@Override
public String getParameterDescription(SufficientStatistics parameters, int numFeatures) {
Tensor featureWeights = getFeatureWeights(parameters);
TableFactor featureValues =
new TableFactor(
initialWeights.getVars(),
initialWeights.getWeights().replaceValues(featureWeights.getValues()));
List<Assignment> biggestAssignments =
featureValues.product(featureValues).getMostLikelyAssignments(numFeatures);
return featureValues.describeAssignments(biggestAssignments);
}
/**
* If this tree contains max-marginals, recover the best parse subtree for a given symbol with the
* specified span.
*/
public CfgParseTree getBestParseTreeWithSpan(Object root, int spanStart, int spanEnd) {
Preconditions.checkState(!sumProduct);
Assignment rootAssignment = parentVar.outcomeArrayToAssignment(root);
int rootNonterminalNum = parentVar.assignmentToIntArray(rootAssignment)[0];
double prob =
insideChart[spanStart][spanEnd][rootNonterminalNum]
* outsideChart[spanStart][spanEnd][rootNonterminalNum];
if (prob == 0.0) {
return null;
}
int splitInd = splitBackpointers[spanStart][spanEnd][rootNonterminalNum];
if (splitInd < 0) {
long terminalKey = backpointers[spanStart][spanEnd][rootNonterminalNum];
int positiveSplitInd = (-1 * splitInd) - 1;
int terminalSpanStart = positiveSplitInd / numTerminals;
int terminalSpanEnd = positiveSplitInd % numTerminals;
// This is a really sucky way to transform the keys back to objects.
VariableNumMap vars = parentVar.union(ruleTypeVar);
int[] dimKey = TableFactor.zero(vars).getWeights().keyNumToDimKey(terminalKey);
Assignment a = vars.intArrayToAssignment(dimKey);
Object ruleType = a.getValue(ruleTypeVar.getOnlyVariableNum());
List<Object> terminalList = Lists.newArrayList();
terminalList.addAll(terminals.subList(terminalSpanStart, terminalSpanEnd + 1));
return new CfgParseTree(root, ruleType, terminalList, prob, spanStart, spanEnd);
} else {
long binaryRuleKey = backpointers[spanStart][spanEnd][rootNonterminalNum];
int[] binaryRuleComponents =
binaryRuleDistribution.coerceToDiscrete().getWeights().keyNumToDimKey(binaryRuleKey);
Assignment best = binaryRuleDistribution.getVars().intArrayToAssignment(binaryRuleComponents);
Object leftRoot = best.getValue(leftVar.getOnlyVariableNum());
Object rightRoot = best.getValue(rightVar.getOnlyVariableNum());
Object ruleType = best.getValue(ruleTypeVar.getOnlyVariableNum());
Preconditions.checkArgument(
spanStart + splitInd != spanEnd,
"CFG parse decoding error: %s %s %s",
spanStart,
spanEnd,
splitInd);
CfgParseTree leftTree = getBestParseTreeWithSpan(leftRoot, spanStart, spanStart + splitInd);
CfgParseTree rightTree =
getBestParseTreeWithSpan(rightRoot, spanStart + splitInd + 1, spanEnd);
Preconditions.checkState(leftTree != null);
Preconditions.checkState(rightTree != null);
return new CfgParseTree(root, ruleType, leftTree, rightTree, prob);
}
}
@Override
public void incrementSufficientStatisticsFromMarginal(
SufficientStatistics gradient,
SufficientStatistics currentParameters,
Factor marginal,
Assignment conditionalAssignment,
double count,
double partitionFunction) {
if (conditionalAssignment.containsAll(getVars().getVariableNumsArray())) {
// Short-circuit the slow computation below if possible.
double multiplier = marginal.getTotalUnnormalizedProbability() * count / partitionFunction;
incrementSufficientStatisticsFromAssignment(
gradient, currentParameters, conditionalAssignment, multiplier);
} else {
VariableNumMap conditionedVars =
initialWeights.getVars().intersection(conditionalAssignment.getVariableNumsArray());
TableFactor productFactor =
(TableFactor)
initialWeights
.product(
TableFactor.pointDistribution(
conditionedVars, conditionalAssignment.intersection(conditionedVars)))
.product(marginal);
Tensor productFactorWeights = productFactor.getWeights();
double[] productFactorValues = productFactorWeights.getValues();
int tensorSize = productFactorWeights.size();
double multiplier = count / partitionFunction;
TensorSufficientStatistics tensorGradient = (TensorSufficientStatistics) gradient;
for (int i = 0; i < tensorSize; i++) {
int builderIndex = (int) productFactorWeights.indexToKeyNum(i);
tensorGradient.incrementFeatureByIndex(productFactorValues[i] * multiplier, builderIndex);
}
}
}
private void runTestAllFeatures(DenseIndicatorLogLinearFactor parametricFactor) {
SufficientStatistics parameters = parametricFactor.getNewSufficientStatistics();
Factor initial = parametricFactor.getModelFromParameters(parameters);
assertEquals(0.0, initial.getUnnormalizedLogProbability("A", "T"), TOLERANCE);
assertEquals(0.0, initial.getUnnormalizedLogProbability("B", "T"), TOLERANCE);
assertEquals(0.0, initial.getUnnormalizedLogProbability("C", "F"), TOLERANCE);
parametricFactor.incrementSufficientStatisticsFromAssignment(
parameters, parameters, vars.outcomeArrayToAssignment("B", "F"), 2.0);
parametricFactor.incrementSufficientStatisticsFromAssignment(
parameters, parameters, vars.outcomeArrayToAssignment("B", "F"), -3.0);
parametricFactor.incrementSufficientStatisticsFromAssignment(
parameters, parameters, vars.outcomeArrayToAssignment("A", "T"), 1.0);
parametricFactor.incrementSufficientStatisticsFromAssignment(
parameters, parameters, vars.outcomeArrayToAssignment("C", "T"), 2.0);
Factor factor = parametricFactor.getModelFromParameters(parameters);
assertEquals(-1.0, factor.getUnnormalizedLogProbability("B", "F"), TOLERANCE);
assertEquals(1.0, factor.getUnnormalizedLogProbability("A", "T"), TOLERANCE);
assertEquals(2.0, factor.getUnnormalizedLogProbability("C", "T"), TOLERANCE);
TableFactorBuilder incrementBuilder =
new TableFactorBuilder(vars, SparseTensorBuilder.getFactory());
incrementBuilder.setWeight(4.0, "A", "F");
incrementBuilder.setWeight(6.0, "C", "F");
Factor increment = incrementBuilder.build();
parametricFactor.incrementSufficientStatisticsFromMarginal(
parameters, parameters, increment, Assignment.EMPTY, 3.0, 2.0);
factor = parametricFactor.getModelFromParameters(parameters);
assertEquals(-1.0, factor.getUnnormalizedLogProbability("B", "F"), TOLERANCE);
assertEquals(1.0, factor.getUnnormalizedLogProbability("A", "T"), TOLERANCE);
assertEquals(2.0, factor.getUnnormalizedLogProbability("C", "T"), TOLERANCE);
assertEquals(6.0, factor.getUnnormalizedLogProbability("A", "F"), TOLERANCE);
assertEquals(9.0, factor.getUnnormalizedLogProbability("C", "F"), TOLERANCE);
TableFactor pointDist =
TableFactor.logPointDistribution(truthVar, truthVar.outcomeArrayToAssignment("T"));
parametricFactor.incrementSufficientStatisticsFromMarginal(
parameters, parameters, pointDist, alphabetVar.outcomeArrayToAssignment("B"), 3, 2.0);
factor = parametricFactor.getModelFromParameters(parameters);
assertEquals(1.5, factor.getUnnormalizedLogProbability("B", "T"), TOLERANCE);
}
/**
* Create a parse chart with the specified number of terminal symbols.
*
* <p>If "sumProduct" is true, then updates for the same symbol add (i.e., the ParseChart computes
* marginals). Otherwise, updates use the maximum probability, meaning ParseChart computes
* max-marginals.
*/
public CfgParseChart(
List<?> terminals,
VariableNumMap parent,
VariableNumMap left,
VariableNumMap right,
VariableNumMap terminal,
VariableNumMap ruleTypeVar,
Factor binaryRuleDistribution,
boolean sumProduct) {
this.terminals = terminals;
this.parentVar = parent;
this.leftVar = left;
this.rightVar = right;
this.ruleTypeVar = ruleTypeVar;
this.terminalVar = terminal;
this.sumProduct = sumProduct;
this.numTerminals = terminals.size();
this.numNonterminals = parentVar.getDiscreteVariables().get(0).numValues();
insideChart = new double[numTerminals][numTerminals][numNonterminals];
outsideChart = new double[numTerminals][numTerminals][numNonterminals];
this.binaryRuleDistribution = binaryRuleDistribution;
binaryRuleExpectations =
new double[binaryRuleDistribution.coerceToDiscrete().getWeights().getValues().length];
terminalRuleExpectations =
TableFactor.zero(VariableNumMap.unionAll(parentVar, terminalVar, ruleTypeVar));
insideCalculated = false;
outsideCalculated = false;
partitionFunction = 0.0;
if (!sumProduct) {
backpointers = new long[numTerminals][numTerminals][numNonterminals];
splitBackpointers = new int[numTerminals][numTerminals][numNonterminals];
} else {
backpointers = null;
splitBackpointers = null;
}
}
public static IndicatorLogLinearFactor createDenseFactor(VariableNumMap variables) {
return new IndicatorLogLinearFactor(variables, TableFactor.unity(variables));
}