICES=0 python tools/train.py \
-c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_student.yml
```
**2）模型评估**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/5a75137c5f924dfeb6956b5818812298cc3dc7992ac84954b4175be9adf83c77"></center>
<center>图8 文本检测方案2-模型评估</center>
使用训练好的模型进行评估，更新模型路径`Global.checkpoints`，这里为大家提供训练好的模型`./pretrain/ch_db_mv3-student1600-finetune/best_accuracy`
```python
%cd /home/aistudio/PaddleOCR/
! python tools/eval.py \
-c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_student.yml \
-o Global.checkpoints="pretrain/ch_db_mv3-student1600-finetune/best_accuracy"
```
同时我们提供了未finetuen的模型，配置文件参数(`pretrained_model`设置为空，`learning_rate` 设置为0.001)
```python
%cd /home/aistudio/PaddleOCR/
! python tools/eval.py \
-c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_student.yml \
-o Global.checkpoints="pretrain/ch_db_mv3-student1600/best_accuracy"
```
使用训练好的模型进行评估，指标如下所示：
| 方案 | hmeans |
| -------- | -------- |
| XFUND数据集 | 79.27% |
| XFUND数据集+fine-tune | 85.24% |
对比仅使用XFUND数据集训练的模型，使用XFUND数据集+finetune训练，在验证集上评估达到85%左右，说明 finetune会提升垂类场景效果。
**3）导出模型**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/07c3b060c54e4b00be7de8d41a8a4696ff53835343cc4981aab0555183306e79"></center>
<center>图9 文本检测方案2-模型导出</center>
在模型训练过程中保存的模型文件是包含前向预测和反向传播的过程，在实际的工业部署则不需要反向传播，因此需要将模型进行导成部署需要的模型格式。 执行下面命令，即可导出模型。
```python
# 加载配置文件`ch_PP-OCRv2_det_student.yml`，从`pretrain/ch_db_mv3-student1600-finetune`目录下加载`best_accuracy`模型
# inference模型保存在`./output/det_db_inference`目录下
%cd /home/aistudio/PaddleOCR/
! python tools/export_model.py \
-c configs/det/ch_PP-OCRv2/ch_PP-OCRv2_det_student.yml \
-o Global.pretrained_model="pretrain/ch_db_mv3-student1600-finetune/best_accuracy" \
Global.save_inference_dir="./output/det_db_inference/"
```
转换成功后，在目录下有三个文件：
```
/inference/rec_crnn/
├── inference.pdiparams # 识别inference模型的参数文件
├── inference.pdiparams.info # 识别inference模型的参数信息，可忽略
└── inference.pdmodel # 识别inference模型的program文件
```
**4）模型预测**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/0d582de9aa46474791e08654f84a614a6510e98bfe5f4ad3a26501cbf49ec151"></center>
<center>图10 文本检测方案2-模型预测</center>
加载上面导出的模型，执行如下命令对验证集或测试集图片进行预测：
```
det_model_dir：预测模型
image_dir：测试图片路径
use_gpu：是否使用GPU
```
检测可视化结果保存在`/home/aistudio/inference_results/`目录下，查看检测效果。
```python
%pwd
!python tools/infer/predict_det.py \
--det_algorithm="DB" \
--det_model_dir="./output/det_db_inference/" \
--image_dir="./doc/vqa/input/zh_val_21.jpg" \
--use_gpu=True
```
总结，我们分别使用PP-OCRv2中英文超轻量检测预训练模型、XFUND数据集+finetune2种方案进行评估、训练等，指标对比如下：
| 方案 | hmeans | 结果分析 |
| -------- | -------- | -------- |
| PP-OCRv2中英文超轻量检测预训练模型 | 77.26% | ppocr提供的预训练模型有一定的泛化能力 |
| XFUND数据集 | 79.27% | |
| XFUND数据集+finetune | 85.24% | finetune会提升垂类场景效果 |
## 4.2 文本识别
我们分别使用如下3种方案进行训练、评估：
- PP-OCRv2中英文超轻量识别预训练模型
- XFUND数据集+fine-tune
- XFUND数据集+fine-tune+真实通用识别数据
### **4.2.1 方案1：预训练模型**
**1）下载预训练模型**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/b7230e9964074181837e1132029f9da8178bf564ac5c43a9a93a30e975c0d8b4"></center>
<center>图11 文本识别方案1-下载预训练模型</center>
我们使用PP-OCRv2中英文超轻量文本识别模型，下载并解压预训练模型：
```python
%cd /home/aistudio/PaddleOCR/pretrain/
! wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_train.tar
! tar -xf ch_PP-OCRv2_rec_train.tar && rm -rf ch_PP-OCRv2_rec_train.tar
% cd ..
```
**2）模型评估**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/166ce56d634c4c7589fe68fbc6e7ae663305dcc82ba144c781507341ffae7fe8"></center>
<center>图12 文本识别方案1-模型评估</center>
首先修改配置文件`configs/det/ch_PP-OCRv2/ch_PP-OCRv2_rec_distillation.yml`中的以下字段：
```
Eval.dataset.data_dir：指向验证集图片存放目录
Eval.dataset.label_file_list：指向验证集标注文件
```
我们使用下载的预训练模型进行评估：
```python
%cd /home/aistudio/PaddleOCR
! CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
-c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec_distillation.yml \
-o Global.checkpoints=./pretrain/ch_PP-OCRv2_rec_train/best_accuracy
```
使用预训练模型进行评估，指标如下所示：
| 方案 | acc |
| -------- | -------- |
| PP-OCRv2中英文超轻量识别预训练模型 | 67.48% |
使用文本预训练模型在XFUND验证集上评估，acc达到67%左右，充分说明ppocr提供的预训练模型有一定的泛化能力。
### **4.2.2 方案2：XFUND数据集+finetune**
同检测模型，我们提取Student模型的参数，在XFUND数据集上进行finetune，可以参考如下代码：
```python
import paddle
# 加载预训练模型
all_params = paddle.load("pretrain/ch_PP-OCRv2_rec_train/best_accuracy.pdparams")
# 查看权重参数的keys
print(all_params.keys())
# 学生模型的权重提取
s_params = {key[len("Student."):]: all_params[key] for key in all_params if "Student." in key}
# 查看学生模型权重参数的keys
print(s_params.keys())
# 保存
paddle.save(s_params, "pretrain/ch_PP-OCRv2_rec_train/student.pdparams")
```
**1)模型训练**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/06dad690219b42a59a27c84a060af1436bcd05de10b843209c6270e04e4dda10"></center>
<center>图13 文本识别方案2-模型训练</center>
修改配置文件`configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml`中的以下字段：
```
Global.pretrained_model：指向预训练模型路径
Global.character_dict_path: 字典路径
Optimizer.lr.values：学习率
Train.dataset.data_dir：指向训练集图片存放目录
Train.dataset.label_file_list：指向训练集标注文件
Eval.dataset.data_dir：指向验证集图片存放目录
Eval.dataset.label_file_list：指向验证集标注文件
```
执行如下命令启动训练：
```python
%cd /home/aistudio/PaddleOCR/
! CUDA_VISIBLE_DEVICES=0 python tools/train.py \
-c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml
```
**2）模型评估**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/c07c88f708ad43cc8cd615861626d0e8333c0e3d4dda49ac8cba1f8939fa8a94"></center>
<center>图14 文本识别方案2-模型评估</center>
使用训练好的模型进行评估，更新模型路径`Global.checkpoints`，这里为大家提供训练好的模型`./pretrain/rec_mobile_pp-OCRv2-student-finetune/best_accuracy`
```python
%cd /home/aistudio/PaddleOCR/
! CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
-c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml \
-o Global.checkpoints=./pretrain/rec_mobile_pp-OCRv2-student-finetune/best_accuracy
```
使用预训练模型进行评估，指标如下所示：
| 方案 | acc |
| -------- | -------- |
| XFUND数据集+finetune | 72.33% |
使用XFUND数据集+finetune训练，在验证集上评估达到72%左右，说明 finetune会提升垂类场景效果。
### **4.2.3 方案3：XFUND数据集+finetune+真实通用识别数据**
接着我们在上述`XFUND数据集+finetune`实验的基础上，添加真实通用识别数据，进一步提升识别效果。首先准备真实通用识别数据，并上传到AIStudio：
**1)模型训练**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/45f288ce8b2c45d8aa5407785b4b40f4876fc3da23744bd7a78060797fba0190"></center>
<center>图15 文本识别方案3-模型训练</center>
在上述`XFUND数据集+finetune`实验中修改配置文件`configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml`的基础上，继续修改以下字段：
```
Train.dataset.label_file_list：指向真实识别训练集图片存放目录
Train.dataset.ratio_list：动态采样
```
执行如下命令启动训练：
```python
%cd /home/aistudio/PaddleOCR/
! CUDA_VISIBLE_DEVICES=0 python tools/train.py \
-c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml
```
**2）模型评估**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/965db9f758614c6f9be301286cd5918f21110603c8aa4a1dbf5371e3afeec782"></center>
<center>图16 文本识别方案3-模型评估</center>
使用训练好的模型进行评估，更新模型路径`Global.checkpoints`，这里为大家提供训练好的模型`./pretrain/rec_mobile_pp-OCRv2-student-readldata/best_accuracy`
```python
! CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
-c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml \
-o Global.checkpoints=./pretrain/rec_mobile_pp-OCRv2-student-realdata/best_accuracy
```
使用预训练模型进行评估，指标如下所示：
| 方案 | acc |
| -------- | -------- |
| XFUND数据集+fine-tune+真实通用识别数据 | 85.29% |
使用XFUND数据集+finetune训练，在验证集上评估达到85%左右，说明真实通用识别数据对于性能提升很有帮助。
**3)导出模型**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/3dc7f69fac174cde96b9d08b5e2353a1d88dc63e7be9410894c0783660b35b76"></center>
<center>图17 文本识别方案3-导出模型</center>
导出模型只保留前向预测的过程：
```python
!python tools/export_model.py \
-c configs/rec/ch_PP-OCRv2/ch_PP-OCRv2_rec.yml \
-o Global.pretrained_model=pretrain/rec_mobile_pp-OCRv2-student-realdata/best_accuracy \
Global.save_inference_dir=./output/rec_crnn_inference/
```
**4)模型预测**
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/60b95b4945954f81a080a8f308cee66f83146479cd1142b9b6b1290938fd1df8"></center>
<center>图18 文本识别方案3-模型预测</center>
加载上面导出的模型，执行如下命令对验证集或测试集图片进行预测，检测可视化结果保存在`/home/aistudio/inference_results/`目录下，查看检测、识别效果。需要通过`--rec_char_dict_path`指定使用的字典路径
```python
! python tools/infer/predict_system.py \
--image_dir="./doc/vqa/input/zh_val_21.jpg" \
--det_model_dir="./output/det_db_inference/" \
--rec_model_dir="./output/rec_crnn_inference/" \
--rec_image_shape="3, 32, 320" \
--rec_char_dict_path="/home/aistudio/XFUND/word_dict.txt"
```
总结，我们分别使用PP-OCRv2中英文超轻量检测预训练模型、XFUND数据集+finetune2种方案进行评估、训练等，指标对比如下：
| 方案 | acc | 结果分析 |
| -------- | -------- | -------- |
| PP-OCRv2中英文超轻量识别预训练模型 | 67.48% | ppocr提供的预训练模型有一定的泛化能力 |
| XFUND数据集+fine-tune |72.33% | finetune会提升垂类场景效果 |
| XFUND数据集+fine-tune+真实通用识别数据 | 85.29% | 真实通用识别数据对于性能提升很有帮助 |
# 5 文档视觉问答(DOC-VQA)
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。
PaddleOCR中DOC-VQA系列算法基于PaddleNLP自然语言处理算法库实现LayoutXLM论文，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。
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。
```python
%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 ../
```
## 5.1 SER
SER: 语义实体识别 (Semantic Entity Recognition）, 可以完成对图像中的文本识别与分类。
<center><img src='https://ai-studio-static-online.cdn.bcebos.com/a3b25766f3074d2facdf88d4a60fc76612f51992fd124cf5bd846b213130665b' width='700'></center>
<center>图19 SER测试效果图</center>
**图19** 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。
#### 5.1.1 模型训练
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/2e45f297c9d44ca5b8718ae100a365f7348eaeed4cb8495b904f28a9c8075d8a"></center>
<center>图20 SER-模型训练</center>
启动训练之前，需要修改配置文件 `configs/vqa/ser/layoutxlm.yml` 以下四个字段：
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
```python
%cd /home/aistudio/PaddleOCR/
! CUDA_VISIBLE_DEVICES=0 python 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。
#### 5.1.2 模型评估
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/5df160ac39ee4d9e92a937094bc53a737272f9f2abeb4ddfaebb48e8eccf1be2"></center>
<center>图21 SER-模型评估</center>
我们使用下载的预训练模型进行评估，如果使用自己训练好的模型进行评估，将待评估的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段即可。
```python
! CUDA_VISIBLE_DEVICES=0 python tools/eval.py \
-c configs/vqa/ser/layoutxlm.yml \
-o Architecture.Backbone.checkpoints=pretrain/ser_LayoutXLM_xfun_zh/
```
最终会打印出`precision`, `recall`, `hmean`等指标，预训练模型评估指标如下：
<center><img src='https://ai-studio-static-online.cdn.bcebos.com/2854aee557a74079a82dd5cd57e48bc2ce97974d5637477fb4deea137d0e312c' width='700'></center>
<center>图 SER预训练模型评估指标</center>
#### 5.1.3 模型预测
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/0f7d50a0fb924b408b93e1fbd6ca64148eed34a2e6724280acd3e113fef7dc48"></center>
<center>图22 SER-模型预测</center>
使用如下命令即可完成`OCR引擎 + SER`的串联预测, 以SER预训练模型为例:
```python
! CUDA_VISIBLE_DEVICES=0 python 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`。通过如下命令查看预测图片：
```python
import cv2
from matplotlib import pyplot as plt
# 在notebook中使用matplotlib.pyplot绘图时，需要添加该命令进行显示
%matplotlib inline
img = cv2.imread('output/ser/zh_val_42_ser.jpg')
plt.figure(figsize=(48,24))
plt.imshow(img)
```
## 5.2 RE
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)。
<center><img src='https://ai-studio-static-online.cdn.bcebos.com/4de19ca3e54343e88961e816cad28bbacdc807f40b9440be914d871b0a914570' width='700'></center>
<center>图23 RE预测效果图</center>
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。
#### 5.2.1 模型训练
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/268c707a62c54e93958d2b2ab29e0932953aad41819e44aaaaa05c8ad85c6491"></center>
<center>图24 RE-模型训练</center>
启动训练之前，需要修改配置文件`configs/vqa/re/layoutxlm.yml`中的以下四个字段
Train.dataset.data_dir：指向训练集图片存放目录
Train.dataset.label_file_list：指向训练集标注文件
Eval.dataset.data_dir：指指向验证集图片存放目录
Eval.dataset.label_file_list：指向验证集标注文件
```python
! 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
#### 5.2.2 模型评估
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/93c66a43a69e472899c1c6732408b7a42e99a43721e94e9ca3c0a64e080306e4"></center>
<center>图25 RE-模型评估</center>
我们使用下载的预训练模型进行评估，如果使用自己训练好的模型进行评估，将待评估的模型所在文件夹路径赋值给 `Architecture.Backbone.checkpoints` 字段即可。
```python
! CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py \
-c configs/vqa/re/layoutxlm.yml \
-o Architecture.Backbone.checkpoints=pretrain/re_LayoutXLM_xfun_zh/
```
最终会打印出`precision`, `recall`, `hmean`等指标，预训练模型评估指标如下：
<center><img src='https://ai-studio-static-online.cdn.bcebos.com/f99af54fb2d14691a73b1a748e0ca22618aeddfded0c4da58bbbb03edb8c2340' width='700'></center>
<center>图 RE预训练模型评估指标</center>
#### 5.2.3 模型预测
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/bab32d32bdec4339b9a3e5f911e4b41f77996f3faabc40bd8309b5b20cad31e4"></center>
<center>图26 RE-模型预测</center>
使用OCR引擎 + SER + RE串联预测
使用如下命令即可完成OCR引擎 + SER + RE的串联预测, 以预训练SER和RE模型为例：
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。
```python
%cd /home/aistudio/PaddleOCR
! 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=test_imgs/ \
-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, 每一行表示一张图片的结果，每张图片的结果如下所示，前面表示测试图片路径，后面为测试结果：key字段及对应的value字段。
```
test_imgs/t131.jpg {"政治面税": "群众", "性别": "男", "籍贯": "河北省邯郸市", "婚姻状况": "亏末婚口已婚口已娇", "通讯地址": "邯郸市阳光苑7号楼003", "民族": "汉族", "毕业院校": "河南工业大学", "户口性质": "口农村城镇", "户口地址": "河北省邯郸市", "联系电话": "13288888888", "健康状况": "健康", "姓名": "小六", "好高cm": "180", "出生年月": "1996年8月9日", "文化程度": "本科", "身份证号码": "458933777777777777"}
````
```python
# 展示预测结果
import cv2
from matplotlib import pyplot as plt
%matplotlib inline
img = cv2.imread('./output/re/t131_ser.jpg')
plt.figure(figsize=(48,24))
plt.imshow(img)
```
# 6 导出Excel
<center><img src="https://ai-studio-static-online.cdn.bcebos.com/ab93d3d90d77437a81c9534b2dd1d3e39ef81e8473054fd3aeff6e837ebfb827"></center>
<center>图27 导出Excel</center>
为了输出信息匹配对，我们修改`tools/infer_vqa_token_ser_re.py`文件中的`line 194-197`。
```
fout.write(img_path + "\t" + json.dumps(
{
"ser_resule": result,
}, ensure_ascii=False) + "\n")
```
更改为
```
result_key = {}
for ocr_info_head, ocr_info_tail in result:
result_key[ocr_info_head['text']] = ocr_info_tail['text']
fout.write(img_path + "\t" + json.dumps(
result_key, ensure_ascii=False) + "\n")
```
同时将输出结果导出到Excel中，效果如 图28 所示：
<center><img src='https://ai-studio-static-online.cdn.bcebos.com/9f45d3eef75e4842a0828bb9e518c2438300264aec0646cc9addfce860a04196' width='700'></center>
<center>图28 Excel效果图</center>
```python
import json
import xlsxwriter as xw
workbook = xw.Workbook('output/re/infer_results.xlsx')
format1 = workbook.add_format({
'align': 'center',
'valign': 'vcenter',
'text_wrap': True,
})
worksheet1 = workbook.add_worksheet('sheet1')
worksheet1.activate()
title = ['姓名', '性别', '民族', '文化程度', '身份证号码', '联系电话', '通讯地址']
worksheet1.write_row('A1', title)
i = 2
with open('output/re/infer_results.txt', 'r', encoding='utf-8') as fin:
lines = fin.readlines()
for line in lines:
img_path, result = line.strip().split('\t')
result_key = json.loads(result)
# 写入Excel
row_data = [result_key['姓名'], result_key['性别'], result_key['民族'], result_key['文化程度'], result_key['身份证号码'],
result_key['联系电话'], result_key['通讯地址']]
row = 'A' + str(i)
worksheet1.write_row(row, row_data, format1)
i+=1
workbook.close()
```
# 更多资源
- 更多深度学习知识、产业案例、面试宝典等，请参考：[awesome-DeepLearning](https://github.com/paddlepaddle/awesome-DeepLearning)
- 更多PaddleOCR使用教程，请参考：[PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR/tree/dygraph)
- 更多PaddleNLP使用教程，请参考：[PaddleNLP](https://github.com/PaddlePaddle/PaddleNLP)
- 飞桨框架相关资料，请参考：[飞桨深度学习平台](https://www.paddlepaddle.org.cn/?fr=paddleEdu_aistudio)
# 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
configs/cls/ch_PP-OCRv3/ch_PP-OCRv3_rotnet.yml 0 → 100644
View file @ aa59fca5
Global:
debug: false
use_gpu: true
epoch_num: 100
log_smooth_window: 20
print_batch_step: 10
save_model_dir: ./output/rec_ppocr_v3_rotnet
save_epoch_step: 3
eval_batch_step: [0, 2000]
cal_metric_during_train: true
pretrained_model: null
checkpoints: null
save_inference_dir: null
use_visualdl: false
infer_img: doc/imgs_words/ch/word_1.jpg
character_dict_path: ppocr/utils/ppocr_keys_v1.txt
max_text_length: 25
infer_mode: false
use_space_char: true
save_res_path: ./output/rec/predicts_chinese_lite_v2.0.txt
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.001
regularizer:
name: L2
factor: 1.0e-05
Architecture:
model_type: cls
algorithm: CLS
Transform: null
Backbone:
name: MobileNetV1Enhance
scale: 0.5
last_conv_stride: [1, 2]
last_pool_type: avg
Neck:
Head:
name: ClsHead
class_dim: 4
Loss:
name: ClsLoss
main_indicator: acc
PostProcess:
name: ClsPostProcess
Metric:
name: ClsMetric
main_indicator: acc
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data
label_file_list:
- ./train_data/train_list.txt
transforms:
- DecodeImage:
img_mode: BGR
channel_first: false
- RecAug:
use_tia: False
- RandAugment:
- SSLRotateResize:
image_shape: [3, 48, 320]
- KeepKeys:
keep_keys: ["image", "label"]
loader:
collate_fn: "SSLRotateCollate"
shuffle: true
batch_size_per_card: 32
drop_last: true
num_workers: 8
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data
label_file_list:
- ./train_data/val_list.txt
transforms:
- DecodeImage:
img_mode: BGR
channel_first: false
- SSLRotateResize:
image_shape: [3, 48, 320]
- KeepKeys:
keep_keys: ["image", "label"]
loader:
collate_fn: "SSLRotateCollate"
shuffle: false
drop_last: false
batch_size_per_card: 64
num_workers: 8
profiler_options: null
configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml 0 → 100644
View file @ aa59fca5
Global:
debug: false
use_gpu: true
epoch_num: 500
log_smooth_window: 20
print_batch_step: 10
save_model_dir: ./output/ch_PP-OCR_v3_det/
save_epoch_step: 100
eval_batch_step:
- 0
- 400
cal_metric_during_train: false
pretrained_model: null
checkpoints: null
save_inference_dir: null
use_visualdl: false
infer_img: doc/imgs_en/img_10.jpg
save_res_path: ./checkpoints/det_db/predicts_db.txt
distributed: true
Architecture:
name: DistillationModel
algorithm: Distillation
model_type: det
Models:
Student:
model_type: det
algorithm: DB
Transform: null
Backbone:
name: MobileNetV3
scale: 0.5
model_name: large
disable_se: true
Neck:
name: RSEFPN
out_channels: 96
shortcut: True
Head:
name: DBHead
k: 50
Student2:
model_type: det
algorithm: DB
Transform: null
Backbone:
name: MobileNetV3
scale: 0.5
model_name: large
disable_se: true
Neck:
name: RSEFPN
out_channels: 96
shortcut: True
Head:
name: DBHead
k: 50
Teacher:
freeze_params: true
return_all_feats: false
model_type: det
algorithm: DB
Backbone:
name: ResNet
in_channels: 3
layers: 50
Neck:
name: LKPAN
out_channels: 256
Head:
name: DBHead
kernel_list: [7,2,2]
k: 50
Loss:
name: CombinedLoss
loss_config_list:
- DistillationDilaDBLoss:
weight: 1.0
model_name_pairs:
- ["Student", "Teacher"]
- ["Student2", "Teacher"]
key: maps
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
- DistillationDMLLoss:
model_name_pairs:
- ["Student", "Student2"]
maps_name: "thrink_maps"
weight: 1.0
# act: None
model_name_pairs: ["Student", "Student2"]
key: maps
- DistillationDBLoss:
weight: 1.0
model_name_list: ["Student", "Student2"]
# key: maps
# name: DBLoss
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.001
warmup_epoch: 2
regularizer:
name: L2
factor: 5.0e-05
PostProcess:
name: DistillationDBPostProcess
model_name: ["Student"]
key: head_out
thresh: 0.3
box_thresh: 0.6
max_candidates: 1000
unclip_ratio: 1.5
Metric:
name: DistillationMetric
base_metric_name: DetMetric
main_indicator: hmean
key: "Student"
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/icdar2015/text_localization/
label_file_list:
- ./train_data/icdar2015/text_localization/train_icdar2015_label.txt
ratio_list: [1.0]
transforms:
- DecodeImage:
img_mode: BGR
channel_first: false
- DetLabelEncode: null
- CopyPaste:
- IaaAugment:
augmenter_args:
- type: Fliplr
args:
p: 0.5
- type: Affine
args:
rotate:
- -10
- 10
- type: Resize
args:
size:
- 0.5
- 3
- EastRandomCropData:
size:
- 960
- 960
max_tries: 50
keep_ratio: true
- MakeBorderMap:
shrink_ratio: 0.4
thresh_min: 0.3
thresh_max: 0.7
- MakeShrinkMap:
shrink_ratio: 0.4
min_text_size: 8
- NormalizeImage:
scale: 1./255.
mean:
- 0.485
- 0.456
- 0.406
std:
- 0.229
- 0.224
- 0.225
order: hwc
- ToCHWImage: null
- KeepKeys:
keep_keys:
- image
- threshold_map
- threshold_mask
- shrink_map
- shrink_mask
loader:
shuffle: true
drop_last: false
batch_size_per_card: 8
num_workers: 4
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data/icdar2015/text_localization/
label_file_list:
- ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
transforms:
- DecodeImage:
img_mode: BGR
channel_first: false
- DetLabelEncode: null
- DetResizeForTest: null
- NormalizeImage:
scale: 1./255.
mean:
- 0.485
- 0.456
- 0.406
std:
- 0.229
- 0.224
- 0.225
order: hwc
- ToCHWImage: null
- KeepKeys:
keep_keys:
- image
- shape
- polys
- ignore_tags
loader:
shuffle: false
drop_last: false
batch_size_per_card: 1
num_workers: 2
test_tipc/configs/ch_ppocr_mobile_v2.0_det/det_mv3_db.yml → configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_student.yml
View file @ aa59fca5
Global:
use_gpu: false
epoch_num: 5
debug: false
use_gpu: true
epoch_num: 500
log_smooth_window: 20
print_batch_step: 1
save_model_dir: ./output/db_mv3/
save_epoch_step: 1200
# evaluation is run every 2000 iterations
eval_batch_step: [0, 400]
cal_metric_during_train: False
pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
checkpoints:
save_inference_dir:
use_visualdl: False
print_batch_step: 10
save_model_dir: ./output/ch_PP-OCR_V3_det/
save_epoch_step: 100
eval_batch_step:
- 0
- 400
cal_metric_during_train: false
pretrained_model: null
checkpoints: null
save_inference_dir: null
use_visualdl: false
infer_img: doc/imgs_en/img_10.jpg
save_res_path: ./output/det_db/predicts_db.txt
save_res_path: ./checkpoints/det_db/predicts_db.txt
distributed: true
Architecture:
model_type: det
......@@ -23,10 +26,11 @@ Architecture:
name: MobileNetV3
scale: 0.5
model_name: large
disable_se: False
disable_se: True
Neck:
name: DBFPN
out_channels: 256
name: RSEFPN
out_channels: 96
shortcut: True
Head:
name: DBHead
k: 50
......@@ -35,32 +39,29 @@ Loss:
name: DBLoss
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
alpha: 5
beta: 10
ohem_ratio: 3
Optimizer:
name: Adam #Momentum
#momentum: 0.9
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.001
warmup_epoch: 2
regularizer:
name: 'L2'
factor: 0
name: L2
factor: 5.0e-05
PostProcess:
name: DBPostProcess
thresh: 0.3
box_thresh: 0.6
max_candidates: 1000
unclip_ratio: 1.5
Metric:
name: DetMetric
main_indicator: hmean
Train:
dataset:
name: SimpleDataSet
......@@ -69,34 +70,62 @@ Train:
- ./train_data/icdar2015/text_localization/train_icdar2015_label.txt
ratio_list: [1.0]
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- Resize:
size: [640, 640]
- MakeBorderMap:
shrink_ratio: 0.4
thresh_min: 0.3
thresh_max: 0.7
- MakeShrinkMap:
shrink_ratio: 0.4
min_text_size: 8
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] # the order of the dataloader list
- DecodeImage:
img_mode: BGR
channel_first: false
- DetLabelEncode: null
- IaaAugment:
augmenter_args:
- type: Fliplr
args:
p: 0.5
- type: Affine
args:
rotate:
- -10
- 10
- type: Resize
args:
size:
- 0.5
- 3
- EastRandomCropData:
size:
- 960
- 960
max_tries: 50
keep_ratio: true
- MakeBorderMap:
shrink_ratio: 0.4
thresh_min: 0.3
thresh_max: 0.7
- MakeShrinkMap:
shrink_ratio: 0.4
min_text_size: 8
- NormalizeImage:
scale: 1./255.
mean:
- 0.485
- 0.456
- 0.406
std:
- 0.229
- 0.224
- 0.225
order: hwc
- ToCHWImage: null
- KeepKeys:
keep_keys:
- image
- threshold_map
- threshold_mask
- shrink_map
- shrink_mask
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1
num_workers: 0
use_shared_memory: False
shuffle: true
drop_last: false
batch_size_per_card: 8
num_workers: 4
Eval:
dataset:
name: SimpleDataSet
......@@ -104,23 +133,31 @@ Eval:
label_file_list:
- ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- DetResizeForTest:
image_shape: [736, 1280]
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
- DecodeImage:
img_mode: BGR
channel_first: false
- DetLabelEncode: null
- DetResizeForTest: null
- NormalizeImage:
scale: 1./255.
mean:
- 0.485
- 0.456
- 0.406
std:
- 0.229
- 0.224
- 0.225
order: hwc
- ToCHWImage: null
- KeepKeys:
keep_keys:
- image
- shape
- polys
- ignore_tags
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1 # must be 1
num_workers: 0
use_shared_memory: False
shuffle: false
drop_last: false
batch_size_per_card: 1
num_workers: 2
configs/rec/ch_PP-OCRv3/ch_PP-OCRv3_rec.yml 0 → 100644
View file @ aa59fca5
Global:
debug: false
use_gpu: true
epoch_num: 500
log_smooth_window: 20
print_batch_step: 10
save_model_dir: ./output/rec_ppocr_v3
save_epoch_step: 3
eval_batch_step: [0, 2000]
cal_metric_during_train: true
pretrained_model:
checkpoints:
save_inference_dir:
use_visualdl: false
infer_img: doc/imgs_words/ch/word_1.jpg
character_dict_path: ppocr/utils/ppocr_keys_v1.txt
max_text_length: &max_text_length 25
infer_mode: false
use_space_char: true
distributed: true
save_res_path: ./output/rec/predicts_ppocrv3.txt
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Cosine
learning_rate: 0.001
warmup_epoch: 5
regularizer:
name: L2
factor: 3.0e-05
Architecture:
model_type: rec
algorithm: SVTR
Transform:
Backbone:
name: MobileNetV1Enhance
scale: 0.5
last_conv_stride: [1, 2]
last_pool_type: avg
Head:
name: MultiHead
head_list:
- CTCHead:
Neck:
name: svtr
dims: 64
depth: 2
hidden_dims: 120
use_guide: True
Head:
fc_decay: 0.00001
- SARHead:
enc_dim: 512
max_text_length: *max_text_length
Loss:
name: MultiLoss
loss_config_list:
- CTCLoss:
- SARLoss:
PostProcess:
name: CTCLabelDecode
Metric:
name: RecMetric
main_indicator: acc
ignore_space: True
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/
ext_op_transform_idx: 1
label_file_list:
- ./train_data/train_list.txt
transforms:
- DecodeImage:
img_mode: BGR
channel_first: false
- RecConAug:
prob: 0.5
ext_data_num: 2
image_shape: [48, 320, 3]
- RecAug:
- MultiLabelEncode:
- RecResizeImg:
image_shape: [3, 48, 320]
- KeepKeys:
keep_keys:
- image
- label_ctc
- label_sar
- length
- valid_ratio
loader:
shuffle: true
batch_size_per_card: 128
drop_last: true
num_workers: 4
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data
label_file_list:
- ./train_data/val_list.txt
transforms:
- DecodeImage:
img_mode: BGR
channel_first: false
- MultiLabelEncode:
- RecResizeImg:
image_shape: [3, 48, 320]
- KeepKeys:
keep_keys:
- image
- label_ctc
- label_sar
- length
- valid_ratio
loader:
shuffle: false
drop_last: false
batch_size_per_card: 128
num_workers: 4
configs/rec/ch_PP-OCRv3/ch_PP-OCRv3_rec_distillation.yml 0 → 100644
View file @ aa59fca5
Global:
debug: false
use_gpu: true
epoch_num: 800
log_smooth_window: 20
print_batch_step: 10
save_model_dir: ./output/rec_ppocr_v3_distillation
save_epoch_step: 3
eval_batch_step: [0, 2000]
cal_metric_during_train: true
pretrained_model:
checkpoints:
save_inference_dir:
use_visualdl: false
infer_img: doc/imgs_words/ch/word_1.jpg
character_dict_path: ppocr/utils/ppocr_keys_v1.txt
max_text_length: &max_text_length 25
infer_mode: false
use_space_char: true
distributed: true
save_res_path: ./output/rec/predicts_ppocrv3_distillation.txt
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
name: Piecewise
decay_epochs : [700, 800]
values : [0.0005, 0.00005]
warmup_epoch: 5
regularizer:
name: L2
factor: 3.0e-05
Architecture:
model_type: &model_type "rec"
name: DistillationModel
algorithm: Distillation
Models:
Teacher:
pretrained:
freeze_params: false
return_all_feats: true
model_type: *model_type
algorithm: SVTR
Transform:
Backbone:
name: MobileNetV1Enhance
scale: 0.5
last_conv_stride: [1, 2]
last_pool_type: avg
Head:
name: MultiHead
head_list:
- CTCHead:
Neck:
name: svtr
dims: 64
depth: 2
hidden_dims: 120
use_guide: True
Head:
fc_decay: 0.00001
- SARHead:
enc_dim: 512
max_text_length: *max_text_length
Student:
pretrained:
freeze_params: false
return_all_feats: true
model_type: *model_type
algorithm: SVTR
Transform:
Backbone:
name: MobileNetV1Enhance
scale: 0.5
last_conv_stride: [1, 2]
last_pool_type: avg
Head:
name: MultiHead
head_list:
- CTCHead:
Neck:
name: svtr
dims: 64
depth: 2
hidden_dims: 120
use_guide: True
Head:
fc_decay: 0.00001
- SARHead:
enc_dim: 512
max_text_length: *max_text_length
Loss:
name: CombinedLoss
loss_config_list:
- DistillationDMLLoss:
weight: 1.0
act: "softmax"
use_log: true
model_name_pairs:
- ["Student", "Teacher"]
key: head_out
multi_head: True
dis_head: ctc
name: dml_ctc
- DistillationDMLLoss:
weight: 0.5
act: "softmax"
use_log: true
model_name_pairs:
- ["Student", "Teacher"]
key: head_out
multi_head: True
dis_head: sar
name: dml_sar
- DistillationDistanceLoss:
weight: 1.0
mode: "l2"
model_name_pairs:
- ["Student", "Teacher"]
key: backbone_out
- DistillationCTCLoss:
weight: 1.0
model_name_list: ["Student", "Teacher"]
key: head_out
multi_head: True
- DistillationSARLoss:
weight: 1.0
model_name_list: ["Student", "Teacher"]
key: head_out
multi_head: True
PostProcess:
name: DistillationCTCLabelDecode
model_name: ["Student", "Teacher"]
key: head_out
multi_head: True
Metric:
name: DistillationMetric
base_metric_name: RecMetric
main_indicator: acc
key: "Student"
ignore_space: True
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/
ext_op_transform_idx: 1
label_file_list:
- ./train_data/train_list.txt
transforms:
- DecodeImage:
img_mode: BGR
channel_first: false
- RecConAug:
prob: 0.5
ext_data_num: 2
image_shape: [48, 320, 3]
- RecAug:
- MultiLabelEncode:
- RecResizeImg:
image_shape: [3, 48, 320]
- KeepKeys:
keep_keys:
- image
- label_ctc
- label_sar
- length
- valid_ratio
loader:
shuffle: true
batch_size_per_card: 128
drop_last: true
num_workers: 4
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data
label_file_list:
- ./train_data/val_list.txt
transforms:
- DecodeImage:
img_mode: BGR
channel_first: false
- MultiLabelEncode:
- RecResizeImg:
image_shape: [3, 48, 320]
- KeepKeys:
keep_keys:
- image
- label_ctc
- label_sar
- length
- valid_ratio
loader:
shuffle: false
drop_last: false
batch_size_per_card: 128
num_workers: 4
configs/rec/rec_svtrnet.yml 0 → 100644
View file @ aa59fca5
Global:
use_gpu: True
epoch_num: 20
log_smooth_window: 20
print_batch_step: 10
save_model_dir: ./output/rec/svtr/
save_epoch_step: 1
# evaluation is run every 2000 iterations after the 0th iteration
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_10.png
# for data or label process
character_dict_path:
character_type: en
max_text_length: 25
infer_mode: False
use_space_char: False
save_res_path: ./output/rec/predicts_svtr_tiny.txt
Optimizer:
name: AdamW
beta1: 0.9
beta2: 0.99
epsilon: 0.00000008
weight_decay: 0.05
no_weight_decay_name: norm pos_embed
one_dim_param_no_weight_decay: true
lr:
name: Cosine
learning_rate: 0.0005
warmup_epoch: 2
Architecture:
model_type: rec
algorithm: SVTR
Transform:
name: STN_ON
tps_inputsize: [32, 64]
tps_outputsize: [32, 100]
num_control_points: 20
tps_margins: [0.05,0.05]
stn_activation: none
Backbone:
name: SVTRNet
img_size: [32, 100]
out_char_num: 25
out_channels: 192
patch_merging: 'Conv'
embed_dim: [64, 128, 256]
depth: [3, 6, 3]
num_heads: [2, 4, 8]
mixer: ['Local','Local','Local','Local','Local','Local','Global','Global','Global','Global','Global','Global']
local_mixer: [[7, 11], [7, 11], [7, 11]]
last_stage: True
prenorm: false
Neck:
name: SequenceEncoder
encoder_type: reshape
Head:
name: CTCHead
Loss:
name: CTCLoss
PostProcess:
name: CTCLabelDecode
Metric:
name: RecMetric
main_indicator: acc
Train:
dataset:
name: LMDBDataSet
data_dir: ./train_data/data_lmdb_release/training/
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- CTCLabelEncode: # Class handling label
- RecResizeImg:
character_dict_path:
image_shape: [3, 64, 256]
padding: False
- KeepKeys:
keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
loader:
shuffle: True
batch_size_per_card: 512
drop_last: True
num_workers: 4
Eval:
dataset:
name: LMDBDataSet
data_dir: ./train_data/data_lmdb_release/validation/
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- CTCLabelEncode: # Class handling label
- RecResizeImg:
character_dict_path:
image_shape: [3, 64, 256]
padding: False
- KeepKeys:
keep_keys: ['image', 'label', 'length'] # dataloader will return list in this order
loader:
shuffle: False
drop_last: False
batch_size_per_card: 256
num_workers: 2
deploy/README.md 0 → 100644
View file @ aa59fca5
English | [简体中文](README_ch.md)
# PP-OCR Deployment
- [Paddle Deployment Introduction](#1)
- [PP-OCR Deployment](#2)
<a name="1"></a>
## Paddle Deployment Introduction
Paddle provides a variety of deployment schemes to meet the deployment requirements of different scenarios. Please choose according to the actual situation:
<div align="center">
<img src="../doc/deployment_en.png" width="800">
</div>
<a name="2"></a>
## PP-OCR Deployment
PP-OCR has supported muti deployment schemes. Click the link to get the specific tutorial.
- [Python Inference](../doc/doc_en/inference_ppocr_en.md)
- [C++ Inference](./cpp_infer/readme.md)
- [Serving (Python/C++)](./pdserving/README.md)
- [Paddle-Lite (ARM CPU/OpenCL ARM GPU/Metal ARM GPU)](./lite/readme.md)
- [Paddle.js](./paddlejs/README.md)
- [Jetson Inference]()
- [XPU Inference]()
- [Paddle2ONNX](./paddle2onnx/readme.md)
If you need the deployment tutorial of academic algorithm models other than PP-OCR, please directly enter the main page of corresponding algorithms, [entrance](../doc/doc_en/algorithm_overview_en.md)。
\ No newline at end of file
deploy/README_ch.md 0 → 100644
View file @ aa59fca5
[English](README.md) | 简体中文
# PP-OCR 模型推理部署
- [Paddle 推理部署方式简介](#1)
- [PP-OCR 推理部署](#2)
<a name="1"></a>
## Paddle 推理部署方式简介
飞桨提供多种部署方案，以满足不同场景的部署需求，请根据实际情况进行选择：
<div align="center">
<img src="../doc/deployment.png" width="800">
</div>
<a name="2"></a>
## PP-OCR 推理部署
PP-OCR模型已打通多种场景部署方案，点击链接获取具体的使用教程。
- [Python 推理](../doc/doc_ch/inference_ppocr.md)
- [C++ 推理](./cpp_infer/readme_ch.md)
- [Serving 服务化部署（Python/C++）](./pdserving/README_CN.md)
- [Paddle-Lite 端侧部署（ARM CPU/OpenCL ARM GPU/Metal ARM GPU）](./lite/readme_ch.md)
- [Paddle.js 部署](./paddlejs/README_ch.md)
- [Jetson 推理]()
- [XPU 推理]()
- [Paddle2ONNX 推理](./paddle2onnx/readme_ch.md)
需要PP-OCR以外的学术算法模型的推理部署，请直接进入相应算法主页面，[入口](../doc/doc_ch/algorithm_overview.md)。
\ No newline at end of file
deploy/cpp_infer/docs/windows_vs2019_build.md
View file @ aa59fca5
......@@ -70,7 +70,7 @@ cmake安装完后后系统里会有一个cmake-gui程序，打开cmake-gui，在
* cpu版本，仅需考虑OPENCV_DIR、OpenCV_DIR、PADDLE_LIB三个参数
- OPENCV_DIR：填写opencv lib文件夹所在位置
- OpenCV_DIR：同填写opencv lib文件夹所在位
- OpenCV_DIR：同填写opencv lib文件夹所在位置
- PADDLE_LIB：paddle_inference文件夹所在位置
* GPU版本，在cpu版本的基础上，还需填写以下变量
......@@ -78,7 +78,7 @@ CUDA_LIB、CUDNN_LIB、TENSORRT_DIR、WITH_GPU、WITH_TENSORRT
- CUDA_LIB: CUDA地址，如 `C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2\lib\x64`
- CUDNN_LIB: 和CUDA_LIB一致
- TENSORRT_DIR：TRT下载后解压缩的位置
- TENSORRT_DIR：TRT下载后解压缩的位置，如 `D:\TensorRT-8.0.1.6`
- WITH_GPU: 打钩
- WITH_TENSORRT：打勾
......@@ -110,10 +110,11 @@ 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
1. `paddle_inference/paddle/lib/paddle_inference.dll`
2. `opencv/build/x64/vc15/bin/opencv_world455.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`
### Step4: 预测
上述`Visual Studio 2019`编译产出的可执行文件在`out\build\x64-Release\Release`目录下，打开`cmd`，并切换到`D:\projects\cpp\PaddleOCR\deploy\cpp_infer\`：
上述`Visual Studio 2019`编译产出的可执行文件在`build/Release/`目录下，打开`cmd`，并切换到`D:\projects\cpp\PaddleOCR\deploy\cpp_infer\`：
```
cd /d D:\projects\cpp\PaddleOCR\deploy\cpp_infer
......@@ -128,7 +129,7 @@ CHCP 65001
```
识别结果如下
![result](imgs/result.png)
![result](imgs/result.jpg)
## FAQ
......
deploy/cpp_infer/include/args.h
View file @ aa59fca5
......@@ -46,6 +46,8 @@ DECLARE_int32(cls_batch_num);
DECLARE_string(rec_model_dir);
DECLARE_int32(rec_batch_num);
DECLARE_string(rec_char_dict_path);
DECLARE_int32(rec_img_h);
DECLARE_int32(rec_img_w);
// forward related
DECLARE_bool(det);
DECLARE_bool(rec);
......
deploy/cpp_infer/include/ocr_rec.h
View file @ aa59fca5
......@@ -45,7 +45,8 @@ public:
const bool &use_mkldnn, const string &label_path,
const bool &use_tensorrt,
const std::string &precision,
const int &rec_batch_num) {
const int &rec_batch_num, const int &rec_img_h,
const int &rec_img_w) {
this->use_gpu_ = use_gpu;
this->gpu_id_ = gpu_id;
this->gpu_mem_ = gpu_mem;
......@@ -54,6 +55,10 @@ public:
this->use_tensorrt_ = use_tensorrt;
this->precision_ = precision;
this->rec_batch_num_ = rec_batch_num;
this->rec_img_h_ = rec_img_h;
this->rec_img_w_ = rec_img_w;
std::vector<int> rec_image_shape = {3, rec_img_h, rec_img_w};
this->rec_image_shape_ = rec_image_shape;
this->label_list_ = Utility::ReadDict(label_path);
this->label_list_.insert(this->label_list_.begin(),
......@@ -86,7 +91,9 @@ private:
bool use_tensorrt_ = false;
std::string precision_ = "fp32";
int rec_batch_num_ = 6;
int rec_img_h_ = 32;
int rec_img_w_ = 320;
std::vector<int> rec_image_shape_ = {3, rec_img_h_, rec_img_w_};
// pre-process
CrnnResizeImg resize_op_;
Normalize normalize_op_;
......
deploy/cpp_infer/include/paddleocr.h
View file @ aa59fca5
......@@ -39,10 +39,10 @@ using namespace paddle_infer;
namespace PaddleOCR {
class PaddleOCR {
class PPOCR {
public:
explicit PaddleOCR();
~PaddleOCR();
explicit PPOCR();
~PPOCR();
std::vector<std::vector<OCRPredictResult>>
ocr(std::vector<cv::String> cv_all_img_names, bool det = true,
bool rec = true, bool cls = true);
......
deploy/cpp_infer/include/utility.h
View file @ aa59fca5
......@@ -65,6 +65,8 @@ public:
static bool PathExists(const std::string &path);
static void CreateDir(const std::string &path);
static void print_result(const std::vector<OCRPredictResult> &ocr_result);
};
......
deploy/cpp_infer/readme.md
View file @ aa59fca5
- [服务器端C++预测](#服务器端c预测)
- [1. 准备环境](#1-准备环境)
- [1.0 运行准备](#10-运行准备)
- [1.1 编译opencv库](#11-编译opencv库)
- [1.2 下载或者编译Paddle预测库](#12-下载或者编译paddle预测库)
- [1.2.1 直接下载安装](#121-直接下载安装)
- [1.2.2 预测库源码编译](#122-预测库源码编译)
- [2 开始运行](#2-开始运行)
- [2.1 将模型导出为inference model](#21-将模型导出为inference-model)
- [2.2 编译PaddleOCR C++预测demo](#22-编译paddleocr-c预测demo)
- [2.3 运行demo](#23-运行demo)
- [1. 检测+分类+识别：](#1-检测分类识别)
- [2. 检测+识别：](#2-检测识别)
- [3. 检测：](#3-检测)
- [4. 分类+识别：](#4-分类识别)
- [5. 识别：](#5-识别)
- [6. 分类：](#6-分类)
- [3. FAQ](#3-faq)
# 服务器端C++预测
本章节介绍PaddleOCR 模型的的C++部署方法，与之对应的python预测部署方式参考[文档](../../doc/doc_ch/inference.md)。
C++在性能计算上优于python，因此，在大多数CPU、GPU部署场景，多采用C++的部署方式，本节将介绍如何在Linux\Windows (CPU\GPU)环境下配置C++环境并完成
PaddleOCR模型部署。
English | [简体中文](readme_ch.md)
# Server-side C++ Inference
<a name="1"></a>
## 1. 准备环境
- [1. Prepare the Environment](#1)
- [1.1 Environment](#11)
- [1.2 Compile OpenCV](#12)
- [1.3 Compile or Download or the Paddle Inference Library](#13)
- [2. Compile and Run the Demo](#2)
- [2.1 Export the inference model](#21)
- [2.2 Compile PaddleOCR C++ inference demo](#22)
- [2.3 Run the demo](#23)
- [3. FAQ](#3)
<a name="10"></a>
### 1.0 运行准备
This chapter introduces the C++ deployment steps of the PaddleOCR model. C++ is better than Python in terms of performance. Therefore, in CPU and GPU deployment scenarios, C++ deployment is mostly used.
This section will introduce how to configure the C++ environment and deploy PaddleOCR in Linux (CPU\GPU) environment. For Windows deployment please refer to [Windows](./docs/windows_vs2019_build.md) compilation guidelines.
- Linux环境，推荐使用docker。
- Windows环境。
* 该文档主要介绍基于Linux环境的PaddleOCR C++预测流程，如果需要在Windows下基于预测库进行C++预测，具体编译方法请参考[Windows下编译教程](./docs/windows_vs2019_build.md)
<a name="1"></a>
## 1. Prepare the Environment
<a name="11"></a>
### 1.1 Environment
- Linux, docker is recommended.
- Windows.
### 1.1 编译opencv库
* 首先需要从opencv官网上下载在Linux环境下源码编译的包，以opencv3.4.7为例，下载命令如下。
<a name="12"></a>
### 1.2 Compile OpenCV
* First of all, you need to download the source code compiled package in the Linux environment from the OpenCV official website. Taking OpenCV 3.4.7 as an example, the download command is as follows.
```bash
cd deploy/cpp_infer
......@@ -49,18 +38,18 @@ wget https://paddleocr.bj.bcebos.com/libs/opencv/opencv-3.4.7.tar.gz
tar -xf opencv-3.4.7.tar.gz
```
最终可以在当前目录下看到`opencv-3.4.7/`的文件夹。
Finally, you will see the folder of `opencv-3.4.7/` in the current directory.
* Compile OpenCV, the OpenCV source path (`root_path`) and installation path (`install_path`) should be set by yourself. Enter the OpenCV source code path and compile it in the following way.
* 编译opencv，设置opencv源码路径(`root_path`)以及安装路径(`install_path`)。进入opencv源码路径下，按照下面的方式进行编译。
```shell
root_path="your_opencv_root_path"
root_path=your_opencv_root_path
install_path=${root_path}/opencv3
build_dir=${root_path}/build
rm -rf ${build_dir}
mkdir ${build_dir}
cd ${build_dir}
rm -rf build
mkdir build
cd build
cmake .. \
-DCMAKE_INSTALL_PREFIX=${install_path} \
......@@ -84,15 +73,11 @@ make -j
make install
```
也可以直接修改`tools/build_opencv.sh`的内容，然后直接运行下面的命令进行编译。
In the above commands, `root_path` is the downloaded OpenCV source code path, and `install_path` is the installation path of OpenCV. After `make install` is completed, the OpenCV header file and library file will be generated in this folder for later OCR source code compilation.
```shell
sh tools/build_opencv.sh
```
其中`root_path`为下载的opencv源码路径，`install_path`为opencv的安装路径，`make install`完成之后，会在该文件夹下生成opencv头文件和库文件，用于后面的OCR代码编译。
最终在安装路径下的文件结构如下所示。
The final file structure under the OpenCV installation path is as follows.
```
opencv3/
......@@ -103,35 +88,35 @@ opencv3/
|-- share
```
<a name="12"></a>
### 1.2 下载或者编译Paddle预测库
<a name="13"></a>
### 1.3 Compile or Download or the Paddle Inference Library
* 有2种方式获取Paddle预测库，下面进行详细介绍。
* There are 2 ways to obtain the Paddle inference library, described in detail below.
#### 1.3.1 Direct download and installation
#### 1.2.1 直接下载安装
[Paddle inference library official website](https://paddleinference.paddlepaddle.org.cn/user_guides/download_lib.html#linux). You can review and select the appropriate version of the inference library on the official website.
* [Paddle预测库官网](https://paddleinference.paddlepaddle.org.cn/user_guides/download_lib.html#linux) 上提供了不同cuda版本的Linux预测库，可以在官网查看并选择合适的预测库版本（*建议选择paddle版本>=2.0.1版本的预测库* ）。
* 下载之后使用下面的方法解压。
* After downloading, use the following command to extract files.
```
tar -xf paddle_inference.tgz
```
最终会在当前的文件夹中生成`paddle_inference/`的子文件夹。
Finally you will see the the folder of `paddle_inference/` in the current path.
#### 1.3.2 Compile the inference source code
* If you want to get the latest Paddle inference library features, you can download the latest code from Paddle GitHub repository and compile the inference library from the source code. It is recommended to download the inference library with paddle version greater than or equal to 2.0.1.
* You can refer to [Paddle inference library] (https://www.paddlepaddle.org.cn/documentation/docs/en/advanced_guide/inference_deployment/inference/build_and_install_lib_en.html) to get the Paddle source code from GitHub, and then compile To generate the latest inference library. The method of using git to access the code is as follows.
#### 1.2.2 预测库源码编译
* 如果希望获取最新预测库特性，可以从Paddle github上克隆最新代码，源码编译预测库。
* 可以参考[Paddle预测库安装编译说明](https://www.paddlepaddle.org.cn/documentation/docs/zh/2.0/guides/05_inference_deployment/inference/build_and_install_lib_cn.html#congyuanmabianyi) 的说明，从github上获取Paddle代码，然后进行编译，生成最新的预测库。使用git获取代码方法如下。
```shell
git clone https://github.com/PaddlePaddle/Paddle.git
git checkout develop
```
* 进入Paddle目录后，编译方法如下。
* Enter the Paddle directory and run the following commands to compile the paddle inference library.
```shell
rm -rf build
......@@ -151,10 +136,10 @@ make -j
make inference_lib_dist
```
更多编译参数选项介绍可以参考[文档说明](https://www.paddlepaddle.org.cn/documentation/docs/zh/2.0/guides/05_inference_deployment/inference/build_and_install_lib_cn.html#congyuanmabianyi)。
For more compilation parameter options, please refer to the [document](https://www.paddlepaddle.org.cn/documentation/docs/zh/2.0/guides/05_inference_deployment/inference/build_and_install_lib_cn.html#congyuanmabianyi).
* 编译完成之后，可以在`build/paddle_inference_install_dir/`文件下看到生成了以下文件及文件夹。
* After the compilation process, you can see the following files in the folder of `build/paddle_inference_install_dir/`.
```
build/paddle_inference_install_dir/
......@@ -164,17 +149,16 @@ build/paddle_inference_install_dir/
|-- version.txt
```
其中`paddle`就是C++预测所需的Paddle库，`version.txt`中包含当前预测库的版本信息。
`paddle` is the Paddle library required for C++ prediction later, and `version.txt` contains the version information of the current inference library.
<a name="2"></a>
## 2 开始运行
<a name="2"></a>
## 2. Compile and Run the Demo
<a name="21"></a>
### 2.1 Export the inference model
### 2.1 将模型导出为inference model
* 可以参考[模型预测章节](../../doc/doc_ch/inference.md)，导出inference model，用于模型预测。模型导出之后，假设放在`inference`目录下，则目录结构如下。
* You can refer to [Model inference](../../doc/doc_ch/inference.md) and export the inference model. After the model is exported, assuming it is placed in the `inference` directory, the directory structure is as follows.
```
inference/
......@@ -189,41 +173,44 @@ inference/
| |--inference.pdmodel
```
<a name="22"></a>
### 2.2 编译PaddleOCR C++预测demo
<a name="22"></a>
### 2.2 Compile PaddleOCR C++ inference demo
* 编译命令如下，其中Paddle C++预测库、opencv等其他依赖库的地址需要换成自己机器上的实际地址。
* The compilation commands are as follows. The addresses of Paddle C++ inference library, opencv and other Dependencies need to be replaced with the actual addresses on your own machines.
```shell
sh tools/build.sh
```
* 具体的，需要修改`tools/build.sh`中环境路径，相关内容如下：
Specifically, you should modify the paths in `tools/build.sh`. The related content is as follows.
```shell
OPENCV_DIR=your_opencv_dir
LIB_DIR=your_paddle_inference_dir
CUDA_LIB_DIR=your_cuda_lib_dir
CUDNN_LIB_DIR=/your_cudnn_lib_dir
CUDNN_LIB_DIR=your_cudnn_lib_dir
```
其中，`OPENCV_DIR`为opencv编译安装的地址；`LIB_DIR`为下载(`paddle_inference`文件夹)或者编译生成的Paddle预测库地址(`build/paddle_inference_install_dir`文件夹)；`CUDA_LIB_DIR`为cuda库文件地址，在docker中为`/usr/local/cuda/lib64`；`CUDNN_LIB_DIR`为cudnn库文件地址，在docker中为`/usr/lib/x86_64-linux-gnu/`。**注意：以上路径都写绝对路径，不要写相对路径。**
`OPENCV_DIR` is the OpenCV installation path; `LIB_DIR` is the download (`paddle_inference` folder)
or the generated Paddle inference library path (`build/paddle_inference_install_dir` folder);
`CUDA_LIB_DIR` is the CUDA library file path, in docker; it is `/usr/local/cuda/lib64`; `CUDNN_LIB_DIR` is the cuDNN library file path, in docker it is `/usr/lib/x86_64-linux-gnu/`.
* 编译完成之后，会在`build`文件夹下生成一个名为`ppocr`的可执行文件。
* After the compilation is completed, an executable file named `ppocr` will be generated in the `build` folder.
<a name="23"></a>
### 2.3 运行demo
<a name="23"></a>
### 2.3 Run the demo
运行方式：
Execute the built executable file:
```shell
./build/ppocr [--param1] [--param2] [...]
```
具体命令如下：
##### 1. 检测+分类+识别：
Specifically,
##### 1. det+cls+rec：
```shell
./build/ppocr --det_model_dir=inference/det_db \
--rec_model_dir=inference/rec_rcnn \
......@@ -235,7 +222,7 @@ CUDNN_LIB_DIR=/your_cudnn_lib_dir
--cls=true \
```
##### 2. 检测+识别：
##### 2. det+rec：
```shell
./build/ppocr --det_model_dir=inference/det_db \
--rec_model_dir=inference/rec_rcnn \
......@@ -246,7 +233,7 @@ CUDNN_LIB_DIR=/your_cudnn_lib_dir
--cls=false \
```
##### 3. 检测：
##### 3. det
```shell
./build/ppocr --det_model_dir=inference/det_db \
--image_dir=../../doc/imgs/12.jpg \
......@@ -254,7 +241,7 @@ CUDNN_LIB_DIR=/your_cudnn_lib_dir
--rec=false
```
##### 4. 分类+识别：
##### 4. cls+rec：
```shell
./build/ppocr --rec_model_dir=inference/rec_rcnn \
--cls_model_dir=inference/cls \
......@@ -265,7 +252,7 @@ CUDNN_LIB_DIR=/your_cudnn_lib_dir
--cls=true \
```
##### 5. 识别：
##### 5. rec
```shell
./build/ppocr --rec_model_dir=inference/rec_rcnn \
--image_dir=../../doc/imgs_words/ch/word_1.jpg \
......@@ -275,7 +262,7 @@ CUDNN_LIB_DIR=/your_cudnn_lib_dir
--cls=false \
```
##### 6. 分类：
##### 6. cls
```shell
./build/ppocr --cls_model_dir=inference/cls \
--cls_model_dir=inference/cls \
......@@ -286,61 +273,64 @@ CUDNN_LIB_DIR=/your_cudnn_lib_dir
--cls=true \
```
更多支持的可调节参数解释如下：
More parameters are as follows,
- 通用参数
- Common parameters
|参数名称|类型|默认参数|意义|
| :---: | :---: | :---: | :---: |
|use_gpu|bool|false|是否使用GPU|
|gpu_id|int|0|GPU id，使用GPU时有效|
|gpu_mem|int|4000|申请的GPU内存|
|cpu_math_library_num_threads|int|10|CPU预测时的线程数，在机器核数充足的情况下，该值越大，预测速度越快|
|enable_mkldnn|bool|true|是否使用mkldnn库|
|output|str|./output|可视化结果保存的路径|
|parameter|data type|default|meaning|
| --- | --- | --- | --- |
|use_gpu|bool|false|Whether to use GPU|
|gpu_id|int|0|GPU id when use_gpu is true|
|gpu_mem|int|4000|GPU memory requested|
|cpu_math_library_num_threads|int|10|Number of threads when using CPU inference. When machine cores is enough, the large the value, the faster the inference speed|
|enable_mkldnn|bool|true|Whether to use mkdlnn library|
|output|str|./output|Path where visualization results are saved|
- 前向相关
|参数名称|类型|默认参数|意义|
- forward
|parameter|data type|default|meaning|
| :---: | :---: | :---: | :---: |
|det|bool|true|前向是否执行文字检测|
|rec|bool|true|前向是否执行文字识别|
|cls|bool|false|前向是否执行文字方向分类|
- 检测模型相关
- Detection related parameters
|参数名称|类型|默认参数|意义|
| :---: | :---: | :---: | :---: |
|det_model_dir|string|-|检测模型inference model地址|
|max_side_len|int|960|输入图像长宽大于960时，等比例缩放图像，使得图像最长边为960|
|det_db_thresh|float|0.3|用于过滤DB预测的二值化图像，设置为0.-0.3对结果影响不明显|
|det_db_box_thresh|float|0.5|DB后处理过滤box的阈值，如果检测存在漏框情况，可酌情减小|
|det_db_unclip_ratio|float|1.6|表示文本框的紧致程度，越小则文本框更靠近文本|
|det_db_score_mode|string|slow|slow:使用多边形框计算bbox score，fast:使用矩形框计算。矩形框计算速度更快，多边形框对弯曲文本区域计算更准确。|
|visualize|bool|true|是否对结果进行可视化，为1时，预测结果会保存在`output`字段指定的文件夹下和输入图像同名的图像上。|
|parameter|data type|default|meaning|
| --- | --- | --- | --- |
|det_model_dir|string|-|Address of detection inference model|
|max_side_len|int|960|Limit the maximum image height and width to 960|
|det_db_thresh|float|0.3|Used to filter the binarized image of DB prediction, setting 0.-0.3 has no obvious effect on the result|
|det_db_box_thresh|float|0.5|DB post-processing filter box threshold, if there is a missing box detected, it can be reduced as appropriate|
|det_db_unclip_ratio|float|1.6|Indicates the compactness of the text box, the smaller the value, the closer the text box to the text|
|det_db_score_mode|string|slow| slow: use polygon box to calculate bbox score, fast: use rectangle box to calculate. Use rectangular box to calculate faster, and polygonal box more accurate for curved text area.|
|visualize|bool|true|Whether to visualize the results，when it is set as true, the prediction results will be saved in the folder specified by the `output` field on an image with the same name as the input image.|
- 方向分类器相关
- Classifier related parameters
|参数名称|类型|默认参数|意义|
| :---: | :---: | :---: | :---: |
|use_angle_cls|bool|false|是否使用方向分类器|
|cls_model_dir|string|-|方向分类器inference model地址|
|cls_thresh|float|0.9|方向分类器的得分阈值|
|cls_batch_num|int|1|方向分类器batchsize|
|parameter|data type|default|meaning|
| --- | --- | --- | --- |
|use_angle_cls|bool|false|Whether to use the direction classifier|
|cls_model_dir|string|-|Address of direction classifier inference model|
|cls_thresh|float|0.9|Score threshold of the direction classifier|
|cls_batch_num|int|1|batch size of classifier|
- 识别模型相关
- Recognition related parameters
|参数名称|类型|默认参数|意义|
| :---: | :---: | :---: | :---: |
|rec_model_dir|string|-|识别模型inference model地址|
|rec_char_dict_path|string|../../ppocr/utils/ppocr_keys_v1.txt|字典文件|
|rec_batch_num|int|6|识别模型batchsize|
|parameter|data type|default|meaning|
| --- | --- | --- | --- |
|rec_model_dir|string|-|Address of recognition inference model|
|rec_char_dict_path|string|../../ppocr/utils/ppocr_keys_v1.txt|dictionary file|
|rec_batch_num|int|6|batch size of recognition|
|rec_img_h|int|32|image height of recognition|
|rec_img_w|int|320|image width of recognition|
* Multi-language inference is also supported in PaddleOCR, you can refer to [recognition tutorial](../../doc/doc_en/recognition_en.md) for more supported languages and models in PaddleOCR. Specifically, if you want to infer using multi-language models, you just need to modify values of `rec_char_dict_path` and `rec_model_dir`.
* PaddleOCR也支持多语言的预测，更多支持的语言和模型可以参考[识别文档](../../doc/doc_ch/recognition.md)中的多语言字典与模型部分，如果希望进行多语言预测，只需将修改`rec_char_dict_path`（字典文件路径）以及`rec_model_dir`（inference模型路径）字段即可。
最终屏幕上会输出检测结果如下。
The detection results will be shown on the screen, which is as follows.
```bash
predict img: ../../doc/imgs/12.jpg
......@@ -352,6 +342,8 @@ predict img: ../../doc/imgs/12.jpg
The detection visualized image saved in ./output//12.jpg
```
<a name="3"></a>
## 3. FAQ
1. 遇到报错 `unable to access 'https://github.com/LDOUBLEV/AutoLog.git/': gnutls_handshake() failed: The TLS connection was non-properly terminated.`， 将 `deploy/cpp_infer/external-cmake/auto-log.cmake` 中的github地址改为 https://gitee.com/Double_V/AutoLog 地址即可。
1. Encountered the error `unable to access 'https://github.com/LDOUBLEV/AutoLog.git/': gnutls_handshake() failed: The TLS connection was non-properly terminated.`, change the github address in `deploy/cpp_infer/external-cmake/auto-log.cmake` to the https://gitee.com/Double_V/AutoLog address.
deploy/cpp_infer/readme_en.md → deploy/cpp_infer/readme_ch.md
View file @ aa59fca5
- [Server-side C++ Inference](#server-side-c-inference)
- [1. Prepare the Environment](#1-prepare-the-environment)
- [Environment](#environment)
- [1.1 Compile OpenCV](#11-compile-opencv)
- [1.2 Compile or Download or the Paddle Inference Library](#12-compile-or-download-or-the-paddle-inference-library)
- [1.2.1 Direct download and installation](#121-direct-download-and-installation)
- [1.2.2 Compile the inference source code](#122-compile-the-inference-source-code)
- [2. Compile and Run the Demo](#2-compile-and-run-the-demo)
- [2.1 Export the inference model](#21-export-the-inference-model)
- [2.2 Compile PaddleOCR C++ inference demo](#22-compile-paddleocr-c-inference-demo)
- [Run the demo](#run-the-demo)
- [1. det+cls+rec：](#1-detclsrec)
- [2. det+rec：](#2-detrec)
- [3. det](#3-det)
- [4. cls+rec：](#4-clsrec)
- [5. rec](#5-rec)
- [6. cls](#6-cls)
- [3. FAQ](#3-faq)
[English](readme.md) | 简体中文
# Server-side C++ Inference
# 服务器端C++预测
This chapter introduces the C++ deployment steps of the PaddleOCR model. The corresponding Python predictive deployment method refers to [document](../../doc/doc_ch/inference.md).
C++ is better than python in terms of performance. Therefore, in CPU and GPU deployment scenarios, C++ deployment is mostly used.
This section will introduce how to configure the C++ environment and deploy PaddleOCR in Linux (CPU\GPU) environment. For Windows deployment please refer to [Windows](./docs/windows_vs2019_build.md) compilation guidelines.
- [1. 准备环境](#1)
- [1.1 运行准备](#11)
- [1.2 编译opencv库](#12)
- [1.3 下载或者编译Paddle预测库](#13)
- [2 开始运行](#2)
- [2.1 准备模型](#21)
- [2.2 编译PaddleOCR C++预测demo](#22)
- [2.3 运行demo](#23)
- [3. FAQ](#3)
本章节介绍PaddleOCR 模型的的C++部署方法。C++在性能计算上优于Python，因此，在大多数CPU、GPU部署场景，多采用C++的部署方式，本节将介绍如何在Linux\Windows (CPU\GPU)环境下配置C++环境并完成PaddleOCR模型部署。
## 1. Prepare the Environment
### Environment
<a name="1"></a>
- Linux, docker is recommended.
- Windows.
## 1. 准备环境
<a name="11"></a>
### 1.1 Compile OpenCV
### 1.1 运行准备
* First of all, you need to download the source code compiled package in the Linux environment from the OpenCV official website. Taking OpenCV 3.4.7 as an example, the download command is as follows.
- Linux环境，推荐使用docker。
- Windows环境。
* 该文档主要介绍基于Linux环境的PaddleOCR C++预测流程，如果需要在Windows下基于预测库进行C++预测，具体编译方法请参考[Windows下编译教程](./docs/windows_vs2019_build.md)
<a name="12"></a>
### 1.2 编译opencv库
* 首先需要从opencv官网上下载在Linux环境下源码编译的包，以opencv3.4.7为例，下载命令如下。
```bash
cd deploy/cpp_infer
......@@ -42,18 +40,18 @@ wget https://paddleocr.bj.bcebos.com/libs/opencv/opencv-3.4.7.tar.gz
tar -xf opencv-3.4.7.tar.gz
```
Finally, you will see the folder of `opencv-3.4.7/` in the current directory.
* Compile OpenCV, the OpenCV source path (`root_path`) and installation path (`install_path`) should be set by yourself. Enter the OpenCV source code path and compile it in the following way.
最终可以在当前目录下看到`opencv-3.4.7/`的文件夹。
* 编译opencv，设置opencv源码路径(`root_path`)以及安装路径(`install_path`)。进入opencv源码路径下，按照下面的方式进行编译。
```shell
root_path=your_opencv_root_path
root_path="your_opencv_root_path"
install_path=${root_path}/opencv3
build_dir=${root_path}/build
rm -rf build
mkdir build
cd build
rm -rf ${build_dir}
mkdir ${build_dir}
cd ${build_dir}
cmake .. \
-DCMAKE_INSTALL_PREFIX=${install_path} \
......@@ -77,11 +75,15 @@ make -j
make install
```
In the above commands, `root_path` is the downloaded OpenCV source code path, and `install_path` is the installation path of OpenCV. After `make install` is completed, the OpenCV header file and library file will be generated in this folder for later OCR source code compilation.
也可以直接修改`tools/build_opencv.sh`的内容，然后直接运行下面的命令进行编译。
```shell
sh tools/build_opencv.sh
```
其中`root_path`为下载的opencv源码路径，`install_path`为opencv的安装路径，`make install`完成之后，会在该文件夹下生成opencv头文件和库文件，用于后面的OCR代码编译。
The final file structure under the OpenCV installation path is as follows.
最终在安装路径下的文件结构如下所示。
```
opencv3/
......@@ -92,34 +94,38 @@ opencv3/
|-- share
```
### 1.2 Compile or Download or the Paddle Inference Library
<a name="13"></a>
* There are 2 ways to obtain the Paddle inference library, described in detail below.
### 1.3 下载或者编译Paddle预测库
#### 1.2.1 Direct download and installation
可以选择直接下载安装或者从源码编译，下文分别进行具体说明。
[Paddle inference library official website](https://paddleinference.paddlepaddle.org.cn/user_guides/download_lib.html#linux). You can review and select the appropriate version of the inference library on the official website.
<a name="131"></a>
#### 1.3.1 直接下载安装
[Paddle预测库官网](https://paddleinference.paddlepaddle.org.cn/user_guides/download_lib.html#linux) 上提供了不同cuda版本的Linux预测库，可以在官网查看并选择合适的预测库版本（*建议选择paddle版本>=2.0.1版本的预测库* ）。
* After downloading, use the following command to extract files.
下载之后解压:
```
```shell
tar -xf paddle_inference.tgz
```
Finally you will see the the folder of `paddle_inference/` in the current path.
最终会在当前的文件夹中生成`paddle_inference/`的子文件夹。
#### 1.2.2 Compile the inference source code
* If you want to get the latest Paddle inference library features, you can download the latest code from Paddle GitHub repository and compile the inference library from the source code. It is recommended to download the inference library with paddle version greater than or equal to 2.0.1.
* You can refer to [Paddle inference library] (https://www.paddlepaddle.org.cn/documentation/docs/en/advanced_guide/inference_deployment/inference/build_and_install_lib_en.html) to get the Paddle source code from GitHub, and then compile To generate the latest inference library. The method of using git to access the code is as follows.
<a name="132"></a>
#### 1.3.2 预测库源码编译
如果希望获取最新预测库特性，可以从github上克隆最新Paddle代码进行编译，生成最新的预测库。
* 使用git获取代码:
```shell
git clone https://github.com/PaddlePaddle/Paddle.git
git checkout develop
```
* Enter the Paddle directory and run the following commands to compile the paddle inference library.
* 进入Paddle目录，进行编译:
```shell
rm -rf build
......@@ -139,10 +145,10 @@ make -j
make inference_lib_dist
```
For more compilation parameter options, please refer to the [document](https://www.paddlepaddle.org.cn/documentation/docs/zh/2.0/guides/05_inference_deployment/inference/build_and_install_lib_cn.html#congyuanmabianyi).
更多编译参数选项介绍可以参考[Paddle预测库编译文档](https://www.paddlepaddle.org.cn/documentation/docs/zh/2.0/guides/05_inference_deployment/inference/build_and_install_lib_cn.html#congyuanmabianyi)。
* After the compilation process, you can see the following files in the folder of `build/paddle_inference_install_dir/`.
* 编译完成之后，可以在`build/paddle_inference_install_dir/`文件下看到生成了以下文件及文件夹。
```
build/paddle_inference_install_dir/
......@@ -152,14 +158,17 @@ build/paddle_inference_install_dir/
|-- version.txt
```
`paddle` is the Paddle library required for C++ prediction later, and `version.txt` contains the version information of the current inference library.
其中`paddle`就是C++预测所需的Paddle库，`version.txt`中包含当前预测库的版本信息。
<a name="2"></a>
## 2. 开始运行
## 2. Compile and Run the Demo
<a name="21"></a>
### 2.1 Export the inference model
### 2.1 准备模型
* You can refer to [Model inference](../../doc/doc_ch/inference.md) and export the inference model. After the model is exported, assuming it is placed in the `inference` directory, the directory structure is as follows.
直接下载PaddleOCR提供的推理模型，或者参考[模型预测章节](../../doc/doc_ch/inference_ppocr.md)，将训练好的模型导出为推理模型。模型导出之后，假设放在`inference`目录下，则目录结构如下。
```
inference/
......@@ -174,42 +183,43 @@ inference/
| |--inference.pdmodel
```
<a name="22"></a>
### 2.2 Compile PaddleOCR C++ inference demo
### 2.2 编译PaddleOCR C++预测demo
* The compilation commands are as follows. The addresses of Paddle C++ inference library, opencv and other Dependencies need to be replaced with the actual addresses on your own machines.
编译命令如下，其中Paddle C++预测库、opencv等其他依赖库的地址需要换成自己机器上的实际地址。
```shell
sh tools/build.sh
```
Specifically, you should modify the paths in `tools/build.sh`. The related content is as follows.
具体的，需要修改`tools/build.sh`中环境路径，相关内容如下：
```shell
OPENCV_DIR=your_opencv_dir
LIB_DIR=your_paddle_inference_dir
CUDA_LIB_DIR=your_cuda_lib_dir
CUDNN_LIB_DIR=your_cudnn_lib_dir
CUDNN_LIB_DIR=/your_cudnn_lib_dir
```
`OPENCV_DIR` is the OpenCV installation path; `LIB_DIR` is the download (`paddle_inference` folder)
or the generated Paddle inference library path (`build/paddle_inference_install_dir` folder);
`CUDA_LIB_DIR` is the CUDA library file path, in docker; it is `/usr/local/cuda/lib64`; `CUDNN_LIB_DIR` is the cuDNN library file path, in docker it is `/usr/lib/x86_64-linux-gnu/`.
其中，`OPENCV_DIR`为opencv编译安装的地址；`LIB_DIR`为下载(`paddle_inference`文件夹)或者编译生成的Paddle预测库地址(`build/paddle_inference_install_dir`文件夹)；`CUDA_LIB_DIR`为cuda库文件地址，在docker中为`/usr/local/cuda/lib64`；`CUDNN_LIB_DIR`为cudnn库文件地址，在docker中为`/usr/lib/x86_64-linux-gnu/`。**注意：以上路径都写绝对路径，不要写相对路径。**
* After the compilation is completed, an executable file named `ppocr` will be generated in the `build` folder.
编译完成之后，会在`build`文件夹下生成一个名为`ppocr`的可执行文件。
<a name="23"></a>
### Run the demo
Execute the built executable file:
### 2.3 运行demo
本demo支持系统串联调用，也支持单个功能的调用，如，只使用检测或识别功能。
运行方式：
```shell
./build/ppocr [--param1] [--param2] [...]
```
具体命令如下：
Specifically,
##### 1. det+cls+rec：
##### 1. 检测+分类+识别：
```shell
./build/ppocr --det_model_dir=inference/det_db \
--rec_model_dir=inference/rec_rcnn \
......@@ -221,7 +231,7 @@ Specifically,
--cls=true \
```
##### 2. det+rec：
##### 2. 检测+识别：
```shell
./build/ppocr --det_model_dir=inference/det_db \
--rec_model_dir=inference/rec_rcnn \
......@@ -232,7 +242,7 @@ Specifically,
--cls=false \
```
##### 3. det
##### 3. 检测：
```shell
./build/ppocr --det_model_dir=inference/det_db \
--image_dir=../../doc/imgs/12.jpg \
......@@ -240,7 +250,7 @@ Specifically,
--rec=false
```
##### 4. cls+rec：
##### 4. 分类+识别：
```shell
./build/ppocr --rec_model_dir=inference/rec_rcnn \
--cls_model_dir=inference/cls \
......@@ -251,7 +261,7 @@ Specifically,
--cls=true \
```
##### 5. rec
##### 5. 识别：
```shell
./build/ppocr --rec_model_dir=inference/rec_rcnn \
--image_dir=../../doc/imgs_words/ch/word_1.jpg \
......@@ -261,7 +271,7 @@ Specifically,
--cls=false \
```
##### 6. cls
##### 6. 分类：
```shell
./build/ppocr --cls_model_dir=inference/cls \
--cls_model_dir=inference/cls \
......@@ -272,62 +282,63 @@ Specifically,
--cls=true \
```
More parameters are as follows,
更多支持的可调节参数解释如下：
- Common parameters
|parameter|data type|default|meaning|
| --- | --- | --- | --- |
|use_gpu|bool|false|Whether to use GPU|
|gpu_id|int|0|GPU id when use_gpu is true|
|gpu_mem|int|4000|GPU memory requested|
|cpu_math_library_num_threads|int|10|Number of threads when using CPU inference. When machine cores is enough, the large the value, the faster the inference speed|
|enable_mkldnn|bool|true|Whether to use mkdlnn library|
|output|str|./output|Path where visualization results are saved|
- 通用参数
|参数名称|类型|默认参数|意义|
| :---: | :---: | :---: | :---: |
|use_gpu|bool|false|是否使用GPU|
|gpu_id|int|0|GPU id，使用GPU时有效|
|gpu_mem|int|4000|申请的GPU内存|
|cpu_math_library_num_threads|int|10|CPU预测时的线程数，在机器核数充足的情况下，该值越大，预测速度越快|
|enable_mkldnn|bool|true|是否使用mkldnn库|
|output|str|./output|可视化结果保存的路径|
- forward
- 前向相关
|parameter|data type|default|meaning|
|参数名称|类型|默认参数|意义|
| :---: | :---: | :---: | :---: |
|det|bool|true|前向是否执行文字检测|
|rec|bool|true|前向是否执行文字识别|
|cls|bool|false|前向是否执行文字方向分类|
- Detection related parameters
- 检测模型相关
|parameter|data type|default|meaning|
| --- | --- | --- | --- |
|det_model_dir|string|-|Address of detection inference model|
|max_side_len|int|960|Limit the maximum image height and width to 960|
|det_db_thresh|float|0.3|Used to filter the binarized image of DB prediction, setting 0.-0.3 has no obvious effect on the result|
|det_db_box_thresh|float|0.5|DB post-processing filter box threshold, if there is a missing box detected, it can be reduced as appropriate|
|det_db_unclip_ratio|float|1.6|Indicates the compactness of the text box, the smaller the value, the closer the text box to the text|
|det_db_score_mode|string|slow| slow: use polygon box to calculate bbox score, fast: use rectangle box to calculate. Use rectangular box to calculate faster, and polygonal box more accurate for curved text area.|
|visualize|bool|true|Whether to visualize the results，when it is set as true, the prediction results will be saved in the folder specified by the `output` field on an image with the same name as the input image.|
|参数名称|类型|默认参数|意义|
| :---: | :---: | :---: | :---: |
|det_model_dir|string|-|检测模型inference model地址|
|max_side_len|int|960|输入图像长宽大于960时，等比例缩放图像，使得图像最长边为960|
|det_db_thresh|float|0.3|用于过滤DB预测的二值化图像，设置为0.-0.3对结果影响不明显|
|det_db_box_thresh|float|0.5|DB后处理过滤box的阈值，如果检测存在漏框情况，可酌情减小|
|det_db_unclip_ratio|float|1.6|表示文本框的紧致程度，越小则文本框更靠近文本|
|det_db_score_mode|string|slow|slow:使用多边形框计算bbox score，fast:使用矩形框计算。矩形框计算速度更快，多边形框对弯曲文本区域计算更准确。|
|visualize|bool|true|是否对结果进行可视化，为1时，预测结果会保存在`output`字段指定的文件夹下和输入图像同名的图像上。|
- Classifier related parameters
- 方向分类器相关
|parameter|data type|default|meaning|
| --- | --- | --- | --- |
|use_angle_cls|bool|false|Whether to use the direction classifier|
|cls_model_dir|string|-|Address of direction classifier inference model|
|cls_thresh|float|0.9|Score threshold of the direction classifier|
|cls_batch_num|int|1|batch size of classifier|
|参数名称|类型|默认参数|意义|
| :---: | :---: | :---: | :---: |
|use_angle_cls|bool|false|是否使用方向分类器|
|cls_model_dir|string|-|方向分类器inference model地址|
|cls_thresh|float|0.9|方向分类器的得分阈值|
|cls_batch_num|int|1|方向分类器batchsize|
- Recognition related parameters
- 识别模型相关
|parameter|data type|default|meaning|
| --- | --- | --- | --- |
|rec_model_dir|string|-|Address of recognition inference model|
|rec_char_dict_path|string|../../ppocr/utils/ppocr_keys_v1.txt|dictionary file|
|rec_batch_num|int|6|batch size of recognition|
|参数名称|类型|默认参数|意义|
| :---: | :---: | :---: | :---: |
|rec_model_dir|string|-|识别模型inference model地址|
|rec_char_dict_path|string|../../ppocr/utils/ppocr_keys_v1.txt|字典文件|
|rec_batch_num|int|6|识别模型batchsize|
|rec_img_h|int|32|识别模型输入图像高度|
|rec_img_w|int|320|识别模型输入图像宽度|
* Multi-language inference is also supported in PaddleOCR, you can refer to [recognition tutorial](../../doc/doc_en/recognition_en.md) for more supported languages and models in PaddleOCR. Specifically, if you want to infer using multi-language models, you just need to modify values of `rec_char_dict_path` and `rec_model_dir`.
* PaddleOCR也支持多语言的预测，更多支持的语言和模型可以参考[识别文档](../../doc/doc_ch/recognition.md)中的多语言字典与模型部分，如果希望进行多语言预测，只需将修改`rec_char_dict_path`（字典文件路径）以及`rec_model_dir`（inference模型路径）字段即可。
The detection results will be shown on the screen, which is as follows.
最终屏幕上会输出检测结果如下。
```bash
predict img: ../../doc/imgs/12.jpg
......@@ -339,7 +350,7 @@ predict img: ../../doc/imgs/12.jpg
The detection visualized image saved in ./output//12.jpg
```
<a name="3"></a>
## 3. FAQ
1. Encountered the error `unable to access 'https://github.com/LDOUBLEV/AutoLog.git/': gnutls_handshake() failed: The TLS connection was non-properly terminated.`, change the github address in `deploy/cpp_infer/external-cmake/auto-log.cmake` to the https://gitee.com/Double_V/AutoLog address.
1. 遇到报错 `unable to access 'https://github.com/LDOUBLEV/AutoLog.git/': gnutls_handshake() failed: The TLS connection was non-properly terminated.`， 将 `deploy/cpp_infer/external-cmake/auto-log.cmake` 中的github地址改为 https://gitee.com/Double_V/AutoLog 地址即可。
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment