https://github.com/pymc-devs/pymc3
Tip revision: f67ff8bc22fcc72ead96eeeaef04173ed53650cb authored by Ricardo Vieira on 30 June 2023, 09:54:11 UTC
Use "unobserved" as imputed variable suffix instead of "missing"
Use "unobserved" as imputed variable suffix instead of "missing"
Tip revision: f67ff8b
distribution.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 contextvars
import functools
import sys
import types
import warnings
from abc import ABCMeta
from functools import singledispatch
from typing import Callable, Optional, Sequence, Tuple, Union
import numpy as np
from pytensor import tensor as pt
from pytensor.compile.builders import OpFromGraph
from pytensor.graph import FunctionGraph, node_rewriter
from pytensor.graph.basic import Node, Variable
from pytensor.graph.replace import clone_replace
from pytensor.graph.rewriting.basic import in2out
from pytensor.graph.utils import MetaType
from pytensor.tensor.basic import as_tensor_variable
from pytensor.tensor.random.op import RandomVariable
from pytensor.tensor.random.rewriting import local_subtensor_rv_lift
from pytensor.tensor.random.utils import normalize_size_param
from pytensor.tensor.var import TensorVariable
from typing_extensions import TypeAlias
from pymc.distributions.shape_utils import (
Dims,
Shape,
_change_dist_size,
convert_dims,
convert_shape,
convert_size,
find_size,
rv_size_is_none,
shape_from_dims,
)
from pymc.exceptions import BlockModelAccessError
from pymc.logprob.abstract import MeasurableVariable, _icdf, _logcdf, _logprob
from pymc.logprob.basic import logp
from pymc.logprob.rewriting import logprob_rewrites_db
from pymc.model import BlockModelAccess
from pymc.printing import str_for_dist
from pymc.pytensorf import collect_default_updates, convert_observed_data, floatX
from pymc.util import UNSET, _add_future_warning_tag
from pymc.vartypes import continuous_types, string_types
__all__ = [
"CustomDist",
"DensityDist",
"DiracDelta",
"Distribution",
"Continuous",
"Discrete",
"SymbolicRandomVariable",
]
DIST_PARAMETER_TYPES: TypeAlias = Union[np.ndarray, int, float, TensorVariable]
vectorized_ppc: contextvars.ContextVar[Optional[Callable]] = contextvars.ContextVar(
"vectorized_ppc", default=None
)
PLATFORM = sys.platform
class _Unpickling:
pass
class DistributionMeta(ABCMeta):
"""
DistributionMeta class
Notes
-----
DistributionMeta currently performs many functions, and will likely be refactored soon.
See issue below for more details
https://github.com/pymc-devs/pymc/issues/5308
"""
def __new__(cls, name, bases, clsdict):
# Forcefully deprecate old v3 `Distribution`s
if "random" in clsdict:
def _random(*args, **kwargs):
warnings.warn(
"The old `Distribution.random` interface is deprecated.",
FutureWarning,
stacklevel=2,
)
return clsdict["random"](*args, **kwargs)
clsdict["random"] = _random
rv_op = clsdict.setdefault("rv_op", None)
rv_type = clsdict.setdefault("rv_type", None)
if isinstance(rv_op, RandomVariable):
if rv_type is not None:
assert isinstance(rv_op, rv_type)
else:
rv_type = type(rv_op)
clsdict["rv_type"] = rv_type
new_cls = super().__new__(cls, name, bases, clsdict)
if rv_type is not None:
# Create dispatch functions
class_logp = clsdict.get("logp")
if class_logp:
@_logprob.register(rv_type)
def logp(op, values, *dist_params, **kwargs):
dist_params = dist_params[3:]
(value,) = values
return class_logp(value, *dist_params)
class_logcdf = clsdict.get("logcdf")
if class_logcdf:
@_logcdf.register(rv_type)
def logcdf(op, value, *dist_params, **kwargs):
dist_params = dist_params[3:]
return class_logcdf(value, *dist_params)
class_icdf = clsdict.get("icdf")
if class_icdf:
@_icdf.register(rv_type)
def icdf(op, value, *dist_params, **kwargs):
dist_params = dist_params[3:]
return class_icdf(value, *dist_params)
class_moment = clsdict.get("moment")
if class_moment:
@_moment.register(rv_type)
def moment(op, rv, rng, size, dtype, *dist_params):
return class_moment(rv, size, *dist_params)
# Register the PyTensor rv_type as a subclass of this
# PyMC Distribution type.
new_cls.register(rv_type)
return new_cls
def _make_nice_attr_error(oldcode: str, newcode: str):
def fn(*args, **kwargs):
raise AttributeError(f"The `{oldcode}` method was removed. Instead use `{newcode}`.`")
return fn
class SymbolicRandomVariable(OpFromGraph):
"""Symbolic Random Variable
This is a subclasse of `OpFromGraph` which is used to encapsulate the symbolic
random graph of complex distributions which are built on top of pure
`RandomVariable`s.
These graphs may vary structurally based on the inputs (e.g., their dimensionality),
and usually require that random inputs have specific shapes for correct outputs
(e.g., avoiding broadcasting of random inputs). Due to this, most distributions that
return SymbolicRandomVariable create their these graphs at runtime via the
classmethod `cls.rv_op`, taking care to clone and resize random inputs, if needed.
"""
ndim_supp: int = None
"""Number of support dimensions as in RandomVariables
(0 for scalar, 1 for vector, ...)
"""
inline_logprob: bool = False
"""Specifies whether the logprob function is derived automatically by introspection
of the inner graph.
If `False`, a logprob function must be dispatched directly to the subclass type.
"""
_print_name: Tuple[str, str] = ("Unknown", "\\operatorname{Unknown}")
"""Tuple of (name, latex name) used for for pretty-printing variables of this type"""
def __init__(self, *args, ndim_supp, **kwargs):
"""Initialitze a SymbolicRandomVariable class."""
self.ndim_supp = ndim_supp
kwargs.setdefault("inline", True)
super().__init__(*args, **kwargs)
def update(self, node: Node):
"""Symbolic update expression for input random state variables
Returns a dictionary with the symbolic expressions required for correct updating
of random state input variables repeated function evaluations. This is used by
`pytensorf.compile_pymc`.
"""
return {}
class Distribution(metaclass=DistributionMeta):
"""Statistical distribution"""
rv_op: [RandomVariable, SymbolicRandomVariable] = None
rv_type: MetaType = None
def __new__(
cls,
name: str,
*args,
rng=None,
dims: Optional[Dims] = None,
initval=None,
observed=None,
total_size=None,
transform=UNSET,
**kwargs,
) -> TensorVariable:
"""Adds a tensor variable corresponding to a PyMC distribution to the current model.
Note that all remaining kwargs must be compatible with ``.dist()``
Parameters
----------
cls : type
A PyMC distribution.
name : str
Name for the new model variable.
rng : optional
Random number generator to use with the RandomVariable.
dims : tuple, optional
A tuple of dimension names known to the model. When shape is not provided,
the shape of dims is used to define the shape of the variable.
initval : optional
Numeric or symbolic untransformed initial value of matching shape,
or one of the following initial value strategies: "moment", "prior".
Depending on the sampler's settings, a random jitter may be added to numeric, symbolic
or moment-based initial values in the transformed space.
observed : optional
Observed data to be passed when registering the random variable in the model.
When neither shape nor dims is provided, the shape of observed is used to
define the shape of the variable.
See ``Model.register_rv``.
total_size : float, optional
See ``Model.register_rv``.
transform : optional
See ``Model.register_rv``.
**kwargs
Keyword arguments that will be forwarded to ``.dist()`` or the PyTensor RV Op.
Most prominently: ``shape`` for ``.dist()`` or ``dtype`` for the Op.
Returns
-------
rv : TensorVariable
The created random variable tensor, registered in the Model.
"""
try:
from pymc.model import Model
model = Model.get_context()
except TypeError:
raise TypeError(
"No model on context stack, which is needed to "
"instantiate distributions. Add variable inside "
"a 'with model:' block, or use the '.dist' syntax "
"for a standalone distribution."
)
if "testval" in kwargs:
initval = kwargs.pop("testval")
warnings.warn(
"The `testval` argument is deprecated; use `initval`.",
FutureWarning,
stacklevel=2,
)
if not isinstance(name, string_types):
raise TypeError(f"Name needs to be a string but got: {name}")
dims = convert_dims(dims)
if observed is not None:
observed = convert_observed_data(observed)
# Preference is given to size or shape. If not specified, we rely on dims and
# finally, observed, to determine the shape of the variable.
if kwargs.get("size") is None and kwargs.get("shape") is None:
if dims is not None:
kwargs["shape"] = shape_from_dims(dims, model)
elif observed is not None:
kwargs["shape"] = tuple(observed.shape)
rv_out = cls.dist(*args, **kwargs)
rv_out = model.register_rv(
rv_out,
name,
observed,
total_size,
dims=dims,
transform=transform,
initval=initval,
)
# add in pretty-printing support
rv_out.str_repr = types.MethodType(str_for_dist, rv_out)
rv_out._repr_latex_ = types.MethodType(
functools.partial(str_for_dist, formatting="latex"), rv_out
)
rv_out.logp = _make_nice_attr_error("rv.logp(x)", "pm.logp(rv, x)")
rv_out.logcdf = _make_nice_attr_error("rv.logcdf(x)", "pm.logcdf(rv, x)")
rv_out.random = _make_nice_attr_error("rv.random()", "pm.draw(rv)")
return rv_out
@classmethod
def dist(
cls,
dist_params,
*,
shape: Optional[Shape] = None,
**kwargs,
) -> TensorVariable:
"""Creates a tensor variable corresponding to the `cls` distribution.
Parameters
----------
dist_params : array-like
The inputs to the `RandomVariable` `Op`.
shape : int, tuple, Variable, optional
A tuple of sizes for each dimension of the new RV.
**kwargs
Keyword arguments that will be forwarded to the PyTensor RV Op.
Most prominently: ``size`` or ``dtype``.
Returns
-------
rv : TensorVariable
The created random variable tensor.
"""
if "testval" in kwargs:
kwargs.pop("testval")
warnings.warn(
"The `.dist(testval=...)` argument is deprecated and has no effect. "
"Initial values for sampling/optimization can be specified with `initval` in a modelcontext. "
"For using PyTensor's test value features, you must assign the `.tag.test_value` yourself.",
FutureWarning,
stacklevel=2,
)
if "initval" in kwargs:
raise TypeError(
"Unexpected keyword argument `initval`. "
"This argument is not available for the `.dist()` API."
)
if "dims" in kwargs:
raise NotImplementedError("The use of a `.dist(dims=...)` API is not supported.")
size = kwargs.pop("size", None)
if shape is not None and size is not None:
raise ValueError(
f"Passing both `shape` ({shape}) and `size` ({size}) is not supported!"
)
shape = convert_shape(shape)
size = convert_size(size)
# SymbolicRVs don't have `ndim_supp` until they are created
ndim_supp = getattr(cls.rv_op, "ndim_supp", None)
if ndim_supp is None:
ndim_supp = cls.rv_op(*dist_params, **kwargs).owner.op.ndim_supp
create_size = find_size(shape=shape, size=size, ndim_supp=ndim_supp)
rv_out = cls.rv_op(*dist_params, size=create_size, **kwargs)
rv_out.logp = _make_nice_attr_error("rv.logp(x)", "pm.logp(rv, x)")
rv_out.logcdf = _make_nice_attr_error("rv.logcdf(x)", "pm.logcdf(rv, x)")
rv_out.random = _make_nice_attr_error("rv.random()", "pm.draw(rv)")
_add_future_warning_tag(rv_out)
return rv_out
# Let PyMC know that the SymbolicRandomVariable has a logprob.
MeasurableVariable.register(SymbolicRandomVariable)
@node_rewriter([SymbolicRandomVariable])
def inline_symbolic_random_variable(fgraph, node):
"""
Optimization that expands the internal graph of a SymbolicRV when obtaining the logp
graph, if the flag `inline_logprob` is True.
"""
op = node.op
if op.inline_logprob:
return clone_replace(op.inner_outputs, {u: v for u, v in zip(op.inner_inputs, node.inputs)})
# Registered before pre-canonicalization which happens at position=-10
logprob_rewrites_db.register(
"inline_SymbolicRandomVariable",
in2out(inline_symbolic_random_variable),
"basic",
position=-20,
)
@singledispatch
def _moment(op, rv, *rv_inputs) -> TensorVariable:
raise NotImplementedError(f"Variable {rv} of type {op} has no moment implementation.")
def moment(rv: TensorVariable) -> TensorVariable:
"""Method for choosing a representative point/value
that can be used to start optimization or MCMC sampling.
The only parameter to this function is the RandomVariable
for which the value is to be derived.
"""
return _moment(rv.owner.op, rv, *rv.owner.inputs).astype(rv.dtype)
class Discrete(Distribution):
"""Base class for discrete distributions"""
def __new__(cls, name, *args, **kwargs):
if kwargs.get("transform", None):
raise ValueError("Transformations for discrete distributions")
return super().__new__(cls, name, *args, **kwargs)
class Continuous(Distribution):
"""Base class for continuous distributions"""
class CustomDistRV(RandomVariable):
"""
Base class for CustomDistRV
This should be subclassed when defining CustomDist objects.
"""
name = "CustomDistRV"
_print_name = ("CustomDist", "\\operatorname{CustomDist}")
@classmethod
def rng_fn(cls, rng, *args):
args = list(args)
size = args.pop(-1)
return cls._random_fn(*args, rng=rng, size=size)
class _CustomDist(Distribution):
"""A distribution that returns a subclass of CustomDistRV"""
rv_type = CustomDistRV
@classmethod
def dist(
cls,
*dist_params,
logp: Optional[Callable] = None,
logcdf: Optional[Callable] = None,
random: Optional[Callable] = None,
moment: Optional[Callable] = None,
ndim_supp: int = 0,
ndims_params: Optional[Sequence[int]] = None,
dtype: str = "floatX",
class_name: str = "CustomDist",
**kwargs,
):
dist_params = [as_tensor_variable(param) for param in dist_params]
# Assume scalar ndims_params
if ndims_params is None:
ndims_params = [0] * len(dist_params)
if logp is None:
logp = default_not_implemented(class_name, "logp")
if logcdf is None:
logcdf = default_not_implemented(class_name, "logcdf")
if moment is None:
moment = functools.partial(
default_moment,
rv_name=class_name,
has_fallback=random is not None,
ndim_supp=ndim_supp,
)
if random is None:
random = default_not_implemented(class_name, "random")
return super().dist(
dist_params,
logp=logp,
logcdf=logcdf,
random=random,
moment=moment,
ndim_supp=ndim_supp,
ndims_params=ndims_params,
dtype=dtype,
class_name=class_name,
**kwargs,
)
@classmethod
def rv_op(
cls,
*dist_params,
logp: Optional[Callable],
logcdf: Optional[Callable],
random: Optional[Callable],
moment: Optional[Callable],
ndim_supp: int,
ndims_params: Optional[Sequence[int]],
dtype: str,
class_name: str,
**kwargs,
):
rv_type = type(
class_name,
(CustomDistRV,),
dict(
name=class_name,
inplace=False,
ndim_supp=ndim_supp,
ndims_params=ndims_params,
dtype=dtype,
# Specific to CustomDist
_random_fn=random,
),
)
# Dispatch custom methods
@_logprob.register(rv_type)
def custom_dist_logp(op, values, rng, size, dtype, *dist_params, **kwargs):
return logp(values[0], *dist_params)
@_logcdf.register(rv_type)
def density_dist_logcdf(op, value, rng, size, dtype, *dist_params, **kwargs):
return logcdf(value, *dist_params, **kwargs)
@_moment.register(rv_type)
def density_dist_get_moment(op, rv, rng, size, dtype, *dist_params):
return moment(rv, size, *dist_params)
rv_op = rv_type()
return rv_op(*dist_params, **kwargs)
class CustomSymbolicDistRV(SymbolicRandomVariable):
"""
Base class for CustomSymbolicDist
This should be subclassed when defining custom CustomDist objects that have
symbolic random methods.
"""
default_output = -1
_print_name = ("CustomSymbolicDist", "\\operatorname{CustomSymbolicDist}")
def update(self, node: Node):
op = node.op
inner_updates = collect_default_updates(
inputs=op.inner_inputs, outputs=op.inner_outputs, must_be_shared=False
)
# Map inner updates to outer inputs/outputs
updates = {}
for rng, update in inner_updates.items():
inp_idx = op.inner_inputs.index(rng)
out_idx = op.inner_outputs.index(update)
updates[node.inputs[inp_idx]] = node.outputs[out_idx]
return updates
class _CustomSymbolicDist(Distribution):
rv_type = CustomSymbolicDistRV
@classmethod
def dist(
cls,
*dist_params,
dist: Callable,
logp: Optional[Callable] = None,
logcdf: Optional[Callable] = None,
moment: Optional[Callable] = None,
ndim_supp: int = 0,
dtype: str = "floatX",
class_name: str = "CustomDist",
**kwargs,
):
dist_params = [as_tensor_variable(param) for param in dist_params]
if logcdf is None:
logcdf = default_not_implemented(class_name, "logcdf")
if moment is None:
moment = functools.partial(
default_moment,
rv_name=class_name,
has_fallback=True,
ndim_supp=ndim_supp,
)
return super().dist(
dist_params,
class_name=class_name,
logp=logp,
logcdf=logcdf,
dist=dist,
moment=moment,
ndim_supp=ndim_supp,
**kwargs,
)
@classmethod
def rv_op(
cls,
*dist_params,
dist: Callable,
logp: Optional[Callable],
logcdf: Optional[Callable],
moment: Optional[Callable],
size=None,
ndim_supp: int,
class_name: str,
):
size = normalize_size_param(size)
dummy_size_param = size.type()
dummy_dist_params = [dist_param.type() for dist_param in dist_params]
with BlockModelAccess(
error_msg_on_access="Model variables cannot be created in the dist function. Use the `.dist` API"
):
dummy_rv = dist(*dummy_dist_params, dummy_size_param)
dummy_params = [dummy_size_param] + dummy_dist_params
dummy_updates_dict = collect_default_updates(inputs=dummy_params, outputs=(dummy_rv,))
rv_type = type(
class_name,
(CustomSymbolicDistRV,),
# If logp is not provided, we try to infer it from the dist graph
dict(
inline_logprob=logp is None,
),
)
# Dispatch custom methods
if logp is not None:
@_logprob.register(rv_type)
def custom_dist_logp(op, values, size, *params, **kwargs):
return logp(values[0], *params[: len(dist_params)])
if logcdf is not None:
@_logcdf.register(rv_type)
def custom_dist_logcdf(op, value, size, *params, **kwargs):
return logcdf(value, *params[: len(dist_params)])
@_moment.register(rv_type)
def custom_dist_get_moment(op, rv, size, *params):
return moment(rv, size, *params[: len(params)])
@_change_dist_size.register(rv_type)
def change_custom_symbolic_dist_size(op, rv, new_size, expand):
node = rv.owner
if expand:
shape = tuple(rv.shape)
old_size = shape[: len(shape) - node.op.ndim_supp]
new_size = tuple(new_size) + tuple(old_size)
new_size = pt.as_tensor(new_size, ndim=1, dtype="int64")
old_size, *old_dist_params = node.inputs[: len(dist_params) + 1]
# OpFromGraph has to be recreated if the size type changes!
dummy_size_param = new_size.type()
dummy_dist_params = [dist_param.type() for dist_param in old_dist_params]
dummy_rv = dist(*dummy_dist_params, dummy_size_param)
dummy_params = [dummy_size_param] + dummy_dist_params
dummy_updates_dict = collect_default_updates(inputs=dummy_params, outputs=(dummy_rv,))
new_rv_op = rv_type(
inputs=dummy_params,
outputs=[*dummy_updates_dict.values(), dummy_rv],
ndim_supp=ndim_supp,
)
new_rv = new_rv_op(new_size, *dist_params)
return new_rv
rv_op = rv_type(
inputs=dummy_params,
outputs=[*dummy_updates_dict.values(), dummy_rv],
ndim_supp=ndim_supp,
)
return rv_op(size, *dist_params)
class CustomDist:
"""A helper class to create custom distributions
This class can be used to wrap black-box random and logp methods for use in
forward and mcmc sampling.
A user can provide a `dist` function that returns a PyTensor graph built from
simpler PyMC distributions, which represents the distribution. This graph is
used to take random draws, and to infer the logp expression automatically
when not provided by the user.
Alternatively, a user can provide a `random` function that returns numerical
draws (e.g., via NumPy routines), and a `logp` function that must return an
Python graph that represents the logp graph when evaluated. This is used for
mcmc sampling.
Additionally, a user can provide a `logcdf` and `moment` functions that must return
an PyTensor graph that computes those quantities. These may be used by other PyMC
routines.
Parameters
----------
name : str
dist_params : Tuple
A sequence of the distribution's parameter. These will be converted into
Pytensor tensor variables internally.
dist: Optional[Callable]
A callable that returns a PyTensor graph built from simpler PyMC distributions
which represents the distribution. This can be used by PyMC to take random draws
as well as to infer the logp of the distribution in some cases. In that case
it's not necessary to implement ``random`` or ``logp`` functions.
It must have the following signature: ``dist(*dist_params, size)``.
The symbolic tensor distribution parameters are passed as positional arguments in
the same order as they are supplied when the ``CustomDist`` is constructed.
random : Optional[Callable]
A callable that can be used to generate random draws from the distribution
It must have the following signature: ``random(*dist_params, rng=None, size=None)``.
The numerical distribution parameters are passed as positional arguments in the
same order as they are supplied when the ``CustomDist`` is constructed.
The keyword arguments are ``rng``, which will provide the random variable's
associated :py:class:`~numpy.random.Generator`, and ``size``, that will represent
the desired size of the random draw. If ``None``, a ``NotImplemented``
error will be raised when trying to draw random samples from the distribution's
prior or posterior predictive.
logp : Optional[Callable]
A callable that calculates the log probability of some given ``value``
conditioned on certain distribution parameter values. It must have the
following signature: ``logp(value, *dist_params)``, where ``value`` is
an PyTensor tensor that represents the distribution value, and ``dist_params``
are the tensors that hold the values of the distribution parameters.
This function must return an PyTensor tensor.
When the `dist` function is specified, PyMC will try to automatically
infer the `logp` when this is not provided.
Otherwise, a ``NotImplementedError`` will be raised when trying to compute the
distribution's logp.
logcdf : Optional[Callable]
A callable that calculates the log cumulative log probability of some given
``value`` conditioned on certain distribution parameter values. It must have the
following signature: ``logcdf(value, *dist_params)``, where ``value`` is
an PyTensor tensor that represents the distribution value, and ``dist_params``
are the tensors that hold the values of the distribution parameters.
This function must return an PyTensor tensor. If ``None``, a ``NotImplementedError``
will be raised when trying to compute the distribution's logcdf.
moment : Optional[Callable]
A callable that can be used to compute the moments of the distribution.
It must have the following signature: ``moment(rv, size, *rv_inputs)``.
The distribution's variable is passed as the first argument ``rv``. ``size``
is the random variable's size implied by the ``dims``, ``size`` and parameters
supplied to the distribution. Finally, ``rv_inputs`` is the sequence of the
distribution parameters, in the same order as they were supplied when the
CustomDist was created. If ``None``, a default ``moment`` function will be
assigned that will always return 0, or an array of zeros.
ndim_supp : int
The number of dimensions in the support of the distribution. Defaults to assuming
a scalar distribution, i.e. ``ndim_supp = 0``.
ndims_params : Optional[Sequence[int]]
The list of number of dimensions in the support of each of the distribution's
parameters. If ``None``, it is assumed that all parameters are scalars, hence
the number of dimensions of their support will be 0. This is not needed if an
PyTensor dist function is provided.
dtype : str
The dtype of the distribution. All draws and observations passed into the
distribution will be cast onto this dtype. This is not needed if an PyTensor
dist function is provided, which should already return the right dtype!
class_name : str
Name for the class which will wrap the CustomDist methods. When not specified,
it will be given the name of the model variable.
kwargs :
Extra keyword arguments are passed to the parent's class ``__new__`` method.
Examples
--------
Create a CustomDist that wraps a black-box logp function. This variable cannot be
used in prior or posterior predictive sampling because no random function was provided
.. code-block:: python
import numpy as np
import pymc as pm
from pytensor.tensor import TensorVariable
def logp(value: TensorVariable, mu: TensorVariable) -> TensorVariable:
return -(value - mu)**2
with pm.Model():
mu = pm.Normal('mu',0,1)
pm.CustomDist(
'custom_dist',
mu,
logp=logp,
observed=np.random.randn(100),
)
idata = pm.sample(100)
Provide a random function that return numerical draws. This allows one to use a
CustomDist in prior and posterior predictive sampling.
.. code-block:: python
from typing import Optional, Tuple
import numpy as np
import pymc as pm
from pytensor.tensor import TensorVariable
def logp(value: TensorVariable, mu: TensorVariable) -> TensorVariable:
return -(value - mu)**2
def random(
mu: np.ndarray | float,
rng: Optional[np.random.Generator] = None,
size : Optional[Tuple[int]]=None,
) -> np.ndarray | float :
return rng.normal(loc=mu, scale=1, size=size)
with pm.Model():
mu = pm.Normal('mu', 0 , 1)
pm.CustomDist(
'custom_dist',
mu,
logp=logp,
random=random,
observed=np.random.randn(100, 3),
size=(100, 3),
)
prior = pm.sample_prior_predictive(10)
Provide a dist function that creates a PyTensor graph built from other
PyMC distributions. PyMC can automatically infer that the logp of this
variable corresponds to a shifted Exponential distribution.
.. code-block:: python
import pymc as pm
from pytensor.tensor import TensorVariable
def dist(
lam: TensorVariable,
shift: TensorVariable,
size: TensorVariable,
) -> TensorVariable:
return pm.Exponential.dist(lam, size=size) + shift
with pm.Model() as m:
lam = pm.HalfNormal("lam")
shift = -1
pm.CustomDist(
"custom_dist",
lam,
shift,
dist=dist,
observed=[-1, -1, 0],
)
prior = pm.sample_prior_predictive()
posterior = pm.sample()
Provide a dist function that creates a PyTensor graph built from other
PyMC distributions. PyMC can automatically infer that the logp of
this variable corresponds to a modified-PERT distribution.
.. code-block:: python
import pymc as pm
from pytensor.tensor import TensorVariable
def pert(
low: Tensorvariable,
peak: Tensorvariable,
high: Tensorvariable,
lmbda: Tensorvariable,
size: Tensorvariable,
) -> Tensorvariable:
range = (high - low)
s_alpha = 1 + lmbda * (peak - low) / range
s_beta = 1 + lmbda * (high - peak) / range
return pm.Beta.dist(s_alpha, s_beta, size=size) * range + low
with pm.Model() as m:
low = pm.Normal("low", 0, 10)
peak = pm.Normal("peak", 50, 10)
high = pm.Normal("high", 100, 10)
lmbda = 4
pm.CustomDist("pert", low, peak, high, lmbda, dist=pert, observed=[30, 35, 73])
m.point_logps()
"""
def __new__(
cls,
name,
*dist_params,
dist: Optional[Callable] = None,
random: Optional[Callable] = None,
logp: Optional[Callable] = None,
logcdf: Optional[Callable] = None,
moment: Optional[Callable] = None,
ndim_supp: int = 0,
ndims_params: Optional[Sequence[int]] = None,
dtype: str = "floatX",
**kwargs,
):
if isinstance(kwargs.get("observed", None), dict):
raise TypeError(
"Since ``v4.0.0`` the ``observed`` parameter should be of type"
" ``pd.Series``, ``np.array``, or ``pm.Data``."
" Previous versions allowed passing distribution parameters as"
" a dictionary in ``observed``, in the current version these "
"parameters are positional arguments."
)
dist_params = cls.parse_dist_params(dist_params)
cls.check_valid_dist_random(dist, random, dist_params)
if dist is not None:
kwargs.setdefault("class_name", f"CustomDist_{name}")
return _CustomSymbolicDist(
name,
*dist_params,
dist=dist,
logp=logp,
logcdf=logcdf,
moment=moment,
ndim_supp=ndim_supp,
**kwargs,
)
else:
kwargs.setdefault("class_name", f"CustomDist_{name}")
return _CustomDist(
name,
*dist_params,
random=random,
logp=logp,
logcdf=logcdf,
moment=moment,
ndim_supp=ndim_supp,
ndims_params=ndims_params,
dtype=dtype,
**kwargs,
)
@classmethod
def dist(
cls,
*dist_params,
dist: Optional[Callable] = None,
random: Optional[Callable] = None,
logp: Optional[Callable] = None,
logcdf: Optional[Callable] = None,
moment: Optional[Callable] = None,
ndim_supp: int = 0,
ndims_params: Optional[Sequence[int]] = None,
dtype: str = "floatX",
**kwargs,
):
dist_params = cls.parse_dist_params(dist_params)
cls.check_valid_dist_random(dist, random, dist_params)
if dist is not None:
return _CustomSymbolicDist.dist(
*dist_params,
dist=dist,
logp=logp,
logcdf=logcdf,
moment=moment,
ndim_supp=ndim_supp,
**kwargs,
)
else:
return _CustomDist.dist(
*dist_params,
random=random,
logp=logp,
logcdf=logcdf,
moment=moment,
ndim_supp=ndim_supp,
ndims_params=ndims_params,
dtype=dtype,
**kwargs,
)
@classmethod
def parse_dist_params(cls, dist_params):
if len(dist_params) > 0 and callable(dist_params[0]):
raise TypeError(
"The DensityDist API has changed, you are using the old API "
"where logp was the first positional argument. In the current API, "
"the logp is a keyword argument, amongst other changes. Please refer "
"to the API documentation for more information on how to use the "
"new DensityDist API."
)
return [as_tensor_variable(param) for param in dist_params]
@classmethod
def check_valid_dist_random(cls, dist, random, dist_params):
if dist is not None and random is not None:
raise ValueError("Cannot provide both dist and random functions")
if random is not None and cls.is_symbolic_random(random, dist_params):
raise TypeError(
"API change: function passed to `random` argument should no longer return a PyTensor graph. "
"Pass such function to the `dist` argument instead."
)
@classmethod
def is_symbolic_random(self, random, dist_params):
if random is None:
return False
# Try calling random with symbolic inputs
try:
size = normalize_size_param(None)
with BlockModelAccess(
error_msg_on_access="Model variables cannot be created in the random function. Use the `.dist` API to create such variables."
):
out = random(*dist_params, size)
except BlockModelAccessError:
raise
except Exception:
# If it fails we assume it was not
return False
# Confirm the output is symbolic
return isinstance(out, Variable)
DensityDist = CustomDist
def default_not_implemented(rv_name, method_name):
message = (
f"Attempted to run {method_name} on the CustomDist '{rv_name}', "
f"but this method had not been provided when the distribution was "
f"constructed. Please re-build your model and provide a callable "
f"to '{rv_name}'s {method_name} keyword argument.\n"
)
def func(*args, **kwargs):
raise NotImplementedError(message)
return func
def default_moment(rv, size, *rv_inputs, rv_name=None, has_fallback=False, ndim_supp=0):
if ndim_supp == 0:
return pt.zeros(size, dtype=rv.dtype)
elif has_fallback:
return pt.zeros_like(rv)
else:
raise TypeError(
"Cannot safely infer the size of a multivariate random variable's moment. "
f"Please provide a moment function when instantiating the {rv_name} "
"random variable."
)
class DiracDeltaRV(RandomVariable):
name = "diracdelta"
ndim_supp = 0
ndims_params = [0]
_print_name = ("DiracDelta", "\\operatorname{DiracDelta}")
def make_node(self, rng, size, dtype, c):
c = pt.as_tensor_variable(c)
return super().make_node(rng, size, c.dtype, c)
@classmethod
def rng_fn(cls, rng, c, size=None):
if size is None:
return c.copy()
return np.full(size, c)
diracdelta = DiracDeltaRV()
class DiracDelta(Discrete):
r"""
DiracDelta log-likelihood.
Parameters
----------
c : tensor_like of float or int
Dirac Delta parameter. The dtype of `c` determines the dtype of the distribution.
This can affect which sampler is assigned to DiracDelta variables, or variables
that use DiracDelta, such as Mixtures.
"""
rv_op = diracdelta
@classmethod
def dist(cls, c, *args, **kwargs):
c = pt.as_tensor_variable(c)
if c.dtype in continuous_types:
c = floatX(c)
return super().dist([c], **kwargs)
def moment(rv, size, c):
if not rv_size_is_none(size):
c = pt.full(size, c)
return c
def logp(value, c):
return pt.switch(
pt.eq(value, c),
pt.zeros_like(value),
-np.inf,
)
def logcdf(value, c):
return pt.switch(
pt.lt(value, c),
-np.inf,
0,
)
class PartialObservedRV(SymbolicRandomVariable):
"""RandomVariable with partially observed subspace, as indicated by a boolean mask.
See `create_partial_observed_rv` for more details.
"""
def create_partial_observed_rv(
rv: TensorVariable,
mask: Union[np.ndarray, TensorVariable],
) -> Tuple[
Tuple[TensorVariable, TensorVariable], Tuple[TensorVariable, TensorVariable], TensorVariable
]:
"""Separate observed and unobserved components of a RandomVariable.
This function may return two independent RandomVariables or, if not possible,
two variables from a common `PartialObservedRV` node
Parameters
----------
rv : TensorVariable
mask : tensor_like
Constant or variable boolean mask. True entries correspond to components of the variable that are not observed.
Returns
-------
observed_rv and mask : Tuple of TensorVariable
The observed component of the RV and respective indexing mask
unobserved_rv and mask : Tuple of TensorVariable
The unobserved component of the RV and respective indexing mask
joined_rv : TensorVariable
The symbolic join of the observed and unobserved components.
"""
if not mask.dtype == "bool":
raise ValueError(
f"mask must be an array or tensor of boolean dtype, got dtype: {mask.dtype}"
)
if mask.ndim > rv.ndim:
raise ValueError(f"mask can't have more dims than rv, got ndim: {mask.ndim}")
antimask = ~mask
can_rewrite = False
# Only pure RVs can be rewritten
if isinstance(rv.owner.op, RandomVariable):
ndim_supp = rv.owner.op.ndim_supp
# All univariate RVs can be rewritten
if ndim_supp == 0:
can_rewrite = True
# Multivariate RVs can be rewritten if masking does not split within support dimensions
else:
batch_dims = rv.type.ndim - ndim_supp
constant_mask = getattr(as_tensor_variable(mask), "data", None)
# Indexing does not overlap with core dimensions
if mask.ndim <= batch_dims:
can_rewrite = True
# Try to handle special case where mask is constant across support dimensions,
# TODO: This could be done by the rewrite itself
elif constant_mask is not None:
# We check if a constant_mask that only keeps the first entry of each support dim
# is equivalent to the original one after re-expanding.
trimmed_mask = constant_mask[(...,) + (0,) * ndim_supp]
expanded_mask = np.broadcast_to(
np.expand_dims(trimmed_mask, axis=tuple(range(-ndim_supp, 0))),
shape=constant_mask.shape,
)
if np.array_equal(constant_mask, expanded_mask):
mask = trimmed_mask
antimask = ~trimmed_mask
can_rewrite = True
if can_rewrite:
# Rewrite doesn't work with boolean masks. Should be fixed after https://github.com/pymc-devs/pytensor/pull/329
mask, antimask = mask.nonzero(), antimask.nonzero()
masked_rv = rv[mask]
fgraph = FunctionGraph(outputs=[masked_rv], clone=False)
[unobserved_rv] = local_subtensor_rv_lift.transform(fgraph, fgraph.outputs[0].owner)
antimasked_rv = rv[antimask]
fgraph = FunctionGraph(outputs=[antimasked_rv], clone=False)
[observed_rv] = local_subtensor_rv_lift.transform(fgraph, fgraph.outputs[0].owner)
# Make a clone of the observedRV, with a distinct rng so that observed and
# unobserved are never treated as equivalent (and mergeable) nodes by pytensor.
_, size, _, *inps = observed_rv.owner.inputs
observed_rv = observed_rv.owner.op(*inps, size=size)
# For all other cases use the more general PartialObservedRV
else:
# The symbolic graph simply splits the observed and unobserved components,
# so they can be given separate values.
dist_, mask_ = rv.type(), as_tensor_variable(mask).type()
observed_rv_, unobserved_rv_ = dist_[~mask_], dist_[mask_]
observed_rv, unobserved_rv = PartialObservedRV(
inputs=[dist_, mask_],
outputs=[observed_rv_, unobserved_rv_],
ndim_supp=rv.owner.op.ndim_supp,
)(rv, mask)
joined_rv = pt.empty(rv.shape, dtype=rv.type.dtype)
joined_rv = pt.set_subtensor(joined_rv[mask], unobserved_rv)
joined_rv = pt.set_subtensor(joined_rv[antimask], observed_rv)
return (observed_rv, antimask), (unobserved_rv, mask), joined_rv
@_logprob.register(PartialObservedRV)
def partial_observed_rv_logprob(op, values, dist, mask, **kwargs):
# For the logp, simply join the values
[obs_value, unobs_value] = values
antimask = ~mask
joined_value = pt.empty_like(dist)
joined_value = pt.set_subtensor(joined_value[mask], unobs_value)
joined_value = pt.set_subtensor(joined_value[antimask], obs_value)
joined_logp = logp(dist, joined_value)
# If we have a univariate RV we can split apart the logp terms
if op.ndim_supp == 0:
return joined_logp[antimask], joined_logp[mask]
# Otherwise, we can't (always/ easily) split apart logp terms.
# We return the full logp for the observed value, and a 0-nd array for the unobserved value
else:
return joined_logp.ravel(), pt.zeros((0,), dtype=joined_logp.type.dtype)
@_moment.register(PartialObservedRV)
def partial_observed_rv_moment(op, partial_obs_rv, rv, mask):
# Unobserved output
if partial_obs_rv.owner.outputs.index(partial_obs_rv) == 1:
return moment(rv)[mask]
# Observed output
else:
return moment(rv)[~mask]
