def test_props(self):
with self.test_session() as sess:
mean = np.random.normal(size=[2, 1, 4]).astype(np.float32)
log_scale = np.random.normal(size=[3, 1, 5, 1]).astype(np.float32)
d = DiscretizedLogistic(
mean=mean,
log_scale=log_scale,
bin_size=1. / 255,
min_val=-1.,
max_val=2.,
biased_edges=False
)
self.assertEqual(d.dtype, tf.float32)
self.assertFalse(d.is_continuous)
self.assertFalse(d.is_reparameterized)
self.assertEqual(d.value_ndims, 0)
self.assertEqual(d.get_batch_shape().as_list(), [3, 2, 5, 4])
self.assertEqual(list(sess.run(d.batch_shape)), [3, 2, 5, 4])
self.assertEqual(d.bin_size.dtype, tf.float32)
self.assertFalse(d.biased_edges)
np.testing.assert_allclose(sess.run(d.mean), mean)
np.testing.assert_allclose(sess.run(d.log_scale), log_scale)
np.testing.assert_allclose(sess.run(d.bin_size), 1. / 255)
np.testing.assert_allclose(sess.run(d.min_val), -1.)
np.testing.assert_allclose(sess.run(d.max_val), 2.)
d = DiscretizedLogistic(mean=mean, log_scale=log_scale, bin_size=1.,
dtype=tf.int32)
self.assertEqual(d.dtype, tf.int32)
self.assertIsNone(d.min_val)
self.assertIsNone(d.max_val)
self.assertTrue(d.biased_edges)
def test_errors(self):
with pytest.raises(ValueError,
match='`bin_size` is a float number, but `dtype` '
'is not a float number type'):
_ = DiscretizedLogistic(tf.zeros([]), tf.zeros([]), 0.1,
dtype=tf.int32)
with pytest.raises(ValueError,
match='`min_val` and `max_val` must be both None '
'or neither None.'):
_ = DiscretizedLogistic(tf.zeros([2, 3]), 0., bin_size=.1,
min_val=-1.)
with pytest.raises(ValueError,
match='`min_val` and `max_val` must be both None '
'or neither None.'):
_ = DiscretizedLogistic(tf.zeros([2, 3]), 0., bin_size=.1,
max_val=1.)
with pytest.raises(ValueError,
match='`min_val - max_val` must be multiples of '
'`bin_size`: max_val - min_val = 1.5 vs '
'bin_size = 1.'):
_ = DiscretizedLogistic(tf.zeros([2, 3]), 0., bin_size=1.0,
min_val=-0.5, max_val=1.)
with pytest.raises(ValueError,
match='The shape of `mean` and `log_scale` cannot '
'be broadcasted'):
_ = DiscretizedLogistic(tf.zeros([2]), tf.zeros([3]), 0.1)
def test_log_prob_extreme(self):
assert_allclose = functools.partial(
np.testing.assert_allclose, rtol=1e-5, atol=1e-6)
x = np.asarray(0., dtype=np.float64)
mean = np.asarray(0., dtype=np.float64)
bin_size = 1 / 256.
# to ensure bin_size / (2*scale) < cdf_delta
cdf_delta = 1e-8
s = np.linspace(0, 20, 101)
t = bin_size / (2 * np.exp(s))
idx = np.where(safe_sigmoid(t) - safe_sigmoid(-t) < cdf_delta)[0]
self.assertGreater(np.size(idx), 2)
log_scale = s[idx]
with self.test_session() as sess:
# now compute the log-probability of this extreme case
d = DiscretizedLogistic(
mean=mean, log_scale=log_scale, bin_size=bin_size,
epsilon=1e-7
)
assert_allclose(
sess.run(d.log_prob(x, group_ndims=0)),
naive_discretized_logistic_pdf(
x, mean, log_scale, bin_size, None, None,
biased_edges=True, group_ndims=0)
)
def test_errors(self):
with pytest.raises(ValueError,
match='`bin_size` is a float number, but `dtype` '
'is not a float number type'):
_ = DiscretizedLogistic(tf.zeros([]), tf.zeros([]), 0.1,
dtype=tf.int32)
with pytest.raises(ValueError,
match='`min_val` must be multiples of `bin_size`'):
_ = DiscretizedLogistic(tf.zeros([]), tf.zeros([]), 0.15,
min_val=-1)
with pytest.raises(ValueError,
match='`max_val` must be multiples of `bin_size`'):
_ = DiscretizedLogistic(tf.zeros([]), tf.zeros([]), 0.15,
max_val=1)
with pytest.raises(ValueError,
match='The shape of `mean` and `log_scale` cannot '
'be broadcasted'):
_ = DiscretizedLogistic(tf.zeros([2]), tf.zeros([3]), 0.1)
def test_prob_sum(self):
assert_allclose = functools.partial(
np.testing.assert_allclose, rtol=1e-5, atol=1e-6)
np.random.seed(1234)
x = np.reshape(np.linspace(0, 1, 11), [-1, 1, 1, 1, 1])
mean = np.random.uniform(size=[2, 1, 4]).astype(np.float64)
log_scale = np.random.normal(size=[3, 1, 5, 1]).astype(np.float64)
bin_size = 1 / 10.
min_val = 0.
max_val = 1.
with self.test_session() as sess:
d = DiscretizedLogistic(
mean=mean, log_scale=log_scale, bin_size=bin_size,
min_val=min_val, max_val=max_val, biased_edges=True,
dtype=tf.float64
)
prob = sess.run(tf.reduce_sum(d.prob(x), 0))
self.assertEqual(prob.shape, (3, 2, 5, 4))
assert_allclose(prob, np.ones_like(prob))
def p_net(observed=None,
n_z=None,
beta=1.0,
mcmc_iterator=0,
log_Z=0.0,
initial_z=None):
net = spt.BayesianNet(observed=observed)
# sample z ~ p(z)
normal = spt.Normal(mean=tf.zeros([1, config.z_dim]),
logstd=tf.zeros([1, config.z_dim]))
normal = normal.batch_ndims_to_value(1)
xi = tf.get_variable(name='xi',
shape=(),
initializer=tf.constant_initializer(
config.initial_xi),
dtype=tf.float32,
trainable=True)
# xi = tf.square(xi)
xi = tf.nn.sigmoid(xi) # TODO
pz = EnergyDistribution(normal,
G=G_theta,
D=D_psi,
log_Z=log_Z,
xi=xi,
mcmc_iterator=mcmc_iterator,
initial_z=initial_z)
z = net.add('z', pz, n_samples=n_z)
x_mean, x_logstd = G_theta(z, return_std=True)
x = net.add('x',
DiscretizedLogistic(
mean=x_mean,
log_scale=spt.ops.maybe_clip_value(
beta if config.uniform_scale else x_logstd,
min_val=config.epsilon),
bin_size=2.0 / 256.0,
min_val=-1.0 + 1.0 / 256.0,
max_val=1.0 - 1.0 / 256.0,
epsilon=1e-10),
n_samples=config.test_x_samples,
group_ndims=3)
return net
def test_log_prob(self):
assert_allclose = functools.partial(
np.testing.assert_allclose, rtol=1e-5, atol=1e-6)
np.random.seed(1234)
x = np.random.normal(size=[7, 3, 2, 5, 4]).astype(np.float64)
self.assertLess(np.min(x), -1.1)
self.assertGreater(np.max(x), 2.1)
self.assertGreater(np.sum(np.logical_and(x > -0.5, x < 1.5)), 0)
mean = 3 * np.random.uniform(size=[2, 1, 4]).astype(np.float64) - 1
log_scale = np.random.normal(size=[3, 1, 5, 1]).astype(np.float64)
bin_size = 1 / 256.
min_val = -1.
max_val = 2.
with self.test_session() as sess:
# biased_edges = False, discretize_given = True
d = DiscretizedLogistic(
mean=mean, log_scale=log_scale, bin_size=bin_size,
min_val=None, max_val=None, biased_edges=False,
dtype=tf.float64
)
assert_allclose(
sess.run(d.log_prob(x, group_ndims=0)),
naive_discretized_logistic_pdf(
x, mean, log_scale, bin_size, None, None,
biased_edges=False, group_ndims=0)
)
# biased_edges = False, discretize_given = True
d = DiscretizedLogistic(
mean=mean, log_scale=log_scale, bin_size=bin_size,
min_val=min_val, max_val=max_val, biased_edges=False,
dtype=tf.float64
)
assert_allclose(
sess.run(d.log_prob(x, group_ndims=0)),
naive_discretized_logistic_pdf(
x, mean, log_scale, bin_size, min_val, max_val,
biased_edges=False, group_ndims=0)
)
# biased_edges = False, discretize_given = False
d = DiscretizedLogistic(
mean=mean, log_scale=log_scale, bin_size=bin_size,
min_val=min_val, max_val=max_val, biased_edges=False,
dtype=tf.float64, discretize_given=False
)
assert_allclose(
sess.run(d.log_prob(x, group_ndims=0)),
naive_discretized_logistic_pdf(
x, mean, log_scale, bin_size, min_val, max_val,
biased_edges=False, discretize_given=False, group_ndims=0)
)
# biased_edges = True, discretize_given = True
d = DiscretizedLogistic(
mean=mean, log_scale=log_scale, bin_size=bin_size,
min_val=min_val, max_val=max_val, biased_edges=True,
dtype=tf.float64
)
assert_allclose(
sess.run(d.log_prob(x, group_ndims=0)),
naive_discretized_logistic_pdf(
x, mean, log_scale, bin_size, min_val, max_val,
biased_edges=True, group_ndims=0)
)
assert_allclose(
sess.run(d.log_prob(x, group_ndims=2)),
naive_discretized_logistic_pdf(
x, mean, log_scale, bin_size, min_val, max_val,
biased_edges=True, group_ndims=2)
)
def test_sample(self):
assert_allclose = functools.partial(
np.testing.assert_allclose, rtol=1e-5, atol=1e-6)
np.random.seed(1234)
mean = 3 * np.random.uniform(size=[2, 1, 4]).astype(np.float64) - 1
log_scale = np.random.normal(size=[3, 1, 5, 1]).astype(np.float64)
bin_size = 1 / 255.
epsilon = 1e-7
with self.test_session() as sess:
# n_samples == None
min_val = -1.5
max_val = 1.5
n_samples = None
sample_shape = [3, 2, 5, 4]
u = np.random.uniform(low=epsilon, high=1. - epsilon,
size=sample_shape).astype(np.float64)
x_ans = naive_discretized_logistic_sample(
u, mean, log_scale, bin_size, min_val, max_val)
def patched_rnd_uniform(*args, **kwargs):
return tf.convert_to_tensor(u)
with mock.patch('tensorflow.random_uniform',
Mock(wraps=patched_rnd_uniform)) as rnd_uniform:
d = DiscretizedLogistic(
mean=mean, log_scale=log_scale, bin_size=bin_size,
min_val=min_val, max_val=max_val, biased_edges=True,
dtype=tf.float64
)
x = d.sample(n_samples=n_samples, compute_density=True)
args = rnd_uniform.call_args[1]
self.assertEqual(list(sess.run(args['shape'])), sample_shape)
self.assertEqual(args['minval'], epsilon)
self.assertEqual(args['maxval'], 1. - epsilon)
self.assertEqual(args['dtype'], tf.float64)
assert_allclose(sess.run(x), x_ans)
assert_allclose(
sess.run(x.log_prob()),
naive_discretized_logistic_pdf(
x_ans, mean, log_scale, bin_size, min_val, max_val,
group_ndims=0
)
)
# n_samples == 11, min_val = max_val = None
min_val = None
max_val = None
n_samples = 11
sample_shape = [11, 3, 2, 5, 4]
u = np.random.uniform(low=epsilon, high=1. - epsilon,
size=sample_shape).astype(np.float64)
x_ans = naive_discretized_logistic_sample(
u, mean, log_scale, bin_size, min_val, max_val,
discretize_sample=False
)
def patched_rnd_uniform(*args, **kwargs):
return tf.convert_to_tensor(u)
with mock.patch('tensorflow.random_uniform',
Mock(wraps=patched_rnd_uniform)) as rnd_uniform:
d = DiscretizedLogistic(
mean=mean, log_scale=log_scale, bin_size=bin_size,
min_val=min_val, max_val=max_val, biased_edges=True,
dtype=tf.float64, discretize_sample=False,
discretize_given=False
)
x = d.sample(n_samples=n_samples, group_ndims=1)
args = rnd_uniform.call_args[1]
self.assertEqual(list(sess.run(args['shape'])), sample_shape)
self.assertEqual(args['minval'], epsilon)
self.assertEqual(args['maxval'], 1. - epsilon)
self.assertEqual(args['dtype'], tf.float64)
assert_allclose(sess.run(x), x_ans)
assert_allclose(
sess.run(x.log_prob()),
naive_discretized_logistic_pdf(
x_ans, mean, log_scale, bin_size, min_val, max_val,
group_ndims=1, discretize_given=False
)
)