This is implementing a couple of simple ranking loss functions, in a way that makes it so we don't have to change our data types at all - you can just replace categorical cross entropy loss with this loss function, and things should just work (though you probably should also modify the last layer of your model to no longer do a softmax, which would make it impossible to get a score of 0 on the hinge loss).
The next PR will actually add some options to a model or two to allow using these loss functions.
This is implementing a couple of simple ranking loss functions, in a way that makes it so we don't have to change our data types at all - you can just replace categorical cross entropy loss with this loss function, and things should just work (though you probably should also modify the last layer of your model to no longer do a softmax, which would make it impossible to get a score of 0 on the hinge loss).
The next PR will actually add some options to a model or two to allow using these loss functions.