def testCovariance(self):
vla = tfd.VectorLaplaceDiag(loc=tf.zeros([2, 3], dtype=tf.float32),
validate_args=True)
self.assertAllClose(2. * np.diag(np.ones([3], dtype=np.float32)),
self.evaluate(vla.covariance()))
vla = tfd.VectorLaplaceDiag(loc=tf.zeros([3], dtype=tf.float32),
scale_identity_multiplier=[3., 2.],
validate_args=True)
self.assertAllEqual([2], vla.batch_shape)
self.assertAllEqual([3], vla.event_shape)
self.assertAllClose(
2. * np.array([[[3., 0, 0], [0, 3, 0], [0, 0, 3]],
[[2, 0, 0], [0, 2, 0], [0, 0, 2]]])**2.,
self.evaluate(vla.covariance()))
vla = tfd.VectorLaplaceDiag(loc=tf.zeros([3], dtype=tf.float32),
scale_diag=[[3., 2, 1], [4, 5, 6]],
validate_args=True)
self.assertAllEqual([2], vla.batch_shape)
self.assertAllEqual([3], vla.event_shape)
self.assertAllClose(
2. * np.array([[[3., 0, 0], [0, 2, 0], [0, 0, 1]],
[[4, 0, 0], [0, 5, 0], [0, 0, 6]]])**2.,
self.evaluate(vla.covariance()))
def testVariance(self):
vla = tfd.VectorLaplaceDiag(loc=tf.zeros([2, 3], dtype=tf.float32))
self.assertAllClose(2. * np.ones([3], dtype=np.float32),
self.evaluate(vla.variance()))
vla = tfd.VectorLaplaceDiag(loc=tf.zeros([3], dtype=tf.float32),
scale_identity_multiplier=[3., 2.])
self.assertAllClose(2. * np.array([[3., 3, 3], [2, 2, 2]])**2.,
self.evaluate(vla.variance()))
vla = tfd.VectorLaplaceDiag(loc=tf.zeros([3], dtype=tf.float32),
scale_diag=[[3., 2, 1], [4, 5, 6]])
self.assertAllClose(2. * np.array([[3., 2, 1], [4, 5, 6]])**2.,
self.evaluate(vla.variance()))
def test_passing_in_laplace_plus_defaults_is_same_as_laplace(self):
d = 10
scale_diag = rng.rand(d) + np.float64(1.2)
loc = rng.randn(d)
vlap = tfd.VectorLaplaceDiag(loc=loc,
scale_diag=scale_diag,
validate_args=True)
sasvlap = tfd.VectorSinhArcsinhDiag(loc=loc,
scale_diag=scale_diag,
distribution=tfd.Laplace(
np.float64(0.),
np.float64(1.)),
validate_args=True)
x = rng.randn(5, d)
vlap_pdf, sasvlap_pdf = self.evaluate([vlap.prob(x), sasvlap.prob(x)])
self.assertAllClose(vlap_pdf, sasvlap_pdf)
vlap_samps, sasvlap_samps = self.evaluate([
vlap.sample(10000, seed=tfp_test_util.test_seed()),
sasvlap.sample(10000, seed=tfp_test_util.test_seed())
])
self.assertAllClose(loc, sasvlap_samps.mean(axis=0), atol=0.1)
self.assertAllClose(vlap_samps.mean(axis=0),
sasvlap_samps.mean(axis=0),
atol=0.1)
self.assertAllClose(vlap_samps.std(axis=0),
sasvlap_samps.std(axis=0),
atol=0.1)
def testStddev(self):
vla = tfd.VectorLaplaceDiag(loc=tf.zeros([2, 3], dtype=tf.float32))
self.assertAllClose(
np.sqrt(2) * np.ones([3], dtype=np.float32),
self.evaluate(vla.stddev()))
vla = tfd.VectorLaplaceDiag(loc=tf.zeros([3], dtype=tf.float32),
scale_identity_multiplier=[3., 2.])
self.assertAllClose(
np.sqrt(2) * np.array([[3., 3, 3], [2, 2, 2]]),
self.evaluate(vla.stddev()))
vla = tfd.VectorLaplaceDiag(loc=tf.zeros([3], dtype=tf.float32),
scale_diag=[[3., 2, 1], [4, 5, 6]])
self.assertAllClose(
np.sqrt(2) * np.array([[3., 2, 1], [4, 5, 6]]),
self.evaluate(vla.stddev()))
def testSample(self):
mu = np.array([-1., 1])
diag = np.array([1., -2])
dist = tfd.VectorLaplaceDiag(mu, diag, validate_args=True)
seed = test_util.test_seed()
samps = self.evaluate(dist.sample(int(2e4), seed=seed))
cov_mat = 2. * self.evaluate(tf.linalg.diag(diag))**2
self.assertAllClose(mu, samps.mean(axis=0), atol=0., rtol=0.10)
self.assertAllClose(cov_mat, np.cov(samps.T), atol=0.15, rtol=0.10)
def testSample(self):
mu = [-1., 1]
diag = [1., -2]
dist = tfd.VectorLaplaceDiag(mu, diag, validate_args=True)
seed = tfp_test_util.test_seed(hardcoded_seed=0, set_eager_seed=False)
samps = self.evaluate(dist.sample(int(1e4), seed=seed))
cov_mat = 2. * self.evaluate(tf.linalg.diag(diag))**2
self.assertAllClose(mu, samps.mean(axis=0), atol=0., rtol=0.05)
self.assertAllClose(cov_mat, np.cov(samps.T), atol=0.05, rtol=0.05)
def testDistWithBatchShapeOneThenTransformedThroughSoftplus(self):
# This complex combination of events resulted in a loss of static shape
# information when tf.get_static_value(self._needs_rotation) was
# being used incorrectly (resulting in always rotating).
# Batch shape = [1], event shape = [3]
mu = tf.zeros((1, 3))
diag = tf.ones((1, 3))
base_dist = tfd.VectorLaplaceDiag(mu, diag, validate_args=True)
dist = tfd.TransformedDistribution(
base_dist, validate_args=True, bijector=tfp.bijectors.Softplus())
samps = dist.sample(5) # Shape [5, 1, 3].
self.assertAllEqual([5, 1], dist.log_prob(samps).shape)
def testSampleWithBroadcastScale(self):
# mu corresponds to a 2-batch of 3-variate normals
mu = np.zeros([2, 3])
# diag corresponds to no batches of 3-variate normals
diag = np.ones([3])
dist = tfd.VectorLaplaceDiag(mu, diag, validate_args=True)
mean = dist.mean()
self.assertAllEqual([2, 3], mean.shape)
self.assertAllClose(mu, self.evaluate(mean))
n = int(1e4)
samps = self.evaluate(dist.sample(n, seed=tfp_test_util.test_seed()))
cov_mat = 2. * self.evaluate(tf.linalg.diag(diag))**2
sample_cov = np.matmul(
samps.transpose([1, 2, 0]), samps.transpose([1, 0, 2])) / n
self.assertAllClose(mu, samps.mean(axis=0), atol=0.10, rtol=0.05)
self.assertAllClose([cov_mat, cov_mat], sample_cov, atol=0.10, rtol=0.05)
def testSingularScaleRaises(self):
mu = [-1., 1]
diag = [1., 0]
dist = tfd.VectorLaplaceDiag(mu, diag, validate_args=True)
with self.assertRaisesOpError("Singular"):
self.evaluate(dist.sample(seed=test_util.test_seed()))
def testMeanWithBroadcastLoc(self):
mu = [-1.]
diag = [1., -5]
dist = tfd.VectorLaplaceDiag(mu, diag, validate_args=True)
self.assertAllEqual([-1., -1.], self.evaluate(dist.mean()))
def testMean(self):
mu = [-1., 1]
diag = [1., -5]
dist = tfd.VectorLaplaceDiag(mu, diag, validate_args=True)
self.assertAllEqual(mu, self.evaluate(dist.mean()))
def testVectorParams(self):
mu = [-1.]
diag = [-5.]
dist = tfd.VectorLaplaceDiag(mu, diag, validate_args=True)
self.assertAllEqual([3, 1],
dist.sample(3, seed=test_util.test_seed()).shape)
def testScalarParams(self):
mu = -1.
diag = -5.
with self.assertRaisesRegexp(ValueError, "at least 1 dimension"):
tfd.VectorLaplaceDiag(mu, diag)
