def test_gaussian():
np.random.seed(88)
n_dims = 6
random_matrix = np.random.rand(n_dims, n_dims)
sigma = random_matrix.dot(random_matrix.T)
mu = np.random.randn(n_dims)
dist = GaussianDistribution(mu, sigma)
for i in range(20):
theta = dist.sample()
exact_diff = dist.diff(theta)
numerical_diff = numerical_diff_dist(dist, theta)
assert np.allclose(exact_diff, numerical_diff)
theta = dist.sample()
log_p = dist.log_pdf(theta)
log_p_test = -5.622792079250801
assert np.isclose(log_p, log_p_test)
theta = np.random.randn(100, n_dims)
weights = np.random.rand(100)
dist.mle(theta)
assert np.array_equal(dist.get_parameters(), theta.mean(axis=0))
dist.mle(theta, weights)
assert np.array_equal(dist.get_parameters(),
weights.dot(theta) / np.sum(weights))
def test_cholesky_gaussian():
np.random.seed(88)
n_dims = 6
random_matrix = np.random.rand(n_dims, n_dims)
sigma = random_matrix.dot(random_matrix.T)
mu = np.random.randn(n_dims)
dist = GaussianCholeskyDistribution(mu, sigma)
for i in range(20):
theta = dist.sample()
exact_diff = dist.diff(theta)
numerical_diff = numerical_diff_dist(dist, theta)
assert np.allclose(exact_diff, numerical_diff)
theta = dist.sample()
log_p = dist.log_pdf(theta)
log_p_test = -5.622792079250836
assert np.isclose(log_p, log_p_test)
theta = np.random.randn(100, n_dims)
weights = np.random.rand(100)
dist.mle(theta)
mle_test = np.array([
0.03898376, 0.07252868, 0.26070238, 0.13782173, 0.18927999,
-0.02548812, 1.00855691, 0.19697324, 1.06381216, -0.07439629,
0.0656251, 1.02907183, 0.03779866, -0.00504703, -0.14902275,
0.99917335, -0.09132656, -0.03225904, -0.13589437, 0.1419549,
0.94040997, -0.00145945, 0.00326904, 0.00291136, -0.07456335,
0.04581934, 1.02750578
])
assert np.allclose(dist.get_parameters(), mle_test)
dist.mle(theta, weights)
wmle_test = np.array([
-0.07797052, 0.08518447, 0.36272218, 0.17409145, 0.26339453,
-0.02891896, 0.98529941, 0.26728657, 1.09177517, -0.03838698,
-0.08395759, 0.98168805, 0.0150622, 0.05611417, -0.09351055, 1.0166716,
-0.06390746, -0.05409177, -0.08944413, 0.17745539, 1.01277413,
0.00923361, 0.05694206, -0.02457328, -0.14141649, 0.1117947, 1.03121418
])
assert np.allclose(dist.get_parameters(), wmle_test)
entropy = dist.entropy()
assert np.isclose(entropy, 8.628109062120682)
def test_diagonal_gaussian():
np.random.seed(88)
n_dims = 6
std = np.abs(np.random.rand(n_dims))
mu = np.random.randn(n_dims)
dist = GaussianDiagonalDistribution(mu, std)
for i in range(20):
theta = dist.sample()
exact_diff = dist.diff(theta)
numerical_diff = numerical_diff_dist(dist, theta)
assert np.allclose(exact_diff, numerical_diff)
theta = dist.sample()
log_p = dist.log_pdf(theta)
log_p_test = -7.818947754486631
assert np.isclose(log_p, log_p_test)
theta = np.random.randn(100, n_dims)
weights = np.random.rand(100)
dist.mle(theta)
assert np.array_equal(dist.get_parameters()[:n_dims], theta.mean(axis=0))
assert np.array_equal(dist.get_parameters()[n_dims:], theta.std(axis=0))
dist.mle(theta, weights)
wmle_test = np.array([
0.14420612, -0.02660736, 0.07439633, -0.00424596, 0.2495252,
0.20968329, 0.97527594, 1.08187144, 1.04907019, 1.0634484, 1.03453275,
1.03961039
])
assert np.allclose(dist.get_parameters(), wmle_test)
entropy = dist.entropy()
assert np.isclose(entropy, 8.749505679983452)