Skip to content

Commit

Permalink
Merge branch 'main' into experimental_minimizer_based_learning
Browse files Browse the repository at this point in the history
  • Loading branch information
@dickensc
dickensc committed Aug 25, 2023
2 parents 1ffba60 + 5bbd231 commit 1d54995
Show file tree
Hide file tree
Showing 34 changed files with 514 additions and 487 deletions.
4 changes: 2 additions & 2 deletions psl-core/src/main/java/org/linqs/psl/application/inference/InferenceApplication.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -114,7 +114,7 @@ protected Reasoner createReasoner() {
}

protected TermStore createTermStore() {
return new ADMMTermStore(database);
return new ADMMTermStore(database.getAtomStore());
}

/**
Expand All @@ -124,7 +124,7 @@ protected TermStore createTermStore() {
*/
protected void completeInitialize() {
log.info("Grounding out model.");
long termCount = Grounding.groundAll(rules, termStore);
long termCount = Grounding.groundAll(rules, termStore, database);
log.info("Grounding complete.");
log.debug("Generated {} terms.", termCount);
}
Expand Down
2 changes: 1 addition & 1 deletion psl-core/src/main/java/org/linqs/psl/application/inference/mpe/ADMMInference.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -41,6 +41,6 @@ protected Reasoner createReasoner() {

@Override
protected TermStore createTermStore() {
return new ADMMTermStore(database);
return new ADMMTermStore(database.getAtomStore());
}
}
2 changes: 1 addition & 1 deletion psl-core/src/main/java/org/linqs/psl/application/inference/mpe/DualBCDInference.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -41,6 +41,6 @@ protected Reasoner createReasoner() {

@Override
protected TermStore createTermStore() {
return new DualLCQPTermStore(database);
return new DualLCQPTermStore(database.getAtomStore());
}
}
2 changes: 1 addition & 1 deletion psl-core/src/main/java/org/linqs/psl/application/inference/mpe/SGDInference.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -41,6 +41,6 @@ protected Reasoner createReasoner() {

@Override
protected TermStore createTermStore() {
return new SGDTermStore(database);
return new SGDTermStore(database.getAtomStore());
}
}
40 changes: 15 additions & 25 deletions ...ore/src/main/java/org/linqs/psl/application/learning/weight/gradient/GradientDescent.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -22,10 +22,7 @@
import org.linqs.psl.config.Options;
import org.linqs.psl.database.AtomStore;
import org.linqs.psl.database.Database;
import org.linqs.psl.model.deep.DeepModelPredicate;
import org.linqs.psl.model.predicate.DeepPredicate;
import org.linqs.psl.model.predicate.Predicate;
import org.linqs.psl.model.rule.GroundRule;
import org.linqs.psl.model.rule.Rule;
import org.linqs.psl.model.rule.WeightedRule;
import org.linqs.psl.reasoner.InitialValue;
Expand All @@ -34,7 +31,6 @@
import org.linqs.psl.util.Logger;
import org.linqs.psl.util.MathUtils;

import java.util.ArrayList;
import java.util.Arrays;
import java.util.HashMap;
import java.util.List;
Expand Down Expand Up @@ -164,7 +160,7 @@ protected void postInitGroundModel() {
throw new IllegalArgumentException("If validation is being run, then an evaluator must be specified for predicates.");
}

if (!((!runValidation) || (validationInferenceApplication.getDatabase().getAtomStore().size() > 0))) {
if (!((!runValidation) || (validationInferenceApplication.getTermStore().getAtomStore().size() > 0))) {
throw new IllegalStateException("If validation is being run, then validation data must be provided in the runtime.json file.");
}

Expand All @@ -178,10 +174,10 @@ protected void postInitGroundModel() {
trainMAPTermState = trainInferenceApplication.getTermStore().saveState();
validationMAPTermState = validationInferenceApplication.getTermStore().saveState();

float[] trainAtomValues = trainInferenceApplication.getDatabase().getAtomStore().getAtomValues();
float[] trainAtomValues = trainInferenceApplication.getTermStore().getAtomStore().getAtomValues();
trainMAPAtomValueState = Arrays.copyOf(trainAtomValues, trainAtomValues.length);

float[] validationAtomValues = validationInferenceApplication.getDatabase().getAtomStore().getAtomValues();
float[] validationAtomValues = validationInferenceApplication.getTermStore().getAtomStore().getAtomValues();
validationMAPAtomValueState = Arrays.copyOf(validationAtomValues, validationAtomValues.length);

rvAtomGradient = new float[trainAtomValues.length];
Expand Down Expand Up @@ -221,18 +217,12 @@ protected void doLearn() {
boolean breakGD = false;
float objective = 0.0f;
float oldObjective = Float.POSITIVE_INFINITY;

float[] bestWeights = new float[mutableRules.size()];


log.info("Gradient Descent Weight Learning Start.");
initForLearning();

for (int i = 0; i < bestWeights.length; i++) {
bestWeights[i] = mutableRules.get(i).getWeight();
}

long totalTime = 0;
int iteration = 0;
int epoch = 0;
while (!breakGD) {
long start = System.currentTimeMillis();

Expand Down Expand Up @@ -261,7 +251,7 @@ protected void doLearn() {
}
}

computeIterationStatistics();
computeIterationStatistics(epoch);

objective = computeTotalLoss();

Expand All @@ -271,26 +261,26 @@ protected void doLearn() {
clipWeightGradient();
}

gradientStep(iteration);
gradientStep(epoch);

long end = System.currentTimeMillis();

totalTime += end - start;
oldObjective = objective;

breakGD = breakOptimization(iteration, objective, oldObjective);
breakGD = breakOptimization(epoch, objective, oldObjective);
log.trace("Iteration {} -- Weight Learning Objective: {}, Gradient Magnitude: {}, Parameter Movement: {}, Iteration Time: {}",
iteration, objective, computeGradientNorm(), parameterMovement, (end - start));
epoch, objective, computeGradientNorm(), parameterMovement, (end - start));

iteration++;
epoch++;
}

log.info("Gradient Descent Weight Learning Finished.");

if (saveBestValidationWeights) {
// Reset rule weights to bestWeights.
for (int i = 0; i < mutableRules.size(); i++) {
mutableRules.get(i).setWeight(bestWeights[i]);
mutableRules.get(i).setWeight(bestValidationWeights[i]);
}
}

Expand Down Expand Up @@ -681,7 +671,7 @@ protected void computeMAPStateWithWarmStart(InferenceApplication inferenceApplic
TermState[] warmStartTermState, float[] warmStartAtomValueState) {
// Warm start inference with previous termState.
inferenceApplication.getTermStore().loadState(warmStartTermState);
AtomStore atomStore = inferenceApplication.getDatabase().getAtomStore();
AtomStore atomStore = inferenceApplication.getTermStore().getAtomStore();
float[] atomValues = atomStore.getAtomValues();
for (int i = 0; i < atomStore.size(); i++) {
if (atomStore.getAtom(i).isFixed()) {
Expand All @@ -697,7 +687,7 @@ protected void computeMAPStateWithWarmStart(InferenceApplication inferenceApplic

// Save the MPE state for future warm starts.
inferenceApplication.getTermStore().saveState(warmStartTermState);
float[] mpeAtomValues = inferenceApplication.getDatabase().getAtomStore().getAtomValues();
float[] mpeAtomValues = inferenceApplication.getTermStore().getAtomStore().getAtomValues();
System.arraycopy(mpeAtomValues, 0, warmStartAtomValueState, 0, mpeAtomValues.length);
}

Expand All @@ -708,7 +698,7 @@ protected void computeCurrentIncompatibility(float[] incompatibilityArray) {
// Zero out the incompatibility first.
Arrays.fill(incompatibilityArray, 0.0f);

float[] atomValues = trainInferenceApplication.getDatabase().getAtomStore().getAtomValues();
float[] atomValues = trainInferenceApplication.getTermStore().getAtomStore().getAtomValues();

// Sums up the incompatibilities.
for (Object rawTerm : trainInferenceApplication.getTermStore()) {
Expand All @@ -732,7 +722,7 @@ protected void computeCurrentIncompatibility(float[] incompatibilityArray) {
* Method called at the start of every gradient descent iteration to
* compute statistics needed for loss and gradient computations.
*/
protected abstract void computeIterationStatistics();
protected abstract void computeIterationStatistics(int epoch);

/**
* Method for computing the total regularized loss.
Expand Down
4 changes: 2 additions & 2 deletions ...java/org/linqs/psl/application/learning/weight/gradient/minimizer/BinaryCrossEntropy.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -39,7 +39,7 @@ public BinaryCrossEntropy(List<Rule> rules, Database trainTargetDatabase, Databa

@Override
protected float computeSupervisedLoss() {
AtomStore atomStore = trainInferenceApplication.getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();

float supervisedLoss = 0.0f;
for (Map.Entry<RandomVariableAtom, ObservedAtom> entry: trainingMap.getLabelMap().entrySet()) {
Expand All @@ -58,7 +58,7 @@ protected float computeSupervisedLoss() {

@Override
protected void addSupervisedProxRuleObservedAtomValueGradient() {
AtomStore atomStore = trainInferenceApplication.getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();

for (Map.Entry<RandomVariableAtom, ObservedAtom> entry: trainingMap.getLabelMap().entrySet()) {
RandomVariableAtom randomVariableAtom = entry.getKey();
Expand Down
30 changes: 13 additions & 17 deletions ...src/main/java/org/linqs/psl/application/learning/weight/gradient/minimizer/Minimizer.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -81,7 +81,6 @@ public abstract class Minimizer extends GradientDescent {
protected float constraintTolerance;
protected float finalConstraintTolerance;

protected boolean initializedProxRuleConstants;
protected int outerIteration;

protected final float initialSquaredPenaltyCoefficient;
Expand Down Expand Up @@ -124,13 +123,12 @@ public Minimizer(List<Rule> rules, Database trainTargetDatabase, Database trainT
finalParameterMovementTolerance = Options.MINIMIZER_FINAL_PARAMETER_MOVEMENT_CONVERGENCE_TOLERANCE.getFloat();
constraintTolerance = (float)(1.0f / Math.pow(initialSquaredPenaltyCoefficient, 0.1f));
finalConstraintTolerance = Options.MINIMIZER_OBJECTIVE_DIFFERENCE_TOLERANCE.getFloat();
initializedProxRuleConstants = false;
outerIteration = 1;
}

@Override
protected void postInitGroundModel() {
AtomStore atomStore = trainInferenceApplication.getTermStore().getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();

// Create and add the augmented inference proximity terms.
int unFixedAtomCount = 0;
Expand Down Expand Up @@ -204,7 +202,7 @@ protected void postInitGroundModel() {
super.postInitGroundModel();

// Initialize latent and augmented inference warm start state objects.
float[] atomValues = trainInferenceApplication.getDatabase().getAtomStore().getAtomValues();
float[] atomValues = trainInferenceApplication.getTermStore().getAtomStore().getAtomValues();

latentInferenceTermState = trainInferenceApplication.getTermStore().saveState();
latentInferenceAtomValueState = Arrays.copyOf(atomValues, atomValues.length);
Expand Down Expand Up @@ -273,7 +271,7 @@ protected void gradientStep(int iteration) {
protected float internalParameterGradientStep(int iteration) {
float proxRuleObservedAtomsValueMovement = 0.0f;
// Take a step in the direction of the negative gradient of the proximity rule constants and project back onto box constraints.
float[] atomValues = trainInferenceApplication.getTermStore().getDatabase().getAtomStore().getAtomValues();
float[] atomValues = trainInferenceApplication.getTermStore().getAtomStore().getAtomValues();
for (int i = 0; i < proxRules.length; i++) {
float newProxRuleObservedAtomsValue = Math.min(Math.max(
proxRuleObservedAtoms[i].getValue() - proxRuleObservedAtomValueStepSize * proxRuleObservedAtomValueGradient[i], 0.0f), 1.0f);
Expand All @@ -291,13 +289,13 @@ protected void initializeProximityRuleConstants() {
// Initialize the proximity rule constants to the truth if it exists or the latent MAP state.
fixLabeledRandomVariables();

log.trace("Performing Latent Inference.");
log.trace("Running Latent Inference.");
computeMAPStateWithWarmStart(trainInferenceApplication, latentInferenceTermState, latentInferenceAtomValueState);
inTrainingMAPState = true;

unfixLabeledRandomVariables();

AtomStore atomStore = trainInferenceApplication.getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();
float[] atomValues = atomStore.getAtomValues();

System.arraycopy(latentInferenceAtomValueState, 0, augmentedInferenceAtomValueState, 0, latentInferenceAtomValueState.length);
Expand All @@ -320,15 +318,13 @@ protected void initializeProximityRuleConstants() {
atomValues[proxRuleObservedAtomIndexes[proxRuleIndex]] = observedAtom.getValue();
augmentedInferenceAtomValueState[proxRuleObservedAtomIndexes[proxRuleIndex]] = observedAtom.getValue();
}

initializedProxRuleConstants = true;
}

@Override
protected void computeIterationStatistics() {
protected void computeIterationStatistics(int epoch) {
computeFullInferenceStatistics();

if (!initializedProxRuleConstants) {
if (epoch == 0) {
initializeProximityRuleConstants();
}

Expand All @@ -341,7 +337,7 @@ protected void computeIterationStatistics() {
protected void computeTotalAtomGradient() {
float totalEnergyDifference = computeObjectiveDifference();

for (int i = 0; i < trainInferenceApplication.getDatabase().getAtomStore().size(); i++) {
for (int i = 0; i < trainInferenceApplication.getTermStore().getAtomStore().size(); i++) {
float rvGradientDifference = augmentedRVAtomGradient[i] - MAPRVAtomGradient[i];
float deepGradientDifference = augmentedDeepAtomGradient[i] - MAPDeepAtomGradient[i];

Expand Down Expand Up @@ -487,10 +483,10 @@ private float computeTotalEnergyDifference(float[] incompatibilityDifference){
}
}

GroundAtom[] atoms = trainInferenceApplication.getDatabase().getAtomStore().getAtoms();
GroundAtom[] atoms = trainInferenceApplication.getTermStore().getAtomStore().getAtoms();
float augmentedInferenceLCQPRegularization = 0.0f;
float fullInferenceLCQPRegularization = 0.0f;
for (int i = 0; i < trainInferenceApplication.getDatabase().getAtomStore().size(); i++) {
for (int i = 0; i < trainInferenceApplication.getTermStore().getAtomStore().size(); i++) {
if (atoms[i].isFixed()) {
continue;
}
Expand Down Expand Up @@ -543,7 +539,7 @@ protected void computeCurrentSquaredIncompatibility(float[] incompatibilityArray
// Zero out the incompatibility first.
Arrays.fill(incompatibilityArray, 0.0f);

float[] atomValues = trainInferenceApplication.getDatabase().getAtomStore().getAtomValues();
float[] atomValues = trainInferenceApplication.getTermStore().getAtomStore().getAtomValues();

// Sums up the incompatibilities.
for (Object rawTerm : trainInferenceApplication.getTermStore()) {
Expand All @@ -570,7 +566,7 @@ protected void computeCurrentSquaredIncompatibility(float[] incompatibilityArray
* with the same predicates and arguments having the same hash.
*/
protected void fixLabeledRandomVariables() {
AtomStore atomStore = trainInferenceApplication.getTermStore().getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();

for (Map.Entry<RandomVariableAtom, ObservedAtom> entry: trainingMap.getLabelMap().entrySet()) {
RandomVariableAtom randomVariableAtom = entry.getKey();
Expand All @@ -592,7 +588,7 @@ protected void fixLabeledRandomVariables() {
* with the same predicates and arguments having the same hash.
*/
protected void unfixLabeledRandomVariables() {
AtomStore atomStore = trainInferenceApplication.getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();

for (Map.Entry<RandomVariableAtom, ObservedAtom> entry: trainingMap.getLabelMap().entrySet()) {
RandomVariableAtom randomVariableAtom = entry.getKey();
Expand Down
4 changes: 2 additions & 2 deletions .../main/java/org/linqs/psl/application/learning/weight/gradient/minimizer/SquaredError.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -38,7 +38,7 @@ public SquaredError(List<Rule> rules, Database trainTargetDatabase, Database tra

@Override
protected float computeSupervisedLoss() {
AtomStore atomStore = trainInferenceApplication.getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();

float supervisedLoss = 0.0f;
for (Map.Entry<RandomVariableAtom, ObservedAtom> entry: trainingMap.getLabelMap().entrySet()) {
Expand All @@ -54,7 +54,7 @@ protected float computeSupervisedLoss() {

@Override
protected void addSupervisedProxRuleObservedAtomValueGradient() {
AtomStore atomStore = trainInferenceApplication.getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();

for (Map.Entry<RandomVariableAtom, ObservedAtom> entry: trainingMap.getLabelMap().entrySet()) {
RandomVariableAtom randomVariableAtom = entry.getKey();
Expand Down
2 changes: 1 addition & 1 deletion ...src/main/java/org/linqs/psl/application/learning/weight/gradient/optimalvalue/Energy.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -45,7 +45,7 @@ protected float computeLearningLoss() {
}

@Override
protected void computeIterationStatistics() {
protected void computeIterationStatistics(int epoch) {
computeLatentInferenceIncompatibility();
}

Expand Down
6 changes: 3 additions & 3 deletions ...in/java/org/linqs/psl/application/learning/weight/gradient/optimalvalue/OptimalValue.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -64,7 +64,7 @@ protected void postInitGroundModel() {

// Initialize latent inference warm start state objects.
latentInferenceTermState = trainInferenceApplication.getTermStore().saveState();
float[] atomValues = trainInferenceApplication.getDatabase().getAtomStore().getAtomValues();
float[] atomValues = trainInferenceApplication.getTermStore().getAtomStore().getAtomValues();
latentInferenceAtomValueState = Arrays.copyOf(atomValues, atomValues.length);

rvLatentAtomGradient = new float[atomValues.length];
Expand Down Expand Up @@ -98,7 +98,7 @@ protected void computeLatentInferenceIncompatibility() {
* with the same predicates and arguments having the same hash.
*/
protected void fixLabeledRandomVariables() {
AtomStore atomStore = trainInferenceApplication.getTermStore().getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();

for (Map.Entry<RandomVariableAtom, ObservedAtom> entry: trainingMap.getLabelMap().entrySet()) {
RandomVariableAtom randomVariableAtom = entry.getKey();
Expand All @@ -120,7 +120,7 @@ protected void fixLabeledRandomVariables() {
* with the same predicates and arguments having the same hash.
*/
protected void unfixLabeledRandomVariables() {
AtomStore atomStore = trainInferenceApplication.getDatabase().getAtomStore();
AtomStore atomStore = trainInferenceApplication.getTermStore().getAtomStore();

for (Map.Entry<RandomVariableAtom, ObservedAtom> entry: trainingMap.getLabelMap().entrySet()) {
RandomVariableAtom randomVariableAtom = entry.getKey();
Expand Down
2 changes: 1 addition & 1 deletion ...org/linqs/psl/application/learning/weight/gradient/optimalvalue/StructuredPerceptron.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -37,7 +37,7 @@ public StructuredPerceptron(List<Rule> rules, Database trainTargetDatabase, Data
}

@Override
protected void computeIterationStatistics() {
protected void computeIterationStatistics(int epoch) {
computeLatentInferenceIncompatibility();
computeFullInferenceIncompatibility();
}
Expand Down
Loading

0 comments on commit 1d54995

Please sign in to comment.