Deep Neural Network Auto Compression Technology Using Dynamic Channel Ranking Strategy

Deep neural networks have excellent inferential predictive power,but the high requirement for matrix computing power limits the scope of use.In order to obtain easyto-use models,model compression has become a widely used technique in academia and industry.Compared with the original model,the compressed model can effectively reduce the storage space and chip computing power requirements,and only incur a small loss in inference accuracy.However,the following aspects of model compression are still problematic: when determining the sparsity of the network,the involvement of professionals generates additional manpower overhead,which is contrary to the purpose of saving costs and lowering the threshold of model usage;when distinguishing between important and redundant substructures,a single importance evaluation algorithm cannot be applied to a variety of model structures,which affects the inference capability of the compressed model.In order to solve the problems arising from human involvement,this thesis proposes an automatic model compression method.Combining model pruning,reinforcement learning,and neural architecture search,this thesis automates model compression by leaving the work that require professional participation to the deep reinforcement learning agent.By mapping the neural network model structure,the compression strategy,and the predictive power of the compressed model onto the state space,action space,and reward function of the deep reinforcement learning intelligent body,this method successfully empowers the intelligent body to automatically search and optimize the compression strategy,as well as to use pruning to obtain sub-models with excellent predictive power.To solve the problems caused by a single importance assessment method,this thesis proposes a dynamic channel ranking strategy that replaces a single importance assessment algorithm with a collection containing many different algorithms and uses the reinforcement learning agent to optimize the algorithm selection for different model structures during compression policy generation.In combination with the automatic model compression method,the intelligent body designed in this thesis is able to give a sparsity distribution close enough to the real situation and a suitable choice of importance evaluation algorithms for different parts of the model,and to obtain a compressed model with excellent prediction level by pruning.Comparative tests on various neural network model structures show that the method proposed in this thesis is able to achieve a stable model compression rate of 50.0%,with an improvement of 5.0% to 10.0% relative to other model compression methods.In terms of the accuracy loss generated by the compressed model,the accuracy loss of the method in this thesis is reduced by 1.0%?3.0% compared to other methods.According to the specification process of software development,this thesis implements the automatic model compression method incorporating the dynamic channel sorting algorithm as a prototype system.A series of test results show that this system has the capability to perform automated model compression.