Source code for mrftools.ApproxMaxLikelihood

"""Class to do generative learning directly on MRF parameters."""
from copy import deepcopy

from .Learner import Learner
from .LogLinearModel import LogLinearModel
from .MatrixBeliefPropagator import MatrixBeliefPropagator


class ApproxMaxLikelihood(Learner):
    """
    Object that runs approximate maximum likelihood parameter training.
    This method creates an indicator model where every feature is an indicator function, also known as an 
    overcomplete representation.
    """

    def __init__(self, markov_net, inference_type=MatrixBeliefPropagator):
        """
        Initialize the learner with the Markov network whose parameters are to be learned.
        
        :param markov_net: MarkovNet object whose parameters are to be learned
        :type markov_net: MarkovNet
        :param inference_type: Inference method to use for estimating the feature expectations during learning
        :type inference_type: Inference
        """
        super(ApproxMaxLikelihood, self).__init__(inference_type)
        self.base_model = LogLinearModel()
        self.base_model.create_indicator_model(markov_net)

    def add_data(self, labels):
        """
        Add observed training data
        
        :param labels: dictionary containing an integer state value for each observed variable 
        :type labels: dict
        :return: None
        """
        model = deepcopy(self.base_model)
        super(ApproxMaxLikelihood, self).add_data(labels, model)

        # as a hack to save time, since these models don't condition on anything, make all belief propagators equal
        self.belief_propagators = [self.belief_propagators[0]]