How to get the top N predictions for an object?

[hqm]

I am trying to build a more common-sense object recognition system which can give feedback to the Mask RCNN
to help train it. In order to do this, I would like, for a given object bbox that it detects, to get not just the classification
which it most confidence in, but all of the confidence levels for different classes for that region. That info is fed into
another system which independently computes which prediction has the 'least surprise' with respect to a different learning system's classification. What is the easiest way to get that list of confidences, for each object bbox, rather than the argmax that is coming out now?