Multi Label Soft Margin Loss

Hi all,

Newbie here, and I am trying to realize a multi label (not multi class) classification network with three classes. My question is, if I would like to use Multilabel softmargin loss(is it recommended?), should i put a sigmoid layer after the last FC layer ?

or should the loss be defined as: loss=multilabel( output of Fc , target)

rather than

loss = multilabel(sigmoid(FC), target)?

and what should the target be like? should it be something like [ 0, 1, 1 ] which means it falls into class 2 and class 3?

Thanks for any help.