基于神经网络的嵌入式系统设计及模型部署方法

The design of an embedded system for neural networks and the software development process

为了应对嵌入式终端智能化应用需求,本研究研制了一种基于Hi35XX SoC的智能处理系统。本研究完成了该嵌入式系统的硬件设计与软件设计。在软件系统的应用软件层面,提出了一种将卷积神经网络在Hi35XX的AI加速引擎NNIE上的部署方法,解决了分割计算层和上采样计算层在Pytorch、onnx、Caffe、wk四种不同框架下的格式转化冲突。软、硬件系统的建立使该系统具有目标检测智能处理能力。

In this paper,an embedded system,which is based on the Hi35XX SoC,is developed to support the intelligent application for terminal equipment.The hardware circuits and its corresponding software system are completed in this research.On the level of application in software system,a method of processing convolutional neural networks on NNIE,which serves as an AI hardware accelerating engine in Hi35XX,is proposed solving the conflictions in model conversion among several frameworks including Pytorch,onnx,Caffe and wk,especially when Split/Slice layer and Upsample layer are involved in a network model.The embedded system is capable of intelligently performing object detection with the accomplishment of the hardware and software system.