cc.mallet.grmm.inference
Class AbstractBeliefPropagation

java.lang.Object
  cc.mallet.grmm.inference.AbstractInferencer
      cc.mallet.grmm.inference.AbstractBeliefPropagation

All Implemented Interfaces:

Inferencer, java.io.Serializable, java.lang.Cloneable

Direct Known Subclasses:

LoopyBP, ResidualBP, TreeBP, TRP

public abstract class AbstractBeliefPropagation
extends AbstractInferencer

Abstract base class for umplementations of belief propagation for general factor graphs. This class manages arrays of messages, computing beliefs from messages, and convergence thresholds.

How to send individual messages (e.g., sum-product, max-product, etc) are mananged by istances of the interface @link{#MessageStrategy}. Concrete subclasses decide which order to send messages in.

Version:

$Id: AbstractBeliefPropagation.java,v 1.1 2007/10/22 21:37:49 mccallum Exp $

Author:

Charles Sutton

See Also:

Serialized Form

Nested Class Summary

static class	AbstractBeliefPropagation.AbstractMessageStrategy
static class	AbstractBeliefPropagation.MaxProductMessageStrategy
static interface	AbstractBeliefPropagation.MessageStrategy
static class	AbstractBeliefPropagation.SumProductMessageStrategy

Field Summary

protected int[]	assignedVertexPtls
protected int	iterUsed
protected static java.util.logging.Logger	logger
protected FactorGraph	mdlCurrent
protected boolean	normalizeBeliefs
protected boolean	useCaching

Constructor Summary

protected	AbstractBeliefPropagation()
protected	AbstractBeliefPropagation(AbstractBeliefPropagation.MessageStrategy messager)

Method Summary

protected void	copyOldMessages()
protected void	doneWithGraph(FactorGraph mdl)
void	dump()
void	dump(java.io.PrintWriter writer)
AbstractBeliefPropagation.MessageStrategy	getMessager()
int	getMessagesSent() Returns the total number of messages this inferencer has sent since its creation.
int	getMessagesUsedLastTime() Returns the number of messages sent during the last call to computeMarginals.
static int	getTotalMessagesSent() Returns the total number of messages all BP inferencers have sent in the current Java image.
protected boolean	hasConverged()
protected boolean	hasConverged(double threshold)
protected void	initForGraph(FactorGraph mdl)
int	iterationsUsed()
double	lookupLogJoint(Assignment assn) Returns the natural logarithm of the joint probability of a given assignment, computed in some factorized fashion.
Factor	lookupMarginal(Variable var) Returns the computed marginal of a given variable.
Factor	lookupMarginal(VarSet c) Returns the computed marginal of a given clique in a graph.
void	reportTime() Outputs some measure of the total time spent in this inferencer.
protected void	resetMessagesSentAtStart()
protected void	sendMessage(FactorGraph mdl, Factor from, Variable to)
protected void	sendMessage(FactorGraph mdl, Variable from, Factor to)
AbstractBeliefPropagation	setMessager(AbstractBeliefPropagation.MessageStrategy messager)

Methods inherited from class cc.mallet.grmm.inference.AbstractInferencer

computeMarginals, duplicate, lookupJoint, query

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

logger

protected static java.util.logging.Logger logger

normalizeBeliefs

protected boolean normalizeBeliefs

useCaching

protected boolean useCaching

iterUsed

protected transient int iterUsed

mdlCurrent

protected transient FactorGraph mdlCurrent

assignedVertexPtls

protected transient int[] assignedVertexPtls

Constructor Detail

AbstractBeliefPropagation

protected AbstractBeliefPropagation()

AbstractBeliefPropagation

protected AbstractBeliefPropagation(AbstractBeliefPropagation.MessageStrategy messager)

Method Detail

getMessager

public AbstractBeliefPropagation.MessageStrategy getMessager()

setMessager

public AbstractBeliefPropagation setMessager(AbstractBeliefPropagation.MessageStrategy messager)

getTotalMessagesSent

public static int getTotalMessagesSent()

Returns the total number of messages all BP inferencers have sent in the current Java image.

getMessagesSent

public int getMessagesSent()

Returns the total number of messages this inferencer has sent since its creation.

getMessagesUsedLastTime

public int getMessagesUsedLastTime()

Returns the number of messages sent during the last call to computeMarginals.

resetMessagesSentAtStart

protected void resetMessagesSentAtStart()

copyOldMessages

protected final void copyOldMessages()

hasConverged

protected final boolean hasConverged()

hasConverged

protected final boolean hasConverged(double threshold)

initForGraph

protected void initForGraph(FactorGraph mdl)

sendMessage

protected void sendMessage(FactorGraph mdl,
                           Variable from,
                           Factor to)

sendMessage

protected void sendMessage(FactorGraph mdl,
                           Factor from,
                           Variable to)

doneWithGraph

protected void doneWithGraph(FactorGraph mdl)

iterationsUsed

public int iterationsUsed()

lookupMarginal

public Factor lookupMarginal(Variable var)

Description copied from interface: Inferencer

Returns the computed marginal of a given variable. Before using this method, computeMarginals must have been previously called on the graphical model that contains v.

Specified by:

lookupMarginal in interface Inferencer

Specified by:

lookupMarginal in class AbstractInferencer

See Also:

Inferencer.computeMarginals(FactorGraph)

dump

public void dump()

Specified by:

dump in interface Inferencer

Overrides:

dump in class AbstractInferencer

reportTime

public void reportTime()

Description copied from interface: Inferencer

Outputs some measure of the total time spent in this inferencer.

Specified by:

reportTime in interface Inferencer

Overrides:

reportTime in class AbstractInferencer

dump

public void dump(java.io.PrintWriter writer)

lookupMarginal

public Factor lookupMarginal(VarSet c)

Description copied from interface: Inferencer

Returns the computed marginal of a given clique in a graph. Before using this method, computeMarginals must have been previously called on the graphical model that contains the clique.

Specified by:

lookupMarginal in interface Inferencer

Overrides:

lookupMarginal in class AbstractInferencer

See Also:

Inferencer.computeMarginals(cc.mallet.grmm.types.FactorGraph), #computeMarginals(JunctionTree)

lookupLogJoint

public double lookupLogJoint(Assignment assn)

Description copied from interface: Inferencer

Returns the natural logarithm of the joint probability of a given assignment, computed in some factorized fashion. Before using this method, computeMarginals must have been previously called on the graphical model that contains the variables of assn.

This method is less likely to underflow than Math.log (lookupJoint (assn)).

Specified by:

lookupLogJoint in interface Inferencer

Overrides:

lookupLogJoint in class AbstractInferencer

See Also:

Inferencer.computeMarginals(cc.mallet.grmm.types.FactorGraph), #computeMarginals(JunctionTree)