Research On Multi-grained Pruning Algorithm Of Convolutional Neural Networks

In recent years,with the rapid development of deep learning technology,deep neural network has made revolutionary breakthroughs in image classification,target detection,image segmentation and other computer vision tasks,and has reached and exceeded the limits of human perception.At the same time,the "wisdom" of the new generation of AIoT is gradually shifting from the centralized high-performance processing platform to the "edge" node,and the industry calls for improving the intelligence level of the resource-limited edge node is also increasingly high.However,at present,the number of layers,the number of parameters and the calculation amount of high-precision network model are very large,and at the same time,it has a large redundancy,which makes it difficult to be placed in the edge nodes with limited resources.Therefore,how to compress and accelerate the model of deep neural network so that it can meet the need of intelligent and real-time application of edge nodes has become a hot issue in the field of deep learning.Compression and acceleration techniques based on deep neural networks can be roughly divided into network model optimization and hardware-based acceleration.For the former,previous studies have included designing lean network models and network pruning,Network pruning technology is directly applied to the existing high-precision network model,which is flexible and efficient.However,filter level pruning is not complete for network pruning,and connection level pruning makes the network model difficult to converge after pruning.In order to better learn from the existing depth model results,this paper proposes a multi-grained convolutional neural network pruning framework based on the idea of filter level and connection level pruning,and designs and implements the corresponding training strategy,which is compatible with the advantages of the two pruning strategies.For the latter,it is usually necessary to expand it according to specific application requirements.In this paper,the depth model in the human action recognition algorithm is optimized and deployed on the new embedded cpu-gpu heterogeneous computing module(NVIDIA Jetson TX2),which effectively improves the intelligence level of terminal nodes.The main work of this paper can be summarized as the following three aspects:1.This paper analyzes the filter level pruning and connection level pruning,and proposes a multi-granularity convolutional neural network pruning framework and corresponding training strategy to solve the problems existing in the two methods.The proposed network pruning algorithm selects AlexNet and VGG16 network for experimental verification,and the experimental results show that the proposed method significantly improves the network compression rate under the premise of better maintaining the network accuracy.2.This paper proposes a human action recognition algorithm based on OpenPose and makes the corresponding coordinate data set of human key points.The algorithm is applied to the intelligent recording system of classroom-based high-quality courses.The correctness and effectiveness of the algorithm are verified by the teacher's blackboard writing action recognition experiment in the real scene.3.This paper uses the multi-grained convolutional neural network pruning framework to optimize the deep network model of the proposed human action recognition algorithm based on OpenPose,and transplants the optimized algorithm into the embedded heterogeneous computing core module(TX2)for experimental verification.Experimental results show that the optimized human motion recognition algorithm can effectively meet the practical application needs.