What losses can I use for Multi Lable Calssification?

I have a multi lable classification problem, meaning each sample might have more than one label associate to it.
I was wondering what kinda loss functions can I use for this task, and how my target should look for each of those loss functions?

You could use nn.BCELoss with a sigmoid as the last non-linearity or directly pass the logits to nn.BCEwithLogitsLoss.
Your target should be a multi-hot encoded tensor, i.e. the true classes should be passed as ones while the others as zeros.