Datastack: Unification of Heterogeneous Machine Learning Dataset Interfaces

Machine learning (ML) dataset preprocessing, cleaning, and integration into ML pipelines is often a cum-bersome endeavor that is susceptible to bugs and leads to unstructured code from the start. While existing frameworks for dataset integration often come with an extensive dataset repository, extending these repositories to new datasets is nontrivial due to lack of dataset retrieval, processing and iterator separation. To simplify the process of dataset integration, we present Datastack, an open-source framework that minimizes these efforts by providing well-defined interfaces that seamlessly integrate into existing machine learning frameworks. Inspired by stream processing frameworks such as Flink or Storm, Datastack decouples dataset-specific peculiarities such as custom data formats from the framework by introducing byte streams on an interface level. Furthermore, Datastack delivers dataset preprocessing functionalities such as stacking, splitting, and merging to alleviate error-prone data processing pipelines.