| Deep convolutional neural network has become a popular technique in computer vision.Generally,the widening and deepening convolutional neural network model can achieve better performance,but it leads to a rapid increase in the FLOPs and parameters of convolutional neural network models.In practical application,the deployment of depth model still has some problems due to the limited amount of the FLOPs and parameters,so the research on reducing the scale through network pruning and compression becomes the primary goal.Two effective pruning methods were proposed in this paper to solve the problems of weak theoretical basis and artificial presetting of pruning rate for each layer in the field of classification model pruning.1.Adaptive global pruning method based on deep learning.In this method,an activation weight generation module(AWGM)adaptively obtained the global activation weight of each filter in the network from the input feature graph,and selected the important filters by global sorting while pruning the less important filters.On this basis,aiming at the high specificity of the deep part of the network structure,the classbased pruning was carried out to retain the filters with large differences in response to different types of input images.Experiments on AUC data sets showed that the accuracy of the two methods was reduced by no more than 1.1% before and after pruning on Res Net,but only about 50%of the FLOPs and parameters of the original model were retained.2.Multi-branch pruning method based on deep learning.The existing single-layer attention module may be lost in the next layer due to the details of the upper layer in the process of network transmission,so it has problems of insufficient learning ability and difficulty in adjusting attention to key areas.Therefore,this method used deep connection channel attention mechanism,the model of joint training,all of the modules and designs the connection on a layer of output layer and the current double branch structure,the fusion of shallow and deep information,improve the learning ability of the focus module,maked different channel more sensitive response to input information;Combined with the adaptive global pruning method proposed in this paper,better model pruning effect is achieved.In addition,this method was not limited by the basic network architecture and has good generalization.Experiments on AUC data set showed that the accuracy of the proposed method was reduced by less than 0.6% before and after pruning on VGG-19,but only about 20%of the FLOPs and 60% of the parameters of the original model were retained. |
