智能算法模型转换工具的设计与实现  

Design and implementation of intelligent algorithm model conversion tool

作　　者：叶亚峰 王昕 寇金桥 李墈婧 YE Ya-feng;WANG Xin;KOU Jin-qiao;LI Kan-jing(Institute 706,Second Academy of China Aerospace Science and Industry Corporation,Beijing 100854,China)

机构地区：[1]中国航天科工集团第二研究院七〇六所,北京100854

出　　处：《计算机工程与设计》2023年第12期3819-3826,共8页Computer Engineering and Design

摘　　要：针对边缘计算中国产化异构智能加速硬件种类繁多,算法模型在不同硬件平台上运行环境的部署过程繁琐、上手难度大等问题,设计并实现一种智能算法模型转换工具。该工具基于QT开源框架开发,运用多线程和线程交互等技术,以界面的形式呈现给用户;支持华为昇腾(CANN)平台、百度飞桨(PaddlePaddle)平台和国产FPGA平台的模型转化功能,可将开源框架下训练的深度学习算法模型一键转换为国产化智能加速硬件平台支持的模型文件,实现智能算法模型的快速部署和高效运行。实验验证了该工具的可行性和有效性,转换前后的推理精度误差不超过1%。In view of the wide variety of home-made heterogeneous intelligent acceleration hardware in edge computing,the cumbersome deployment process of algorithm models in different hardware platforms and the difficulty of getting started,an intelligent algorithm model conversion tool was designed and implemented.The tool was developed based on the QT open source framework and technologies such as multithreading and thread interaction were used to present an interface.The model conversion tool supported Huawei CANN platform,Baidu PaddlePaddle platform and domestic FPGA platform,which allowing the deep learning algorithm model being trained under the open source framework tansformed to the domestic hardware platform in one click to realize the rapid deployment and efficient operation of the intelligent algorithm model.Experiments verify the feasibility and effectiveness of the tool,and the accuracy error before and after deployment is no more than 1%.

关 键 词：模型转换工具 华为昇腾 百度飞桨 智能加速 边缘计算 多线程技术 国产化 

分 类 号：TP311.56[自动化与计算机技术—计算机软件与理论]