/// <summary>
/// Gets the negative log-likelihood at the given input vector.
/// </summary>
/// <param name="x">The input vector.</param>
/// <returns>The negative log-likelihood.</returns>
/// <exception cref="ArgumentException">The <paramref name="x"/> is invalid, its dimension is not equal to domain dimension.</exception>
public virtual double ValueAt(double[] x)
{
if (x.Length != dimension)
{
throw new ArgumentException("x is invalid, its dimension is not equal to domain dimension.", nameof(x));
}
int ci;
double negLogLikelihood = 0;
for (ci = 0; ci < numContexts; ci++)
{
int oi;
for (oi = 0; oi < numOutcomes; oi++)
{
tempSums[oi] = 0;
int ai;
for (ai = 0; ai < contexts[ci].Length; ai++)
{
var vectorIndex = IndexOf(oi, contexts[ci][ai]);
var predValue = values != null ? values[ci][ai] : 1.0;
tempSums[oi] += predValue * x[vectorIndex];
}
}
var logSumOfExps = ArrayMath.LogSumOfExps(tempSums);
negLogLikelihood -= (tempSums[outcomeList[ci]] - logSumOfExps) * numTimesEventsSeen[ci];
}
return(negLogLikelihood);
}
private void NegLLCompute(int threadIndex, int startIndex, int length, double[] x)
{
negLogLikelihoodThread[threadIndex] = 0;
// Knuppe: In parallel we can't use the tempSums variable ;)
var temp = new double[numOutcomes];
for (var ci = startIndex; ci < startIndex + length; ci++)
{
for (var oi = 0; oi < numOutcomes; oi++)
{
temp[oi] = 0;
for (var ai = 0; ai < contexts[ci].Length; ai++)
{
var vectorIndex = IndexOf(oi, contexts[ci][ai]);
var predValue = values != null ? values[ci][ai] : 1.0;
temp[oi] += predValue * x[vectorIndex];
}
}
var logSumOfExps = ArrayMath.LogSumOfExps(temp);
var outcome = outcomeList[ci];
negLogLikelihoodThread[threadIndex] -= (temp[outcome] - logSumOfExps) * numTimesEventsSeen[ci];
}
}
/// <summary>
/// Gets the gradient at the given input vector.
/// </summary>
/// <param name="x">The input vector.</param>
/// <returns>The gradient value.</returns>
/// <exception cref="System.ArgumentException">x is invalid, its dimension is not equal to domain dimension.;x</exception>
/// <exception cref="ArgumentException">The <paramref name="x" /> is invalid, its dimension is not equal to domain dimension.</exception>
public virtual double[] GradientAt(double[] x)
{
if (x.Length != dimension)
{
throw new ArgumentException("x is invalid, its dimension is not equal to domain dimension.", nameof(x));
}
int ci;
// Reset gradient
for (var i = 0; i < gradient.Length; i++)
{
gradient[i] = 0;
}
for (ci = 0; ci < numContexts; ci++)
{
int oi;
double predValue;
int vectorIndex;
int ai;
for (oi = 0; oi < numOutcomes; oi++)
{
expectation[oi] = 0;
for (ai = 0; ai < contexts[ci].Length; ai++)
{
vectorIndex = IndexOf(oi, contexts[ci][ai]);
predValue = values != null ? values[ci][ai] : 1.0;
expectation[oi] += predValue * x[vectorIndex];
}
}
var logSumOfExps = ArrayMath.LogSumOfExps(expectation);
for (oi = 0; oi < numOutcomes; oi++)
{
expectation[oi] = Math.Exp(expectation[oi] - logSumOfExps);
}
for (oi = 0; oi < numOutcomes; oi++)
{
var empirical = outcomeList[ci] == oi ? 1 : 0;
for (ai = 0; ai < contexts[ci].Length; ai++)
{
vectorIndex = IndexOf(oi, contexts[ci][ai]);
predValue = values != null ? values[ci][ai] : 1.0;
gradient[vectorIndex] +=
predValue * (expectation[oi] - empirical) * numTimesEventsSeen[ci];
}
}
}
return(gradient);
}
private void GradientCompute(int threadIndex, int startIndex, int length, double[] x)
{
var exp = new double[numOutcomes];
// Reset gradientThread
Array.Clear(gradientThread[threadIndex], 0, gradientThread[threadIndex].Length);
for (var ci = startIndex; ci < startIndex + length; ci++)
{
double predValue;
int vectorIndex;
for (var oi = 0; oi < numOutcomes; oi++)
{
exp[oi] = 0;
for (var ai = 0; ai < contexts[ci].Length; ai++)
{
vectorIndex = IndexOf(oi, contexts[ci][ai]);
predValue = values != null ? values[ci][ai] : 1.0;
exp[oi] += predValue * x[vectorIndex];
}
}
var logSumOfExps = ArrayMath.LogSumOfExps(exp);
for (var oi = 0; oi < numOutcomes; oi++)
{
exp[oi] = Math.Exp(exp[oi] - logSumOfExps);
}
for (var oi = 0; oi < numOutcomes; oi++)
{
var empirical = outcomeList[ci] == oi ? 1 : 0;
for (var ai = 0; ai < contexts[ci].Length; ai++)
{
vectorIndex = IndexOf(oi, contexts[ci][ai]);
predValue = values != null ? values[ci][ai] : 1.0;
gradientThread[threadIndex][vectorIndex] += predValue * (exp[oi] - empirical) *
numTimesEventsSeen[ci];
}
}
}
}
/// <summary>
/// Model evaluation which should be used during training to report model accuracy.
/// </summary>
/// <param name="context">Indices of the predicates which have been observed at the present decision point.</param>
/// <param name="values">The weights of the predicates which have been observed at the present decision point.</param>
/// <param name="probs">The probability for outcomes.</param>
/// <param name="nOutcomes">The number of outcomes.</param>
/// <param name="nPredLabels">The number of unique predicates.</param>
/// <param name="parameters">The model parameters.</param>
/// <returns>The normalized probabilities for the outcomes given the context.</returns>
public static double[] Eval(int[] context, float[] values, double[] probs, int nOutcomes, int nPredLabels, double[] parameters)
{
for (var i = 0; i < context.Length; i++)
{
var predIdx = context[i];
var predValue = values != null ? values[i] : 1d;
for (var oi = 0; oi < nOutcomes; oi++)
{
probs[oi] += predValue * parameters[oi * nPredLabels + predIdx];
}
}
var logSumExp = ArrayMath.LogSumOfExps(probs);
for (var oi = 0; oi < nOutcomes; oi++)
{
probs[oi] = Math.Exp(probs[oi] - logSumExp);
}
return(probs);
}
/// <summary>
/// Model evaluation which should be used during inference.
/// </summary>
/// <param name="context">The predicates which have been observed at the present decision point.</param>
/// <param name="values">The weights of the predicates which have been observed at the present decision point.</param>
/// <param name="probs">The probability for outcomes.</param>
/// <returns>The normalized probabilities for the outcomes given the context.</returns>
public double[] Eval(string[] context, float[] values, double[] probs)
{
var ep = evalParameters.Parameters;
for (var ci = 0; ci < context.Length; ci++)
{
var predIdx = GetPredIndex(context[ci]);
if (predIdx < 0)
{
continue;
}
var predValue = 1d;
if (values != null)
{
predValue = values[ci];
}
var outcomes = ep[predIdx].Outcomes;
var parameters = ep[predIdx].Parameters;
for (var i = 0; i < outcomes.Length; i++)
{
probs[outcomes[i]] += predValue * parameters[i];
}
}
var logSumExp = ArrayMath.LogSumOfExps(probs);
for (var oi = 0; oi < outcomeNames.Length; oi++)
{
probs[oi] = Math.Exp(probs[oi] - logSumExp);
}
return(probs);
}
