Face models trained on VGGFace2 (3,057,398 images / 8,414 identities): 31-attribute multi-task predictors and VGGFace2-derived recognizers.