public virtual void TrainWeightedData(GeneralDataset <L, F> data, float[] dataWeights)
{
//Use LogisticClassifierFactory to train instead.
if (data.labelIndex.Size() != 2)
{
throw new Exception("LogisticClassifier is only for binary classification!");
}
IMinimizer <IDiffFunction> minim;
LogisticObjectiveFunction lof = null;
if (data is Dataset <object, object> )
{
lof = new LogisticObjectiveFunction(data.NumFeatureTypes(), data.GetDataArray(), data.GetLabelsArray(), prior, dataWeights);
}
else
{
if (data is RVFDataset <object, object> )
{
lof = new LogisticObjectiveFunction(data.NumFeatureTypes(), data.GetDataArray(), data.GetValuesArray(), data.GetLabelsArray(), prior, dataWeights);
}
}
minim = new QNMinimizer(lof);
weights = minim.Minimize(lof, 1e-4, new double[data.NumFeatureTypes()]);
featureIndex = data.featureIndex;
classes[0] = data.labelIndex.Get(0);
classes[1] = data.labelIndex.Get(1);
}
private NaiveBayesClassifierFactory.NBWeights TrainWeightsUCL(int[][] data, int[] labels, int numFeatures, int numClasses)
{
int[] numValues = NumberValues(data, numFeatures);
int[] sumValues = new int[numFeatures];
//how many feature-values are before this feature
for (int j = 1; j < numFeatures; j++)
{
sumValues[j] = sumValues[j - 1] + numValues[j - 1];
}
int[][] newdata = new int[data.Length][];
for (int i = 0; i < data.Length; i++)
{
newdata[i][0] = 0;
for (int j_1 = 0; j_1 < numFeatures; j_1++)
{
newdata[i][j_1 + 1] = sumValues[j_1] + data[i][j_1] + 1;
}
}
int totalFeatures = sumValues[numFeatures - 1] + numValues[numFeatures - 1] + 1;
logger.Info("total feats " + totalFeatures);
LogConditionalObjectiveFunction <L, F> objective = new LogConditionalObjectiveFunction <L, F>(totalFeatures, numClasses, newdata, labels, prior, sigma, 0.0);
IMinimizer <IDiffFunction> min = new QNMinimizer();
double[] argmin = min.Minimize(objective, 1e-4, objective.Initial());
double[][] wts = objective.To2D(argmin);
System.Console.Out.WriteLine("weights have dimension " + wts.Length);
return(new NaiveBayesClassifierFactory.NBWeights(wts, numValues));
}
public virtual LogisticClassifier <L, F> TrainWeightedData(GeneralDataset <L, F> data, float[] dataWeights)
{
if (data is RVFDataset)
{
((RVFDataset <L, F>)data).EnsureRealValues();
}
if (data.labelIndex.Size() != 2)
{
throw new Exception("LogisticClassifier is only for binary classification!");
}
IMinimizer <IDiffFunction> minim;
LogisticObjectiveFunction lof = null;
if (data is Dataset <object, object> )
{
lof = new LogisticObjectiveFunction(data.NumFeatureTypes(), data.GetDataArray(), data.GetLabelsArray(), new LogPrior(LogPrior.LogPriorType.Quadratic), dataWeights);
}
else
{
if (data is RVFDataset <object, object> )
{
lof = new LogisticObjectiveFunction(data.NumFeatureTypes(), data.GetDataArray(), data.GetValuesArray(), data.GetLabelsArray(), new LogPrior(LogPrior.LogPriorType.Quadratic), dataWeights);
}
}
minim = new QNMinimizer(lof);
weights = minim.Minimize(lof, 1e-4, new double[data.NumFeatureTypes()]);
featureIndex = data.featureIndex;
classes[0] = data.labelIndex.Get(0);
classes[1] = data.labelIndex.Get(1);
return(new LogisticClassifier <L, F>(weights, featureIndex, classes));
}
private NaiveBayesClassifierFactory.NBWeights TrainWeightsCL(int[][] data, int[] labels, int numFeatures, int numClasses)
{
LogConditionalEqConstraintFunction objective = new LogConditionalEqConstraintFunction(numFeatures, numClasses, data, labels, prior, sigma, 0.0);
IMinimizer <IDiffFunction> min = new QNMinimizer();
double[] argmin = min.Minimize(objective, 1e-4, objective.Initial());
double[][][] wts = objective.To3D(argmin);
double[] priors = objective.Priors(argmin);
return(new NaiveBayesClassifierFactory.NBWeights(priors, wts));
}
/// <summary>Solves the problem using a quasi-newton method (L-BFGS).</summary>
/// <remarks>
/// Solves the problem using a quasi-newton method (L-BFGS). The solution
/// is stored in the
/// <c>lambda</c>
/// array of
/// <c>prob</c>
/// .
/// </remarks>
public virtual void SolveQN()
{
CGRunner.LikelihoodFunction df = new CGRunner.LikelihoodFunction(prob, tol, useGaussianPrior, priorSigmaS, sigmaSquareds);
CGRunner.MonitorFunction monitor = new CGRunner.MonitorFunction(prob, df, filename);
IMinimizer <IDiffFunction> cgm = new QNMinimizer(monitor, 10);
// all parameters are started at 0.0
prob.lambda = cgm.Minimize(df, tol, new double[df.DomainDimension()]);
PrintOptimizationResults(df, monitor);
}
public virtual LogisticClassifier <L, F> TrainClassifier(GeneralDataset <L, F> data, double l1reg, double tol, LogPrior prior, bool biased)
{
if (data is RVFDataset)
{
((RVFDataset <L, F>)data).EnsureRealValues();
}
if (data.labelIndex.Size() != 2)
{
throw new Exception("LogisticClassifier is only for binary classification!");
}
IMinimizer <IDiffFunction> minim;
if (!biased)
{
LogisticObjectiveFunction lof = null;
if (data is Dataset <object, object> )
{
lof = new LogisticObjectiveFunction(data.NumFeatureTypes(), data.GetDataArray(), data.GetLabelsArray(), prior);
}
else
{
if (data is RVFDataset <object, object> )
{
lof = new LogisticObjectiveFunction(data.NumFeatureTypes(), data.GetDataArray(), data.GetValuesArray(), data.GetLabelsArray(), prior);
}
}
if (l1reg > 0.0)
{
minim = ReflectionLoading.LoadByReflection("edu.stanford.nlp.optimization.OWLQNMinimizer", l1reg);
}
else
{
minim = new QNMinimizer(lof);
}
weights = minim.Minimize(lof, tol, new double[data.NumFeatureTypes()]);
}
else
{
BiasedLogisticObjectiveFunction lof = new BiasedLogisticObjectiveFunction(data.NumFeatureTypes(), data.GetDataArray(), data.GetLabelsArray(), prior);
if (l1reg > 0.0)
{
minim = ReflectionLoading.LoadByReflection("edu.stanford.nlp.optimization.OWLQNMinimizer", l1reg);
}
else
{
minim = new QNMinimizer(lof);
}
weights = minim.Minimize(lof, tol, new double[data.NumFeatureTypes()]);
}
featureIndex = data.featureIndex;
classes[0] = data.labelIndex.Get(0);
classes[1] = data.labelIndex.Get(1);
return(new LogisticClassifier <L, F>(weights, featureIndex, classes));
}
public virtual void TestQNMinimizerRosenbrock()
{
double[] initial = new double[] { 0.0, 0.0 };
IDiffFunction rf = new MinimizerTest.RosenbrockFunction();
QNMinimizer qn = new QNMinimizer();
double[] answer = qn.Minimize(rf, 1e-10, initial);
System.Console.Error.WriteLine("Answer is: " + Arrays.ToString(answer));
NUnit.Framework.Assert.AreEqual(1.0, answer[0], 1e-8);
NUnit.Framework.Assert.AreEqual(1.0, answer[1], 1e-8);
}
public void TestRosenbrockFunction()
{
var minimizer = new QNMinimizer();
var f = new RosenbrockFunction();
var x = minimizer.Minimize(f);
var minValue = f.ValueAt(x);
Assert.AreEqual(x[0], 1.0, 1e-5);
Assert.AreEqual(x[1], 1.0, 1e-5);
Assert.AreEqual(minValue, 0, 1e-10);
}
public void TestQuadraticFunction()
{
var minimizer = new QNMinimizer();
var f = new QuadraticFunction();
var x = minimizer.Minimize(f);
var minValue = f.ValueAt(x);
Assert.AreEqual(x[0], 1.0, 0.000001);
Assert.AreEqual(x[1], 5.0, 0.000001);
Assert.AreEqual(minValue, 10.0, 0.000001);
}
protected internal override double[] TrainWeights(int[][][][] data, int[][] labels, IEvaluator[] evaluators, int pruneFeatureItr, double[][][][] featureVals)
{
CRFLogConditionalObjectiveFloatFunction func = new CRFLogConditionalObjectiveFloatFunction(data, labels, windowSize, classIndex, labelIndices, map, flags.backgroundSymbol, flags.sigma);
cliquePotentialFunctionHelper = func;
QNMinimizer minimizer;
if (flags.interimOutputFreq != 0)
{
IFloatFunction monitor = new ResultStoringFloatMonitor(flags.interimOutputFreq, flags.serializeTo);
minimizer = new QNMinimizer(monitor);
}
else
{
minimizer = new QNMinimizer();
}
if (pruneFeatureItr == 0)
{
minimizer.SetM(flags.QNsize);
}
else
{
minimizer.SetM(flags.QNsize2);
}
float[] initialWeights;
if (flags.initialWeights == null)
{
initialWeights = func.Initial();
}
else
{
try
{
log.Info("Reading initial weights from file " + flags.initialWeights);
using (DataInputStream dis = new DataInputStream(new BufferedInputStream(new GZIPInputStream(new FileInputStream(flags.initialWeights)))))
{
initialWeights = ConvertByteArray.ReadFloatArr(dis);
}
}
catch (IOException)
{
throw new Exception("Could not read from float initial weight file " + flags.initialWeights);
}
}
log.Info("numWeights: " + initialWeights.Length);
float[] weightsArray = minimizer.Minimize(func, (float)flags.tolerance, initialWeights);
return(ArrayMath.FloatArrayToDoubleArray(weightsArray));
}
public virtual void FinishTraining()
{
IntCounter <string> tagCounter = new IntCounter <string>();
WeightedDataset data = new WeightedDataset(datumCounter.Size());
foreach (TaggedWord word in datumCounter.KeySet())
{
int count = datumCounter.GetIntCount(word);
if (trainOnLowCount && count > trainCountThreshold)
{
continue;
}
if (functionWordTags.Contains(word.Word()))
{
continue;
}
tagCounter.IncrementCount(word.Tag());
if (trainByType)
{
count = 1;
}
data.Add(new BasicDatum(featExtractor.MakeFeatures(word.Word()), word.Tag()), count);
}
datumCounter = null;
tagDist = Distribution.LaplaceSmoothedDistribution(tagCounter, tagCounter.Size(), 0.5);
tagCounter = null;
ApplyThresholds(data);
Verbose("Making classifier...");
QNMinimizer minim = new QNMinimizer();
//new ResultStoringMonitor(5, "weights"));
// minim.shutUp();
LinearClassifierFactory factory = new LinearClassifierFactory(minim);
factory.SetTol(tol);
factory.SetSigma(sigma);
scorer = factory.TrainClassifier(data);
Verbose("Done training.");
}
public virtual double[] Minimize(IDiffFunction function, double functionTolerance, double[] initial, int maxIterations)
{
Sayln("SGDToQNMinimizer called on function of " + function.DomainDimension() + " variables;");
// check for stochastic derivatives
if (!(function is AbstractStochasticCachingDiffFunction))
{
throw new NotSupportedException();
}
AbstractStochasticCachingDiffFunction dfunction = (AbstractStochasticCachingDiffFunction)function;
dfunction.method = StochasticCalculateMethods.GradientOnly;
ScaledSGDMinimizer sgd = new ScaledSGDMinimizer(this.gain, this.bSize, this.SGDPasses, 1, this.outputIterationsToFile);
QNMinimizer qn = new QNMinimizer(this.QNMem, true);
double[] x = sgd.Minimize(dfunction, functionTolerance, initial, this.SGDPasses);
QNMinimizer.QNInfo qnInfo = new QNMinimizer.QNInfo(this, sgd.sList, sgd.yList);
qnInfo.d = sgd.diag;
qn.Minimize(dfunction, functionTolerance, x, this.QNPasses, qnInfo);
log.Info(string.Empty);
log.Info("Minimization complete.");
log.Info(string.Empty);
log.Info("Exiting for Debug");
return(x);
}
private double[][] TrainWeights()
{
QNMinimizer minimizer = new QNMinimizer(15, true);
minimizer.UseOWLQN(true, lambda);
IDiffFunction objective = new ShiftParamsLogisticObjectiveFunction(data, dataValues, ConvertLabels(labels), numClasses, numFeatures + data.Length, numFeatures, prior);
double[] augmentedThetas = new double[(numClasses - 1) * (numFeatures + data.Length)];
augmentedThetas = minimizer.Minimize(objective, 1e-4, augmentedThetas);
// calculate number of non-zero parameters, for debugging
int count = 0;
for (int j = numFeatures; j < augmentedThetas.Length; j++)
{
if (augmentedThetas[j] != 0)
{
count++;
}
}
Redwood.Log("NUM NONZERO PARAMETERS: " + count);
double[][] thetas = new double[][] { };
LogisticUtils.Unflatten(augmentedThetas, thetas);
return(thetas);
}
public virtual void ExecuteOneTrainingBatch(IList <Tree> trainingBatch, IdentityHashMap <Tree, byte[]> compressedParses, double[] sumGradSquare)
{
Timing convertTiming = new Timing();
convertTiming.Doing("Converting trees");
IdentityHashMap <Tree, IList <Tree> > topParses = CacheParseHypotheses.ConvertToTrees(trainingBatch, compressedParses, op.trainOptions.trainingThreads);
convertTiming.Done();
DVParserCostAndGradient gcFunc = new DVParserCostAndGradient(trainingBatch, topParses, dvModel, op);
double[] theta = dvModel.ParamsToVector();
switch (Minimizer)
{
case (1):
{
//maxFuncIter = 10;
// 1: QNMinimizer, 2: SGD
QNMinimizer qn = new QNMinimizer(op.trainOptions.qnEstimates, true);
qn.UseMinPackSearch();
qn.UseDiagonalScaling();
qn.TerminateOnAverageImprovement(true);
qn.TerminateOnNumericalZero(true);
qn.TerminateOnRelativeNorm(true);
theta = qn.Minimize(gcFunc, op.trainOptions.qnTolerance, theta, op.trainOptions.qnIterationsPerBatch);
break;
}
case 2:
{
//Minimizer smd = new SGDMinimizer(); double tol = 1e-4; theta = smd.minimize(gcFunc,tol,theta,op.trainOptions.qnIterationsPerBatch);
double lastCost = 0;
double currCost = 0;
bool firstTime = true;
for (int i = 0; i < op.trainOptions.qnIterationsPerBatch; i++)
{
//gcFunc.calculate(theta);
double[] grad = gcFunc.DerivativeAt(theta);
currCost = gcFunc.ValueAt(theta);
log.Info("batch cost: " + currCost);
// if(!firstTime){
// if(currCost > lastCost){
// System.out.println("HOW IS FUNCTION VALUE INCREASING????!!! ... still updating theta");
// }
// if(Math.abs(currCost - lastCost) < 0.0001){
// System.out.println("function value is not decreasing. stop");
// }
// }else{
// firstTime = false;
// }
lastCost = currCost;
ArrayMath.AddMultInPlace(theta, grad, -1 * op.trainOptions.learningRate);
}
break;
}
case 3:
{
// AdaGrad
double eps = 1e-3;
double currCost = 0;
for (int i = 0; i < op.trainOptions.qnIterationsPerBatch; i++)
{
double[] gradf = gcFunc.DerivativeAt(theta);
currCost = gcFunc.ValueAt(theta);
log.Info("batch cost: " + currCost);
for (int feature = 0; feature < gradf.Length; feature++)
{
sumGradSquare[feature] = sumGradSquare[feature] + gradf[feature] * gradf[feature];
theta[feature] = theta[feature] - (op.trainOptions.learningRate * gradf[feature] / (System.Math.Sqrt(sumGradSquare[feature]) + eps));
}
}
break;
}
default:
{
throw new ArgumentException("Unsupported minimizer " + Minimizer);
}
}
dvModel.VectorToParams(theta);
}
