private double Inference(
VariationalInferenceParameter varInference,
CorpusDocument doc,
LLNAModel model)
{
double oldLHood = 0, convergence = 0;
varInference.UpdateLikelihoodBound(doc, model);
do
{
varInference.IncrementIter();
varInference.OptimizeZeta(model);
varInference.OptimizeLambda(model, doc);
varInference.OptimizeZeta(model);
varInference.OptimizeNu(model, doc);
varInference.OptimizeZeta(model);
varInference.OptimizePhi(model, doc);
oldLHood = varInference.LHood;
varInference.UpdateLikelihoodBound(doc, model);
convergence = Math.Abs((oldLHood - varInference.LHood) / oldLHood);
}while ((convergence > _parameters.VarConvergence) && (_parameters.VarMaxIter < 0 || varInference.NIter < _parameters.VarMaxIter));
if (convergence > _parameters.VarConvergence)
{
varInference.Converged = false;
}
else
{
varInference.Converged = true;
}
return(varInference.LHood);
}
private static void DisplayModel(LLNAModel model)
{
Console.WriteLine("Covariance between topics");
for (int i = 0; i < model.Covariance.RowCount; i++)
{
for (int j = 0; j < model.Covariance.ColumnCount; j++)
{
Console.Write($"{model.Covariance[i, j]} ");
}
Console.WriteLine();
}
}
internal void Fit(IDataView input)
{
double convergence = 1, lHoodOld = 0;
bool resetVar = true;
var corpus = ExtractCorpus(input);
var llnaModel = LLNAModel.Create(_nbTopics, corpus, _parameters);
_model = llnaModel;
var sufficientStatistic = new LLNASufficientStatistics(llnaModel);
var corpusLambda = Matrix <double> .Build.Dense(corpus.NbDocuments, llnaModel.K, 0);
var corpusNu = Matrix <double> .Build.Dense(corpus.NbDocuments, llnaModel.K, 0);
var corpusPhiSum = Matrix <double> .Build.Dense(corpus.NbDocuments, llnaModel.K, 0);
int iteration = 0;
// Expectation - maximization algorithm (EM) algorithm.
do
{
var result = Expectation(corpus,
llnaModel,
sufficientStatistic,
corpusLambda,
corpusNu,
corpusPhiSum,
resetVar);
convergence = (lHoodOld - result.LHood) / lHoodOld;
if (convergence < 0)
{
resetVar = false;
if (_parameters.VarMaxIter > 0)
{
_parameters.VarMaxIter += 10;
}
else
{
_parameters.VarConvergence /= 10;
}
}
else
{
llnaModel.Maximize(sufficientStatistic);
lHoodOld = result.LHood;
resetVar = true;
iteration++;
}
sufficientStatistic.Reset();
}while ((iteration < _parameters.EmMaxIter) && ((convergence > _parameters.EmConvergence) || (convergence < 0)));
}
private ExpectationResult Expectation(
Corpus corpus,
LLNAModel llnaModel,
LLNASufficientStatistics llnaSufficientStatistics,
Matrix <double> corpusLambda,
Matrix <double> corpusNu,
Matrix <double> corpusPhiSum,
bool resetVar)
{
double avgNIter = 0, avgConverged = 0;
double total = 0;
for (int i = 0; i < corpus.NbDocuments; i++)
{
Debug.WriteLine($"Document {i}");
var doc = corpus.GetDocument(i);
var varInference = new VariationalInferenceParameter(doc.NbTerms, llnaModel.K);
if (resetVar)
{
varInference.Init(llnaModel);
}
else
{
varInference.Lambda = corpusLambda.Row(i);
varInference.Nu = corpusNu.Row(i);
varInference.OptimizeZeta(llnaModel);
varInference.OptimizePhi(llnaModel, doc);
varInference.NIter = 0;
}
var lHood = Inference(varInference, doc, llnaModel);
llnaSufficientStatistics.Update(varInference, doc);
total += lHood;
avgNIter += varInference.NIter;
if (varInference.Converged)
{
avgConverged++;
}
corpusLambda.SetRow(i, varInference.Lambda);
corpusNu.SetRow(i, varInference.Nu);
var phiSum = varInference.Phi.ColumnSum();
corpusPhiSum.SetRow(i, phiSum);
}
avgNIter = avgNIter / corpus.NbDocuments;
avgConverged = avgConverged / corpus.NbDocuments;
return(new ExpectationResult(avgNIter, avgConverged, total));
}
