/// <summary>
/// Initializes a new instance of the <see cref="ChunkerME"/> with the specified <see cref="ChunkerModel"/>.
/// </summary>
/// <param name="model">The chunker model.</param>
public ChunkerME(ChunkerModel model)
{
contextGenerator = model.Factory.GetContextGenerator();
sequenceValidator = model.Factory.GetSequenceValidator();
this.model = model.ChunkerSequenceModel ?? new BeamSearch(model.BeamSize, model.MaxentModel);
}
/// <summary>
/// Initializes the current instance with the specified model and the specified beam size.
/// </summary>
/// <param name="model">The model for this chunker</param>
/// <param name="beamSize">The size of the beam that should be used when decoding sequences.</param>
/// <param name="sequenceValidator">The <see cref="ISequenceValidator{String}"/> to determines whether the outcome is valid for the preceding sequence. This can be used to implement constraints on what sequences are valid..</param>
/// <param name="contextGenerator">The context generator.</param>
internal ChunkerME(ChunkerModel model, int beamSize, ISequenceValidator<string> sequenceValidator, IChunkerContextGenerator contextGenerator) {
// This method is marked as deprecated in the OpenNLP, but it is required in the Parser,
// I could change the cg in the factory, but its not ideal in this situation (i think) :P
this.sequenceValidator = sequenceValidator;
this.contextGenerator = contextGenerator;
this.model = model.ChunkerSequenceModel ?? new BeamSearch(beamSize, model.MaxentModel);
}
/// <summary>
/// Initializes the current instance with the specified model and the specified beam size.
/// </summary>
/// <param name="model">The model for this chunker</param>
/// <param name="beamSize">The size of the beam that should be used when decoding sequences.</param>
/// <param name="sequenceValidator">The <see cref="ISequenceValidator{String}"/> to determines whether the outcome is valid for the preceding sequence. This can be used to implement constraints on what sequences are valid..</param>
/// <param name="contextGenerator">The context generator.</param>
internal ChunkerME(ChunkerModel model, int beamSize, ISequenceValidator <string> sequenceValidator, IChunkerContextGenerator contextGenerator)
{
// This method is marked as deprecated in the OpenNLP, but it is required in the Parser,
// I could change the cg in the factory, but its not ideal in this situation (i think) :P
this.sequenceValidator = sequenceValidator;
this.contextGenerator = contextGenerator;
this.model = model.ChunkerSequenceModel ?? new BeamSearch(beamSize, model.MaxentModel);
}
/// <summary>
/// Finds the sequence with the highest probability.
/// </summary>
/// <param name="sequence">The sequence.</param>
/// <param name="additionalContext">The additional context.</param>
/// <param name="beamSearch">The beam search.</param>
/// <param name="validator">The validator.</param>
/// <returns>The sequence with the highest probability.</returns>
public Sequence BestSequence(T[] sequence, object[] additionalContext,
IBeamSearchContextGenerator <T> beamSearch,
ISequenceValidator <T> validator)
{
var sequences = BestSequences(1, sequence, additionalContext, beamSearch, validator);
if (sequence.Length > 0)
{
return(sequences[0]);
}
return(null);
}
public LemmatizerME(LemmatizerModel model)
{
if (model == null)
{
throw new ArgumentNullException(nameof(model));
}
contextGenerator = model.Factory.GetContextGenerator();
sequenceValidator = model.Factory.GetSequenceValidator();
// Knuppe: In the original implementation there is condition to recreate the beamsearch object, but
// the condition is impossible to occur, due to the getLemmatizerSequenceModel() method logic
this.model = model.LemmatizerSequenceModel;
}
/// <summary>
/// Creates new search object.
/// </summary>
/// <param name="size">The size of the beam (k).</param>
/// <param name="cg">The context generator for the model.</param>
/// <param name="model">The model for assigning probabilities to the sequence outcomes.</param>
/// <param name="validator">The sequence validator.</param>
/// <param name="cacheSize">Size of the cache.</param>
public BeamSearch(int size, IBeamSearchContextGenerator <T> cg, IMaxentModel model, ISequenceValidator <T> validator, int cacheSize)
{
this.cg = cg;
this.size = size;
this.model = model;
this.validator = validator;
if (cacheSize > 0)
{
contextsCache = new Cache(cacheSize);
}
probs = new double[model.GetNumOutcomes()];
}
/// <summary>
/// Initializes a new instance of the <see cref="NameFinderME"/> using the given <see cref="TokenNameFinderModel"/>.
/// </summary>
/// <param name="model">The model.</param>
/// <exception cref="System.ArgumentNullException">model</exception>
public NameFinderME(TokenNameFinderModel model)
{
if (model == null)
{
throw new ArgumentNullException(nameof(model));
}
sequenceCodec = model.Factory.CreateSequenceCodec();
sequenceValidator = sequenceCodec.CreateSequenceValidator();
this.model = model.NameFinderSequenceModel;
contextGenerator = model.Factory.CreateContextGenerator();
// TODO: We should deprecate this. And come up with a better solution!
additionalContextFeatureGenerator = new AdditionalContextFeatureGenerator();
contextGenerator.AddFeatureGenerator(
new WindowFeatureGenerator(additionalContextFeatureGenerator, 8, 8));
}
/// <summary>
/// Finds the n most probable sequences.
/// </summary>
/// <param name="numSequences">The number sequences.</param>
/// <param name="sequence">The sequence.</param>
/// <param name="additionalContext">The additional context.</param>
/// <param name="minSequenceScore">The minimum sequence score.</param>
/// <param name="beamSearch">The beam search.</param>
/// <param name="validator">The validator.</param>
public Sequence[] BestSequences(int numSequences, T[] sequence, object[] additionalContext,
double minSequenceScore,
IBeamSearchContextGenerator <T> beamSearch, ISequenceValidator <T> validator)
{
IHeap <Sequence> prev = new ListHeap <Sequence>(size);
IHeap <Sequence> next = new ListHeap <Sequence>(size);
prev.Add(new Sequence());
if (additionalContext == null)
{
additionalContext = new object[] {}; // EMPTY_ADDITIONAL_CONTEXT
}
for (var i = 0; i < sequence.Length; i++)
{
var sz = Math.Min(size, prev.Size());
for (var sc = 0; prev.Size() > 0 && sc < sz; sc++)
{
var top = prev.Extract();
var tmpOutcomes = top.Outcomes;
var outcomes = tmpOutcomes.ToArray();
var contexts = beamSearch.GetContext(i, sequence, outcomes, additionalContext);
double[] scores;
if (contextsCache != null)
{
scores = (double[])contextsCache.Get(contexts);
if (scores == null)
{
scores = model.Eval(contexts, probs);
contextsCache.Put(contexts, scores);
}
}
else
{
scores = model.Eval(contexts, probs);
}
var tempScores = new double[scores.Length];
for (var c = 0; c < scores.Length; c++)
{
tempScores[c] = scores[c];
}
Array.Sort(tempScores);
var min = tempScores[Math.Max(0, scores.Length - size)];
for (var p = 0; p < scores.Length; p++)
{
if (scores[p] < min)
{
continue; //only advance first "size" outcomes
}
var outcome = model.GetOutcome(p);
if (validator.ValidSequence(i, sequence, outcomes, outcome))
{
var ns = new Sequence(top, outcome, scores[p]);
if (ns.Score > minSequenceScore)
{
next.Add(ns);
}
}
}
if (next.Size() == 0)
{
//if no advanced sequences, advance all valid
for (var p = 0; p < scores.Length; p++)
{
var outcome = model.GetOutcome(p);
if (validator.ValidSequence(i, sequence, outcomes, outcome))
{
var ns = new Sequence(top, outcome, scores[p]);
if (ns.Score > minSequenceScore)
{
next.Add(ns);
}
}
}
}
}
// make prev = next; and re-init next (we reuse existing prev set once we clear it)
prev.Clear();
var tmp = prev;
prev = next;
next = tmp;
}
var numSeq = Math.Min(numSequences, prev.Size());
var topSequences = new Sequence[numSeq];
for (var seqIndex = 0; seqIndex < numSeq; seqIndex++)
{
topSequences[seqIndex] = prev.Extract();
}
return(topSequences);
}
/// <summary>
/// Finds the n most probable sequences.
/// </summary>
/// <param name="numSequences">The number sequences.</param>
/// <param name="sequence">The sequence.</param>
/// <param name="additionalContext">The additional context.</param>
/// <param name="beamSearch">The beam search.</param>
/// <param name="validator">The validator.</param>
/// <returns>The n most probable sequences.</returns>
public Sequence[] BestSequences(int numSequences, T[] sequence, object[] additionalContext,
IBeamSearchContextGenerator <T> beamSearch, ISequenceValidator <T> validator)
{
return(BestSequences(numSequences, sequence, additionalContext, zeroLog, beamSearch, validator));
}
/// <summary>
/// Initializes a new instance of the <see cref="NameFinderME"/> using the given <see cref="TokenNameFinderModel"/>.
/// </summary>
/// <param name="model">The model.</param>
/// <exception cref="System.ArgumentNullException">model</exception>
public NameFinderME(TokenNameFinderModel model) {
if (model == null)
throw new ArgumentNullException("model");
sequenceCodec = model.Factory.CreateSequenceCodec();
sequenceValidator = sequenceCodec.CreateSequenceValidator();
this.model = model.NameFinderSequenceModel;
contextGenerator = model.Factory.CreateContextGenerator();
// TODO: We should deprecate this. And come up with a better solution!
additionalContextFeatureGenerator = new AdditionalContextFeatureGenerator();
contextGenerator.AddFeatureGenerator(
new WindowFeatureGenerator(additionalContextFeatureGenerator, 8, 8));
}
/// <summary>
/// Initializes a new instance of the <see cref="ChunkerME"/> with the specified <see cref="ChunkerModel"/>.
/// </summary>
/// <param name="model">The chunker model.</param>
public ChunkerME(ChunkerModel model) {
contextGenerator = model.Factory.GetContextGenerator();
sequenceValidator = model.Factory.GetSequenceValidator();
this.model = model.ChunkerSequenceModel ?? new BeamSearch(model.BeamSize, model.MaxentModel);
}
