Project repositories for machine learning with TensorFlow

PS Janardhanan
2020 Procedia Computer Science  
In machine learning, models capture intelligence from data using algorithms implemented on frameworks like TensorFlow. Models learn during the training phase; an iterative process in which parameters are tuned to improve the prediction accuracy. Software repositories are used to save the artifacts of model development so that they can be modified in subsequent releases and shared between development teams. Issues in saving the state of machine learning projects is different from standard
more » ... e practices. Machine learning models are equivalent to binary executable programs and ideally one should be able to recreate the model from the information preserved in project repositories. Recreation of the model becomes necessary when there is a change in the development team as part of the product transition. Retraining of models also becomes necessary when the inaccuracy in the predictions made on new data increases beyond the acceptable limit. When traditional source code management systems are used for maintaining repositories of machine learning projects, we face challenges in keeping complete information required for model development starting from saved states. This paper presents the results of the studies conducted to identify the challenges in maintaining TensorFlow machine learning projects in repositories. Some of the existing tools are compared and recommendations are made to improve the ease of recreation of machine learning models by saving complete information in project repositories maintained in normal source code control systems. Abstract In machine learning, models capture intelligence from data using algorithms implemented on frameworks like TensorFlow. Models learn during the training phase; an iterative process in which parameters are tuned to improve the prediction accuracy. Software repositories are used to save the artifacts of model development so that they can be modified in subsequent releases and shared between development teams. Issues in saving the state of machine learning projects is different from standard software practices. Machine learning models are equivalent to binary executable programs and ideally one should be able to recreate the model from the information preserved in project repositories. Recreation of the model becomes necessary when there is a change in the development team as part of the product transition. Retraining of models also becomes necessary when the inaccuracy in the predictions made on new data increases beyond the acceptable limit. When traditional source code management systems are used for maintaining repositories of machine learning projects, we face challenges in keeping complete information required for model development starting from saved states. This paper presents the results of the studies conducted to identify the challenges in maintaining TensorFlow machine learning projects in repositories. Some of the existing tools are compared and recommendations are made to improve the ease of recreation of machine learning models by saving complete information in project repositories maintained in normal source code control systems.
doi:10.1016/j.procs.2020.04.020 fatcat:tin3ivao3fanlbzywhd6lc72ty