|
| 1 | +# Copyright 2025 - present The PyMC Developers |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | +from collections.abc import Callable |
| 15 | +from itertools import chain |
| 16 | + |
| 17 | +from pytensor.xtensor import as_xtensor |
| 18 | +from pytensor.xtensor.type import XTensorVariable |
| 19 | + |
| 20 | +from pymc import modelcontext |
| 21 | +from pymc.distributions import transforms |
| 22 | +from pymc.distributions.shape_utils import Dims, convert_dims |
| 23 | +from pymc.util import UNSET |
| 24 | + |
| 25 | + |
| 26 | +class DimDistribution: |
| 27 | + """Base class for PyMC distribution that wrap pytensor.xtensor.random operations, and follow xarray-like semantics.""" |
| 28 | + |
| 29 | + xrv_op: Callable |
| 30 | + default_transform: Callable | None |
| 31 | + |
| 32 | + def __new__( |
| 33 | + cls, |
| 34 | + name: str, |
| 35 | + *dist_params, |
| 36 | + dims: Dims | None = None, |
| 37 | + initval=None, |
| 38 | + observed=None, |
| 39 | + total_size=None, |
| 40 | + transform=UNSET, |
| 41 | + default_transform=UNSET, |
| 42 | + model=None, |
| 43 | + **kwargs, |
| 44 | + ) -> XTensorVariable: |
| 45 | + try: |
| 46 | + model = modelcontext(model) |
| 47 | + except TypeError: |
| 48 | + raise TypeError( |
| 49 | + "No model on context stack, which is needed to instantiate distributions. " |
| 50 | + "Add variable inside a 'with model:' block, or use the '.dist' syntax for a standalone distribution." |
| 51 | + ) |
| 52 | + |
| 53 | + if not isinstance(name, str): |
| 54 | + raise TypeError(f"Name needs to be a string but got: {name}") |
| 55 | + |
| 56 | + if dims is None: |
| 57 | + dims_dict = {} |
| 58 | + else: |
| 59 | + dims = convert_dims(dims) |
| 60 | + try: |
| 61 | + dims_dict = { |
| 62 | + dim: model.dim_lengths[dim] for dim in dims if dim in model.dim_lengths |
| 63 | + } |
| 64 | + except KeyError: |
| 65 | + raise ValueError( |
| 66 | + f"Not all dims {dims} are part of the model coords. " |
| 67 | + f"Add them at initialization time or use `model.add_coord` before defining the distribution." |
| 68 | + ) |
| 69 | + |
| 70 | + if observed is not None: |
| 71 | + try: |
| 72 | + observed = as_xtensor(observed) |
| 73 | + except TypeError as e: |
| 74 | + raise TypeError( |
| 75 | + "Observed value must have dims associated with it.\n" |
| 76 | + "To avoid subtle bugs, PyMC does not make any assumptions about the dims of the observed value. " |
| 77 | + "Convert observed value to an xarray.DataArray, pymc.dims.Data or pytensor.xtensor.as_xtensor with explicit dims." |
| 78 | + ) from e |
| 79 | + |
| 80 | + # Propagate observed dims to dims_dict |
| 81 | + for observed_dim in observed.type.dims: |
| 82 | + if observed_dim not in dims_dict: |
| 83 | + dims_dict[observed_dim] = model.dim_lengths[observed_dim] |
| 84 | + |
| 85 | + rv = cls.dist(*dist_params, dims_dict=dims_dict, **kwargs) |
| 86 | + |
| 87 | + # User provided dims must specify all dims or use ellipsis |
| 88 | + if dims is not None: |
| 89 | + if (... not in dims) and (set(dims) != set(rv.type.dims)): |
| 90 | + raise ValueError( |
| 91 | + f"Provided dims {dims} do not match the distribution's output dims {rv.type.dims}. " |
| 92 | + "Use ellipsis to specify all other dimensions." |
| 93 | + ) |
| 94 | + # Use provided dims to transpose the output to the desired order |
| 95 | + rv = rv.transpose(*dims) |
| 96 | + |
| 97 | + rv_dims = rv.type.dims |
| 98 | + if observed is None: |
| 99 | + if default_transform is UNSET: |
| 100 | + default_transform = cls.default_transform |
| 101 | + else: |
| 102 | + # Align observed dims with those of the RV |
| 103 | + observed = observed.transpose(*rv_dims).values |
| 104 | + |
| 105 | + rv = model.register_rv( |
| 106 | + rv.values, |
| 107 | + name=name, |
| 108 | + observed=observed, |
| 109 | + total_size=total_size, |
| 110 | + dims=rv_dims, |
| 111 | + transform=transform, |
| 112 | + default_transform=default_transform, |
| 113 | + initval=initval, |
| 114 | + ) |
| 115 | + |
| 116 | + xrv = as_xtensor(rv, dims=rv_dims) |
| 117 | + return xrv |
| 118 | + |
| 119 | + @classmethod |
| 120 | + def dist( |
| 121 | + cls, |
| 122 | + dist_params, |
| 123 | + *, |
| 124 | + dims_dict: dict[str, int] | None = None, |
| 125 | + **kwargs, |
| 126 | + ) -> XTensorVariable: |
| 127 | + for invalid_kwarg in ("size", "shape", "dims"): |
| 128 | + if invalid_kwarg in kwargs: |
| 129 | + raise TypeError(f"DimDistribution does not accept {invalid_kwarg} argument.") |
| 130 | + |
| 131 | + # XRV requires only extra_dims, not dims |
| 132 | + try: |
| 133 | + dist_params = [as_xtensor(param) for param in dist_params] |
| 134 | + except TypeError as e: |
| 135 | + raise TypeError( |
| 136 | + "Distribution parameters must have dims associated with it.\n" |
| 137 | + "To avoid subtle bugs, PyMC does not make any assumptions about the dims of the parameters. " |
| 138 | + "Convert parameters to an xarray.DataArray, pymc.dims.Data or pytensor.xtensor.as_xtensor with explicit dims." |
| 139 | + ) |
| 140 | + |
| 141 | + if dims_dict is None: |
| 142 | + extra_dims = None |
| 143 | + else: |
| 144 | + parameter_implied_dims = set( |
| 145 | + chain.from_iterable(param.type.dims for param in dist_params) |
| 146 | + ) |
| 147 | + extra_dims = { |
| 148 | + dim: length |
| 149 | + for dim, length in dims_dict.items() |
| 150 | + if dim not in parameter_implied_dims |
| 151 | + } |
| 152 | + return cls.xrv_op(*dist_params, extra_dims=extra_dims, **kwargs) |
| 153 | + |
| 154 | + |
| 155 | +class ContinuousDimDistribution(DimDistribution): |
| 156 | + """Base class for real-valued distributions.""" |
| 157 | + |
| 158 | + default_transform = None |
| 159 | + |
| 160 | + |
| 161 | +class PositiveContinuousDimDistribution(DimDistribution): |
| 162 | + """Base class for positive continuous distributions.""" |
| 163 | + |
| 164 | + default_transform = transforms.log |
| 165 | + |
| 166 | + |
| 167 | +class UnitDimDistribution(DimDistribution): |
| 168 | + """Base class for unit-valued distributions.""" |
| 169 | + |
| 170 | + default_transform = transforms.logodds |
0 commit comments