Convolutional Neural Network Compression Algorithm Based On Tucker Decomposition Of Shared Factor Matrices

Convolutional Neural Network(CNN)is a very widely used neural network that has dominated various computer vision tasks such as image classification and object detection.With the continuous emergence of new application scenarios,more and more vision tasks become complex.To meet the demands of these tasks,the number of network layers as well as the number of parameters in CNN-based models is gradually increasing,which makes these tasks difficult to achieve on mobile devices with limited memory and computing power.Therefore,reducing the number of parameters and computational complexity of CNN-based models is crucial for practical use.In recent years,scholars have proposed many algorithms for compressing convolutional neural networks based on Tucker decomposition.In multi-task scenarios,these algorithms achieve model compression by performing independent Tucker decomposition on the convolution kernels in each CNN model.However,these algorithms usually select Tucker rank manually through time-consuming trial and error,which not only consumes a lot of energy but also makes it difficult to guarantee the compression effect and model accuracy.At the same time,since the correlation between multiple models is ignored,the compression rate of the model is not high.Therefore,this paper proposes a joint compression method for multi-task convolutional neural networks.The main results are as follows:(1)Instead of selecting Tucker rank by a time-consuming trial-and-error method,an automatic VBMF-based Tucker rank selection method is proposed.It mainly includes two steps: firstly,the two-dimensional expansion of the convolution kernel weight tensor is performed to obtain the two-dimensional expansion matrix of the convolution kernel weight tensor on the input channel and the output channel.Then use the VBMF method to decompose the two-dimensional expansion matrix to obtain the extreme Tucker rank of the convolution kernel weight tensor.In addition,since the factor matrix shared by multiple tasks requires the same Tucker rank,to achieve this,we further propose a relaxation-based algorithm to select the common Tucker rank shared by multiple tasks in J-Tucker.(2)A joint compression scheme(J-Tucker)is proposed to compress CNN models in multi-task scenarios.We utilize the core tensor obtained after Tucker decomposition to represent the specific information in each task,and utilize the shared factor matrix to represent the correlation and common features among multiple tasks,thus improving the compression ratio of the multi-task model.(3)A compression scheduling algorithm based on ARPR is proposed to select the convolution layer for compression.Compressing the convolution layer with a smaller ARPR value preferentially can bring greater gain in parameter compression,but has little impact on the accuracy,so that the compression model of the CNN model can preserve the information in the original CNN model to the greatest extent,and can realize the flexible compression scheduling of the CNN model to ensure the compression effect and accuracy of the model,especially considering the storage limitation of the target device.We have done extensive experiments on two neural networks(AlexNet and VGG-16)with two real datasets(CIFAR-10 and CIFAR-100)to evaluate the effectiveness of the proposed algorithms.The experimental results demonstrate that our J-Tucker can achieve significant reductions in model size and run-time,at the cost of a small loss in model accuracy.Especially,after performing our J-Tucker on AlexNet(or VGG-16)based multi-task neural networks,we can achieve up to 11.50 ×(or 8.55 ×)compression ratio.