def test_all_finite_raises(self):
with self.test_session():
x = np.inf * tf.constant([-1.0, -2.0, -3.0, -4.0])
with self.assertRaisesOpError('Inf'):
cumprod(x).eval()
x = tf.constant([-1.0, np.nan, -3.0, -4.0])
with self.assertRaisesOpError('NaN'):
cumprod(x).eval()
def test_all_finite_raises(self):
with self.test_session():
x = np.inf * tf.constant([-1.0, -2.0, -3.0, -4.0])
with self.assertRaisesOpError('Inf'):
cumprod(x).eval()
x = tf.constant([-1.0, np.nan, -3.0, -4.0])
with self.assertRaisesOpError('NaN'):
cumprod(x).eval()
def mapping(self, x):
# Transform a real (K-1)-vector to K-dimensional simplex.
pi = Variable("pi", [self.num_factors, self.K-1])
eq = -tf.log(tf.cast(self.K - 1 - tf.range(self.K-1), dtype=tf.float32))
z = tf.sigmoid(eq + pi)
pil = tf.concat(1, [z, tf.ones([self.num_factors, 1])])
piu = tf.concat(1, [tf.ones([self.num_factors, 1]), 1.0 - z])
# cumulative product along 1st axis
S = tf.pack([cumprod(piu_x) for piu_x in tf.unpack(piu)])
return [S * pil]
def mapping(self, x):
# Transform a real (K-1)-vector to K-dimensional simplex.
pi = Variable("pi", [self.num_factors, self.K - 1])
eq = -tf.log(
tf.cast(self.K - 1 - tf.range(self.K - 1), dtype=tf.float32))
z = tf.sigmoid(eq + pi)
pil = tf.concat(1, [z, tf.ones([self.num_factors, 1])])
piu = tf.concat(1, [tf.ones([self.num_factors, 1]), 1.0 - z])
# cumulative product along 1st axis
S = tf.pack([cumprod(piu_x) for piu_x in tf.unpack(piu)])
return [S * pil]
def __init__(self, shape, pi=None):
num_factors = shape[0]
K = shape[-1]
if K == 1:
raise ValueError("Multinomial is not supported for K=1. Use Bernoulli.")
Distribution.__init__(self, num_factors)
self.num_vars = K*num_factors
self.num_params = K*num_factors
self.K = K # dimension of each factor
self.sample_tensor = False
if pi is None:
# Transform a real (K-1)-vector to K-dimensional simplex.
pi_unconst = tf.Variable(tf.random_normal([self.num_factors, self.K-1]))
eq = -tf.log(tf.cast(self.K - 1 - tf.range(self.K-1), dtype=tf.float32))
x = tf.sigmoid(eq + pi_unconst)
pil = tf.concat(1, [x, tf.ones([self.num_factors, 1])])
piu = tf.concat(1, [tf.ones([self.num_factors, 1]), 1.0 - x])
# cumulative product along 1st axis
S = tf.pack([cumprod(piu_x) for piu_x in tf.unpack(piu)])
pi = S * pil
self.pi = pi
def test_cumprod_2d(self):
with self.test_session():
x = tf.constant([[-1.0], [-2.0], [-3.0], [-4.0]])
self.assertAllClose(cumprod(x).eval(),
np.array([[-1.], [2.], [-6.], [24.]]))
def test_cumprod_1d(self):
with self.test_session():
x = tf.constant([-1.0, -2.0, -3.0, -4.0])
self.assertAllEqual(cumprod(x).eval(),
np.array([-1., 2., -6., 24.]))
def test_cumprod_2d(self):
with self.test_session():
x = tf.constant([[-1.0], [-2.0], [-3.0], [-4.0]])
self.assertAllClose(cumprod(x).eval(),
np.array([ [-1.], [2.], [-6.], [24.]]))
def test_cumprod_1d(self):
with self.test_session():
x = tf.constant([-1.0, -2.0, -3.0, -4.0])
self.assertAllEqual(cumprod(x).eval(),
np.array([ -1., 2., -6., 24.]))