@Override
public double getUnnormalizedLogProbability(Assignment assignment) {
Preconditions.checkArgument(assignment.containsAll(getVars().getVariableNumsArray()));
Tensor inputFeatureVector =
(Tensor) assignment.getValue(getInputVariable().getOnlyVariableNum());
if (conditionalVars.size() == 0) {
// No normalization for any conditioned-on variables. This case
// allows a more efficient implementation than the default
// in ClassifierFactor.
VariableNumMap outputVars = getOutputVariables();
Tensor outputTensor =
SparseTensor.singleElement(
outputVars.getVariableNumsArray(),
outputVars.getVariableSizes(),
outputVars.assignmentToIntArray(assignment),
1.0);
Tensor featureIndicator = outputTensor.outerProduct(inputFeatureVector);
return logWeights.innerProduct(featureIndicator).getByDimKey();
} else {
// Default to looking up the answer in the output log probabilities
int[] outputIndexes = getOutputVariables().assignmentToIntArray(assignment);
Tensor logProbs = getOutputLogProbTensor(inputFeatureVector);
return logProbs.getByDimKey(outputIndexes);
}
}
@Override
public void incrementSufficientStatisticsFromAssignment(
SufficientStatistics gradient,
SufficientStatistics currentParameters,
Assignment assignment,
double count) {
Preconditions.checkArgument(assignment.containsAll(getVars().getVariableNumsArray()));
Assignment subAssignment = assignment.intersection(getVars().getVariableNumsArray());
long keyNum =
initialWeights
.getWeights()
.dimKeyToKeyNum(initialWeights.getVars().assignmentToIntArray(subAssignment));
int index = initialWeights.getWeights().keyNumToIndex(keyNum);
((TensorSufficientStatistics) gradient).incrementFeatureByIndex(count, index);
}
@Override
public void log(long iteration, int exampleNum, Assignment example, FactorGraph graph) {
if (showExamples) {
if (iteration % logInterval == 0) {
String prob = "";
if (example.containsAll(graph.getVariables().getVariableNumsArray())) {
prob = Double.toString(graph.getUnnormalizedLogProbability(example));
}
print(
iteration
+ "."
+ exampleNum
+ " "
+ prob
+ ": example: "
+ graph.assignmentToObject(example));
}
}
}
@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);
}
}
}