Why MultiLabelMarginLoss take torch.long has arguments?

I think nn.BCELoss could be a good alternative. Here is a small example.