Guaranteeing Reproducibility in Deep Learning Competitions