Supplementary: Characterizing Generalization under Out-Of-Distribution Shifts in Deep Metric Learning AAnalyzing the model bias for selecting train-test splits