def testShapeGetterRaisesException(self):
x = tf.Variable(-1.)
self.evaluate(tf1.global_variables_initializer())
with self.assertRaisesOpError(
'Argument `concentration1` must be positive.'):
b = tfb.Kumaraswamy(concentration0=1.,
concentration1=x,
validate_args=True)
self.evaluate(b.forward_event_shape_tensor([1, 2, 3]))
with self.assertRaisesOpError(
'Argument `concentration0` must be positive.'):
b = tfb.Kumaraswamy(concentration0=x,
concentration1=1.,
validate_args=True)
self.evaluate(b.forward_event_shape_tensor(tf.constant([1, 2, 3])))
def testBijector(self):
with self.cached_session():
a = 2.
b = 0.3
bijector = tfb.Kumaraswamy(concentration1=a,
concentration0=b,
validate_args=True)
self.assertEqual("kumaraswamy", bijector.name)
x = np.array([[[0.1], [0.2], [0.3], [0.4], [0.5]]],
dtype=np.float32)
# Kumaraswamy cdf. This is the same as inverse(x).
y = 1. - (1. - x**a)**b
self.assertAllClose(y, self.evaluate(bijector.inverse(x)))
self.assertAllClose(x, self.evaluate(bijector.forward(y)))
kumaraswamy_log_pdf = (np.log(a) + np.log(b) +
(a - 1) * np.log(x) +
(b - 1) * np.log1p(-x**a))
self.assertAllClose(
np.squeeze(kumaraswamy_log_pdf, axis=-1),
self.evaluate(
bijector.inverse_log_det_jacobian(x, event_ndims=1)))
self.assertAllClose(self.evaluate(
-bijector.inverse_log_det_jacobian(x, event_ndims=1)),
self.evaluate(
bijector.forward_log_det_jacobian(
y, event_ndims=1)),
rtol=1e-4,
atol=0.)
def testScalarCongruency(self):
bijector_test_util.assert_scalar_congruency(tfb.Kumaraswamy(
concentration1=0.5, concentration0=1.1),
lower_x=0.,
upper_x=1.,
eval_func=self.evaluate,
n=int(10e3),
rtol=0.02)
def testScalarCongruency(self):
with self.test_session():
assert_scalar_congruency(tfb.Kumaraswamy(concentration1=0.5,
concentration0=1.1),
lower_x=0.,
upper_x=1.,
n=int(10e3),
rtol=0.02)
def testBijectorConcentration1LogDetJacobianFiniteAtZero(self):
# When concentration = 1., forward_log_det_jacobian should be finite at
# zero.
concentration0 = np.logspace(0.1, 10., num=20).astype(np.float32)
bijector = tfb.Kumaraswamy(concentration1=1., concentration0=concentration0)
fldj = self.evaluate(
bijector.forward_log_det_jacobian(0., event_ndims=0))
self.assertAllEqual(np.ones_like(fldj, dtype=np.bool), np.isfinite(fldj))
def testGradient(self):
x = tf.Variable(1.)
y = tf.Variable(2.)
b = tfb.Kumaraswamy(concentration0=x,
concentration1=y,
validate_args=True)
with tf.GradientTape() as tape:
loss = b.forward(1.)
g = tape.gradient(loss, b.trainable_variables)
self.assertLen(g, 2)
self.assertAllNotNone(g)
def transform_by_kumaraswamy(x, feature_ndims, example_ndims):
"""Apply a Kumaraswamy bijector to features."""
concentration1 = util.pad_shape_with_ones(
self.concentration1, example_ndims, start=-(feature_ndims + 1))
concentration0 = util.pad_shape_with_ones(
self.concentration0, example_ndims, start=-(feature_ndims + 1))
bij = bijectors.Kumaraswamy(concentration1,
concentration0,
validate_args=validate_args)
# Apply the inverse as this is the Kumaraswamy CDF.
return bij.inverse(x)
def testVariableConcentration1(self):
x = tf.Variable(1.)
b = tfb.Kumaraswamy(concentration0=1.,
concentration1=x,
validate_args=True)
self.evaluate(tf1.global_variables_initializer())
self.assertIs(x, b.concentration1)
self.assertEqual((), self.evaluate(b.forward(1.)).shape)
with self.assertRaisesOpError(
'Argument `concentration1` must be positive.'):
with tf.control_dependencies([x.assign(-1.)]):
self.assertEqual((), self.evaluate(b.forward(1.)).shape)
def testBijectiveAndFinite(self):
with self.test_session():
concentration1 = 1.2
concentration0 = 2.
bijector = tfb.Kumaraswamy(
concentration1=concentration1,
concentration0=concentration0,
validate_args=True)
# Omitting the endpoints 0 and 1, since idlj will be infinity at these
# endpoints.
y = np.linspace(.01, 0.99, num=10).astype(np.float32)
x = 1 - (1 - y ** concentration1) ** concentration0
assert_bijective_and_finite(bijector, x, y, event_ndims=0, rtol=1e-3)
def __init__(self,
concentration1=None,
concentration0=None,
validate_args=False,
allow_nan_stats=True,
name="Kumaraswamy"):
"""Initialize a batch of Kumaraswamy distributions.
Args:
concentration1: Positive floating-point `Tensor` indicating mean
number of successes; aka "alpha". Implies `self.dtype` and
`self.batch_shape`, i.e.,
`concentration1.shape = [N1, N2, ..., Nm] = self.batch_shape`.
concentration0: Positive floating-point `Tensor` indicating mean
number of failures; aka "beta". Otherwise has same semantics as
`concentration1`.
validate_args: Python `bool`, default `False`. When `True` distribution
parameters are checked for validity despite possibly degrading runtime
performance. When `False` invalid inputs may silently render incorrect
outputs.
allow_nan_stats: Python `bool`, default `True`. When `True`, statistics
(e.g., mean, mode, variance) use the value "`NaN`" to indicate the
result is undefined. When `False`, an exception is raised if one or
more of the statistic's batch members are undefined.
name: Python `str` name prefixed to Ops created by this class.
"""
with tf.name_scope(name, values=[concentration1,
concentration0]) as name:
dtype = dtype_util.common_dtype([concentration1, concentration0],
tf.float32)
concentration1 = tf.convert_to_tensor(concentration1,
name="concentration1",
dtype=dtype)
concentration0 = tf.convert_to_tensor(concentration0,
name="concentration0",
dtype=dtype)
super(Kumaraswamy, self).__init__(
distribution=tf.distributions.Uniform(
low=tf.zeros([], dtype=concentration1.dtype),
high=tf.ones([], dtype=concentration1.dtype),
allow_nan_stats=allow_nan_stats),
bijector=bijectors.Kumaraswamy(concentration1=concentration1,
concentration0=concentration0,
validate_args=validate_args),
batch_shape=distribution_util.get_broadcast_shape(
concentration1, concentration0),
name=name)
self._reparameterization_type = tf.distributions.FULLY_REPARAMETERIZED
