def test_noise_dbns():
X = np.random.standard_normal((10, 5))
Q = X.T.dot(X)
noises = [randomization.isotropic_gaussian((5,), 1.),
randomization.laplace((5,), 1.),
randomization.logistic((5,), 1.),
randomization.gaussian(Q)]
for i, noise in enumerate(noises):
x = np.random.standard_normal(5)
u = np.random.standard_normal(5)
noise.log_density(x)
np.testing.assert_allclose(np.exp(noise.log_density(x)), noise._density(x))
noise.smooth_objective(x, 'func')
noise.smooth_objective(x, 'grad')
noise.smooth_objective(x, 'both')
noise.gradient(x)
S = noise.sample()
if i != 3:
np.testing.assert_allclose(float(noise.log_density(S)), float(np.log(noise._density(S)).sum()))
nt.assert_equal(noise.sample().shape, (5,))
nt.assert_equal(noise.sample().shape, (5,))
if hasattr(noise, "CGF"):
val, grad = noise.CGF
u = np.zeros(5)
u[:2] = 0.1
val(u), grad(u)
if hasattr(noise, "CGF_grad"):
val, grad = noise.CGF_grad
val(x), grad(x)
def test_noise_dbns():
X = np.random.standard_normal((10, 5))
Q = X.T.dot(X)
noises = [
randomization.isotropic_gaussian((5, ), 1.),
randomization.laplace((5, ), 1.),
randomization.logistic((5, ), 1.),
randomization.gaussian(Q)
]
v1, v2 = [], []
for i, noise in enumerate(noises):
x = np.random.standard_normal(5)
u = np.random.standard_normal(5)
v1.append(np.exp(noise.log_density(x)))
v2.append(noise._density(x))
noise.smooth_objective(x, 'func')
noise.smooth_objective(x, 'grad')
noise.smooth_objective(x, 'both')
noise.gradient(x)
nt.assert_equal(noise.sample().shape, (5, ))
nt.assert_equal(noise.sample().shape, (5, ))
if noise.CGF is not None:
u = np.zeros(5)
u[:2] = 0.1
noise.CGF.smooth_objective(u, 'both')
if noise.CGF_conjugate is not None:
noise.CGF_conjugate.smooth_objective(x, 'both')
def __init__(self, X, Y, l_theory, l_min, l_1se, sigma_reid):
randomized_lasso.__init__(self, X, Y, l_theory, l_min, l_1se,
sigma_reid)
n, p = X.shape
ARrho = []
for s in np.random.sample(100):
Xr = X[int(s * n)]
ARrho.append(np.corrcoef(Xr[1:], Xr[:-1])[0, 1])
ARrho = np.mean(ARrho)
print("AR parameter", ARrho)
mean_diag = np.mean((X**2).sum(0))
randomizer_scale = np.sqrt(mean_diag) * np.std(
Y) * self.randomizer_scale
ARcov = ARrho**(np.abs(np.subtract.outer(
np.arange(p), np.arange(p)))) * randomizer_scale**2
self._randomizer = randomization.gaussian(ARcov)