AVH Deployment Practice (1) | Deploying the Flying Paddle Model on Arm Virtual Hardware

Model training and deployment has always beenAI应用的核心.Aiming at the application and challenges of deep learning algorithms in the process of industrialization,百度飞桨与Arm合作,针对Arm Cortex-MDeep learning application scenarios for processors,完成了飞桨模型在Cortex-MHardware adaptation.百度与Arm的此次合作,填补了飞桨模型在Cortex-MThe adaptation on the hardware is blank,同时也增加了Cortex-MNumber of deep learning models on hardware,More deployment options are provided for developers.

本次飞桨与Cortex-MThe adaptation is usedArm虚拟硬件(Arm Virtual Hardware, AVH)Model deployment as a platform.AVHDesigned to enable software development earlier than chip development,为物联网及嵌入式平台带来现代化敏捷的软件开发方法,包括持续集成/持续开发 (CI/CD)、DevOps与MLOps,而无需大型定制硬件集群.

本期,Xiaobian will take you to review Baidu飞桨The developers say it“Arm virtual hardware platform飞桨Model deployment in practice”The main content of the live course,Review with everyoneAVH上部署飞桨The end-to-end flow of the model.

This course is also the first to show how to飞桨Models are deployed directly in the containsArm Cortex-M处理器IP的AVH平台上.pass this course,You can not only learn in depth how to use飞桨Train for more efficient training and inference models,You will also learn how to compile with deep learningTVM将飞桨The model is compiled to fit inArm Cortex-M上运行的目标代码,and deploy it in the containing Cortex-M55的AVH平台上.

For more exciting content, please click the course playback link: 

https://www.bilibili.com/video/BV1Q34y1n74a?share_source=copy_web

This live broadcast course mainly covers the following five parts:

第一部分  课程概述

A brief introduction to the era of the Internet of Things,Various challenges faced by deep learning applications in the process of industrialization,并详细分析AVH如何与飞桨Combined to inspire more efficient MLOps工作流,This accelerates the development of edge-side machine learning applications.

图 1：AVH与PaddlePaddleCombined to inspire more efficientMLOps工作流

第二部分   PaddleDetection on Arm

Target detection tasks in the field of computer vision(Detection),Detailed introduction to Baidu飞桨的 PaddleDetectionFeatures and Benefits of the Development Kit,and explained in depthPP-PicoDet模型如何与Cortex-MUse cases for processor adaptation.

图 2：PaddleDetection Development kit panorama

第三部分PaddleOCR on Arm

Text recognition for computer vision(OCR),Baidu is introduced in detail飞桨的PaddleOCRFeatures and Benefits of the Development Kit,and explained in depthPP-OCRv3模型如何与Cortex-MUse cases for processor adaptation.

图 3：PaddleOCR Development kit panorama

 第四部分TVMC编译模型

A brief introduction to deep learning compilersTVMThe basic concept of , and its and 飞桨前端、CMSIS-NNBackend adaptation,and explain how to use itTVM编译器完成对飞桨Compile transformation for inference models,Make it fitCortex-Mcode running on the processor.

图 4：Schematic diagram of the compilation process

(了解TVM更多信息,请访问https://tvm.apache.org)

 第五部分   AVHto complete the model deployment 

简要介绍AVH的特点与优势,and used for this courseArm Corstone和  Cortex CPUFeatures of virtual hardware、其与Arm处理器IPThe relationship and the currently supported virtual hardware models are further introduced.Followed by a specific demonstration session,A step-by-step visual demonstration of how the PP-OCRv3English recognition model published in (After completing the operator adaptation) 部署在Arm Corstone-300on the virtual hardware platform (内含Arm Cortex-M55处理器,Arm Ethos-U55processor and some basic peripherals) .

图 5：AVH Corstone和Cortex CPU简介

In this issue, we mainly lead you to review the main course chapters.在下期推送中,We will take the task of text recognition in the field of computer vision(OCR)为目标,Step by step, I will lead you to complete the entire end-to-end development process from model training optimization to deep learning application deployment,敬请期待下一期内容：AVH部署实践 (二) | 在ArmDeploy on virtual hardwarePP-OCR模型.

不熟悉OCR？Want to unlock early to learn more about itOCRcutting-edge theory and code practice？

欢迎大家scan rightFang QR code pays attention to the public account,Get it for free from Baidu飞桨PaddleOCRThe team, together with colleagues from academia and industry technical experts, covers everything from detection and identification to document analysisOCR全栈技术书籍《动手学OCR》.

温馨提示: 

• This deployment hands-on course is hosted on AWS/AWS ChinaAmazon Machine MirrorAMI中的 Corstone和Cortex CPU虚拟硬件.

• 欢迎各位开发者扫描下方二维码,即刻注册AVHThird-party hardware platforms,抢先体验AVHThe ultimate convenience for your embedded development journey！There's more to comeAVHInteresting examples of development on third-party hardware platforms await.

• Use your company email or school email to register your application,Approval is more efficient！

  

关注【飞桨PaddlePaddle】公众号

获取更多技术内容~

本文同步分享在 博客“飞桨PaddlePaddle”（CSDN）.
如有侵权,请联系 [email protected] 删除.
本文参与“OSC源创计划”,欢迎正在阅读的你也加入,一起分享.