from __future__ import division, print_function, absolute_import from wafo import stats import numpy as np from numpy.testing import assert_almost_equal, assert_, assert_raises, \ assert_array_almost_equal, assert_array_almost_equal_nulp, run_module_suite def test_kde_1d(): #some basic tests comparing to normal distribution np.random.seed(8765678) n_basesample = 500 xn = np.random.randn(n_basesample) xnmean = xn.mean() xnstd = xn.std(ddof=1) # get kde for original sample gkde = stats.gaussian_kde(xn) # evaluate the density function for the kde for some points xs = np.linspace(-7,7,501) kdepdf = gkde.evaluate(xs) normpdf = stats.norm.pdf(xs, loc=xnmean, scale=xnstd) intervall = xs[1] - xs[0] assert_(np.sum((kdepdf - normpdf)**2)*intervall < 0.01) prob1 = gkde.integrate_box_1d(xnmean, np.inf) prob2 = gkde.integrate_box_1d(-np.inf, xnmean) assert_almost_equal(prob1, 0.5, decimal=1) assert_almost_equal(prob2, 0.5, decimal=1) assert_almost_equal(gkde.integrate_box(xnmean, np.inf), prob1, decimal=13) assert_almost_equal(gkde.integrate_box(-np.inf, xnmean), prob2, decimal=13) assert_almost_equal(gkde.integrate_kde(gkde), (kdepdf**2).sum()*intervall, decimal=2) assert_almost_equal(gkde.integrate_gaussian(xnmean, xnstd**2), (kdepdf*normpdf).sum()*intervall, decimal=2) def test_kde_bandwidth_method(): def scotts_factor(kde_obj): """Same as default, just check that it works.""" return np.power(kde_obj.n, -1./(kde_obj.d+4)) np.random.seed(8765678) n_basesample = 50 xn = np.random.randn(n_basesample) # Default gkde = stats.gaussian_kde(xn) # Supply a callable gkde2 = stats.gaussian_kde(xn, bw_method=scotts_factor) # Supply a scalar gkde3 = stats.gaussian_kde(xn, bw_method=gkde.factor) xs = np.linspace(-7,7,51) kdepdf = gkde.evaluate(xs) kdepdf2 = gkde2.evaluate(xs) assert_almost_equal(kdepdf, kdepdf2) kdepdf3 = gkde3.evaluate(xs) assert_almost_equal(kdepdf, kdepdf3) assert_raises(ValueError, stats.gaussian_kde, xn, bw_method='wrongstring') # Subclasses that should stay working (extracted from various sources). # Unfortunately the earlier design of gaussian_kde made it necessary for users # to create these kinds of subclasses, or call _compute_covariance() directly. class _kde_subclass1(stats.gaussian_kde): def __init__(self, dataset): self.dataset = np.atleast_2d(dataset) self.d, self.n = self.dataset.shape self.covariance_factor = self.scotts_factor self._compute_covariance() class _kde_subclass2(stats.gaussian_kde): def __init__(self, dataset): self.covariance_factor = self.scotts_factor super(_kde_subclass2, self).__init__(dataset) class _kde_subclass3(stats.gaussian_kde): def __init__(self, dataset, covariance): self.covariance = covariance stats.gaussian_kde.__init__(self, dataset) def _compute_covariance(self): self.inv_cov = np.linalg.inv(self.covariance) self._norm_factor = np.sqrt(np.linalg.det(2*np.pi * self.covariance)) \ * self.n class _kde_subclass4(stats.gaussian_kde): def covariance_factor(self): return 0.5 * self.silverman_factor() def test_gaussian_kde_subclassing(): x1 = np.array([-7, -5, 1, 4, 5], dtype=np.float) xs = np.linspace(-10, 10, num=50) # gaussian_kde itself kde = stats.gaussian_kde(x1) ys = kde(xs) # subclass 1 kde1 = _kde_subclass1(x1) y1 = kde1(xs) assert_array_almost_equal_nulp(ys, y1, nulp=10) # subclass 2 kde2 = _kde_subclass2(x1) y2 = kde2(xs) assert_array_almost_equal_nulp(ys, y2, nulp=10) # subclass 3 kde3 = _kde_subclass3(x1, kde.covariance) y3 = kde3(xs) assert_array_almost_equal_nulp(ys, y3, nulp=10) # subclass 4 kde4 = _kde_subclass4(x1) y4 = kde4(x1) y_expected = [0.06292987, 0.06346938, 0.05860291, 0.08657652, 0.07904017] assert_array_almost_equal(y_expected, y4, decimal=6) # Not a subclass, but check for use of _compute_covariance() kde5 = kde kde5.covariance_factor = lambda: kde.factor kde5._compute_covariance() y5 = kde5(xs) assert_array_almost_equal_nulp(ys, y5, nulp=10) def test_gaussian_kde_covariance_caching(): x1 = np.array([-7, -5, 1, 4, 5], dtype=np.float) xs = np.linspace(-10, 10, num=5) # These expected values are from scipy 0.10, before some changes to # gaussian_kde. They were not compared with any external reference. y_expected = [0.02463386, 0.04689208, 0.05395444, 0.05337754, 0.01664475] # Set the bandwidth, then reset it to the default. kde = stats.gaussian_kde(x1) kde.set_bandwidth(bw_method=0.5) kde.set_bandwidth(bw_method='scott') y2 = kde(xs) assert_array_almost_equal(y_expected, y2, decimal=7) def test_gaussian_kde_monkeypatch(): """Ugly, but people may rely on this. See scipy pull request 123, specifically the linked ML thread "Width of the Gaussian in stats.kde". If it is necessary to break this later on, that is to be discussed on ML. """ x1 = np.array([-7, -5, 1, 4, 5], dtype=np.float) xs = np.linspace(-10, 10, num=50) # The old monkeypatched version to get at Silverman's Rule. kde = stats.gaussian_kde(x1) kde.covariance_factor = kde.silverman_factor kde._compute_covariance() y1 = kde(xs) # The new saner version. kde2 = stats.gaussian_kde(x1, bw_method='silverman') y2 = kde2(xs) assert_array_almost_equal_nulp(y1, y2, nulp=10) def test_kde_integer_input(): """Regression test for #1181.""" x1 = np.arange(5) kde = stats.gaussian_kde(x1) y_expected = [0.13480721, 0.18222869, 0.19514935, 0.18222869, 0.13480721] assert_array_almost_equal(kde(x1), y_expected, decimal=6) def test_pdf_logpdf(): np.random.seed(1) n_basesample = 50 xn = np.random.randn(n_basesample) # Default gkde = stats.gaussian_kde(xn) xs = np.linspace(-15, 12, 25) pdf = gkde.evaluate(xs) pdf2 = gkde.pdf(xs) assert_almost_equal(pdf, pdf2, decimal=12) logpdf = np.log(pdf) logpdf2 = gkde.logpdf(xs) assert_almost_equal(logpdf, logpdf2, decimal=12) if __name__ == "__main__": run_module_suite()