CONIX Publication

We introduce a production scale library called Mango, designed to do parallel hyperparameter tuning of machine learning classifiers. Mango enables developers with rich abstractions to define complex parameter spaces, scheduling capabilities on the cloud, fault tolerance, and state-of-the-art tuning algorithms. Our evaluation of Mango on several datasets shows it performs better than other available libraries and is designed with the ease of usability. Mango is currently used in production at ARM. In this poster, we also discuss the new capabilities available for experimentation in Mango-ver2, including the optimal search of hyperparameters across a set of classifiers parallelly.