def stochastic_node_mean_symbolic(approx: pm.MeanField,
node,
size=100,
more_replacements=None,
shape=None,
dtype=None):
"""Symbolic mean of a given PyMC3 stochastic node with respect to a given variational
posterior approximation.
Args:
approx: an instance of PyMC3 approximation
node: stochastic node
size: the number of samples to use for calculating the mean
more_replacements: (optional) an ordered dictionary of node replacements to be
applied to the computational graph before sampling
shape: (optional) shape of the node
dtype: (optional) dtype of the node
Raises:
ValueError: If `node.tag.test_value` is not present and `shape` and `dtype` are
not provided
Returns:
Symbolic approximate mean of the stochastic node
"""
assert size > 0
if shape is not None and dtype is not None:
cum_sum = tt.zeros(shape, dtype)
elif hasattr(node.tag, 'test_value') and node.tag.test_value is not None:
cum_sum = tt.zeros(node.tag.test_value.shape,
node.tag.test_value.dtype)
else:
raise ValueError(
"Can not determine the shape of the node to be sampled")
if more_replacements is not None:
node = th.clone(node, more_replacements, strict=False)
posterior_samples = approx.random(size)
node = approx.to_flat_input(node)
def add_sample_to_cum_sum(posterior_sample, _cum_sum):
new_sample = th.clone(node, {approx.input: posterior_sample},
strict=False)
return _cum_sum + tt.patternbroadcast(new_sample,
_cum_sum.broadcastable)
outputs, _ = th.scan(add_sample_to_cum_sum,
sequences=[posterior_samples],
outputs_info=[cum_sum],
n_steps=size)
return outputs[-1] / size
def stochastic_node_mean_symbolic(approx: pm.MeanField, node, size=100,
more_replacements=None, shape=None, dtype=None):
"""Symbolic mean of a given PyMC3 stochastic node with respect to a given variational
posterior approximation.
Args:
approx: an instance of PyMC3 approximation
node: stochastic node
size: the number of samples to use for calculating the mean
more_replacements: (optional) an ordered dictionary of node replacements to be
applied to the computational graph before sampling
shape: (optional) shape of the node
dtype: (optional) dtype of the node
Raises:
ValueError: If `node.tag.test_value` is not present and `shape` and `dtype` are
not provided
Returns:
Symbolic approximate mean of the stochastic node
"""
assert size > 0
if shape is not None and dtype is not None:
cum_sum = tt.zeros(shape, dtype)
elif hasattr(node.tag, 'test_value') and node.tag.test_value is not None:
cum_sum = tt.zeros(node.tag.test_value.shape, node.tag.test_value.dtype)
else:
raise ValueError("Can not determine the shape of the node to be sampled")
if more_replacements is not None:
node = th.clone(node, more_replacements, strict=False)
posterior_samples = approx.random(size)
node = approx.to_flat_input(node)
def add_sample_to_cum_sum(posterior_sample, _cum_sum):
new_sample = th.clone(node, {approx.input: posterior_sample}, strict=False)
return _cum_sum + tt.patternbroadcast(new_sample, _cum_sum.broadcastable)
outputs, _ = th.scan(add_sample_to_cum_sum,
sequences=[posterior_samples],
outputs_info=[cum_sum],
n_steps=size)
return outputs[-1] / size
def get_sampling_generator_for_model_approximation(model_approx: pm.MeanField, model_var_name: str,
num_samples: int = 250) -> Generator:
"""Get a generator that returns samples of a precomputed model approximation for a specific variable in that model
Args:
model_approx: an instance of PyMC3 meanfield approximation
model_var_name: a stochastic node in the model
num_samples: number of samples to draw
Returns:
A generator that will yield `num_samples` samples from an approximation to a posterior
"""
return (model_approx.sample()[model_var_name] for _ in range(num_samples))
def get_sampling_generator_for_model_approximation(model_approx: pm.MeanField, node,
num_samples: int = 20) -> Generator:
"""Get a generator that returns samples of a precomputed model approximation for a specific variable in that model
Args:
model_approx: an instance of PyMC3 mean-field approximation
node: a stochastic node in the model
num_samples: number of samples to draw
Returns:
A generator that will yield `num_samples` samples from an approximation to a posterior
"""
sample = model_approx.sample_node(node, size=1)[0]
return (sample.eval() for _ in range(num_samples))
def get_sampling_generator_for_model_approximation(model_approx: pm.MeanField, node,
num_samples: int = 20) -> Generator:
"""Get a generator that returns samples of a precomputed model approximation for a specific variable in that model
Args:
model_approx: an instance of PyMC3 mean-field approximation
node: a stochastic node in the model
num_samples: number of samples to draw
Returns:
A generator that will yield `num_samples` samples from an approximation to a posterior
"""
sample = model_approx.sample_node(node, size=1)[0]
return (sample.eval() for _ in range(num_samples))