A Study On Fast Neural Architecture Search For Affinity Chip Microarchitecture

In recent years,the rapid development of computer technology makes deep neural networks widely used in artificial intelligence(AI).The popularity of smart terminals and the development of 5G technology,also make it easier to obtain big data such as pictures and videos.The applications of data-driven artificial intelligence technology are gradually expanded.Therefore,it is of practical significance to design high-performance deep neural networks according to different application scenarios.The performance indicators of deep neural networks include accuracy,calculation amount,the number of parameters,reasoning speed,and so on.The key factors affecting the performance of deep neural networks mainly include :(1)Neural network structure design: in the process of designing the neural network structure,it is necessary to balance the accuracy,calculation amount,and the number of parameters.(2)The hardware platform where the deep neural network operates: different hardware devices carry different chip microarchitectures,which makes the same computing unit operate at different speeds on different hardware devices.Therefore,it is necessary to design a high-performance neural network structure that is compatible with the hardware platform.Neural Architecture Search(NAS)has also become one of the important problems in the context of Automated Machine Learning(Auto ML).This thesis focuses on how to apply the neural network structure search method to the task of the image classification and object detection in computer vision.The main contributions include:(1)The NAS method is designed for the affinity chip microarchitecture based on the One-shot model and evolutionary algorithm.Experimental verification is conducted in Huawei Atlas 300.Experimental results show that compared with popular image classification models Res Net50,Mobile Netv2,and object detection model YOLOV3,the neural network structure got by the proposed method can significantly improve the reasoning speed with the same accuracy.(2)In the image classification task,a high-performance search space of the neural network structure is constructed,and a dynamic and hierarchical performance evaluation method of candidate neural network modules is proposed to guide the training process of the hyper network to improve search efficiency.The real-time hardware-sensing search strategy based on affinity chip microarchitecture is adopted.(3)In the object detection task,some strategies of neural network structure search for the image classification task are applied to the target detection task.Combining with the characteristics of target detection tasks,a supernet design scheme of the single-stage object detection network structure is proposed,and a method to search network structure of FPN is proposed according to the target resolution in application scenarios.This thesis focuses on the application of the neural network structure search method in the image classification task and object detection task.Several improved methods are proposed,and the effectiveness of the method is proved by experiments.