Font Size: a A A

Research And Implementation Of Convolutional Neural Network Compression Technique

Posted on:2020-09-29Degree:MasterType:Thesis
Country:ChinaCandidate:X J WangFull Text:PDF
GTID:2428330575457129Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
With the research and development of deep learning,convolutional neural networks have made breakthroughs in many fields like computer vision.However,there are large number of parameters in the convolutional neural network,which takes up large amounts of computing resources at runtime and brings challenges to the deployment of the model on mobile devices with limited resources.Therefore,the convolutional neural network pruning technology has been widely studied in order to reduce the computational resource consumption of the model.In the related research,pruning methods based on parameter importance and parameter sparsity are two important research directions.The main contributions of this thesis are as follows:This thesis presents an algorithm for measuring the importance of parameters based on attention mechanism.With the help of attention module,the algorithm applies weights to the output channels of filters in the same layer,and calculates the importance of the filter to guide network pruning at the filter level.The experimental results show that the proposed algorithm can achieve higher accuracy under the same pruning ratio.In the aspect of pruning based on parameter sparsity,a global pruning algorithm based on regularization is proposed to control the pruning ratio of each layer in convolutional neural network adaptively.By applying L1 regularization to channel weights,channels with weights approaching 0 are removed.Convolutional neural networks performs channel-level feature selection to enhance the sparsity of the structure.In addition,aiming at sloving the performance deficiency of pruning based on mask operation in Tensorflow framework,this thesis proposes an effective pruning method to realize parameter reduction.The experimental results show that the global pruning method based on regularization can further improve the pruning effect.The application of plant recognition in mobile device based on compressed neural network is designed and implemented.The requirement analysis,system overall design,database design and system detailed design are carried out,and the function and performance tests are completed.The test results show that the application can complete plant species identification in off-line environment,and has the characteristics of simple operation and good performance.
Keywords/Search Tags:CNN, compression, pruning, attention mechanism, sparsity
PDF Full Text Request
Related items