mrftools.ConvexBeliefPropagator module

Convexified Belief Propagation Class

class mrftools.ConvexBeliefPropagator.ConvexBeliefPropagator(markov_net, counting_numbers=None)

Bases: mrftools.MatrixBeliefPropagator.MatrixBeliefPropagator

Class to perform convexified belief propagation based on counting numbers. The class allows for non-Bethe counting numbers for the different factors in the MRF. If the factors are all non-negative, then the adjusted Bethe free energy is convex, providing better guarantees about the convergence and bounds of the primal and dual objective values.

compute_beliefs()

compute_bethe_entropy()

compute_pairwise_beliefs()

update_messages()