Research On Channel Pruning Algorithm Of Convolutional Neural Networks

In recent years,the technology about convolutional neural networks(CNN)has been widely used in various scenarios such as face recognition,pedestrian detection,and automated production.However,due to the large demand for computing and storage resources of the CNN models,it is difficult to directly deploy the original CNN models on some mobile devices and edge computing devices with limited resources,which limits the application scope of CNN.Therefore,CNN compression technique is an important research direction for the actual applications of CNN.Among many CNN compression techniques,the channel pruning algorithm is widely used in the actual CNN compression task because it can reduce the size of the existing model without changing the computational method of the convolutional neural network.However,the existing channel pruning algorithms usually have the problem of channel measure locality in the stage of channel importance measurement or have the problem of adopting fixed channel pruning rate in the stage of channel pruning strategy.Besides,when pruning the CNN models containing BN layers,most channel pruning algorithms adopt the strategy of recalculating the mean and the variance in the BN layer to improve the performance of the pruned model,but lack of analysis on the causes of this phenomenon.This article studies these three issues,and proposes the fast neural importance propagation algorithm(F-NISP),automatic channel pruning algorithm based on multi-objective optimization algorithm(APMO),and theoretical analysis of the impact of the fixed mean and variance of the BN layer on the performance of the pruning model(referred to as BN pruning effect)to solve the above problems.The research content is as follows:(1)Aiming at the locality problem of the existing channel importance measurement algorithm,the Neural Importance Score Propagation(NISP)algorithm can obtain the CNN's global channel importance assessment in a layer-by-layer manner,but its calculation efficiency is slow.Therefore,by studying the principle of the NISP algorithm,the sparse matrix multiplication calculation method in the original NISP algorithm is replaced by a transposed convolution calculation method.And the F-NISP algorithm is proposed.At the same time,the F-NISP algorithm also converts the deep separable convolution structure into a traditional convolution structure,thereby simplifying the calculation process.And the experimental results show that the compute efficiency of the F-NISP algorithm is greatly improved compared to the NISP algorithm while maintaining the NISP algorithm's ability to measure the global channel importance of the CNN.(2)Aiming at the problem of adopting a fixed channel pruning rate in the step of channel pruning strategy.According to the inconsistent sensitivity to channel pruning at different layers of CNN,the channel pruning problem is modeled as a multi-objective optimization problem,which takes the accuracy performance and resource consumption of the pruned models as two optimization goals and takes the pruning rate of each layer as the decision variable.And the multi-objective optimization problem is solved through a multi-objective evolutionary algorithm based on decomposition.On this basis,the APMO algorithm is put forward.Compared with the existing channel pruning strategy,the APMO algorithm can automatically obtain a series of Pareto optimal pruning models in terms of model size and accuracy,thereby facilitating the analysis of the redundancy of the convolutional neural network model.The experimental results show that the APMO algorithm can obtain a more accurate pruning model with the same amount of calculations.(3)Aiming at the lack of theoretical analysis of the BN pruning effect in the existing CNN channel pruning algorithms,a theoretical analysis of the BN pruning problem is carried out by using the formula derivation method,and the conclusion of the analysis is as follow: the mean and variance of the corresponding original CNN model and the expected real mean and real variance expected of the pruned model are different,so that the output feature map distribution and the expected distribution are inconsistent,resulting in a significant decrease in the performance of the pruned CNN model.And the experimental results are consistent with the conclusion of the theoretical analysis of the BN pruning problem.