private MutableContext[] FindParameters()
{
var allOutcomesPattern = new int[numOutcomes];
for (var oi = 0; oi < numOutcomes; oi++)
{
allOutcomesPattern[oi] = oi;
}
/* Stores the estimated parameter value of each predicate during iteration. */
var parameters = new MutableContext[numPreds];
for (var pi = 0; pi < numPreds; pi++)
{
parameters[pi] = new MutableContext(allOutcomesPattern, new double[numOutcomes]);
for (var aoi = 0; aoi < numOutcomes; aoi++)
{
parameters[pi].SetParameter(aoi, 0.0);
}
}
// ReSharper disable once CoVariantArrayConversion
var evalParams = new EvalParameters(parameters, numOutcomes);
const double stepSize = 1;
for (var ei = 0; ei < numUniqueEvents; ei++)
{
var targetOutcome = outcomeList[ei];
for (var ni = 0; ni < numTimesEventsSeen[ei]; ni++)
{
for (var ci = 0; ci < contexts[ei].Length; ci++)
{
var pi = contexts[ei][ci];
if (values == null)
{
parameters[pi].UpdateParameter(targetOutcome, stepSize);
}
else
{
parameters[pi].UpdateParameter(targetOutcome, stepSize * values[ei][ci]);
}
}
}
}
// Output the final training stats.
TrainingStats(evalParams);
return(parameters);
}
public virtual AbstractModel trainModel(int iterations, SequenceStream sequenceStream, int cutoff,
bool useAverage)
{
this.iterations = iterations;
this.sequenceStream = sequenceStream;
DataIndexer di = new OnePassDataIndexer(new SequenceStreamEventStream(sequenceStream), cutoff, false);
numSequences = 0;
foreach (Sequence <Event> s in sequenceStream)
{
numSequences++;
}
outcomeList = di.OutcomeList;
predLabels = di.PredLabels;
pmap = new IndexHashTable <string>(predLabels, 0.7d);
display("Incorporating indexed data for training... \n");
this.useAverage = useAverage;
numEvents = di.NumEvents;
this.iterations = iterations;
outcomeLabels = di.OutcomeLabels;
omap = new Dictionary <string, int?>();
for (int oli = 0; oli < outcomeLabels.Length; oli++)
{
omap[outcomeLabels[oli]] = oli;
}
outcomeList = di.OutcomeList;
numPreds = predLabels.Length;
numOutcomes = outcomeLabels.Length;
if (useAverage)
{
updates = RectangularArrays.ReturnRectangularIntArray(numPreds, numOutcomes, 3);
}
display("done.\n");
display("\tNumber of Event Tokens: " + numEvents + "\n");
display("\t Number of Outcomes: " + numOutcomes + "\n");
display("\t Number of Predicates: " + numPreds + "\n");
parameters = new MutableContext[numPreds];
if (useAverage)
{
averageParams = new MutableContext[numPreds];
}
allOutcomesPattern = new int[numOutcomes];
for (int oi = 0; oi < numOutcomes; oi++)
{
allOutcomesPattern[oi] = oi;
}
for (int pi = 0; pi < numPreds; pi++)
{
parameters[pi] = new MutableContext(allOutcomesPattern, new double[numOutcomes]);
if (useAverage)
{
averageParams[pi] = new MutableContext(allOutcomesPattern, new double[numOutcomes]);
}
for (int aoi = 0; aoi < numOutcomes; aoi++)
{
parameters[pi].setParameter(aoi, 0.0);
if (useAverage)
{
averageParams[pi].setParameter(aoi, 0.0);
}
}
}
modelDistribution = new double[numOutcomes];
display("Computing model parameters...\n");
findParameters(iterations);
display("...done.\n");
/// <summary>
///************* Create and return the model ***************** </summary>
string[] updatedPredLabels = predLabels;
/*
* String[] updatedPredLabels = new String[pmap.size()];
* for (String pred : pmap.keySet()) {
* updatedPredLabels[pmap.get(pred)]=pred;
* }
*/
if (useAverage)
{
return(new PerceptronModel(averageParams, updatedPredLabels, outcomeLabels));
}
else
{
return(new PerceptronModel(parameters, updatedPredLabels, outcomeLabels));
}
}
/// <summary>
/// Train a model using the GIS algorithm.
/// </summary>
/// <param name="iterations">The number of GIS iterations to perform.</param>
/// <param name="di">The data indexer used to compress events in memory.</param>
/// <param name="modelPrior">The prior distribution used to train this model.</param>
/// <param name="modelCutoff">The number of times a feature must occur to be included.</param>
/// <param name="threads">The number of threads used to train this model.</param>
/// <returns>The newly trained model, which can be used immediately or saved to disk using an <see cref="GISModelWriter"/> object.</returns>
public GISModel TrainModel(int iterations, IDataIndexer di, IPrior modelPrior, int modelCutoff, int threads) {
if (threads <= 0)
throw new ArgumentOutOfRangeException("threads", threads, @"Threads must be at least one or greater.");
modelExpects = new MutableContext[threads][];
info.Append("Trained using GIS algorithm.\n\n");
// Executes the data indexer
di.Execute();
// Incorporate all of the needed info.
Display("Incorporating indexed data for training...");
contexts = di.GetContexts();
values = di.Values;
cutoff = modelCutoff;
predicateCounts = di.GetPredCounts();
numTimesEventsSeen = di.GetNumTimesEventsSeen();
numUniqueEvents = contexts.Length;
prior = modelPrior;
// determine the correction constant and its inverse
double correctionConstant = 0;
for (int ci = 0; ci < contexts.Length; ci++) {
if (values == null || values[ci] == null) {
if (contexts[ci].Length > correctionConstant) {
correctionConstant = contexts[ci].Length;
}
} else {
var cl = values[ci][0];
for (var vi = 1; vi < values[ci].Length; vi++) {
cl += values[ci][vi];
}
if (cl > correctionConstant) {
correctionConstant = cl;
}
}
}
Display("done.");
outcomeLabels = di.GetOutcomeLabels();
outcomeList = di.GetOutcomeList();
numOutcomes = outcomeLabels.Length;
predLabels = di.GetPredLabels();
prior.SetLabels(outcomeLabels, predLabels);
numPreds = predLabels.Length;
info.Append("Number of Event Tokens: {0}\n", numUniqueEvents);
info.Append(" Number of Outcomes: {0}\n", numOutcomes);
info.Append(" Number of Predicates: {0}\n", numPreds);
Display("\tNumber of Event Tokens: " + numUniqueEvents);
Display("\t Number of Outcomes: " + numOutcomes);
Display("\t Number of Predicates: " + numPreds);
// set up feature arrays
//var predCount = new float[numPreds][numOutcomes];
var predCount = new float[numPreds][];
for (int ti = 0; ti < numUniqueEvents; ti++) {
for (int j = 0; j < contexts[ti].Length; j++) {
if (predCount[contexts[ti][j]] == null) {
predCount[contexts[ti][j]] = new float[numOutcomes];
}
if (values != null && values[ti] != null) {
predCount[contexts[ti][j]][outcomeList[ti]] += numTimesEventsSeen[ti]*values[ti][j];
} else {
predCount[contexts[ti][j]][outcomeList[ti]] += numTimesEventsSeen[ti];
}
}
}
// ReSharper disable once RedundantAssignment
di = null;
// Get the observed expectations of the features. Strictly speaking,
// we should divide the counts by the number of Tokens, but because of
// the way the model's expectations are approximated in the
// implementation, this is canceled out when we compute the next
// iteration of a parameter, making the extra divisions wasteful.
param = new MutableContext[numPreds];
for (var i = 0; i < modelExpects.Length; i++)
modelExpects[i] = new MutableContext[numPreds];
observedExpects = new MutableContext[numPreds];
// The model does need the correction constant and the correction feature. The correction constant
// is only needed during training, and the correction feature is not necessary.
// For compatibility reasons the model contains form now on a correction constant of 1,
// and a correction param 0.
// ReSharper disable once CoVariantArrayConversion
evalParams = new EvalParameters(param, 0, 1, numOutcomes);
var activeOutcomes = new int[numOutcomes];
var allOutcomesPattern = new int[numOutcomes];
for (var oi = 0; oi < numOutcomes; oi++) {
allOutcomesPattern[oi] = oi;
}
for (var pi = 0; pi < numPreds; pi++) {
var numActiveOutcomes = 0;
int[] outcomePattern;
if (Smoothing) {
numActiveOutcomes = numOutcomes;
outcomePattern = allOutcomesPattern;
} else {
//determine active outcomes
for (var oi = 0; oi < numOutcomes; oi++) {
if (predCount[pi][oi] > 0 && predicateCounts[pi] >= cutoff) {
activeOutcomes[numActiveOutcomes] = oi;
numActiveOutcomes++;
}
}
if (numActiveOutcomes == numOutcomes) {
outcomePattern = allOutcomesPattern;
} else {
outcomePattern = new int[numActiveOutcomes];
for (var aoi = 0; aoi < numActiveOutcomes; aoi++) {
outcomePattern[aoi] = activeOutcomes[aoi];
}
}
}
param[pi] = new MutableContext(outcomePattern, new double[numActiveOutcomes]);
foreach (MutableContext[] me in modelExpects)
me[pi] = new MutableContext(outcomePattern, new double[numActiveOutcomes]);
observedExpects[pi] = new MutableContext(outcomePattern, new double[numActiveOutcomes]);
for (var aoi = 0; aoi < numActiveOutcomes; aoi++) {
var oi = outcomePattern[aoi];
param[pi].SetParameter(aoi, 0.0);
foreach (var modelExpect in modelExpects) {
modelExpect[pi].SetParameter(aoi, 0.0);
}
if (predCount[pi][oi] > 0) {
observedExpects[pi].SetParameter(aoi, predCount[pi][oi]);
} else if (Smoothing) {
observedExpects[pi].SetParameter(aoi, SmoothingObservation);
}
}
}
Display("...done.");
/***************** Find the parameters ************************/
if (threads == 1)
Display("Computing model parameters ...");
else
Display("Computing model parameters in " + threads + " threads...");
FindParameters(iterations, correctionConstant);
/*************** Create and return the model ******************/
// To be compatible with old models the correction constant is always 1
// ReSharper disable once CoVariantArrayConversion
return new GISModel(param, predLabels, outcomeLabels, 1, evalParams.CorrectionParam) {
info = TrainingInfo
};
}
public AbstractModel TrainModel(int trainIterations, ISequenceStream trainStream, int cutoff,
bool trainUseAverage)
{
iterations = trainIterations;
useAverage = trainUseAverage;
sequenceStream = trainStream;
info.Append("Trained using Perceptron Sequence algorithm.\n\n");
var di = new OnePassDataIndexer(new SequenceStreamEventStream(trainStream), cutoff, false);
trainStream.Reset();
numSequences = 0;
while (trainStream.Read() != null)
{
numSequences++;
}
outcomeList = di.GetOutcomeList();
predLabels = di.GetPredLabels();
pMap = new IndexHashTable <string>(predLabels, 0.7d);
// Incorporation indexed data for training...
numEvents = di.GetNumEvents();
outcomeLabels = di.GetOutcomeLabels();
oMap = new Dictionary <string, int>();
for (var i = 0; i < outcomeLabels.Length; i++)
{
oMap.Add(outcomeLabels[i], i);
}
outcomeList = di.GetOutcomeList();
numPreds = predLabels.Length;
numOutcomes = outcomeLabels.Length;
if (trainUseAverage)
{
updates = new int[numPreds][][];
for (var i = 0; i < numPreds; i++)
{
updates[i] = new int[numOutcomes][];
for (var j = 0; j < numOutcomes; j++)
{
updates[i][j] = new int[3];
}
}
}
// done.
Display("done.\n");
info.Append("Number of Event Tokens: {0}\n" +
" Number of Outcomes: {1}\n" +
" Number of Predicates: {2}\n", numEvents, numOutcomes, numPreds);
Display("\tNumber of Event Tokens: " + numEvents);
Display("\t Number of Outcomes: " + numOutcomes);
Display("\t Number of Predicates: " + numPreds);
param = new MutableContext[numPreds];
if (trainUseAverage)
{
averageParams = new MutableContext[numPreds];
}
allOutcomesPattern = new int[numOutcomes];
for (var i = 0; i < numOutcomes; i++)
{
allOutcomesPattern[i] = i;
}
for (var pi = 0; pi < numPreds; pi++)
{
param[pi] = new MutableContext(allOutcomesPattern, new double[numOutcomes]);
if (trainUseAverage)
{
averageParams[pi] = new MutableContext(allOutcomesPattern, new double[numOutcomes]);
}
for (var aoi = 0; aoi < numOutcomes; aoi++)
{
param[pi].SetParameter(aoi, 0.0d);
if (trainUseAverage)
{
averageParams[pi].SetParameter(aoi, 0.0d);
}
}
}
Display("Computing model parameters...");
FindParameters();
Display("...done.");
/*************** Create and return the model ******************/
// ReSharper disable CoVariantArrayConversion
if (trainUseAverage)
{
return new PerceptronModel(averageParams, predLabels, outcomeLabels)
{
info = info
}
}
;
return(new PerceptronModel(param, predLabels, outcomeLabels)
{
info = info
});
// ReSharper restore CoVariantArrayConversion
}
/// <summary>
/// Train a model using the GIS algorithm.
/// </summary>
/// <param name="iterations">The number of GIS iterations to perform.</param>
/// <param name="di">The data indexer used to compress events in memory.</param>
/// <param name="modelPrior">The prior distribution used to train this model.</param>
/// <param name="modelCutoff">The number of times a feature must occur to be included.</param>
/// <param name="threads">The number of threads used to train this model.</param>
/// <returns>The newly trained model, which can be used immediately or saved to disk using an <see cref="GISModelWriter"/> object.</returns>
public GISModel TrainModel(int iterations, IDataIndexer di, IPrior modelPrior, int modelCutoff, int threads)
{
if (threads <= 0)
{
throw new ArgumentOutOfRangeException("threads", threads, @"Threads must be at least one or greater.");
}
modelExpects = new MutableContext[threads][];
info.Append("Trained using GIS algorithm.\n\n");
// Executes the data indexer
di.Execute();
// Incorporate all of the needed info.
Display("Incorporating indexed data for training...");
contexts = di.GetContexts();
values = di.Values;
cutoff = modelCutoff;
predicateCounts = di.GetPredCounts();
numTimesEventsSeen = di.GetNumTimesEventsSeen();
numUniqueEvents = contexts.Length;
prior = modelPrior;
// determine the correction constant and its inverse
double correctionConstant = 0;
for (int ci = 0; ci < contexts.Length; ci++)
{
if (values == null || values[ci] == null)
{
if (contexts[ci].Length > correctionConstant)
{
correctionConstant = contexts[ci].Length;
}
}
else
{
var cl = values[ci][0];
for (var vi = 1; vi < values[ci].Length; vi++)
{
cl += values[ci][vi];
}
if (cl > correctionConstant)
{
correctionConstant = cl;
}
}
}
Display("done.");
outcomeLabels = di.GetOutcomeLabels();
outcomeList = di.GetOutcomeList();
numOutcomes = outcomeLabels.Length;
predLabels = di.GetPredLabels();
prior.SetLabels(outcomeLabels, predLabels);
numPreds = predLabels.Length;
info.Append("Number of Event Tokens: {0}\n", numUniqueEvents);
info.Append(" Number of Outcomes: {0}\n", numOutcomes);
info.Append(" Number of Predicates: {0}\n", numPreds);
Display("\tNumber of Event Tokens: " + numUniqueEvents);
Display("\t Number of Outcomes: " + numOutcomes);
Display("\t Number of Predicates: " + numPreds);
// set up feature arrays
//var predCount = new float[numPreds][numOutcomes];
var predCount = new float[numPreds][];
for (int ti = 0; ti < numUniqueEvents; ti++)
{
for (int j = 0; j < contexts[ti].Length; j++)
{
if (predCount[contexts[ti][j]] == null)
{
predCount[contexts[ti][j]] = new float[numOutcomes];
}
if (values != null && values[ti] != null)
{
predCount[contexts[ti][j]][outcomeList[ti]] += numTimesEventsSeen[ti] * values[ti][j];
}
else
{
predCount[contexts[ti][j]][outcomeList[ti]] += numTimesEventsSeen[ti];
}
}
}
// ReSharper disable once RedundantAssignment
di = null;
// Get the observed expectations of the features. Strictly speaking,
// we should divide the counts by the number of Tokens, but because of
// the way the model's expectations are approximated in the
// implementation, this is canceled out when we compute the next
// iteration of a parameter, making the extra divisions wasteful.
param = new MutableContext[numPreds];
for (var i = 0; i < modelExpects.Length; i++)
{
modelExpects[i] = new MutableContext[numPreds];
}
observedExpects = new MutableContext[numPreds];
// The model does need the correction constant and the correction feature. The correction constant
// is only needed during training, and the correction feature is not necessary.
// For compatibility reasons the model contains form now on a correction constant of 1,
// and a correction param 0.
// ReSharper disable once CoVariantArrayConversion
evalParams = new EvalParameters(param, 0, 1, numOutcomes);
var activeOutcomes = new int[numOutcomes];
var allOutcomesPattern = new int[numOutcomes];
for (var oi = 0; oi < numOutcomes; oi++)
{
allOutcomesPattern[oi] = oi;
}
for (var pi = 0; pi < numPreds; pi++)
{
var numActiveOutcomes = 0;
int[] outcomePattern;
if (Smoothing)
{
numActiveOutcomes = numOutcomes;
outcomePattern = allOutcomesPattern;
}
else
{
//determine active outcomes
for (var oi = 0; oi < numOutcomes; oi++)
{
if (predCount[pi][oi] > 0 && predicateCounts[pi] >= cutoff)
{
activeOutcomes[numActiveOutcomes] = oi;
numActiveOutcomes++;
}
}
if (numActiveOutcomes == numOutcomes)
{
outcomePattern = allOutcomesPattern;
}
else
{
outcomePattern = new int[numActiveOutcomes];
for (var aoi = 0; aoi < numActiveOutcomes; aoi++)
{
outcomePattern[aoi] = activeOutcomes[aoi];
}
}
}
param[pi] = new MutableContext(outcomePattern, new double[numActiveOutcomes]);
foreach (MutableContext[] me in modelExpects)
{
me[pi] = new MutableContext(outcomePattern, new double[numActiveOutcomes]);
}
observedExpects[pi] = new MutableContext(outcomePattern, new double[numActiveOutcomes]);
for (var aoi = 0; aoi < numActiveOutcomes; aoi++)
{
var oi = outcomePattern[aoi];
param[pi].SetParameter(aoi, 0.0);
foreach (var modelExpect in modelExpects)
{
modelExpect[pi].SetParameter(aoi, 0.0);
}
if (predCount[pi][oi] > 0)
{
observedExpects[pi].SetParameter(aoi, predCount[pi][oi]);
}
else if (Smoothing)
{
observedExpects[pi].SetParameter(aoi, SmoothingObservation);
}
}
}
Display("...done.");
/***************** Find the parameters ************************/
if (threads == 1)
{
Display("Computing model parameters ...");
}
else
{
Display("Computing model parameters in " + threads + " threads...");
}
FindParameters(iterations, correctionConstant);
/*************** Create and return the model ******************/
// To be compatible with old models the correction constant is always 1
// ReSharper disable once CoVariantArrayConversion
return(new GISModel(param, predLabels, outcomeLabels, 1, evalParams.CorrectionParam)
{
info = TrainingInfo
});
}
private MutableContext[] FindParameters(int iterations, bool useAverage)
{
info.Append(" Number of Iterations: {0}\n", iterations);
Display("\nPerforming " + iterations + " iterations.\n");
var allOutcomesPattern = new int[numOutcomes];
for (var oi = 0; oi < numOutcomes; oi++)
{
allOutcomesPattern[oi] = oi;
}
/* Stores the estimated parameter value of each predicate during iteration. */
var param = new MutableContext[numPreds];
for (var pi = 0; pi < numPreds; pi++)
{
param[pi] = new MutableContext(allOutcomesPattern, new double[numOutcomes]);
for (var aoi = 0; aoi < numOutcomes; aoi++)
{
param[pi].SetParameter(aoi, 0.0);
}
}
// ReSharper disable once CoVariantArrayConversion
var evalParams = new EvalParameters(param, numOutcomes);
// Stores the sum of parameter values of each predicate over many iterations.
var summedParams = new MutableContext[numPreds];
if (useAverage)
{
for (var pi = 0; pi < numPreds; pi++)
{
summedParams[pi] = new MutableContext(allOutcomesPattern, new double[numOutcomes]);
for (var aoi = 0; aoi < numOutcomes; aoi++)
{
summedParams[pi].SetParameter(aoi, 0.0);
}
}
}
// Keep track of the previous three accuracies. The difference of
// the mean of these and the current training set accuracy is used
// with tolerance to decide whether to stop.
var prevAccuracy1 = 0.0;
var prevAccuracy2 = 0.0;
var prevAccuracy3 = 0.0;
// A counter for the denominator for averaging.
var numTimesSummed = 0;
double stepSize = 1;
for (var i = 1; i <= iterations; i++)
{
// Decrease the step size by a small amount.
if (stepSizeDecrease > 0)
{
stepSize *= 1 - stepSizeDecrease;
}
if (Monitor != null && Monitor.Token.CanBeCanceled)
{
Monitor.Token.ThrowIfCancellationRequested();
}
var numCorrect = 0;
for (var ei = 0; ei < numUniqueEvents; ei++)
{
var targetOutcome = outcomeList[ei];
for (var ni = 0; ni < numTimesEventsSeen[ei]; ni++)
{
// Compute the model's prediction according to the current parameters.
var modelDistribution = new double[numOutcomes];
PerceptronModel.Eval(
contexts[ei],
values != null ? values[ei] : null,
modelDistribution,
evalParams, false);
var maxOutcome = MaxIndex(modelDistribution);
// If the predicted outcome is different from the target
// outcome, do the standard update: boost the parameters
// associated with the target and reduce those associated
// with the incorrect predicted outcome.
if (maxOutcome != targetOutcome)
{
for (var ci = 0; ci < contexts[ei].Length; ci++)
{
var pi = contexts[ei][ci];
if (values == null)
{
param[pi].UpdateParameter(targetOutcome, stepSize);
param[pi].UpdateParameter(maxOutcome, -stepSize);
}
else
{
param[pi].UpdateParameter(targetOutcome, stepSize * values[ei][ci]);
param[pi].UpdateParameter(maxOutcome, -stepSize * values[ei][ci]);
}
}
}
// Update the counts for accuracy.
if (maxOutcome == targetOutcome)
{
numCorrect++;
}
}
}
// Calculate the training accuracy and display.
var trainingAccuracy = (double)numCorrect / numEvents;
Display($"{i,-4} {numCorrect} of {numEvents} - {trainingAccuracy}");
// TODO: Make averaging configurable !!!
bool doAveraging;
if (useAverage && UseSkippedAveraging && (i < 20 || IsPerfectSquare(i)))
{
doAveraging = true;
}
else if (useAverage)
{
doAveraging = true;
}
else
{
doAveraging = false;
}
if (doAveraging)
{
numTimesSummed++;
for (var pi = 0; pi < numPreds; pi++)
{
for (var aoi = 0; aoi < numOutcomes; aoi++)
{
summedParams[pi].UpdateParameter(aoi, param[pi].Parameters[aoi]);
}
}
}
// If the tolerance is greater than the difference between the
// current training accuracy and all of the previous three
// training accuracies, stop training.
if (Math.Abs(prevAccuracy1 - trainingAccuracy) < tolerance &&
Math.Abs(prevAccuracy2 - trainingAccuracy) < tolerance &&
Math.Abs(prevAccuracy3 - trainingAccuracy) < tolerance)
{
Display("Stopping: change in training set accuracy less than " + tolerance + "\n");
break;
}
// Update the previous training accuracies.
prevAccuracy1 = prevAccuracy2;
prevAccuracy2 = prevAccuracy3;
prevAccuracy3 = trainingAccuracy;
}
// Output the final training stats.
TrainingStats(evalParams);
if (!useAverage)
{
return(param);
}
if (numTimesSummed == 0) // Improbable but possible according to the Coverity.
{
numTimesSummed = 1;
}
// Create averaged parameters
for (var pi = 0; pi < numPreds; pi++)
{
for (var aoi = 0; aoi < numOutcomes; aoi++)
{
summedParams[pi].SetParameter(aoi, summedParams[pi].Parameters[aoi] / numTimesSummed);
}
}
return(summedParams);
}
private MutableContext[] findParameters(int iterations, bool useAverage)
{
display("Performing " + iterations + " iterations.\n");
int[] allOutcomesPattern = new int[numOutcomes];
for (int oi = 0; oi < numOutcomes; oi++)
{
allOutcomesPattern[oi] = oi;
}
/// <summary>
/// Stores the estimated parameter value of each predicate during iteration. </summary>
MutableContext[] parameters = new MutableContext[numPreds];
for (int pi = 0; pi < numPreds; pi++)
{
parameters[pi] = new MutableContext(allOutcomesPattern, new double[numOutcomes]);
for (int aoi = 0; aoi < numOutcomes; aoi++)
{
parameters[pi].setParameter(aoi, 0.0);
}
}
EvalParameters evalParams = new EvalParameters(parameters, numOutcomes);
/// <summary>
/// Stores the sum of parameter values of each predicate over many iterations. </summary>
MutableContext[] summedParams = new MutableContext[numPreds];
if (useAverage)
{
for (int pi = 0; pi < numPreds; pi++)
{
summedParams[pi] = new MutableContext(allOutcomesPattern, new double[numOutcomes]);
for (int aoi = 0; aoi < numOutcomes; aoi++)
{
summedParams[pi].setParameter(aoi, 0.0);
}
}
}
// Keep track of the previous three accuracies. The difference of
// the mean of these and the current training set accuracy is used
// with tolerance to decide whether to stop.
double prevAccuracy1 = 0.0;
double prevAccuracy2 = 0.0;
double prevAccuracy3 = 0.0;
// A counter for the denominator for averaging.
int numTimesSummed = 0;
double stepsize = 1;
for (int i = 1; i <= iterations; i++)
{
// Decrease the stepsize by a small amount.
if (stepSizeDecrease != null)
{
stepsize *= 1 - stepSizeDecrease.GetValueOrDefault();
}
displayIteration(i);
int numCorrect = 0;
for (int ei = 0; ei < numUniqueEvents; ei++)
{
int targetOutcome = outcomeList[ei];
for (int ni = 0; ni < this.numTimesEventsSeen[ei]; ni++)
{
// Compute the model's prediction according to the current parameters.
double[] modelDistribution = new double[numOutcomes];
if (values != null)
{
PerceptronModel.eval(contexts[ei], values[ei], modelDistribution, evalParams, false);
}
else
{
PerceptronModel.eval(contexts[ei], null, modelDistribution, evalParams, false);
}
int maxOutcome = maxIndex(modelDistribution);
// If the predicted outcome is different from the target
// outcome, do the standard update: boost the parameters
// associated with the target and reduce those associated
// with the incorrect predicted outcome.
if (maxOutcome != targetOutcome)
{
for (int ci = 0; ci < contexts[ei].Length; ci++)
{
int pi = contexts[ei][ci];
if (values == null)
{
parameters[pi].updateParameter(targetOutcome, stepsize);
parameters[pi].updateParameter(maxOutcome, -stepsize);
}
else
{
parameters[pi].updateParameter(targetOutcome, stepsize * values[ei][ci]);
parameters[pi].updateParameter(maxOutcome, -stepsize * values[ei][ci]);
}
}
}
// Update the counts for accuracy.
if (maxOutcome == targetOutcome)
{
numCorrect++;
}
}
}
// Calculate the training accuracy and display.
double trainingAccuracy = (double)numCorrect / numEvents;
if (i < 10 || (i % 10) == 0)
{
display(". (" + numCorrect + "/" + numEvents + ") " + trainingAccuracy + "\n");
}
// TODO: Make averaging configurable !!!
bool doAveraging;
if (useAverage && useSkippedlAveraging && (i < 20 || isPerfectSquare(i)))
{
doAveraging = true;
}
else if (useAverage)
{
doAveraging = true;
}
else
{
doAveraging = false;
}
if (doAveraging)
{
numTimesSummed++;
for (int pi = 0; pi < numPreds; pi++)
{
for (int aoi = 0; aoi < numOutcomes; aoi++)
{
summedParams[pi].updateParameter(aoi, parameters[pi].Parameters[aoi]);
}
}
}
// If the tolerance is greater than the difference between the
// current training accuracy and all of the previous three
// training accuracies, stop training.
if (Math.Abs(prevAccuracy1 - trainingAccuracy) < tolerance &&
Math.Abs(prevAccuracy2 - trainingAccuracy) < tolerance &&
Math.Abs(prevAccuracy3 - trainingAccuracy) < tolerance)
{
display("Stopping: change in training set accuracy less than " + tolerance + "\n");
break;
}
// Update the previous training accuracies.
prevAccuracy1 = prevAccuracy2;
prevAccuracy2 = prevAccuracy3;
prevAccuracy3 = trainingAccuracy;
}
// Output the final training stats.
trainingStats(evalParams);
// Create averaged parameters
if (useAverage)
{
for (int pi = 0; pi < numPreds; pi++)
{
for (int aoi = 0; aoi < numOutcomes; aoi++)
{
summedParams[pi].setParameter(aoi, summedParams[pi].Parameters[aoi] / numTimesSummed);
}
}
return(summedParams);
}
else
{
return(parameters);
}
}
/// <summary>
/// Train a model using the GIS algorithm.
/// </summary>
/// <param name="iterations"> The number of GIS iterations to perform. </param>
/// <param name="di"> The data indexer used to compress events in memory. </param>
/// <param name="modelPrior"> The prior distribution used to train this model. </param>
/// <returns> The newly trained model, which can be used immediately or saved
/// to disk using an opennlp.maxent.io.GISModelWriter object. </returns>
public virtual GISModel trainModel(int iterations, DataIndexer di, Prior modelPrior, int cutoff, int threads)
{
if (threads <= 0)
{
throw new System.ArgumentException("threads must be at least one or greater but is " + threads + "!");
}
modelExpects = new MutableContext[threads][];
/// <summary>
///************ Incorporate all of the needed info ***************** </summary>
display("Incorporating indexed data for training... \n");
contexts = di.Contexts;
values = di.Values;
this.cutoff = cutoff;
predicateCounts = di.PredCounts;
numTimesEventsSeen = di.NumTimesEventsSeen;
numUniqueEvents = contexts.Length;
this.prior = modelPrior;
//printTable(contexts);
// determine the correction constant and its inverse
double correctionConstant = 0;
for (int ci = 0; ci < contexts.Length; ci++)
{
if (values == null || values[ci] == null)
{
if (contexts[ci].Length > correctionConstant)
{
correctionConstant = contexts[ci].Length;
}
}
else
{
float cl = values[ci][0];
for (int vi = 1; vi < values[ci].Length; vi++)
{
cl += values[ci][vi];
}
if (cl > correctionConstant)
{
correctionConstant = cl;
}
}
}
display("done.\n");
outcomeLabels = di.OutcomeLabels;
outcomeList = di.OutcomeList;
numOutcomes = outcomeLabels.Length;
predLabels = di.PredLabels;
prior.setLabels(outcomeLabels, predLabels);
numPreds = predLabels.Length;
display("\tNumber of Event Tokens: " + numUniqueEvents + "\n");
display("\t Number of Outcomes: " + numOutcomes + "\n");
display("\t Number of Predicates: " + numPreds + "\n");
// set up feature arrays
float[][] predCount = RectangularArrays.ReturnRectangularFloatArray(numPreds, numOutcomes);
for (int ti = 0; ti < numUniqueEvents; ti++)
{
for (int j = 0; j < contexts[ti].Length; j++)
{
if (values != null && values[ti] != null)
{
predCount[contexts[ti][j]][outcomeList[ti]] += numTimesEventsSeen[ti] * values[ti][j];
}
else
{
predCount[contexts[ti][j]][outcomeList[ti]] += numTimesEventsSeen[ti];
}
}
}
//printTable(predCount);
di = null; // don't need it anymore
// A fake "observation" to cover features which are not detected in
// the data. The default is to assume that we observed "1/10th" of a
// feature during training.
double smoothingObservation = _smoothingObservation;
// Get the observed expectations of the features. Strictly speaking,
// we should divide the counts by the number of Tokens, but because of
// the way the model's expectations are approximated in the
// implementation, this is cancelled out when we compute the next
// iteration of a parameter, making the extra divisions wasteful.
parameters = new MutableContext[numPreds];
for (int i = 0; i < modelExpects.Length; i++)
{
modelExpects[i] = new MutableContext[numPreds];
}
observedExpects = new MutableContext[numPreds];
// The model does need the correction constant and the correction feature. The correction constant
// is only needed during training, and the correction feature is not necessary.
// For compatibility reasons the model contains form now on a correction constant of 1,
// and a correction param 0.
evalParams = new EvalParameters(parameters, 0, 1, numOutcomes);
int[] activeOutcomes = new int[numOutcomes];
int[] outcomePattern;
int[] allOutcomesPattern = new int[numOutcomes];
for (int oi = 0; oi < numOutcomes; oi++)
{
allOutcomesPattern[oi] = oi;
}
int numActiveOutcomes = 0;
for (int pi = 0; pi < numPreds; pi++)
{
numActiveOutcomes = 0;
if (useSimpleSmoothing)
{
numActiveOutcomes = numOutcomes;
outcomePattern = allOutcomesPattern;
}
else //determine active outcomes
{
for (int oi = 0; oi < numOutcomes; oi++)
{
if (predCount[pi][oi] > 0 && predicateCounts[pi] >= cutoff)
{
activeOutcomes[numActiveOutcomes] = oi;
numActiveOutcomes++;
}
}
if (numActiveOutcomes == numOutcomes)
{
outcomePattern = allOutcomesPattern;
}
else
{
outcomePattern = new int[numActiveOutcomes];
for (int aoi = 0; aoi < numActiveOutcomes; aoi++)
{
outcomePattern[aoi] = activeOutcomes[aoi];
}
}
}
parameters[pi] = new MutableContext(outcomePattern, new double[numActiveOutcomes]);
for (int i = 0; i < modelExpects.Length; i++)
{
modelExpects[i][pi] = new MutableContext(outcomePattern, new double[numActiveOutcomes]);
}
observedExpects[pi] = new MutableContext(outcomePattern, new double[numActiveOutcomes]);
for (int aoi = 0; aoi < numActiveOutcomes; aoi++)
{
int oi = outcomePattern[aoi];
parameters[pi].setParameter(aoi, 0.0);
foreach (MutableContext[] modelExpect in modelExpects)
{
modelExpect[pi].setParameter(aoi, 0.0);
}
if (predCount[pi][oi] > 0)
{
observedExpects[pi].setParameter(aoi, predCount[pi][oi]);
}
else if (useSimpleSmoothing)
{
observedExpects[pi].setParameter(aoi, smoothingObservation);
}
}
}
predCount = null; // don't need it anymore
display("...done.\n");
/// <summary>
///*************** Find the parameters *********************** </summary>
if (threads == 1)
{
display("Computing model parameters ...\n");
}
else
{
display("Computing model parameters in " + threads + " threads...\n");
}
findParameters(iterations, correctionConstant);
/// <summary>
///************* Create and return the model ***************** </summary>
// To be compatible with old models the correction constant is always 1
return(new GISModel(parameters, predLabels, outcomeLabels, 1, evalParams.CorrectionParam));
}
