public void TestDecisionFunction()
{
// multi class:
var clf = new Svc <int>(kernel: Kernel.Linear, c: 0.1);
clf.Fit(iris.Data, iris.Target);
var dec = (iris.Data * clf.Coef.Transpose()).AddRowVector(clf.Intercept);
Assert.IsTrue(dec.AlmostEquals(clf.DecisionFunction(iris.Data)));
// binary:
clf.Fit(X, Y);
dec = (X * clf.Coef.Transpose()).AddRowVector(clf.Intercept);
int[] prediction = clf.Predict(X);
Assert.IsTrue(dec.AlmostEquals(clf.DecisionFunction(X)));
var b = clf.DecisionFunction(X).Column(0).Select(v => clf.Classes[v > 0 ? 1 : 0]);
Assert.IsTrue(prediction.SequenceEqual(b));
var expected = DenseMatrix.OfArray(new[, ] {
{ -1.0 }, { -0.66 }, { -1.0 }, { 0.66 }, { 1.0 }, { 1.0 }
});
Assert.IsTrue(clf.DecisionFunction(X).AlmostEquals(expected, 1E-2));
}
public void TestSvc()
{
var clf = new Svc <int>(kernel: Kernel.Linear, probability: true);
clf.Fit(X, Y);
var spClf = new Svc <int>(kernel: Kernel.Linear, probability: true);
spClf.Fit(SparseMatrix.OfMatrix(X), Y);
Assert.IsTrue(spClf.Predict(T).SequenceEqual(true_result));
Assert.IsTrue(spClf.SupportVectors is SparseMatrix);
Assert.IsTrue(clf.SupportVectors.AlmostEquals(spClf.SupportVectors));
Assert.IsTrue(spClf.DualCoef is SparseMatrix);
Assert.IsTrue(clf.DualCoef.AlmostEquals(spClf.DualCoef));
Assert.IsTrue(spClf.Coef is SparseMatrix);
Assert.IsTrue(clf.Coef.AlmostEquals(spClf.Coef));
Assert.IsTrue(clf.Support.SequenceEqual(spClf.Support));
Assert.IsTrue(clf.Predict(T).SequenceEqual(spClf.Predict(T)));
// refit with a different dataset
clf.Fit(X2, Y2);
spClf.Fit(SparseMatrix.OfMatrix(X2), Y2);
Assert.IsTrue(clf.SupportVectors.AlmostEquals(spClf.SupportVectors));
Assert.IsTrue(clf.DualCoef.AlmostEquals(spClf.DualCoef));
Assert.IsTrue(clf.Coef.AlmostEquals(spClf.Coef));
Assert.IsTrue(clf.Support.SequenceEqual(spClf.Support));
Assert.IsTrue(clf.Predict(T2).SequenceEqual(spClf.Predict(T2)));
Assert.IsTrue(clf.PredictProba(T2).AlmostEquals(spClf.PredictProba(T2), 0.001));
}
public void TestSparseRealdata()
{
var x = new SparseMatrix(80, 36);
x[7, 6] = 0.03771744;
x[39, 5] = 0.1003567;
x[77, 35] = 0.01174647;
x[77, 31] = 0.027069;
var y = new[]
{
1.0, 0.0, 2.0, 2.0, 1.0, 1.0, 1.0, 2.0, 2.0, 0.0, 1.0, 2.0, 2.0,
0.0, 2.0, 0.0, 3.0, 0.0, 3.0, 0.0, 1.0, 1.0, 3.0, 2.0, 3.0, 2.0,
0.0, 3.0, 1.0, 0.0, 2.0, 1.0, 2.0, 0.0, 1.0, 0.0, 2.0, 3.0, 1.0,
3.0, 0.0, 1.0, 0.0, 0.0, 2.0, 0.0, 1.0, 2.0, 2.0, 2.0, 3.0, 2.0,
0.0, 3.0, 2.0, 1.0, 2.0, 3.0, 2.0, 2.0, 0.0, 1.0, 0.0, 1.0, 2.0,
3.0, 0.0, 0.0, 2.0, 2.0, 1.0, 3.0, 1.0, 1.0, 0.0, 1.0, 2.0, 1.0,
1.0, 3.0
};
var clf = new Svc <double>(kernel: Kernel.Linear);
clf.Fit(DenseMatrix.OfMatrix(x), y);
var spClf = new Svc <double>(kernel: Kernel.Linear);
spClf.Fit(x, y);
Assert.IsTrue(clf.SupportVectors.AlmostEquals(spClf.SupportVectors));
Assert.IsTrue(clf.DualCoef.AlmostEquals(spClf.DualCoef));
}
/// <summary>
/// Make some classification predictions on a toy dataset using a SVC
///
/// If binary is True restrict to a binary classification problem instead of a
/// multiclass classification problem
/// </summary>
/// <param name="x"></param>
/// <param name="y"></param>
/// <param name="binary"></param>
private static Tuple <int[], int[], Matrix <double> > MakePrediction(
Matrix <double> x = null,
int[] y = null,
bool binary = false)
{
if (x == null && y == null)
{
// import some data to play with
var dataset = IrisDataset.Load();
x = dataset.Data;
y = dataset.Target;
}
if (binary)
{
// restrict to a binary classification task
x = x.RowsAt(y.Indices(v => v < 2));
y = y.Where(v => v < 2).ToArray();
}
int nSamples = x.RowCount;
int nFeatures = x.ColumnCount;
var rng = new Random(37);
int[] p = Shuffle(rng, Enumerable.Range(0, nSamples).ToArray());
x = x.RowsAt(p);
y = y.ElementsAt(p);
var half = nSamples / 2;
// add noisy features to make the problem harder and avoid perfect results
rng = new Random(0);
x = x.HStack(DenseMatrix.CreateRandom(nSamples, 200, new Normal {
RandomSource = rng
}));
// run classifier, get class probabilities and label predictions
var clf = new Svc <int>(kernel: Kernel.Linear, probability: true);
clf.Fit(x.SubMatrix(0, half, 0, x.ColumnCount), y.Take(half).ToArray());
Matrix <double> probasPred = clf.PredictProba(x.SubMatrix(half, x.RowCount - half, 0, x.ColumnCount));
if (binary)
{
// only interested in probabilities of the positive case
// XXX: do we really want a special API for the binary case?
probasPred = probasPred.SubMatrix(0, probasPred.RowCount, 1, 1);
}
var yPred = clf.Predict(x.SubMatrix(half, x.RowCount - half, 0, x.ColumnCount));
var yTrue = y.Skip(half).ToArray();
return(Tuple.Create(yTrue, yPred, probasPred));
}
public void TestSvcWithCustomKernel()
{
var clfLin = new Svc <int>(kernel: Kernel.Linear);
clfLin.Fit(SparseMatrix.OfMatrix(X), Y);
var clfMylin =
new Svc <int>(kernel: Kernel.FromFunction((x, y) => x * y.Transpose()));
clfMylin.Fit(SparseMatrix.OfMatrix(X), Y);
Assert.IsTrue(
clfLin.Predict(SparseMatrix.OfMatrix(X)).SequenceEqual(clfMylin.Predict(SparseMatrix.OfMatrix(X))));
}
public void TestLibsvmIris()
{
// shuffle the dataset so that labels are not ordered
foreach (var k in new[] { Kernel.Linear, Kernel.Rbf })
{
var clf = new Svc <int>(kernel: k);
clf.Fit(iris.Data, iris.Target);
var pred = clf.Predict(iris.Data);
var matchingN = pred.Zip(iris.Target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0 * matchingN / pred.Length > 0.9);
Assert.IsTrue(clf.Classes.SequenceEqual(clf.Classes.OrderBy(v => v)));
}
}
public void TestSingleSample_1D()
{
var clf = new Svc <int>();
clf.Fit(X, Y);
var p = clf.Predict(X.Row(0).ToRowMatrix());
Assert.AreEqual(Y[0], p[0]);
//todo:
//clf = svm.LinearSVC(random_state=0).fit(X, Y)
//clf.predict(X[0])
}
public void TestLibsvmParameters()
{
var clf = new Svc<int>(kernel: Kernel.Linear);
clf.Fit(X, Y);
Assert.IsTrue(clf.DualCoef.AlmostEquals(DenseMatrix.OfArray(new[,] {{0.25, -.25}})));
Assert.IsTrue(clf.Support.SequenceEqual(new[] {1, 3}));
Assert.IsTrue(
clf.SupportVectors.AlmostEquals(
DenseMatrix.OfRows(2, X.ColumnCount, new[] {X.Row(1), X.Row(3)})));
Assert.IsTrue(clf.Intercept.SequenceEqual(new[] {0.0}));
Assert.IsTrue(clf.Predict(X).SequenceEqual(Y));
}
/// <summary>
/// Make some classification predictions on a toy dataset using a SVC
///
/// If binary is True restrict to a binary classification problem instead of a
/// multiclass classification problem
/// </summary>
/// <param name="x"></param>
/// <param name="y"></param>
/// <param name="binary"></param>
private static Tuple<int[], int[], Matrix<double>> MakePrediction(
Matrix<double> x = null,
int[] y = null,
bool binary = false)
{
if (x == null && y == null)
{
// import some data to play with
var dataset = IrisDataset.Load();
x = dataset.Data;
y = dataset.Target;
}
if (binary)
{
// restrict to a binary classification task
x = x.RowsAt(y.Indices(v => v < 2));
y = y.Where(v => v < 2).ToArray();
}
int nSamples = x.RowCount;
int nFeatures = x.ColumnCount;
var rng = new Random(37);
int[] p = Shuffle(rng, Enumerable.Range(0, nSamples).ToArray());
x = x.RowsAt(p);
y = y.ElementsAt(p);
var half = nSamples/2;
// add noisy features to make the problem harder and avoid perfect results
rng = new Random(0);
x = x.HStack(DenseMatrix.CreateRandom(nSamples, 200, new Normal{RandomSource = rng}));
// run classifier, get class probabilities and label predictions
var clf = new Svc<int>(kernel: Kernel.Linear, probability: true);
clf.Fit(x.SubMatrix(0, half, 0, x.ColumnCount), y.Take(half).ToArray());
Matrix<double> probasPred = clf.PredictProba(x.SubMatrix(half, x.RowCount - half, 0, x.ColumnCount));
if (binary)
{
// only interested in probabilities of the positive case
// XXX: do we really want a special API for the binary case?
probasPred = probasPred.SubMatrix(0, probasPred.RowCount, 1, 1);
}
var yPred = clf.Predict(x.SubMatrix(half, x.RowCount - half, 0, x.ColumnCount));
var yTrue = y.Skip(half).ToArray();
return Tuple.Create(yTrue, yPred, probasPred);
}
public void TestLibsvmParameters()
{
var clf = new Svc <int>(kernel: Kernel.Linear);
clf.Fit(X, Y);
Assert.IsTrue(clf.DualCoef.AlmostEquals(DenseMatrix.OfArray(new[, ] {
{ 0.25, -.25 }
})));
Assert.IsTrue(clf.Support.SequenceEqual(new[] { 1, 3 }));
Assert.IsTrue(
clf.SupportVectors.AlmostEquals(
DenseMatrix.OfRows(2, X.ColumnCount, new[] { X.Row(1), X.Row(3) })));
Assert.IsTrue(clf.Intercept.SequenceEqual(new[] { 0.0 }));
Assert.IsTrue(clf.Predict(X).SequenceEqual(Y));
}
/// <summary>
/// Make sure some tweaking of parameters works.
///
/// We change clf.dual_coef_ at run time and expect .predict() to change
/// accordingly. Notice that this is not trivial since it involves a lot
/// of C/Python copying in the libsvm bindings.
///
/// The success of this test ensures that the mapping between libsvm and
/// the python classifier is complete.
/// </summary>
public void TestTweakParams()
{
var clf = new Svc <int>(kernel: Kernel.Linear, c: 1.0);
clf.Fit(X, Y);
Assert.IsTrue(clf.DualCoef.AlmostEquals(DenseMatrix.OfArray(new[, ] {
{ 0.25, -0.25 }
})));
Assert.IsTrue(clf.Predict(DenseMatrix.OfArray(new[, ] {
{ -.1, -.1 }
})).SequenceEqual(new[] { 1 }));
clf.DualCoef = DenseMatrix.OfArray(new[, ] {
{ 0.0, 1.0 }
});
Assert.IsTrue(clf.Predict(DenseMatrix.OfArray(new[, ] {
{ -.1, -.1 }
})).SequenceEqual(new[] { 2 }));
}
public void TestSvcWithCallableKernel()
{
// create SVM with callable linear kernel, check that results are the same
// as with built-in linear kernel
var svmCallable = new Svc <int>(kernel: Kernel.FromFunction((x, y) => x * (y.Transpose())),
probability: true);
svmCallable.Fit(X, Y);
var svmBuiltin = new Svc <int>(kernel: Kernel.Linear, probability: true);
svmBuiltin.Fit(X, Y);
Assert.IsTrue(svmCallable.DualCoef.AlmostEquals(svmBuiltin.DualCoef));
Assert.IsTrue(svmCallable.Intercept.AlmostEquals(svmBuiltin.Intercept));
Assert.IsTrue(svmCallable.Predict(X).SequenceEqual(svmBuiltin.Predict(X)));
Assert.IsTrue(svmCallable.PredictProba(X).AlmostEquals(svmBuiltin.PredictProba(X), 1));
Assert.IsTrue(svmCallable.DecisionFunction(X).AlmostEquals(svmBuiltin.DecisionFunction(X), 2));
}
public void TestSvcIris()
{
foreach (var k in new[] { Kernel.Linear, Kernel.Poly, Kernel.Rbf })
{
var spClf = new Svc <int>(kernel: k);
spClf.Fit(SparseMatrix.OfMatrix(iris.Data), iris.Target);
var clf = new Svc <int>(kernel: k);
clf.Fit(DenseMatrix.OfMatrix(iris.Data), iris.Target);
Assert.IsTrue(clf.SupportVectors.AlmostEquals(spClf.SupportVectors));
Assert.IsTrue(clf.DualCoef.AlmostEquals(spClf.DualCoef));
Assert.IsTrue(
clf.Predict(DenseMatrix.OfMatrix(iris.Data)).SequenceEqual(
spClf.Predict(SparseMatrix.OfMatrix(iris.Data))));
if (k == Kernel.Linear)
{
Assert.IsTrue(clf.Coef.AlmostEquals(spClf.Coef));
}
}
}
public void test_weight()
{
var clf = new Svc <int>(classWeightEstimator: ClassWeightEstimator <int> .Explicit(new Dictionary <int, double> {
{ 1, 0.1 }
}));
// we give a small weights to class 1
clf.Fit(X, Y);
// so all predicted values belong to class 2
Assert.IsTrue(clf.Predict(X).SequenceEqual(Enumerable.Repeat(2, 6)));
/*
* X_, y_ = make_classification(n_samples=200, n_features=10,
* weights=[0.833, 0.167], random_state=2)
*
* for clf in (linear_model.LogisticRegression(),
* svm.LinearSVC(random_state=0), svm.SVC()):
* clf.set_params(class_weight={0: .1, 1: 10})
* clf.fit(X_[:100], y_[:100])
* y_pred = clf.predict(X_[100:])
* assert_true(f1_score(y_[100:], y_pred) > .3)
* */
}
public void TestSingleSample_1D()
{
var clf = new Svc<int>();
clf.Fit(X, Y);
var p = clf.Predict(X.Row(0).ToRowMatrix());
Assert.AreEqual(Y[0], p[0]);
//todo:
//clf = svm.LinearSVC(random_state=0).fit(X, Y)
//clf.predict(X[0])
}
public void TestSparseRealdata()
{
var x = new SparseMatrix(80, 36);
x[7, 6] = 0.03771744;
x[39, 5] = 0.1003567;
x[77, 35] = 0.01174647;
x[77, 31] = 0.027069;
var y = new[]
{
1.0, 0.0, 2.0, 2.0, 1.0, 1.0, 1.0, 2.0, 2.0, 0.0, 1.0, 2.0, 2.0,
0.0, 2.0, 0.0, 3.0, 0.0, 3.0, 0.0, 1.0, 1.0, 3.0, 2.0, 3.0, 2.0,
0.0, 3.0, 1.0, 0.0, 2.0, 1.0, 2.0, 0.0, 1.0, 0.0, 2.0, 3.0, 1.0,
3.0, 0.0, 1.0, 0.0, 0.0, 2.0, 0.0, 1.0, 2.0, 2.0, 2.0, 3.0, 2.0,
0.0, 3.0, 2.0, 1.0, 2.0, 3.0, 2.0, 2.0, 0.0, 1.0, 0.0, 1.0, 2.0,
3.0, 0.0, 0.0, 2.0, 2.0, 1.0, 3.0, 1.0, 1.0, 0.0, 1.0, 2.0, 1.0,
1.0, 3.0
};
var clf = new Svc<double>(kernel: Kernel.Linear);
clf.Fit(DenseMatrix.OfMatrix(x), y);
var spClf = new Svc<double>(kernel: Kernel.Linear);
spClf.Fit(x, y);
Assert.IsTrue(clf.SupportVectors.AlmostEquals(spClf.SupportVectors));
Assert.IsTrue(clf.DualCoef.AlmostEquals(spClf.DualCoef));
}
public void TestSvcIris()
{
foreach (var k in new[] {Kernel.Linear, Kernel.Poly, Kernel.Rbf})
{
var spClf = new Svc<int>(kernel: k);
spClf.Fit(SparseMatrix.OfMatrix(iris.Data), iris.Target);
var clf = new Svc<int>(kernel: k);
clf.Fit(DenseMatrix.OfMatrix(iris.Data), iris.Target);
Assert.IsTrue(clf.SupportVectors.AlmostEquals(spClf.SupportVectors));
Assert.IsTrue(clf.DualCoef.AlmostEquals(spClf.DualCoef));
Assert.IsTrue(
clf.Predict(DenseMatrix.OfMatrix(iris.Data)).SequenceEqual(
spClf.Predict(SparseMatrix.OfMatrix(iris.Data))));
if (k == Kernel.Linear)
{
Assert.IsTrue(clf.Coef.AlmostEquals(spClf.Coef));
}
}
}
public void TestSvcWithCustomKernel()
{
var clfLin = new Svc<int>(kernel: Kernel.Linear);
clfLin.Fit(SparseMatrix.OfMatrix(X), Y);
var clfMylin =
new Svc<int>(kernel: Kernel.FromFunction((x, y) => x*y.Transpose()));
clfMylin.Fit(SparseMatrix.OfMatrix(X), Y);
Assert.IsTrue(
clfLin.Predict(SparseMatrix.OfMatrix(X)).SequenceEqual(clfMylin.Predict(SparseMatrix.OfMatrix(X))));
}
public void TestSvcBadKernel()
{
var svc = new Svc<int>(kernel: Kernel.FromFunction((x, y) => x));
svc.Fit(X, Y);
}
public void TestSvcWithCallableKernel()
{
// create SVM with callable linear kernel, check that results are the same
// as with built-in linear kernel
var svmCallable = new Svc<int>(kernel: Kernel.FromFunction((x, y) => x*(y.Transpose())),
probability: true);
svmCallable.Fit(X, Y);
var svmBuiltin = new Svc<int>(kernel: Kernel.Linear, probability: true);
svmBuiltin.Fit(X, Y);
Assert.IsTrue(svmCallable.DualCoef.AlmostEquals(svmBuiltin.DualCoef));
Assert.IsTrue(svmCallable.Intercept.AlmostEquals(svmBuiltin.Intercept));
Assert.IsTrue(svmCallable.Predict(X).SequenceEqual(svmBuiltin.Predict(X)));
Assert.IsTrue(svmCallable.PredictProba(X).AlmostEquals(svmBuiltin.PredictProba(X), 1));
Assert.IsTrue(svmCallable.DecisionFunction(X).AlmostEquals(svmBuiltin.DecisionFunction(X), 2));
}
public void test_weight()
{
var clf = new Svc<int>(classWeightEstimator: ClassWeightEstimator<int>.Explicit(new Dictionary<int, double> {{1, 0.1}}));
// we give a small weights to class 1
clf.Fit(X, Y);
// so all predicted values belong to class 2
Assert.IsTrue(clf.Predict(X).SequenceEqual(Enumerable.Repeat(2, 6)));
/*
X_, y_ = make_classification(n_samples=200, n_features=10,
weights=[0.833, 0.167], random_state=2)
for clf in (linear_model.LogisticRegression(),
svm.LinearSVC(random_state=0), svm.SVC()):
clf.set_params(class_weight={0: .1, 1: 10})
clf.fit(X_[:100], y_[:100])
y_pred = clf.predict(X_[100:])
assert_true(f1_score(y_[100:], y_pred) > .3)
* */
}
public void TestDecisionFunction()
{
// multi class:
var clf = new Svc<int>(kernel: Kernel.Linear, c: 0.1);
clf.Fit(iris.Data, iris.Target);
var dec = (iris.Data*clf.Coef.Transpose()).AddRowVector(clf.Intercept);
Assert.IsTrue(dec.AlmostEquals(clf.DecisionFunction(iris.Data)));
// binary:
clf.Fit(X, Y);
dec = (X*clf.Coef.Transpose()).AddRowVector(clf.Intercept);
int[] prediction = clf.Predict(X);
Assert.IsTrue(dec.AlmostEquals(clf.DecisionFunction(X)));
var b = clf.DecisionFunction(X).Column(0).Select(v => clf.Classes[v > 0 ? 1 : 0]);
Assert.IsTrue(prediction.SequenceEqual(b));
var expected = DenseMatrix.OfArray(new[,] {{-1.0}, {-0.66}, {-1.0}, {0.66}, {1.0}, {1.0}});
Assert.IsTrue(clf.DecisionFunction(X).AlmostEquals(expected, 1E-2));
}
/// <summary>
/// Make sure some tweaking of parameters works.
///
/// We change clf.dual_coef_ at run time and expect .predict() to change
/// accordingly. Notice that this is not trivial since it involves a lot
/// of C/Python copying in the libsvm bindings.
///
/// The success of this test ensures that the mapping between libsvm and
/// the python classifier is complete.
/// </summary>
public void TestTweakParams()
{
var clf = new Svc<int>(kernel: Kernel.Linear, c: 1.0);
clf.Fit(X, Y);
Assert.IsTrue(clf.DualCoef.AlmostEquals(DenseMatrix.OfArray(new[,]{{0.25, -0.25}})));
Assert.IsTrue(clf.Predict(DenseMatrix.OfArray(new[,] {{-.1, -.1}})).SequenceEqual(new[] {1}));
clf.DualCoef = DenseMatrix.OfArray(new[,] {{0.0, 1.0}});
Assert.IsTrue(clf.Predict(DenseMatrix.OfArray(new[,] {{-.1, -.1}})).SequenceEqual(new[] {2}));
}
public void TestPrecomputed()
{
var clf = new Svc <int>(kernel: Kernel.Precomputed);
// Gram matrix for train data (square matrix)
// (we use just a linear kernel)
var k = X * (X.Transpose());
clf.Fit(k, Y);
// Gram matrix for test data (rectangular matrix)
var kt = T * X.Transpose();
var pred = clf.Predict(kt);
try
{
clf.Predict(kt.Transpose());
Assert.Fail();
}
catch (ArgumentException)
{
}
Assert.IsTrue(clf.DualCoef.AlmostEquals(DenseMatrix.OfArray(new[, ] {
{ 0.25, -.25 }
})));
Assert.IsTrue(clf.Support.SequenceEqual(new[] { 1, 3 }));
Assert.IsTrue(clf.Intercept.SequenceEqual(new[] { 0.0 }));
Assert.IsTrue(pred.SequenceEqual(true_result));
// Gram matrix for test data but compute KT[i,j]
// for support vectors j only.
kt = kt.CreateMatrix(kt.RowCount, kt.ColumnCount);
for (int i = 0; i < T.RowCount; i++)
{
foreach (var j in clf.Support)
{
kt[i, j] = T.Row(i) * X.Row(j);
}
}
pred = clf.Predict(kt);
Assert.IsTrue(pred.SequenceEqual(true_result));
// same as before, but using a callable function instead of the kernel
// matrix. kernel is just a linear kernel
clf = new Svc <int>(kernel: Kernel.FromFunction((x, y) => x * y.Transpose()));
clf.Fit(X, Y);
pred = clf.Predict(T);
Assert.IsTrue(clf.DualCoef.AlmostEquals(DenseMatrix.OfArray(new[, ] {
{ 0.25, -.25 }
})));
Assert.IsTrue(clf.Support.SequenceEqual(new[] { 1, 3 }));
Assert.IsTrue(clf.Intercept.SequenceEqual(new[] { 0.0 }));
Assert.IsTrue(pred.SequenceEqual(true_result));
// test a precomputed kernel with the iris dataset
// and check parameters against a linear SVC
clf = new Svc <int>(kernel: Kernel.Precomputed);
var clf2 = new Svc <int>(kernel: Kernel.Linear);
k = iris.Data * iris.Data.Transpose();
clf.Fit(k, iris.Target);
clf2.Fit(iris.Data, iris.Target);
pred = clf.Predict(k);
Assert.IsTrue(clf.Support.SequenceEqual(clf2.Support));
Assert.IsTrue(clf.DualCoef.AlmostEquals(clf2.DualCoef));
Assert.IsTrue(clf.Intercept.AlmostEquals(clf2.Intercept));
var matchingN = pred.Zip(iris.Target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0 * matchingN / pred.Length > 0.99);
// Gram matrix for test data but compute KT[i,j]
// for support vectors j only.
k = k.CreateMatrix(k.RowCount, k.ColumnCount);
for (int i = 0; i < iris.Data.RowCount; i++)
{
foreach (var j in clf.Support)
{
k[i, j] = iris.Data.Row(i) * iris.Data.Row(j);
}
}
pred = clf.Predict(k);
matchingN = pred.Zip(iris.Target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0 * matchingN / pred.Length > 0.99);
clf = new Svc <int>(kernel: Kernel.FromFunction((x, y) => x * y.Transpose()));
clf.Fit(iris.Data, iris.Target);
matchingN = pred.Zip(iris.Target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0 * matchingN / pred.Length > 0.99);
}
public void TestSvc()
{
var clf = new Svc<int>(kernel: Kernel.Linear, probability: true);
clf.Fit(X, Y);
var spClf = new Svc<int>(kernel: Kernel.Linear, probability: true);
spClf.Fit(SparseMatrix.OfMatrix(X), Y);
Assert.IsTrue(spClf.Predict(T).SequenceEqual(true_result));
Assert.IsTrue(spClf.SupportVectors is SparseMatrix);
Assert.IsTrue(clf.SupportVectors.AlmostEquals(spClf.SupportVectors));
Assert.IsTrue(spClf.DualCoef is SparseMatrix);
Assert.IsTrue(clf.DualCoef.AlmostEquals(spClf.DualCoef));
Assert.IsTrue(spClf.Coef is SparseMatrix);
Assert.IsTrue(clf.Coef.AlmostEquals(spClf.Coef));
Assert.IsTrue(clf.Support.SequenceEqual(spClf.Support));
Assert.IsTrue(clf.Predict(T).SequenceEqual(spClf.Predict(T)));
// refit with a different dataset
clf.Fit(X2, Y2);
spClf.Fit(SparseMatrix.OfMatrix(X2), Y2);
Assert.IsTrue(clf.SupportVectors.AlmostEquals(spClf.SupportVectors));
Assert.IsTrue(clf.DualCoef.AlmostEquals(spClf.DualCoef));
Assert.IsTrue(clf.Coef.AlmostEquals(spClf.Coef));
Assert.IsTrue(clf.Support.SequenceEqual(spClf.Support));
Assert.IsTrue(clf.Predict(T2).SequenceEqual(spClf.Predict(T2)));
Assert.IsTrue(clf.PredictProba(T2).AlmostEquals(spClf.PredictProba(T2), 0.001));
}
public void TestSvcBadKernel()
{
var svc = new Svc <int>(kernel: Kernel.FromFunction((x, y) => x));
svc.Fit(X, Y);
}
public void TestLibsvmIris()
{
// shuffle the dataset so that labels are not ordered
foreach (var k in new[] {Kernel.Linear, Kernel.Rbf})
{
var clf = new Svc<int>(kernel: k);
clf.Fit(iris.Data, iris.Target);
var pred = clf.Predict(iris.Data);
var matchingN = pred.Zip(iris.Target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0*matchingN/pred.Length > 0.9);
Assert.IsTrue(clf.Classes.SequenceEqual(clf.Classes.OrderBy(v => v)));
}
}
public void TestPrecomputed()
{
var clf = new Svc<int>(kernel: Kernel.Precomputed);
// Gram matrix for train data (square matrix)
// (we use just a linear kernel)
var k = X*(X.Transpose());
clf.Fit(k, Y);
// Gram matrix for test data (rectangular matrix)
var kt = T*X.Transpose();
var pred = clf.Predict(kt);
try
{
clf.Predict(kt.Transpose());
Assert.Fail();
}
catch (ArgumentException)
{
}
Assert.IsTrue(clf.DualCoef.AlmostEquals(DenseMatrix.OfArray(new[,] {{0.25, -.25}})));
Assert.IsTrue(clf.Support.SequenceEqual(new[] {1, 3}));
Assert.IsTrue(clf.Intercept.SequenceEqual(new[] {0.0}));
Assert.IsTrue(pred.SequenceEqual(true_result));
// Gram matrix for test data but compute KT[i,j]
// for support vectors j only.
kt = kt.CreateMatrix(kt.RowCount, kt.ColumnCount);
for (int i = 0; i < T.RowCount; i++)
{
foreach (var j in clf.Support)
{
kt[i, j] = T.Row(i)*X.Row(j);
}
}
pred = clf.Predict(kt);
Assert.IsTrue(pred.SequenceEqual(true_result));
// same as before, but using a callable function instead of the kernel
// matrix. kernel is just a linear kernel
clf = new Svc<int>(kernel: Kernel.FromFunction((x, y) => x*y.Transpose()));
clf.Fit(X, Y);
pred = clf.Predict(T);
Assert.IsTrue(clf.DualCoef.AlmostEquals(DenseMatrix.OfArray(new[,] {{0.25, -.25}})));
Assert.IsTrue(clf.Support.SequenceEqual(new[] {1, 3}));
Assert.IsTrue(clf.Intercept.SequenceEqual(new[] {0.0}));
Assert.IsTrue(pred.SequenceEqual(true_result));
// test a precomputed kernel with the iris dataset
// and check parameters against a linear SVC
clf = new Svc<int>(kernel: Kernel.Precomputed);
var clf2 = new Svc<int>(kernel: Kernel.Linear);
k = iris.Data*iris.Data.Transpose();
clf.Fit(k, iris.Target);
clf2.Fit(iris.Data, iris.Target);
pred = clf.Predict(k);
Assert.IsTrue(clf.Support.SequenceEqual(clf2.Support));
Assert.IsTrue(clf.DualCoef.AlmostEquals(clf2.DualCoef));
Assert.IsTrue(clf.Intercept.AlmostEquals(clf2.Intercept));
var matchingN = pred.Zip(iris.Target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0*matchingN/pred.Length > 0.99);
// Gram matrix for test data but compute KT[i,j]
// for support vectors j only.
k = k.CreateMatrix(k.RowCount, k.ColumnCount);
for (int i = 0; i < iris.Data.RowCount; i++)
{
foreach (var j in clf.Support)
k[i, j] = iris.Data.Row(i)*iris.Data.Row(j);
}
pred = clf.Predict(k);
matchingN = pred.Zip(iris.Target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0*matchingN/pred.Length > 0.99);
clf = new Svc<int>(kernel: Kernel.FromFunction((x, y) => x*y.Transpose()));
clf.Fit(iris.Data, iris.Target);
matchingN = pred.Zip(iris.Target, Tuple.Create).Where(t => t.Item1 == t.Item2).Count();
Assert.IsTrue(1.0*matchingN/pred.Length > 0.99);
}
