def test_no_conditioning(self, joint_distribution):
model = joint_distribution([
tfd.Normal(1., 1.),
tfd.Normal(3., 2.), lambda b, a: tfd.Normal(a + b, 5.)
])
approx = rethinking.quap(model)
self.assertAllCloseNested(approx.mean(), [1., 3., 4.])
self.assertAllCloseNested(approx.stddev(), [1., 2., 30**0.5])
def test_no_conditioning(self, joint_distribution):
model = joint_distribution({
'a': tfd.Normal(1., 1.),
'b': tfd.Normal(3., 2.),
'c': lambda a, b: tfd.Normal(a + b, 5.)
})
approx = rethinking.quap(model)
self.assertAllCloseNested(approx.mean(), dict(a=1., b=3., c=4.))
self.assertAllCloseNested(approx.stddev(), dict(a=1., b=2., c=30**0.5))
def test_no_conditioning(self, joint_distribution):
ModelSpec = collections.namedtuple('ModelSpec', 'a,b,c')
model = joint_distribution(
ModelSpec(a=tfd.Normal(1., 1.),
b=tfd.Normal(3., 2.),
c=lambda a, b: tfd.Normal(a + b, 5.)))
approx = rethinking.quap(model)
self.assertAllCloseNested(approx.mean(), ModelSpec(a=1., b=3., c=4.))
self.assertAllCloseNested(approx.stddev(),
ModelSpec(a=1., b=2., c=30**0.5))
def test_conditioning(self, joint_distribution):
model = joint_distribution(
[tfd.Normal(1., 2.), lambda a: tfd.Normal(2 * a + 1., 5.)])
approx = rethinking.quap(model,
data=[None, 2.],
initial_position=[1.5])
dists, _ = approx.sample_distributions()
# See, e.g., Bishop, equation 2.116
self.assertAllClose(dists[0].mean(), 33. / 41.)
self.assertAllClose(dists[0].stddev(), (100. / 41.)**0.5)
def test_conditioning(self, joint_distribution):
model = joint_distribution({
'a': tfd.Normal(1., 2.),
'b': lambda a: tfd.Normal(2 * a + 1., 5.)})
approx = rethinking.quap(model,
data={'b': 2.},
initial_position={'a': 1.5})
dists, _ = approx.sample_distributions()
# See, e.g., Bishop, equation 2.116
self.assertAllClose(dists['a'].mean(), 33. / 41.)
self.assertAllClose(dists['a'].stddev(), (100. / 41.) ** 0.5)
def test_conditioning(self, joint_distribution):
ModelSpec = collections.namedtuple('ModelSpec', 'a,b')
model = joint_distribution(ModelSpec(
a=tfd.Normal(1., 2.),
b=lambda a: tfd.Normal(2 * a + 1., 5.)))
approx = rethinking.quap(model,
data=ModelSpec(a=None, b=2.),
initial_position=ModelSpec(a=1.5, b=None))
dists, _ = approx.sample_distributions()
# See, e.g., Bishop, equation 2.116
self.assertAllClose(dists.a.mean(), 33. / 41.)
self.assertAllClose(dists.a.stddev(), (100. / 41.) ** 0.5)
