Does ULMFiT require custom loss for multi-class?

I want to use the ULMFiT textcategorizer for multi-class, multi-label classification. Do I need to set a different-than-standard loss function?

Probably not, but it’s easy to set a custom loss for a learner. Perhaps this thread will help you.

@Pablo What I want is sigmoid applied to the last outputs; not softmax, since there might be more than one correct label. Is this default when there’s at least 1 input row with multiple labels?

I believe so, yes. If your labels have multiple active labels per row, then the loss will be automatically chosen so that it uses sigmoid instead of softmax.

Remember you can set the loss function manually and also, for your peace of mind, that you can simply print your current loss (try learner.loss_func).