scANVI classification uncertainty extraction method

[LuckyMD]

It would be very useful to have some quantification of model uncertainty for the scANVI cell type classification output. In the simplest form this could be the softmax output from the penultimate layer. Ideally, it would be a separately trained linear activation function output as softmax will probably exaggerate the differences in uncertainties leading to poor calibration. The simple output would already be useful I think.

[PierreBoyeau]

Hi, it sounds like this already returns the softmax outputs.

[LuckyMD]

Ah, I must have missed that, thanks! I'll test it out soon! I guess then my feature request slims down to the calibrated linear layer for probabilities :).