Mobile Deep Learning Model Selection And Library Recommendation

In recent years,with the development of mobile Internet and the increasing popularity of mobile terminal devices(such as smart phones,tablets,etc.),users have more diversified ways to access the Internet,and end-side devices have surpassed computers to become the largest number of Internet terminals.Meanwhile,the third wave of artificial intelligence promoted by deep learning gradually improves the ecology of deep learning model for endto-end devices.Under this background,on the one hand,in the face of rich deep learning model,how to dig deep learning model and the developers demand,the relationship between the business scenario,quickly select deep learning model is crucial.on the other hand,due to the end side equipment running environment is complicated,and is used to model reasoning characteristic to the performance of deep learning library.How to effectively associate the inference performance of the libs with the selected model,so as to select the model deployment framework and deployment mode with the best performance,reduce the cost of model deployment,and improve the inference performance is an urgent problem to be solved.In view of the above problems,the paper studies the deep learning model selection and library recommendation for mobile devices,and has achieved the following results:The deep learning library ecology is deeply explored,and the mobile deep learning library is comprehensively tested.Firstly,MDLParse,a mobile deep learning mining analysis tool and DLbench,a performance evaluation library,are designed and implemented.Based on MDLParse mining depth study and analysis technology application status and trends are implement,and based on DLBench in 10 sets of equipment and different hardware extensive experiments on deep learning model,explore the actuators,hardware equipment impact on the performance of deep learning library,and reveal the underlying resources use details and end side the internal connection of deep learning framework performance,Particularly salient issues such as performance fragmentation and slow cold-start reasoning.To solve the problem that it is difficult to efficiently filter models that conform to specific scenarios and target user groups from complex and diverse deep learning technologies,research on function-driven deep learning model recommendation methods.Firstly,a crude classification algorithm was proposed based on TF-IDF and Labeled-LDA.Then,based on ALBERT pretraining language model,a light twin four-level model with multi-scale features is proposed to describe the similarity between applications.Finally,compound recommendation of deep learning model is carried out based on user demand and cost.Experimental results show that the distance model with multi-dimensional classification and multi-scale feature aggregation has significant performance advantages.Considering performance fragmentation phenomenon and effect of complex factors on mobile model inference performance,it is difficult to associate the model with best performance libs.Firstly,we build feature engineering that transforms the structural information of high-dimensional complex models into heterogeneous graphs and low-dimensional numerical information.Secondly,because misclassification of libs recommendation is directly sensitive to the performance loss error and the data set’ category imbalance,we propose PSO-W-CatBoost based on particle swarm optimization.Experiment result show that,PSO-W-CatBoost significantly improves the prediction accuracy.