Research On Channel Flexible Pruning On Deep Convolutional Neural Networks

The arrival of the era of big data brings huge amounts of data information,including video image information.Due to the specificity of the image data,it is difficult to analyze it efficiently using traditional statistical methods.Statistical machine learning realizes the analysis and prediction of massive data by computer,and now is widely used in practical problems such as image recognition,semantic segmentation,and object detection.In particular,neural networks have become mainstream methods for processing video and image data,more than other computer learning methods.The development of neural networks benefits from the emergence of different neural network models.However,as the neural network scale grows,depth deepens and complexity increases,and its computation and parameters grow.The rapid development of edge intelligence is in need of compressing neural networks to for edge devices with limited computational and storage resources.This paper presents a series of studies on model pruning,one of model compression methods.The specific contents of this article are summarized as follows:(1)In view of model pruning prone to excessive pruning,a progressive channel pre-pruning method based on batch normalization layer is proposed.Unlike traditional pruning methods which divide channel importance into two important and unimportant classes,our method defines an intermediate state between important and unimportant.By performing the pre-pruning operation of this part of the middle state channel,it effectively avoids excessive pruning and improves the performance of the model pruning.Experimental results on the CIFAR-10 and CIFAR-100 datasets demonstrate the effectiveness of the method by tailoring nearly half the computational amount of VGG16 and Resnet56 to achieve near or even exceeding the accuracy of the original model.Experiments on the Image Net dataset and Mobile Net V2 demonstrate specific scalability.(2)For the problem of too long model pruning process,we propose a two-step channel flexible pruning process.Traditional pruning often uses iterative methods to gradually crop the model and gradually approximate the target compression rate,resulting in a substantial computational cost of the pruning process.Our approach provides a more efficient pruning scheme that rapidly approaches the model to the target compression rate by smoothing the pre-pruning of the network.Subsequently,through the channel equivalent recombination process,the compression model architecture is quickly obtained.Experimental results based on VGG16 and Res Net56 in CIFAR-10 show that the flexible pruning strategy achieves floating-point operational compression ratios of 3.49 and 2.19 respectively,while the accuracy drops by only 0.15 percentage points and0.20 percentage points respectively,compared to the benchmark model.