Research On Optimization Technology Of Deep Learning Model For Edge Computing

With the gradual maturity of 5G technology and the popularization of its commercial application products,edge computing,which had been developed for several years,has further become a research hotspot.The fusion of artificial intelligence technology and edge computing has attracted intense interest from researchers.However,memory-consuming and computationally intensive artificial intelligence algorithms are in great conflict with resource-starved edge terminal hardware.This paper analyzes the key technologies of model compression,model transplantation,automatic coding,and model storage for the integration requirements of edge computing and deep learning,and on this basis,proposes and implements an integrated system for model transformation and transplantation.The main research contents and innovations of this article include:(1)A method of deep learning model parameter extraction and automatic filling is proposed.The model disassembly method is used to extract the parameters in the model to lay the foundation for the deep learning model to get rid of environmental variables and dependent function libraries,and then to analyze the model parameter types and quantities in detail to form a parameter dictionary to facilitate automatic parameter filling.The specific expression language of the inference algorithm,See Dot,is studied,which can reduce the model size and increase the model calculation speed to a certain extent.(2)A unified model conversion and transplantation system MoTransFrame is proposed.Through in-depth analysis of the calculation process of the deep learning classification algorithm,the "automatic coding" method is used to re-encode the extracted model parameter dictionary according to the classification rules.According to the initial network calculation graph output,the classification model of the resource-poor edge device is adapted.MoTransFrame builds a unified model conversion integrated system,which provides a feasible solution for the transplantation of deep learning models to resource-constrained edge nodes.(3)A model storage optimization method based on the storage requirements of edge nodes is proposed.For the Arduino device,a storage method combining multiple types of storage media is proposed to maximize the use of the limited storage resources of the edge device.The model is read through the address binding method,which saves memory usage and also guarantees the completeness of model storage and the correctness of reasoning,so that the transplanted model can run accurately on the edge terminal.