def testXlogyWithZero(self):
for dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
x = constant_op.constant(np.zeros((2, 3)), dtype=dtype)
y = constant_op.constant([[0.1, 0.2, 3.5], [0., 1., 2.]], dtype=dtype)
with self.cached_session(use_gpu=True):
xlogy_tf_np = self.evaluate(math_ops.xlogy(x, y))
zeros_np = self.evaluate(array_ops.zeros_like(y))
self.assertAllClose(xlogy_tf_np, zeros_np)
def testXlogyNoZero(self):
for dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
x = constant_op.constant([[0.1, 0.2, 3.5], [-2., -5., 30.]], dtype=dtype)
y = constant_op.constant([[0.1, 0.2, 3.5], [3.1, 4., 2.]], dtype=dtype)
with self.cached_session(use_gpu=True):
xlogy = self.evaluate(math_ops.xlogy(x, y))
xtimeslogy = self.evaluate(x * math_ops.log(y))
self.assertAllClose(xlogy, xtimeslogy)
def testXlogyWithZeroBroadcast(self):
for dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
x = constant_op.constant([[0.], [1.]], dtype=dtype)
y = constant_op.constant([[0.1, 0.2, 3.5], [0., 1., 2.]], dtype=dtype)
with self.cached_session(use_gpu=True):
xlogy_tf_np = self.evaluate(math_ops.xlogy(x, y))
zeros_np = self.evaluate(array_ops.zeros_like(y[0]))
xtimes_logy = self.evaluate(math_ops.log(y[1]))
self.assertAllClose(zeros_np, xlogy_tf_np[0])
self.assertAllClose(xtimes_logy, xlogy_tf_np[1])
def testXlogyNoZero(self):
for dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
x = constant_op.constant([[0.1, 0.2, 3.5], [-2., -5., 30.]],
dtype=dtype)
y = constant_op.constant([[0.1, 0.2, 3.5], [3.1, 4., 2.]],
dtype=dtype)
with test_util.use_gpu():
xlogy = self.evaluate(math_ops.xlogy(x, y))
xtimeslogy = self.evaluate(x * math_ops.log(y))
self.assertAllClose(xlogy, xtimeslogy)
def testXlogyWithZeroBroadcast(self):
for dtype in [dtypes.float16, dtypes.float32, dtypes.float64]:
x = constant_op.constant([[0.], [1.]], dtype=dtype)
y = constant_op.constant([[0.1, 0.2, 3.5], [0., 1., 2.]], dtype=dtype)
with test_util.use_gpu():
xlogy_tf_np = self.evaluate(math_ops.xlogy(x, y))
zeros_np = self.evaluate(array_ops.zeros_like(y[0]))
xtimes_logy = self.evaluate(math_ops.log(y[1]))
self.assertAllClose(zeros_np, xlogy_tf_np[0])
self.assertAllClose(xtimes_logy, xlogy_tf_np[1])
def implicit_reparameterization_grad(a, x):
log_prob = math_ops.xlogy(a - 1., x) - math_ops.lgamma(a) - x
prob = math_ops.exp(log_prob)
return -gen_math_ops.igamma_grad_a(a, x) / prob
def _xlogy_gradients(self, x, y): xlogy_xgrad = self.evaluate(gradients.gradients(math_ops.xlogy(x, y), x)[0]) xlogy_ygrad = self.evaluate(gradients.gradients(math_ops.xlogy(x, y), y)[0]) return xlogy_xgrad, xlogy_ygrad
def _log_unnormalized_prob(self, x): x = self._maybe_assert_valid_sample(x) return math_ops.reduce_sum(math_ops.xlogy(self.concentration - 1., x), -1)
def _log_unnormalized_prob(self, x):
x = self._maybe_assert_valid_sample(x)
return (math_ops.xlogy(self.concentration1 - 1., x) +
(self.concentration0 - 1.) * math_ops.log1p(-x))
def _log_unnormalized_prob(self, x):
x = self._maybe_assert_valid_sample(x)
return (math_ops.xlogy(self.concentration1 - 1., x) +
(self.concentration0 - 1.) * math_ops.log1p(-x)) # pylint: disable=invalid-unary-operand-type
