부산대학교 도서관

We present Ultron-AutoML, an open-source, distributed framework for efficient hyper-parameter optimization (HPO) of ML models. Considering that hyper-parameter optimization is compute intensive and time-consuming, the framework has been designed for reliability – the ability to successfully complete an HPO Job in a multi-tenant, failure prone environment, as well as efficiency – completing the job with minimum compute cost and wall-clock time. From a user’s perspective, the framework emphasizes ease of use and customizability. The user can declaratively specify and execute an HPO Job, while ancillary tasks – containerizing and running the user’s scripts, model checkpointing, monitoring progress, parallelization – are handled by the framework. At the same time, the user has complete flexibility in composing the code-base for specifying the ML model training algorithm as well as, optionally, any custom HPO algorithm. The framework supports the creation of data-pipelines to stream batches of shuffled and augmented data from a distributed file system. This comes in handy for training Deep Learning models based on self-supervised, semi-supervised or representation learning algorithms over large training datasets. We demonstrate the framework’s reliability and efficiency by running a BERT pre-training job over a large training corpus using pre-emptible GPU compute targets. Despite the inherent unreliability of the underlying compute nodes, the framework is able to complete such long running jobs at 30% of the cost with a marginal increase in wall-clock time. The framework also comes with a service to monitor jobs and ensures reproducibility of any result.