def testModifiedVariableAssertion(self):
loc = 0.3
scale = 2.
tailweight = tf.Variable(1.9, dtype=self.dtype)
skewness = tf.Variable(1.2, dtype=self.dtype)
self.evaluate([tailweight.initializer, skewness.initializer])
normal_inverse_gaussian = tfd.NormalInverseGaussian(
loc=loc,
scale=scale,
tailweight=tailweight,
skewness=skewness,
validate_args=True)
with self.assertRaisesOpError('`tailweight > |skewness|`'):
with tf.control_dependencies([tailweight.assign(1.)]):
self.evaluate(normal_inverse_gaussian.mean())
tailweight = tf.Variable(1.9, dtype=self.dtype)
skewness = tf.Variable(1.2, dtype=self.dtype)
self.evaluate([tailweight.initializer, skewness.initializer])
normal_inverse_gaussian = tfd.NormalInverseGaussian(
loc=loc,
scale=scale,
tailweight=tailweight,
skewness=skewness,
validate_args=True)
with self.assertRaisesOpError('`tailweight > |skewness|`'):
with tf.control_dependencies([skewness.assign(-2.)]):
self.evaluate(normal_inverse_gaussian.mean())
def testNormalInverseGaussianLogPdfScipyMultidimensional(self):
loc_v = np.linspace(-10., 10., 5).astype(self.dtype).reshape(
(5, 1, 1, 1))
scale_v = self.dtype(1.)
tailweight_v = np.array([2., 0.5], dtype=self.dtype).reshape((2, 1))
skewness_v = 0.5 * tailweight_v
x_v = np.linspace(-9., 9., 20).astype(self.dtype)
normal_inverse_gaussian = tfd.NormalInverseGaussian(loc_v,
scale_v,
tailweight_v,
skewness_v,
validate_args=True)
log_prob = normal_inverse_gaussian.log_prob(x_v)
self.assertAllClose(
self.evaluate(log_prob),
stats.norminvgauss(a=tailweight_v,
b=skewness_v,
loc=loc_v,
scale=scale_v).logpdf(x_v))
prob = normal_inverse_gaussian.prob(x_v)
self.assertAllClose(
self.evaluate(prob),
stats.norminvgauss(a=tailweight_v,
b=skewness_v,
loc=loc_v,
scale=scale_v).pdf(x_v))
def testNormalInverseGaussianScipyLogPdf(self):
loc_v = self.dtype(0.)
scale_v = self.dtype(1.)
tailweight_v = self.dtype(1.)
skewness_v = self.dtype(0.2)
x_v = np.array([3., 3.1, 4., 5., 6., 7.]).astype(self.dtype)
normal_inverse_gaussian = tfd.NormalInverseGaussian(loc_v,
scale_v,
tailweight_v,
skewness_v,
validate_args=True)
log_prob = normal_inverse_gaussian.log_prob(x_v)
self.assertAllClose(
self.evaluate(log_prob),
stats.norminvgauss(a=tailweight_v,
b=skewness_v,
loc=loc_v,
scale=scale_v).logpdf(x_v))
pdf = normal_inverse_gaussian.prob(x_v)
self.assertAllClose(
self.evaluate(pdf),
stats.norminvgauss(a=tailweight_v,
b=skewness_v,
loc=loc_v,
scale=scale_v).pdf(x_v))
def testInvalidSkewnessTailweight(self):
with self.assertRaisesOpError('`tailweight > |skewness|`'):
normal_inverse_gaussian = tfd.NormalInverseGaussian(
loc=0.,
scale=1.,
tailweight=2.,
skewness=3.,
validate_args=True)
self.evaluate(normal_inverse_gaussian.mean())
def testInvalidScale(self):
with self.assertRaisesOpError('`scale` must be positive'):
normal_inverse_gaussian = tfd.NormalInverseGaussian(
loc=0.,
scale=-1.,
tailweight=2.,
skewness=1.,
validate_args=True)
self.evaluate(normal_inverse_gaussian.mean())
def testNormalInverseGaussianVarianceScale1(self):
loc_v = np.linspace(-10., 10., 5).astype(self.dtype).reshape((5, 1, 1, 1))
scale_v = self.dtype(1.)
tailweight_v = np.array([2., 0.5], dtype=self.dtype).reshape((2, 1))
skewness_v = 0.5 * tailweight_v
normal_inverse_gaussian = tfd.NormalInverseGaussian(
loc_v, scale_v, tailweight_v, skewness_v, validate_args=True)
self.assertAllClose(
self.evaluate(normal_inverse_gaussian.variance()),
stats.norminvgauss(
a=tailweight_v, b=skewness_v, loc=loc_v, scale=scale_v).var())
def testNormalInverseGaussianShape(self):
loc = tf.ones([2, 1, 1, 1], dtype=self.dtype)
scale = tf.ones([1, 3, 1, 1], dtype=self.dtype)
tailweight = 2 * tf.ones([1, 1, 5, 1], dtype=self.dtype)
skewness = tf.ones([1, 1, 1, 7], dtype=self.dtype)
normal_inverse_gaussian = tfd.NormalInverseGaussian(
loc, scale, tailweight, skewness, validate_args=True)
self.assertAllEqual(self.evaluate(
normal_inverse_gaussian.batch_shape_tensor()), (2, 3, 5, 7))
self.assertEqual(
normal_inverse_gaussian.batch_shape, tf.TensorShape([2, 3, 5, 7]))
self.assertAllEqual(
self.evaluate(normal_inverse_gaussian.event_shape_tensor()), [])
self.assertEqual(normal_inverse_gaussian.event_shape, tf.TensorShape([]))
def testNormalInverseGaussianSampleMultidimensionalVariance(self):
loc_v = np.linspace(-10., 10., 5).astype(self.dtype).reshape((5, 1, 1))
scale_v = np.linspace(1., 3., 3).astype(self.dtype).reshape((3, 1))
tailweight_v = np.array([2., 0.5], dtype=self.dtype)
skewness_v = 0.5 * tailweight_v
n = int(3e5)
normal_inverse_gaussian = tfd.NormalInverseGaussian(
loc_v, scale_v, tailweight_v, skewness_v, validate_args=True)
samples = normal_inverse_gaussian.sample(n, seed=test_util.test_seed())
sample_values = self.evaluate(samples)
self.assertEqual(sample_values.shape, (n, 5, 3, 2))
self.assertAllClose(
sample_values.var(axis=0),
self.evaluate(normal_inverse_gaussian.variance()),
rtol=.05,
atol=0)
def testNormalInverseGaussianSampleVariance(self):
loc_v = self.dtype(-2.)
scale_v = self.dtype(np.pi)
tailweight_v = self.dtype(1.87)
skewness_v = self.dtype(-1.)
n = int(3e5)
normal_inverse_gaussian = tfd.NormalInverseGaussian(
loc_v, scale_v, tailweight_v, skewness_v, validate_args=True)
samples = normal_inverse_gaussian.sample(n, seed=test_util.test_seed())
sample_values = self.evaluate(samples)
self.assertEqual(sample_values.shape, (n,))
self.assertAllClose(
sample_values.var(),
self.evaluate(normal_inverse_gaussian.variance()),
rtol=.02,
atol=0)
