Research On High-efficiency Compression Technology For Convolutional Neural Networks

With the improvement of computer hardware capabilities and the advent of the era of big data,deep learning technology has developed rapidly.Compared with other intelligent algorithms in machine learning,deep learning models use a more complex model structure and a larger amount of data to simulate more problems in actual scenarios.As a representative algorithm of deep learning,Convolutional Neural Network is widely used in natural language processing,image recognition,target detection and other fields.With the increase of task complexity,the structure of the model will usually be more complicated.For example,the number of model layers and the number of parameters will increase exponentially,consuming a large amount of storage and computing resources,making it difficult to deploy to hardware resourceconstrained Embedded devices limit the further development of Convolutional Neural Network.In order to deal with this dilemma,this thesis proposes a Convolutional Neural Network compression pruning scheme based on automated search,which reduces the consumption of computing resources and memory space while ensuring the models predictive performance.The public image data sets and actual vulnerabilities The algorithms effectiveness is verified on th e detection data set,and good results have been achieved.Specifically,the main contributions of this article are as follows:(1)Propose a channel pruning algorithm that automatically searches the pruning structure of the model.Firstly,the target constraints are increased through hyperparameters,and the vast optimization structure combination is reduced.Then,the search process is regarded as an optimization problem,and the sparrow search algorithm is used to automate the search and optimize the pruning structure.The algorithm realizes model compression and acceleration by determining the number of channels in each layer of the neural network,which can reduce manual operations and increase the compression rate of the model,avoiding the selection of important channels based on rules or experience and realizing automated search selection.(2)We experimentally verify the channel pruning algorithm proposed in this paper on ResNet-56 and VGG-16,two convolutional neural networks,and compare it with several other classic compression algorithms.For ResNet-56,the accuracy is reduced by no more than 1%,the parameter compression rate reaches 55.27%,and FLOPs are reduced by 57.58%.For VGG-16,when the accuracy is reduced by no more than 0.5%,the parameter compression rate goes 80.62%,and FLOPs are reduced by 70.35%.(3)To verify the effectiveness and practicability of the algorithm,apply the algorithm to the field of software vulnerability detection.By analyzing the classification effect,model parameter amount,FLOPs number,and channel number of the ResNet-34 convolutional neural network model in the binary file before and after compression,it is proved that the channel pruning algorithm is applied while ensuring the detection rate of the vulnerability detection model,it can achieve model compression and acceleration and get a lightweight vulnerability detection model.