def testValueOutsideAndOnEdgeOfSupport(self, dtype):
finfo = np.finfo(dtype)
x = tf.convert_to_tensor(
[
# Sqrt(finfo.max)**2 = finfo.max < Inf, so
# round_exponential_bump_function == 0 here.
-np.sqrt(finfo.max),
# -2 is just outside the support, so round_exponential_bump_function
# should == 0.
-2.,
# -1 is on boundary of support, so round_exponential_bump_function
# should == 0.
# The gradient should also equal 0.
-1.,
1.,
2.0,
np.sqrt(finfo.max),
],
dtype=dtype)
y = tfp_math.round_exponential_bump_function(x)
self.assertDTypeEqual(y, dtype)
y_ = self.evaluate(y)
# We hard-set y to 0 outside support.
self.assertAllFinite(y_)
self.assertAllEqual(y_, np.zeros((6, )))
def testValueOnSupportInterior(self, dtype):
# round_exponential_bump_function(x) = 0 for x right at the edge of the
# support, e.g. x = -0.999. This is expected, due to the exponential and
# division.
x = tf.convert_to_tensor([-0.9925, -0.5, 0., 0.5, 0.9925], dtype=dtype)
y = tfp_math.round_exponential_bump_function(x)
self.assertDTypeEqual(y, dtype)
y_ = self.evaluate(y)
self.assertAllFinite(y_)
# round_exponential_bump_function(0) = 1.
self.assertAllClose(1., y_[2])
# round_exponential_bump_function(x) > 0 for |x| < 1.
self.assertAllGreater(y_, 0)
