def testEventNdimsLargerThanOneRaises(self):
with self.test_session():
mu = [1., -1]
with self.assertRaisesRegexp(
ValueError, (r"event_ndims\(2\) was not 0 or 1")):
# Scale corresponds to 2x2 identity matrix.
bijector = Affine(shift=mu, event_ndims=2, validate_args=True)
bijector.forward([1., 1.]).eval()
def testNoBatchMultivariateRaisesWhenSingular(self):
with self.test_session():
mu = [1., -1]
bijector = Affine(
shift=mu,
# Has zero on the diagonal.
scale_diag=[0., 1],
validate_args=True)
with self.assertRaisesOpError("Condition x > 0"):
bijector.forward([1., 1.]).eval()
def testNoBatchMultivariateRaisesWhenSingular(self):
with self.test_session():
mu = [1., -1]
bijector = Affine(
shift=mu,
# Has zero on the diagonal.
scale_diag=[0., 1],
validate_args=True)
with self.assertRaisesOpError("diagonal part must be non-zero"):
bijector.forward([1., 1.]).eval()
def testScaleZeroScalarRaises(self):
with self.test_session():
mu = -1.
# Check Identity matrix with zero scaling.
bijector = Affine(
shift=mu,
scale_identity_multiplier=0.,
event_ndims=0,
validate_args=True)
with self.assertRaisesOpError("identity_multiplier should be non-zero"):
bijector.forward(1.).eval()
def testBatchMultivariateFullDynamic(self):
with self.test_session() as sess:
x = array_ops.placeholder(dtypes.float32, name="x")
mu = array_ops.placeholder(dtypes.float32, name="mu")
scale_diag = array_ops.placeholder(dtypes.float32,
name="scale_diag")
event_ndims = array_ops.placeholder(dtypes.int32,
name="event_ndims")
x_value = np.array([[[1., 1]]], dtype=np.float32)
mu_value = np.array([[1., -1]], dtype=np.float32)
scale_diag_value = np.array([[2., 2]], dtype=np.float32)
event_ndims_value = 1
feed_dict = {
x: x_value,
mu: mu_value,
scale_diag: scale_diag_value,
event_ndims: event_ndims_value
}
bijector = Affine(shift=mu,
scale_diag=scale_diag,
event_ndims=event_ndims)
self.assertEqual(1, sess.run(bijector.event_ndims, feed_dict))
self.assertAllClose([[[3., 1]]],
sess.run(bijector.forward(x), feed_dict))
self.assertAllClose([[[0., 1]]],
sess.run(bijector.inverse(x), feed_dict))
self.assertAllClose([-np.log(4)],
sess.run(bijector.inverse_log_det_jacobian(x),
feed_dict))
def testCompareToBijector(self):
"""Demonstrates equivalence between TD, Bijector approach and AR dist."""
sample_shape = np.int32([4, 5])
batch_shape = np.int32([])
event_size = np.int32(2)
with self.cached_session() as sess:
batch_event_shape = np.concatenate([batch_shape, [event_size]],
axis=0)
sample0 = array_ops.zeros(batch_event_shape)
affine = Affine(scale_tril=self._random_scale_tril(event_size))
ar = autoregressive_lib.Autoregressive(self._normal_fn(affine),
sample0,
validate_args=True)
ar_flow = MaskedAutoregressiveFlow(is_constant_jacobian=True,
shift_and_log_scale_fn=lambda x:
[None, affine.forward(x)],
validate_args=True)
td = transformed_distribution_lib.TransformedDistribution(
distribution=normal_lib.Normal(loc=0., scale=1.),
bijector=ar_flow,
event_shape=[event_size],
batch_shape=batch_shape,
validate_args=True)
x_shape = np.concatenate([sample_shape, batch_shape, [event_size]],
axis=0)
x = 2. * self._rng.random_sample(x_shape).astype(np.float32) - 1.
td_log_prob_, ar_log_prob_ = sess.run(
[td.log_prob(x), ar.log_prob(x)])
self.assertAllClose(td_log_prob_, ar_log_prob_, atol=0., rtol=1e-6)
def testNoBatchMultivariateFullDynamic(self):
with self.cached_session() as sess:
x = array_ops.placeholder(dtypes.float32, name="x")
mu = array_ops.placeholder(dtypes.float32, name="mu")
scale_diag = array_ops.placeholder(dtypes.float32,
name="scale_diag")
x_value = np.array([[1., 1]], dtype=np.float32)
mu_value = np.array([1., -1], dtype=np.float32)
scale_diag_value = np.array([2., 2], dtype=np.float32)
feed_dict = {
x: x_value,
mu: mu_value,
scale_diag: scale_diag_value,
}
bijector = Affine(shift=mu, scale_diag=scale_diag)
self.assertAllClose([[3., 1]],
sess.run(bijector.forward(x), feed_dict))
self.assertAllClose([[0., 1]],
sess.run(bijector.inverse(x), feed_dict))
self.assertAllClose(
-np.log(4),
sess.run(bijector.inverse_log_det_jacobian(x, event_ndims=1),
feed_dict))
def testBatchMultivariateFullDynamic(self):
with self.test_session() as sess:
x = array_ops.placeholder(dtypes.float32, name="x")
mu = array_ops.placeholder(dtypes.float32, name="mu")
scale_diag = array_ops.placeholder(dtypes.float32, name="scale_diag")
event_ndims = array_ops.placeholder(dtypes.int32, name="event_ndims")
x_value = np.array([[[1., 1]]], dtype=np.float32)
mu_value = np.array([[1., -1]], dtype=np.float32)
scale_diag_value = np.array([[2., 2]], dtype=np.float32)
event_ndims_value = 1
feed_dict = {
x: x_value,
mu: mu_value,
scale_diag: scale_diag_value,
event_ndims: event_ndims_value
}
bijector = Affine(
shift=mu, scale_diag=scale_diag, event_ndims=event_ndims)
self.assertEqual(1, sess.run(bijector.event_ndims, feed_dict))
self.assertAllClose([[[3., 1]]], sess.run(bijector.forward(x), feed_dict))
self.assertAllClose([[[0., 1]]], sess.run(bijector.inverse(x), feed_dict))
self.assertAllClose([-np.log(4)],
sess.run(
bijector.inverse_log_det_jacobian(x), feed_dict))
def testCompareToBijector(self):
"""Demonstrates equivalence between TD, Bijector approach and AR dist."""
sample_shape = np.int32([4, 5])
batch_shape = np.int32([])
event_size = np.int32(2)
with self.test_session() as sess:
batch_event_shape = np.concatenate([batch_shape, [event_size]], axis=0)
sample0 = array_ops.zeros(batch_event_shape)
affine = Affine(scale_tril=self._random_scale_tril(event_size))
ar = autoregressive_lib.Autoregressive(
self._normal_fn(affine), sample0, validate_args=True)
ar_flow = MaskedAutoregressiveFlow(
is_constant_jacobian=True,
shift_and_log_scale_fn=lambda x: [None, affine.forward(x)],
validate_args=True)
td = transformed_distribution_lib.TransformedDistribution(
distribution=normal_lib.Normal(loc=0., scale=1.),
bijector=ar_flow,
event_shape=[event_size],
batch_shape=batch_shape,
validate_args=True)
x_shape = np.concatenate(
[sample_shape, batch_shape, [event_size]], axis=0)
x = 2. * self._rng.random_sample(x_shape).astype(np.float32) - 1.
td_log_prob_, ar_log_prob_ = sess.run([td.log_prob(x), ar.log_prob(x)])
self.assertAllClose(td_log_prob_, ar_log_prob_, atol=0., rtol=1e-6)
def testScaleZeroScalarRaises(self):
with self.test_session():
mu = -1.
# Check Identity matrix with zero scaling.
bijector = Affine(
shift=mu,
scale_identity_multiplier=0.0,
event_ndims=0,
validate_args=True)
with self.assertRaisesOpError("Condition x > 0"):
bijector.forward(1.).eval()
# Check Diag matrix with zero scaling.
bijector = Affine(
shift=mu, scale_diag=[0.0], event_ndims=0, validate_args=True)
with self.assertRaisesOpError("Condition x > 0"):
bijector.forward(1.).eval()
def testScaleZeroScalarRaises(self):
with self.test_session():
mu = -1.
# Check Identity matrix with zero scaling.
bijector = Affine(shift=mu,
scale_identity_multiplier=0.0,
event_ndims=0,
validate_args=True)
with self.assertRaisesOpError("Condition x > 0"):
bijector.forward(1.).eval()
# Check Diag matrix with zero scaling.
bijector = Affine(shift=mu,
scale_diag=[0.0],
event_ndims=0,
validate_args=True)
with self.assertRaisesOpError("Condition x > 0"):
bijector.forward(1.).eval()
def _testLegalInputs(self, shift=None, scale_params=None, x=None):
def _powerset(x):
s = list(x)
return itertools.chain.from_iterable(
itertools.combinations(s, r) for r in range(len(s) + 1))
for args in _powerset(scale_params.items()):
with self.test_session():
args = dict(args)
scale_args = dict({"x": x}, **args)
scale = self._makeScale(**scale_args)
# We haven't specified enough information for the scale.
if scale is None:
with self.assertRaisesRegexp(ValueError,
("must be specified.")):
bijector = Affine(shift=shift, **args)
else:
bijector = Affine(shift=shift, **args)
np_x = x
# For the case a vector is passed in, we need to make the shape
# match the matrix for matmul to work.
if x.ndim == scale.ndim - 1:
np_x = np.expand_dims(x, axis=-1)
forward = np.matmul(scale, np_x) + shift
if x.ndim == scale.ndim - 1:
forward = np.squeeze(forward, axis=-1)
self.assertAllClose(forward, bijector.forward(x).eval())
backward = np.linalg.solve(scale, np_x - shift)
if x.ndim == scale.ndim - 1:
backward = np.squeeze(backward, axis=-1)
self.assertAllClose(backward, bijector.inverse(x).eval())
scale *= np.ones(shape=x.shape[:-1], dtype=scale.dtype)
ildj = -np.log(np.abs(np.linalg.det(scale)))
# TODO (jvdillon): We need to make it so the scale_identity_multiplier id:1102
# https://github.com/imdone/tensorflow/issues/1103
# case does not deviate in expected shape. Fixing this will get rid of
# these special cases.
if (ildj.ndim > 0 and
(len(scale_args) == 1 or
(len(scale_args) == 2 and scale_args.get(
"scale_identity_multiplier", None) is not None))):
ildj = np.squeeze(ildj[0])
elif ildj.ndim < scale.ndim - 2:
ildj = np.reshape(ildj, scale.shape[0:-2])
self.assertAllClose(
ildj,
bijector.inverse_log_det_jacobian(
x, event_ndims=1).eval())
def _testLegalInputs(self, shift=None, scale_params=None, x=None):
def _powerset(x):
s = list(x)
return itertools.chain.from_iterable(
itertools.combinations(s, r) for r in range(len(s) + 1))
for args in _powerset(scale_params.items()):
with self.test_session():
args = dict(args)
scale_args = dict({"x": x}, **args)
scale = self._makeScale(**scale_args)
bijector_args = dict({"event_ndims": 1}, **args)
# We haven't specified enough information for the scale.
if scale is None:
with self.assertRaisesRegexp(ValueError, ("must be specified.")):
bijector = Affine(shift=shift, **bijector_args)
else:
bijector = Affine(shift=shift, **bijector_args)
np_x = x
# For the case a vector is passed in, we need to make the shape
# match the matrix for matmul to work.
if x.ndim == scale.ndim - 1:
np_x = np.expand_dims(x, axis=-1)
forward = np.matmul(scale, np_x) + shift
if x.ndim == scale.ndim - 1:
forward = np.squeeze(forward, axis=-1)
self.assertAllClose(forward, bijector.forward(x).eval())
backward = np.linalg.solve(scale, np_x - shift)
if x.ndim == scale.ndim - 1:
backward = np.squeeze(backward, axis=-1)
self.assertAllClose(backward, bijector.inverse(x).eval())
ildj = -np.log(np.abs(np.linalg.det(scale)))
# TODO(jvdillon): We need to make it so the scale_identity_multiplier
# case does not deviate in expected shape. Fixing this will get rid of
# these special cases.
if (ildj.ndim > 0 and (len(scale_args) == 1 or (
len(scale_args) == 2 and
scale_args.get("scale_identity_multiplier", None) is not None))):
ildj = np.squeeze(ildj[0])
elif ildj.ndim < scale.ndim - 2:
ildj = np.reshape(ildj, scale.shape[0:-2])
self.assertAllClose(ildj, bijector.inverse_log_det_jacobian(x).eval())
def testNoBatchMultivariateFullDynamic(self):
with self.test_session() as sess:
x = array_ops.placeholder(dtypes.float32, name="x")
mu = array_ops.placeholder(dtypes.float32, name="mu")
scale_diag = array_ops.placeholder(dtypes.float32, name="scale_diag")
x_value = np.array([[1., 1]], dtype=np.float32)
mu_value = np.array([1., -1], dtype=np.float32)
scale_diag_value = np.array([2., 2], dtype=np.float32)
feed_dict = {
x: x_value,
mu: mu_value,
scale_diag: scale_diag_value,
}
bijector = Affine(shift=mu, scale_diag=scale_diag)
self.assertAllClose([[3., 1]], sess.run(bijector.forward(x), feed_dict))
self.assertAllClose([[0., 1]], sess.run(bijector.inverse(x), feed_dict))
self.assertAllClose(
-np.log(4),
sess.run(bijector.inverse_log_det_jacobian(x), feed_dict))
def testEventNdimsLargerThanOneRaises(self):
with self.test_session():
mu = [1., -1]
# Scale corresponds to 2x2 identity matrix.
bijector = Affine(shift=mu, event_ndims=2, validate_args=True)
bijector.forward([1., 1.]).eval()
def testEventNdimsLargerThanOneRaises(self):
with self.test_session():
mu = [1., -1]
# Scale corresponds to 2x2 identity matrix.
bijector = Affine(shift=mu, event_ndims=2, validate_args=True)
bijector.forward([1., 1.]).eval()