Research On Pruning Algorithm Of Neural Network Model Based On Reinforcement Learning

With the improvement of computer hardware,the field of deep learning,which has extremely high requirements for hardware computing capabilities,has also developed rapidly.However,with the deepening of various models and the diversification of application scenarios,the requirements of excellent deep learning models on hardware facilities have also increased sharply.Therefore,most deep learning models can only be performed on servers with super computing power or in the cloud.deploy.In the process of actual production and life,an excellent deep learning model cannot be effectively applied due to the limitation of the computing power of the edge hardware equipment.This topic takes the research of deep neural network model pruning technology as the goal,and the reinforcement learning method is an excellent tool to realize automatic model pruning.It evaluates the importance and redundancy of the filters in the neural network model.After the current mainstream soft pruning method,finally completed the research of neural network model pruning technology under the guidance of reinforcement learning.The main research is as follows:First,compare and analyze the advantages and disadvantages of mainstream pruning methods and traditional pruning methods in various aspects.Aiming at the problem of large loss of model accuracy after pruning in the current manual pruning methods,combined with cutting-edge reinforcement learning technology,Using the automatic model pruning method to give different compression ratios to different layers of the deep neural network and using the L2 norm as the filter importance guide,an automatic pruning algorithm based on filter importance ranking is completed.Second,filter pruning can compress the model very well,but it will inevitably come at the cost of a certain loss of accuracy.Therefore,for applications with high precision requirements,this paper abstracts the filter into a set of filters in space,and abstracts all filters in a certain layer of the convolutional neural network into a Euclidean space,and selects Among them,the filters that can be jointly represented by other filters perform pruning operations,thereby reducing the number of redundant filters in the neural network while avoiding accuracy loss.After training and pruning through reinforcement learning,its performance is restored through fine-tuning.Using this method for pruning and fine-tuning the model after completion,the amount of parameters and floating point calculations are greatly reduced while the accuracy loss is small.In order to further explore the rationality and feasibility of the proposed algorithm,an experimental comparison between the two methods and related algorithms is completed on a typical neural network model.Both methods can show good model pruning effects.At the same time,according to the computing capabilities of the existing edge devices,the actual deployment of the model after the pruning is completed is completed,and the model after the proposed algorithm pruning is completed also shows a good reasoning acceleration ability on the edge smart device.