I have imbalanced dataset for multilabel classification, and I am not sure which loss function is a good one BinaryFocalCrossentropy or SigmoidFocalCrossEntropy. Another question is should I use class_weight parameter in model.fit or is this loss going to handle this itself?
You should use the Sigmoid one. Class_weight seem to work within model fit. You probably want to set alpha =None when using it, since alpha parameter in FC loss sets class weight itself.