def test_laplace(self):
TRUE_MEAN = -10.0
TRUE_SCALE = 0.07
sampler = NUTS()
X = T.scalar('X')
output = -1.0 * (abs(X - TRUE_MEAN) / TRUE_SCALE)
f = function([X], output, allow_input_downcast=True)
grad_f = function([X],
grad(output, X),
allow_input_downcast=True)
E = energy.Energy(eval=f,
gradient=grad_f)
samples = sampler.nuts_with_initial_epsilon(0.0, E, iterations=10000, burn_in=100)
plt.hist(samples, bins=100)
plt.show()
def test_normal(self):
TRUE_MEAN = -10.0
TRUE_SIGMA = 3.7
sampler = NUTS()
X = T.scalar('X')
output = -0.5 * ((X - TRUE_MEAN)**2) * (TRUE_SIGMA ** (-2))
f = function([X], output, allow_input_downcast=True)
grad_f = function([X],
grad(output, X),
allow_input_downcast=True)
E = energy.Energy(eval=f,
gradient=grad_f)
samples = sampler.nuts_with_initial_epsilon(0.0, E, iterations=10000, burn_in=100)
w,p = shapiro(samples)
self.assertGreater(p, 0.01)