def test_all_finite_raises(self):
with self.test_session():
x = tf.constant([12.5, np.inf])
with self.assertRaisesOpError('Inf'):
to_simplex(x).eval()
x = tf.constant([12.5, np.nan])
with self.assertRaisesOpError('NaN'):
to_simplex(x).eval()
def test_to_simplex_1d(self):
with self.test_session():
x = tf.constant([0.0])
self.assertAllClose(to_simplex(x).eval(), [0.5, 0.5])
x = tf.constant([0.0, 10.0])
self.assertAllClose(
to_simplex(x).eval(),
[3.333333e-01, 6.666363e-01, 3.027916e-05])
def test_all_finite_raises(self):
with self.test_session():
x = tf.constant([12.5, np.inf])
with self.assertRaisesOpError('Inf'):
to_simplex(x).eval()
x = tf.constant([12.5, np.nan])
with self.assertRaisesOpError('NaN'):
to_simplex(x).eval()
def test_to_simplex_1d(self):
with self.test_session():
x = tf.constant([0.0])
self.assertAllClose(to_simplex(x).eval(),
[0.5, 0.5])
x = tf.constant([0.0, 10.0])
self.assertAllClose(to_simplex(x).eval(),
[3.333333e-01, 6.666363e-01, 3.027916e-05])
def __init__(self, shape, pi=None):
if shape[-1] == 1:
raise ValueError("Multinomial is not supported for K=1. Use Bernoulli.")
Distribution.__init__(self, shape)
self.num_params = np.prod(shape[:-1]) * (shape[-1] -1)
self.sample_tensor = False
self.is_multivariate = True
if pi is None:
real_shape = self.shape[:-1]
K_minus_one = self.shape[-1] - 1
pi_unconst = tf.Variable(tf.random_normal([real_shape + (K_minus_one, )]))
pi = to_simplex(pi_unconst)
self.pi = pi
def __init__(self, shape, pi=None):
super(Multinomial, self).__init__(shape)
if self.shape[-1] == 1:
raise ValueError("Multinomial is not supported for K=1. Use Bernoulli.")
self.n_params = np.prod(self.shape[:-1]) * (self.shape[-1] -1)
self.is_differentiable = False
self.is_multivariate = True
self.is_reparameterized = False
if pi is None:
real_shape = self.shape[:-1]
K_minus_one = self.shape[-1] - 1
pi_unconst = tf.Variable(tf.random_normal([real_shape + (K_minus_one, )]))
pi = to_simplex(pi_unconst)
self.pi = pi
def __init__(self, shape, pi=None):
super(Multinomial, self).__init__(shape)
if self.shape[-1] == 1:
raise ValueError(
"Multinomial is not supported for K=1. Use Bernoulli.")
self.n_params = np.prod(self.shape[:-1]) * (self.shape[-1] - 1)
self.is_differentiable = False
self.is_multivariate = True
self.is_reparameterized = False
if pi is None:
real_shape = self.shape[:-1]
K_minus_one = self.shape[-1] - 1
pi_unconst = tf.Variable(
tf.random_normal([real_shape + (K_minus_one, )]))
pi = to_simplex(pi_unconst)
self.pi = pi
