def test_other_properties(self):
with self.get_session():
mean, stddev = self.get_mean_stddev(**self.simple_params)
# test the parameters of the distribution
dist = Normal(**self.simple_params)
np.testing.assert_allclose(dist.mean.eval(), mean)
np.testing.assert_allclose(dist.stddev.eval(), stddev)
np.testing.assert_allclose(dist.logstd.eval(), np.log(stddev))
np.testing.assert_allclose(dist.var.eval(), np.square(stddev))
np.testing.assert_allclose(dist.logvar.eval(), 2. * np.log(stddev))
np.testing.assert_allclose(dist.precision.eval(),
1. / np.square(stddev))
np.testing.assert_allclose(dist.log_precision.eval(),
-2. * np.log(stddev))
# test the parameters of the distribution when logstd is specified
dist = Normal(mean, logstd=np.log(stddev))
np.testing.assert_allclose(dist.mean.eval(), mean)
np.testing.assert_allclose(dist.stddev.eval(), stddev)
np.testing.assert_allclose(dist.logstd.eval(), np.log(stddev))
np.testing.assert_allclose(dist.var.eval(), np.square(stddev))
np.testing.assert_allclose(dist.logvar.eval(), 2. * np.log(stddev))
np.testing.assert_allclose(dist.precision.eval(),
1. / np.square(stddev))
np.testing.assert_allclose(dist.log_precision.eval(),
-2. * np.log(stddev))
def test_construction_error(self):
with self.get_session():
# test construction due to no std specified
with self.assertRaisesRegex(
ValueError,
'One and only one of `stddev`, `logstd` should '
'be specified.'):
Normal(1.)
with self.assertRaisesRegex(
ValueError,
'One and only one of `stddev`, `logstd` should '
'be specified.'):
Normal(1., 2., 3.)
# test construction due to data type error
with self.assertRaisesRegex(
TypeError, 'Normal distribution parameters must be float '
'numbers'):
Normal(1, 2)
# test construction error due to shape mismatch
with self.assertRaisesRegex(
ValueError, '`mean` and `stddev`/`logstd` should be '
'broadcastable'):
Normal(np.arange(2, dtype=np.float32),
np.arange(3, dtype=np.float32))
def test_sgvb(self):
with self.get_session():
a = StochasticTensor(Normal(0., 1.),
observed=np.asarray([[0., 1., 2.]]))
b = StochasticTensor(Normal(1., 2.),
observed=np.asarray([[1., 2., 3.]]))
c = StochasticTensor(Normal(2., 3.),
observed=np.asarray([[2., 3., 4.]]))
lower_bound = sgvb([a, b], [c])
self.assertEqual(lower_bound.get_shape().as_list(), [1, 3])
np.testing.assert_almost_equal(lower_bound.eval(),
(a.log_prob() + b.log_prob() -
c.log_prob()).eval())
lower_bound = sgvb([a], [b, c], latent_axis=0)
self.assertEqual(lower_bound.get_shape().as_list(), [3])
np.testing.assert_almost_equal(lower_bound.eval(),
(a.log_prob() - b.log_prob() -
c.log_prob()).eval().reshape([3]))
lower_bound = sgvb([a], [b, c], latent_axis=[0, 1])
self.assertEqual(lower_bound.get_shape().as_list(), [])
np.testing.assert_almost_equal(
lower_bound.eval(),
np.mean((a.log_prob() - b.log_prob() - c.log_prob()).eval()))
def test_prob_and_log_prob(self):
with self.get_session():
distrib = Normal(np.asarray(0., dtype=np.float32),
np.asarray([1.0, 2.0, 3.0], dtype=np.float32))
observed = np.arange(24, dtype=np.float32).reshape([4, 2, 3])
t = StochasticTensor(distrib, observed=observed)
np.testing.assert_almost_equal(t.log_prob().eval(),
distrib.log_prob(observed).eval())
np.testing.assert_almost_equal(t.prob().eval(),
distrib.prob(observed).eval())
def test_prob_with_group_events_ndims(self):
with self.get_session():
distrib = Normal(np.asarray(0., dtype=np.float32),
np.asarray([1.0, 2.0, 3.0], dtype=np.float32),
group_event_ndims=1)
observed = np.asarray([[-1., 1., 2.], [0., 0., 0.]])
t = StochasticTensor(distrib, observed=observed)
np.testing.assert_allclose(
t.prob(group_event_ndims=0).eval(),
distrib.prob(t, group_event_ndims=0).eval())
np.testing.assert_allclose(
t.prob(group_event_ndims=1).eval(),
distrib.prob(t, group_event_ndims=1).eval())
def test_attributes_from_distribution(self):
with self.get_session():
distrib = Normal(0., 1.)
t = StochasticTensor(distrib, tf.constant(0.))
for k in ['is_continuous', 'is_reparameterized']:
self.assertEqual(getattr(distrib, k), getattr(t, k),
msg='attribute %r mismatch.' % k)
def test_session_run_issue_49(self):
# test fix for the bug at https://github.com/thu-ml/zhusuan/issues/49
x_mean = tf.zeros([1, 2])
x_logstd = tf.zeros([1, 2])
x = Normal(mean=x_mean, logstd=x_logstd).sample()
with self.get_session() as sess:
sess.run(tf.global_variables_initializer())
_ = sess.run(x)
def test_convert_to_tensor_if_dynamic(self):
for v in [
tf.placeholder(tf.int32, ()),
tf.get_variable('v', shape=(), dtype=tf.int32),
StochasticTensor(Normal(0., 1.), 1.)
]:
self.assertIsInstance(convert_to_tensor_if_dynamic(v), tf.Tensor)
for v in [1, 1.0, object(), (), [], {}, np.array([1, 2, 3])]:
self.assertIs(convert_to_tensor_if_dynamic(v), v)
def test_basic(self):
with self.get_session() as sess:
a = Normal(0., np.asarray([1., 2., 3.]),
group_event_ndims=1).sample_n(16)
b = Normal(1., 2.).sample_n(16)
a_prob, b_prob = gather_log_lower_bound([a, b])
self.assertIsInstance(a_prob, tf.Tensor)
self.assertEqual(a_prob.get_shape().as_list(), [16])
res, ans = sess.run(
[a_prob,
a.distribution.log_prob(a, group_event_ndims=1)])
np.testing.assert_almost_equal(res, ans)
self.assertIsInstance(b_prob, tf.Tensor)
self.assertEqual(b_prob.get_shape().as_list(), [16])
res, ans = sess.run(
[b_prob,
b.distribution.log_prob(b, group_event_ndims=0)])
np.testing.assert_almost_equal(res, ans)
def test_is_dynamic_tensor_like(self):
for v in [
tf.placeholder(tf.int32, ()),
tf.get_variable('v', shape=(), dtype=tf.int32),
StochasticTensor(Normal(0., 1.), 1.)
]:
self.assertTrue(
is_dynamic_tensor_like(v),
msg='%r should be interpreted as a dynamic tensor.' % (v, ))
for v in [1, 1.0, object(), (), [], {}, np.array([1, 2, 3])]:
self.assertFalse(
is_dynamic_tensor_like(v),
msg='%r should not be interpreted as a dynamic tensor.' %
(v, ))
def test_integrated(self):
with self.get_session(use_gpu=True) as sess:
tf.set_random_seed(1234)
p = Normal(0., 1.)
q = Normal(1., 2.)
x = q.sample_n(10000)
log_p = p.log_prob(x)
log_q = q.log_prob(x)
# test integrated equality
output = sess.run(
importance_sampling(log_p, log_p, log_q, latent_axis=0))
self.assertLess(np.abs(output + 0.5 * (np.log(2 * np.pi) + 1)),
1e-2)
def test_element_wise(self):
with self.get_session(use_gpu=True) as sess:
tf.set_random_seed(1234)
p = Normal(0., 1.)
q = Normal(1., 2.)
x = q.sample_n(1000)
log_p = p.log_prob(x)
log_q = q.log_prob(x)
# test element-wise equality
output, expected = sess.run([
importance_sampling(log_p, log_p, log_q),
log_p * tf.exp(log_p - log_q)
])
np.testing.assert_almost_equal(output, expected)