public void computeGaussianPerplexity_1()
{
double[][] points =
{
new double[] { 2, 3, 2 },
new double[] { 5, 4, 5 },
new double[] { 9, 6, 4 },
new double[] { 4, 7, 5 },
new double[] { 8, 1, 1 },
new double[] { 1, 2, 4 },
};
double perplexity = 0.5;
int N = points.Length;
int D = 3;
var X = points.ToMatrix();
double[,] expected = new double[N, N];
TSNEWrapper.computeGaussianPerplexity(X, N, D, expected, perplexity);
double[][] actual = Jagged.Zeros(N, N);
TSNE.computeGaussianPerplexity(points, N, D, ref actual, perplexity);
Assert.IsTrue(actual.IsEqual(expected, rtol: 1e-5));
}
public void computeGaussianPerplexity2_larger()
{
var points = yinyang.Submatrix(null, 0, 1).ToJagged();
double perplexity = 0.5;
int N = points.Rows();
int D = points.Columns();
int K = (int)(3 * perplexity);
double[,] X = points.ToMatrix();
double[][] x = X.ToJagged();
uint[] expected_row = Vector.Zeros <uint>(200);
uint[] expected_col = Vector.Zeros <uint>(200);
double[] expected_val = Vector.Zeros <double>(200);
TSNEWrapper.computeGaussianPerplexity(X, N, D, expected_row, expected_col, expected_val, perplexity, K);
int[] actual_row = null;
int[] actual_col = null;
double[] actual_val = null;
TSNE.computeGaussianPerplexity(x, N, D, ref actual_row, ref actual_col, ref actual_val, perplexity, K);
expected_row = expected_row.First(101);
expected_col = expected_col.First(100);
expected_val = expected_val.First(100);
Assert.IsTrue(actual_row.IsEqual(expected_row));
Assert.IsTrue(actual_col.IsEqual(expected_col));
Assert.IsTrue(actual_val.IsEqual(expected_val, 1e-4));
}
public void compute_squared_distance_larger()
{
var points = yinyang.Submatrix(null, 0, 1).ToJagged();
var X = points.ToMatrix();
int N = X.Rows();
int D = X.Columns();
double[,] expected = new double[N, N];
TSNEWrapper.computeSquaredEuclideanDistance(X, expected);
double[][] actual = Jagged.Zeros(N, N);
TSNE.computeSquaredEuclideanDistance(points, N, D, actual);
Assert.IsTrue(actual.IsEqual(expected));
}
public void computeGradient_1()
{
Accord.Math.Random.Generator.Seed = 0;
double perplexity = 0.5;
double theta = 0.5;
int N = 100;
int K = (int)(3 * perplexity);
int D = 3;
uint[] row_P = Vector.Create(N + 1, new uint[] { 0, 1, 2, 3, 4, 5, 6 });
uint[] col_P = Vector.Create(N * K, new uint[] { 5, 3, 1, 1, 2, 1 });
double[] val_P = Vector.Create(N * K, new double[]
{
0.83901046609114708,
0.39701047304189827,
0.19501046869768451,
0.59401047304189827,
0.49301046869768484,
0.59901046869768451,
});
double[,] P = Matrix.Random(N, N, new NormalDistribution());
double[][] p = P.ToJagged();
double[,] Y = Matrix.Random(N, D, new NormalDistribution());
double[][] y = Y.ToJagged();
uint[] expected_row = Vector.Create(row_P);
uint[] expected_col = Vector.Create(col_P);
double[] expected_val = Vector.Create(val_P);
double[,] expected = Matrix.Zeros(N, D);
TSNEWrapper.computeGradient(P, expected_row, expected_col, expected_val, Y, N, D, expected, theta);
int[] actual_row = row_P.To <int[]>();
int[] actual_col = col_P.To <int[]>();
double[] actual_val = (double[])val_P.Clone();
double[][] actual = Jagged.Zeros(N, D);
TSNE.computeGradient(p, actual_row, actual_col, actual_val, y, N, D, actual, theta);
Assert.IsTrue(actual.IsEqual(expected));
Assert.IsTrue(actual_row.IsEqual(expected_row));
Assert.IsTrue(actual_col.IsEqual(expected_col));
Assert.IsTrue(actual_val.IsEqual(expected_val, 1e-4));
}
public void evaluateError_2()
{
int N = 6;
int D = 2;
double[,] P = Matrix.Random(6, 6, new NormalDistribution());
double[][] p = P.ToJagged();
double[,] Y = Matrix.Random(6, 2, new NormalDistribution());
double[][] y = Y.ToJagged();
double expected = TSNEWrapper.evaluateError(P, Y, N, D);
double actual = TSNE.evaluateError(p, y, N, D);
Assert.AreEqual(expected, actual);
}
public void computeGaussianPerplexity_larger()
{
var points = yinyang.Submatrix(null, 0, 1).ToJagged();
double perplexity = 0.5;
int N = points.Rows();
int D = points.Columns();
var X = points.ToMatrix();
double[,] expected = new double[N, N];
TSNEWrapper.computeGaussianPerplexity(X, N, D, expected, perplexity);
double[][] actual = Jagged.Zeros(N, N);
TSNE.computeGaussianPerplexity(points, N, D, ref actual, perplexity);
Assert.IsTrue(actual.IsEqual(expected, rtol: 1e-5));
}
public void evaluateError_1()
{
double perplexity = 0.5;
double theta = 0.5;
int N = 6;
int K = (int)(3 * perplexity);
int D = 2;
uint[] row_P = Vector.Create(N + 1, new uint[] { 0, 1, 2, 3, 4, 5, 6 });
uint[] col_P = Vector.Create(N * K, new uint[] { 5, 3, 1, 1, 2, 1 });
double[] val_P = Vector.Create(N * K, new double[]
{
0.99901046609114708,
0.99901047304189827,
0.99901046869768451,
0.99901047304189827,
0.99901046869768484,
0.99901046869768451,
});
double[,] Y = Matrix.Random(6, 2, new NormalDistribution());
double[][] y = Y.ToJagged();
uint[] expected_row = Vector.Create(row_P);
uint[] expected_col = Vector.Create(col_P);
double[] expected_val = Vector.Create(val_P);
double expected = TSNEWrapper.evaluateError(expected_row, expected_col, expected_val, Y, N, D, theta);
int[] actual_row = row_P.To <int[]>();
int[] actual_col = col_P.To <int[]>();
double[] actual_val = (double[])val_P.Clone();
double actual = TSNE.evaluateError(actual_row, actual_col, actual_val, y, N, D, theta);
Assert.AreEqual(expected, actual);
Assert.IsTrue(actual_col.IsEqual(expected_col));
Assert.IsTrue(actual_row.IsEqual(expected_row));
Assert.IsTrue(actual_val.IsEqual(expected_val));
}
public void computeGaussianPerplexity_2()
{
double[][] points =
{
new double[] { 2, 3, 20 },
new double[] { 5, 4, 5 },
new double[] { 9, 6, 500 },
new double[] { 4, 7, -100 },
new double[] { 8, 1, 67 },
new double[] { 1, 2, -888 },
};
double perplexity = 0.5;
int N = points.Length;
int D = 3;
int K = (int)(3 * perplexity);
double[,] X = points.ToMatrix();
double[][] x = X.ToJagged();
uint[] expected_row = Vector.Zeros <uint>(100);
uint[] expected_col = Vector.Zeros <uint>(100);
double[] expected_val = Vector.Zeros <double>(100);
TSNEWrapper.computeGaussianPerplexity(X, N, D, expected_row, expected_col, expected_val, perplexity, K);
int[] actual_row = null;
int[] actual_col = null;
double[] actual_val = null;
TSNE.computeGaussianPerplexity(x, N, D, ref actual_row, ref actual_col, ref actual_val, perplexity, K);
expected_row = expected_row.First(7);
expected_col = expected_col.First(6);
expected_val = expected_val.First(6);
Assert.IsTrue(actual_row.IsEqual(expected_row));
Assert.IsTrue(actual_col.IsEqual(expected_col));
Assert.IsTrue(actual_val.IsEqual(expected_val, 1e-4));
}
public void FromDataTest2()
{
Accord.Math.Random.Generator.Seed = 0;
int N = yinyang.GetLength(0);
var init = Matrix.Random(N, 1, new NormalDistribution(0, 0.001));
var X = yinyang.Submatrix(null, 0, 1);
var Y = (double[, ])init.Clone();
var x = X.ToJagged();
var y = Y.ToJagged();
double perplexity = 20;
double theta = 0.5;
TSNEWrapper.run(X, Y, perplexity, theta);
var expected = Y.Flatten();
TSNE.run(x, y, perplexity, theta, true);
var actual = y.Flatten();
Assert.IsTrue(actual.IsEqual(expected));
}
public void symmetrizeMatrix_1()
{
double perplexity = 0.5;
int n = 6;
int k = (int)(3 * perplexity);
uint[] row_P = Vector.Create(n + 1, new uint[] { 0, 1, 2, 3, 4, 5, 6 });
uint[] col_P = Vector.Create(n * k, new uint[] { 5, 3, 1, 1, 2, 1 });
double[] val_P = Vector.Create(n * k, new double[]
{
0.99901046609114708,
0.99901047304189827,
0.99901046869768451,
0.99901047304189827,
0.99901046869768484,
0.99901046869768451,
});
uint[] expected_row = Vector.Create(100, row_P);
uint[] expected_col = Vector.Create(100, col_P);
double[] expected_val = Vector.Create(100, val_P);
TSNEWrapper.symmetrizeMatrix(expected_row, expected_col, expected_val, n);
int[] actual_row = row_P.To <int[]>();
int[] actual_col = col_P.To <int[]>();
double[] actual_val = (double[])val_P.Clone();
TSNE.symmetrizeMatrix(ref actual_row, ref actual_col, ref actual_val, n);
expected_row = expected_row.First(7);
expected_col = expected_col.First(10);
expected_val = expected_val.First(10);
Assert.IsTrue(actual_col.IsEqual(expected_col));
Assert.IsTrue(actual_row.IsEqual(expected_row));
Assert.IsTrue(actual_val.IsEqual(expected_val));
}
public void compute_squared_distance_1()
{
double[][] points =
{
new double[] { 2, 3, 2 },
new double[] { 5, 4, 5 },
new double[] { 9, 6, 4 },
new double[] { 4, 7, 5 },
new double[] { 8, 1, 1 },
new double[] { 1, 2, 4 },
};
var X = points.ToMatrix();
int N = X.Rows();
int D = X.Columns();
double[,] expected = new double[N, N];
TSNEWrapper.computeSquaredEuclideanDistance(X, expected);
double[][] actual = Jagged.Zeros(N, N);
TSNE.computeSquaredEuclideanDistance(points, N, D, actual);
Assert.IsTrue(actual.IsEqual(expected));
}