Allow parametrization through either shape, dims or size

Closes pymc-devs#4552.

RELEASE-NOTES.md

@@ -8,6 +8,11 @@
 
 ### New Features
 - The `CAR` distribution has been added to allow for use of conditional autoregressions which often are used in spatial and network models.
+- The dimensionality of model variables can now be parametrized through either of `shape`, `dims` or `size` (see [#4625](https://github.com/pymc-devs/pymc3/pull/4625)):
+  - With `shape` the length of dimensions must be given numerically or as scalar Aesara `Variables`.
+  - `dims` is arguably the most elegant parametrization, because it allows you to resize `pm.Data` variables and leads to well defined coordinates in `InferenceData` objects.
+  - The `size` kwarg creates new dimensions in addition to what is implied by RV parameters.
+  - An `Ellipsis` (`...`) in the last position of `shape` or `dims` can be used as short-hand notation for implied dimensions.
 - ...
 
 ### Maintenance

pymc3/distributions/distribution.py

@@ -13,28 +13,25 @@
 #   limitations under the License.
 import contextvars
 import inspect
+import logging
 import multiprocessing
 import sys
 import types
 import warnings
 
 from abc import ABCMeta
 from copy import copy
-from typing import TYPE_CHECKING
+from typing import Any, Optional, Sequence, Tuple, Union
 
+import aesara
+import aesara.tensor as at
 import dill
 
+from aesara.graph.basic import Variable
 from aesara.tensor.random.op import RandomVariable
 
+from pymc3.aesaraf import change_rv_size
 from pymc3.distributions import _logcdf, _logp
-
-if TYPE_CHECKING:
-    from typing import Optional, Callable
-
-import aesara
-import aesara.graph.basic
-import aesara.tensor as at
-
 from pymc3.util import UNSET, get_repr_for_variable
 from pymc3.vartypes import string_types
 
@@ -46,12 +43,18 @@
     "NoDistribution",
 ]
 
+_log = logging.getLogger(__file__)
+
 vectorized_ppc = contextvars.ContextVar(
     "vectorized_ppc", default=None
 )  # type: contextvars.ContextVar[Optional[Callable]]
 
 PLATFORM = sys.platform
 
+Shape = Union[int, Sequence[Union[str, type(Ellipsis)]], Variable]
+Dims = Union[str, Sequence[Union[str, None, type(Ellipsis)]]]
+Size = Union[int, Tuple[int, ...]]
+
 
 class _Unpickling:
     pass
@@ -122,13 +125,111 @@ def logcdf(op, var, rvs_to_values, *dist_params, **kwargs):
         return new_cls
 
 
+def _valid_ellipsis_position(items: Union[None, Shape, Dims, Size]) -> bool:
+    if items is not None and not isinstance(items, Variable) and Ellipsis in items:
+        if any(i == Ellipsis for i in items[:-1]):
+            return False
+    return True
+
+
+def _validate_shape_dims_size(
+    shape: Any = None, dims: Any = None, size: Any = None
+) -> Tuple[Optional[Shape], Optional[Dims], Optional[Size]]:
+    # Raise on unsupported parametrization
+    if shape is not None and dims is not None:
+        raise ValueError("Passing both `shape` ({shape}) and `dims` ({dims}) is not supported!")
+    if dims is not None and size is not None:
+        raise ValueError("Passing both `dims` ({dims}) and `size` ({size}) is not supported!")
+    if shape is not None and size is not None:
+        raise ValueError("Passing both `shape` ({shape}) and `size` ({size}) is not supported!")
+
+    # Raise on invalid types
+    if not isinstance(shape, (type(None), int, list, tuple, Variable)):
+        raise ValueError("The `shape` parameter must be an int, list or tuple.")
+    if not isinstance(dims, (type(None), str, list, tuple)):
+        raise ValueError("The `dims` parameter must be a str, list or tuple.")
+    if not isinstance(size, (type(None), int, list, tuple)):
+        raise ValueError("The `size` parameter must be an int, list or tuple.")
+
+    # Auto-convert non-tupled parameters
+    if isinstance(shape, int):
+        shape = (shape,)
+    if isinstance(dims, str):
+        dims = (dims,)
+    if isinstance(size, int):
+        size = (size,)
+
+    # Convert to actual tuples
+    if not isinstance(shape, (type(None), tuple, Variable)):
+        shape = tuple(shape)
+    if not isinstance(dims, (type(None), tuple)):
+        dims = tuple(dims)
+    if not isinstance(size, (type(None), tuple)):
+        size = tuple(size)
+
+    if not _valid_ellipsis_position(shape):
+        raise ValueError(
+            f"Ellipsis in `shape` may only appear in the last position. Actual: {shape}"
+        )
+    if not _valid_ellipsis_position(dims):
+        raise ValueError(f"Ellipsis in `dims` may only appear in the last position. Actual: {dims}")
+    if size is not None and Ellipsis in size:
+        raise ValueError("The `size` parameter cannot contain an Ellipsis. Actual: {size}")
+    return shape, dims, size
+
+
 class Distribution(metaclass=DistributionMeta):
     """Statistical distribution"""
 
     rv_class = None
     rv_op = None
 
-    def __new__(cls, name, *args, **kwargs):
+    def __new__(
+        cls,
+        name: str,
+        *args,
+        rng=None,
+        dims: Optional[Dims] = None,
+        testval=None,
+        observed=None,
+        total_size=None,
+        transform=UNSET,
+        **kwargs,
+    ) -> RandomVariable:
+        """Adds a RandomVariable corresponding to a PyMC3 distribution to the current model.
+
+        Note that all remaining kwargs must be compatible with .dist()
+
+        Parameters
+        ----------
+        cls : type
+            A PyMC3 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.
+        testval : optional
+            Test value to be attached to the output RV.
+            Must match its shape exactly.
+        observed : optional
+            Observed data to be passed when registering the random variable in the model.
+            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().
+            Most prominently: `shape` and `size`
+
+        Returns
+        -------
+        rv : RandomVariable
+            The created RV, registered in the Model.
+        """
+
         try:
             from pymc3.model import Model
 
@@ -141,40 +242,121 @@ def __new__(cls, name, *args, **kwargs):
                 "for a standalone distribution."
             )
 
-        rng = kwargs.pop("rng", None)
+        if not isinstance(name, string_types):
+            raise TypeError(f"Name needs to be a string but got: {name}")
 
         if rng is None:
             rng = model.default_rng
 
-        if not isinstance(name, string_types):
-            raise TypeError(f"Name needs to be a string but got: {name}")
+        _, dims, _ = _validate_shape_dims_size(dims=dims)
 
-        data = kwargs.pop("observed", None)
+        # Create the RV without specifying testval, because the testval may
+        # have a shape that only matches after replicating with a size implied
+        # by dims (see below).
+        rv_out = cls.dist(*args, rng=rng, testval=None, **kwargs)
 
         total_size = kwargs.pop("total_size", None)
 
-        dims = kwargs.pop("dims", None)
+        # `dims` are only available with this API, because `.dist()` can be used
+        # without a modelcontext and dims are not yet tracked at the Aesara level.
+        if dims is not None:
+            if Ellipsis in dims:
+                # Auto-complete the dims tuple to the full length
+                dims = (*dims[:-1], *[None] * rv_out.ndim)
+
+            n_implied = rv_out.ndim
+            n_size = len(dims) - n_implied
+
+            # All size dims must be known already (numerically or symbolically).
+            unknown_size_dims = set(dims[:n_size]) - set(model.dim_lengths)
+            if unknown_size_dims:
+                raise KeyError(
+                    f"Dimensions {unknown_size_dims} are unknown to the model and cannot be used to specify a `size`."
+                )
 
-        if "shape" in kwargs:
-            raise DeprecationWarning("The `shape` keyword is deprecated; use `size`.")
+            # The numeric/symbolic size tuple can be created using model.RV_dim_lengths
+            size = tuple(model.dim_lengths[dname] for dname in dims[:n_size])
 
-        transform = kwargs.pop("transform", UNSET)
+            if size:
+                # A batch size was specified through dims!
+                rv_out = change_rv_size(rv_var=rv_out, new_size=size, expand=True)
 
-        rv_out = cls.dist(*args, rng=rng, **kwargs)
+            # Now that we have a handle on the output RV, we can register named implied dimensions that
+            # were not yet known to the model, such that they can be used for size further downstream.
+            for di, dname in enumerate(dims[n_size:]):
+                if not dname in model.dim_lengths:
+                    model.add_coord(dname, values=None, length=rv_out.shape[n_size + di])
+
+        if testval is not None:
+            # Assigning the testval earlier causes trouble because the RV may not be created with the final shape already.
+            rv_out.tag.test_value = testval
 
-        return model.register_rv(rv_out, name, data, total_size, dims=dims, transform=transform)
+        return model.register_rv(rv_out, name, observed, total_size, dims=dims, transform=transform)
 
     @classmethod
-    def dist(cls, dist_params, **kwargs):
+    def dist(
+        cls,
+        dist_params,
+        *,
+        shape: Optional[Shape] = None,
+        size: Optional[Size] = None,
+        testval=None,
+        **kwargs,
+    ) -> RandomVariable:
+        """Creates a RandomVariable corresponding to the `cls` distribution.
+
+        Parameters
+        ----------
+        dist_params
+        shape : tuple, optional
+            A tuple of sizes for each dimension of the new RV.
+
+            Ellipsis (...) may be used in the last position of the tuple,
+            and automatically expand to the shape implied by RV inputs.
 
-        testval = kwargs.pop("testval", None)
+        size : int, tuple, Variable, optional
+            A scalar or tuple for replicating the RV in addition
+            to its implied shape/dimensionality.
+        testval : optional
+            Test value to be attached to the output RV.
+            Must match its shape exactly.
 
-        rv_var = cls.rv_op(*dist_params, **kwargs)
+        Returns
+        -------
+        rv : RandomVariable
+            The created RV.
+        """
+        if "dims" in kwargs:
+            raise NotImplementedError("The use of a `.dist(dims=...)` API is not yet supported.")
+
+        shape, _, size = _validate_shape_dims_size(shape=shape, size=size)
+        # Create the RV without specifying size or testval.
+        # The size will be expanded later (if necessary) and only then the testval fits.
+        rv_native = cls.rv_op(*dist_params, size=None, **kwargs)
+
+        if shape is None and size is None:
+            size = ()
+        elif shape is not None:
+            if isinstance(shape, Variable):
+                size = ()
+            else:
+                if Ellipsis in shape:
+                    size = tuple(shape[:-1])
+                else:
+                    size = tuple(shape[: len(shape) - rv_native.ndim])
+        # no-op conditions:
+        # `elif size is not None` (User already specified how to expand the RV)
+        # `else` (Unreachable)
+
+        if size:
+            rv_out = change_rv_size(rv_var=rv_native, new_size=size, expand=True)
+        else:
+            rv_out = rv_native
 
         if testval is not None:
-            rv_var.tag.test_value = testval
+            rv_out.tag.test_value = testval
 
-        return rv_var
+        return rv_out
 
     def _distr_parameters_for_repr(self):
         """Return the names of the parameters for this distribution (e.g. "mu"

pymc3/model.py

@@ -959,17 +959,18 @@ def add_coord(
         ----------
         name : str
             Name of the dimension.
-            Forbidden: {"chain", "draw"}
+            Forbidden: {"chain", "draw", "__samaple__"}
         values : optional, array-like
             Coordinate values or ``None`` (for auto-numbering).
             If ``None`` is passed, a ``length`` must be specified.
         length : optional, scalar
             A symbolic scalar of the dimensions length.
             Defaults to ``aesara.shared(len(values))``.
         """
-        if name in {"draw", "chain"}:
+        if name in {"draw", "chain", "__sample__"}:
             raise ValueError(
-                "Dimensions can not be named `draw` or `chain`, as they are reserved for the sampler's outputs."
+                "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(
@@ -981,7 +982,7 @@ def add_coord(
             )
         if name in self.coords:
             if not values.equals(self.coords[name]):
-                raise ValueError("Duplicate and incompatiple coordinate: %s." % name)
+                raise ValueError(f"Duplicate and incompatiple coordinate: {name}.")
         else:
             self._coords[name] = values
             self._dim_lengths[name] = length or aesara.shared(len(values))
@@ -1019,7 +1020,8 @@ def set_data(
             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.
+            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):

pymc3/tests/sampler_fixtures.py

@@ -92,7 +92,7 @@ class BetaBinomialFixture(KnownCDF):
     @classmethod
     def make_model(cls):
         with pm.Model() as model:
-            p = pm.Beta("p", [0.5, 0.5, 1.0], [0.5, 0.5, 1.0], size=3)
+            p = pm.Beta("p", [0.5, 0.5, 1.0], [0.5, 0.5, 1.0])
             pm.Binomial("y", p=p, n=[4, 12, 9], observed=[1, 2, 9])
         return model