scipy 1.16.2__cp312-cp312-win_arm64.whl

scipy/_lib/array_api_compat/torch/_info.py

@@ -0,0 +1,369 @@
+"""
+Array API Inspection namespace
+
+This is the namespace for inspection functions as defined by the array API
+standard. See
+https://data-apis.org/array-api/latest/API_specification/inspection.html for
+more details.
+
+"""
+import torch
+
+from functools import cache
+
+class __array_namespace_info__:
+    """
+    Get the array API inspection namespace for PyTorch.
+
+    The array API inspection namespace defines the following functions:
+
+    - capabilities()
+    - default_device()
+    - default_dtypes()
+    - dtypes()
+    - devices()
+
+    See
+    https://data-apis.org/array-api/latest/API_specification/inspection.html
+    for more details.
+
+    Returns
+    -------
+    info : ModuleType
+        The array API inspection namespace for PyTorch.
+
+    Examples
+    --------
+    >>> info = xp.__array_namespace_info__()
+    >>> info.default_dtypes()
+    {'real floating': numpy.float64,
+     'complex floating': numpy.complex128,
+     'integral': numpy.int64,
+     'indexing': numpy.int64}
+
+    """
+
+    __module__ = 'torch'
+
+    def capabilities(self):
+        """
+        Return a dictionary of array API library capabilities.
+
+        The resulting dictionary has the following keys:
+
+        - **"boolean indexing"**: boolean indicating whether an array library
+          supports boolean indexing. Always ``True`` for PyTorch.
+
+        - **"data-dependent shapes"**: boolean indicating whether an array
+          library supports data-dependent output shapes. Always ``True`` for
+          PyTorch.
+
+        See
+        https://data-apis.org/array-api/latest/API_specification/generated/array_api.info.capabilities.html
+        for more details.
+
+        See Also
+        --------
+        __array_namespace_info__.default_device,
+        __array_namespace_info__.default_dtypes,
+        __array_namespace_info__.dtypes,
+        __array_namespace_info__.devices
+
+        Returns
+        -------
+        capabilities : dict
+            A dictionary of array API library capabilities.
+
+        Examples
+        --------
+        >>> info = xp.__array_namespace_info__()
+        >>> info.capabilities()
+        {'boolean indexing': True,
+         'data-dependent shapes': True,
+         'max dimensions': 64}
+
+        """
+        return {
+            "boolean indexing": True,
+            "data-dependent shapes": True,
+            "max dimensions": 64,
+        }
+
+    def default_device(self):
+        """
+        The default device used for new PyTorch arrays.
+
+        See Also
+        --------
+        __array_namespace_info__.capabilities,
+        __array_namespace_info__.default_dtypes,
+        __array_namespace_info__.dtypes,
+        __array_namespace_info__.devices
+
+        Returns
+        -------
+        device : Device
+            The default device used for new PyTorch arrays.
+
+        Examples
+        --------
+        >>> info = xp.__array_namespace_info__()
+        >>> info.default_device()
+        device(type='cpu')
+
+        Notes
+        -----
+        This method returns the static default device when PyTorch is initialized.
+        However, the *current* device used by creation functions (``empty`` etc.)
+        can be changed at runtime.
+
+        See Also
+        --------
+        https://github.com/data-apis/array-api/issues/835
+        """
+        return torch.device("cpu")
+
+    def default_dtypes(self, *, device=None):
+        """
+        The default data types used for new PyTorch arrays.
+
+        Parameters
+        ----------
+        device : Device, optional
+            The device to get the default data types for.
+            Unused for PyTorch, as all devices use the same default dtypes.
+
+        Returns
+        -------
+        dtypes : dict
+            A dictionary describing the default data types used for new PyTorch
+            arrays.
+
+        See Also
+        --------
+        __array_namespace_info__.capabilities,
+        __array_namespace_info__.default_device,
+        __array_namespace_info__.dtypes,
+        __array_namespace_info__.devices
+
+        Examples
+        --------
+        >>> info = xp.__array_namespace_info__()
+        >>> info.default_dtypes()
+        {'real floating': torch.float32,
+         'complex floating': torch.complex64,
+         'integral': torch.int64,
+         'indexing': torch.int64}
+
+        """
+        # Note: if the default is set to float64, the devices like MPS that
+        # don't support float64 will error. We still return the default_dtype
+        # value here because this error doesn't represent a different default
+        # per-device.
+        default_floating = torch.get_default_dtype()
+        default_complex = torch.complex64 if default_floating == torch.float32 else torch.complex128
+        default_integral = torch.int64
+        return {
+            "real floating": default_floating,
+            "complex floating": default_complex,
+            "integral": default_integral,
+            "indexing": default_integral,
+        }
+
+
+    def _dtypes(self, kind):
+        bool = torch.bool
+        int8 = torch.int8
+        int16 = torch.int16
+        int32 = torch.int32
+        int64 = torch.int64
+        uint8 = torch.uint8
+        # uint16, uint32, and uint64 are present in newer versions of pytorch,
+        # but they aren't generally supported by the array API functions, so
+        # we omit them from this function.
+        float32 = torch.float32
+        float64 = torch.float64
+        complex64 = torch.complex64
+        complex128 = torch.complex128
+
+        if kind is None:
+            return {
+                "bool": bool,
+                "int8": int8,
+                "int16": int16,
+                "int32": int32,
+                "int64": int64,
+                "uint8": uint8,
+                "float32": float32,
+                "float64": float64,
+                "complex64": complex64,
+                "complex128": complex128,
+            }
+        if kind == "bool":
+            return {"bool": bool}
+        if kind == "signed integer":
+            return {
+                "int8": int8,
+                "int16": int16,
+                "int32": int32,
+                "int64": int64,
+            }
+        if kind == "unsigned integer":
+            return {
+                "uint8": uint8,
+            }
+        if kind == "integral":
+            return {
+                "int8": int8,
+                "int16": int16,
+                "int32": int32,
+                "int64": int64,
+                "uint8": uint8,
+            }
+        if kind == "real floating":
+            return {
+                "float32": float32,
+                "float64": float64,
+            }
+        if kind == "complex floating":
+            return {
+                "complex64": complex64,
+                "complex128": complex128,
+            }
+        if kind == "numeric":
+            return {
+                "int8": int8,
+                "int16": int16,
+                "int32": int32,
+                "int64": int64,
+                "uint8": uint8,
+                "float32": float32,
+                "float64": float64,
+                "complex64": complex64,
+                "complex128": complex128,
+            }
+        if isinstance(kind, tuple):
+            res = {}
+            for k in kind:
+                res.update(self.dtypes(kind=k))
+            return res
+        raise ValueError(f"unsupported kind: {kind!r}")
+
+    @cache
+    def dtypes(self, *, device=None, kind=None):
+        """
+        The array API data types supported by PyTorch.
+
+        Note that this function only returns data types that are defined by
+        the array API.
+
+        Parameters
+        ----------
+        device : Device, optional
+            The device to get the data types for.
+            Unused for PyTorch, as all devices use the same dtypes.
+        kind : str or tuple of str, optional
+            The kind of data types to return. If ``None``, all data types are
+            returned. If a string, only data types of that kind are returned.
+            If a tuple, a dictionary containing the union of the given kinds
+            is returned. The following kinds are supported:
+
+            - ``'bool'``: boolean data types (i.e., ``bool``).
+            - ``'signed integer'``: signed integer data types (i.e., ``int8``,
+              ``int16``, ``int32``, ``int64``).
+            - ``'unsigned integer'``: unsigned integer data types (i.e.,
+              ``uint8``, ``uint16``, ``uint32``, ``uint64``).
+            - ``'integral'``: integer data types. Shorthand for ``('signed
+              integer', 'unsigned integer')``.
+            - ``'real floating'``: real-valued floating-point data types
+              (i.e., ``float32``, ``float64``).
+            - ``'complex floating'``: complex floating-point data types (i.e.,
+              ``complex64``, ``complex128``).
+            - ``'numeric'``: numeric data types. Shorthand for ``('integral',
+              'real floating', 'complex floating')``.
+
+        Returns
+        -------
+        dtypes : dict
+            A dictionary mapping the names of data types to the corresponding
+            PyTorch data types.
+
+        See Also
+        --------
+        __array_namespace_info__.capabilities,
+        __array_namespace_info__.default_device,
+        __array_namespace_info__.default_dtypes,
+        __array_namespace_info__.devices
+
+        Examples
+        --------
+        >>> info = xp.__array_namespace_info__()
+        >>> info.dtypes(kind='signed integer')
+        {'int8': numpy.int8,
+         'int16': numpy.int16,
+         'int32': numpy.int32,
+         'int64': numpy.int64}
+
+        """
+        res = self._dtypes(kind)
+        for k, v in res.copy().items():
+            try:
+                torch.empty((0,), dtype=v, device=device)
+            except:
+                del res[k]
+        return res
+
+    @cache
+    def devices(self):
+        """
+        The devices supported by PyTorch.
+
+        Returns
+        -------
+        devices : list[Device]
+            The devices supported by PyTorch.
+
+        See Also
+        --------
+        __array_namespace_info__.capabilities,
+        __array_namespace_info__.default_device,
+        __array_namespace_info__.default_dtypes,
+        __array_namespace_info__.dtypes
+
+        Examples
+        --------
+        >>> info = xp.__array_namespace_info__()
+        >>> info.devices()
+        [device(type='cpu'), device(type='mps', index=0), device(type='meta')]
+
+        """
+        # Torch doesn't have a straightforward way to get the list of all
+        # currently supported devices. To do this, we first parse the error
+        # message of torch.device to get the list of all possible types of
+        # device:
+        try:
+            torch.device('notadevice')
+            raise AssertionError("unreachable")  # pragma: nocover
+        except RuntimeError as e:
+            # The error message is something like:
+            # "Expected one of cpu, cuda, ipu, xpu, mkldnn, opengl, opencl, ideep, hip, ve, fpga, ort, xla, lazy, vulkan, mps, meta, hpu, mtia, privateuseone device type at start of device string: notadevice"
+            devices_names = e.args[0].split('Expected one of ')[1].split(' device type')[0].split(', ')
+
+        # Next we need to check for different indices for different devices.
+        # device(device_name, index=index) doesn't actually check if the
+        # device name or index is valid. We have to try to create a tensor
+        # with it (which is why this function is cached).
+        devices = []
+        for device_name in devices_names:
+            i = 0
+            while True:
+                try:
+                    a = torch.empty((0,), device=torch.device(device_name, index=i))
+                    if a.device in devices:
+                        break
+                    devices.append(a.device)
+                except:
+                    break
+                i += 1
+
+        return devices

scipy/_lib/array_api_compat/torch/_typing.py

@@ -0,0 +1,3 @@
+__all__ = ["Array", "Device", "DType"]
+
+from torch import device as Device, dtype as DType, Tensor as Array

scipy/_lib/array_api_compat/torch/fft.py

@@ -0,0 +1,85 @@
+from __future__ import annotations
+
+from typing import Union, Sequence, Literal
+
+import torch
+import torch.fft
+from torch.fft import * # noqa: F403
+
+from ._typing import Array
+
+# Several torch fft functions do not map axes to dim
+
+def fftn(
+    x: Array,
+    /,
+    *,
+    s: Sequence[int] = None,
+    axes: Sequence[int] = None,
+    norm: Literal["backward", "ortho", "forward"] = "backward",
+    **kwargs,
+) -> Array:
+    return torch.fft.fftn(x, s=s, dim=axes, norm=norm, **kwargs)
+
+def ifftn(
+    x: Array,
+    /,
+    *,
+    s: Sequence[int] = None,
+    axes: Sequence[int] = None,
+    norm: Literal["backward", "ortho", "forward"] = "backward",
+    **kwargs,
+) -> Array:
+    return torch.fft.ifftn(x, s=s, dim=axes, norm=norm, **kwargs)
+
+def rfftn(
+    x: Array,
+    /,
+    *,
+    s: Sequence[int] = None,
+    axes: Sequence[int] = None,
+    norm: Literal["backward", "ortho", "forward"] = "backward",
+    **kwargs,
+) -> Array:
+    return torch.fft.rfftn(x, s=s, dim=axes, norm=norm, **kwargs)
+
+def irfftn(
+    x: Array,
+    /,
+    *,
+    s: Sequence[int] = None,
+    axes: Sequence[int] = None,
+    norm: Literal["backward", "ortho", "forward"] = "backward",
+    **kwargs,
+) -> Array:
+    return torch.fft.irfftn(x, s=s, dim=axes, norm=norm, **kwargs)
+
+def fftshift(
+    x: Array,
+    /,
+    *,
+    axes: Union[int, Sequence[int]] = None,
+    **kwargs,
+) -> Array:
+    return torch.fft.fftshift(x, dim=axes, **kwargs)
+
+def ifftshift(
+    x: Array,
+    /,
+    *,
+    axes: Union[int, Sequence[int]] = None,
+    **kwargs,
+) -> Array:
+    return torch.fft.ifftshift(x, dim=axes, **kwargs)
+
+
+__all__ = torch.fft.__all__ + [
+    "fftn",
+    "ifftn",
+    "rfftn",
+    "irfftn",
+    "fftshift",
+    "ifftshift",
+]
+
+_all_ignore = ['torch']

scipy/_lib/array_api_compat/torch/linalg.py

@@ -0,0 +1,121 @@
+from __future__ import annotations
+
+import torch
+from typing import Optional, Union, Tuple
+
+from torch.linalg import * # noqa: F403
+
+# torch.linalg doesn't define __all__
+# from torch.linalg import __all__ as linalg_all
+from torch import linalg as torch_linalg
+linalg_all = [i for i in dir(torch_linalg) if not i.startswith('_')]
+
+# outer is implemented in torch but aren't in the linalg namespace
+from torch import outer
+from ._aliases import _fix_promotion, sum
+# These functions are in both the main and linalg namespaces
+from ._aliases import matmul, matrix_transpose, tensordot
+from ._typing import Array, DType
+from ..common._typing import JustInt, JustFloat
+
+# Note: torch.linalg.cross does not default to axis=-1 (it defaults to the
+# first axis with size 3), see https://github.com/pytorch/pytorch/issues/58743
+
+# torch.cross also does not support broadcasting when it would add new
+# dimensions https://github.com/pytorch/pytorch/issues/39656
+def cross(x1: Array, x2: Array, /, *, axis: int = -1) -> Array:
+    x1, x2 = _fix_promotion(x1, x2, only_scalar=False)
+    if not (-min(x1.ndim, x2.ndim) <= axis < max(x1.ndim, x2.ndim)):
+        raise ValueError(f"axis {axis} out of bounds for cross product of arrays with shapes {x1.shape} and {x2.shape}")
+    if not (x1.shape[axis] == x2.shape[axis] == 3):
+        raise ValueError(f"cross product axis must have size 3, got {x1.shape[axis]} and {x2.shape[axis]}")
+    x1, x2 = torch.broadcast_tensors(x1, x2)
+    return torch_linalg.cross(x1, x2, dim=axis)
+
+def vecdot(x1: Array, x2: Array, /, *, axis: int = -1, **kwargs) -> Array:
+    from ._aliases import isdtype
+
+    x1, x2 = _fix_promotion(x1, x2, only_scalar=False)
+
+    # torch.linalg.vecdot incorrectly allows broadcasting along the contracted dimension
+    if x1.shape[axis] != x2.shape[axis]:
+        raise ValueError("x1 and x2 must have the same size along the given axis")
+
+    # torch.linalg.vecdot doesn't support integer dtypes
+    if isdtype(x1.dtype, 'integral') or isdtype(x2.dtype, 'integral'):
+        if kwargs:
+            raise RuntimeError("vecdot kwargs not supported for integral dtypes")
+
+        x1_ = torch.moveaxis(x1, axis, -1)
+        x2_ = torch.moveaxis(x2, axis, -1)
+        x1_, x2_ = torch.broadcast_tensors(x1_, x2_)
+
+        res = x1_[..., None, :] @ x2_[..., None]
+        return res[..., 0, 0]
+    return torch.linalg.vecdot(x1, x2, dim=axis, **kwargs)
+
+def solve(x1: Array, x2: Array, /, **kwargs) -> Array:
+    x1, x2 = _fix_promotion(x1, x2, only_scalar=False)
+    # Torch tries to emulate NumPy 1 solve behavior by using batched 1-D solve
+    # whenever
+    # 1. x1.ndim - 1 == x2.ndim
+    # 2. x1.shape[:-1] == x2.shape
+    #
+    # See linalg_solve_is_vector_rhs in
+    # aten/src/ATen/native/LinearAlgebraUtils.h and
+    # TORCH_META_FUNC(_linalg_solve_ex) in
+    # aten/src/ATen/native/BatchLinearAlgebra.cpp in the PyTorch source code.
+    #
+    # The easiest way to work around this is to prepend a size 1 dimension to
+    # x2, since x2 is already one dimension less than x1.
+    #
+    # See https://github.com/pytorch/pytorch/issues/52915
+    if x2.ndim != 1 and x1.ndim - 1 == x2.ndim and x1.shape[:-1] == x2.shape:
+        x2 = x2[None]
+    return torch.linalg.solve(x1, x2, **kwargs)
+
+# torch.trace doesn't support the offset argument and doesn't support stacking
+def trace(x: Array, /, *, offset: int = 0, dtype: Optional[DType] = None) -> Array:
+    # Use our wrapped sum to make sure it does upcasting correctly
+    return sum(torch.diagonal(x, offset=offset, dim1=-2, dim2=-1), axis=-1, dtype=dtype)
+
+def vector_norm(
+    x: Array,
+    /,
+    *,
+    axis: Optional[Union[int, Tuple[int, ...]]] = None,
+    keepdims: bool = False,
+    # JustFloat stands for inf | -inf, which are not valid for Literal
+    ord: JustInt | JustFloat = 2,
+    **kwargs,
+) -> Array:
+    # torch.vector_norm incorrectly treats axis=() the same as axis=None
+    if axis == ():
+        out = kwargs.get('out')
+        if out is None:
+            dtype = None
+            if x.dtype == torch.complex64:
+                dtype = torch.float32
+            elif x.dtype == torch.complex128:
+                dtype = torch.float64
+
+            out = torch.zeros_like(x, dtype=dtype)
+
+        # The norm of a single scalar works out to abs(x) in every case except
+        # for ord=0, which is x != 0.
+        if ord == 0:
+            out[:] = (x != 0)
+        else:
+            out[:] = torch.abs(x)
+        return out
+    return torch.linalg.vector_norm(x, ord=ord, axis=axis, keepdim=keepdims, **kwargs)
+
+__all__ = linalg_all + ['outer', 'matmul', 'matrix_transpose', 'tensordot',
+                        'cross', 'vecdot', 'solve', 'trace', 'vector_norm']
+
+_all_ignore = ['torch_linalg', 'sum']
+
+del linalg_all
+
+def __dir__() -> list[str]:
+    return __all__

scipy/_lib/array_api_extra/__init__.py

@@ -0,0 +1,38 @@
+"""Extra array functions built on top of the array API standard."""
+
+from ._delegation import isclose, pad
+from ._lib._at import at
+from ._lib._funcs import (
+    apply_where,
+    atleast_nd,
+    broadcast_shapes,
+    cov,
+    create_diagonal,
+    expand_dims,
+    kron,
+    nunique,
+    setdiff1d,
+    sinc,
+)
+from ._lib._lazy import lazy_apply
+
+__version__ = "0.8.0.dev0"
+
+# pylint: disable=duplicate-code
+__all__ = [
+    "__version__",
+    "apply_where",
+    "at",
+    "atleast_nd",
+    "broadcast_shapes",
+    "cov",
+    "create_diagonal",
+    "expand_dims",
+    "isclose",
+    "kron",
+    "lazy_apply",
+    "nunique",
+    "pad",
+    "setdiff1d",
+    "sinc",
+]