Research On Deep Neural Network Model Compression Algorithm Based On Convolution Kernel Pruning

With the development of deep learning,natural language processing,computer vision and other work have made breakthrough progress.As an important branch in the field of computer vision,model compression task has also been widely concerned.Although the current model compression algorithm has achieved some achievements,there are still many problems.For example,most algorithms need to prepare a pretrained model,then select a convolution kernel importance measurement standard,cut the convolution kernel which is lower than the standard directly,and then carrying out the Fine-tune training to restore the accuracy of the model.This method has high requirements for the pretrained model,and the capacity of the model will be greatly reduced after pruning,which affects the accuracy of the compression model.This kind of algorithm can achieve better results in simple tasks,but when the task is a little complex,the accuracy and compression ratio will be greatly reduced.In order to solve these problems,we propose a new pruning while training model compression algorithm,which can keep the accuracy of the original model unchanged and get a higher compression ratio.Compared with the traditional pruning method,the algorithm has three characteristics: Firstly,through the modification of pruning method,the model compression does not need pretrained model and Finetune training operation,which improves the efficiency;Secondly,combining the training model and pruning operation,the training process of the model can be constrained by the "pruning" constraint method,which can train a compressed model from scratch;Thirdly,before the final pruning operation,all convolution kernels are updated normally,and the capacity of the model will not be reduced.Therefore,the accuracy of the original algorithm model can be maintained to the greatest extent,and the compression ratio of the model can be improved at the same time.In this paper,this method is applied to image classification and single-stage object detection tasks.The experimental results on several benchmark datasets(including CIFAR-10,ILSVRC-2012,PASCAL VOC and Microsoft COCO)show that the algorithm has achieved the current leading effect and can compress convolution network model more effectively.