How to use DICE on multi-label tasks?

[xueyagaga]

Dear author
In many universal scenarios, there are multi-label tasks with output dimensions greater than one, such as recommendation and image recognition tasks.
How can I get DICE to apply in multi-label cases? Because most DICE tutorials are mainly discussing cases where the output dimension is 1, even for multi-class classification tasks (e.g. the IRIS dataset).