Reproducible Research Environments with Repo2Docker

Jessica Forde, Tim Head, Chris Holdgraf, Yuvi Panda, Gladys Nalvarete, Benjamin Ragan-Kelley, Erik Sundell

Abstract: Reproducibility challenges in machine learning often center on questions of software engineering practices. Researchers struggle to reproduce another scientist's work because they cannot translate a paper into code with similar results or run an author's code. repo2docker provides a simple tool for checking the minimum requirements to reproduce a paper by building a Docker image based on a repository path or URL. Its goal is to minimize the effort needed to convert a static repository into a working software environment. By inspecting a repository for standard configuration files used in contemporary software engineering and leveraging containerization methods, repo2docker deterministically reproduces the environment of the author so the researcher can reproduce the author's experiments.

TL;DR: repo2docker uses standard configuration files to reproduce the environment of a repository in a Docker image.

Keywords: reproducibility, docker, python