Focal Loss in loss.py

[mr-mainak]

Hi,
In loss.py there is FocalLoss which has alpha=0.25. alpha_t = alpha * true + (1-alpha)(1-true). This is alright for a binary scenario. I have a multilabel scenario highly imbalanced. How can I use model.class_weights (which calculates the inverse frequency) as class weights for Focal Loss. When I pass model.class_weights as alpha parameter in FocalLoss I get dimension mismatch error. How can I do it?
P.S: I intend to use Focal Loss, thus not resorting to --image-weights flag for sampling while training.
Any help is highly appreciated. @glenn-jocher