Design And Application Of Lightweight Neural Network Based On Multi-dimensional Architecture Search

With the emergence of a series of novel network structures designed by humans,deep learning has made continuous breakthroughs in various fields.However,the shortcomings such as complex network structure,large amount of model parameters,excessive prior knowledge and the difficulty in deploying edge devices are also increasingly prominent.As one of the important sub-fields of automatic machine learning,neural network architecture search has attracted much attention in recent years.With the goal of automatically designing lightweight models that meet the requirements of resource-constrained equipment,this thesis focuses on the evolutionary search method based on freely scaling strategy and the architecture search method based on differentiable space from multiple dimensions of the network structure.At the same time,it designs and implements a model self-design system for fire identification and personnel detection.The main work of this thesis includes the following four parts:(1)Aiming at the problem of resource-constrained edge devices,this thesis optimizes the evaluation index function and search process of the neural network architecture search.It no longer only takes the quantity of model parameters as an indicator of whether the network is lightweight,but introduces the actual reasoning time to guide the algorithm to find a more suitable network structure for resource-constrained devices.At the same time,the search process is accelerated and optimized,such as search agent,to reduce the search time.(2)In view of the problem that traditional NAS generally searches the entire super-network or focuses on designing novel structures in image recognition,this thesis proposes a free scaling strategy based on the original structure of the network itself and multiple dimensions,which greatly enriched the search space.At the same time,the search strategy based on evolutionary algorithm is adopted to automatically design a network structure with good performance.(3)Aiming at the problem that the search part of NAS in the target detection task is single and difficult to converge,this thesis proposes a three-layer differentiable structure unit including operation category,network width,and activation function,which can realize the end-to-end search of the entire target detection network.The results have achieved a good balance between the lightweight of the model and the accuracy of the model.(4)Based on the work(2)and(3),for edge devices,this thesis designs and implements a NAS system.According to the lightweight requirements of users,the system automatically designs a network structure that meets the requirements,and provides users with a network model with trained parameters.At the same time,the application verification and test of the model are carried out on the UAV platform.The two algorithms proposed in this thesis have achieved a good balance between model lightweight and model accuracy on CIFAR100 and COCO data sets respectively,which is advanced and innovative to a certain extent.At the same time,the designed NAS system can efficiently search neural network models that meet the needs of edge devices,which has certain practical significance.