public void Init(LLNAModel model)
{
var phiVal = (double)1 / (double)model.K;
var logPhiVal = -Math.Log(model.K);
Phi.SetValue(phiVal);
LogPhi.SetValue(logPhiVal);
Zeta = 10;
int i = 0;
for (i = 0; i < (model.K - 1); i++)
{
Nu[i] = 10;
Lambda[i] = 0;
}
Nu[i] = 0;
Lambda[i] = 0;
NIter = 0;
LHood = 0;
for (i = 0; i < 4; i++)
{
_temp.Add(Vector <double> .Build.Dense(model.K - 1));
}
}
public void OptimizePhi(LLNAModel model, CorpusDocument doc)
{
// Compute phi proportion in log space.
for (int n = 0; n < doc.NbTerms; n++)
{
double logSumN = 0;
for (int i = 0; i < model.K; i++)
{
var word = doc.GetWord(n);
LogPhi[n, i] = Lambda[i] + model.LogBeta[i, word];
if (i == 0)
{
logSumN = LogPhi[n, i];
}
else
{
logSumN = MathHelper.LogSum(logSumN, LogPhi[n, i]);
}
}
for (int i = 0; i < model.K; i++)
{
LogPhi[n, i] = LogPhi[n, i] - logSumN;
Phi[n, i] = Math.Exp(LogPhi[n, i]);
}
}
}
public void OptimizeNu(LLNAModel model, CorpusDocument doc)
{
for (int i = 0; i < model.K - 1; i++)
{
double initNu = 10;
double nuI = 0, logNuI = 0, df = 0, d2f = 0;
int iter = 0;
logNuI = Math.Log(initNu);
do
{
iter++;
nuI = Math.Exp(logNuI);
if (double.IsNaN(nuI))
{
initNu = initNu * 2;
logNuI = Math.Log(initNu);
nuI = initNu;
}
df = -(model.InvCovariance[i, i] * 0.5)
- (0.5 * ((doc.Total / Zeta) * Math.Exp(Lambda[i] + nuI / 2)))
+ (0.5 * (1 / nuI));
d2f = -(0.25 * (doc.Total / Zeta) * Math.Exp(Lambda[i] + nuI / 2))
- (0.5 * (1 / (nuI * nuI)));
logNuI = logNuI - (df * nuI) / (d2f * nuI * nuI + df * nuI);
}while (Math.Abs(df) > NEWTON_THRESH);
Nu[i] = Math.Exp(logNuI);
}
}
public void OptimizeZeta(LLNAModel model)
{
Zeta = 1;
for (int i = 0; i < model.K - 1; i++)
{
Zeta += Math.Exp(Lambda[i] + 0.5 * (Nu[i]));
}
}
public LLNASufficientStatistics(LLNAModel model)
{
MuSufficientStatistic = new double[model.K - 1];
CovarianceSufficientStatistic = Matrix <double> .Build.Dense(model.K - 1, model.K - 1, 0);
BetaSufficientStatistic = Matrix <double> .Build.Dense(model.K, model.LogBeta.ColumnCount, 0);
NData = 0;
}
public Bundle(int k, CorpusDocument document, LLNAModel model, VariationalInferenceParameter inference)
{
SumPhi = Vector <double> .Build.Dense(k - 1, 0);
Document = document;
Model = model;
VarInference = inference;
Init(k);
}
public void OptimizeLambda(LLNAModel model, CorpusDocument doc)
{
int i = 0, n = model.K - 1, iter = 0;
double fOld = 0;
var bundle = new Bundle(model.K, doc, model, this);
var x = Vector <double> .Build.Dense(model.K - 1);
for (i = 0; i < model.K - 1; i++)
{
x[i] = Lambda[i];
}
/*
* var iterator = new MultidimensionalIterator(n);
* var fdf = new MultidimensionalFunctionFDF
* {
* F = (vector) =>
* {
* return ComputeFunction(vector, bundle);
* },
* DF = (vector) =>
* {
* return ComputeGradient(vector, bundle);
* },
* N = n
* };
* var multidimensionalMinimizer = new MultidimensionalMinimizer(n, iterator, fdf);
* multidimensionalMinimizer.Set(x, 0.01, 1e-3);
* do
* {
* iter++;
* fOld = multidimensionalMinimizer.F;
* multidimensionalMinimizer.Iterate();
* }
* while (true);
*/
var obj = ObjectiveFunction.Gradient((vector) => {
return(ComputeFunction(vector, bundle));
}, (vector) =>
{
return(ComputeGradient(vector, bundle));
});
var solver = new ConjugateGradientMinimizer(1e-8, 5000);
var result = solver.FindMinimum(obj, x);
for (i = 0; i < model.K - 1; i++)
{
Lambda[i] = result.MinimizingPoint[i];
}
Lambda[i] = 0;
}
public static LLNAModel Create(int k, Corpus corpus, CorrelatedTopicModelParameters parameters)
{
var result = new LLNAModel(parameters)
{
K = k,
Mu = Vector <double> .Build.Dense(k - 1),
LogBeta = Matrix <double> .Build.Dense(k, corpus.VocabularySize, 0),
};
result.Covariance = Matrix <double> .Build.DenseIdentity(k - 1, k - 1);
result.InvCovariance = result.Covariance.Inverse();
result.InvCovariance.Display();
result.LogDeterminantInvCovariance = result.InvCovariance.ComputeLNDeterminant();
result.Init(corpus);
return(result);
}
public void UpdateLikelihoodBound(CorpusDocument doc, LLNAModel model)
{
TopicScores.SetValue(0);
// p(η | μ, Σ) (distribution of topic proportions) && q(η | λ, ν)
// E[log p(\eta | \mu, \Sigma)] + H(q(\eta | \lambda, \nu)
double lHood = 0.5 * model.LogDeterminantInvCovariance + 0.5 * (model.K - 1);
for (int i = 0; i < (model.K - 1); i++)
{
double v = -0.5 * Nu[i] * model.InvCovariance[i, i];
for (int j = 0; j < model.K - 1; j++)
{
v -= 0.5 *
(Lambda[i] - model.Mu[i]) *
model.InvCovariance[i, j] *
(Lambda[j] - model.Mu[j]);
}
v += 0.5 * Math.Log(Nu[i]);
lHood += v;
}
// E[log p(z_n | \eta)] + E[log p(w_n | \beta)] + H(q(z_n | \phi_n))
lHood -= CalculateLikelihoodBound() * doc.Total;
for (int i = 0; i < doc.NbTerms; i++)
{
for (int j = 0; j < model.K; j++)
{
var phi = Phi[i, j];
var logPhi = LogPhi[i, j];
if (phi > 0)
{
TopicScores[j] += phi * doc.GetCount(i);
double a1 = Lambda[j];
int ww = doc.GetWord(i);
double a2 = model.LogBeta[j, doc.GetWord(i)];
var ss = doc.GetCount(i) * phi *
((Lambda[j] + model.LogBeta[j, doc.GetWord(i)]) - logPhi);
lHood += ss;
}
}
}
LHood = lHood;
}