Feature request: computation of multi-label accuracy

[wendlerc]

Recent work [1] has teased apart all kinds of labelling errors in imagenet and (rightfully) suggests that multi-label accuracy (labels for imagenet available nowadays [2]) is the better metric for assessing model performance on imagenet.

[1] https://openreview.net/pdf?id=zEoP4vzFKy
[2] https://github.com/naver-ai/relabel_imagenet