Code Generation for Ablation Technique
Samat Rakimbekuulu, Kanatbek Shambetaliev, Gulzada Esenalieva, Al Khan
Abstract
Using a machine learning system for testing involves several attempts in an iterative approach. Executing these trials concurrently on an Apache Spark cluster is a popular method. Due to Apache Spark's adherence to the Bulk Synchronous Parallel paradigm, trials are executed in parallel over several phases, separated by obstacles. This implies that all trials from the previous stage have to be completed in order to start a new set of trials. Because of this, we frequently wind up squandering a great deal of time and computational resources on unproductive experiments that should have been terminated quickly. In an effort to tackle these issues, we have developed an open-source system using MAGGY, which uses TensorFlow and Apache Spark to facilitate asynchronous and concurrent hyperparameter tuning and ablation investigations. Better resource use, ablation studies, and hyperparameter tuning are all made possible by this framework under a single expandable API