/// <summary>
/// Constructs a new potential function modeling Hidden Markov Models.
/// </summary>
///
/// <param name="states">The number of states.</param>
/// <param name="symbols">The number of symbols.</param>
/// <param name="outputClasses">The number of output classes.</param>
///
public DiscreteMarkovClassifierFunction(int states, int symbols, int outputClasses)
{
this.Outputs = outputClasses;
this.Symbols = symbols;
int factorIndex = 0;
var factorParams = new List <double>();
var factorFeatures = new List <IFeature <int> >();
this.Factors = new FactorPotential <int> [Outputs];
// Create features for initial class probabilities
for (int c = 0; c < outputClasses; c++)
{
var stateParams = new List <double>();
var stateFeatures = new List <IFeature <int> >();
var edgeParams = new List <double>();
var edgeFeatures = new List <IFeature <int> >();
var classParams = new List <double>();
var classFeatures = new List <IFeature <int> >();
// Create features for class labels
classParams.Add(Math.Log(1.0 / outputClasses));
classFeatures.Add(new OutputFeature <int>(this, c, c));
// Create features for initial state probabilities
for (int i = 0; i < states; i++)
{
edgeParams.Add((i == 0) ? Math.Log(1.0) : Math.Log(0.0));
edgeFeatures.Add(new InitialFeature <int>(this, c, i));
}
// Create features for state transition probabilities
for (int i = 0; i < states; i++)
{
for (int j = 0; j < states; j++)
{
edgeParams.Add(Math.Log(1.0 / states));
edgeFeatures.Add(new TransitionFeature <int>(this, c, i, j));
}
}
// Create features for symbol emission probabilities
for (int i = 0; i < states; i++)
{
for (int k = 0; k < symbols; k++)
{
stateParams.Add(Math.Log(1.0 / symbols));
stateFeatures.Add(new EmissionFeature(this, c, i, k));
}
}
int startClassIndex = factorIndex;
int startEdgeIndex = factorIndex + classParams.Count;
int startStateIndex = factorIndex + classParams.Count + edgeParams.Count;
// First features and params are always belonging to classes
Factors[c] = new DiscreteMarkovModelFactor(this, states, c, symbols,
startClassIndex, classParams.Count, // 1. classes
startEdgeIndex, edgeParams.Count, // 2. edges
startStateIndex, stateParams.Count); // 3. states
// 1. classes
factorFeatures.AddRange(classFeatures);
factorParams.AddRange(classParams);
// 2. edges
factorFeatures.AddRange(edgeFeatures);
factorParams.AddRange(edgeParams);
// 3. states
factorFeatures.AddRange(stateFeatures);
factorParams.AddRange(stateParams);
factorIndex += classParams.Count + stateParams.Count + edgeParams.Count;
}
System.Diagnostics.Debug.Assert(factorIndex == factorParams.Count);
System.Diagnostics.Debug.Assert(factorIndex == factorFeatures.Count);
this.Weights = factorParams.ToArray();
this.Features = factorFeatures.ToArray();
}
/// <summary>
/// Constructs a new potential function modeling Hidden Markov Models.
/// </summary>
///
/// <param name="classifier">The classifier model.</param>
/// <param name="includeClassFeatures">True to include class features (priors), false otherwise.</param>
///
public DiscreteMarkovClassifierFunction(HiddenMarkovClassifier classifier, bool includeClassFeatures = true)
{
this.Symbols = classifier.Symbols;
this.Outputs = classifier.Classes;
int factorIndex = 0;
var factorParams = new List <double>();
var factorFeatures = new List <IFeature <int> >();
this.Factors = new FactorPotential <int> [Outputs];
// Create features for initial class probabilities
for (int c = 0; c < classifier.Classes; c++)
{
var stateParams = new List <double>();
var stateFeatures = new List <IFeature <int> >();
var edgeParams = new List <double>();
var edgeFeatures = new List <IFeature <int> >();
var classParams = new List <double>();
var classFeatures = new List <IFeature <int> >();
var model = classifier[c];
if (includeClassFeatures)
{
// Create features for class labels
classParams.Add(Math.Log(classifier.Priors[c]));
classFeatures.Add(new OutputFeature <int>(this, c, c));
}
// Create features for initial state probabilities
for (int i = 0; i < model.States; i++)
{
edgeParams.Add(model.Probabilities[i]);
edgeFeatures.Add(new InitialFeature <int>(this, c, i));
}
// Create features for state transition probabilities
for (int i = 0; i < model.States; i++)
{
for (int j = 0; j < model.States; j++)
{
edgeParams.Add(model.Transitions[i, j]);
edgeFeatures.Add(new TransitionFeature <int>(this, c, i, j));
}
}
// Create features for symbol emission probabilities
for (int i = 0; i < model.States; i++)
{
for (int k = 0; k < model.Symbols; k++)
{
stateParams.Add(model.Emissions[i, k]);
stateFeatures.Add(new EmissionFeature(this, c, i, k));
}
}
int startClassIndex = factorIndex;
int startEdgeIndex = factorIndex + classParams.Count;
int startStateIndex = factorIndex + classParams.Count + edgeParams.Count;
// First features and params are always belonging to classes
Factors[c] = new DiscreteMarkovModelFactor(this, model.States, c, Symbols,
startClassIndex, classParams.Count, // 1. classes
startEdgeIndex, edgeParams.Count, // 2. edges
startStateIndex, stateParams.Count); // 3. states
// 1. classes
factorFeatures.AddRange(classFeatures);
factorParams.AddRange(classParams);
// 2. edges
factorFeatures.AddRange(edgeFeatures);
factorParams.AddRange(edgeParams);
// 3. states
factorFeatures.AddRange(stateFeatures);
factorParams.AddRange(stateParams);
factorIndex += classParams.Count + stateParams.Count + edgeParams.Count;
}
System.Diagnostics.Debug.Assert(factorIndex == factorParams.Count);
System.Diagnostics.Debug.Assert(factorIndex == factorFeatures.Count);
this.Weights = factorParams.ToArray();
this.Features = factorFeatures.ToArray();
}