mrftools.MatrixBeliefPropagator module

BeliefPropagator class.

class mrftools.MatrixBeliefPropagator.MatrixBeliefPropagator(markov_net)

Bases: mrftools.Inference.Inference

Object that can run belief propagation on a MarkovNet. Uses sparse matrices to encode the indexing underlying belief propagation.

augment_loss(var, state)

Adds a loss penalty to the MRF energy function. Used to create loss-augmented inference for max-margin learning

Parameters:

• var (object) – variable to add loss to
• state (int) – state of variable in ground truth labels

Returns:

None

compute_beliefs()

Compute unary log beliefs based on current messages and store them in belief_mat

Returns:None

compute_bethe_entropy()

Compute Bethe entropy from current beliefs. This method assumes that the beliefs have been computed and are fresh.

Returns:computed Bethe entropy

compute_dual_objective()

Compute the value of the BP Lagrangian.

Returns:Lagrangian objective function

compute_energy()

Compute the log-linear energy. Assume that the beliefs have been computed and are fresh.

Returns:computed energy

compute_energy_functional()

Compute the energy functional, which is the variational approximation of the log-partition function.

Returns:computed energy functional

compute_inconsistency()

Return the total disagreement between each unary belief and its pairwise beliefs. When message passing converges, the inconsistency should be within numerical error.

Returns:the total absolute disagreement between each unary belief and its pairwise beliefs

compute_pairwise_beliefs()

Compute pairwise log beliefs based on current messages, and stores them in pair_belief_tensor

Returns:None

condition(var, state)

Condition a variable, usually because of statistical evidence, to be in a subset of its possible states

Parameters:

• var (object) – variable to condition
• state (int or array) – state to condition to or array of states that the variable may be in

Returns:

None

disallow_impossible_states()

Force variables to only allow nonzero probability on their possible states.

Returns:None

get_feature_expectations()

Computes the feature expectations under the currently estimated marginal probabilities. Only works when the model is a LogLinearModel class with features for edges.

Returns:vector of the marginals in order of the flattened unary features first, then the flattened pairwise features

infer(tolerance=1e-08, display='iter')

Run belief propagation until messages change less than tolerance.

Parameters:

• tolerance – the minimum amount that the messages can change while message passing can be considered not converged
• display –

  string parameter indicating how much to display. Options are ‘full’ and ‘iter’ ‘full’ prints the energy functional and dual objective each iteration,

  which requires extra computation

  ‘iter’ prints just the change in messages each iteration

Returns:

None

initialize_messages()

Initialize messages to default initialization (set to zeros).

Returns:None

load_beliefs()

Update the belief dictionaries var_beliefs and pair_beliefs using the current messages.

Returns:None

set_max_iter(max_iter)

Set the maximum iterations of belief propagation to run before early stopping :param max_iter: integer maximum iterations :return: None

set_messages(messages)

Set the message vector. Useful for warm-starting inference if a previously computed message matrix is available.

Parameters:messages (ndarray) – message matrix Returns:None

update_messages()

Update all messages between variables and store them in message_mat

Returns:the float change in messages from previous iteration.

mrftools.MatrixBeliefPropagator.logsumexp(matrix, dim=None)

Compute log(sum(exp(matrix), dim)) in a numerically stable way.

Parameters:

• matrix (ndarray) – input ndarray
• dim (int) – integer indicating which dimension to sum along

Returns:

numerically stable equivalent of np.log(np.sum(np.exp(matrix), dim)))

Return type:

ndarray

mrftools.MatrixBeliefPropagator.sparse_dot(full_matrix, sparse_matrix)

Convenience function to compute the dot product of a full matrix and a sparse matrix. Useful to avoid writing code with a lot of transposes.

Parameters:

• full_matrix (ndarray) – dense matrix
• sparse_matrix (csc_matrix) – sparse matrix

Returns:

full_matrix.dot(sparse_matrix)

Return type:

ndarray