The first argument to inference
The second argument to inference
The loss function. It needs to score variables.
The inference routine
Put objective value and gradient into the accumulators.
Put objective value and gradient into the accumulators. Either argument can be null if they don't need to be computed.
Accumulator to hold value
Accumulator to hold gradient
Implements the structured SVM objective function, by doing loss-augmented inference.
It needs to be a loss, not an objective.
If you want to pass a specialized inference method use the structured perceptron example with loss-augmented inference directly.
The type of the labels