def auto_assign_sampler(
model: Model,
sampler_type: Optional[str] = None,
):
"""
The toy implementation of sampler assigner
Parameters
----------
model : pymc4.Model
Model to sample posterior for
sampler_type : Optional[str]
The step method type for the model
Returns
-------
sampler_type : str
Sampler type name
"""
if sampler_type:
_log.info("Working with {} sampler".format(reg_samplers[sampler_type].__name__))
return sampler_type
_, _, free_disc_names, free_cont_names, _, _ = initialize_state(model)
if not free_disc_names:
_log.info("Auto-assigning NUTS sampler")
return "nuts"
else:
_log.info("The model contains discrete distributions. " "\nCompound step is chosen.")
return "compound"
def __init__(
self,
model: Model,
**kwargs,
):
if not isinstance(model, Model):
raise TypeError(
"`sample` function only supports `pymc4.Model` objects, but \
you've passed `{}`".format(type(model)))
_, _, disc_names, cont_names, _, _ = initialize_state(model)
# if sampler has the gradient calculation during `one_step`
# and the model contains discrete distributions then we throw the
# error.
if self._grad is True and disc_names:
raise ValueError("Discrete distributions can't be used with \
gradient-based sampler")
self.model = model
self.stat_names: List[str] = []
self.parent_inds: List[int] = []
# assign arguments from **kwargs to distinct kwargs for
# `kernel`, `adaptation_kernel`, `chain_sampler`
self._assign_arguments(kwargs)
# check arguments for correctness
self._check_arguments()
self._bound_kwargs()
def check_proposal_functions(
model: Model,
state: Optional[flow.SamplingState] = None,
observed: Optional[dict] = None,
) -> bool:
"""
Check for the non-default proposal generation functions
Parameters
----------
model : pymc4.Model
Model to sample posterior for
state : Optional[flow.SamplingState]
Current state
observed : Optional[Dict[str, Any]]
Observed values (optional)
"""
(_, state, _, _, continuous_distrs, discrete_distrs) = initialize_state(
model, observed=observed, state=state
)
init = state.all_unobserved_values
init_state = list(init.values())
init_keys = list(init.keys())
for i, state_part in enumerate(init_state):
untrs_var, unscoped_tr_var = scope_remove_transformed_part_if_required(
init_keys[i], state.transformed_values
)
# get the distribution for the random variable name
distr = continuous_distrs.get(untrs_var, None)
if distr is None:
distr = discrete_distrs[untrs_var]
func = distr._default_new_state_part
if callable(func):
return True
return False
def _assign_default_methods(
self,
*,
sampler_methods: Optional[List] = None,
state: Optional[flow.SamplingState] = None,
observed: Optional[dict] = None,
):
converted_sampler_methods: List = CompoundStep._convert_sampler_methods(
sampler_methods)
(_, state, _, _, continuous_distrs,
discrete_distrs) = initialize_state(self.model,
observed=observed,
state=state)
init = state.all_unobserved_values
init_state = list(init.values())
init_keys = list(init.keys())
# assignd samplers for free variables
make_kernel_fn: list = []
# user passed kwargs for each sampler in `make_kernel_fn`
part_kernel_kwargs: list = []
# keep the list for proposal func names
func_names: list = []
for i, state_part in enumerate(init_state):
untrs_var, unscoped_tr_var = scope_remove_transformed_part_if_required(
init_keys[i], state.transformed_values)
# get the distribution for the random variable name
distr = continuous_distrs.get(untrs_var, None)
if distr is None:
distr = discrete_distrs[untrs_var]
# get custom `new_state_fn` for the distribution
func = distr._default_new_state_part
# simplest way of assigning sampling methods
# if the sampler_methods was passed and if a var is provided
# then the var will be assigned to the given sampler
# but will also be checked if the sampler supports the distr
# 1. If sampler is provided by the user, we create new sampler
# and add to `make_kernel_fn`
# 2. If the distribution has `new_state_fn` then the new sampler
# should be create also. Because sampler is initialized with
# the `new_state_fn` argument.
if unscoped_tr_var in converted_sampler_methods:
sampler, kwargs = converted_sampler_methods[unscoped_tr_var]
# check for the sampler able to sampler from the distribution
if not distr._grad_support and sampler._grad:
raise ValueError(
"The `{}` doesn't support gradient, please provide an appropriate sampler method"
.format(unscoped_tr_var))
# add sampler to the dict
make_kernel_fn.append(sampler)
part_kernel_kwargs.append({})
# update with user provided kwargs
part_kernel_kwargs[-1].update(kwargs)
# if proposal function is provided then replace
func = part_kernel_kwargs[-1].get("new_state_fn", func)
# add the default `new_state_fn` for the distr
# `new_state_fn` is supported for only RandomWalkMetropolis transition
# kernel.
if func and sampler._name == "rwm":
part_kernel_kwargs[-1]["new_state_fn"] = partial(func)()
elif callable(func):
# If distribution has defined `new_state_fn` attribute then we need
# to assign `RandomWalkMetropolis` transition kernel
make_kernel_fn.append(RandomWalkM)
part_kernel_kwargs.append({})
part_kernel_kwargs[-1]["new_state_fn"] = partial(func)()
else:
# by default if user didn't not provide any sampler
# we choose NUTS for the variable with gradient and
# RWM for the variable without the gradient
sampler = NUTS if distr._grad_support else RandomWalkM
make_kernel_fn.append(sampler)
part_kernel_kwargs.append({})
# _log.info("Auto-assigning NUTS sampler...")
# save proposal func names
func_names.append(func._name if func else "default")
# `make_kernel_fn` contains (len(state)) sampler methods, this could lead
# to more overhed when we are iterating at each call of `one_step` in the
# compound step kernel. For that we need to merge some of the samplers.
kernels, set_lengths = self._merge_samplers(make_kernel_fn,
part_kernel_kwargs)
# log variable sampler mapping
CompoundStep._log_variables(init_keys, kernels, set_lengths,
self.parent_inds, func_names)
# save to use late for compound kernel init
self.kernel_kwargs["compound_samplers"] = kernels
self.kernel_kwargs["compound_set_lengths"] = set_lengths