def test_full_adapt_not_invertible():
window = 10
pot = quadpotential.QuadPotentialFullAdapt(
2, np.zeros(2), np.eye(2), 0, adaptation_window=window
)
for i in range(window + 1):
pot.update(np.ones(2), None, True)
with pytest.raises(ValueError):
pot.raise_ok(None)
def test_full_adapt_adaptation_window(seed=8978):
np.random.seed(seed)
window = 10
pot = quadpotential.QuadPotentialFullAdapt(
2, np.zeros(2), np.eye(2), 1, adaptation_window=window
)
for i in range(window + 1):
pot.update(np.random.randn(2), None, True)
assert pot._previous_update == window
assert pot.adaptation_window == window * pot.adaptation_window_multiplier
pot = quadpotential.QuadPotentialFullAdapt(
2, np.zeros(2), np.eye(2), 1, adaptation_window=window
)
for i in range(window + 1):
pot.update(np.random.randn(2), None, True)
assert pot._previous_update == window
assert pot.adaptation_window == window * pot.adaptation_window_multiplier
def test_full_adapt_update_window(seed=1123):
np.random.seed(seed)
init_cov = np.array([[1.0, 0.02], [0.02, 0.8]])
pot = quadpotential.QuadPotentialFullAdapt(2, np.zeros(2), init_cov, 1, update_window=50)
assert np.allclose(pot._cov, init_cov)
for i in range(49):
pot.update(np.random.randn(2), None, True)
assert np.allclose(pot._cov, init_cov)
pot.update(np.random.randn(2), None, True)
assert not np.allclose(pot._cov, init_cov)
def test_full_adapt_sampling(seed=289586):
np.random.seed(seed)
L = np.random.randn(5, 5)
L[np.diag_indices_from(L)] = np.exp(L[np.diag_indices_from(L)])
L[np.triu_indices_from(L, 1)] = 0.0
with pymc3.Model() as model:
pymc3.MvNormal("a", mu=np.zeros(len(L)), chol=L, shape=len(L))
pot = quadpotential.QuadPotentialFullAdapt(model.ndim, np.zeros(model.ndim))
step = pymc3.NUTS(model=model, potential=pot)
pymc3.sample(draws=10, tune=1000, random_seed=seed, step=step, cores=1, chains=1)
def test_full_adapt_sample_p(seed=4566):
# ref: https://github.com/stan-dev/stan/pull/2672
np.random.seed(seed)
m = np.array([[3.0, -2.0], [-2.0, 4.0]])
m_inv = np.linalg.inv(m)
var = np.array([
[2 * m[0, 0], m[1, 0] * m[1, 0] + m[1, 1] * m[0, 0]],
[m[0, 1] * m[0, 1] + m[1, 1] * m[0, 0], 2 * m[1, 1]],
])
n_samples = 1000
pot = quadpotential.QuadPotentialFullAdapt(2, np.zeros(2), m_inv, 1)
samples = [pot.random() for n in range(n_samples)]
sample_cov = np.cov(samples, rowvar=0)
# Covariance matrix within 5 sigma of expected value
# (comes from a Wishart distribution)
assert np.all(np.abs(m - sample_cov) < 5 * np.sqrt(var / n_samples))
def test_full_adapt_sampling(seed=289586):
np.random.seed(seed)
L = np.random.randn(5, 5)
L[np.diag_indices_from(L)] = np.exp(L[np.diag_indices_from(L)])
L[np.triu_indices_from(L, 1)] = 0.0
with pymc3.Model() as model:
pymc3.MvNormal("a", mu=np.zeros(len(L)), chol=L, size=len(L))
initial_point = model.initial_point
initial_point_size = sum(initial_point[n.name].size
for n in model.value_vars)
pot = quadpotential.QuadPotentialFullAdapt(
initial_point_size, np.zeros(initial_point_size))
step = pymc3.NUTS(model=model, potential=pot)
pymc3.sample(draws=10,
tune=1000,
random_seed=seed,
step=step,
cores=1,
chains=1)
def test_full_adapt_warn():
with pytest.warns(UserWarning):
quadpotential.QuadPotentialFullAdapt(2, np.zeros(2), np.eye(2), 0)
