import numpy import numpy.random import csb.test as test from csb.statistics.pdf import ParameterValueError, EstimationFailureError from csb.statistics.pdf import Normal, GeneralizedNormal, Laplace, Gamma, InverseGamma, GumbelMaximum, GumbelMinimum from csb.statistics.pdf import MultivariateGaussian, Dirichlet, InverseGaussian, GeneralizedInverseGaussian @test.regression class TestDensityRegressions(test.Case): def testParameterValidation(self): """ @see: [CSB 0000132] """ self.assertRaises(ParameterValueError, InverseGaussian, -1, 1) pdf = InverseGaussian() data = [ 3.1, 1.15, 6.86, 0.69, -1.73, -2.91, -2.2800744] def testMu(v): pdf['mu'] = v pdf.mu = v self.assertRaises(ParameterValueError, testMu, -1) self.assertRaises(EstimationFailureError, pdf.estimate, data) @test.unit class TestNormal(test.Case): def testParameters(self): pdf = Normal(2.5, 1.5) pdf.mu = 33 pdf.sigma = 44 self.assertEqual(pdf.mu, 33) self.assertEqual(pdf.sigma, 44) self.assertEqual(pdf.sigma, pdf['sigma']) def testLogProb(self): n = Normal(0, 1) self.assertAlmostEqual(n(0.), 1. / numpy.sqrt(2 * numpy.pi), delta=1e-1) self.assertAlmostEqual(n(1.), numpy.exp(-0.5) / numpy.sqrt(2 * numpy.pi), delta=1e-1) def testParameterEstimation(self): mu, sigma = 2.2, 0.3 data = numpy.random.normal(mu, sigma, 100000) pdf = Normal(1, 1) pdf.estimate(data) self.assertAlmostEqual(pdf.mu, mu, delta=0.05) self.assertAlmostEqual(pdf.sigma, sigma, delta=0.05) def testRandom(self): pdf = Normal(4, 2) data = pdf.random(100000) self.assertAlmostEqual(numpy.mean(data), pdf.mu, delta=0.05) self.assertAlmostEqual(numpy.std(data), pdf.sigma, delta=0.05) @test.unit class TestLaplace(test.Case): def testParameters(self): pdf = Laplace(1.5, 2.5) pdf.mu = 33 pdf.b = 44 self.assertEqual(pdf.mu, 33) self.assertEqual(pdf.b, 44) self.assertEqual(pdf.b, pdf['b']) def propertyAssignment(): pdf.b = -1 def directAssignment(): pdf['b'] = -1 self.assertRaises(ParameterValueError, propertyAssignment) self.assertRaises(ParameterValueError, directAssignment) def testLogProb(self): pdf = Laplace(0, 2) self.assertEqual(pdf(0), 1.0 / (2 * pdf.b)) self.assertEqual(pdf(1), 1 / (2 * pdf.b) * numpy.exp(-1 / pdf.b)) def testParameterEstimation(self): mu, b = 2.2, 2 data = numpy.random.laplace(mu, b, 100000) pdf = Laplace(1, 1) pdf.estimate(data) self.assertAlmostEqual(pdf.mu, mu, places=1) self.assertAlmostEqual(pdf.b, b, places=1) def testRandom(self): pdf = Laplace(4, 2) data = pdf.random(100000) self.assertAlmostEqual(numpy.median(data), pdf.mu, delta=0.05) @test.unit class TestInverseGaussian(test.Case): def testParameters(self): pdf = InverseGaussian(1.5, 2.5) pdf.mu = 3 pdf.shape = 4 self.assertEqual(pdf.mu, 3) self.assertEqual(pdf.shape, 4) self.assertEqual(pdf.mu, pdf['mu']) def propertyAssignment(): pdf.shape = -1 def directAssignment(): pdf['shape'] = -1 self.assertRaises(ParameterValueError, propertyAssignment) self.assertRaises(ParameterValueError, directAssignment) def testLogProb(self): ig = InverseGaussian(1., 1.) self.assertAlmostEqual(ig(1.), numpy.sqrt(1 / (2 * numpy.pi)), delta=1e-5) self.assertAlmostEqual(ig(2.), (numpy.sqrt(1 / (2 * numpy.pi * 2 ** 3)) * numpy.exp(-1. / 4.)), delta=1e-5) def testParameterEstimation(self): mu, shape = 1, 3 data = numpy.random.wald(mu, shape, 100000) pdf = InverseGaussian(1, 1) pdf.estimate(data) self.assertAlmostEqual(pdf.mu, mu, places=2) self.assertAlmostEqual(pdf.shape, shape, delta=0.05) def testRandom(self): ig = InverseGaussian(1., 1.) samples = ig.random(1000000) mu = numpy.mean(samples) var = numpy.var(samples) self.assertAlmostEqual(ig.mu, mu, delta=1e-1) self.assertAlmostEqual(ig.mu ** 3 / ig.shape, var, delta=1e-1) ig = InverseGaussian(3., 6.) samples = ig.random(1000000) mu = numpy.mean(samples) var = numpy.var(samples) self.assertAlmostEqual(ig.mu, mu, delta=1e-1) self.assertAlmostEqual(ig.mu ** 3 / ig.shape, var, delta=5e-1) @test.unit class TestGeneralizedNormal(test.Case): def testParameters(self): pdf = GeneralizedNormal(0, 1, 1) pdf.mu = 11 pdf.alpha = 2 pdf.beta = 3 self.assertEqual(pdf.mu, 11) self.assertEqual(pdf.alpha, 2) self.assertEqual(pdf.beta, 3) self.assertEqual(pdf.beta, pdf['beta']) def propertyAssignment(): pdf.beta = -1 def directAssignment(p): pdf[p] = -1 self.assertRaises(ParameterValueError, propertyAssignment) self.assertRaises(ParameterValueError, directAssignment, 'alpha') self.assertRaises(ParameterValueError, directAssignment, 'beta') def testLogProb(self): from scipy.special import gamma pdf = GeneralizedNormal(0, 1, 1) a = pdf.alpha b = pdf.beta self.assertEqual(pdf(0), b / (2 * a * gamma(1 / b))) self.assertAlmostEqual(pdf(1), 1. / (2 * gamma(1)) * numpy.exp(-1), delta=1e-5) def testParameterEstimation(self): mu, alpha, beta = -0.04, 2.11, 1.90 data = [-1.1712, -2.5608, -0.7143, 2.6218, -2.0655, 0.7544, 1.208, -0.5289, 0.0045, 1.1746, -1.0766, 1.1198, 1.2785, -0.6051, 2.2913, -3.6672, -0.2525, 0.8782, -0.0617, -0.0239] pdf = GeneralizedNormal(1, 1, 1) pdf.estimate(data) self.assertAlmostEqual(pdf.mu, mu, places=2) self.assertAlmostEqual(pdf.alpha, alpha, places=2) self.assertAlmostEqual(pdf.beta, beta, places=1) @test.unit class TestGeneralizedInverseGaussian(test.Case): def testParameters(self): pdf = GeneralizedInverseGaussian(1, 1, 1) pdf.a = 2 pdf.b = 3 pdf.p = 4 self.assertEqual(pdf.a, 2) self.assertEqual(pdf.b, 3) self.assertEqual(pdf.p, 4) self.assertEqual(pdf.p, pdf['p']) def propertyAssignment(): pdf.a = -1 def directAssignment(p): pdf[p] = -1 self.assertRaises(ParameterValueError, propertyAssignment) self.assertRaises(ParameterValueError, directAssignment, 'a') self.assertRaises(ParameterValueError, directAssignment, 'b') self.assertRaises(ParameterValueError, directAssignment, 'p') def testLogProb(self): fx = [2.7939e-24, 1.5305e-03, 2.3028e-02, 6.6302e-02, 1.1759e-01, 1.6810e-01, 2.1365e-01, 2.5255e-01, 2.8438e-01, 3.0937e-01, 3.2806e-01, 3.4111e-01, 3.4922e-01, 3.5307e-01, 3.5329e-01, 3.5048e-01, 3.4515e-01, 3.3778e-01, 3.2878e-01, 3.1850e-01, 3.0726e-01, 2.9531e-01, 2.8289e-01, 2.7018e-01, 2.5736e-01, 2.4454e-01, 2.3185e-01, 2.1937e-01, 2.0717e-01, 1.9532e-01, 1.8385e-01, 1.7280e-01, 1.6220e-01, 1.5205e-01, 1.4236e-01, 1.3315e-01, 1.2440e-01, 1.1611e-01, 1.0828e-01, 1.0088e-01, 9.3917e-02, 8.7363e-02, 8.1207e-02, 7.5432e-02, 7.0021e-02, 6.4958e-02, 6.0224e-02, 5.5804e-02] x = numpy.arange(0.01, 5., 0.1) a = 2. b = 1. p = 2 gig = GeneralizedInverseGaussian(a, b, p) fx2 = gig(x) for i in range(len(fx)): self.assertAlmostEqual(fx[i], fx2[i], delta=1e-1) def testRandom(self): from scipy.special import kv from numpy import sqrt a = 2. b = 1. p = 1 gig = GeneralizedInverseGaussian(a, b, p) samples = gig.random(10000) mu_analytical = sqrt(b) * kv(p + 1, sqrt(a * b)) / (sqrt(a) * kv(p, sqrt(a * b))) var_analytical = b * kv(p + 2, sqrt(a * b)) / a / kv(p, sqrt(a * b)) - mu_analytical ** 2 mu = numpy.mean(samples) var = numpy.var(samples) self.assertAlmostEqual(mu_analytical, mu, delta=1e-1) self.assertAlmostEqual(var_analytical, var, delta=1e-1) @test.unit class TestGamma(test.Case): def testParameters(self): pdf = Gamma(1, 1) pdf.alpha = 2 pdf.beta = 3 self.assertEqual(pdf.alpha, 2) self.assertEqual(pdf.beta, 3) self.assertEqual(pdf.beta, pdf['beta']) def propertyAssignment(): pdf.beta = -1 def directAssignment(p): pdf[p] = -1 self.assertRaises(ParameterValueError, propertyAssignment) self.assertRaises(ParameterValueError, directAssignment, 'alpha') self.assertRaises(ParameterValueError, directAssignment, 'beta') def testLogProb(self): g = Gamma(1., 1.) self.assertAlmostEqual(g(1.), numpy.exp(-1.), delta=1e-5) self.assertAlmostEqual(g(2.), numpy.exp(-2.), delta=1e-5) g = Gamma(2., 1.) self.assertAlmostEqual(g(1.), 1. * numpy.exp(-1), delta=1e-5) self.assertAlmostEqual(g(2.), 2. * numpy.exp(-2), delta=1e-5) def testParameterEstimation(self): alpha = 0.1 beta = 0.1 data = numpy.random.gamma(alpha, 1. / beta, 10000) pdf = Gamma(1, 1) pdf.estimate(data) self.assertAlmostEqual(pdf.alpha, alpha, places=2) self.assertAlmostEqual(pdf.beta, beta, places=1) def testRandom(self): gamma = Gamma(1., 1.) samples = gamma.random(10000) mu = numpy.mean(samples) var = numpy.var(samples) self.assertAlmostEqual(gamma.alpha / gamma.beta, mu, delta=1e-1) self.assertAlmostEqual(gamma.alpha / gamma.beta ** 2, var, delta=1e-1) @test.unit class TestInverseGamma(test.Case): def testParameters(self): pdf = InverseGamma(1, 1) pdf.alpha = 2 pdf.beta = 3 self.assertEqual(pdf.alpha, 2) self.assertEqual(pdf.beta, 3) self.assertEqual(pdf.beta, pdf['beta']) def propertyAssignment(): pdf.beta = -1 def directAssignment(p): pdf[p] = -1 self.assertRaises(ParameterValueError, propertyAssignment) self.assertRaises(ParameterValueError, directAssignment, 'alpha') self.assertRaises(ParameterValueError, directAssignment, 'beta') def testLogProb(self): ig = InverseGamma(1., 1.) self.assertAlmostEqual(ig(1.), numpy.exp(-1.), delta=1e-5) self.assertAlmostEqual(ig(2.), 2 ** -2 * numpy.exp(-0.5), delta=1e-5) ig = InverseGamma(2., 1.) self.assertAlmostEqual(ig(1.), 1. * numpy.exp(-1), delta=1e-5) self.assertAlmostEqual(ig(2.), 2.** -3 * numpy.exp(-0.5), delta=1e-5) @test.unit class TestMultivariateGaussian(test.Case): def testParameters(self): pdf = MultivariateGaussian(numpy.zeros(3), numpy.eye(3)) d = 2 pdf.mu = numpy.zeros(d) pdf.sigma = numpy.eye(d) self.assertEqual(list(pdf.mu), [0, 0]) self.assertEqual(pdf.sigma.tolist(), numpy.eye(d).tolist()) def testLogProb(self): mg = MultivariateGaussian(numpy.zeros(2), numpy.eye(2)) self.assertAlmostEqual(mg(numpy.zeros(2)), 1. / (2 * numpy.pi), delta=1e-1) self.assertAlmostEqual(mg(numpy.ones(2)), 1. / (2 * numpy.pi) * numpy.exp(-0.5), delta=1e-1) def testParameterEstimation(self): d = 3 mu = numpy.ones(d) sigma = numpy.eye(d) pdf = MultivariateGaussian(mu, sigma) samples = pdf.random(100000) pdf.estimate(samples) for i in range(d): self.assertAlmostEqual(pdf.mu[i], mu[i], delta=1e-1) for j in range(d): self.assertAlmostEqual(pdf.sigma[i, j], sigma[i, j], delta=1e-1) d = 3 mu = numpy.array([0., 1., 2.]) sigma = numpy.random.random((d, d)) sigma = numpy.dot(sigma, sigma.T) pdf = MultivariateGaussian(mu, sigma) samples = pdf.random(1000000) pdf.estimate(samples) for i in range(d): self.assertAlmostEqual(pdf.mu[i], mu[i], delta=1e-1) for j in range(d): self.assertAlmostEqual(pdf.sigma[i, j], sigma[i, j], delta=1e-1) @test.unit class TestDirichlet(test.Case): def testParameters(self): pdf = Dirichlet([1, 1]) alpha = numpy.array([2, 2]) pdf.alpha = numpy.array([2, 2]) self.assertEqual(pdf.alpha.tolist(), alpha.tolist()) self.assertEqual(pdf.alpha.tolist(), pdf['alpha'].tolist()) def testLogProb(self): alpha = numpy.array([2, 2]) pdf = Dirichlet(alpha) x = numpy.array([0.5, 0.5]) self.assertAlmostEqual(pdf(x), 1.5, places=2) x = numpy.array([0.1, 0.9]) self.assertAlmostEqual(pdf(x), .54, places=2) x = numpy.array([0.9, 0.1]) self.assertAlmostEqual(pdf(x), .54, places=2) x = numpy.array([[0.9, 0.1], [0.1, 0.9], [0.5, 0.5]]) self.assertAlmostEqual(pdf(x)[0], .54, places=2) self.assertAlmostEqual(pdf(x)[1], .54, places=2) self.assertAlmostEqual(pdf(x)[2], 1.5, places=2) def testParameterEstimation(self): alpha = numpy.array([1., 1., 1.]) pdf = Dirichlet(alpha) self.assertAlmostEqual(pdf(numpy.array([1., 0., 0])), pdf(numpy.array([0., 1., 0])), delta=1e-1) self.assertAlmostEqual(pdf(numpy.array([1. / 3., 1. / 3., 1. / 3])), pdf(numpy.array([0., 1., 0])), delta=1e-1) alpha = numpy.array([16., 16., 16.]) pdf = Dirichlet(alpha) self.assertAlmostEqual(pdf(numpy.array([1., 0., 0])), pdf(numpy.array([0., 1., 0])), delta=1e-1) self.assertAlmostEqual(pdf(numpy.array([1., 0., 0])), pdf(numpy.array([0., 0., 1])), delta=1e-1) self.assertNotEqual(pdf(numpy.array([1. / 3., 1. / 3., 1. / 3])), pdf(numpy.array([0., 1., 0]))) alpha = numpy.array([1., 1., 1.]) pdf = Dirichlet(alpha) samples = pdf.random(100000) pdf.estimate(samples) for i in range(len(alpha)): self.assertAlmostEqual(pdf.alpha[i], alpha[i], delta=1e-1) alpha = numpy.array([200., 60., 20.]) pdf = Dirichlet(alpha) samples = pdf.random(100000) pdf.estimate(samples) for i in range(len(alpha)): self.assertAlmostEqual(pdf.alpha[i], alpha[i], delta=1e1) @test.unit class TestGumbelMinimum(test.Case): @property def klass(self): return GumbelMinimum def testParameters(self): pdf = self.klass(1, 1) pdf.mu = 2 pdf.beta = 3 self.assertEqual(pdf.mu, 2) self.assertEqual(pdf.beta, 3) self.assertEqual(pdf.beta, pdf['beta']) def propertyAssignment(): pdf.beta = -1 def directAssignment(): pdf['beta'] = -1 self.assertRaises(ParameterValueError, propertyAssignment) self.assertRaises(ParameterValueError, directAssignment) def testLogProb(self): pdf = self.klass(0, 2) self.assertAlmostEqual(pdf(0), 1.0 / (pdf.beta * numpy.e)) def testParameterEstimation(self): mu, beta = 4, 2 data = -numpy.random.gumbel(-mu, beta, 1000000) pdf = GumbelMinimum(1, 1) pdf.estimate(data) self.assertAlmostEqual(pdf.mu, mu, delta=0.01) self.assertAlmostEqual(pdf.beta, beta, delta=0.01) def testRandom(self): pdf = self.klass(4, 2) data = pdf.random(100000) beta = numpy.sqrt(6 * numpy.var(data)) / numpy.pi self.assertAlmostEqual(pdf.beta, beta, delta=0.05) @test.unit class TestGumbelMaximum(TestGumbelMinimum): @property def klass(self): return GumbelMaximum def testParameterEstimation(self): mu, beta = 1, 2 data = numpy.random.gumbel(mu, beta, 1000000) pdf = GumbelMaximum(1, 1) pdf.estimate(data) self.assertAlmostEqual(pdf.mu, mu, delta=0.01) self.assertAlmostEqual(pdf.beta, beta, delta=0.01) if __name__ == '__main__': test.Console()