Research And Application Of Structured Model Pruning Algorithm In Deep Convolutional Neural Networks

Network pruning has proven to be an extremely effective approach to compress convolutional neural networks.Pruning more redundant network structures while ensuring network accuracy is challenging.For better performance of computer vision tasks,designing deeper and broader networks has become the most common way.However,these methods increase dramatically float point operations,parameters and memory footprint,which restricts their usage on a mobile or embedded device where computation and storage are limited.When it is difficult to improve the performance of hardware resources,model compression has received extensive attention.This paper conducts analysis on the structured pruning and studies the application of this technology in pedestrian detection on embedded devices.The main research results obtained are as follows:(1)Aiming at the problem that iterative pruning the “least important” filters and retraining the pruned network layer-by-layer may consume significant computational resources and time.an end-to-end structured network pruning method based on adversarial multi-indicator architecture selection is presented.The pruning is implemented by striving to align the output of the baseline network with the output of the pruned network in a generative adversarial framework.Furthermore,to efficiently find optimal pruned architecture under constrained resources,an adversarial fine-tuning network selection strategy is designed,in which two contradictory indicators,namely pruned channel number and network classification accuracy,are considered.Experiments on SVHN show that the method reduces75.4% of FLOPs and 74.4% of parameters with even 0.36% accuracy improvement for Res Net-110.On CIFAR-10,it obtains a 56.9% pruned rate in FLOPs and 59.2% parameters reduction,while with an increase of 0.49% accuracy for Res Net-110.(2)For iterative pruning,after changing the network structure for the first time,only through a small amount of retraining and pruning again,it is easy to form a "sub-optimal iterative sub-optimal solution" problem.A convolutional neural network compression method based on adaptive layer entropy is proposed.Firstly,the retention rate of each convolutional layer filter is directly determined based on the correlation between layer entropy,taking the convolutional layer as a whole.It belongs to an end-to-end structured network pruning and avoids information loss due to iterative pruning.Then,the pruned network is retrained by adaptive joint grafting of the convolutional and batch normalization layers to learn more information from the network,in which the layer entropy used in the compression process is considered.Extensive experiments are conducted to demonstrate the efficiency and superiority of the proposed method.On CIFAR-10,it obtains a 56.9% pruned rate in FLOPs and 59.2% parameters reduction,while with an increase of 0.49% accuracy for Res Net-110.For Res Net-56,compared with the baseline network,the accuracy is improved by 1% when FLOPs is compressed by 36.2%.It reduces 55.2% of FLOPs with even 1.29% accuracy improvement for Mobile Net V2.(3)Aiming at the problem that only using layer entropy to measure layer information richness is not perfect and does not deal with extreme values of indicators.A convolutional neural network compression method based on abnormal information richness constraints is proposed,considering a global network structure.First,the convolutional layer is regarded as a whole,and the extreme value of the layer entropy and average L1 norm is processed by the box plot method.Then,the processed layer entropy is fused with the average L1 norm to measure the layer information richness.Finally,the correlation between the layer information richness is used to determine the retention rate of each convolutional layer filter,and the pruned network is retrained using the adaptive joint grafting method to improve the network performance.Extensive experiments on different datasets for mainstream convolutional neural networks show the superiority,especially in networks with short connections.On CIFAR-10,it obtains a 53.9% pruned rate in FLOPs and 58.4% parameters reduction,while with an increase of 1.56% accuracy for Mobile Net V2,which significantly outperforms state-of-the-art methods.(4)A pedestrian detection system is built on the embedded device Jetson Nano.First,the PASCAL VOC 2007 dataset is used for model training and pruning to obtain the original model and the pruned model.Then the system is built and debugged on the server side.Finally,the system is embedded in the embedded device.The system is tested in real scenarios,which verifies the effectiveness of the research content in practical applications.