Models

BaseModel([domain])	Base model.
LinearModel([domain, w, features])	A linear model.

Pyconstructi abstracts predictive models using the BaseModel class. Classes that inherit from BaseModel are objects holding the parameters of some mathematical model and that can “communicate” with domains in order to make prediction by solving inference problems.

When making inference, the model’s parameters are passed to the Domain, which needs to be able to interpret the semantics of the parameters. For instance, the standard model provided by Pyconstruct is a LinearModel, whose parameters consist in just a weight vector (and some optional additional features). Domains using the linear_model macro from the linear.pmzn file (see the domains module documentation for details) are readily capable of interpreting this weight vector and use it to compute the dot product with the feature vector.

While linear models cover many of the cases in structured-output prediction, one can easily think of cases in which an ad-hoc model may be beneficial, e.g. when the model has some hyper-paramenters or when performing some sort of feature learning. That said, in most cases you should not need to manipulate models directly.

The model is the middle-men between a learner and a domain. The learner fits the parameter of a model, without knowing how inference is carried out. The domain, on the other hand, receives the model’s parameters, without knowing how these are learned. Usually a learner is only capable of learning one type of models (and derived ones), while a domain may conditionally be used with different models (if properly configured).