https://github.com/pymc-devs/pymc3
Tip revision: 5c600c7117c98d296244c16f27b2fcdb0d37cd70 authored by Ricardo Vieira on 26 July 2023, 09:31:30 UTC
Use graph_replace instead of clone_replace in VI
Use graph_replace instead of clone_replace in VI
Tip revision: 5c600c7
model.py
# Copyright 2023 The PyMC Developers
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import functools
import sys
import threading
import types
import warnings
from sys import modules
from typing import (
TYPE_CHECKING,
Any,
Callable,
Dict,
List,
Literal,
Optional,
Sequence,
Tuple,
Type,
TypeVar,
Union,
cast,
)
import numpy as np
import pytensor
import pytensor.sparse as sparse
import pytensor.tensor as pt
import scipy.sparse as sps
from pytensor.compile import DeepCopyOp, get_mode
from pytensor.compile.sharedvalue import SharedVariable
from pytensor.graph.basic import Constant, Variable, graph_inputs
from pytensor.scalar import Cast
from pytensor.tensor.elemwise import Elemwise
from pytensor.tensor.random.op import RandomVariable
from pytensor.tensor.random.type import RandomType
from pytensor.tensor.sharedvar import ScalarSharedVariable
from pytensor.tensor.var import TensorConstant, TensorVariable
from typing_extensions import Self
from pymc.blocking import DictToArrayBijection, RaveledVars
from pymc.data import GenTensorVariable, is_minibatch
from pymc.distributions.transforms import _default_transform
from pymc.exceptions import (
BlockModelAccessError,
ImputationWarning,
SamplingError,
ShapeError,
ShapeWarning,
)
from pymc.initial_point import make_initial_point_fn
from pymc.logprob.basic import transformed_conditional_logp
from pymc.logprob.utils import ParameterValueError
from pymc.pytensorf import (
PointFunc,
SeedSequenceSeed,
compile_pymc,
convert_observed_data,
gradient,
hessian,
inputvars,
replace_rvs_by_values,
rewrite_pregrad,
)
from pymc.util import (
UNSET,
WithMemoization,
_add_future_warning_tag,
get_transformed_name,
get_value_vars_from_user_vars,
get_var_name,
treedict,
treelist,
)
from pymc.vartypes import continuous_types, discrete_types, typefilter
__all__ = [
"Model",
"modelcontext",
"Deterministic",
"Potential",
"set_data",
"Point",
"compile_fn",
]
T = TypeVar("T", bound="ContextMeta")
class ContextMeta(type):
"""Functionality for objects that put themselves in a context using
the `with` statement.
"""
def __new__(cls, name, bases, dct, **kwargs): # pylint: disable=unused-argument
"""Add __enter__ and __exit__ methods to the class."""
def __enter__(self):
self.__class__.context_class.get_contexts().append(self)
# self._pytensor_config is set in Model.__new__
self._config_context = None
if hasattr(self, "_pytensor_config"):
self._config_context = pytensor.config.change_flags(**self._pytensor_config)
self._config_context.__enter__()
return self
def __exit__(self, typ, value, traceback): # pylint: disable=unused-argument
self.__class__.context_class.get_contexts().pop()
# self._pytensor_config is set in Model.__new__
if self._config_context:
self._config_context.__exit__(typ, value, traceback)
dct[__enter__.__name__] = __enter__
dct[__exit__.__name__] = __exit__
# We strip off keyword args, per the warning from
# StackExchange:
# DO NOT send "**kwargs" to "type.__new__". It won't catch them and
# you'll get a "TypeError: type() takes 1 or 3 arguments" exception.
return super().__new__(cls, name, bases, dct)
# FIXME: is there a more elegant way to automatically add methods to the class that
# are instance methods instead of class methods?
def __init__(
cls, name, bases, nmspc, context_class: Optional[Type] = None, **kwargs
): # pylint: disable=unused-argument
"""Add ``__enter__`` and ``__exit__`` methods to the new class automatically."""
if context_class is not None:
cls._context_class = context_class
super().__init__(name, bases, nmspc)
def get_context(cls, error_if_none=True, allow_block_model_access=False) -> Optional[T]:
"""Return the most recently pushed context object of type ``cls``
on the stack, or ``None``. If ``error_if_none`` is True (default),
raise a ``TypeError`` instead of returning ``None``."""
try:
candidate: Optional[T] = cls.get_contexts()[-1]
except IndexError:
# Calling code expects to get a TypeError if the entity
# is unfound, and there's too much to fix.
if error_if_none:
raise TypeError(f"No {cls} on context stack")
return None
if isinstance(candidate, BlockModelAccess) and not allow_block_model_access:
raise BlockModelAccessError(candidate.error_msg_on_access)
return candidate
def get_contexts(cls) -> List[T]:
"""Return a stack of context instances for the ``context_class``
of ``cls``."""
# This lazily creates the context class's contexts
# thread-local object, as needed. This seems inelegant to me,
# but since the context class is not guaranteed to exist when
# the metaclass is being instantiated, I couldn't figure out a
# better way. [2019/10/11:rpg]
# no race-condition here, contexts is a thread-local object
# be sure not to override contexts in a subclass however!
context_class = cls.context_class
assert isinstance(
context_class, type
), f"Name of context class, {context_class} was not resolvable to a class"
if not hasattr(context_class, "contexts"):
context_class.contexts = threading.local()
contexts = context_class.contexts
if not hasattr(contexts, "stack"):
contexts.stack = []
return contexts.stack
# the following complex property accessor is necessary because the
# context_class may not have been created at the point it is
# specified, so the context_class may be a class *name* rather
# than a class.
@property
def context_class(cls) -> Type:
def resolve_type(c: Union[Type, str]) -> Type:
if isinstance(c, str):
c = getattr(modules[cls.__module__], c)
if isinstance(c, type):
return c
raise ValueError(f"Cannot resolve context class {c}")
assert cls is not None
if isinstance(cls._context_class, str):
cls._context_class = resolve_type(cls._context_class)
if not isinstance(cls._context_class, (str, type)):
raise ValueError(
f"Context class for {cls.__name__}, {cls._context_class}, is not of the right type"
)
return cls._context_class
# Inherit context class from parent
def __init_subclass__(cls, **kwargs):
super().__init_subclass__(**kwargs)
cls.context_class = super().context_class
# Initialize object in its own context...
# Merged from InitContextMeta in the original.
def __call__(cls, *args, **kwargs):
# We type hint Model here so type checkers understand that Model is a context manager.
# This metaclass is only used for Model, so this is safe to do. See #6809 for more info.
instance: "Model" = cls.__new__(cls, *args, **kwargs)
with instance: # appends context
instance.__init__(*args, **kwargs)
return instance
def modelcontext(model: Optional["Model"]) -> "Model":
"""
Return the given model or, if none was supplied, try to find one in
the context stack.
"""
if model is None:
model = Model.get_context(error_if_none=False)
if model is None:
# TODO: This should be a ValueError, but that breaks
# ArviZ (and others?), so might need a deprecation.
raise TypeError("No model on context stack.")
return model
class ValueGradFunction:
"""Create an PyTensor function that computes a value and its gradient.
Parameters
----------
costs: list of PyTensor variables
We compute the weighted sum of the specified PyTensor values, and the gradient
of that sum. The weights can be specified with `ValueGradFunction.set_weights`.
grad_vars: list of named PyTensor variables or None
The arguments with respect to which the gradient is computed.
extra_vars_and_values: dict of PyTensor variables and their initial values
Other arguments of the function that are assumed constant and their
values. They are stored in shared variables and can be set using
`set_extra_values`.
dtype: str, default=pytensor.config.floatX
The dtype of the arrays.
casting: {'no', 'equiv', 'save', 'same_kind', 'unsafe'}, default='no'
Casting rule for casting `grad_args` to the array dtype.
See `numpy.can_cast` for a description of the options.
Keep in mind that we cast the variables to the array *and*
back from the array dtype to the variable dtype.
compute_grads: bool, default=True
If False, return only the logp, not the gradient.
kwargs
Extra arguments are passed on to `pytensor.function`.
Attributes
----------
profile: PyTensor profiling object or None
The profiling object of the PyTensor function that computes value and
gradient. This is None unless `profile=True` was set in the
kwargs.
"""
def __init__(
self,
costs,
grad_vars,
extra_vars_and_values=None,
*,
dtype=None,
casting="no",
compute_grads=True,
**kwargs,
):
if extra_vars_and_values is None:
extra_vars_and_values = {}
names = [arg.name for arg in grad_vars + list(extra_vars_and_values.keys())]
if any(name is None for name in names):
raise ValueError("Arguments must be named.")
if len(set(names)) != len(names):
raise ValueError("Names of the arguments are not unique.")
self._grad_vars = grad_vars
self._extra_vars = list(extra_vars_and_values.keys())
self._extra_var_names = {var.name for var in extra_vars_and_values.keys()}
if dtype is None:
dtype = pytensor.config.floatX
self.dtype = dtype
self._n_costs = len(costs)
if self._n_costs == 0:
raise ValueError("At least one cost is required.")
weights = np.ones(self._n_costs - 1, dtype=self.dtype)
self._weights = pytensor.shared(weights, "__weights")
cost = costs[0]
for i, val in enumerate(costs[1:]):
if cost.ndim > 0 or val.ndim > 0:
raise ValueError("All costs must be scalar.")
cost = cost + self._weights[i] * val
self._extra_are_set = False
for var in self._grad_vars:
if not np.can_cast(var.dtype, self.dtype, casting):
raise TypeError(
f"Invalid dtype for variable {var.name}. Can not "
f"cast to {self.dtype} with casting rule {casting}."
)
if not np.issubdtype(var.dtype, np.floating):
raise TypeError(
f"Invalid dtype for variable {var.name}. Must be "
f"floating point but is {var.dtype}."
)
givens = []
self._extra_vars_shared = {}
for var, value in extra_vars_and_values.items():
shared = pytensor.shared(
value, var.name + "_shared__", shape=[1 if s == 1 else None for s in value.shape]
)
self._extra_vars_shared[var.name] = shared
givens.append((var, shared))
cost = rewrite_pregrad(cost)
if compute_grads:
grads = pytensor.grad(cost, grad_vars, disconnected_inputs="ignore")
for grad_wrt, var in zip(grads, grad_vars):
grad_wrt.name = f"{var.name}_grad"
outputs = [cost] + grads
else:
outputs = [cost]
inputs = grad_vars
self._pytensor_function = compile_pymc(inputs, outputs, givens=givens, **kwargs)
def set_weights(self, values):
if values.shape != (self._n_costs - 1,):
raise ValueError("Invalid shape. Must be (n_costs - 1,).")
self._weights.set_value(values)
def set_extra_values(self, extra_vars):
self._extra_are_set = True
for var in self._extra_vars:
self._extra_vars_shared[var.name].set_value(extra_vars[var.name])
def get_extra_values(self):
if not self._extra_are_set:
raise ValueError("Extra values are not set.")
return {var.name: self._extra_vars_shared[var.name].get_value() for var in self._extra_vars}
def __call__(self, grad_vars, grad_out=None, extra_vars=None):
if extra_vars is not None:
self.set_extra_values(extra_vars)
if not self._extra_are_set:
raise ValueError("Extra values are not set.")
if isinstance(grad_vars, RaveledVars):
grad_vars = list(DictToArrayBijection.rmap(grad_vars).values())
cost, *grads = self._pytensor_function(*grad_vars)
if grads:
grads_raveled = DictToArrayBijection.map(
{v.name: gv for v, gv in zip(self._grad_vars, grads)}
)
if grad_out is None:
return cost, grads_raveled.data
else:
np.copyto(grad_out, grads_raveled.data)
return cost
else:
return cost
@property
def profile(self):
"""Profiling information of the underlying PyTensor function."""
return self._pytensor_function.profile
class Model(WithMemoization, metaclass=ContextMeta):
"""Encapsulates the variables and likelihood factors of a model.
Model class can be used for creating class based models. To create
a class based model you should inherit from :class:`~pymc.Model` and
override the `__init__` method with arbitrary definitions (do not
forget to call base class :meth:`pymc.Model.__init__` first).
Parameters
----------
name: str
name that will be used as prefix for names of all random
variables defined within model
check_bounds: bool
Ensure that input parameters to distributions are in a valid
range. If your model is built in a way where you know your
parameters can only take on valid values you can set this to
False for increased speed. This should not be used if your model
contains discrete variables.
Examples
--------
How to define a custom model
.. code-block:: python
class CustomModel(Model):
# 1) override init
def __init__(self, mean=0, sigma=1, name=''):
# 2) call super's init first, passing model and name
# to it name will be prefix for all variables here if
# no name specified for model there will be no prefix
super().__init__(name, model)
# now you are in the context of instance,
# `modelcontext` will return self you can define
# variables in several ways note, that all variables
# will get model's name prefix
# 3) you can create variables with the register_rv method
self.register_rv(Normal.dist(mu=mean, sigma=sigma), 'v1', initval=1)
# this will create variable named like '{name::}v1'
# and assign attribute 'v1' to instance created
# variable can be accessed with self.v1 or self['v1']
# 4) this syntax will also work as we are in the
# context of instance itself, names are given as usual
Normal('v2', mu=mean, sigma=sigma)
# something more complex is allowed, too
half_cauchy = HalfCauchy('sigma', beta=10, initval=1.)
Normal('v3', mu=mean, sigma=half_cauchy)
# Deterministic variables can be used in usual way
Deterministic('v3_sq', self.v3 ** 2)
# Potentials too
Potential('p1', pt.constant(1))
# After defining a class CustomModel you can use it in several
# ways
# I:
# state the model within a context
with Model() as model:
CustomModel()
# arbitrary actions
# II:
# use new class as entering point in context
with CustomModel() as model:
Normal('new_normal_var', mu=1, sigma=0)
# III:
# just get model instance with all that was defined in it
model = CustomModel()
# IV:
# use many custom models within one context
with Model() as model:
CustomModel(mean=1, name='first')
CustomModel(mean=2, name='second')
# variables inside both scopes will be named like `first::*`, `second::*`
"""
if TYPE_CHECKING:
def __enter__(self: Self) -> Self:
...
def __exit__(self, exc_type: None, exc_val: None, exc_tb: None) -> None:
...
def __new__(cls, *args, **kwargs):
# resolves the parent instance
instance = super().__new__(cls)
if kwargs.get("model") is not None:
instance._parent = kwargs.get("model")
else:
instance._parent = cls.get_context(error_if_none=False)
instance._pytensor_config = kwargs.get("pytensor_config", {})
return instance
@staticmethod
def _validate_name(name):
if name.endswith(":"):
raise KeyError("name should not end with `:`")
return name
def __init__(
self,
name="",
coords=None,
check_bounds=True,
*,
coords_mutable=None,
pytensor_config=None,
model=None,
):
del pytensor_config, model # used in __new__
self.name = self._validate_name(name)
self.check_bounds = check_bounds
if self.parent is not None:
self.named_vars = treedict(parent=self.parent.named_vars)
self.named_vars_to_dims = treedict(parent=self.parent.named_vars_to_dims)
self.values_to_rvs = treedict(parent=self.parent.values_to_rvs)
self.rvs_to_values = treedict(parent=self.parent.rvs_to_values)
self.rvs_to_transforms = treedict(parent=self.parent.rvs_to_transforms)
self.rvs_to_initial_values = treedict(parent=self.parent.rvs_to_initial_values)
self.free_RVs = treelist(parent=self.parent.free_RVs)
self.observed_RVs = treelist(parent=self.parent.observed_RVs)
self.deterministics = treelist(parent=self.parent.deterministics)
self.potentials = treelist(parent=self.parent.potentials)
self._coords = self.parent._coords
self._dim_lengths = self.parent._dim_lengths
else:
self.named_vars = treedict()
self.named_vars_to_dims = treedict()
self.values_to_rvs = treedict()
self.rvs_to_values = treedict()
self.rvs_to_transforms = treedict()
self.rvs_to_initial_values = treedict()
self.free_RVs = treelist()
self.observed_RVs = treelist()
self.deterministics = treelist()
self.potentials = treelist()
self._coords = {}
self._dim_lengths = {}
self.add_coords(coords)
if coords_mutable is not None:
for name, values in coords_mutable.items():
self.add_coord(name, values, mutable=True)
from pymc.printing import str_for_model
self.str_repr = types.MethodType(str_for_model, self)
self._repr_latex_ = types.MethodType(
functools.partial(str_for_model, formatting="latex"), self
)
@property
def model(self):
return self
@property
def parent(self):
return self._parent
@property
def root(self):
model = self
while not model.isroot:
model = model.parent
return model
@property
def isroot(self):
return self.parent is None
def logp_dlogp_function(self, grad_vars=None, tempered=False, **kwargs):
"""Compile an PyTensor function that computes logp and gradient.
Parameters
----------
grad_vars: list of random variables, optional
Compute the gradient with respect to those variables. If None,
use all free random variables of this model.
tempered: bool
Compute the tempered logp `free_logp + alpha * observed_logp`.
`alpha` can be changed using `ValueGradFunction.set_weights([alpha])`.
"""
if grad_vars is None:
grad_vars = self.continuous_value_vars
else:
grad_vars = get_value_vars_from_user_vars(grad_vars, self)
for i, var in enumerate(grad_vars):
if var.dtype not in continuous_types:
raise ValueError(f"Can only compute the gradient of continuous types: {var}")
if tempered:
costs = [self.varlogp, self.datalogp]
else:
costs = [self.logp()]
input_vars = {i for i in graph_inputs(costs) if not isinstance(i, Constant)}
ip = self.initial_point(0)
extra_vars_and_values = {
var: ip[var.name]
for var in self.value_vars
if var in input_vars and var not in grad_vars
}
return ValueGradFunction(costs, grad_vars, extra_vars_and_values, **kwargs)
def compile_logp(
self,
vars: Optional[Union[Variable, Sequence[Variable]]] = None,
jacobian: bool = True,
sum: bool = True,
) -> PointFunc:
"""Compiled log probability density function.
Parameters
----------
vars: list of random variables or potential terms, optional
Compute the gradient with respect to those variables. If None, use all
free and observed random variables, as well as potential terms in model.
jacobian:
Whether to include jacobian terms in logprob graph. Defaults to True.
sum:
Whether to sum all logp terms or return elemwise logp for each variable.
Defaults to True.
"""
return self.model.compile_fn(self.logp(vars=vars, jacobian=jacobian, sum=sum))
def compile_dlogp(
self,
vars: Optional[Union[Variable, Sequence[Variable]]] = None,
jacobian: bool = True,
) -> PointFunc:
"""Compiled log probability density gradient function.
Parameters
----------
vars: list of random variables or potential terms, optional
Compute the gradient with respect to those variables. If None, use all
free and observed random variables, as well as potential terms in model.
jacobian:
Whether to include jacobian terms in logprob graph. Defaults to True.
"""
return self.model.compile_fn(self.dlogp(vars=vars, jacobian=jacobian))
def compile_d2logp(
self,
vars: Optional[Union[Variable, Sequence[Variable]]] = None,
jacobian: bool = True,
) -> PointFunc:
"""Compiled log probability density hessian function.
Parameters
----------
vars: list of random variables or potential terms, optional
Compute the gradient with respect to those variables. If None, use all
free and observed random variables, as well as potential terms in model.
jacobian:
Whether to include jacobian terms in logprob graph. Defaults to True.
"""
return self.model.compile_fn(self.d2logp(vars=vars, jacobian=jacobian))
def logp(
self,
vars: Optional[Union[Variable, Sequence[Variable]]] = None,
jacobian: bool = True,
sum: bool = True,
) -> Union[Variable, List[Variable]]:
"""Elemwise log-probability of the model.
Parameters
----------
vars: list of random variables or potential terms, optional
Compute the gradient with respect to those variables. If None, use all
free and observed random variables, as well as potential terms in model.
jacobian:
Whether to include jacobian terms in logprob graph. Defaults to True.
sum:
Whether to sum all logp terms or return elemwise logp for each variable.
Defaults to True.
Returns
-------
Logp graph(s)
"""
varlist: List[TensorVariable]
if vars is None:
varlist = self.free_RVs + self.observed_RVs + self.potentials
elif not isinstance(vars, (list, tuple)):
varlist = [vars]
else:
varlist = cast(List[TensorVariable], vars)
# We need to separate random variables from potential terms, and remember their
# original order so that we can merge them together in the same order at the end
rvs = []
potentials = []
rv_order, potential_order = [], []
for i, var in enumerate(varlist):
rv = self.values_to_rvs.get(var, var)
if rv in self.basic_RVs:
rvs.append(rv)
rv_order.append(i)
else:
if var in self.potentials:
potentials.append(var)
potential_order.append(i)
else:
raise ValueError(
f"Requested variable {var} not found among the model variables"
)
rv_logps: List[TensorVariable] = []
if rvs:
rv_logps = transformed_conditional_logp(
rvs=rvs,
rvs_to_values=self.rvs_to_values,
rvs_to_transforms=self.rvs_to_transforms,
jacobian=jacobian,
)
assert isinstance(rv_logps, list)
# Replace random variables by their value variables in potential terms
potential_logps = []
if potentials:
potential_logps = self.replace_rvs_by_values(potentials)
logp_factors = [None] * len(varlist)
for logp_order, logp in zip((rv_order + potential_order), (rv_logps + potential_logps)):
logp_factors[logp_order] = logp
if not sum:
return logp_factors
logp_scalar = pt.sum([pt.sum(factor) for factor in logp_factors])
logp_scalar_name = "__logp" if jacobian else "__logp_nojac"
if self.name:
logp_scalar_name = f"{logp_scalar_name}_{self.name}"
logp_scalar.name = logp_scalar_name
return logp_scalar
def dlogp(
self,
vars: Optional[Union[Variable, Sequence[Variable]]] = None,
jacobian: bool = True,
) -> Variable:
"""Gradient of the models log-probability w.r.t. ``vars``.
Parameters
----------
vars: list of random variables or potential terms, optional
Compute the gradient with respect to those variables. If None, use all
free and observed random variables, as well as potential terms in model.
jacobian:
Whether to include jacobian terms in logprob graph. Defaults to True.
Returns
-------
dlogp graph
"""
if vars is None:
value_vars = None
else:
if not isinstance(vars, (list, tuple)):
vars = [vars]
value_vars = []
for i, var in enumerate(vars):
value_var = self.rvs_to_values.get(var)
if value_var is not None:
value_vars.append(value_var)
else:
raise ValueError(
f"Requested variable {var} not found among the model variables"
)
cost = self.logp(jacobian=jacobian)
cost = rewrite_pregrad(cost)
return gradient(cost, value_vars)
def d2logp(
self,
vars: Optional[Union[Variable, Sequence[Variable]]] = None,
jacobian: bool = True,
) -> Variable:
"""Hessian of the models log-probability w.r.t. ``vars``.
Parameters
----------
vars: list of random variables or potential terms, optional
Compute the gradient with respect to those variables. If None, use all
free and observed random variables, as well as potential terms in model.
jacobian:
Whether to include jacobian terms in logprob graph. Defaults to True.
Returns
-------
d²logp graph
"""
if vars is None:
value_vars = None
else:
if not isinstance(vars, (list, tuple)):
vars = [vars]
value_vars = []
for i, var in enumerate(vars):
value_var = self.rvs_to_values.get(var)
if value_var is not None:
value_vars.append(value_var)
else:
raise ValueError(
f"Requested variable {var} not found among the model variables"
)
cost = self.logp(jacobian=jacobian)
cost = rewrite_pregrad(cost)
return hessian(cost, value_vars)
@property
def datalogp(self) -> Variable:
"""PyTensor scalar of log-probability of the observed variables and
potential terms"""
return self.observedlogp + self.potentiallogp
@property
def varlogp(self) -> Variable:
"""PyTensor scalar of log-probability of the unobserved random variables
(excluding deterministic)."""
return self.logp(vars=self.free_RVs)
@property
def varlogp_nojac(self) -> Variable:
"""PyTensor scalar of log-probability of the unobserved random variables
(excluding deterministic) without jacobian term."""
return self.logp(vars=self.free_RVs, jacobian=False)
@property
def observedlogp(self) -> Variable:
"""PyTensor scalar of log-probability of the observed variables"""
return self.logp(vars=self.observed_RVs)
@property
def potentiallogp(self) -> Variable:
"""PyTensor scalar of log-probability of the Potential terms"""
# Convert random variables in Potential expression into their log-likelihood
# inputs and apply their transforms, if any
potentials = self.replace_rvs_by_values(self.potentials)
if potentials:
return pt.sum([pt.sum(factor) for factor in potentials])
else:
return pt.constant(0.0)
@property
def value_vars(self):
"""List of unobserved random variables used as inputs to the model's
log-likelihood (which excludes deterministics).
"""
return [self.rvs_to_values[v] for v in self.free_RVs]
@property
def unobserved_value_vars(self):
"""List of all random variables (including untransformed projections),
as well as deterministics used as inputs and outputs of the model's
log-likelihood graph
"""
vars = []
transformed_rvs = []
for rv in self.free_RVs:
value_var = self.rvs_to_values[rv]
transform = self.rvs_to_transforms[rv]
if transform is not None:
transformed_rvs.append(rv)
vars.append(value_var)
# Remove rvs from untransformed values graph
untransformed_vars = self.replace_rvs_by_values(transformed_rvs)
# Remove rvs from deterministics graph
deterministics = self.replace_rvs_by_values(self.deterministics)
return vars + untransformed_vars + deterministics
@property
def disc_vars(self):
warnings.warn(
"Model.disc_vars has been deprecated. Use Model.discrete_value_vars instead.",
FutureWarning,
)
return self.discrete_value_vars
@property
def discrete_value_vars(self):
"""All the discrete value variables in the model"""
return list(typefilter(self.value_vars, discrete_types))
@property
def cont_vars(self):
warnings.warn(
"Model.cont_vars has been deprecated. Use Model.continuous_value_vars instead.",
FutureWarning,
)
return self.continuous_value_vars
@property
def continuous_value_vars(self):
"""All the continuous value variables in the model"""
return list(typefilter(self.value_vars, continuous_types))
@property
def basic_RVs(self):
"""List of random variables the model is defined in terms of
(which excludes deterministics).
These are the actual random variable terms that make up the
"sample-space" graph (i.e. you can sample these graphs by compiling them
with `pytensor.function`). If you want the corresponding log-likelihood terms,
use `model.value_vars` instead.
"""
return self.free_RVs + self.observed_RVs
@property
def unobserved_RVs(self):
"""List of all random variables, including deterministic ones.
These are the actual random variable terms that make up the
"sample-space" graph (i.e. you can sample these graphs by compiling them
with `pytensor.function`). If you want the corresponding log-likelihood terms,
use `var.unobserved_value_vars` instead.
"""
return self.free_RVs + self.deterministics
@property
def RV_dims(self) -> Dict[str, Tuple[Union[str, None], ...]]:
"""Tuples of dimension names for specific model variables.
Entries in the tuples may be ``None``, if the RV dimension was not given a name.
"""
warnings.warn(
"Model.RV_dims is deprecated. User Model.named_vars_to_dims instead.",
FutureWarning,
)
return self.named_vars_to_dims
@property
def coords(self) -> Dict[str, Union[Tuple, None]]:
"""Coordinate values for model dimensions."""
return self._coords
@property
def dim_lengths(self) -> Dict[str, Variable]:
"""The symbolic lengths of dimensions in the model.
The values are typically instances of ``TensorVariable`` or ``ScalarSharedVariable``.
"""
return self._dim_lengths
def shape_from_dims(self, dims):
shape = []
if len(set(dims)) != len(dims):
raise ValueError("Can not contain the same dimension name twice.")
for dim in dims:
if dim not in self.coords:
raise ValueError(
f"Unknown dimension name '{dim}'. All dimension "
"names must be specified in the `coords` "
"argument of the model or through a pm.Data "
"variable."
)
shape.extend(np.shape(self.coords[dim]))
return tuple(shape)
def add_coord(
self,
name: str,
values: Optional[Sequence] = None,
mutable: bool = False,
*,
length: Optional[Union[int, Variable]] = None,
):
"""Registers a dimension coordinate with the model.
Parameters
----------
name : str
Name of the dimension.
Forbidden: {"chain", "draw", "__sample__"}
values : optional, array-like
Coordinate values or ``None`` (for auto-numbering).
If ``None`` is passed, a ``length`` must be specified.
mutable : bool
Whether the created dimension should be resizable.
Default is False.
length : optional, scalar
A scalar of the dimensions length.
Defaults to ``pytensor.tensor.constant(len(values))``.
"""
if name in {"draw", "chain", "__sample__"}:
raise ValueError(
"Dimensions can not be named `draw`, `chain` or `__sample__`, "
"as those are reserved for use in `InferenceData`."
)
if values is None and length is None:
raise ValueError(
f"Either `values` or `length` must be specified for the '{name}' dimension."
)
if values is not None:
# Conversion to a tuple ensures that the coordinate values are immutable.
# Also unlike numpy arrays the's tuple.index(...) which is handy to work with.
values = tuple(values)
if name in self.coords:
if not np.array_equal(values, self.coords[name]):
raise ValueError(f"Duplicate and incompatible coordinate: {name}.")
if length is not None and not isinstance(length, (int, Variable)):
raise ValueError(
f"The `length` passed for the '{name}' coord must be an int, PyTensor Variable or None."
)
if length is None:
length = len(values)
if not isinstance(length, Variable):
if mutable:
length = pytensor.shared(length, name=name)
else:
length = pytensor.tensor.constant(length)
self._dim_lengths[name] = length
self._coords[name] = values
def add_coords(
self,
coords: Dict[str, Optional[Sequence]],
*,
lengths: Optional[Dict[str, Optional[Union[int, Variable]]]] = None,
):
"""Vectorized version of ``Model.add_coord``."""
if coords is None:
return
lengths = lengths or {}
for name, values in coords.items():
self.add_coord(name, values, length=lengths.get(name, None))
def set_dim(self, name: str, new_length: int, coord_values: Optional[Sequence] = None):
"""Update a mutable dimension.
Parameters
----------
name
Name of the dimension.
new_length
New length of the dimension.
coord_values
Optional sequence of coordinate values.
"""
if not isinstance(self.dim_lengths[name], ScalarSharedVariable):
raise ValueError(f"The dimension '{name}' is immutable.")
if coord_values is None and self.coords.get(name, None) is not None:
raise ValueError(
f"'{name}' has coord values. Pass `set_dim(..., coord_values=...)` to update them."
)
if coord_values is not None:
len_cvals = len(coord_values)
if len_cvals != new_length:
raise ShapeError(
"Length of new coordinate values does not match the new dimension length.",
actual=len_cvals,
expected=new_length,
)
self._coords[name] = tuple(coord_values)
self.dim_lengths[name].set_value(new_length)
return
def initial_point(self, random_seed: SeedSequenceSeed = None) -> Dict[str, np.ndarray]:
"""Computes the initial point of the model.
Parameters
----------
random_seed : SeedSequenceSeed, default None
Seed(s) for generating initial point from the model. Passed into :func:`pymc.pytensorf.reseed_rngs`
Returns
-------
ip : dict of {str : array_like}
Maps names of transformed variables to numeric initial values in the transformed space.
"""
fn = make_initial_point_fn(model=self, return_transformed=True)
return Point(fn(random_seed), model=self)
@property
def initial_values(self) -> Dict[TensorVariable, Optional[Union[np.ndarray, Variable, str]]]:
"""Maps transformed variables to initial value placeholders.
Keys are the random variables (as returned by e.g. ``pm.Uniform()``) and
values are the numeric/symbolic initial values, strings denoting the strategy to get them, or None.
"""
warnings.warn(
"Model.initial_values is deprecated. Use Model.rvs_to_initial_values instead."
)
return self.rvs_to_initial_values
def set_initval(self, rv_var, initval):
"""Sets an initial value (strategy) for a random variable."""
if initval is not None and not isinstance(initval, (Variable, str)):
# Convert scalars or array-like inputs to ndarrays
initval = rv_var.type.filter(initval)
self.rvs_to_initial_values[rv_var] = initval
def set_data(
self,
name: str,
values: Union[Sequence, np.ndarray],
coords: Optional[Dict[str, Sequence]] = None,
):
"""Changes the values of a data variable in the model.
In contrast to pm.MutableData().set_value, this method can also
update the corresponding coordinates.
Parameters
----------
name : str
Name of a shared variable in the model.
values : array-like
New values for the shared variable.
coords : optional, dict
New coordinate values for dimensions of the shared variable.
Must be provided for all named dimensions that change in length
and already have coordinate values.
"""
shared_object = self[name]
if not isinstance(shared_object, SharedVariable):
raise TypeError(
f"The variable `{name}` must be a `SharedVariable`"
" (created through `pm.MutableData()` or `pm.Data(mutable=True)`) to allow updating. "
f"The current type is: {type(shared_object)}"
)
if isinstance(values, list):
values = np.array(values)
values = convert_observed_data(values)
dims = self.named_vars_to_dims.get(name, None) or ()
coords = coords or {}
if values.ndim != shared_object.ndim:
raise ValueError(
f"New values for '{name}' must have {shared_object.ndim} dimensions, just like the original."
)
for d, dname in enumerate(dims):
length_tensor = self.dim_lengths[dname]
old_length = length_tensor.eval()
new_length = values.shape[d]
original_coords = self.coords.get(dname, None)
new_coords = coords.get(dname, None)
length_changed = new_length != old_length
# Reject resizing if we already know that it would create shape problems.
# NOTE: If there are multiple pm.MutableData containers sharing this dim, but the user only
# changes the values for one of them, they will run into shape problems nonetheless.
if length_changed:
if original_coords is not None:
if new_coords is None:
raise ValueError(
f"The '{name}' variable already had {len(original_coords)} coord values defined for "
f"its {dname} dimension. With the new values this dimension changes to length "
f"{new_length}, so new coord values for the {dname} dimension are required."
)
if isinstance(length_tensor, TensorConstant):
# The dimension was fixed in length.
# Resizing a data variable in this dimension would
# definitely lead to shape problems.
raise ShapeError(
f"Resizing dimension '{dname}' is impossible, because "
"a `TensorConstant` stores its length. To be able "
"to change the dimension length, pass `mutable=True` when "
"registering the dimension via `model.add_coord`, "
"or define it via a `pm.MutableData` variable."
)
elif length_tensor.owner is not None:
# The dimension was created from another variable:
length_tensor_origin = length_tensor.owner.inputs[0]
# Get a handle on the tensor from which this dimension length was
# obtained by doing subindexing on the shape as in `.shape[i]`.
if isinstance(length_tensor_origin, TensorConstant):
raise ShapeError(
f"Resizing dimension '{dname}' with values of length {new_length} would lead to incompatibilities, "
f"because the dimension length is tied to a TensorConstant. "
f"Check if the dimension was defined implicitly before the shared variable '{name}' was created, "
f"for example by another model variable.",
actual=new_length,
expected=old_length,
)
# The shape entry this dimension is tied to is not a TensorConstant.
# Whether the dimension can be resized depends on the kind of Variable the shape belongs to.
# TODO: Consider checking the graph is what we are assuming it is
# isinstance(length_tensor.owner.op, Subtensor)
# isinstance(length_tensor.owner.inputs[0].owner.op, Shape)
length_belongs_to = length_tensor_origin.owner.inputs[0]
if length_belongs_to is shared_object:
# This is the shared variable that's being updated!
# No surprise it's changing.
pass
elif isinstance(length_belongs_to, SharedVariable):
# The dimension is mutable through a SharedVariable other than the one being modified.
# But the other variable was not yet re-sized! Warn the user to do that!
warnings.warn(
f"You are resizing a variable with dimension '{dname}' which was initialized "
f"as a mutable dimension by another variable ('{length_belongs_to}')."
" Remember to update that variable with the correct shape to avoid shape issues.",
ShapeWarning,
stacklevel=2,
)
else:
# The dimension is immutable.
raise ShapeError(
f"Resizing dimension '{dname}' with values of length {new_length} would lead to incompatibilities, "
f"because the dimension was initialized from '{length_belongs_to}' which is not a shared variable. "
f"Check if the dimension was defined implicitly before the shared variable '{name}' was created, "
f"for example by another model variable.",
actual=new_length,
expected=old_length,
)
if isinstance(length_tensor, ScalarSharedVariable):
# The dimension is mutable, but was defined without being linked
# to a shared variable. This is allowed, but a little less robust.
self.set_dim(dname, new_length, coord_values=new_coords)
if new_coords is not None:
# Update the registered coord values (also if they were None)
if len(new_coords) != new_length:
raise ShapeError(
f"Length of new coordinate values for dimension '{dname}' does not match the provided values.",
actual=len(new_coords),
expected=new_length,
)
# store it as tuple for immutability as in add_coord
self._coords[dname] = tuple(new_coords)
shared_object.set_value(values)
def register_rv(
self, rv_var, name, observed=None, total_size=None, dims=None, transform=UNSET, initval=None
):
"""Register an (un)observed random variable with the model.
Parameters
----------
rv_var: TensorVariable
name: str
Intended name for the model variable.
observed: array_like (optional)
Data values for observed variables.
total_size: scalar
upscales logp of variable with ``coef = total_size/var.shape[0]``
dims: tuple
Dimension names for the variable.
transform
A transform for the random variable in log-likelihood space.
initval
The initial value of the random variable.
Returns
-------
TensorVariable
"""
name = self.name_for(name)
rv_var.name = name
_add_future_warning_tag(rv_var)
# Associate previously unknown dimension names with
# the length of the corresponding RV dimension.
if dims is not None:
for d, dname in enumerate(dims):
if not isinstance(dname, str):
raise TypeError(f"Dims must be string. Got {dname} of type {type(dname)}")
if dname not in self.dim_lengths:
self.add_coord(dname, values=None, length=rv_var.shape[d])
if observed is None:
if total_size is not None:
raise ValueError("total_size can only be passed to observed RVs")
self.free_RVs.append(rv_var)
self.create_value_var(rv_var, transform)
self.add_named_variable(rv_var, dims)
self.set_initval(rv_var, initval)
else:
if (
isinstance(observed, Variable)
and not isinstance(observed, GenTensorVariable)
and observed.owner is not None
# The only PyTensor operation we allow on observed data is type casting
# Although we could allow for any graph that does not depend on other RVs
and not (
isinstance(observed.owner.op, Elemwise)
and isinstance(observed.owner.op.scalar_op, Cast)
)
and not is_minibatch(observed)
):
raise TypeError(
"Variables that depend on other nodes cannot be used for observed data."
f"The data variable was: {observed}"
)
# `rv_var` is potentially changed by `make_obs_var`,
# for example into a new graph for imputation of missing data.
rv_var = self.make_obs_var(rv_var, observed, dims, transform, total_size)
return rv_var
def make_obs_var(
self,
rv_var: TensorVariable,
data: np.ndarray,
dims,
transform: Union[Any, None],
total_size: Union[int, None],
) -> TensorVariable:
"""Create a `TensorVariable` for an observed random variable.
Parameters
----------
rv_var
The random variable that is observed.
Its dimensionality must be compatible with the data already.
data
The observed data.
dims: tuple
Dimension names for the variable.
transform
A transform for the random variable in log-likelihood space.
"""
name = rv_var.name
data = convert_observed_data(data).astype(rv_var.dtype)
if data.ndim != rv_var.ndim:
raise ShapeError(
"Dimensionality of data and RV don't match.", actual=data.ndim, expected=rv_var.ndim
)
if pytensor.config.compute_test_value != "off":
test_value = getattr(rv_var.tag, "test_value", None)
if test_value is not None:
# We try to reuse the old test value
rv_var.tag.test_value = np.broadcast_to(test_value, rv_var.shape)
else:
rv_var.tag.test_value = data
mask = getattr(data, "mask", None)
if mask is not None:
impute_message = (
f"Data in {rv_var} contains missing values and"
" will be automatically imputed from the"
" sampling distribution."
)
warnings.warn(impute_message, ImputationWarning)
if total_size is not None:
raise ValueError("total_size is not compatible with imputed variables")
from pymc.distributions.distribution import create_partial_observed_rv
(
(observed_rv, observed_mask),
(unobserved_rv, _),
joined_rv,
) = create_partial_observed_rv(rv_var, mask)
observed_data = pt.as_tensor(data.data[observed_mask])
# Register ObservedRV corresponding to observed component
observed_rv.name = f"{name}_observed"
self.create_value_var(observed_rv, transform=None, value_var=observed_data)
self.add_named_variable(observed_rv)
self.observed_RVs.append(observed_rv)
# Register FreeRV corresponding to unobserved components
self.register_rv(unobserved_rv, f"{name}_unobserved", transform=transform)
# Register Deterministic that combines observed and missing
# Note: This can widely increase memory consumption during sampling for large datasets
rv_var = Deterministic(name, joined_rv, self, dims)
else:
if sps.issparse(data):
data = sparse.basic.as_sparse(data, name=name)
else:
data = pt.as_tensor_variable(data, name=name)
if total_size:
from pymc.variational.minibatch_rv import create_minibatch_rv
rv_var = create_minibatch_rv(rv_var, total_size)
rv_var.name = name
rv_var.tag.observations = data
self.create_value_var(rv_var, transform=None, value_var=data)
self.add_named_variable(rv_var, dims)
self.observed_RVs.append(rv_var)
return rv_var
def create_value_var(
self, rv_var: TensorVariable, transform: Any, value_var: Optional[Variable] = None
) -> TensorVariable:
"""Create a ``TensorVariable`` that will be used as the random
variable's "value" in log-likelihood graphs.
In general, we'll call this type of variable the "value" variable.
In all other cases, the role of the value variable is taken by
observed data. That's why value variables are only referenced in
this branch of the conditional.
"""
# Make the value variable a transformed value variable,
# if there's an applicable transform
if transform is UNSET:
if rv_var.owner is None:
transform = None
else:
transform = _default_transform(rv_var.owner.op, rv_var)
if value_var is not None:
if transform is not None:
raise ValueError("Cannot use transform when providing a pre-defined value_var")
elif transform is None:
# Create value variable with the same type as the RV
value_var = rv_var.type()
value_var.name = rv_var.name
if pytensor.config.compute_test_value != "off":
value_var.tag.test_value = rv_var.tag.test_value
else:
# Create value variable with the same type as the transformed RV
value_var = transform.forward(rv_var, *rv_var.owner.inputs).type()
value_var.name = f"{rv_var.name}_{transform.name}__"
value_var.tag.transform = transform
if pytensor.config.compute_test_value != "off":
value_var.tag.test_value = transform.forward(
rv_var, *rv_var.owner.inputs
).tag.test_value
_add_future_warning_tag(value_var)
rv_var.tag.value_var = value_var
self.rvs_to_transforms[rv_var] = transform
self.rvs_to_values[rv_var] = value_var
self.values_to_rvs[value_var] = rv_var
return value_var
def add_named_variable(self, var, dims: Optional[Tuple[Union[str, None], ...]] = None):
"""Add a random graph variable to the named variables of the model.
This can include several types of variables such basic_RVs, Data, Deterministics,
and Potentials.
"""
if var.name is None:
raise ValueError("Variable is unnamed.")
if self.named_vars.tree_contains(var.name):
raise ValueError(f"Variable name {var.name} already exists.")
if dims is not None:
if isinstance(dims, str):
dims = (dims,)
for dim in dims:
if dim not in self.coords and dim is not None:
raise ValueError(f"Dimension {dim} is not specified in `coords`.")
if any(var.name == dim for dim in dims if dim is not None):
raise ValueError(f"Variable `{var.name}` has the same name as its dimension label.")
self.named_vars_to_dims[var.name] = dims
self.named_vars[var.name] = var
if not hasattr(self, self.name_of(var.name)):
setattr(self, self.name_of(var.name), var)
@property
def prefix(self) -> str:
if self.isroot or not self.parent.prefix:
name = self.name
else:
name = f"{self.parent.prefix}::{self.name}"
return name
def name_for(self, name):
"""Checks if name has prefix and adds if needed"""
name = self._validate_name(name)
if self.prefix:
if not name.startswith(self.prefix + "::"):
return f"{self.prefix}::{name}"
else:
return name
else:
return name
def name_of(self, name):
"""Checks if name has prefix and deletes if needed"""
name = self._validate_name(name)
if not self.prefix or not name:
return name
elif name.startswith(self.prefix + "::"):
return name[len(self.prefix) + 2 :]
else:
return name
def __getitem__(self, key):
try:
return self.named_vars[key]
except KeyError as e:
try:
return self.named_vars[self.name_for(key)]
except KeyError:
raise e
def __contains__(self, key):
return key in self.named_vars or self.name_for(key) in self.named_vars
def replace_rvs_by_values(
self,
graphs: Sequence[TensorVariable],
**kwargs,
) -> List[TensorVariable]:
"""Clone and replace random variables in graphs with their value variables.
This will *not* recompute test values in the resulting graphs.
Parameters
----------
graphs
The graphs in which to perform the replacements.
"""
return replace_rvs_by_values(
graphs,
rvs_to_values=self.rvs_to_values,
rvs_to_transforms=self.rvs_to_transforms,
)
def compile_fn(
self,
outs: Union[Variable, Sequence[Variable]],
*,
inputs: Optional[Sequence[Variable]] = None,
mode=None,
point_fn: bool = True,
**kwargs,
) -> Union[PointFunc, Callable[[Sequence[np.ndarray]], Sequence[np.ndarray]]]:
"""Compiles an PyTensor function
Parameters
----------
outs
PyTensor variable or iterable of PyTensor variables.
inputs
PyTensor input variables, defaults to pytensorf.inputvars(outs).
mode
PyTensor compilation mode, default=None.
point_fn : bool
Whether to wrap the compiled function in a PointFunc, which takes a Point
dictionary with model variable names and values as input.
Returns
-------
Compiled PyTensor function
"""
if inputs is None:
inputs = inputvars(outs)
with self:
fn = compile_pymc(
inputs,
outs,
allow_input_downcast=True,
accept_inplace=True,
mode=mode,
**kwargs,
)
if point_fn:
return PointFunc(fn)
return fn
def profile(self, outs, *, n=1000, point=None, profile=True, **kwargs):
"""Compiles and profiles an PyTensor function which returns ``outs`` and
takes values of model vars as a dict as an argument.
Parameters
----------
outs: PyTensor variable or iterable of PyTensor variables
n: int, default 1000
Number of iterations to run
point: point
Point to pass to the function
profile: True or ProfileStats
args, kwargs
Compilation args
Returns
-------
ProfileStats
Use .summary() to print stats.
"""
kwargs.setdefault("on_unused_input", "ignore")
f = self.compile_fn(outs, inputs=self.value_vars, point_fn=False, profile=profile, **kwargs)
if point is None:
point = self.initial_point()
for _ in range(n):
f(**point)
return f.profile
def update_start_vals(self, a: Dict[str, np.ndarray], b: Dict[str, np.ndarray]):
r"""Update point `a` with `b`, without overwriting existing keys.
Values specified for transformed variables in `a` will be recomputed
conditional on the values of `b` and stored in `b`.
"""
raise FutureWarning(
"The `Model.update_start_vals` method was removed."
" To change initial values you may set the items of `Model.initial_values` directly."
)
def eval_rv_shapes(self) -> Dict[str, Tuple[int, ...]]:
"""Evaluates shapes of untransformed AND transformed free variables.
Returns
-------
shapes : dict
Maps untransformed and transformed variable names to shape tuples.
"""
names = []
outputs = []
for rv in self.free_RVs:
transform = self.rvs_to_transforms[rv]
if transform is not None:
names.append(get_transformed_name(rv.name, transform))
outputs.append(transform.forward(rv, *rv.owner.inputs).shape)
names.append(rv.name)
outputs.append(rv.shape)
f = pytensor.function(
inputs=[],
outputs=outputs,
givens=[(obs, self.rvs_to_values[obs]) for obs in self.observed_RVs],
mode=pytensor.compile.mode.FAST_COMPILE,
on_unused_input="ignore",
)
return {name: tuple(shape) for name, shape in zip(names, f())}
def check_start_vals(self, start):
r"""Check that the starting values for MCMC do not cause the relevant log probability
to evaluate to something invalid (e.g. Inf or NaN)
Parameters
----------
start : dict, or array of dict
Starting point in parameter space (or partial point)
Defaults to ``trace.point(-1))`` if there is a trace provided and
``model.initial_point`` if not (defaults to empty dict). Initialization
methods for NUTS (see ``init`` keyword) can overwrite the default.
Raises
------
``KeyError`` if the parameters provided by `start` do not agree with the
parameters contained within the model.
``pymc.exceptions.SamplingError`` if the evaluation of the parameters
in ``start`` leads to an invalid (i.e. non-finite) state
Returns
-------
None
"""
start_points = [start] if isinstance(start, dict) else start
value_names_to_dtypes = {value.name: value.dtype for value in self.value_vars}
value_names_set = set(value_names_to_dtypes.keys())
for elem in start_points:
for k, v in elem.items():
elem[k] = np.asarray(v, dtype=value_names_to_dtypes[k])
if not set(elem.keys()).issubset(value_names_set):
extra_keys = ", ".join(set(elem.keys()) - value_names_set)
valid_keys = ", ".join(value_names_set)
raise KeyError(
"Some start parameters do not appear in the model!\n"
f"Valid keys are: {valid_keys}, but {extra_keys} was supplied"
)
initial_eval = self.point_logps(point=elem)
if not all(np.isfinite(v) for v in initial_eval.values()):
raise SamplingError(
"Initial evaluation of model at starting point failed!\n"
f"Starting values:\n{elem}\n\n"
f"Logp initial evaluation results:\n{initial_eval}\n"
"You can call `model.debug()` for more details."
)
def point_logps(self, point=None, round_vals=2):
"""Computes the log probability of `point` for all random variables in the model.
Parameters
----------
point: Point, optional
Point to be evaluated. If ``None``, then ``model.initial_point``
is used.
round_vals: int, default 2
Number of decimals to round log-probabilities.
Returns
-------
log_probability_of_point : dict
Log probability of `point`.
"""
if point is None:
point = self.initial_point()
factors = self.basic_RVs + self.potentials
factor_logps_fn = [pt.sum(factor) for factor in self.logp(factors, sum=False)]
return {
factor.name: np.round(np.asarray(factor_logp), round_vals)
for factor, factor_logp in zip(
factors,
self.compile_fn(factor_logps_fn)(point),
)
}
def debug(
self,
point: Optional[Dict[str, np.ndarray]] = None,
fn: Literal["logp", "dlogp", "random"] = "logp",
verbose: bool = False,
):
"""Debug model function at point.
The method will evaluate the `fn` for each variable at a time.
When an evaluation fails or produces a non-finite value we print:
1. The graph of the parameters
2. The value of the parameters (if those can be evaluated)
3. The output of `fn` (if it can be evaluated)
This function should help to quickly narrow down invalid parametrizations.
Parameters
----------
point : Point
Point at which model function should be evaluated
fn : str, default "logp"
Function to be used for debugging. Can be one of [logp, dlogp, random].
verbose : bool, default False
Whether to show a more verbose PyTensor output when function cannot be evaluated
"""
print_ = functools.partial(print, file=sys.stdout)
def first_line(exc):
return exc.args[0].split("\n")[0]
def debug_parameters(rv):
if isinstance(rv.owner.op, RandomVariable):
inputs = rv.owner.inputs[3:]
else:
inputs = [inp for inp in rv.owner.inputs if not isinstance(inp.type, RandomType)]
rv_inputs = pytensor.function(
self.value_vars,
self.replace_rvs_by_values(inputs),
on_unused_input="ignore",
mode=get_mode(None).excluding("inplace", "fusion"),
)
print_(f"The variable {rv} has the following parameters:")
# done and used_ids are used to keep the same ids across distinct dprint calls
done = {}
used_ids = {}
for i, out in enumerate(rv_inputs.maker.fgraph.outputs):
print_(f"{i}: ", end=""),
# Don't print useless deepcopys
if out.owner and isinstance(out.owner.op, DeepCopyOp):
out = out.owner.inputs[0]
pytensor.dprint(out, print_type=True, done=done, used_ids=used_ids)
try:
print_("The parameters evaluate to:")
for i, rv_input_eval in enumerate(rv_inputs(**point)):
print_(f"{i}: {rv_input_eval}")
except Exception as exc:
print_(
f"The parameters of the variable {rv} cannot be evaluated: {first_line(exc)}"
)
if verbose:
print_(exc, "\n")
if fn not in ("logp", "dlogp", "random"):
raise ValueError(f"fn must be one of [logp, dlogp, random], got {fn}")
if point is None:
point = self.initial_point()
print_(f"point={point}\n")
rvs_to_check = list(self.basic_RVs)
if fn in ("logp", "dlogp"):
rvs_to_check += [self.replace_rvs_by_values(p) for p in self.potentials]
found_problem = False
for rv in rvs_to_check:
if fn == "logp":
rv_fn = pytensor.function(
self.value_vars, self.logp(vars=rv, sum=False)[0], on_unused_input="ignore"
)
elif fn == "dlogp":
rv_fn = pytensor.function(
self.value_vars, self.dlogp(vars=rv), on_unused_input="ignore"
)
else:
[rv_inputs_replaced] = replace_rvs_by_values(
[rv],
# Don't include itself, or the function will just the the value variable
rvs_to_values={
rv_key: value
for rv_key, value in self.rvs_to_values.items()
if rv_key is not rv
},
rvs_to_transforms=self.rvs_to_transforms,
)
rv_fn = pytensor.function(
self.value_vars, rv_inputs_replaced, on_unused_input="ignore"
)
try:
rv_fn_eval = rv_fn(**point)
except ParameterValueError as exc:
found_problem = True
debug_parameters(rv)
print_(
f"This does not respect one of the following constraints: {first_line(exc)}\n"
)
if verbose:
print_(exc)
except Exception as exc:
found_problem = True
debug_parameters(rv)
print_(
f"The variable {rv} {fn} method raised the following exception: {first_line(exc)}\n"
)
if verbose:
print_(exc)
else:
if not np.all(np.isfinite(rv_fn_eval)):
found_problem = True
debug_parameters(rv)
if fn == "random" or rv is self.potentials:
print_("This combination seems able to generate non-finite values")
else:
# Find which values are associated with non-finite evaluation
values = self.rvs_to_values[rv]
if rv in self.observed_RVs:
values = values.eval()
else:
values = point[values.name]
observed = " observed " if rv in self.observed_RVs else " "
print_(
f"Some of the{observed}values of variable {rv} are associated with a non-finite {fn}:"
)
mask = ~np.isfinite(rv_fn_eval)
for value, fn_eval in zip(values[mask], rv_fn_eval[mask]):
print_(f" value = {value} -> {fn} = {fn_eval}")
print_()
if not found_problem:
print_("No problems found")
elif not verbose:
print_("You can set `verbose=True` for more details")
# this is really disgusting, but it breaks a self-loop: I can't pass Model
# itself as context class init arg.
Model._context_class = Model
class BlockModelAccess(Model):
"""Can be used to prevent user access to Model contexts"""
def __init__(self, *args, error_msg_on_access="Model access is blocked", **kwargs):
self.error_msg_on_access = error_msg_on_access
def new_or_existing_block_model_access(*args, **kwargs):
"""Return a BlockModelAccess in the stack or create a new one if none is found."""
model = Model.get_context(error_if_none=False, allow_block_model_access=True)
if isinstance(model, BlockModelAccess):
return model
return BlockModelAccess(*args, **kwargs)
def set_data(new_data, model=None, *, coords=None):
"""Sets the value of one or more data container variables. Note that the shape is also
dynamic, it is updated when the value is changed. See the examples below for two common
use-cases that take advantage of this behavior.
Parameters
----------
new_data: dict
New values for the data containers. The keys of the dictionary are
the variables' names in the model and the values are the objects
with which to update.
model: Model (optional if in `with` context)
Examples
--------
This example shows how to change the shape of the likelihood to correspond automatically with
`x`, the predictor in a regression model.
.. code-block:: python
import pymc as pm
with pm.Model() as model:
x = pm.MutableData('x', [1., 2., 3.])
y = pm.MutableData('y', [1., 2., 3.])
beta = pm.Normal('beta', 0, 1)
obs = pm.Normal('obs', x * beta, 1, observed=y, shape=x.shape)
idata = pm.sample()
Then change the value of `x` to predict on new data.
.. code-block:: python
with model:
pm.set_data({'x': [5., 6., 9., 12., 15.]})
y_test = pm.sample_posterior_predictive(idata)
print(y_test.posterior_predictive['obs'].mean(('chain', 'draw')))
>>> array([4.6088569 , 5.54128318, 8.32953844, 11.14044852, 13.94178173])
This example shows how to reuse the same model without recompiling on a new data set. The
shape of the likelihood, `obs`, automatically tracks the shape of the observed data, `y`.
.. code-block:: python
import numpy as np
import pymc as pm
rng = np.random.default_rng()
data = rng.normal(loc=1.0, scale=2.0, size=100)
with pm.Model() as model:
y = pm.MutableData('y', data)
theta = pm.Normal('theta', mu=0.0, sigma=10.0)
obs = pm.Normal('obs', theta, 2.0, observed=y, shape=y.shape)
idata = pm.sample()
Now update the model with a new data set.
.. code-block:: python
with model:
pm.set_data({'y': rng.normal(loc=1.0, scale=2.0, size=200)})
idata = pm.sample()
"""
model = modelcontext(model)
for variable_name, new_value in new_data.items():
model.set_data(variable_name, new_value, coords=coords)
def compile_fn(
outs: Union[Variable, Sequence[Variable]],
*,
inputs: Optional[Sequence[Variable]] = None,
mode=None,
point_fn: bool = True,
model: Optional[Model] = None,
**kwargs,
) -> Union[PointFunc, Callable[[Sequence[np.ndarray]], Sequence[np.ndarray]]]:
"""Compiles an PyTensor function
Parameters
----------
outs
PyTensor variable or iterable of PyTensor variables.
inputs
PyTensor input variables, defaults to pytensorf.inputvars(outs).
mode
PyTensor compilation mode, default=None.
point_fn : bool
Whether to wrap the compiled function in a PointFunc, which takes a Point
dictionary with model variable names and values as input.
model : Model, optional
Current model on stack.
Returns
-------
Compiled PyTensor function
"""
model = modelcontext(model)
return model.compile_fn(
outs,
inputs=inputs,
mode=mode,
point_fn=point_fn,
**kwargs,
)
def Point(*args, filter_model_vars=False, **kwargs) -> Dict[str, np.ndarray]:
"""Build a point. Uses same args as dict() does.
Filters out variables not in the model. All keys are strings.
Parameters
----------
args, kwargs
arguments to build a dict
filter_model_vars : bool
If `True`, only model variables are included in the result.
"""
model = modelcontext(kwargs.pop("model", None))
args = list(args)
try:
d = dict(*args, **kwargs)
except Exception as e:
raise TypeError(f"can't turn {args} and {kwargs} into a dict. {e}")
return {
get_var_name(k): np.array(v)
for k, v in d.items()
if not filter_model_vars or (get_var_name(k) in map(get_var_name, model.value_vars))
}
def Deterministic(name, var, model=None, dims=None):
"""Create a named deterministic variable.
Deterministic nodes are only deterministic given all of their inputs, i.e.
they don't add randomness to the model. They are generally used to record
an intermediary result.
Parameters
----------
name : str
Name of the deterministic variable to be registered in the model.
var : tensor_like
Expression for the calculation of the variable.
model : Model, optional
The model object to which the Deterministic variable is added.
If ``None`` is provided, the current model in the context stack is used.
dims : str or tuple of str, optional
Dimension names for the variable.
Returns
-------
var : tensor_like
The registered, named variable wrapped in Deterministic.
Examples
--------
Indeed, PyMC allows for arbitrary combinations of random variables, for
example in the case of a logistic regression
.. code:: python
with pm.Model():
alpha = pm.Normal("alpha", 0, 1)
intercept = pm.Normal("intercept", 0, 1)
p = pm.math.invlogit(alpha * x + intercept)
outcome = pm.Bernoulli("outcome", p, observed=outcomes)
but doesn't memorize the fact that the expression ``pm.math.invlogit(alpha *
x + intercept)`` has been affected to the variable ``p``. If the quantity
``p`` is important and one would like to track its value in the sampling
trace, then one can use a deterministic node:
.. code:: python
with pm.Model():
alpha = pm.Normal("alpha", 0, 1)
intercept = pm.Normal("intercept", 0, 1)
p = pm.Deterministic("p", pm.math.invlogit(alpha * x + intercept))
outcome = pm.Bernoulli("outcome", p, observed=outcomes)
These two models are strictly equivalent from a mathematical point of view.
However, in the first case, the inference data will only contain values for
the variables ``alpha``, ``intercept`` and ``outcome``. In the second, it
will also contain sampled values of ``p`` for each of the observed points.
Notes
-----
Even though adding a Deterministic node forces PyMC to compute this
expression, which could have been optimized away otherwise, this doesn't come
with a performance cost. Indeed, Deterministic nodes are computed outside
the main computation graph, which can be optimized as though there was no
Deterministic nodes. Whereas the optimized graph can be evaluated thousands
of times during a NUTS step, the Deterministic quantities are just
computeed once at the end of the step, with the final values of the other
random variables.
"""
model = modelcontext(model)
var = var.copy(model.name_for(name))
model.deterministics.append(var)
model.add_named_variable(var, dims)
from pymc.printing import str_for_potential_or_deterministic
var.str_repr = types.MethodType(
functools.partial(str_for_potential_or_deterministic, dist_name="Deterministic"), var
)
var._repr_latex_ = types.MethodType(
functools.partial(
str_for_potential_or_deterministic, dist_name="Deterministic", formatting="latex"
),
var,
)
return var
def Potential(name, var: TensorVariable, model=None, dims=None) -> TensorVariable:
"""Add an arbitrary term to the model log-probability.
Warnings
--------
Potential terms only influence probability-based sampling, such as ``pm.sample``, but not forward sampling like
``pm.sample_prior_predictive`` or ``pm.sample_posterior_predictive``. A warning is raised when doing forward
sampling with models containing Potential terms.
Parameters
----------
name : str
Name of the potential variable to be registered in the model.
var : tensor_like
Expression to be added to the model joint logp.
model : Model, optional
The model object to which the potential function is added.
If ``None`` is provided, the current model in the context stack is used.
dims : str or tuple of str, optional
Dimension names for the variable.
Returns
-------
var : tensor_like
The registered, named model variable.
Examples
--------
In this example, we define a constraint on ``x`` to be greater or equal to 0.
The statement ``pm.math.log(pm.math.switch(constraint, 0, 1))`` adds either 0 or -inf to the model logp,
depending on whether the constraint is met. During sampling, any proposals where ``x`` is negative will be rejected.
.. code:: python
import pymc as pm
with pm.Model() as model:
x = pm.Normal("x", mu=0, sigma=1)
constraint = x >= 0
potential = pm.Potential("x_constraint", pm.math.log(pm.math.switch(constraint, 1, 0)))
Instead, with a soft constraint like ``pm.math.log(pm.math.switch(constraint, 1, 0.5))``,
the sampler will be less likely, but not forbidden, from accepting negative values for `x`.
.. code:: python
import pymc as pm
with pm.Model() as model:
x = pm.Normal("x", mu=0, sigma=1)
constraint = x >= 0
potential = pm.Potential("x_constraint", pm.math.log(pm.math.switch(constraint, 1.0, 0.5)))
A Potential term can depend on multiple variables.
In the following example, the ``soft_sum_constraint`` potential encourages ``x`` and ``y`` to have a small sum.
The more the sum deviates from zero, the more negative the penalty value of ``(-((x + y)**2))``.
.. code:: python
import pymc as pm
with pm.Model() as model:
x = pm.Normal("x", mu=0, sigma=10)
y = pm.Normal("y", mu=0, sigma=10)
soft_sum_constraint = pm.Potential("soft_sum_constraint", -((x + y)**2))
A Potential can be used to define a specific prior term.
The following example imposes a power law prior on `max_items`, under the form ``log(1/max_items)``,
which penalizes very large values of `max_items`.
.. code:: python
import pymc as pm
with pm.Model() as model:
# p(max_items) = 1 / max_items
max_items = pm.Uniform("max_items", lower=1, upper=100)
pm.Potential("power_prior", pm.math.log(1/max_items))
n_items = pm.Uniform("n_items", lower=1, upper=max_items, observed=60)
A Potential can be used to define a specific likelihood term.
In the following example, a normal likelihood term is added to fixed data.
The same result would be obtained by using an observed `Normal` variable.
.. code:: python
import pymc as pm
def normal_logp(value, mu, sigma):
return -0.5 * ((value - mu) / sigma) ** 2 - pm.math.log(sigma)
with pm.Model() as model:
mu = pm.Normal("x")
sigma = pm.HalfNormal("sigma")
data = [0.1, 0.5, 0.9]
llike = pm.Potential("llike", normal_logp(data, mu, sigma))
"""
model = modelcontext(model)
var.name = model.name_for(name)
model.potentials.append(var)
model.add_named_variable(var, dims)
from pymc.printing import str_for_potential_or_deterministic
var.str_repr = types.MethodType(
functools.partial(str_for_potential_or_deterministic, dist_name="Potential"), var
)
var._repr_latex_ = types.MethodType(
functools.partial(
str_for_potential_or_deterministic, dist_name="Potential", formatting="latex"
),
var,
)
return var
