What I learned today: you can't use if-conditionals in a Keras custom loss function. I managed to come with a solution...

All the operations in the custom loss function need to be vectorised, so an if-conditional breaks it and throws an exception. The solution is to create Boolean vectors corresponding to each condition, turn their types into float (must be float), multiply each Boolean by the corresponding loss function and sum.

Example:

This is a fascinating and illuminating read, with implications well beyond Wikipedia – think of all the careful narrative framing and dubious claims that you see on social media and hear from politicians – even supposedly mainstream politicians, such as Jacob Rees-Mogg, who attempted to minimise the British use of concentration camps during the Boer War by falsely claiming that the death rate in those camps was the same as the death rate in Glasgow at the time.

https://getpocket.com/explore/item/one-woman-s-mission-to-rewrite-nazi-history-on-wikipedia?utm_source=pocket-newtab-global-en-GB