According to the rules, upon reception of a prize, the software we use must be able to replicate the winning solution. Does that mean that if we train a model N times, they must give the exact same predictions? Or do you expect slight deviations given GPU randomness and other possible effects?Posted by: Galeros93 @ Aug. 2, 2020, 1:11 p.m.
You should make a reasonable effort to assure that your final results will be reproducible, small and explainable variations (for example, due to GPU randomness) are acceptable.
Please see the measures taken in the "Baseline Model" notebook to assure the reproducibility of the baseline results.
(MAFAT Challenge Team)