Currently, running an experiment on view 2 of the dataset requires to have 10 different experiments run, and at the end concatenate all score files into one in order to plot ROC curves.

Since modern deep-learning evaluations do not require to use the training split of the 10-fold LFW data splits, it would be easier to implement an eval-only (or maybe dev-only) protocol on view 2 that simply concatenates all 10 folds of view 2.