The Design And Implementation Of Deep Learning Model And Data Set Hosting Platform

Recent years,deep learning has achieved breakthrough in speech recognition,computer vision and other areas.As the research fields become more and more extensive,more and more models are built.Researchers need to do a lot of contrast tests when they study in one of the fields,moreover,learning about the existing models and contrasting their results has become more and more complicated.However,existing computing platforms for deep learning just pay attention to the use of computing resources.These platforms are cumbersome to use and lack in model sharing.To address these shortages,this thesis designs and implements a deep learning model and dataset hosting platform that is easy to use and easy to share.Firstly,in order to reduce the user configuration and the tasks' waiting time,we analyse the characteristics of the deep learning model's structure,such as the type and size of the composition structure.And we use these characteristics to predict the running time and combine the genetic algorithm and the localization delay algorithm to achieve the scheduling scheme in the platform.Secondly,the hosting platform is also a common model comparison platform,using multiple nodes to provide computing services.Users can manage their own models and datasets in the platform and compare with the models exposed by users so as to realize the sharing of models between users.Above all,this thesis designs and implements the scheduling scheme used by the platform,introduces the characteristics of the deep learning model used in the scheduling scheme,and carrys out the genetic algorithm and the localization delay algorithm in this platform.After that,we formulate the overall design scheme of the platform.The platform is divided into two subsystems:service subsystem and computing subsystem.Service subsystem provides web services.Computing subsystem uses master-slave mode to manage all computing resources more conveniently.Finally,through the system function and performance tests,it is verified that the hosting platform meets the expected goals.