org.linqs.psl.application.learning.weight.maxlikelihood

Class MaxPiecewisePseudoLikelihood

java.lang.Object

org.linqs.psl.application.learning.weight.WeightLearningApplication

org.linqs.psl.application.learning.weight.VotedPerceptron

org.linqs.psl.application.learning.weight.maxlikelihood.MaxPiecewisePseudoLikelihood

All Implemented Interfaces:

ModelApplication

public class MaxPiecewisePseudoLikelihood
extends VotedPerceptron

Learns weights by optimizing the piecewise-pseudo-log-likelihood of the data using the voted perceptron algorithm.

Field Summary

Fields

Modifier and Type	Field and Description
static String	CONFIG_PREFIX Prefix of property keys used by this class.
static int	NUM_SAMPLES_DEFAULT
static String	NUM_SAMPLES_KEY Key for positive integer property.

Fields inherited from class org.linqs.psl.application.learning.weight.VotedPerceptron

AVERAGE_STEPS_DEFAULT, AVERAGE_STEPS_KEY, averageSteps, baseStepSize, CLIP_NEGATIVE_WEIGHTS_DEFAULT, CLIP_NEGATIVE_WEIGHTS_KEY, clipNegativeWeights, CUT_OBJECTIVE_DEFAULT, CUT_OBJECTIVE_KEY, cutObjective, inertia, INERTIA_DEFAULT, INERTIA_KEY, L1_REGULARIZATION_DEFAULT, L1_REGULARIZATION_KEY, l1Regularization, L2_REGULARIZATION_DEFAULT, L2_REGULARIZATION_KEY, l2Regularization, maxNumSteps, NUM_STEPS_DEFAULT, NUM_STEPS_KEY, numSteps, SCALE_GRADIENT_DEFAULT, SCALE_GRADIENT_KEY, SCALE_STEP_SIZE_DEFAULT, SCALE_STEP_SIZE_KEY, scaleGradient, scaleStepSize, STEP_SIZE_DEFAULT, STEP_SIZE_KEY, ZERO_INITIAL_WEIGHTS_DEFAULT, ZERO_INITIAL_WEIGHTS_KEY, zeroInitialWeights

Fields inherited from class org.linqs.psl.application.learning.weight.WeightLearningApplication

allRules, atomManager, evaluator, EVALUATOR_DEFAULT, EVALUATOR_KEY, expectedIncompatibility, GROUND_RULE_STORE_DEFAULT, GROUND_RULE_STORE_KEY, groundRuleStore, inLatentMPEState, inMPEState, latentGroundRuleStore, latentTermStore, MAX_RANDOM_WEIGHT, MIN_ADMM_STEPS, mutableRules, observedDB, observedIncompatibility, RANDOM_WEIGHTS_DEFAULT, RANDOM_WEIGHTS_KEY, reasoner, REASONER_DEFAULT, REASONER_KEY, rvDB, supportsLatentVariables, TERM_GENERATOR_DEFAULT, TERM_GENERATOR_KEY, TERM_STORE_DEFAULT, TERM_STORE_KEY, termGenerator, termStore, trainingMap

Constructor Summary

Constructors

Constructor and Description
MaxPiecewisePseudoLikelihood(List<Rule> rules, Database rvDB, Database observedDB)
MaxPiecewisePseudoLikelihood(Model model, Database rvDB, Database observedDB)

Method Summary

Modifier and Type	Method and Description
protected void	computeExpectedIncompatibility() Compute the expected incompatibility using the piecewisepseudolikelihood.
double	computeLoss() Internal method for computing the loss at the current point before taking a step.
protected void	computeObservedIncompatibility() Compute the incompatibility in the model using the labels (truth values) from the observed (truth) database.
protected void	postInitGroundModel() A convenient place for children to do additional ground model initialization.
void	setBudget(double budget) Set a budget (give as a proportion of the max budget).

Methods inherited from class org.linqs.psl.application.learning.weight.VotedPerceptron

computeRegularizer, computeScalingFactor, doLearn, getLoss

Methods inherited from class org.linqs.psl.application.learning.weight.WeightLearningApplication

close, computeLatentMPEState, computeMPEState, createAtomManager, getWLA, initGroundModel, initGroundModel, initGroundModel, initLatentGroundModel, learn, setDefaultRandomVariables, setLabeledRandomVariables

Methods inherited from class java.lang.Object

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

Field Detail

CONFIG_PREFIX

public static final String CONFIG_PREFIX

Prefix of property keys used by this class.

See Also:

Constant Field Values

NUM_SAMPLES_KEY

public static final String NUM_SAMPLES_KEY

Key for positive integer property. MaxPiecewisePseudoLikelihood will sample this many values to approximate the expectations.

See Also:

Constant Field Values

NUM_SAMPLES_DEFAULT

public static final int NUM_SAMPLES_DEFAULT

See Also:

Constant Field Values

Constructor Detail

MaxPiecewisePseudoLikelihood

public MaxPiecewisePseudoLikelihood(Model model,
                                    Database rvDB,
                                    Database observedDB)

MaxPiecewisePseudoLikelihood

public MaxPiecewisePseudoLikelihood(List<Rule> rules,
                                    Database rvDB,
                                    Database observedDB)

Method Detail

postInitGroundModel

protected void postInitGroundModel()

Description copied from class: WeightLearningApplication

A convenient place for children to do additional ground model initialization.

Overrides:

postInitGroundModel in class WeightLearningApplication

computeExpectedIncompatibility

protected void computeExpectedIncompatibility()

Compute the expected incompatibility using the piecewisepseudolikelihood. Use Monte Carlo integration to approximate epectations.

Overrides:

computeExpectedIncompatibility in class WeightLearningApplication

computeLoss

public double computeLoss()

Description copied from class: WeightLearningApplication

Internal method for computing the loss at the current point before taking a step. Child methods may override. The default implementation just sums the product of the difference between the expected and observed incompatibility.

Overrides:

computeLoss in class WeightLearningApplication

Returns:

current learning loss

computeObservedIncompatibility

protected void computeObservedIncompatibility()

Description copied from class: WeightLearningApplication

Compute the incompatibility in the model using the labels (truth values) from the observed (truth) database. This method is responsible for filling the observedIncompatibility member variable. This may call setLabeledRandomVariables() and not reset any ground atoms to their original value. The default implementation just calls setLabeledRandomVariables() and sums the incompatibility for each rule.

Overrides:

computeObservedIncompatibility in class WeightLearningApplication

setBudget

public void setBudget(double budget)

Description copied from class: WeightLearningApplication

Set a budget (give as a proportion of the max budget). Child implementations should make sure to override this and call up the super chain.

Overrides:

setBudget in class VotedPerceptron