Open-set implementation for Tensorflow

I’m reading some papers about openset classification in witch a model can predict what it knows (trained classes) and can gives an output “unknown” class when input is something it never trained on.

But it seems is still a state-of-art approach because I can’t find anything on the net about this output mode. Right now there is only “softmax” that is used for closed dataset. I’m wondering if there is some implementation or customization of softmax to add this new class (“unknown”) at test time.

I suggest to take a look at GitHub - iCGY96/awesome_OpenSetRecognition_list: A curated list of papers & resources linked to open set recognition, out-of-distribution, open set domain adaptation and open world recognition

You could try to look at some reference code in the list or on the mentioned challenge.

1 Like