internal void GetBestTagSequence(IList <FeatureVector> vectors, out int[] sysClasses, out double[] distribution)
{
// Since we are working with logarithmic numbers, we want the weight of the root node to be zero.
var beam = new BeamSearch <IdValuePair <double> >(0D);
for (int beamDepth = 0; beamDepth < vectors.Count; beamDepth++)
{
Debug.Assert(beam.Level[beamDepth].Count > 0);
foreach (BeamNode <IdValuePair <double> > node in beam.Level[beamDepth])
{
double[] probs_v_c = new double[classToClassId.Count];
for (int c_i = 0; c_i < classToClassId.Count; c_i++)
{
probs_v_c[c_i] = CalculateProbability_v_c(vectors[beamDepth], c_i, node, beamDepth);
}
NormalizationHelper.NormalizeLogs(probs_v_c, Math.E);
// Prune: Idenitify N classes with highest probability:
IList <int> topNClasses = SearchHelper.GetMaxNItems(topN, probs_v_c);
for (int topN_i = 0; topN_i < topNClasses.Count; topN_i++)
{
int c_i = topNClasses[topN_i];
double prob = probs_v_c[c_i];
node.AddNextNode(new IdValuePair <double>(c_i, prob), Math.Log(prob, Math.E) + node.Weight);
}
}
beam.Prune(topK, beam_size);
}
// Repackage the sequence we just received in such a way that the consuming code will find it most digestable.
var results = beam.GetBestSequence();
sysClasses = new int[results.Length];
distribution = new double[results.Length];
for (int i = 0; i < results.Length; i++)
{
sysClasses[i] = results[i].Id;
distribution[i] = results[i].Value;
}
}