I believe the remaining two test errors are not mine.

Thanks a lot for the PR and for my slow reaction. I am still super busy :-/ I'll try to have a look asap!

Hi, I've also implemented weighted loss, but without breaking the API again.
I've implemented a class Label, which stores full and weak labelling (I'm working on LatentSSVM and I have different losses for full and weak labels), and it also stores weights.
Weighting is a very important feature (e.g. when you work with superpixels in semantic segmentation tasks) but I don't like the idea to pass X to loss function, it's not intuitive.
Thanks

@kondra thank you for your feedback. would you mind sharing some code?
Why don't you like the idea of having x in the loss? In principal everything is conditioned on x, right? In particular, the graph structure is encoded there.

@amueller because in common formulation of ssvm there is no x in loss ^_^
It's ok to pass x to the loss.
But may be we can implement this in a backward compatible way? (e.g. make x as a third parameter with default value None)
Thanks

Yes, it should be as backward compatible as possible.
You are right, there is no x in the loss in the papers, but I guess that is only for convenient notation.I don't see a reason why the loss should not be able to include x.