Research And Prototype Implementation Of Key Technologies For Automatic Generation Of Deep Models For Light Computing Power

In recent years,various industries have relied on machine learning models for technological improvement and innovation.However,the task of model generation is very heavy,because manual model design for a specific task is very time-consuming and labor-intensive,and when training the model,the computer hardware configuration requirements are high,and the domain experts are highly dependent on the professional knowledge of the domain experts.The above reasons have hindered the current development of deep learning,so the research that makes the process of machine learning lightweight and automated has gradually become the focus of researchers.This subject revolves around the main research content of automatic machine learning,including two important links: automatic data enhancement and automatic model design.The research on generating in-depth models in a short time with limited experimental resources is carried out,with the purpose of changing the task of model generation.“light”.In addition,a platform for automatic generation of deep models for light computing power has been built,and the automatically generated models have been deployed to the mobile terminal.This subject focuses on the key technology of automatic generation of depth models for light computing power,focusing on the light computing power automatic data enhancement method based on differentiable gradient estimation,the light computing power automatic model design method and platform design that reduce the depth interval,and the main contents are as follows:(1)Research on automatic data enhancement methods based on light computing power.Different from the current more mature Auto Augment algorithm,this topic treats the search space of automatic data enhancement as continuous,and uses the sigmoid function to break the dilemma of mutual inhibition between multiple operations,and proposes the use of small batch gradients in the selection of search strategies The method of descent,and the use of gradient estimation techniques,proposes a new automatic data enhancement strategy for light computing power,which greatly reduces the time cost,and the quality of the generated data set is still the same as the quality of the data set generated by the Auto Augment algorithm.Comparability.(2)Research on automatic model design methods based on light computing power.Different from the current mature DARTS algorithm,this project divides the search process of neural network structure search into 4 stages,uses a feedback update mechanism,and uses early stopping and dropout to limit the number of jump connections.Reduce the depth interval between networks.At the same time,the type of candidate operation is optimized,a new loss function is designed,and a new automatic model design strategy oriented to light computing power is proposed,which greatly reduces the time of neural network structure search compared with DARTS,and produces The quality of the model is also comparable to that produced by DARTS.(3)Finally,a deep model automatic generation platform for light computing power is constructed,and the generated model is deployed to the mobile terminal.