def get_fpr(self, measure, significance):
fpr = numpy.zeros(self.samples)
for sam in range(self.samples):
# data, links = var_process(self.links_coeffs, T=self.T,
# use='inno_cov', verbosity=verbosity)
data = numpy.random.randn(self.T, 2)
# .argsort(axis=0).argsort(axis=0)
res = te._calculate_lag_function(
measure=measure,
data=data,
var_x=0, var_y=1,
conds_x=None, conds_y=None,
measure_params=self.measure_params,
tau_max=1,
selected_lags=[1],
significance=significance,
sig_lev=self.sig_lev,
sig_samples=self.sig_samples,
verbosity=0)
fpr[sam] = res['cmi_pval'][1]
return (fpr <= (1. - self.sig_lev)).mean()
def get_fpr(self, measure, significance):
fpr = numpy.zeros(self.samples)
for sam in range(self.samples):
# data, links = var_process(self.links_coeffs, T=self.T,
# use='inno_cov', verbosity=verbosity)
data = numpy.random.randn(self.T, 2)
# .argsort(axis=0).argsort(axis=0)
res = te._calculate_lag_function(
measure=measure,
data=data,
var_x=0,
var_y=1,
conds_x=None,
conds_y=None,
measure_params=self.measure_params,
tau_max=1,
selected_lags=[1],
significance=significance,
sig_lev=self.sig_lev,
sig_samples=self.sig_samples,
verbosity=0)
fpr[sam] = res['cmi_pval'][1]
return (fpr <= (1. - self.sig_lev)).mean()
def test_measures():
print("\nTesting function '_calculate_lag_function' to check dependence "
"measures are correctly estimated.")
measure_params = {'knn': 10, }
ax = 0.6
ay = 0.3
cxy = .7
links_coeffs = {0: [((0, -1), ax)],
1: [((1, -1), ay), ((0, -1), cxy)],
}
numpy.random.seed(42)
data, links = pp.var_process(links_coeffs, T=5000,
use='inno_cov', verbosity=verbosity)
T, N = data.shape
expected_cmi = 0.5 * numpy.log(1. + (cxy**2 * 1.**2) / (1.**2))
expected_parcorr = cmi2parcorr_trafo(expected_cmi)
gamma_x = 1.**2 / (1. - ax**2)
gamma_xy = (ax * cxy * gamma_x) / (1. - ax * ay)
gamma_y = (cxy**2 * gamma_x + 1.**2 + 2. *
ay * cxy * gamma_xy) / (1. - ay**2)
expected_reg = cxy * numpy.sqrt(gamma_x / gamma_y)
for measure in ['par_corr', 'reg', 'cmi_knn', 'cmi_gauss']:
res = te._calculate_lag_function(
measure=measure,
data=data,
var_x=0, var_y=1,
conds_x=links[0], conds_y=links[1],
measure_params=measure_params,
tau_max=1,
selected_lags=[1],
verbosity=verbosity)['cmi']
if measure == 'par_corr':
print("%s = %.3f (expected = %.3f)"
% (measure, res[1], expected_parcorr))
numpy.testing.assert_allclose(res[1], expected_parcorr, rtol=0.1)
elif measure == 'reg':
print("%s = %.3f (expected = %.3f)"
% (measure, res[1], expected_reg))
numpy.testing.assert_allclose(res[1], expected_reg, rtol=0.1)
elif (measure == 'cmi_knn' or measure == 'cmi_gauss'):
print("%s = %.3f (expected = %.3f)"
% (measure, res[1], expected_cmi))
numpy.testing.assert_allclose(res[1], expected_cmi, rtol=0.1)
# binning estimator
symb_data = pp.quantile_bin_array(data, bins=5)
res = te._calculate_lag_function(
measure='cmi_symb',
data=symb_data,
var_x=0, var_y=1,
conds_x=links[0], conds_y=links[1],
measure_params=measure_params,
tau_max=1,
selected_lags=[1],
verbosity=verbosity)['cmi']
print("%s = %.3f (expected = %.3f)"
% ('symb', res[1], expected_cmi))
numpy.testing.assert_allclose(res[1], expected_cmi, rtol=0.3)
# ordinal pattern estimator
symb_data = pp.ordinal_patt_array(data,
array_mask=numpy.ones(data.shape,
dtype='int32'),
dim=3, step=2)[0]
res = te._calculate_lag_function(
measure='cmi_symb',
data=symb_data,
var_x=0, var_y=1,
conds_x=links[0], conds_y=links[1],
measure_params=measure_params,
tau_max=1,
selected_lags=[1],
verbosity=verbosity)['cmi']
print("%s = %.3f (expected = %.3f)"
% ('ordinal-symb', res[1], expected_cmi))
numpy.testing.assert_allclose(res[1], expected_cmi, rtol=0.4)
def test_measures():
print("\nTesting function '_calculate_lag_function' to check dependence "
"measures are correctly estimated.")
measure_params = {
'knn': 10,
}
ax = 0.6
ay = 0.3
cxy = .7
links_coeffs = {
0: [((0, -1), ax)],
1: [((1, -1), ay), ((0, -1), cxy)],
}
numpy.random.seed(42)
data, links = pp.var_process(links_coeffs,
T=10000,
use='inno_cov',
verbosity=verbosity)
T, N = data.shape
expected_cmi = 0.5 * numpy.log(1. + (cxy**2 * 1.**2) / (1.**2))
expected_parcorr = cmi2parcorr_trafo(expected_cmi)
gamma_x = 1.**2 / (1. - ax**2)
gamma_xy = (ax * cxy * gamma_x) / (1. - ax * ay)
gamma_y = (cxy**2 * gamma_x + 1.**2 + 2. * ay * cxy * gamma_xy) / (1. -
ay**2)
expected_reg = cxy * numpy.sqrt(gamma_x / gamma_y)
for measure in ['par_corr', 'reg', 'cmi_knn', 'cmi_gauss']:
res = te._calculate_lag_function(measure=measure,
data=data,
var_x=0,
var_y=1,
conds_x=links[0],
conds_y=links[1],
measure_params=measure_params,
tau_max=1,
selected_lags=[1],
verbosity=verbosity)['cmi']
if measure == 'par_corr':
print("%s = %.3f (expected = %.3f)" %
(measure, res[1], expected_parcorr))
numpy.testing.assert_allclose(res[1], expected_parcorr, rtol=0.1)
elif measure == 'reg':
print("%s = %.3f (expected = %.3f)" %
(measure, res[1], expected_reg))
numpy.testing.assert_allclose(res[1], expected_reg, rtol=0.1)
elif (measure == 'cmi_knn' or measure == 'cmi_gauss'):
print("%s = %.3f (expected = %.3f)" %
(measure, res[1], expected_cmi))
numpy.testing.assert_allclose(res[1], expected_cmi, rtol=0.1)
# binning estimator
symb_data = pp.quantile_bin_array(data, bins=6)
res = te._calculate_lag_function(measure='cmi_symb',
data=symb_data,
var_x=0,
var_y=1,
conds_x=links[0],
conds_y=links[1],
measure_params=measure_params,
tau_max=1,
selected_lags=[1],
verbosity=verbosity)['cmi']
print("%s = %.3f (expected = %.3f)" % ('symb', res[1], expected_cmi))
numpy.testing.assert_allclose(res[1], expected_cmi, rtol=0.3)
# ordinal pattern estimator
symb_data = pp.ordinal_patt_array(data,
array_mask=numpy.ones(data.shape,
dtype='int32'),
dim=3,
step=2)[0]
res = te._calculate_lag_function(measure='cmi_symb',
data=symb_data,
var_x=0,
var_y=1,
conds_x=links[0],
conds_y=links[1],
measure_params=measure_params,
tau_max=1,
selected_lags=[1],
verbosity=verbosity)['cmi']
print("%s = %.3f (expected = %.3f)" %
('ordinal-symb', res[1], expected_cmi))
numpy.testing.assert_allclose(res[1], expected_cmi, rtol=0.4)
