def state_space_model_likelihood(**param_vals):
ssm = self.make_state_space_model(
param_vals=param_vals,
num_timesteps=num_timesteps,
initial_step=initial_step,
mask=mask,
experimental_parallelize=experimental_parallelize)
# Looping LGSSM methods are really expensive in eager mode; wrap them
# to keep this from slowing things down in interactive use.
ssm = tfe_util.JitPublicMethods(ssm, trace_only=True)
if distribution_util.shape_may_be_nontrivial(trajectories_shape):
return sample.Sample(ssm, sample_shape=trajectories_shape)
return ssm
def build_split_flow_surrogate_posterior(event_shape,
trainable_bijector,
constraining_bijector=None,
base_distribution=normal.Normal,
batch_shape=(),
dtype=tf.float32,
validate_args=False,
name=None):
"""Builds a joint variational posterior by splitting a normalizing flow.
Args:
event_shape: (Nested) event shape of the surrogate posterior.
trainable_bijector: A trainable `tfb.Bijector` instance that operates on
`Tensor`s (not structures), e.g. `tfb.MaskedAutoregressiveFlow` or
`tfb.RealNVP`. This bijector transforms the base distribution before it is
split.
constraining_bijector: `tfb.Bijector` instance, or nested structure of
`tfb.Bijector` instances, that maps (nested) values in R^n to the support
of the posterior. (This can be the
`experimental_default_event_space_bijector` of the distribution over the
prior latent variables.)
Default value: `None` (i.e., the posterior is over R^n).
base_distribution: A `tfd.Distribution` subclass parameterized by `loc` and
`scale`. The base distribution for the transformed surrogate has `loc=0.`
and `scale=1.`.
Default value: `tfd.Normal`.
batch_shape: The `batch_shape` of the output distribution.
Default value: `()`.
dtype: The `dtype` of the surrogate posterior.
Default value: `tf.float32`.
validate_args: Python `bool`. Whether to validate input with asserts. This
imposes a runtime cost. If `validate_args` is `False`, and the inputs are
invalid, correct behavior is not guaranteed.
Default value: `False`.
name: Python `str` name prefixed to ops created by this function.
Default value: `None` (i.e., 'build_split_flow_surrogate_posterior').
Returns:
surrogate_distribution: Trainable `tfd.TransformedDistribution` with event
shape equal to `event_shape`.
### Examples
```python
# Train a normalizing flow on the Eight Schools model [1].
treatment_effects = [28., 8., -3., 7., -1., 1., 18., 12.]
treatment_stddevs = [15., 10., 16., 11., 9., 11., 10., 18.]
model = tfd.JointDistributionNamed({
'avg_effect':
tfd.Normal(loc=0., scale=10., name='avg_effect'),
'log_stddev':
tfd.Normal(loc=5., scale=1., name='log_stddev'),
'school_effects':
lambda log_stddev, avg_effect: (
tfd.Independent(
tfd.Normal(
loc=avg_effect[..., None] * tf.ones(8),
scale=tf.exp(log_stddev[..., None]) * tf.ones(8),
name='school_effects'),
reinterpreted_batch_ndims=1)),
'treatment_effects': lambda school_effects: tfd.Independent(
tfd.Normal(loc=school_effects, scale=treatment_stddevs),
reinterpreted_batch_ndims=1)
})
# Pin the observed values in the model.
target_model = model.experimental_pin(treatment_effects=treatment_effects)
# Create a Masked Autoregressive Flow bijector.
net = tfb.AutoregressiveNetwork(2, hidden_units=[16, 16], dtype=tf.float32)
maf = tfb.MaskedAutoregressiveFlow(shift_and_log_scale_fn=net)
# Build and fit the surrogate posterior.
surrogate_posterior = (
tfp.experimental.vi.build_split_flow_surrogate_posterior(
event_shape=target_model.event_shape_tensor(),
trainable_bijector=maf,
constraining_bijector=(
target_model.experimental_default_event_space_bijector())))
losses = tfp.vi.fit_surrogate_posterior(
target_model.unnormalized_log_prob,
surrogate_posterior,
num_steps=100,
optimizer=tf.optimizers.Adam(0.1),
sample_size=10)
```
#### References
[1] Andrew Gelman, John Carlin, Hal Stern, David Dunson, Aki Vehtari, and
Donald Rubin. Bayesian Data Analysis, Third Edition.
Chapman and Hall/CRC, 2013.
"""
with tf.name_scope(name or 'build_split_flow_surrogate_posterior'):
shallow_structure = _get_event_shape_shallow_structure(event_shape)
event_shape = nest.map_structure_up_to(shallow_structure,
ps.convert_to_shape_tensor,
event_shape)
if nest.is_nested(constraining_bijector):
constraining_bijector = joint_map.JointMap(
nest.map_structure(
lambda b: identity.Identity()
if b is None else b, constraining_bijector),
validate_args=validate_args)
if constraining_bijector is None:
unconstrained_event_shape = event_shape
else:
unconstrained_event_shape = (
constraining_bijector.inverse_event_shape_tensor(event_shape))
flat_base_event_shape = nest.flatten(unconstrained_event_shape)
flat_base_event_size = nest.map_structure(tf.reduce_prod,
flat_base_event_shape)
event_size = tf.reduce_sum(flat_base_event_size)
base_distribution = sample.Sample(
base_distribution(tf.zeros(batch_shape, dtype=dtype), scale=1.),
[event_size])
# After transforming base distribution samples with `trainable_bijector`,
# split them into vector-valued components.
split_bijector = split.Split(flat_base_event_size,
validate_args=validate_args)
# Reshape the vectors to the correct posterior event shape.
event_reshape = joint_map.JointMap(nest.map_structure(
reshape.Reshape, unconstrained_event_shape),
validate_args=validate_args)
# Restructure the flat list of components to the correct posterior
# structure.
event_unflatten = restructure.Restructure(
nest.pack_sequence_as(unconstrained_event_shape,
range(len(flat_base_event_shape))))
bijectors = [] if constraining_bijector is None else [
constraining_bijector
]
bijectors.extend([
event_reshape, event_unflatten, split_bijector, trainable_bijector
])
bijector = chain.Chain(bijectors, validate_args=validate_args)
return transformed_distribution.TransformedDistribution(
base_distribution, bijector=bijector, validate_args=validate_args)
def _affine_surrogate_posterior(event_shape,
operators='diag',
bijector=None,
base_distribution=normal.Normal,
dtype=tf.float32,
batch_shape=(),
validate_args=False,
name=None):
"""Builds a joint variational posterior with a given `event_shape`.
This function builds a surrogate posterior by applying a trainable
transformation to a standard base distribution and constraining the samples
with `bijector`. The surrogate posterior has event shape equal to
the input `event_shape`.
This function is a convenience wrapper around
`build_affine_surrogate_posterior_from_base_distribution` that allows the
user to pass in the desired posterior `event_shape` instead of
pre-constructed base distributions (at the expense of full control over the
base distribution types and parameterizations).
Args:
event_shape: (Nested) event shape of the posterior.
operators: Either a string or a list/tuple containing `LinearOperator`
subclasses, `LinearOperator` instances, or callables returning
`LinearOperator` instances. Supported string values are "diag" (to create
a mean-field surrogate posterior) and "tril" (to create a full-covariance
surrogate posterior). A list/tuple may be passed to induce other
posterior covariance structures. If the list is flat, a
`tf.linalg.LinearOperatorBlockDiag` instance will be created and applied
to the base distribution. Otherwise the list must be singly-nested and
have a first element of length 1, second element of length 2, etc.; the
elements of the outer list are interpreted as rows of a lower-triangular
block structure, and a `tf.linalg.LinearOperatorBlockLowerTriangular`
instance is created. For complete documentation and examples, see
`tfp.experimental.vi.util.build_trainable_linear_operator_block`, which
receives the `operators` arg if it is list-like.
Default value: `"diag"`.
bijector: `tfb.Bijector` instance, or nested structure of `tfb.Bijector`
instances, that maps (nested) values in R^n to the support of the
posterior. (This can be the `experimental_default_event_space_bijector` of
the distribution over the prior latent variables.)
Default value: `None` (i.e., the posterior is over R^n).
base_distribution: A `tfd.Distribution` subclass parameterized by `loc` and
`scale`. The base distribution of the transformed surrogate has `loc=0.`
and `scale=1.`.
Default value: `tfd.Normal`.
dtype: The `dtype` of the surrogate posterior.
Default value: `tf.float32`.
batch_shape: Batch shape (Python tuple, list, or int) of the surrogate
posterior, to enable parallel optimization from multiple initializations.
Default value: `()`.
validate_args: Python `bool`. Whether to validate input with asserts. This
imposes a runtime cost. If `validate_args` is `False`, and the inputs are
invalid, correct behavior is not guaranteed.
Default value: `False`.
name: Python `str` name prefixed to ops created by this function.
Default value: `None` (i.e., 'build_affine_surrogate_posterior').
Yields:
*parameters: sequence of `trainable_state_util.Parameter` namedtuples.
These are intended to be consumed by
`trainable_state_util.as_stateful_builder` and
`trainable_state_util.as_stateless_builder` to define stateful and
stateless variants respectively.
#### Examples
```python
tfd = tfp.distributions
tfb = tfp.bijectors
# Define a joint probabilistic model.
Root = tfd.JointDistributionCoroutine.Root
def model_fn():
concentration = yield Root(tfd.Exponential(1.))
rate = yield Root(tfd.Exponential(1.))
y = yield tfd.Sample(
tfd.Gamma(concentration=concentration, rate=rate),
sample_shape=4)
model = tfd.JointDistributionCoroutine(model_fn)
# Assume the `y` are observed, such that the posterior is a joint distribution
# over `concentration` and `rate`. The posterior event shape is then equal to
# the first two components of the model's event shape.
posterior_event_shape = model.event_shape_tensor()[:-1]
# Constrain the posterior values to be positive using the `Exp` bijector.
bijector = [tfb.Exp(), tfb.Exp()]
# Build a full-covariance surrogate posterior.
surrogate_posterior = (
tfp.experimental.vi.build_affine_surrogate_posterior(
event_shape=posterior_event_shape,
operators='tril',
bijector=bijector))
# For an example defining `'operators'` as a list to express an alternative
# covariance structure, see
# `build_affine_surrogate_posterior_from_base_distribution`.
# Fit the model.
y = [0.2, 0.5, 0.3, 0.7]
target_model = model.experimental_pin(y=y)
losses = tfp.vi.fit_surrogate_posterior(
target_model.unnormalized_log_prob,
surrogate_posterior,
num_steps=100,
optimizer=tf.optimizers.Adam(0.1),
sample_size=10)
```
"""
with tf.name_scope(name or 'build_affine_surrogate_posterior'):
event_shape = nest.map_structure_up_to(
_get_event_shape_shallow_structure(event_shape),
lambda s: tf.convert_to_tensor(s, dtype=tf.int32), event_shape)
if nest.is_nested(bijector):
bijector = joint_map.JointMap(nest.map_structure(
lambda b: identity.Identity() if b is None else b, bijector),
validate_args=validate_args)
if bijector is None:
unconstrained_event_shape = event_shape
else:
unconstrained_event_shape = (
bijector.inverse_event_shape_tensor(event_shape))
standard_base_distribution = nest.map_structure(
lambda s: base_distribution(loc=tf.zeros([], dtype=dtype),
scale=1.), unconstrained_event_shape)
standard_base_distribution = nest.map_structure(
lambda d, s: ( # pylint: disable=g-long-lambda
sample.Sample(d, sample_shape=s, validate_args=validate_args)
if distribution_util.shape_may_be_nontrivial(s) else d),
standard_base_distribution,
unconstrained_event_shape)
if distribution_util.shape_may_be_nontrivial(batch_shape):
standard_base_distribution = nest.map_structure(
lambda d: batch_broadcast.BatchBroadcast( # pylint: disable=g-long-lambda
d,
to_shape=batch_shape,
validate_args=validate_args),
standard_base_distribution)
surrogate_posterior = yield from _affine_surrogate_posterior_from_base_distribution(
standard_base_distribution,
operators=operators,
bijector=bijector,
validate_args=validate_args)
return surrogate_posterior
def __init__(self,
loc=None,
scale=None,
validate_args=False,
allow_nan_stats=True,
experimental_use_kahan_sum=False,
name='MultivariateNormalLinearOperator'):
"""Construct Multivariate Normal distribution on `R^k`.
The `batch_shape` is the broadcast shape between `loc` and `scale`
arguments.
The `event_shape` is given by last dimension of the matrix implied by
`scale`. The last dimension of `loc` (if provided) must broadcast with this.
Recall that `covariance = scale @ scale.T`.
Additional leading dimensions (if any) will index batches.
Args:
loc: Floating-point `Tensor`. If this is set to `None`, `loc` is
implicitly `0`. When specified, may have shape `[B1, ..., Bb, k]` where
`b >= 0` and `k` is the event size.
scale: Instance of `LinearOperator` with same `dtype` as `loc` and shape
`[B1, ..., Bb, k, k]`.
validate_args: Python `bool`, default `False`. Whether to validate input
with asserts. If `validate_args` is `False`, and the inputs are
invalid, correct behavior is not guaranteed.
allow_nan_stats: Python `bool`, default `True`. If `False`, raise an
exception if a statistic (e.g. mean/mode/etc...) is undefined for any
batch member If `True`, batch members with valid parameters leading to
undefined statistics will return NaN for this statistic.
experimental_use_kahan_sum: Python `bool`. When `True`, we use Kahan
summation to aggregate independent underlying log_prob values. For best
results, Kahan summation should also be applied when computing the
log-determinant of the `LinearOperator` representing the scale matrix.
Kahan summation improves against the precision of a naive float32 sum.
This can be noticeable in particular for large dimensions in float32.
See CPU caveat on `tfp.math.reduce_kahan_sum`.
name: The name to give Ops created by the initializer.
Raises:
ValueError: if `scale` is unspecified.
TypeError: if not `scale.dtype.is_floating`
"""
parameters = dict(locals())
self._experimental_use_kahan_sum = experimental_use_kahan_sum
if scale is None:
raise ValueError('Missing required `scale` parameter.')
if not dtype_util.is_floating(scale.dtype):
raise TypeError('`scale` parameter must have floating-point dtype.')
with tf.name_scope(name) as name:
dtype = dtype_util.common_dtype([loc, scale], dtype_hint=tf.float32)
# Since expand_dims doesn't preserve constant-ness, we obtain the
# non-dynamic value if possible.
loc = tensor_util.convert_nonref_to_tensor(
loc, dtype=dtype, name='loc')
batch_shape, event_shape = distribution_util.shapes_from_loc_and_scale(
loc, scale)
self._loc = loc
self._scale = scale
bijector = scale_matvec_linear_operator.ScaleMatvecLinearOperator(
scale, validate_args=validate_args)
if loc is not None:
bijector = shift_bijector.Shift(
shift=loc, validate_args=validate_args)(bijector)
super(MultivariateNormalLinearOperator, self).__init__(
# TODO(b/137665504): Use batch-adding meta-distribution to set the batch
# shape instead of tf.zeros.
# We use `Sample` instead of `Independent` because `Independent`
# requires concatenating `batch_shape` and `event_shape`, which loses
# static `batch_shape` information when `event_shape` is not statically
# known.
distribution=sample.Sample(
normal.Normal(
loc=tf.zeros(batch_shape, dtype=dtype),
scale=tf.ones([], dtype=dtype)),
event_shape,
experimental_use_kahan_sum=experimental_use_kahan_sum),
bijector=bijector,
validate_args=validate_args,
name=name)
self._parameters = parameters
def posterior_generator():
prior_gen = prior._model_coroutine() # pylint: disable=protected-access
dist = next(prior_gen)
i = 0
try:
while True:
original_dist = dist.distribution if isinstance(
dist, Root) else dist
if isinstance(original_dist,
joint_distribution.JointDistribution):
# TODO(kateslin): Build inner JD surrogate in
# _make_asvi_trainable_variables to avoid rebuilding variables.
raise TypeError(
'Argument `prior` cannot be a nested `JointDistribution`.'
)
else:
original_dist = _as_trainable_family(original_dist)
try:
actual_dist = original_dist.distribution
except AttributeError:
actual_dist = original_dist
dist_params = actual_dist.parameters
temp_params_dict = {}
for param, value in dist_params.items():
if param in (
_NON_STATISTICAL_PARAMS +
_NON_TRAINABLE_PARAMS) or value is None:
temp_params_dict[param] = value
else:
prior_weight = param_dicts[i][
param].prior_weight
mean_field_parameter = param_dicts[i][
param].mean_field_parameter
if mean_field:
temp_params_dict[
param] = mean_field_parameter
else:
temp_params_dict[
param] = prior_weight * value + (
1. - prior_weight
) * mean_field_parameter
if isinstance(original_dist, sample.Sample):
surrogate_dist = sample.Sample(
type(actual_dist)(**temp_params_dict))
else:
surrogate_dist = type(actual_dist)(
**temp_params_dict)
if isinstance(
original_dist, transformed_distribution.
TransformedDistribution):
surrogate_dist = transformed_distribution.TransformedDistribution(
surrogate_dist,
bijector=original_dist.bijector)
if isinstance(original_dist, independent.Independent):
surrogate_dist = independent.Independent(
surrogate_dist,
reinterpreted_batch_ndims=original_dist.
reinterpreted_batch_ndims)
if isinstance(dist, Root):
value_out = yield Root(surrogate_dist)
else:
value_out = yield surrogate_dist
dist = prior_gen.send(value_out)
i += 1
except StopIteration:
pass
def __init__(self,
loc=None,
scale=None,
validate_args=False,
allow_nan_stats=True,
name='VectorExponentialLinearOperator'):
"""Construct Vector Exponential distribution supported on a subset of `R^k`.
The `batch_shape` is the broadcast shape between `loc` and `scale`
arguments.
The `event_shape` is given by last dimension of the matrix implied by
`scale`. The last dimension of `loc` (if provided) must broadcast with this.
Recall that `covariance = scale @ scale.T`.
Additional leading dimensions (if any) will index batches.
Args:
loc: Floating-point `Tensor`. If this is set to `None`, `loc` is
implicitly `0`. When specified, may have shape `[B1, ..., Bb, k]` where
`b >= 0` and `k` is the event size.
scale: Instance of `LinearOperator` with same `dtype` as `loc` and shape
`[B1, ..., Bb, k, k]`.
validate_args: Python `bool`, default `False`. Whether to validate input
with asserts. If `validate_args` is `False`, and the inputs are
invalid, correct behavior is not guaranteed.
allow_nan_stats: Python `bool`, default `True`. If `False`, raise an
exception if a statistic (e.g. mean/mode/etc...) is undefined for any
batch member If `True`, batch members with valid parameters leading to
undefined statistics will return NaN for this statistic.
name: The name to give Ops created by the initializer.
Raises:
ValueError: if `scale` is unspecified.
TypeError: if not `scale.dtype.is_floating`
"""
parameters = dict(locals())
if loc is None:
loc = 0.0 # Implicit value for backwards compatibility.
if scale is None:
raise ValueError('Missing required `scale` parameter.')
if not dtype_util.is_floating(scale.dtype):
raise TypeError(
'`scale` parameter must have floating-point dtype.')
with tf.name_scope(name) as name:
# Since expand_dims doesn't preserve constant-ness, we obtain the
# non-dynamic value if possible.
loc = loc if loc is None else tf.convert_to_tensor(
loc, name='loc', dtype=scale.dtype)
batch_shape, event_shape = distribution_util.shapes_from_loc_and_scale(
loc, scale)
self._loc = loc
self._scale = scale
super(VectorExponentialLinearOperator, self).__init__(
# TODO(b/137665504): Use batch-adding meta-distribution to set the
# batch shape instead of tf.ones.
# We use `Sample` instead of `Independent` because `Independent`
# requires concatenating `batch_shape` and `event_shape`, which loses
# static `batch_shape` information when `event_shape` is not
# statically known.
distribution=sample.Sample(
exponential.Exponential(rate=tf.ones(batch_shape,
dtype=scale.dtype),
allow_nan_stats=allow_nan_stats),
event_shape),
bijector=shift_bijector.Shift(shift=loc)(
scale_matvec_linear_operator.ScaleMatvecLinearOperator(
scale=scale, validate_args=validate_args)),
validate_args=validate_args,
name=name)
self._parameters = parameters
def __init__(self,
design_matrix,
nonzero_prior_prob=0.5,
weights_prior_precision=None,
default_pseudo_observations=1.,
observation_noise_variance_prior_concentration=0.005,
observation_noise_variance_prior_scale=0.0025,
observation_noise_variance_upper_bound=None,
num_missing=0.):
"""Initializes priors for the spike and slab sampler.
Args:
design_matrix: (batch of) float `Tensor`(s) regression design matrix (`X`
in [1]) having shape `[num_outputs, num_features]`.
nonzero_prior_prob: scalar float `Tensor` prior probability of the 'slab',
i.e., prior probability that any given feature has nonzero weight (`pi`
in [1]). Default value: `0.5`.
weights_prior_precision: (batch of) float `Tensor` complete prior
precision matrix(s) over the weights, of shape `[num_features,
num_features]`. If not specified, defaults to the Zellner g-prior
specified in `[1]` as `Omega^{-1} = kappa * (X'X + diag(X'X)) / (2 *
num_outputs)`, in which we've plugged in the suggested default of `w =
0.5`. The parameter `kappa` is controlled by the
`default_pseudo_observations` argument. Default value: `None`.
default_pseudo_observations: scalar float `Tensor` Controls the number of
pseudo-observations for the prior precision matrix over the weights.
Corresponds to `kappa` in [1]. See also `weights_prior_precision`.
observation_noise_variance_prior_concentration: scalar float `Tensor`
concentration parameter of the inverse gamma prior on the noise
variance. Corresponds to `nu / 2` in [1]. Default value: 0.005.
observation_noise_variance_prior_scale: scalar float `Tensor` scale
parameter of the inverse gamma prior on the noise variance. Corresponds
to `ss / 2` in [1]. Default value: 0.0025.
observation_noise_variance_upper_bound: optional scalar float `Tensor`
maximum value of sampled observation noise variance. Specifying a bound
can help avoid divergence when the sampler is initialized far from the
posterior. Default value: `None`.
num_missing: Optional scalar float `Tensor`. Corrects for how many missing
values are are coded as zero in the design matrix.
"""
with tf.name_scope('spike_slab_sampler'):
dtype = dtype_util.common_dtype([
design_matrix, nonzero_prior_prob, weights_prior_precision,
observation_noise_variance_prior_concentration,
observation_noise_variance_prior_scale,
observation_noise_variance_upper_bound, num_missing
],
dtype_hint=tf.float32)
design_matrix = tf.convert_to_tensor(design_matrix, dtype=dtype)
nonzero_prior_prob = tf.convert_to_tensor(nonzero_prior_prob,
dtype=dtype)
observation_noise_variance_prior_concentration = tf.convert_to_tensor(
observation_noise_variance_prior_concentration, dtype=dtype)
observation_noise_variance_prior_scale = tf.convert_to_tensor(
observation_noise_variance_prior_scale, dtype=dtype)
num_missing = tf.convert_to_tensor(num_missing, dtype=dtype)
if observation_noise_variance_upper_bound is not None:
observation_noise_variance_upper_bound = tf.convert_to_tensor(
observation_noise_variance_upper_bound, dtype=dtype)
design_shape = ps.shape(design_matrix)
num_outputs = tf.cast(design_shape[-2], dtype=dtype) - num_missing
num_features = design_shape[-1]
x_transpose_x = tf.matmul(design_matrix,
design_matrix,
adjoint_a=True)
if weights_prior_precision is None:
# Default prior: 'Zellner’s g−prior' from section 3.2.1 of [1]:
# `omega^{-1} = kappa * (w X'X + (1 − w) diag(X'X))/n`
# with default `w = 0.5`.
padded_inputs = broadcast_util.left_justified_expand_dims_like(
num_outputs, x_transpose_x)
weights_prior_precision = default_pseudo_observations * tf.linalg.set_diag(
0.5 * x_transpose_x,
tf.linalg.diag_part(x_transpose_x)) / padded_inputs
observation_noise_variance_posterior_concentration = (
observation_noise_variance_prior_concentration +
tf.convert_to_tensor(num_outputs / 2., dtype=dtype))
self.num_outputs = num_outputs
self.num_features = num_features
self.design_matrix = design_matrix
self.x_transpose_x = x_transpose_x
self.dtype = dtype
self.nonzeros_prior = sample_dist.Sample(
bernoulli.Bernoulli(probs=nonzero_prior_prob),
sample_shape=[num_features])
self.weights_prior_precision = weights_prior_precision
self.observation_noise_variance_prior_concentration = (
observation_noise_variance_prior_concentration)
self.observation_noise_variance_prior_scale = (
observation_noise_variance_prior_scale)
self.observation_noise_variance_upper_bound = (
observation_noise_variance_upper_bound)
self.observation_noise_variance_posterior_concentration = (
observation_noise_variance_posterior_concentration)
def _batched_isotropic_normal_like(state_part):
return sample.Sample(
normal.Normal(ps.zeros([], dtype=state_part.dtype), 1.),
ps.shape(state_part)[batch_rank:])
