Neural network lightweight deployment method based on three-objective joint optimization

Abstract

The invention relates to a neural network lightweight deployment method based on three-objective joint optimization, and the method comprises the steps: mounting a terminal device on an edge side device which is provided with a calculation unit and a connection contact capable of achieving data transmission, inputting an initial network model M and an objective function into the calculation unit, pruning the initial network model M to obtain an optimized optimal network model M '; transmitting the optimal network model M'to a terminal device through a connection contact; acquiring an image sequence of the to-be-recognized target in real time by using a camera device on the terminal equipment; and inputting the image sequence into the optimal network model M ', and outputting an identification result. According to the method, the pruned and optimized lightweight network model is deployed on the terminal equipment in a low-calculation-power or low-energy-consumption-required scene, so that lightweight operation of the terminal equipment is realized, and the limitation that a large-scale neural network model cannot be deployed on the terminal equipment with limited performance and electric quantity is overcome.

Description

technical field [0001] The invention relates to the technical field of neural network compression in edge intelligence scenarios, in particular to a lightweight deployment method of neural networks based on three-objective joint optimization. Background technique [0002] At present, Convolutional Neural Networks (CNN) have achieved great success in tasks such as computer vision and natural language processing, and have been applied in many practical applications. For example, UAV equipment, etc., UAV equipment can be used to perform monitoring tasks such as fire rescue, cyanobacteria outbreak, and watershed pollution. When the UAV equipment performs tasks, it needs to capture the target image through the camera, and input the target image into the convolutional neural network for target recognition. In order to improve the accuracy of target recognition, the layers of this type of convolutional neural network are deeper and the structure is more and more complex. Such a m...

AI Technical Summary

Problems solved by technology

Therefore, most of the current pruning algorithms are structured pruning of network models, but structured pruning needs to set hyperparameters for each layer, and when pruning, the network needs to iterate multiple times to converge

Non-Patent Document 1 (Shangqian Gao, Feihu Huang, Jian Pei, and Heng Huang. Discrete modelcompression with resource constraint for deep neural networks. In Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition , pages1899–1908, 2020.) Using discrete particle swarm optimization and differentiable discrete gates to search for the best precision subnetwork under a given FLOPs budget, its disadvantage is that it searches for the solution with the highest precision under a given FLOPs budget, which is a For single-objective optimization, since FLOPs are fixed, the optimal value of FLOPs cannot be adjusted during optimization, which leads to poor generalization

Images