def test_running_urbk(self):
prior_rv = normal.med(mu=0.5, sigma=1.0)
n = normal.fast_p
prior = jit.jit_probability(prior_rv)
@numba.jit(fastmath=True, nopython=True, forceobj=False)
def likelihood(y, w):
return n(y - w, mu=0.0, sigma=1.0)
data = normal.sample(mu=3.0, sigma=1.0, size=100)
log_likelihood, log_prior = jit_log_probabilities((data,), likelihood, prior)
samples, densities = search_posterior_estimation(
size=300, log_likelihood=log_likelihood,
log_prior=log_prior,
initial=prior_rv.sample(size=10),
energy=0.1,
volume=100
)
density = URBK(variance=5.0, verbose=True)
density.fit(samples, densities)
lb, ub = -6, 6
n = 2000
x = np.linspace(lb, ub, n)
y = density.p(x)
self.assertEqual(y.size, 2000)
fast_p = density.get_fast_p()
fast_p(x) # is slower than normal p.. but numba need numba functions
def _execute_test(distribution=None, parameters=None, names=None):
for test_case in itertools.product([0, 1], repeat=len(parameters)):
kwargs, args = {}, []
for i, pick in enumerate(test_case):
if pick == 1:
kwargs[names[i]] = parameters[i]
else:
args.append(parameters[i])
rv = distribution.med(**kwargs)
samples = rv.sample(*args, size=10)
fast_p = jit.jit_probability(rv)
fast_p(samples, *args)
def test_parameter_posterior_mcmc(self):
prior_rv = normal.med(mu=0.5, sigma=1.0)
n = normal.fast_p
prior = jit.jit_probability(prior_rv)
@numba.jit(fastmath=True, nopython=True, forceobj=False)
def likelihood(y, w):
return n(y - w, mu=0.0, sigma=1.0)
data = normal.sample(mu=3.0, sigma=1.0, size=100)
log_likelihood, log_prior = jit_log_probabilities((data, ), likelihood,
prior)
result = fast_metropolis_hastings_log_space_parameter_posterior_estimation(
size=2000,
log_likelihood=log_likelihood,
log_prior=log_prior,
initial=prior_rv.sample(size=10),
energy=0.1)
self.assertAlmostEqual(result.mean(), 3.0, delta=1.0)
def test_parameter_posterior_search(self):
prior_rv = normal.med(mu=0.5, sigma=1.0)
n = normal.fast_p
prior = jit.jit_probability(prior_rv)
@numba.jit(fastmath=True, nopython=True, forceobj=False)
def likelihood(y, w):
return n(y - w, mu=0.0, sigma=1.0)
data = normal.sample(mu=3.0, sigma=1.0, size=100)
log_likelihood, log_prior = jit_log_probabilities((data, ), likelihood,
prior)
points, densities = search_posterior_estimation(
size=1000,
log_likelihood=log_likelihood,
log_prior=log_prior,
initial=prior_rv.sample(size=10),
energy=0.1,
volume=1000)
self.assertTrue(True)
def jitted_likelihood(likelihood: Union[RandomVariable, Callable[[np.ndarray],
np.ndarray]]):
if type(likelihood) == RandomVariable:
return jit_probability(likelihood)
return numba.jit(nopython=True, fastmath=True, forceobj=False)(likelihood)
def jitted_prior(rv: RandomVariable):
return jit_probability(rv)