Rectify docs for MultiLabelSoftMarginLoss

[andy-yangz]

📚 Documentation

Concern about MultiLabelSoftMarginLoss, I think there is still a problem with its doc.
As @DNGros pointed out, there is a lot of confusion when using MultiLabelMarginLoss and MultiLabelSoftMarginLoss, and he improved the docs. #15863
However, for the docs of MultiLabelSoftMarginLoss it's still not correct.
Since if you set the label as the docs

Target: (N, C)(N,C) , label targets padded by -1 ensuring same shape as the input.

pad it like [0,3,-1,-1], the result will be wrong.
If we look at the equation

we will find it, in fact, write clearly that label y need to be 0 or 1.
If we feed the -1 padded label, the loss actually sometimes will become negative number, and this is shown by my experiment.

So I think the right Target format for MultiLabelSoftMarginLoss is just one_hot vector.
And we can give a good example code as MultiLabelMarginLoss has.

[andy-yangz]

Do I need to submit a pull request to fix this issue?
I have no such open-source contributing experience before, so I wish to get some guidance.

[shaibagon]

Just came across this SO question raising the same issue.

[zhiqihuang]

I think the target of this loss function should be binary vectors indicating the positive class. The doc of this loss function is somehow the same as the MultiLabelMarginLoss. But MultiLabelSoftMarginLoss is more like the multi-label version of cross-entropy.