private static LogisticRegression<bool> GetLogisticRegression(double[,] samples,
bool[] knownOutput)
{
var clf = new LogisticRegression<bool>();
clf.Fit(DenseMatrix.OfArray(samples), knownOutput);
return clf;
}
private static void TestLogisticRegression(FeatureVector training, FeatureVector test)
{
LogisticRegression lr = new LogisticRegression(0.1, 3000);
LogisticRegressionModel lrModel = (LogisticRegressionModel)lr.Fit(training);
FeatureVector predictions = lrModel.transform(test);
PrintPredictionsAndEvaluate(predictions);
}
public void TestWriteParameters()
{
var clf = new LogisticRegression <int>(random: new Random(0));
clf.Fit(X, Y1);
clf.Coef.MapInplace(v => 0);
clf.Intercept.MapInplace(v => 0);
Assert.IsTrue(clf.DecisionFunction(X).Column(0).AlmostEquals(new DenseVector(new double[] { 0, 0, 0 })));
}
public void TestLiblinearRandomState()
{
var classification = SampleGenerator.MakeClassification(nSamples: 20);
var lr1 = new LogisticRegression <int>(random: new Random(0));
lr1.Fit(classification.X, classification.Y);
var lr2 = new LogisticRegression <int>(random: new Random(0));
lr2.Fit(classification.X, classification.Y);
Assert.IsTrue(lr1.Coef.AlmostEquals(lr2.Coef));
}
public void TestEstimatorLogisticRegression()
{
(IEstimator <ITransformer> pipe, IDataView dataView) = GetBinaryClassificationPipeline();
var trainer = new LogisticRegression(Env, "Label", "Features");
var pipeWithTrainer = pipe.Append(trainer);
TestEstimatorCore(pipeWithTrainer, dataView);
var transformedDataView = pipe.Fit(dataView).Transform(dataView);
var model = trainer.Fit(transformedDataView);
trainer.Train(transformedDataView, model.Model);
Done();
}
/// <summary>
/// Check that the model is able to fit the classification data
/// </summary>
private void check_predictions(LogisticRegression<int> clf, Matrix<double> X, int[] y)
{
int nSamples = y.Length;
int[] classes = y.Distinct().OrderBy(v => v).ToArray();
int nClasses = classes.Length;
clf.Fit(X, y);
var predicted = clf.Predict(X);
Assert.IsTrue(classes.SequenceEqual(clf.Classes));
Assert.AreEqual(nSamples, predicted.Length);
Assert.IsTrue(y.SequenceEqual(predicted));
Matrix<double> probabilities = clf.PredictProba(X);
Assert.AreEqual(Tuple.Create(nSamples, nClasses), probabilities.Shape());
Assert.IsTrue(probabilities.SumOfEveryRow().AlmostEquals(DenseVector.Create(probabilities.RowCount, i => 1.0)));
Assert.IsTrue(y.SequenceEqual(probabilities.ArgmaxColumns()));
}
private async Task DrawRoc( )
{
CanInteract = false;
Status = "Training classifier...";
await Task.Factory.StartNew(() => logistic.Fit(train,
trainAns.Select(x => isAimClass(x))
.ToArray( )));
var series = new LineSeries( )
{
Smooth = false, LineStyle = LineStyle.Solid
};
Status = "Iterating over thresholds...";
for (double tr = -10; tr <= 10; tr += 0.005)
{
logistic.Threshold = tr;
var ans = await Task.Factory.StartNew(() => logistic.Classify(test));
series.Points.Add(new DataPoint(CalcFPR(ans), CalcTPR(ans)));
}
var newModel = new PlotModel
{
Title = $"ROC for class {CurrentClass}",
};
newModel.Series.Add(series);
newModel.Series.Add(new FunctionSeries(x => x, 0, 1, 0.01));
Model = newModel;
NotifyOfPropertyChange(() => Model);
var plot = series.Points.Distinct( ).OrderBy(x => x.X * 10 + x.Y).ToList( );
Status = $"Area under curve: {CalcArea( plot )}";
CanInteract = true;
}
/// <summary>
/// Check that the model is able to fit the classification data
/// </summary>
private void check_predictions(LogisticRegression <int> clf, Matrix <double> X, int[] y)
{
int nSamples = y.Length;
int[] classes = y.Distinct().OrderBy(v => v).ToArray();
int nClasses = classes.Length;
clf.Fit(X, y);
var predicted = clf.Predict(X);
Assert.IsTrue(classes.SequenceEqual(clf.Classes));
Assert.AreEqual(nSamples, predicted.Length);
Assert.IsTrue(y.SequenceEqual(predicted));
Matrix <double> probabilities = clf.PredictProba(X);
Assert.AreEqual(Tuple.Create(nSamples, nClasses), probabilities.Shape());
Assert.IsTrue(probabilities.SumOfEveryRow().AlmostEquals(DenseVector.Create(probabilities.RowCount, i => 1.0)));
Assert.IsTrue(y.SequenceEqual(probabilities.ArgmaxColumns()));
}
public void TestSparsify()
{
string[] target = iris.Target.Select(t => iris.TargetNames[t]).ToArray();
var clf = new LogisticRegression <string>(random: new Random(0));
clf.Fit(iris.Data, target);
Matrix <double> predDD = clf.DecisionFunction(iris.Data);
clf.Sparsify();
Assert.IsInstanceOfType(clf.Coef, typeof(SparseMatrix));
Matrix <double> predSD = clf.DecisionFunction(iris.Data);
Matrix spData = SparseMatrix.OfMatrix(iris.Data);
Matrix <double> predSS = clf.DecisionFunction(spData);
clf.Densify();
Matrix <double> predDS = clf.DecisionFunction(spData);
Assert.IsTrue(predDD.AlmostEquals(predSD));
Assert.IsTrue(predDD.AlmostEquals(predSS));
Assert.IsTrue(predDD.AlmostEquals(predDS));
}
public void TestPredictIris()
{
int nSamples = iris.Data.RowCount;
string[] target = iris.Target.Select(v => iris.TargetNames[v]).ToArray();
var clf = new LogisticRegression <string>(c: iris.Data.RowCount);
clf.Fit(iris.Data, target);
Assert.IsTrue(target.Distinct().OrderBy(t => t).SequenceEqual(clf.Classes));
var pred = clf.Predict(iris.Data);
var matchingN = pred.Zip(target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0 * matchingN / pred.Length > 0.95);
var probabilities = clf.PredictProba(iris.Data);
Assert.IsTrue(probabilities.SumOfEveryRow().AlmostEquals(DenseVector.Create(nSamples, i => 1.0)));
pred = probabilities.RowEnumerator().Select(r => iris.TargetNames[r.Item2.MaximumIndex()]).ToArray();
matchingN = pred.Zip(target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0 * matchingN / pred.Length > 0.95);
}
public void TestPredictIris()
{
int nSamples = iris.Data.RowCount;
string[] target = iris.Target.Select(v => iris.TargetNames[v]).ToArray();
var clf = new LogisticRegression<string>(c: iris.Data.RowCount);
clf.Fit(iris.Data, target);
Assert.IsTrue(target.Distinct().OrderBy(t => t).SequenceEqual(clf.Classes));
var pred = clf.Predict(iris.Data);
var matchingN = pred.Zip(target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0*matchingN/pred.Length > 0.95);
var probabilities = clf.PredictProba(iris.Data);
Assert.IsTrue(probabilities.SumOfEveryRow().AlmostEquals(DenseVector.Create(nSamples, i => 1.0)));
pred = probabilities.RowEnumerator().Select(r => iris.TargetNames[r.Item2.MaximumIndex()]).ToArray();
matchingN = pred.Zip(target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0 * matchingN / pred.Length > 0.95);
}
public void TestLiblinearRandomState()
{
var classification = SampleGenerator.MakeClassification(nSamples: 20);
var lr1 = new LogisticRegression<int>(random: new Random(0));
lr1.Fit(classification.X, classification.Y);
var lr2 = new LogisticRegression<int>(random: new Random(0));
lr2.Fit(classification.X, classification.Y);
Assert.IsTrue(lr1.Coef.AlmostEquals(lr2.Coef));
}
public void TestWriteParameters()
{
var clf = new LogisticRegression<int>(random: new Random(0));
clf.Fit(X, Y1);
clf.Coef.MapInplace(v => 0);
clf.Intercept.MapInplace(v => 0);
Assert.IsTrue(clf.DecisionFunction(X).Column(0).AlmostEquals(new DenseVector(new double[]{0, 0, 0})));
}
public void TestSparsify()
{
string[] target = iris.Target.Select(t => iris.TargetNames[t]).ToArray();
var clf = new LogisticRegression<string>(random: new Random(0));
clf.Fit(iris.Data, target);
Matrix<double> predDD = clf.DecisionFunction(iris.Data);
clf.Sparsify();
Assert.IsInstanceOfType(clf.Coef, typeof(SparseMatrix));
Matrix<double> predSD = clf.DecisionFunction(iris.Data);
Matrix spData = SparseMatrix.OfMatrix(iris.Data);
Matrix<double> predSS = clf.DecisionFunction(spData);
clf.Densify();
Matrix<double> predDS = clf.DecisionFunction(spData);
Assert.IsTrue(predDD.AlmostEquals(predSD));
Assert.IsTrue(predDD.AlmostEquals(predSS));
Assert.IsTrue(predDD.AlmostEquals(predDS));
}
private void buttonForDataSplitNext_Click(object sender, EventArgs e)
{
trainingSetPercentage = (double)numericUpDownForTrainingSetPercent.Value / 100.0;
numFolds = (int)numericUpDownForNumFolds.Value;
double[] smaOut = null;
double[] wmaOut = null;
double[] emaOut = null;
double[] macdOut = null;
double[] stochasticsOut = null;
double[] williamsROut = null;
double[] rsiOut = null;
double[] closesOut = null;
var data = IndicatorService.GetData(code, targetDate, new string[] { "Tarih", "Kapanis" }, numberOfData + 1);
if (isSMAChecked)
{
smaOut = IndicatorDataPreprocessor.GetSMAOut(MovingAverage.Simple(code, targetDate, smaPeriod, numberOfData));
}
if (isWMAChecked)
{
wmaOut = IndicatorDataPreprocessor.GetWMAOut(MovingAverage.Weighted(code, targetDate, wmaPeriod, numberOfData));
}
if (isEMAChecked)
{
emaOut = IndicatorDataPreprocessor.GetEMAOut(MovingAverage.Exponential(code, targetDate, emaPeriod, numberOfData));
}
if (isMACDChecked)
{
macdOut = IndicatorDataPreprocessor.GetMACDOut(new MovingAverageConvergenceDivergence(code, targetDate, firstPeriod, secondPeriod, triggerPeriod, numberOfData));
}
if (isStochasticsChecked)
{
stochasticsOut = IndicatorDataPreprocessor.GetStochasticsOut(new Stochastics(code, targetDate, fastKPeriod, fastDPeriod, slowDPeriod, numberOfData));
}
if (isWilliamsRChecked)
{
williamsROut = IndicatorDataPreprocessor.GetWilliamsROut(WilliamsR.Wsr(code, targetDate, williamsRPeriod, numberOfData));
}
if (isRSIChecked)
{
rsiOut = IndicatorDataPreprocessor.GetRSIOut(RelativeStrengthIndex.Rsi(code, targetDate, rsiPeriod, numberOfData));
}
closesOut = IndicatorDataPreprocessor.GetClosesOut(numberOfData, data);
int minRowCount = 1000000;
if (smaOut != null)
{
minRowCount = smaOut.Length;
}
if (wmaOut != null)
{
minRowCount = minRowCount < wmaOut.Length ? minRowCount : wmaOut.Length;
}
if (emaOut != null)
{
minRowCount = minRowCount < emaOut.Length ? minRowCount : emaOut.Length;
}
if (macdOut != null)
{
minRowCount = minRowCount < macdOut.Length ? minRowCount : macdOut.Length;
}
if (rsiOut != null)
{
minRowCount = minRowCount < rsiOut.Length ? minRowCount : rsiOut.Length;
}
if (williamsROut != null)
{
minRowCount = minRowCount < williamsROut.Length ? minRowCount : williamsROut.Length;
}
if (stochasticsOut != null)
{
minRowCount = minRowCount < stochasticsOut.Length ? minRowCount : stochasticsOut.Length;
}
if (closesOut != null)
{
minRowCount = minRowCount < closesOut.Length ? minRowCount : closesOut.Length;
}
var fv = new FeatureVector();
if (isSMAChecked)
{
fv.AddColumn("SMA", smaOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (isWMAChecked)
{
fv.AddColumn("WMA", wmaOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (isEMAChecked)
{
fv.AddColumn("EMA", emaOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (isMACDChecked)
{
fv.AddColumn("MACD", macdOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (isRSIChecked)
{
fv.AddColumn("RSI", rsiOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (isWilliamsRChecked)
{
fv.AddColumn("WilliamsR", williamsROut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (isStochasticsChecked)
{
fv.AddColumn("Stochastics", stochasticsOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
fv.AddColumn("label", closesOut.Select(p => (object)string.Format("{0:0.0}", p).ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
var training = new FeatureVector();
var test = new FeatureVector();
int count = fv.Values[0].Length;
for (int i = 0; i < fv.ColumnName.Count; i++)
{
training.AddColumn(fv.ColumnName[i], fv.Values[i].Take((int)(count * trainingSetPercentage)).ToArray());
}
for (int i = 0; i < fv.ColumnName.Count; i++)
{
test.AddColumn(fv.ColumnName[i], fv.Values[i].Skip((int)(count * trainingSetPercentage)).Take(count).ToArray()); // Take(count) means take the rest of all elements, number of the rest of the elements is smaller than count.
}
if (numFolds > 0)
{
BinaryClassificationEvaluator bce1 = new BinaryClassificationEvaluator();
LinearRegression linearRegression = new LinearRegression();
CrossValidator cvLinReg = new CrossValidator(linearRegression, bce1, numFolds);
CrossValidatorModel cvLinRegModel = (CrossValidatorModel)cvLinReg.Fit(training);
FeatureVector linRegPredictions = cvLinRegModel.transform(test);
bce1.evaluate(linRegPredictions);
linRegAcc = bce1.Accuracy;
BinaryClassificationEvaluator bce2 = new BinaryClassificationEvaluator();
LogisticRegression logisticRegression = new LogisticRegression();
CrossValidator cvLogReg = new CrossValidator(logisticRegression, bce2, numFolds);
CrossValidatorModel cvLogRegModel = (CrossValidatorModel)cvLogReg.Fit(training);
FeatureVector logRegPredictions = cvLogRegModel.transform(test);
bce2.evaluate(logRegPredictions);
logRegAcc = bce2.Accuracy;
BinaryClassificationEvaluator bce3 = new BinaryClassificationEvaluator();
NaiveBayes naiveBayes = new NaiveBayes();
CrossValidator cvNaiBay = new CrossValidator(naiveBayes, bce3, numFolds);
CrossValidatorModel cvNaiBayModel = (CrossValidatorModel)cvNaiBay.Fit(training);
FeatureVector naiBayPredictions = cvNaiBayModel.transform(test);
bce3.evaluate(naiBayPredictions);
naiBayAcc = bce3.Accuracy;
}
else
{
BinaryClassificationEvaluator bce1 = new BinaryClassificationEvaluator();
LinearRegression linearRegression = new LinearRegression();
LinearRegressionModel linearRegressionModel = (LinearRegressionModel)linearRegression.Fit(training);
FeatureVector linRegPredictions = linearRegressionModel.transform(test);
bce1.evaluate(linRegPredictions);
linRegAcc = bce1.Accuracy;
BinaryClassificationEvaluator bce2 = new BinaryClassificationEvaluator();
LogisticRegression logicticRegression = new LogisticRegression();
LogisticRegressionModel logisticRegressionModel = (LogisticRegressionModel)logicticRegression.Fit(training);
FeatureVector logRegPredictions = logisticRegressionModel.transform(test);
bce2.evaluate(logRegPredictions);
logRegAcc = bce2.Accuracy;
BinaryClassificationEvaluator bce3 = new BinaryClassificationEvaluator();
NaiveBayes naiveBayes = new NaiveBayes();
NaiveBayesModel naiveBayesModel = (NaiveBayesModel)naiveBayes.Fit(training);
FeatureVector naiBayPredictions = naiveBayesModel.transform(test);
bce3.evaluate(naiBayPredictions);
naiBayAcc = bce3.Accuracy;
}
labelForLinRegAcc.Text = linRegAcc.ToString();
labelForLogRegAcc.Text = logRegAcc.ToString();
labelForNaiBayAcc.Text = naiBayAcc.ToString();
panelForResults.BringToFront();
}
private static double CalculateAccuracy(List <int> indicators, int mlAlgorithm, bool isCrossValidationEnabled, int minRowCount, double trainingSetPercentage, double[] smaOut, double[] wmaOut, double[] emaOut, double[] macdOut, double[] rsiOut, double[] williamsROut, double[] stochasticsOut, double[] closesOut)
{
FeatureVector vector = new FeatureVector();
if (indicators.Contains(IndicatorService.SMA))
{
vector.AddColumn("SMA", smaOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (indicators.Contains(IndicatorService.WMA))
{
vector.AddColumn("WMA", wmaOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (indicators.Contains(IndicatorService.EMA))
{
vector.AddColumn("EMA", emaOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (indicators.Contains(IndicatorService.MACD))
{
vector.AddColumn("MACD", macdOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (indicators.Contains(IndicatorService.RSI))
{
vector.AddColumn("RSI", rsiOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (indicators.Contains(IndicatorService.WilliamsR))
{
vector.AddColumn("WilliamsR", williamsROut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
if (indicators.Contains(IndicatorService.Stochastics))
{
vector.AddColumn("Stochastics", stochasticsOut.Select(p => (object)p.ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
}
vector.AddColumn("label", closesOut.Select(p => (object)string.Format("{0:0.0}", p).ToString(CultureInfo.InvariantCulture)).Take(minRowCount).ToArray());
new CSVExporter(vector).Export("c:\\users\\yasin\\indicatorOutput.csv");
int count = vector.Values[0].Length;
FeatureVector training = new FeatureVector();
for (int i = 0; i < vector.ColumnName.Count; i++)
{
training.AddColumn(vector.ColumnName[i], vector.Values[i].Take((int)(count * trainingSetPercentage)).ToArray());
}
FeatureVector test = new FeatureVector();
for (int i = 0; i < vector.ColumnName.Count; i++)
{
test.AddColumn(vector.ColumnName[i], vector.Values[i].Skip((int)(count * trainingSetPercentage)).Take(count).ToArray());
}
double accuracy = 0;
if (mlAlgorithm == MLAService.LIN_REG)
{
var linReg = new LinearRegression();
var bce = new BinaryClassificationEvaluator();
if (isCrossValidationEnabled)
{
var cv = new CrossValidator(linReg, bce, 10);
var cvModel = (CrossValidatorModel)cv.Fit(training);
var predictions = cvModel.transform(test);
bce.evaluate(predictions);
accuracy = bce.Accuracy;
}
else
{
var linRegModel = (LinearRegressionModel)linReg.Fit(training);
var predictions = linRegModel.transform(test);
bce.evaluate(predictions);
accuracy = bce.Accuracy;
}
}
else if (mlAlgorithm == MLAService.LOG_REG)
{
var logReg = new LogisticRegression();
var bce = new BinaryClassificationEvaluator();
if (isCrossValidationEnabled)
{
var cv = new CrossValidator(logReg, bce, 10);
var cvModel = (CrossValidatorModel)cv.Fit(training);
var predictions = cvModel.transform(test);
bce.evaluate(predictions);
accuracy = bce.Accuracy;
}
else
{
var logRegModel = (LogisticRegressionModel)logReg.Fit(training);
var predictions = logRegModel.transform(test);
bce.evaluate(predictions);
accuracy = bce.Accuracy;
}
}
else if (mlAlgorithm == MLAService.NAI_BAY)
{
var naiBay = new NaiveBayes();
var bce = new BinaryClassificationEvaluator();
if (isCrossValidationEnabled)
{
var cv = new CrossValidator(naiBay, bce, 10);
var cvModel = (CrossValidatorModel)cv.Fit(training);
var predictions = cvModel.transform(test);
bce.evaluate(predictions);
accuracy = bce.Accuracy;
}
else
{
var naiBayModel = (NaiveBayesModel)naiBay.Fit(training);
var predictions = naiBayModel.transform(test);
bce.evaluate(predictions);
accuracy = bce.Accuracy;
}
}
return(accuracy);
}
