multi-label classification

[bruffridge]

I read your paper, and was impressed by your approach. I'm considering using this approach for a multi-label classification problem I'm working on, but from the paper it seems that ASTRA is only doing multi-class classification. Do you know if it is possible to modify ASTRA to extract confidence scores for each class for each instance? My use-case involves generating a list of documents ranked by predicted similarity to a given class.

[gkaramanolakis]

The current version of ASTRA supports multi-class classification but it is possible to modify it for multi-label classification problems.

In our paper, you would need to replace the loss functions for the Teacher and Student (Equations 1 and 4) with the corresponding multi-label versions.

Our code directly provides access to the Teacher's probabilities (i.e., soft pseudo-labels instead of hard pseudo-labels) here. You can use the Teacher's probabilities and define a multi-label loss as part of the Student's trainer code here.

I hope this helps. :)