Research On Convolutional Neural Network Acceleration Based On Tensor Decomposition

Convolutional neural network is an important artificial intelligence technique.Compared with traditional algorithms,it shows great advantages in many fields,such as computer vision,natural language processing,etc.,and is widely used in various life scenarios.Although the convolutional neural network has excellent performance,the huge amount of calculations and parameters in the network limit its application to mobile phones,Io T terminal devices and other hardware with limited computing and storage capabilities.Under this background,this thesis studies various neural network compression and acceleration algorithms,and proposes a convolutional neural network acceleration algorithm based on tensor decomposition.In addition,when accelerating multiple convolutional layers,the thesis proposes two Tucker rank selection methods to determine the acceleration ratio of different convolutional layers properly.The main work and contributions of this thesis include:1.A convolutional neural network acceleration method is proposed,which joints Tucker decomposition with CP(CANDECOM/PARAFAC)decomposition,called as Tucker-CP decomposition.This method first uses Tucker method to decompose the convolution kernel into a kernel tensor and factor matrices,and then uses the CP method to decompose the kernel tensor obtained by Tucker decomposition into multiple rank 1 tensors.Experiments show that when VGG is accelerated by 8times,the classification accuracy drop caused by Tucker-CP decomposition is only2.065%.Compared with using Tucker decomposition alone and using CP decomposition alone,the classification accuracy drop is reduced by 26.36% and50.62%,respectively.2.A Tucker rank selection method is proposed.This method adjusts Tucker ranks obtained by Variational Bayesian Matrix Factorization(VBMF),so it is called VBMF-V(VBMF Variant)method.Considering the differences in the sparsity and the impact on network performance of different convolution kernels,VBMF-V first uses VBMF to obtain Tucker ranks Which reflect the sparsity of each convolution kernel,then decomposes the convolutional kernel using the Tucker rank obtained by VBMF and calculates the classification accuracy drop caused by unit calculation reduction,finally adjusts Tucker ranks based on the accuracy drop.Experiments show that when Alex Net is accelerated by 4 times with ranks selected by VBMF-V,the classification accuracy drop of accelerated network is 3.63%.Compared with speeding up each convolutional layer at same ratio,the accuracy drop is reduced by13%3.Another Tucker rank selection method is proposed,which based on a function fitting accuracy drop with singular value proportion(FAS).Considering that with the change of the Tucker rank,the accuracy drop caused by unit calculation reduction also changes,this method adopts an iterative algorithm.In each iteration,the method uses FAS function to calculate the accuracy drop caused by unit calculation reduction when the Tucker rank decreases by 1,then select the Tucker rank that needs to be adjusted based on the accuracy drop.Experiments show that when Alex Net is accelerated by 4 times with ranks selected by FAS method,the classification accuracy drop of accelerated network is 3.13%.Compared with speeding up each convolutional layer at same ratio,the accuracy drop is reduced by 25%.