def log_joint_fn(*args, **kwargs):
"""Log-probability of inputs according to a joint probability distribution.
Args:
*args: Positional arguments. They are the model's original inputs and can
alternatively be specified as part of `kwargs`.
**kwargs: Keyword arguments, where for each key-value pair `k` and `v`,
`v` is passed as a `value` to the random variable(s) whose keyword
argument `name` during construction is equal to `k`.
Returns:
Scalar tf.Tensor, which represents the model's log-probability summed
over all Edward random variables and their dimensions.
Raises:
TypeError: If a random variable in the model has no specified value in
`**kwargs`.
"""
log_probs = []
def interceptor(rv_constructor, *rv_args, **rv_kwargs):
"""Overrides a random variable's `value` and accumulates its log-prob."""
# Set value to keyword argument indexed by `name` (an input tensor).
rv_name = rv_kwargs.get("name")
if rv_name is None:
raise KeyError("Random variable constructor {} has no name "
"in its arguments.".format(
rv_constructor.__name__))
# If no value is explicitly passed in for an RV, default to the value
# from the RV constructor. This may have been set explicitly by the user
# or forwarded from a lower-level interceptor.
previously_specified_value = rv_kwargs.get("value")
value = kwargs.get(rv_name, previously_specified_value)
if value is None:
raise LookupError(
"Keyword argument specifying value for {} is "
"missing.".format(rv_name))
rv_kwargs["value"] = value
rv = rv_constructor(*rv_args, **rv_kwargs)
log_prob = tf.reduce_sum(
input_tensor=rv.distribution.log_prob(rv.value))
log_probs.append(log_prob)
return rv
model_kwargs = _get_function_inputs(model, kwargs)
with interception(interceptor):
model(*args, **model_kwargs)
log_prob = sum(log_probs)
return log_prob
def log_joint_fn(*args, **kwargs):
"""Log-probability of inputs according to a joint probability distribution.
Args:
*args: Positional arguments. They are the model's original inputs and can
alternatively be specified as part of `kwargs`.
**kwargs: Keyword arguments, where for each key-value pair `k` and `v`,
`v` is passed as a `value` to the random variable(s) whose keyword
argument `name` during construction is equal to `k`.
Returns:
Scalar tf.Tensor, which represents the model's log-probability summed
over all Edward random variables and their dimensions.
Raises:
TypeError: If a random variable in the model has no specified value in
`**kwargs`.
"""
log_probs = []
def interceptor(rv_constructor, *rv_args, **rv_kwargs):
"""Overrides a random variable's `value` and accumulates its log-prob."""
# Set value to keyword argument indexed by `name` (an input tensor).
rv_name = rv_kwargs.get('name')
if rv_name is None:
raise KeyError('Random variable constructor {} has no name '
'in its arguments.'.format(
rv_constructor.__name__))
value = kwargs.get(rv_name)
if value is None:
raise LookupError(
'Keyword argument specifying value for {} is '
'missing.'.format(rv_name))
rv_kwargs['value'] = value
rv = rv_constructor(*rv_args, **rv_kwargs)
log_prob = tf.reduce_sum(
input_tensor=rv.distribution.log_prob(rv.value))
log_probs.append(log_prob)
return rv
model_kwargs = _get_function_inputs(model, **kwargs)
with interception(interceptor):
try:
model(*args, **model_kwargs)
except TypeError as err:
raise Exception(
'Wrong number of arguments in log_joint function definition. {}'
.format(err))
log_prob = sum(log_probs)
return log_prob
def get_trace(model, *args, **kwargs):
trace_result = {}
def trace(rv_constructor, *rv_args, **rv_kwargs):
rv = interceptable(rv_constructor)(*rv_args, **rv_kwargs)
name = rv_kwargs['name']
trace_result[name] = rv.value
return rv
with interception(trace):
model(*args, **kwargs)
return trace_result
def variational_model(*args):
with interception(mean_field):
return model(*args)