Question
I'm currently in the process of building an custom processor with a sequence_classification & a token_classification head. After some small adjustments to the FARM codebase and the creation of a new type of processor, I was able to sucessfully train and evaluate my model on the eval & test data but I'm stuck predicting on new inputs...

My task at hand is about Joint Intent Classification & Slot Filling, e.g. basically text classification + NER but could probably easily extended to any other task involving joint token & sequence classification and I would be open to provide a PR once it is working properly (at the very least for the more specific Joint Intent Classification & Slot Filling Task).

I would like to know what's the easiest/best way to get the inference output result for a custom task consisting of 1 token_classification & 1 sequence_classification head.
Currently I'm running into an KeyError because it's using the else clause used for the Natural Questions in formatted_preds of the trained model.

Additional context