// FROM trainer
public QNModel(LogLikelihoodFunction monitor, double[] parameters)
: base(null, monitor.PredLabels, monitor.OutcomeLabels)
{
int[][] outcomePatterns = monitor.OutcomePatterns;
Context[] cParameters = new Context[monitor.PredLabels.Length];
for (int ci = 0; ci < parameters.Length; ci++)
{
int[] outcomePattern = outcomePatterns[ci];
double[] alpha = new double[outcomePattern.Length];
for (int oi = 0; oi < outcomePattern.Length; oi++)
{
alpha[oi] = parameters[ci + (outcomePattern[oi] * monitor.PredLabels.Length)];
}
cParameters[ci] = new Context(outcomePattern, alpha);
}
this.evalParams = new EvalParameters(cParameters, monitor.OutcomeLabels.Length);
this.prior = new UniformPrior();
this.modelType = ModelTypeEnum.MaxentQn;
this.parameters = parameters;
}
public virtual QNModel trainModel(DataIndexer indexer)
{
LogLikelihoodFunction objectiveFunction = generateFunction(indexer);
this.dimension = objectiveFunction.DomainDimension;
this.updateInfo = new QNInfo(this, this.m, this.dimension);
double[] initialPoint = objectiveFunction.InitialPoint;
double initialValue = objectiveFunction.valueAt(initialPoint);
double[] initialGrad = objectiveFunction.gradientAt(initialPoint);
LineSearchResult lsr = LineSearchResult.getInitialObject(initialValue, initialGrad, initialPoint, 0);
int z = 0;
while (true)
{
if (verbose)
{
Console.Write(z++);
}
double[] direction = null;
direction = computeDirection(objectiveFunction, lsr);
lsr = LineSearch.doLineSearch(objectiveFunction, direction, lsr, verbose);
updateInfo.updateInfo(lsr);
if (isConverged(lsr))
{
break;
}
}
return(new QNModel(objectiveFunction, lsr.NextPoint));
}
