This commit is contained in:
cjw
2026-02-12 23:22:11 +08:00
parent 7b09eb3d89
commit 89660bba4e
5988 changed files with 2517516 additions and 0 deletions
@@ -0,0 +1,161 @@
# This file is generated by SciPy's build process
# It contains system_info results at the time of building this package.
from enum import Enum
__all__ = ["show"]
_built_with_meson = True
class DisplayModes(Enum):
stdout = "stdout"
dicts = "dicts"
def _cleanup(d):
"""
Removes empty values in a `dict` recursively
This ensures we remove values that Meson could not provide to CONFIG
"""
if isinstance(d, dict):
return { k: _cleanup(v) for k, v in d.items() if v != '' and _cleanup(v) != '' }
else:
return d
CONFIG = _cleanup(
{
"Compilers": {
"c": {
"name": "gcc",
"linker": r"ld.bfd",
"version": "10.2.1",
"commands": r"cc",
"args": r"",
"linker args": r"",
},
"cython": {
"name": r"cython",
"linker": r"cython",
"version": r"3.1.6",
"commands": r"cython",
"args": r"",
"linker args": r"",
},
"c++": {
"name": "gcc",
"linker": r"ld.bfd",
"version": "10.2.1",
"commands": r"c++",
"args": r"",
"linker args": r"",
},
"fortran": {
"name": "gcc",
"linker": r"ld.bfd",
"version": "10.2.1",
"commands": r"gfortran",
"args": r"",
"linker args": r"",
},
"pythran": {
"version": r"0.18.0",
"include directory": r"../../tmp/build-env-b9dts1ag/lib/python3.11/site-packages/pythran"
},
},
"Machine Information": {
"host": {
"cpu": r"x86_64",
"family": r"x86_64",
"endian": r"little",
"system": r"linux",
},
"build": {
"cpu": r"x86_64",
"family": r"x86_64",
"endian": r"little",
"system": r"linux",
},
"cross-compiled": bool("False".lower().replace('false', '')),
},
"Build Dependencies": {
"blas": {
"name": "scipy-openblas",
"found": bool("True".lower().replace('false', '')),
"version": "0.3.29.dev",
"detection method": "pkgconfig",
"include directory": r"/opt/_internal/cpython-3.11.13/lib/python3.11/site-packages/scipy_openblas32/include",
"lib directory": r"/opt/_internal/cpython-3.11.13/lib/python3.11/site-packages/scipy_openblas32/lib",
"openblas configuration": r"OpenBLAS 0.3.29.dev DYNAMIC_ARCH NO_AFFINITY SkylakeX MAX_THREADS=64",
"pc file directory": r"/project",
},
"lapack": {
"name": "scipy-openblas",
"found": bool("True".lower().replace('false', '')),
"version": "0.3.29.dev",
"detection method": "pkgconfig",
"include directory": r"/opt/_internal/cpython-3.11.13/lib/python3.11/site-packages/scipy_openblas32/include",
"lib directory": r"/opt/_internal/cpython-3.11.13/lib/python3.11/site-packages/scipy_openblas32/lib",
"openblas configuration": r"OpenBLAS 0.3.29.dev DYNAMIC_ARCH NO_AFFINITY SkylakeX MAX_THREADS=64",
"pc file directory": r"/project",
},
"pybind11": {
"name": "pybind11",
"version": "3.0.1",
"detection method": "config-tool",
"include directory": r"unknown",
},
},
"Python Information": {
"path": r"/tmp/build-env-b9dts1ag/bin/python",
"version": "3.11",
},
}
)
def _check_pyyaml():
import yaml
return yaml
def show(mode=DisplayModes.stdout.value):
"""
Show libraries and system information on which SciPy was built
and is being used
Parameters
----------
mode : {`'stdout'`, `'dicts'`}, optional.
Indicates how to display the config information.
`'stdout'` prints to console, `'dicts'` returns a dictionary
of the configuration.
Returns
-------
out : {`dict`, `None`}
If mode is `'dicts'`, a dict is returned, else None
Notes
-----
1. The `'stdout'` mode will give more readable
output if ``pyyaml`` is installed
"""
if mode == DisplayModes.stdout.value:
try: # Non-standard library, check import
yaml = _check_pyyaml()
print(yaml.dump(CONFIG))
except ModuleNotFoundError:
import warnings
import json
warnings.warn("Install `pyyaml` for better output", stacklevel=1)
print(json.dumps(CONFIG, indent=2))
elif mode == DisplayModes.dicts.value:
return CONFIG
else:
raise AttributeError(
f"Invalid `mode`, use one of: {', '.join([e.value for e in DisplayModes])}"
)
@@ -0,0 +1,138 @@
"""
SciPy: A scientific computing package for Python
================================================
Documentation is available in the docstrings and
online at https://docs.scipy.org/doc/scipy/
Subpackages
-----------
::
cluster --- Vector Quantization / Kmeans
constants --- Physical and mathematical constants and units
datasets --- Dataset methods
differentiate --- Finite difference differentiation tools
fft --- Discrete Fourier transforms
fftpack --- Legacy discrete Fourier transforms
integrate --- Integration routines
interpolate --- Interpolation Tools
io --- Data input and output
linalg --- Linear algebra routines
ndimage --- N-D image package
odr --- Orthogonal Distance Regression
optimize --- Optimization Tools
signal --- Signal Processing Tools
sparse --- Sparse Matrices
spatial --- Spatial data structures and algorithms
special --- Special functions
stats --- Statistical Functions
Public API in the main SciPy namespace
--------------------------------------
::
__version__ --- SciPy version string
LowLevelCallable --- Low-level callback function
show_config --- Show scipy build configuration
test --- Run scipy unittests
"""
import importlib as _importlib
from numpy import __version__ as __numpy_version__
try:
from scipy.__config__ import show as show_config
except ImportError as e:
msg = """Error importing SciPy: you cannot import SciPy while
being in scipy source directory; please exit the SciPy source
tree first and relaunch your Python interpreter."""
raise ImportError(msg) from e
from scipy.version import version as __version__
# Allow distributors to run custom init code
from . import _distributor_init
del _distributor_init
from scipy._lib import _pep440
# In maintenance branch, change to np_maxversion N+3 if numpy is at N
np_minversion = '1.25.2'
np_maxversion = '2.6.0'
if (_pep440.parse(__numpy_version__) < _pep440.Version(np_minversion) or
_pep440.parse(__numpy_version__) >= _pep440.Version(np_maxversion)):
import warnings
warnings.warn(f"A NumPy version >={np_minversion} and <{np_maxversion}"
f" is required for this version of SciPy (detected "
f"version {__numpy_version__})",
UserWarning, stacklevel=2)
del _pep440
# This is the first import of an extension module within SciPy. If there's
# a general issue with the install, such that extension modules are missing
# or cannot be imported, this is where we'll get a failure - so give an
# informative error message.
try:
from scipy._lib._ccallback import LowLevelCallable
except ImportError as e:
msg = "The `scipy` install you are using seems to be broken, " + \
"(extension modules cannot be imported), " + \
"please try reinstalling."
raise ImportError(msg) from e
from scipy._lib._testutils import PytestTester
test = PytestTester(__name__)
del PytestTester
submodules = [
'cluster',
'constants',
'datasets',
'differentiate',
'fft',
'fftpack',
'integrate',
'interpolate',
'io',
'linalg',
'ndimage',
'odr',
'optimize',
'signal',
'sparse',
'spatial',
'special',
'stats'
]
__all__ = submodules + [
'LowLevelCallable',
'test',
'show_config',
'__version__',
]
def __dir__():
return __all__
def __getattr__(name):
if name in submodules:
return _importlib.import_module(f'scipy.{name}')
else:
try:
return globals()[name]
except KeyError:
raise AttributeError(
f"Module 'scipy' has no attribute '{name}'"
)
@@ -0,0 +1,18 @@
""" Distributor init file
Distributors: you can replace the contents of this file with your own custom
code to support particular distributions of SciPy.
For example, this is a good place to put any checks for hardware requirements
or BLAS/LAPACK library initialization.
The SciPy standard source distribution will not put code in this file beyond
the try-except import of `_distributor_init_local` (which is not part of a
standard source distribution), so you can safely replace this file with your
own version.
"""
try:
from . import _distributor_init_local # noqa: F401
except ImportError:
pass
@@ -0,0 +1,14 @@
"""
Module containing private utility functions
===========================================
The ``scipy._lib`` namespace is empty (for now). Tests for all
utilities in submodules of ``_lib`` can be run with::
from scipy import _lib
_lib.test()
"""
from scipy._lib._testutils import PytestTester
test = PytestTester(__name__)
del PytestTester
@@ -0,0 +1,931 @@
"""Utility functions to use Python Array API compatible libraries.
For the context about the Array API see:
https://data-apis.org/array-api/latest/purpose_and_scope.html
The SciPy use case of the Array API is described on the following page:
https://data-apis.org/array-api/latest/use_cases.html#use-case-scipy
"""
import contextlib
import dataclasses
import functools
import os
import textwrap
from collections.abc import Generator, Iterable, Iterator
from contextlib import contextmanager
from contextvars import ContextVar
from types import ModuleType
from typing import Any, Literal, TypeAlias
import numpy as np
import numpy.typing as npt
from scipy._lib import array_api_compat
from scipy._lib.array_api_compat import (
is_array_api_obj,
is_lazy_array,
size as xp_size,
numpy as np_compat,
device as xp_device,
is_numpy_namespace as is_numpy,
is_cupy_namespace as is_cupy,
is_torch_namespace as is_torch,
is_jax_namespace as is_jax,
is_dask_namespace as is_dask,
is_array_api_strict_namespace as is_array_api_strict
)
from scipy._lib._sparse import issparse
from scipy._lib._docscrape import FunctionDoc
__all__ = [
'_asarray', 'array_namespace', 'assert_almost_equal', 'assert_array_almost_equal',
'default_xp', 'eager_warns', 'is_lazy_array', 'is_marray',
'is_array_api_strict', 'is_complex', 'is_cupy', 'is_jax', 'is_numpy', 'is_torch',
'SCIPY_ARRAY_API', 'SCIPY_DEVICE', 'scipy_namespace_for',
'xp_assert_close', 'xp_assert_equal', 'xp_assert_less',
'xp_copy', 'xp_device', 'xp_ravel', 'xp_size',
'xp_unsupported_param_msg', 'xp_vector_norm', 'xp_capabilities',
'xp_result_type', 'xp_promote'
]
# To enable array API and strict array-like input validation
SCIPY_ARRAY_API: str | bool = os.environ.get("SCIPY_ARRAY_API", False)
# To control the default device - for use in the test suite only
SCIPY_DEVICE = os.environ.get("SCIPY_DEVICE", "cpu")
_GLOBAL_CONFIG = {
"SCIPY_ARRAY_API": SCIPY_ARRAY_API,
"SCIPY_DEVICE": SCIPY_DEVICE,
}
Array: TypeAlias = Any # To be changed to a Protocol later (see array-api#589)
ArrayLike: TypeAlias = Array | npt.ArrayLike
def _compliance_scipy(arrays: Iterable[ArrayLike]) -> Iterator[Array]:
"""Raise exceptions on known-bad subclasses. Discard 0-dimensional ArrayLikes
and convert 1+-dimensional ArrayLikes to numpy.
The following subclasses are not supported and raise and error:
- `numpy.ma.MaskedArray`
- `numpy.matrix`
- NumPy arrays which do not have a boolean or numerical dtype
- Any array-like which is neither array API compatible nor coercible by NumPy
- Any array-like which is coerced by NumPy to an unsupported dtype
"""
for array in arrays:
if array is None:
continue
# this comes from `_util._asarray_validated`
if issparse(array):
msg = ('Sparse arrays/matrices are not supported by this function. '
'Perhaps one of the `scipy.sparse.linalg` functions '
'would work instead.')
raise ValueError(msg)
if isinstance(array, np.ma.MaskedArray):
raise TypeError("Inputs of type `numpy.ma.MaskedArray` are not supported.")
if isinstance(array, np.matrix):
raise TypeError("Inputs of type `numpy.matrix` are not supported.")
if isinstance(array, np.ndarray | np.generic):
dtype = array.dtype
if not (np.issubdtype(dtype, np.number) or np.issubdtype(dtype, np.bool_)):
raise TypeError(f"An argument has dtype `{dtype!r}`; "
f"only boolean and numerical dtypes are supported.")
if is_array_api_obj(array):
yield array
else:
try:
array = np.asanyarray(array)
except TypeError:
raise TypeError("An argument is neither array API compatible nor "
"coercible by NumPy.")
dtype = array.dtype
if not (np.issubdtype(dtype, np.number) or np.issubdtype(dtype, np.bool_)):
message = (
f"An argument was coerced to an unsupported dtype `{dtype!r}`; "
f"only boolean and numerical dtypes are supported."
)
raise TypeError(message)
# Ignore 0-dimensional arrays, coherently with array-api-compat.
# Raise if there are 1+-dimensional array-likes mixed with non-numpy
# Array API objects.
if array.ndim:
yield array
def _check_finite(array: Array, xp: ModuleType) -> None:
"""Check for NaNs or Infs."""
if not xp.all(xp.isfinite(array)):
msg = "array must not contain infs or NaNs"
raise ValueError(msg)
def array_namespace(*arrays: Array) -> ModuleType:
"""Get the array API compatible namespace for the arrays xs.
Parameters
----------
*arrays : sequence of array_like
Arrays used to infer the common namespace.
Returns
-------
namespace : module
Common namespace.
Notes
-----
Thin wrapper around `array_api_compat.array_namespace`.
1. Check for the global switch: SCIPY_ARRAY_API. This can also be accessed
dynamically through ``_GLOBAL_CONFIG['SCIPY_ARRAY_API']``.
2. `_compliance_scipy` raise exceptions on known-bad subclasses. See
its definition for more details.
When the global switch is False, it defaults to the `numpy` namespace.
In that case, there is no compliance check. This is a convenience to
ease the adoption. Otherwise, arrays must comply with the new rules.
"""
if not _GLOBAL_CONFIG["SCIPY_ARRAY_API"]:
# here we could wrap the namespace if needed
return np_compat
api_arrays = list(_compliance_scipy(arrays))
# In case of a mix of array API compliant arrays and scalars, return
# the array API namespace. If there are only ArrayLikes (e.g. lists),
# return NumPy (wrapped by array-api-compat).
if api_arrays:
return array_api_compat.array_namespace(*api_arrays)
return np_compat
def _asarray(
array: ArrayLike,
dtype: Any = None,
order: Literal['K', 'A', 'C', 'F'] | None = None,
copy: bool | None = None,
*,
xp: ModuleType | None = None,
check_finite: bool = False,
subok: bool = False,
) -> Array:
"""SciPy-specific replacement for `np.asarray` with `order`, `check_finite`, and
`subok`.
Memory layout parameter `order` is not exposed in the Array API standard.
`order` is only enforced if the input array implementation
is NumPy based, otherwise `order` is just silently ignored.
`check_finite` is also not a keyword in the array API standard; included
here for convenience rather than that having to be a separate function
call inside SciPy functions.
`subok` is included to allow this function to preserve the behaviour of
`np.asanyarray` for NumPy based inputs.
"""
if xp is None:
xp = array_namespace(array)
if is_numpy(xp):
# Use NumPy API to support order
if copy is True:
array = np.array(array, order=order, dtype=dtype, subok=subok)
elif subok:
array = np.asanyarray(array, order=order, dtype=dtype)
else:
array = np.asarray(array, order=order, dtype=dtype)
else:
try:
array = xp.asarray(array, dtype=dtype, copy=copy)
except TypeError:
coerced_xp = array_namespace(xp.asarray(3))
array = coerced_xp.asarray(array, dtype=dtype, copy=copy)
if check_finite:
_check_finite(array, xp)
return array
def xp_copy(x: Array, *, xp: ModuleType | None = None) -> Array:
"""
Copies an array.
Parameters
----------
x : array
xp : array_namespace
Returns
-------
copy : array
Copied array
Notes
-----
This copy function does not offer all the semantics of `np.copy`, i.e. the
`subok` and `order` keywords are not used.
"""
# Note: for older NumPy versions, `np.asarray` did not support the `copy` kwarg,
# so this uses our other helper `_asarray`.
if xp is None:
xp = array_namespace(x)
return _asarray(x, copy=True, xp=xp)
_default_xp_ctxvar: ContextVar[ModuleType] = ContextVar("_default_xp")
@contextmanager
def default_xp(xp: ModuleType) -> Generator[None, None, None]:
"""In all ``xp_assert_*`` and ``assert_*`` function calls executed within this
context manager, test by default that the array namespace is
the provided across all arrays, unless one explicitly passes the ``xp=``
parameter or ``check_namespace=False``.
Without this context manager, the default value for `xp` is the namespace
for the desired array (the second parameter of the tests).
"""
token = _default_xp_ctxvar.set(xp)
try:
yield
finally:
_default_xp_ctxvar.reset(token)
def eager_warns(x, warning_type, match=None):
"""pytest.warns context manager, but only if x is not a lazy array."""
import pytest
# This attribute is interpreted by pytest-run-parallel, ensuring that tests that use
# `eager_warns` aren't run in parallel (since pytest.warns isn't thread-safe).
__thread_safe__ = False # noqa: F841
if is_lazy_array(x):
return contextlib.nullcontext()
return pytest.warns(warning_type, match=match)
def _strict_check(actual, desired, xp, *,
check_namespace=True, check_dtype=True, check_shape=True,
check_0d=True):
__tracebackhide__ = True # Hide traceback for py.test
if xp is None:
try:
xp = _default_xp_ctxvar.get()
except LookupError:
xp = array_namespace(desired)
if check_namespace:
_assert_matching_namespace(actual, desired, xp)
# only NumPy distinguishes between scalars and arrays; we do if check_0d=True.
# do this first so we can then cast to array (and thus use the array API) below.
if is_numpy(xp) and check_0d:
_msg = ("Array-ness does not match:\n Actual: "
f"{type(actual)}\n Desired: {type(desired)}")
assert ((xp.isscalar(actual) and xp.isscalar(desired))
or (not xp.isscalar(actual) and not xp.isscalar(desired))), _msg
actual = xp.asarray(actual)
desired = xp.asarray(desired)
if check_dtype:
_msg = f"dtypes do not match.\nActual: {actual.dtype}\nDesired: {desired.dtype}"
assert actual.dtype == desired.dtype, _msg
if check_shape:
if is_dask(xp):
actual.compute_chunk_sizes()
desired.compute_chunk_sizes()
_msg = f"Shapes do not match.\nActual: {actual.shape}\nDesired: {desired.shape}"
assert actual.shape == desired.shape, _msg
desired = xp.broadcast_to(desired, actual.shape)
return actual, desired, xp
def _assert_matching_namespace(actual, desired, xp):
__tracebackhide__ = True # Hide traceback for py.test
desired_arr_space = array_namespace(desired)
_msg = ("Namespace of desired array does not match expectations "
"set by the `default_xp` context manager or by the `xp`"
"pytest fixture.\n"
f"Desired array's space: {desired_arr_space.__name__}\n"
f"Expected namespace: {xp.__name__}")
assert desired_arr_space == xp, _msg
actual_arr_space = array_namespace(actual)
_msg = ("Namespace of actual and desired arrays do not match.\n"
f"Actual: {actual_arr_space.__name__}\n"
f"Desired: {xp.__name__}")
assert actual_arr_space == xp, _msg
def xp_assert_equal(actual, desired, *, check_namespace=True, check_dtype=True,
check_shape=True, check_0d=True, err_msg='', xp=None):
__tracebackhide__ = True # Hide traceback for py.test
actual, desired, xp = _strict_check(
actual, desired, xp, check_namespace=check_namespace,
check_dtype=check_dtype, check_shape=check_shape,
check_0d=check_0d
)
if is_cupy(xp):
return xp.testing.assert_array_equal(actual, desired, err_msg=err_msg)
elif is_torch(xp):
# PyTorch recommends using `rtol=0, atol=0` like this
# to test for exact equality
err_msg = None if err_msg == '' else err_msg
return xp.testing.assert_close(actual, desired, rtol=0, atol=0, equal_nan=True,
check_dtype=False, msg=err_msg)
# JAX uses `np.testing`
return np.testing.assert_array_equal(actual, desired, err_msg=err_msg)
def xp_assert_close(actual, desired, *, rtol=None, atol=0, check_namespace=True,
check_dtype=True, check_shape=True, check_0d=True,
err_msg='', xp=None):
__tracebackhide__ = True # Hide traceback for py.test
actual, desired, xp = _strict_check(
actual, desired, xp,
check_namespace=check_namespace, check_dtype=check_dtype,
check_shape=check_shape, check_0d=check_0d
)
floating = xp.isdtype(actual.dtype, ('real floating', 'complex floating'))
if rtol is None and floating:
# multiplier of 4 is used as for `np.float64` this puts the default `rtol`
# roughly half way between sqrt(eps) and the default for
# `numpy.testing.assert_allclose`, 1e-7
rtol = xp.finfo(actual.dtype).eps**0.5 * 4
elif rtol is None:
rtol = 1e-7
if is_cupy(xp):
return xp.testing.assert_allclose(actual, desired, rtol=rtol,
atol=atol, err_msg=err_msg)
elif is_torch(xp):
err_msg = None if err_msg == '' else err_msg
return xp.testing.assert_close(actual, desired, rtol=rtol, atol=atol,
equal_nan=True, check_dtype=False, msg=err_msg)
# JAX uses `np.testing`
return np.testing.assert_allclose(actual, desired, rtol=rtol,
atol=atol, err_msg=err_msg)
def xp_assert_less(actual, desired, *, check_namespace=True, check_dtype=True,
check_shape=True, check_0d=True, err_msg='', verbose=True, xp=None):
__tracebackhide__ = True # Hide traceback for py.test
actual, desired, xp = _strict_check(
actual, desired, xp, check_namespace=check_namespace,
check_dtype=check_dtype, check_shape=check_shape,
check_0d=check_0d
)
if is_cupy(xp):
return xp.testing.assert_array_less(actual, desired,
err_msg=err_msg, verbose=verbose)
elif is_torch(xp):
if actual.device.type != 'cpu':
actual = actual.cpu()
if desired.device.type != 'cpu':
desired = desired.cpu()
# JAX uses `np.testing`
return np.testing.assert_array_less(actual, desired,
err_msg=err_msg, verbose=verbose)
def assert_array_almost_equal(actual, desired, decimal=6, *args, **kwds):
"""Backwards compatible replacement. In new code, use xp_assert_close instead.
"""
rtol, atol = 0, 1.5*10**(-decimal)
return xp_assert_close(actual, desired,
atol=atol, rtol=rtol, check_dtype=False, check_shape=False,
*args, **kwds)
def assert_almost_equal(actual, desired, decimal=7, *args, **kwds):
"""Backwards compatible replacement. In new code, use xp_assert_close instead.
"""
rtol, atol = 0, 1.5*10**(-decimal)
return xp_assert_close(actual, desired,
atol=atol, rtol=rtol, check_dtype=False, check_shape=False,
*args, **kwds)
def xp_unsupported_param_msg(param: Any) -> str:
return f'Providing {param!r} is only supported for numpy arrays.'
def is_complex(x: Array, xp: ModuleType) -> bool:
return xp.isdtype(x.dtype, 'complex floating')
def scipy_namespace_for(xp: ModuleType) -> ModuleType | None:
"""Return the `scipy`-like namespace of a non-NumPy backend
That is, return the namespace corresponding with backend `xp` that contains
`scipy` sub-namespaces like `linalg` and `special`. If no such namespace
exists, return ``None``. Useful for dispatching.
"""
if is_cupy(xp):
import cupyx # type: ignore[import-not-found,import-untyped]
return cupyx.scipy
if is_jax(xp):
import jax # type: ignore[import-not-found]
return jax.scipy
if is_torch(xp):
return xp
return None
# maybe use `scipy.linalg` if/when array API support is added
def xp_vector_norm(x: Array, /, *,
axis: int | tuple[int] | None = None,
keepdims: bool = False,
ord: int | float = 2,
xp: ModuleType | None = None) -> Array:
xp = array_namespace(x) if xp is None else xp
if SCIPY_ARRAY_API:
# check for optional `linalg` extension
if hasattr(xp, 'linalg'):
return xp.linalg.vector_norm(x, axis=axis, keepdims=keepdims, ord=ord)
else:
if ord != 2:
raise ValueError(
"only the Euclidean norm (`ord=2`) is currently supported in "
"`xp_vector_norm` for backends not implementing the `linalg` "
"extension."
)
# return (x @ x)**0.5
# or to get the right behavior with nd, complex arrays
return xp.sum(xp.conj(x) * x, axis=axis, keepdims=keepdims)**0.5
else:
# to maintain backwards compatibility
return np.linalg.norm(x, ord=ord, axis=axis, keepdims=keepdims)
def xp_ravel(x: Array, /, *, xp: ModuleType | None = None) -> Array:
# Equivalent of np.ravel written in terms of array API
# Even though it's one line, it comes up so often that it's worth having
# this function for readability
xp = array_namespace(x) if xp is None else xp
return xp.reshape(x, (-1,))
def xp_swapaxes(a, axis1, axis2, xp=None):
# Equivalent of np.swapaxes written in terms of array API
xp = array_namespace(a) if xp is None else xp
axes = list(range(a.ndim))
axes[axis1], axes[axis2] = axes[axis2], axes[axis1]
a = xp.permute_dims(a, axes)
return a
# utility to find common dtype with option to force floating
def xp_result_type(*args, force_floating=False, xp):
"""
Returns the dtype that results from applying type promotion rules
(see Array API Standard Type Promotion Rules) to the arguments. Augments
standard `result_type` in a few ways:
- There is a `force_floating` argument that ensures that the result type
is floating point, even when all args are integer.
- When a TypeError is raised (e.g. due to an unsupported promotion)
and `force_floating=True`, we define a custom rule: use the result type
of the default float and any other floats passed. See
https://github.com/scipy/scipy/pull/22695/files#r1997905891
for rationale.
- This function accepts array-like iterables, which are immediately converted
to the namespace's arrays before result type calculation. Consequently, the
result dtype may be different when an argument is `1.` vs `[1.]`.
Typically, this function will be called shortly after `array_namespace`
on a subset of the arguments passed to `array_namespace`.
"""
args = [(_asarray(arg, subok=True, xp=xp) if np.iterable(arg) else arg)
for arg in args]
args_not_none = [arg for arg in args if arg is not None]
if force_floating:
args_not_none.append(1.0)
if is_numpy(xp) and xp.__version__ < '2.0':
# Follow NEP 50 promotion rules anyway
args_not_none = [arg.dtype if getattr(arg, 'size', 0) == 1 else arg
for arg in args_not_none]
return xp.result_type(*args_not_none)
try: # follow library's preferred promotion rules
return xp.result_type(*args_not_none)
except TypeError: # mixed type promotion isn't defined
if not force_floating:
raise
# use `result_type` of default floating point type and any floats present
# This can be revisited, but right now, the only backends that get here
# are array-api-strict (which is not for production use) and PyTorch
# (due to data-apis/array-api-compat#279).
float_args = []
for arg in args_not_none:
arg_array = xp.asarray(arg) if np.isscalar(arg) else arg
dtype = getattr(arg_array, 'dtype', arg)
if xp.isdtype(dtype, ('real floating', 'complex floating')):
float_args.append(arg)
return xp.result_type(*float_args, xp_default_dtype(xp))
def xp_promote(*args, broadcast=False, force_floating=False, xp):
"""
Promotes elements of *args to result dtype, ignoring `None`s.
Includes options for forcing promotion to floating point and
broadcasting the arrays, again ignoring `None`s.
Type promotion rules follow `xp_result_type` instead of `xp.result_type`.
Typically, this function will be called shortly after `array_namespace`
on a subset of the arguments passed to `array_namespace`.
This function accepts array-like iterables, which are immediately converted
to the namespace's arrays before result type calculation. Consequently, the
result dtype may be different when an argument is `1.` vs `[1.]`.
See Also
--------
xp_result_type
"""
args = [(_asarray(arg, subok=True, xp=xp) if np.iterable(arg) else arg)
for arg in args] # solely to prevent double conversion of iterable to array
dtype = xp_result_type(*args, force_floating=force_floating, xp=xp)
args = [(_asarray(arg, dtype=dtype, subok=True, xp=xp) if arg is not None else arg)
for arg in args]
if not broadcast:
return args[0] if len(args)==1 else tuple(args)
args_not_none = [arg for arg in args if arg is not None]
# determine result shape
shapes = {arg.shape for arg in args_not_none}
try:
shape = (np.broadcast_shapes(*shapes) if len(shapes) != 1
else args_not_none[0].shape)
except ValueError as e:
message = "Array shapes are incompatible for broadcasting."
raise ValueError(message) from e
out = []
for arg in args:
if arg is None:
out.append(arg)
continue
# broadcast only if needed
# Even if two arguments need broadcasting, this is faster than
# `broadcast_arrays`, especially since we've already determined `shape`
if arg.shape != shape:
kwargs = {'subok': True} if is_numpy(xp) else {}
arg = xp.broadcast_to(arg, shape, **kwargs)
# This is much faster than xp.astype(arg, dtype, copy=False)
if arg.dtype != dtype:
arg = xp.astype(arg, dtype)
out.append(arg)
return out[0] if len(out)==1 else tuple(out)
def xp_float_to_complex(arr: Array, xp: ModuleType | None = None) -> Array:
xp = array_namespace(arr) if xp is None else xp
arr_dtype = arr.dtype
# The standard float dtypes are float32 and float64.
# Convert float32 to complex64,
# and float64 (and non-standard real dtypes) to complex128
if xp.isdtype(arr_dtype, xp.float32):
arr = xp.astype(arr, xp.complex64)
elif xp.isdtype(arr_dtype, 'real floating'):
arr = xp.astype(arr, xp.complex128)
return arr
def xp_default_dtype(xp):
"""Query the namespace-dependent default floating-point dtype.
"""
if is_torch(xp):
# historically, we allow pytorch to keep its default of float32
return xp.get_default_dtype()
else:
# we default to float64
return xp.float64
def xp_result_device(*args):
"""Return the device of an array in `args`, for the purpose of
input-output device propagation.
If there are multiple devices, return an arbitrary one.
If there are no arrays, return None (this typically happens only on NumPy).
"""
for arg in args:
# Do not do a duck-type test for the .device attribute, as many backends today
# don't have it yet. See workarouunds in array_api_compat.device().
if is_array_api_obj(arg):
return xp_device(arg)
return None
def is_marray(xp):
"""Returns True if `xp` is an MArray namespace; False otherwise."""
return "marray" in xp.__name__
@dataclasses.dataclass(repr=False)
class _XPSphinxCapability:
cpu: bool | None # None if not applicable
gpu: bool | None
warnings: list[str] = dataclasses.field(default_factory=list)
def _render(self, value):
if value is None:
return "n/a"
if not value:
return "⛔"
if self.warnings:
res = "⚠️ " + '; '.join(self.warnings)
assert len(res) <= 20, "Warnings too long"
return res
return "✅"
def __str__(self):
cpu = self._render(self.cpu)
gpu = self._render(self.gpu)
return f"{cpu:20} {gpu:20}"
def _make_sphinx_capabilities(
# lists of tuples [(module name, reason), ...]
skip_backends=(), xfail_backends=(),
# @pytest.mark.skip/xfail_xp_backends kwargs
cpu_only=False, np_only=False, exceptions=(),
# xpx.lazy_xp_backends kwargs
allow_dask_compute=False, jax_jit=True,
# list of tuples [(module name, reason), ...]
warnings = (),
# unused in documentation
reason=None,
):
exceptions = set(exceptions)
# Default capabilities
capabilities = {
"numpy": _XPSphinxCapability(cpu=True, gpu=None),
"array_api_strict": _XPSphinxCapability(cpu=True, gpu=None),
"cupy": _XPSphinxCapability(cpu=None, gpu=True),
"torch": _XPSphinxCapability(cpu=True, gpu=True),
"jax.numpy": _XPSphinxCapability(cpu=True, gpu=True,
warnings=[] if jax_jit else ["no JIT"]),
# Note: Dask+CuPy is currently untested and unsupported
"dask.array": _XPSphinxCapability(cpu=True, gpu=None,
warnings=["computes graph"] if allow_dask_compute else []),
}
# documentation doesn't display the reason
for module, _ in list(skip_backends) + list(xfail_backends):
backend = capabilities[module]
if backend.cpu is not None:
backend.cpu = False
if backend.gpu is not None:
backend.gpu = False
for module, backend in capabilities.items():
if np_only and module not in exceptions | {"numpy"}:
if backend.cpu is not None:
backend.cpu = False
if backend.gpu is not None:
backend.gpu = False
elif cpu_only and module not in exceptions and backend.gpu is not None:
backend.gpu = False
for module, warning in warnings:
backend = capabilities[module]
backend.warnings.append(warning)
return capabilities
def _make_capabilities_note(fun_name, capabilities):
# Note: deliberately not documenting array-api-strict
note = f"""
`{fun_name}` has experimental support for Python Array API Standard compatible
backends in addition to NumPy. Please consider testing these features
by setting an environment variable ``SCIPY_ARRAY_API=1`` and providing
CuPy, PyTorch, JAX, or Dask arrays as array arguments. The following
combinations of backend and device (or other capability) are supported.
==================== ==================== ====================
Library CPU GPU
==================== ==================== ====================
NumPy {capabilities['numpy'] }
CuPy {capabilities['cupy'] }
PyTorch {capabilities['torch'] }
JAX {capabilities['jax.numpy'] }
Dask {capabilities['dask.array'] }
==================== ==================== ====================
See :ref:`dev-arrayapi` for more information.
"""
return textwrap.dedent(note)
def xp_capabilities(
*,
# Alternative capabilities table.
# Used only for testing this decorator.
capabilities_table=None,
# Generate pytest.mark.skip/xfail_xp_backends.
# See documentation in conftest.py.
# lists of tuples [(module name, reason), ...]
skip_backends=(), xfail_backends=(),
cpu_only=False, np_only=False, reason=None, exceptions=(),
# lists of tuples [(module name, reason), ...]
warnings=(),
# xpx.testing.lazy_xp_function kwargs.
# Refer to array-api-extra documentation.
allow_dask_compute=False, jax_jit=True,
):
"""Decorator for a function that states its support among various
Array API compatible backends.
This decorator has two effects:
1. It allows tagging tests with ``@make_xp_test_case`` or
``make_xp_pytest_param`` (see below) to automatically generate
SKIP/XFAIL markers and perform additional backend-specific
testing, such as extra validation for Dask and JAX;
2. It automatically adds a note to the function's docstring, containing
a table matching what has been tested.
See Also
--------
make_xp_test_case
make_xp_pytest_param
array_api_extra.testing.lazy_xp_function
"""
capabilities_table = (xp_capabilities_table if capabilities_table is None
else capabilities_table)
capabilities = dict(
skip_backends=skip_backends,
xfail_backends=xfail_backends,
cpu_only=cpu_only,
np_only=np_only,
reason=reason,
exceptions=exceptions,
allow_dask_compute=allow_dask_compute,
jax_jit=jax_jit,
warnings=warnings,
)
sphinx_capabilities = _make_sphinx_capabilities(**capabilities)
def decorator(f):
# Don't use a wrapper, as in some cases @xp_capabilities is
# applied to a ufunc
capabilities_table[f] = capabilities
note = _make_capabilities_note(f.__name__, sphinx_capabilities)
doc = FunctionDoc(f)
doc['Notes'].append(note)
doc = str(doc).split("\n", 1)[1] # remove signature
try:
f.__doc__ = doc
except AttributeError:
# Can't update __doc__ on ufuncs if SciPy
# was compiled against NumPy < 2.2.
pass
return f
return decorator
def _make_xp_pytest_marks(*funcs, capabilities_table=None):
capabilities_table = (xp_capabilities_table if capabilities_table is None
else capabilities_table)
import pytest
from scipy._lib.array_api_extra.testing import lazy_xp_function
marks = []
for func in funcs:
capabilities = capabilities_table[func]
exceptions = capabilities['exceptions']
reason = capabilities['reason']
if capabilities['cpu_only']:
marks.append(pytest.mark.skip_xp_backends(
cpu_only=True, exceptions=exceptions, reason=reason))
if capabilities['np_only']:
marks.append(pytest.mark.skip_xp_backends(
np_only=True, exceptions=exceptions, reason=reason))
for mod_name, reason in capabilities['skip_backends']:
marks.append(pytest.mark.skip_xp_backends(mod_name, reason=reason))
for mod_name, reason in capabilities['xfail_backends']:
marks.append(pytest.mark.xfail_xp_backends(mod_name, reason=reason))
lazy_kwargs = {k: capabilities[k]
for k in ('allow_dask_compute', 'jax_jit')}
lazy_xp_function(func, **lazy_kwargs)
return marks
def make_xp_test_case(*funcs, capabilities_table=None):
capabilities_table = (xp_capabilities_table if capabilities_table is None
else capabilities_table)
"""Generate pytest decorator for a test function that tests functionality
of one or more Array API compatible functions.
Read the parameters of the ``@xp_capabilities`` decorator applied to the
listed functions and:
- Generate the ``@pytest.mark.skip_xp_backends`` and
``@pytest.mark.xfail_xp_backends`` decorators
for the decorated test function
- Tag the function with `xpx.testing.lazy_xp_function`
See Also
--------
xp_capabilities
make_xp_pytest_param
array_api_extra.testing.lazy_xp_function
"""
marks = _make_xp_pytest_marks(*funcs, capabilities_table=capabilities_table)
return lambda func: functools.reduce(lambda f, g: g(f), marks, func)
def make_xp_pytest_param(func, *args, capabilities_table=None):
"""Variant of ``make_xp_test_case`` that returns a pytest.param for a function,
with all necessary skip_xp_backends and xfail_xp_backends marks applied::
@pytest.mark.parametrize(
"func", [make_xp_pytest_param(f1), make_xp_pytest_param(f2)]
)
def test(func, xp):
...
The above is equivalent to::
@pytest.mark.parametrize(
"func", [
pytest.param(f1, marks=[
pytest.mark.skip_xp_backends(...),
pytest.mark.xfail_xp_backends(...), ...]),
pytest.param(f2, marks=[
pytest.mark.skip_xp_backends(...),
pytest.mark.xfail_xp_backends(...), ...]),
)
def test(func, xp):
...
Parameters
----------
func : Callable
Function to be tested. It must be decorated with ``@xp_capabilities``.
*args : Any, optional
Extra pytest parameters for the use case, e.g.::
@pytest.mark.parametrize("func,verb", [
make_xp_pytest_param(f1, "hello"),
make_xp_pytest_param(f2, "world")])
def test(func, verb, xp):
# iterates on (func=f1, verb="hello")
# and (func=f2, verb="world")
See Also
--------
xp_capabilities
make_xp_test_case
array_api_extra.testing.lazy_xp_function
"""
import pytest
marks = _make_xp_pytest_marks(func, capabilities_table=capabilities_table)
return pytest.param(func, *args, marks=marks, id=func.__name__)
# Is it OK to have a dictionary that is mutated (once upon import) in many places?
xp_capabilities_table = {} # type: ignore[var-annotated]
@@ -0,0 +1,9 @@
# DO NOT RENAME THIS FILE
# This is a hook for array_api_extra/src/array_api_extra/_lib/_compat.py
# to override functions of array_api_compat.
from .array_api_compat import * # noqa: F403
from ._array_api import array_namespace as scipy_array_namespace
# overrides array_api_compat.array_namespace inside array-api-extra
array_namespace = scipy_array_namespace # type: ignore[assignment]
@@ -0,0 +1,103 @@
"""
Extra testing functions that forbid 0d-input, see #21044
While the xp_assert_* functions generally aim to follow the conventions of the
underlying `xp` library, NumPy in particular is inconsistent in its handling
of scalars vs. 0d-arrays, see https://github.com/numpy/numpy/issues/24897.
For example, this means that the following operations (as of v2.0.1) currently
return scalars, even though a 0d-array would often be more appropriate:
import numpy as np
np.array(0) * 2 # scalar, not 0d array
- np.array(0) # scalar, not 0d-array
np.sin(np.array(0)) # scalar, not 0d array
np.mean([1, 2, 3]) # scalar, not 0d array
Libraries like CuPy tend to return a 0d-array in scenarios like those above,
and even `xp.asarray(0)[()]` remains a 0d-array there. To deal with the reality
of the inconsistencies present in NumPy, as well as 20+ years of code on top,
the `xp_assert_*` functions here enforce consistency in the only way that
doesn't go against the tide, i.e. by forbidding 0d-arrays as the return type.
However, when scalars are not generally the expected NumPy return type,
it remains preferable to use the assert functions from
the `scipy._lib._array_api` module, which have less surprising behaviour.
"""
from scipy._lib._array_api import array_namespace, is_numpy
from scipy._lib._array_api import (xp_assert_close as xp_assert_close_base,
xp_assert_equal as xp_assert_equal_base,
xp_assert_less as xp_assert_less_base)
__all__: list[str] = []
def _check_scalar(actual, desired, *, xp=None, **kwargs):
__tracebackhide__ = True # Hide traceback for py.test
if xp is None:
xp = array_namespace(actual)
# necessary to handle non-numpy scalars, e.g. bare `0.0` has no shape
desired = xp.asarray(desired)
# Only NumPy distinguishes between scalars and arrays;
# shape check in xp_assert_* is sufficient except for shape == ()
if not (is_numpy(xp) and desired.shape == ()):
return
_msg = ("Result is a NumPy 0d-array. Many SciPy functions intend to follow "
"the convention of many NumPy functions, returning a scalar when a "
"0d-array would be correct. The specialized `xp_assert_*` functions "
"in the `scipy._lib._array_api_no_0d` module err on the side of "
"caution and do not accept 0d-arrays by default. If the correct "
"result may legitimately be a 0d-array, pass `check_0d=True`, "
"or use the `xp_assert_*` functions from `scipy._lib._array_api`.")
assert xp.isscalar(actual), _msg
def xp_assert_equal(actual, desired, *, check_0d=False, **kwargs):
# in contrast to xp_assert_equal_base, this defaults to check_0d=False,
# but will do an extra check in that case, which forbids 0d-arrays for `actual`
__tracebackhide__ = True # Hide traceback for py.test
# array-ness (check_0d == True) is taken care of by the *_base functions
if not check_0d:
_check_scalar(actual, desired, **kwargs)
return xp_assert_equal_base(actual, desired, check_0d=check_0d, **kwargs)
def xp_assert_close(actual, desired, *, check_0d=False, **kwargs):
# as for xp_assert_equal
__tracebackhide__ = True
if not check_0d:
_check_scalar(actual, desired, **kwargs)
return xp_assert_close_base(actual, desired, check_0d=check_0d, **kwargs)
def xp_assert_less(actual, desired, *, check_0d=False, **kwargs):
# as for xp_assert_equal
__tracebackhide__ = True
if not check_0d:
_check_scalar(actual, desired, **kwargs)
return xp_assert_less_base(actual, desired, check_0d=check_0d, **kwargs)
def assert_array_almost_equal(actual, desired, decimal=6, *args, **kwds):
"""Backwards compatible replacement. In new code, use xp_assert_close instead.
"""
rtol, atol = 0, 1.5*10**(-decimal)
return xp_assert_close(actual, desired,
atol=atol, rtol=rtol, check_dtype=False, check_shape=False,
*args, **kwds)
def assert_almost_equal(actual, desired, decimal=7, *args, **kwds):
"""Backwards compatible replacement. In new code, use xp_assert_close instead.
"""
rtol, atol = 0, 1.5*10**(-decimal)
return xp_assert_close(actual, desired,
atol=atol, rtol=rtol, check_dtype=False, check_shape=False,
*args, **kwds)
@@ -0,0 +1,229 @@
import sys as _sys
from keyword import iskeyword as _iskeyword
def _validate_names(typename, field_names, extra_field_names):
"""
Ensure that all the given names are valid Python identifiers that
do not start with '_'. Also check that there are no duplicates
among field_names + extra_field_names.
"""
for name in [typename] + field_names + extra_field_names:
if not isinstance(name, str):
raise TypeError('typename and all field names must be strings')
if not name.isidentifier():
raise ValueError('typename and all field names must be valid '
f'identifiers: {name!r}')
if _iskeyword(name):
raise ValueError('typename and all field names cannot be a '
f'keyword: {name!r}')
seen = set()
for name in field_names + extra_field_names:
if name.startswith('_'):
raise ValueError('Field names cannot start with an underscore: '
f'{name!r}')
if name in seen:
raise ValueError(f'Duplicate field name: {name!r}')
seen.add(name)
# Note: This code is adapted from CPython:Lib/collections/__init__.py
def _make_tuple_bunch(typename, field_names, extra_field_names=None,
module=None):
"""
Create a namedtuple-like class with additional attributes.
This function creates a subclass of tuple that acts like a namedtuple
and that has additional attributes.
The additional attributes are listed in `extra_field_names`. The
values assigned to these attributes are not part of the tuple.
The reason this function exists is to allow functions in SciPy
that currently return a tuple or a namedtuple to returned objects
that have additional attributes, while maintaining backwards
compatibility.
This should only be used to enhance *existing* functions in SciPy.
New functions are free to create objects as return values without
having to maintain backwards compatibility with an old tuple or
namedtuple return value.
Parameters
----------
typename : str
The name of the type.
field_names : list of str
List of names of the values to be stored in the tuple. These names
will also be attributes of instances, so the values in the tuple
can be accessed by indexing or as attributes. At least one name
is required. See the Notes for additional restrictions.
extra_field_names : list of str, optional
List of names of values that will be stored as attributes of the
object. See the notes for additional restrictions.
Returns
-------
cls : type
The new class.
Notes
-----
There are restrictions on the names that may be used in `field_names`
and `extra_field_names`:
* The names must be unique--no duplicates allowed.
* The names must be valid Python identifiers, and must not begin with
an underscore.
* The names must not be Python keywords (e.g. 'def', 'and', etc., are
not allowed).
Examples
--------
>>> from scipy._lib._bunch import _make_tuple_bunch
Create a class that acts like a namedtuple with length 2 (with field
names `x` and `y`) that will also have the attributes `w` and `beta`:
>>> Result = _make_tuple_bunch('Result', ['x', 'y'], ['w', 'beta'])
`Result` is the new class. We call it with keyword arguments to create
a new instance with given values.
>>> result1 = Result(x=1, y=2, w=99, beta=0.5)
>>> result1
Result(x=1, y=2, w=99, beta=0.5)
`result1` acts like a tuple of length 2:
>>> len(result1)
2
>>> result1[:]
(1, 2)
The values assigned when the instance was created are available as
attributes:
>>> result1.y
2
>>> result1.beta
0.5
"""
if len(field_names) == 0:
raise ValueError('field_names must contain at least one name')
if extra_field_names is None:
extra_field_names = []
_validate_names(typename, field_names, extra_field_names)
typename = _sys.intern(str(typename))
field_names = tuple(map(_sys.intern, field_names))
extra_field_names = tuple(map(_sys.intern, extra_field_names))
all_names = field_names + extra_field_names
arg_list = ', '.join(field_names)
full_list = ', '.join(all_names)
repr_fmt = ''.join(('(',
', '.join(f'{name}=%({name})r' for name in all_names),
')'))
tuple_new = tuple.__new__
_dict, _tuple, _zip = dict, tuple, zip
# Create all the named tuple methods to be added to the class namespace
s = f"""\
def __new__(_cls, {arg_list}, **extra_fields):
return _tuple_new(_cls, ({arg_list},))
def __init__(self, {arg_list}, **extra_fields):
for key in self._extra_fields:
if key not in extra_fields:
raise TypeError("missing keyword argument '%s'" % (key,))
for key, val in extra_fields.items():
if key not in self._extra_fields:
raise TypeError("unexpected keyword argument '%s'" % (key,))
self.__dict__[key] = val
def __setattr__(self, key, val):
if key in {repr(field_names)}:
raise AttributeError("can't set attribute %r of class %r"
% (key, self.__class__.__name__))
else:
self.__dict__[key] = val
"""
del arg_list
namespace = {'_tuple_new': tuple_new,
'__builtins__': dict(TypeError=TypeError,
AttributeError=AttributeError),
'__name__': f'namedtuple_{typename}'}
exec(s, namespace)
__new__ = namespace['__new__']
__new__.__doc__ = f'Create new instance of {typename}({full_list})'
__init__ = namespace['__init__']
__init__.__doc__ = f'Instantiate instance of {typename}({full_list})'
__setattr__ = namespace['__setattr__']
def __repr__(self):
'Return a nicely formatted representation string'
return self.__class__.__name__ + repr_fmt % self._asdict()
def _asdict(self):
'Return a new dict which maps field names to their values.'
out = _dict(_zip(self._fields, self))
out.update(self.__dict__)
return out
def __getnewargs_ex__(self):
'Return self as a plain tuple. Used by copy and pickle.'
return _tuple(self), self.__dict__
# Modify function metadata to help with introspection and debugging
for method in (__new__, __repr__, _asdict, __getnewargs_ex__):
method.__qualname__ = f'{typename}.{method.__name__}'
# Build-up the class namespace dictionary
# and use type() to build the result class
class_namespace = {
'__doc__': f'{typename}({full_list})',
'_fields': field_names,
'__new__': __new__,
'__init__': __init__,
'__repr__': __repr__,
'__setattr__': __setattr__,
'_asdict': _asdict,
'_extra_fields': extra_field_names,
'__getnewargs_ex__': __getnewargs_ex__,
# _field_defaults and _replace are added to get Polars to detect
# a bunch object as a namedtuple. See gh-22450
'_field_defaults': {},
'_replace': None,
}
for index, name in enumerate(field_names):
def _get(self, index=index):
return self[index]
class_namespace[name] = property(_get)
for name in extra_field_names:
def _get(self, name=name):
return self.__dict__[name]
class_namespace[name] = property(_get)
result = type(typename, (tuple,), class_namespace)
# For pickling to work, the __module__ variable needs to be set to the
# frame where the named tuple is created. Bypass this step in environments
# where sys._getframe is not defined (Jython for example) or sys._getframe
# is not defined for arguments greater than 0 (IronPython), or where the
# user has specified a particular module.
if module is None:
try:
module = _sys._getframe(1).f_globals.get('__name__', '__main__')
except (AttributeError, ValueError):
pass
if module is not None:
result.__module__ = module
__new__.__module__ = module
return result
@@ -0,0 +1,251 @@
from . import _ccallback_c
import ctypes
PyCFuncPtr = ctypes.CFUNCTYPE(ctypes.c_void_p).__bases__[0]
ffi = None
class CData:
pass
def _import_cffi():
global ffi, CData
if ffi is not None:
return
try:
import cffi
ffi = cffi.FFI()
CData = ffi.CData
except ImportError:
ffi = False
class LowLevelCallable(tuple):
"""
Low-level callback function.
Some functions in SciPy take as arguments callback functions, which
can either be python callables or low-level compiled functions. Using
compiled callback functions can improve performance somewhat by
avoiding wrapping data in Python objects.
Such low-level functions in SciPy are wrapped in `LowLevelCallable`
objects, which can be constructed from function pointers obtained from
ctypes, cffi, Cython, or contained in Python `PyCapsule` objects.
.. seealso::
Functions accepting low-level callables:
`scipy.integrate.quad`, `scipy.ndimage.generic_filter`,
`scipy.ndimage.generic_filter1d`, `scipy.ndimage.geometric_transform`
Usage examples:
:ref:`ndimage-ccallbacks`, :ref:`quad-callbacks`
Parameters
----------
function : {PyCapsule, ctypes function pointer, cffi function pointer}
Low-level callback function.
user_data : {PyCapsule, ctypes void pointer, cffi void pointer}
User data to pass on to the callback function.
signature : str, optional
Signature of the function. If omitted, determined from *function*,
if possible.
Attributes
----------
function
Callback function given.
user_data
User data given.
signature
Signature of the function.
Methods
-------
from_cython
Class method for constructing callables from Cython C-exported
functions.
Notes
-----
The argument ``function`` can be one of:
- PyCapsule, whose name contains the C function signature
- ctypes function pointer
- cffi function pointer
The signature of the low-level callback must match one of those expected
by the routine it is passed to.
If constructing low-level functions from a PyCapsule, the name of the
capsule must be the corresponding signature, in the format::
return_type (arg1_type, arg2_type, ...)
For example::
"void (double)"
"double (double, int *, void *)"
The context of a PyCapsule passed in as ``function`` is used as ``user_data``,
if an explicit value for ``user_data`` was not given.
"""
# Make the class immutable
__slots__ = ()
def __new__(cls, function, user_data=None, signature=None):
# We need to hold a reference to the function & user data,
# to prevent them going out of scope
item = cls._parse_callback(function, user_data, signature)
return tuple.__new__(cls, (item, function, user_data))
def __repr__(self):
return f"LowLevelCallable({self.function!r}, {self.user_data!r})"
@property
def function(self):
return tuple.__getitem__(self, 1)
@property
def user_data(self):
return tuple.__getitem__(self, 2)
@property
def signature(self):
return _ccallback_c.get_capsule_signature(tuple.__getitem__(self, 0))
def __getitem__(self, idx):
raise ValueError()
@classmethod
def from_cython(cls, module, name, user_data=None, signature=None):
"""
Create a low-level callback function from an exported Cython function.
Parameters
----------
module : module
Cython module where the exported function resides
name : str
Name of the exported function
user_data : {PyCapsule, ctypes void pointer, cffi void pointer}, optional
User data to pass on to the callback function.
signature : str, optional
Signature of the function. If omitted, determined from *function*.
"""
try:
function = module.__pyx_capi__[name]
except AttributeError as e:
message = "Given module is not a Cython module with __pyx_capi__ attribute"
raise ValueError(message) from e
except KeyError as e:
message = f"No function {name!r} found in __pyx_capi__ of the module"
raise ValueError(message) from e
return cls(function, user_data, signature)
@classmethod
def _parse_callback(cls, obj, user_data=None, signature=None):
_import_cffi()
if isinstance(obj, LowLevelCallable):
func = tuple.__getitem__(obj, 0)
elif isinstance(obj, PyCFuncPtr):
func, signature = _get_ctypes_func(obj, signature)
elif isinstance(obj, CData):
func, signature = _get_cffi_func(obj, signature)
elif _ccallback_c.check_capsule(obj):
func = obj
else:
raise ValueError("Given input is not a callable or a "
"low-level callable (pycapsule/ctypes/cffi)")
if isinstance(user_data, ctypes.c_void_p):
context = _get_ctypes_data(user_data)
elif isinstance(user_data, CData):
context = _get_cffi_data(user_data)
elif user_data is None:
context = 0
elif _ccallback_c.check_capsule(user_data):
context = user_data
else:
raise ValueError("Given user data is not a valid "
"low-level void* pointer (pycapsule/ctypes/cffi)")
return _ccallback_c.get_raw_capsule(func, signature, context)
#
# ctypes helpers
#
def _get_ctypes_func(func, signature=None):
# Get function pointer
func_ptr = ctypes.cast(func, ctypes.c_void_p).value
# Construct function signature
if signature is None:
signature = _typename_from_ctypes(func.restype) + " ("
for j, arg in enumerate(func.argtypes):
if j == 0:
signature += _typename_from_ctypes(arg)
else:
signature += ", " + _typename_from_ctypes(arg)
signature += ")"
return func_ptr, signature
def _typename_from_ctypes(item):
if item is None:
return "void"
elif item is ctypes.c_void_p:
return "void *"
name = item.__name__
pointer_level = 0
while name.startswith("LP_"):
pointer_level += 1
name = name[3:]
if name.startswith('c_'):
name = name[2:]
if pointer_level > 0:
name += " " + "*"*pointer_level
return name
def _get_ctypes_data(data):
# Get voidp pointer
return ctypes.cast(data, ctypes.c_void_p).value
#
# CFFI helpers
#
def _get_cffi_func(func, signature=None):
# Get function pointer
func_ptr = ffi.cast('uintptr_t', func)
# Get signature
if signature is None:
signature = ffi.getctype(ffi.typeof(func)).replace('(*)', ' ')
return func_ptr, signature
def _get_cffi_data(data):
# Get pointer
return ffi.cast('uintptr_t', data)
@@ -0,0 +1,254 @@
"""
Disjoint set data structure
"""
class DisjointSet:
""" Disjoint set data structure for incremental connectivity queries.
.. versionadded:: 1.6.0
Attributes
----------
n_subsets : int
The number of subsets.
Methods
-------
add
merge
connected
subset
subset_size
subsets
__getitem__
Notes
-----
This class implements the disjoint set [1]_, also known as the *union-find*
or *merge-find* data structure. The *find* operation (implemented in
`__getitem__`) implements the *path halving* variant. The *merge* method
implements the *merge by size* variant.
References
----------
.. [1] https://en.wikipedia.org/wiki/Disjoint-set_data_structure
Examples
--------
>>> from scipy.cluster.hierarchy import DisjointSet
Initialize a disjoint set:
>>> disjoint_set = DisjointSet([1, 2, 3, 'a', 'b'])
Merge some subsets:
>>> disjoint_set.merge(1, 2)
True
>>> disjoint_set.merge(3, 'a')
True
>>> disjoint_set.merge('a', 'b')
True
>>> disjoint_set.merge('b', 'b')
False
Find root elements:
>>> disjoint_set[2]
1
>>> disjoint_set['b']
3
Test connectivity:
>>> disjoint_set.connected(1, 2)
True
>>> disjoint_set.connected(1, 'b')
False
List elements in disjoint set:
>>> list(disjoint_set)
[1, 2, 3, 'a', 'b']
Get the subset containing 'a':
>>> disjoint_set.subset('a')
{'a', 3, 'b'}
Get the size of the subset containing 'a' (without actually instantiating
the subset):
>>> disjoint_set.subset_size('a')
3
Get all subsets in the disjoint set:
>>> disjoint_set.subsets()
[{1, 2}, {'a', 3, 'b'}]
"""
def __init__(self, elements=None):
self.n_subsets = 0
self._sizes = {}
self._parents = {}
# _nbrs is a circular linked list which links connected elements.
self._nbrs = {}
# _indices tracks the element insertion order in `__iter__`.
self._indices = {}
if elements is not None:
for x in elements:
self.add(x)
def __iter__(self):
"""Returns an iterator of the elements in the disjoint set.
Elements are ordered by insertion order.
"""
return iter(self._indices)
def __len__(self):
return len(self._indices)
def __contains__(self, x):
return x in self._indices
def __getitem__(self, x):
"""Find the root element of `x`.
Parameters
----------
x : hashable object
Input element.
Returns
-------
root : hashable object
Root element of `x`.
"""
if x not in self._indices:
raise KeyError(x)
# find by "path halving"
parents = self._parents
while self._indices[x] != self._indices[parents[x]]:
parents[x] = parents[parents[x]]
x = parents[x]
return x
def add(self, x):
"""Add element `x` to disjoint set
"""
if x in self._indices:
return
self._sizes[x] = 1
self._parents[x] = x
self._nbrs[x] = x
self._indices[x] = len(self._indices)
self.n_subsets += 1
def merge(self, x, y):
"""Merge the subsets of `x` and `y`.
The smaller subset (the child) is merged into the larger subset (the
parent). If the subsets are of equal size, the root element which was
first inserted into the disjoint set is selected as the parent.
Parameters
----------
x, y : hashable object
Elements to merge.
Returns
-------
merged : bool
True if `x` and `y` were in disjoint sets, False otherwise.
"""
xr = self[x]
yr = self[y]
if self._indices[xr] == self._indices[yr]:
return False
sizes = self._sizes
if (sizes[xr], self._indices[yr]) < (sizes[yr], self._indices[xr]):
xr, yr = yr, xr
self._parents[yr] = xr
self._sizes[xr] += self._sizes[yr]
self._nbrs[xr], self._nbrs[yr] = self._nbrs[yr], self._nbrs[xr]
self.n_subsets -= 1
return True
def connected(self, x, y):
"""Test whether `x` and `y` are in the same subset.
Parameters
----------
x, y : hashable object
Elements to test.
Returns
-------
result : bool
True if `x` and `y` are in the same set, False otherwise.
"""
return self._indices[self[x]] == self._indices[self[y]]
def subset(self, x):
"""Get the subset containing `x`.
Parameters
----------
x : hashable object
Input element.
Returns
-------
result : set
Subset containing `x`.
"""
if x not in self._indices:
raise KeyError(x)
result = [x]
nxt = self._nbrs[x]
while self._indices[nxt] != self._indices[x]:
result.append(nxt)
nxt = self._nbrs[nxt]
return set(result)
def subset_size(self, x):
"""Get the size of the subset containing `x`.
Note that this method is faster than ``len(self.subset(x))`` because
the size is directly read off an internal field, without the need to
instantiate the full subset.
Parameters
----------
x : hashable object
Input element.
Returns
-------
result : int
Size of the subset containing `x`.
"""
return self._sizes[self[x]]
def subsets(self):
"""Get all the subsets in the disjoint set.
Returns
-------
result : list
Subsets in the disjoint set.
"""
result = []
visited = set()
for x in self:
if x not in visited:
xset = self.subset(x)
visited.update(xset)
result.append(xset)
return result
@@ -0,0 +1,761 @@
# copied from numpydoc/docscrape.py, commit 97a6026508e0dd5382865672e9563a72cc113bd2
"""Extract reference documentation from the NumPy source tree."""
import copy
import inspect
import pydoc
import re
import sys
import textwrap
from collections import namedtuple
from collections.abc import Callable, Mapping
from functools import cached_property
from warnings import warn
def strip_blank_lines(l):
"Remove leading and trailing blank lines from a list of lines"
while l and not l[0].strip():
del l[0]
while l and not l[-1].strip():
del l[-1]
return l
class Reader:
"""A line-based string reader."""
def __init__(self, data):
"""
Parameters
----------
data : str
String with lines separated by '\\n'.
"""
if isinstance(data, list):
self._str = data
else:
self._str = data.split("\n") # store string as list of lines
self.reset()
def __getitem__(self, n):
return self._str[n]
def reset(self):
self._l = 0 # current line nr
def read(self):
if not self.eof():
out = self[self._l]
self._l += 1
return out
else:
return ""
def seek_next_non_empty_line(self):
for l in self[self._l :]:
if l.strip():
break
else:
self._l += 1
def eof(self):
return self._l >= len(self._str)
def read_to_condition(self, condition_func):
start = self._l
for line in self[start:]:
if condition_func(line):
return self[start : self._l]
self._l += 1
if self.eof():
return self[start : self._l + 1]
return []
def read_to_next_empty_line(self):
self.seek_next_non_empty_line()
def is_empty(line):
return not line.strip()
return self.read_to_condition(is_empty)
def read_to_next_unindented_line(self):
def is_unindented(line):
return line.strip() and (len(line.lstrip()) == len(line))
return self.read_to_condition(is_unindented)
def peek(self, n=0):
if self._l + n < len(self._str):
return self[self._l + n]
else:
return ""
def is_empty(self):
return not "".join(self._str).strip()
class ParseError(Exception):
def __str__(self):
message = self.args[0]
if hasattr(self, "docstring"):
message = f"{message} in {self.docstring!r}"
return message
Parameter = namedtuple("Parameter", ["name", "type", "desc"])
class NumpyDocString(Mapping):
"""Parses a numpydoc string to an abstract representation
Instances define a mapping from section title to structured data.
"""
sections = {
"Signature": "",
"Summary": [""],
"Extended Summary": [],
"Parameters": [],
"Attributes": [],
"Methods": [],
"Returns": [],
"Yields": [],
"Receives": [],
"Other Parameters": [],
"Raises": [],
"Warns": [],
"Warnings": [],
"See Also": [],
"Notes": [],
"References": "",
"Examples": "",
"index": {},
}
def __init__(self, docstring, config=None):
orig_docstring = docstring
docstring = textwrap.dedent(docstring).split("\n")
self._doc = Reader(docstring)
self._parsed_data = copy.deepcopy(self.sections)
try:
self._parse()
except ParseError as e:
e.docstring = orig_docstring
raise
def __getitem__(self, key):
return self._parsed_data[key]
def __setitem__(self, key, val):
if key not in self._parsed_data:
self._error_location(f"Unknown section {key}", error=False)
else:
self._parsed_data[key] = val
def __iter__(self):
return iter(self._parsed_data)
def __len__(self):
return len(self._parsed_data)
def _is_at_section(self):
self._doc.seek_next_non_empty_line()
if self._doc.eof():
return False
l1 = self._doc.peek().strip() # e.g. Parameters
if l1.startswith(".. index::"):
return True
l2 = self._doc.peek(1).strip() # ---------- or ==========
if len(l2) >= 3 and (set(l2) in ({"-"}, {"="})) and len(l2) != len(l1):
snip = "\n".join(self._doc._str[:2]) + "..."
self._error_location(
f"potentially wrong underline length... \n{l1} \n{l2} in \n{snip}",
error=False,
)
return l2.startswith("-" * len(l1)) or l2.startswith("=" * len(l1))
def _strip(self, doc):
i = 0
j = 0
for i, line in enumerate(doc):
if line.strip():
break
for j, line in enumerate(doc[::-1]):
if line.strip():
break
return doc[i : len(doc) - j]
def _read_to_next_section(self):
section = self._doc.read_to_next_empty_line()
while not self._is_at_section() and not self._doc.eof():
if not self._doc.peek(-1).strip(): # previous line was empty
section += [""]
section += self._doc.read_to_next_empty_line()
return section
def _read_sections(self):
while not self._doc.eof():
data = self._read_to_next_section()
name = data[0].strip()
if name.startswith(".."): # index section
yield name, data[1:]
elif len(data) < 2:
yield StopIteration
else:
yield name, self._strip(data[2:])
def _parse_param_list(self, content, single_element_is_type=False):
content = dedent_lines(content)
r = Reader(content)
params = []
while not r.eof():
header = r.read().strip()
if " : " in header:
arg_name, arg_type = header.split(" : ", maxsplit=1)
else:
# NOTE: param line with single element should never have a
# a " :" before the description line, so this should probably
# warn.
if header.endswith(" :"):
header = header[:-2]
if single_element_is_type:
arg_name, arg_type = "", header
else:
arg_name, arg_type = header, ""
desc = r.read_to_next_unindented_line()
desc = dedent_lines(desc)
desc = strip_blank_lines(desc)
params.append(Parameter(arg_name, arg_type, desc))
return params
# See also supports the following formats.
#
# <FUNCNAME>
# <FUNCNAME> SPACE* COLON SPACE+ <DESC> SPACE*
# <FUNCNAME> ( COMMA SPACE+ <FUNCNAME>)+ (COMMA | PERIOD)? SPACE*
# <FUNCNAME> ( COMMA SPACE+ <FUNCNAME>)* SPACE* COLON SPACE+ <DESC> SPACE*
# <FUNCNAME> is one of
# <PLAIN_FUNCNAME>
# COLON <ROLE> COLON BACKTICK <PLAIN_FUNCNAME> BACKTICK
# where
# <PLAIN_FUNCNAME> is a legal function name, and
# <ROLE> is any nonempty sequence of word characters.
# Examples: func_f1 :meth:`func_h1` :obj:`~baz.obj_r` :class:`class_j`
# <DESC> is a string describing the function.
_role = r":(?P<role>(py:)?\w+):"
_funcbacktick = r"`(?P<name>(?:~\w+\.)?[a-zA-Z0-9_\.-]+)`"
_funcplain = r"(?P<name2>[a-zA-Z0-9_\.-]+)"
_funcname = r"(" + _role + _funcbacktick + r"|" + _funcplain + r")"
_funcnamenext = _funcname.replace("role", "rolenext")
_funcnamenext = _funcnamenext.replace("name", "namenext")
_description = r"(?P<description>\s*:(\s+(?P<desc>\S+.*))?)?\s*$"
_func_rgx = re.compile(r"^\s*" + _funcname + r"\s*")
_line_rgx = re.compile(
r"^\s*"
+ r"(?P<allfuncs>"
+ _funcname # group for all function names
+ r"(?P<morefuncs>([,]\s+"
+ _funcnamenext
+ r")*)"
+ r")"
+ r"(?P<trailing>[,\.])?" # end of "allfuncs"
+ _description # Some function lists have a trailing comma (or period) '\s*'
)
# Empty <DESC> elements are replaced with '..'
empty_description = ".."
def _parse_see_also(self, content):
"""
func_name : Descriptive text
continued text
another_func_name : Descriptive text
func_name1, func_name2, :meth:`func_name`, func_name3
"""
content = dedent_lines(content)
items = []
def parse_item_name(text):
"""Match ':role:`name`' or 'name'."""
m = self._func_rgx.match(text)
if not m:
self._error_location(f"Error parsing See Also entry {line!r}")
role = m.group("role")
name = m.group("name") if role else m.group("name2")
return name, role, m.end()
rest = []
for line in content:
if not line.strip():
continue
line_match = self._line_rgx.match(line)
description = None
if line_match:
description = line_match.group("desc")
if line_match.group("trailing") and description:
self._error_location(
"Unexpected comma or period after function list at index %d of "
'line "%s"' % (line_match.end("trailing"), line),
error=False,
)
if not description and line.startswith(" "):
rest.append(line.strip())
elif line_match:
funcs = []
text = line_match.group("allfuncs")
while True:
if not text.strip():
break
name, role, match_end = parse_item_name(text)
funcs.append((name, role))
text = text[match_end:].strip()
if text and text[0] == ",":
text = text[1:].strip()
rest = list(filter(None, [description]))
items.append((funcs, rest))
else:
self._error_location(f"Error parsing See Also entry {line!r}")
return items
def _parse_index(self, section, content):
"""
.. index:: default
:refguide: something, else, and more
"""
def strip_each_in(lst):
return [s.strip() for s in lst]
out = {}
section = section.split("::")
if len(section) > 1:
out["default"] = strip_each_in(section[1].split(","))[0]
for line in content:
line = line.split(":")
if len(line) > 2:
out[line[1]] = strip_each_in(line[2].split(","))
return out
def _parse_summary(self):
"""Grab signature (if given) and summary"""
if self._is_at_section():
return
# If several signatures present, take the last one
while True:
summary = self._doc.read_to_next_empty_line()
summary_str = " ".join([s.strip() for s in summary]).strip()
compiled = re.compile(r"^([\w., ]+=)?\s*[\w\.]+\(.*\)$")
if compiled.match(summary_str):
self["Signature"] = summary_str
if not self._is_at_section():
continue
break
if summary is not None:
self["Summary"] = summary
if not self._is_at_section():
self["Extended Summary"] = self._read_to_next_section()
def _parse(self):
self._doc.reset()
self._parse_summary()
sections = list(self._read_sections())
section_names = {section for section, content in sections}
has_yields = "Yields" in section_names
# We could do more tests, but we are not. Arbitrarily.
if not has_yields and "Receives" in section_names:
msg = "Docstring contains a Receives section but not Yields."
raise ValueError(msg)
for section, content in sections:
if not section.startswith(".."):
section = (s.capitalize() for s in section.split(" "))
section = " ".join(section)
if self.get(section):
self._error_location(
"The section %s appears twice in %s"
% (section, "\n".join(self._doc._str))
)
if section in ("Parameters", "Other Parameters", "Attributes", "Methods"):
self[section] = self._parse_param_list(content)
elif section in ("Returns", "Yields", "Raises", "Warns", "Receives"):
self[section] = self._parse_param_list(
content, single_element_is_type=True
)
elif section.startswith(".. index::"):
self["index"] = self._parse_index(section, content)
elif section == "See Also":
self["See Also"] = self._parse_see_also(content)
else:
self[section] = content
@property
def _obj(self):
if hasattr(self, "_cls"):
return self._cls
elif hasattr(self, "_f"):
return self._f
return None
def _error_location(self, msg, error=True):
if self._obj is not None:
# we know where the docs came from:
try:
filename = inspect.getsourcefile(self._obj)
except TypeError:
filename = None
# Make UserWarning more descriptive via object introspection.
# Skip if introspection fails
name = getattr(self._obj, "__name__", None)
if name is None:
name = getattr(getattr(self._obj, "__class__", None), "__name__", None)
if name is not None:
msg += f" in the docstring of {name}"
msg += f" in {filename}." if filename else ""
if error:
raise ValueError(msg)
else:
warn(msg, stacklevel=3)
# string conversion routines
def _str_header(self, name, symbol="-"):
return [name, len(name) * symbol]
def _str_indent(self, doc, indent=4):
return [" " * indent + line for line in doc]
def _str_signature(self):
if self["Signature"]:
return [self["Signature"].replace("*", r"\*")] + [""]
return [""]
def _str_summary(self):
if self["Summary"]:
return self["Summary"] + [""]
return []
def _str_extended_summary(self):
if self["Extended Summary"]:
return self["Extended Summary"] + [""]
return []
def _str_param_list(self, name):
out = []
if self[name]:
out += self._str_header(name)
for param in self[name]:
parts = []
if param.name:
parts.append(param.name)
if param.type:
parts.append(param.type)
out += [" : ".join(parts)]
if param.desc and "".join(param.desc).strip():
out += self._str_indent(param.desc)
out += [""]
return out
def _str_section(self, name):
out = []
if self[name]:
out += self._str_header(name)
out += self[name]
out += [""]
return out
def _str_see_also(self, func_role):
if not self["See Also"]:
return []
out = []
out += self._str_header("See Also")
out += [""]
last_had_desc = True
for funcs, desc in self["See Also"]:
assert isinstance(funcs, list)
links = []
for func, role in funcs:
if role:
link = f":{role}:`{func}`"
elif func_role:
link = f":{func_role}:`{func}`"
else:
link = f"`{func}`_"
links.append(link)
link = ", ".join(links)
out += [link]
if desc:
out += self._str_indent([" ".join(desc)])
last_had_desc = True
else:
last_had_desc = False
out += self._str_indent([self.empty_description])
if last_had_desc:
out += [""]
out += [""]
return out
def _str_index(self):
idx = self["index"]
out = []
output_index = False
default_index = idx.get("default", "")
if default_index:
output_index = True
out += [f".. index:: {default_index}"]
for section, references in idx.items():
if section == "default":
continue
output_index = True
out += [f" :{section}: {', '.join(references)}"]
if output_index:
return out
return ""
def __str__(self, func_role=""):
out = []
out += self._str_signature()
out += self._str_summary()
out += self._str_extended_summary()
out += self._str_param_list("Parameters")
for param_list in ("Attributes", "Methods"):
out += self._str_param_list(param_list)
for param_list in (
"Returns",
"Yields",
"Receives",
"Other Parameters",
"Raises",
"Warns",
):
out += self._str_param_list(param_list)
out += self._str_section("Warnings")
out += self._str_see_also(func_role)
for s in ("Notes", "References", "Examples"):
out += self._str_section(s)
out += self._str_index()
return "\n".join(out)
def dedent_lines(lines):
"""Deindent a list of lines maximally"""
return textwrap.dedent("\n".join(lines)).split("\n")
class FunctionDoc(NumpyDocString):
def __init__(self, func, role="func", doc=None, config=None):
self._f = func
self._role = role # e.g. "func" or "meth"
if doc is None:
if func is None:
raise ValueError("No function or docstring given")
doc = inspect.getdoc(func) or ""
if config is None:
config = {}
NumpyDocString.__init__(self, doc, config)
def get_func(self):
func_name = getattr(self._f, "__name__", self.__class__.__name__)
if inspect.isclass(self._f):
func = getattr(self._f, "__call__", self._f.__init__)
else:
func = self._f
return func, func_name
def __str__(self):
out = ""
func, func_name = self.get_func()
roles = {"func": "function", "meth": "method"}
if self._role:
if self._role not in roles:
print(f"Warning: invalid role {self._role}")
out += f".. {roles.get(self._role, '')}:: {func_name}\n \n\n"
out += super().__str__(func_role=self._role)
return out
class ObjDoc(NumpyDocString):
def __init__(self, obj, doc=None, config=None):
self._f = obj
if config is None:
config = {}
NumpyDocString.__init__(self, doc, config=config)
class ClassDoc(NumpyDocString):
extra_public_methods = ["__call__"]
def __init__(self, cls, doc=None, modulename="", func_doc=FunctionDoc, config=None):
if not inspect.isclass(cls) and cls is not None:
raise ValueError(f"Expected a class or None, but got {cls!r}")
self._cls = cls
if "sphinx" in sys.modules:
from sphinx.ext.autodoc import ALL
else:
ALL = object()
if config is None:
config = {}
self.show_inherited_members = config.get("show_inherited_class_members", True)
if modulename and not modulename.endswith("."):
modulename += "."
self._mod = modulename
if doc is None:
if cls is None:
raise ValueError("No class or documentation string given")
doc = pydoc.getdoc(cls)
NumpyDocString.__init__(self, doc)
_members = config.get("members", [])
if _members is ALL:
_members = None
_exclude = config.get("exclude-members", [])
if config.get("show_class_members", True) and _exclude is not ALL:
def splitlines_x(s):
if not s:
return []
else:
return s.splitlines()
for field, items in [
("Methods", self.methods),
("Attributes", self.properties),
]:
if not self[field]:
doc_list = []
for name in sorted(items):
if name in _exclude or (_members and name not in _members):
continue
try:
doc_item = pydoc.getdoc(getattr(self._cls, name))
doc_list.append(Parameter(name, "", splitlines_x(doc_item)))
except AttributeError:
pass # method doesn't exist
self[field] = doc_list
@property
def methods(self):
if self._cls is None:
return []
return [
name
for name, func in inspect.getmembers(self._cls)
if (
(not name.startswith("_") or name in self.extra_public_methods)
and isinstance(func, Callable)
and self._is_show_member(name)
)
]
@property
def properties(self):
if self._cls is None:
return []
return [
name
for name, func in inspect.getmembers(self._cls)
if (
not name.startswith("_")
and not self._should_skip_member(name, self._cls)
and (
func is None
or isinstance(func, property | cached_property)
or inspect.isdatadescriptor(func)
)
and self._is_show_member(name)
)
]
@staticmethod
def _should_skip_member(name, klass):
return (
# Namedtuples should skip everything in their ._fields as the
# docstrings for each of the members is: "Alias for field number X"
issubclass(klass, tuple)
and hasattr(klass, "_asdict")
and hasattr(klass, "_fields")
and name in klass._fields
)
def _is_show_member(self, name):
return (
# show all class members
self.show_inherited_members
# or class member is not inherited
or name in self._cls.__dict__
)
def get_doc_object(
obj,
what=None,
doc=None,
config=None,
class_doc=ClassDoc,
func_doc=FunctionDoc,
obj_doc=ObjDoc,
):
if what is None:
if inspect.isclass(obj):
what = "class"
elif inspect.ismodule(obj):
what = "module"
elif isinstance(obj, Callable):
what = "function"
else:
what = "object"
if config is None:
config = {}
if what == "class":
return class_doc(obj, func_doc=func_doc, doc=doc, config=config)
elif what in ("function", "method"):
return func_doc(obj, doc=doc, config=config)
else:
if doc is None:
doc = pydoc.getdoc(obj)
return obj_doc(obj, doc, config=config)
@@ -0,0 +1,346 @@
# `_elementwise_iterative_method.py` includes tools for writing functions that
# - are vectorized to work elementwise on arrays,
# - implement non-trivial, iterative algorithms with a callback interface, and
# - return rich objects with iteration count, termination status, etc.
#
# Examples include:
# `scipy.optimize._chandrupatla._chandrupatla for scalar rootfinding,
# `scipy.optimize._chandrupatla._chandrupatla_minimize for scalar minimization,
# `scipy.optimize._differentiate._differentiate for numerical differentiation,
# `scipy.optimize._bracket._bracket_root for finding rootfinding brackets,
# `scipy.optimize._bracket._bracket_minimize for finding minimization brackets,
# `scipy.integrate._tanhsinh._tanhsinh` for numerical quadrature,
# `scipy.differentiate.derivative` for finite difference based differentiation.
import math
import numpy as np
from ._util import _RichResult, _call_callback_maybe_halt
from ._array_api import array_namespace, xp_size, xp_result_type
import scipy._lib.array_api_extra as xpx
_ESIGNERR = -1
_ECONVERR = -2
_EVALUEERR = -3
_ECALLBACK = -4
_EINPUTERR = -5
_ECONVERGED = 0
_EINPROGRESS = 1
def _initialize(func, xs, args, complex_ok=False, preserve_shape=None, xp=None):
"""Initialize abscissa, function, and args arrays for elementwise function
Parameters
----------
func : callable
An elementwise function with signature
func(x: ndarray, *args) -> ndarray
where each element of ``x`` is a finite real and ``args`` is a tuple,
which may contain an arbitrary number of arrays that are broadcastable
with ``x``.
xs : tuple of arrays
Finite real abscissa arrays. Must be broadcastable.
args : tuple, optional
Additional positional arguments to be passed to `func`.
preserve_shape : bool, default:False
When ``preserve_shape=False`` (default), `func` may be passed
arguments of any shape; `_scalar_optimization_loop` is permitted
to reshape and compress arguments at will. When
``preserve_shape=False``, arguments passed to `func` must have shape
`shape` or ``shape + (n,)``, where ``n`` is any integer.
xp : namespace
Namespace of array arguments in `xs`.
Returns
-------
xs, fs, args : tuple of arrays
Broadcasted, writeable, 1D abscissa and function value arrays (or
NumPy floats, if appropriate). The dtypes of the `xs` and `fs` are
`xfat`; the dtype of the `args` are unchanged.
shape : tuple of ints
Original shape of broadcasted arrays.
xfat : NumPy dtype
Result dtype of abscissae, function values, and args determined using
`np.result_type`, except integer types are promoted to `np.float64`.
Raises
------
ValueError
If the result dtype is not that of a real scalar
Notes
-----
Useful for initializing the input of SciPy functions that accept
an elementwise callable, abscissae, and arguments; e.g.
`scipy.optimize._chandrupatla`.
"""
nx = len(xs)
xp = array_namespace(*xs) if xp is None else xp
# Try to preserve `dtype`, but we need to ensure that the arguments are at
# least floats before passing them into the function; integers can overflow
# and cause failure.
# There might be benefit to combining the `xs` into a single array and
# calling `func` once on the combined array. For now, keep them separate.
xat = xp_result_type(*xs, force_floating=True, xp=xp)
xas = xp.broadcast_arrays(*xs, *args) # broadcast and rename
xs, args = xas[:nx], xas[nx:]
xs = [xp.asarray(x, dtype=xat) for x in xs] # use copy=False when implemented
fs = [xp.asarray(func(x, *args)) for x in xs]
shape = xs[0].shape
fshape = fs[0].shape
if preserve_shape:
# bind original shape/func now to avoid late-binding gotcha
def func(x, *args, shape=shape, func=func, **kwargs):
i = (0,)*(len(fshape) - len(shape))
return func(x[i], *args, **kwargs)
shape = np.broadcast_shapes(fshape, shape) # just shapes; use of NumPy OK
xs = [xp.broadcast_to(x, shape) for x in xs]
args = [xp.broadcast_to(arg, shape) for arg in args]
message = ("The shape of the array returned by `func` must be the same as "
"the broadcasted shape of `x` and all other `args`.")
if preserve_shape is not None: # only in tanhsinh for now
message = f"When `preserve_shape=False`, {message.lower()}"
shapes_equal = [f.shape == shape for f in fs]
if not all(shapes_equal): # use Python all to reduce overhead
raise ValueError(message)
# These algorithms tend to mix the dtypes of the abscissae and function
# values, so figure out what the result will be and convert them all to
# that type from the outset.
xfat = xp.result_type(*([f.dtype for f in fs] + [xat]))
if not complex_ok and not xp.isdtype(xfat, "real floating"):
raise ValueError("Abscissae and function output must be real numbers.")
xs = [xp.asarray(x, dtype=xfat, copy=True) for x in xs]
fs = [xp.asarray(f, dtype=xfat, copy=True) for f in fs]
# To ensure that we can do indexing, we'll work with at least 1d arrays,
# but remember the appropriate shape of the output.
xs = [xp.reshape(x, (-1,)) for x in xs]
fs = [xp.reshape(f, (-1,)) for f in fs]
args = [xp.reshape(xp.asarray(arg, copy=True), (-1,)) for arg in args]
return func, xs, fs, args, shape, xfat, xp
def _loop(work, callback, shape, maxiter, func, args, dtype, pre_func_eval,
post_func_eval, check_termination, post_termination_check,
customize_result, res_work_pairs, xp, preserve_shape=False):
"""Main loop of a vectorized scalar optimization algorithm
Parameters
----------
work : _RichResult
All variables that need to be retained between iterations. Must
contain attributes `nit`, `nfev`, and `success`. All arrays are
subject to being "compressed" if `preserve_shape is False`; nest
arrays that should not be compressed inside another object (e.g.
`dict` or `_RichResult`).
callback : callable
User-specified callback function
shape : tuple of ints
The shape of all output arrays
maxiter :
Maximum number of iterations of the algorithm
func : callable
The user-specified callable that is being optimized or solved
args : tuple
Additional positional arguments to be passed to `func`.
dtype : NumPy dtype
The common dtype of all abscissae and function values
pre_func_eval : callable
A function that accepts `work` and returns `x`, the active elements
of `x` at which `func` will be evaluated. May modify attributes
of `work` with any algorithmic steps that need to happen
at the beginning of an iteration, before `func` is evaluated,
post_func_eval : callable
A function that accepts `x`, `func(x)`, and `work`. May modify
attributes of `work` with any algorithmic steps that need to happen
in the middle of an iteration, after `func` is evaluated but before
the termination check.
check_termination : callable
A function that accepts `work` and returns `stop`, a boolean array
indicating which of the active elements have met a termination
condition.
post_termination_check : callable
A function that accepts `work`. May modify `work` with any algorithmic
steps that need to happen after the termination check and before the
end of the iteration.
customize_result : callable
A function that accepts `res` and `shape` and returns `shape`. May
modify `res` (in-place) according to preferences (e.g. rearrange
elements between attributes) and modify `shape` if needed.
res_work_pairs : list of (str, str)
Identifies correspondence between attributes of `res` and attributes
of `work`; i.e., attributes of active elements of `work` will be
copied to the appropriate indices of `res` when appropriate. The order
determines the order in which _RichResult attributes will be
pretty-printed.
preserve_shape : bool, default: False
Whether to compress the attributes of `work` (to avoid unnecessary
computation on elements that have already converged).
Returns
-------
res : _RichResult
The final result object
Notes
-----
Besides providing structure, this framework provides several important
services for a vectorized optimization algorithm.
- It handles common tasks involving iteration count, function evaluation
count, a user-specified callback, and associated termination conditions.
- It compresses the attributes of `work` to eliminate unnecessary
computation on elements that have already converged.
"""
if xp is None:
raise NotImplementedError("Must provide xp.")
cb_terminate = False
# Initialize the result object and active element index array
n_elements = math.prod(shape)
active = xp.arange(n_elements) # in-progress element indices
res_dict = {i: xp.zeros(n_elements, dtype=dtype) for i, j in res_work_pairs}
res_dict['success'] = xp.zeros(n_elements, dtype=xp.bool)
res_dict['status'] = xp.full(n_elements, xp.asarray(_EINPROGRESS), dtype=xp.int32)
res_dict['nit'] = xp.zeros(n_elements, dtype=xp.int32)
res_dict['nfev'] = xp.zeros(n_elements, dtype=xp.int32)
res = _RichResult(res_dict)
work.args = args
active = _check_termination(work, res, res_work_pairs, active,
check_termination, preserve_shape, xp)
if callback is not None:
temp = _prepare_result(work, res, res_work_pairs, active, shape,
customize_result, preserve_shape, xp)
if _call_callback_maybe_halt(callback, temp):
cb_terminate = True
while work.nit < maxiter and xp_size(active) and not cb_terminate and n_elements:
x = pre_func_eval(work)
if work.args and work.args[0].ndim != x.ndim:
# `x` always starts as 1D. If the SciPy function that uses
# _loop added dimensions to `x`, we need to
# add them to the elements of `args`.
args = []
for arg in work.args:
n_new_dims = x.ndim - arg.ndim
new_shape = arg.shape + (1,)*n_new_dims
args.append(xp.reshape(arg, new_shape))
work.args = args
x_shape = x.shape
if preserve_shape:
x = xp.reshape(x, (shape + (-1,)))
f = func(x, *work.args)
f = xp.asarray(f, dtype=dtype)
if preserve_shape:
x = xp.reshape(x, x_shape)
f = xp.reshape(f, x_shape)
work.nfev += 1 if x.ndim == 1 else x.shape[-1]
post_func_eval(x, f, work)
work.nit += 1
active = _check_termination(work, res, res_work_pairs, active,
check_termination, preserve_shape, xp)
if callback is not None:
temp = _prepare_result(work, res, res_work_pairs, active, shape,
customize_result, preserve_shape, xp)
if _call_callback_maybe_halt(callback, temp):
cb_terminate = True
break
if xp_size(active) == 0:
break
post_termination_check(work)
work.status = xpx.at(work.status)[:].set(_ECALLBACK if cb_terminate else _ECONVERR)
return _prepare_result(work, res, res_work_pairs, active, shape,
customize_result, preserve_shape, xp)
def _check_termination(work, res, res_work_pairs, active, check_termination,
preserve_shape, xp):
# Checks termination conditions, updates elements of `res` with
# corresponding elements of `work`, and compresses `work`.
stop = check_termination(work)
if xp.any(stop):
# update the active elements of the result object with the active
# elements for which a termination condition has been met
_update_active(work, res, res_work_pairs, active, stop, preserve_shape, xp)
if preserve_shape:
stop = stop[active]
proceed = ~stop
active = active[proceed]
if not preserve_shape:
# compress the arrays to avoid unnecessary computation
for key, val in work.items():
# `continued_fraction` hacks `n`; improve if this becomes a problem
if key in {'args', 'n'}:
continue
work[key] = val[proceed] if getattr(val, 'ndim', 0) > 0 else val
work.args = [arg[proceed] for arg in work.args]
return active
def _update_active(work, res, res_work_pairs, active, mask, preserve_shape, xp):
# Update `active` indices of the arrays in result object `res` with the
# contents of the scalars and arrays in `update_dict`. When provided,
# `mask` is a boolean array applied both to the arrays in `update_dict`
# that are to be used and to the arrays in `res` that are to be updated.
update_dict = {key1: work[key2] for key1, key2 in res_work_pairs}
update_dict['success'] = work.status == 0
if mask is not None:
if preserve_shape:
active_mask = xp.zeros_like(mask)
active_mask = xpx.at(active_mask)[active].set(True)
active_mask = active_mask & mask
for key, val in update_dict.items():
val = val[active_mask] if getattr(val, 'ndim', 0) > 0 else val
res[key] = xpx.at(res[key])[active_mask].set(val)
else:
active_mask = active[mask]
for key, val in update_dict.items():
val = val[mask] if getattr(val, 'ndim', 0) > 0 else val
res[key] = xpx.at(res[key])[active_mask].set(val)
else:
for key, val in update_dict.items():
if preserve_shape and getattr(val, 'ndim', 0) > 0:
val = val[active]
res[key] = xpx.at(res[key])[active].set(val)
def _prepare_result(work, res, res_work_pairs, active, shape, customize_result,
preserve_shape, xp):
# Prepare the result object `res` by creating a copy, copying the latest
# data from work, running the provided result customization function,
# and reshaping the data to the original shapes.
res = res.copy()
_update_active(work, res, res_work_pairs, active, None, preserve_shape, xp)
shape = customize_result(res, shape)
for key, val in res.items():
# this looks like it won't work for xp != np if val is not numeric
temp = xp.reshape(val, shape)
res[key] = temp[()] if temp.ndim == 0 else temp
res['_order_keys'] = ['success'] + [i for i, j in res_work_pairs]
return _RichResult(**res)
@@ -0,0 +1,105 @@
"""
Module for testing automatic garbage collection of objects
.. autosummary::
:toctree: generated/
set_gc_state - enable or disable garbage collection
gc_state - context manager for given state of garbage collector
assert_deallocated - context manager to check for circular references on object
"""
import weakref
import gc
from contextlib import contextmanager
from platform import python_implementation
__all__ = ['set_gc_state', 'gc_state', 'assert_deallocated']
IS_PYPY = python_implementation() == 'PyPy'
class ReferenceError(AssertionError):
pass
def set_gc_state(state):
""" Set status of garbage collector """
if gc.isenabled() == state:
return
if state:
gc.enable()
else:
gc.disable()
@contextmanager
def gc_state(state):
""" Context manager to set state of garbage collector to `state`
Parameters
----------
state : bool
True for gc enabled, False for disabled
Examples
--------
>>> with gc_state(False):
... assert not gc.isenabled()
>>> with gc_state(True):
... assert gc.isenabled()
"""
orig_state = gc.isenabled()
set_gc_state(state)
yield
set_gc_state(orig_state)
@contextmanager
def assert_deallocated(func, *args, **kwargs):
"""Context manager to check that object is deallocated
This is useful for checking that an object can be freed directly by
reference counting, without requiring gc to break reference cycles.
GC is disabled inside the context manager.
This check is not available on PyPy.
Parameters
----------
func : callable
Callable to create object to check
\\*args : sequence
positional arguments to `func` in order to create object to check
\\*\\*kwargs : dict
keyword arguments to `func` in order to create object to check
Examples
--------
>>> class C: pass
>>> with assert_deallocated(C) as c:
... # do something
... del c
>>> class C:
... def __init__(self):
... self._circular = self # Make circular reference
>>> with assert_deallocated(C) as c: #doctest: +IGNORE_EXCEPTION_DETAIL
... # do something
... del c
Traceback (most recent call last):
...
ReferenceError: Remaining reference(s) to object
"""
if IS_PYPY:
raise RuntimeError("assert_deallocated is unavailable on PyPy")
with gc_state(False):
obj = func(*args, **kwargs)
ref = weakref.ref(obj)
yield obj
del obj
if ref() is not None:
raise ReferenceError("Remaining reference(s) to object")
@@ -0,0 +1,487 @@
"""Utility to compare pep440 compatible version strings.
The LooseVersion and StrictVersion classes that distutils provides don't
work; they don't recognize anything like alpha/beta/rc/dev versions.
"""
# Copyright (c) Donald Stufft and individual contributors.
# All rights reserved.
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
# 1. Redistributions of source code must retain the above copyright notice,
# this list of conditions and the following disclaimer.
# 2. Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
# POSSIBILITY OF SUCH DAMAGE.
import collections
import itertools
import re
__all__ = [
"parse", "Version", "LegacyVersion", "InvalidVersion", "VERSION_PATTERN",
]
# BEGIN packaging/_structures.py
class Infinity:
def __repr__(self):
return "Infinity"
def __hash__(self):
return hash(repr(self))
def __lt__(self, other):
return False
def __le__(self, other):
return False
def __eq__(self, other):
return isinstance(other, self.__class__)
def __ne__(self, other):
return not isinstance(other, self.__class__)
def __gt__(self, other):
return True
def __ge__(self, other):
return True
def __neg__(self):
return NegativeInfinity
Infinity = Infinity()
class NegativeInfinity:
def __repr__(self):
return "-Infinity"
def __hash__(self):
return hash(repr(self))
def __lt__(self, other):
return True
def __le__(self, other):
return True
def __eq__(self, other):
return isinstance(other, self.__class__)
def __ne__(self, other):
return not isinstance(other, self.__class__)
def __gt__(self, other):
return False
def __ge__(self, other):
return False
def __neg__(self):
return Infinity
# BEGIN packaging/version.py
NegativeInfinity = NegativeInfinity()
_Version = collections.namedtuple(
"_Version",
["epoch", "release", "dev", "pre", "post", "local"],
)
def parse(version):
"""
Parse the given version string and return either a :class:`Version` object
or a :class:`LegacyVersion` object depending on if the given version is
a valid PEP 440 version or a legacy version.
"""
try:
return Version(version)
except InvalidVersion:
return LegacyVersion(version)
class InvalidVersion(ValueError):
"""
An invalid version was found, users should refer to PEP 440.
"""
class _BaseVersion:
def __hash__(self):
return hash(self._key)
def __lt__(self, other):
return self._compare(other, lambda s, o: s < o)
def __le__(self, other):
return self._compare(other, lambda s, o: s <= o)
def __eq__(self, other):
return self._compare(other, lambda s, o: s == o)
def __ge__(self, other):
return self._compare(other, lambda s, o: s >= o)
def __gt__(self, other):
return self._compare(other, lambda s, o: s > o)
def __ne__(self, other):
return self._compare(other, lambda s, o: s != o)
def _compare(self, other, method):
if not isinstance(other, _BaseVersion):
return NotImplemented
return method(self._key, other._key)
class LegacyVersion(_BaseVersion):
def __init__(self, version):
self._version = str(version)
self._key = _legacy_cmpkey(self._version)
def __str__(self):
return self._version
def __repr__(self):
return f"<LegacyVersion({repr(str(self))})>"
@property
def public(self):
return self._version
@property
def base_version(self):
return self._version
@property
def local(self):
return None
@property
def is_prerelease(self):
return False
@property
def is_postrelease(self):
return False
_legacy_version_component_re = re.compile(
r"(\d+ | [a-z]+ | \.| -)", re.VERBOSE,
)
_legacy_version_replacement_map = {
"pre": "c", "preview": "c", "-": "final-", "rc": "c", "dev": "@",
}
def _parse_version_parts(s):
for part in _legacy_version_component_re.split(s):
part = _legacy_version_replacement_map.get(part, part)
if not part or part == ".":
continue
if part[:1] in "0123456789":
# pad for numeric comparison
yield part.zfill(8)
else:
yield "*" + part
# ensure that alpha/beta/candidate are before final
yield "*final"
def _legacy_cmpkey(version):
# We hardcode an epoch of -1 here. A PEP 440 version can only have an epoch
# greater than or equal to 0. This will effectively put the LegacyVersion,
# which uses the defacto standard originally implemented by setuptools,
# as before all PEP 440 versions.
epoch = -1
# This scheme is taken from pkg_resources.parse_version setuptools prior to
# its adoption of the packaging library.
parts = []
for part in _parse_version_parts(version.lower()):
if part.startswith("*"):
# remove "-" before a prerelease tag
if part < "*final":
while parts and parts[-1] == "*final-":
parts.pop()
# remove trailing zeros from each series of numeric parts
while parts and parts[-1] == "00000000":
parts.pop()
parts.append(part)
parts = tuple(parts)
return epoch, parts
# Deliberately not anchored to the start and end of the string, to make it
# easier for 3rd party code to reuse
VERSION_PATTERN = r"""
v?
(?:
(?:(?P<epoch>[0-9]+)!)? # epoch
(?P<release>[0-9]+(?:\.[0-9]+)*) # release segment
(?P<pre> # pre-release
[-_\.]?
(?P<pre_l>(a|b|c|rc|alpha|beta|pre|preview))
[-_\.]?
(?P<pre_n>[0-9]+)?
)?
(?P<post> # post release
(?:-(?P<post_n1>[0-9]+))
|
(?:
[-_\.]?
(?P<post_l>post|rev|r)
[-_\.]?
(?P<post_n2>[0-9]+)?
)
)?
(?P<dev> # dev release
[-_\.]?
(?P<dev_l>dev)
[-_\.]?
(?P<dev_n>[0-9]+)?
)?
)
(?:\+(?P<local>[a-z0-9]+(?:[-_\.][a-z0-9]+)*))? # local version
"""
class Version(_BaseVersion):
_regex = re.compile(
r"^\s*" + VERSION_PATTERN + r"\s*$",
re.VERBOSE | re.IGNORECASE,
)
def __init__(self, version):
# Validate the version and parse it into pieces
match = self._regex.search(version)
if not match:
raise InvalidVersion(f"Invalid version: '{version}'")
# Store the parsed out pieces of the version
self._version = _Version(
epoch=int(match.group("epoch")) if match.group("epoch") else 0,
release=tuple(int(i) for i in match.group("release").split(".")),
pre=_parse_letter_version(
match.group("pre_l"),
match.group("pre_n"),
),
post=_parse_letter_version(
match.group("post_l"),
match.group("post_n1") or match.group("post_n2"),
),
dev=_parse_letter_version(
match.group("dev_l"),
match.group("dev_n"),
),
local=_parse_local_version(match.group("local")),
)
# Generate a key which will be used for sorting
self._key = _cmpkey(
self._version.epoch,
self._version.release,
self._version.pre,
self._version.post,
self._version.dev,
self._version.local,
)
def __repr__(self):
return f"<Version({repr(str(self))})>"
def __str__(self):
parts = []
# Epoch
if self._version.epoch != 0:
parts.append(f"{self._version.epoch}!")
# Release segment
parts.append(".".join(str(x) for x in self._version.release))
# Pre-release
if self._version.pre is not None:
parts.append("".join(str(x) for x in self._version.pre))
# Post-release
if self._version.post is not None:
parts.append(f".post{self._version.post[1]}")
# Development release
if self._version.dev is not None:
parts.append(f".dev{self._version.dev[1]}")
# Local version segment
if self._version.local is not None:
parts.append(
"+{}".format(".".join(str(x) for x in self._version.local))
)
return "".join(parts)
@property
def public(self):
return str(self).split("+", 1)[0]
@property
def base_version(self):
parts = []
# Epoch
if self._version.epoch != 0:
parts.append(f"{self._version.epoch}!")
# Release segment
parts.append(".".join(str(x) for x in self._version.release))
return "".join(parts)
@property
def local(self):
version_string = str(self)
if "+" in version_string:
return version_string.split("+", 1)[1]
@property
def is_prerelease(self):
return bool(self._version.dev or self._version.pre)
@property
def is_postrelease(self):
return bool(self._version.post)
def _parse_letter_version(letter, number):
if letter:
# We assume there is an implicit 0 in a pre-release if there is
# no numeral associated with it.
if number is None:
number = 0
# We normalize any letters to their lower-case form
letter = letter.lower()
# We consider some words to be alternate spellings of other words and
# in those cases we want to normalize the spellings to our preferred
# spelling.
if letter == "alpha":
letter = "a"
elif letter == "beta":
letter = "b"
elif letter in ["c", "pre", "preview"]:
letter = "rc"
elif letter in ["rev", "r"]:
letter = "post"
return letter, int(number)
if not letter and number:
# We assume that if we are given a number but not given a letter,
# then this is using the implicit post release syntax (e.g., 1.0-1)
letter = "post"
return letter, int(number)
_local_version_seperators = re.compile(r"[\._-]")
def _parse_local_version(local):
"""
Takes a string like abc.1.twelve and turns it into ("abc", 1, "twelve").
"""
if local is not None:
return tuple(
part.lower() if not part.isdigit() else int(part)
for part in _local_version_seperators.split(local)
)
def _cmpkey(epoch, release, pre, post, dev, local):
# When we compare a release version, we want to compare it with all of the
# trailing zeros removed. So we'll use a reverse the list, drop all the now
# leading zeros until we come to something non-zero, then take the rest,
# re-reverse it back into the correct order, and make it a tuple and use
# that for our sorting key.
release = tuple(
reversed(list(
itertools.dropwhile(
lambda x: x == 0,
reversed(release),
)
))
)
# We need to "trick" the sorting algorithm to put 1.0.dev0 before 1.0a0.
# We'll do this by abusing the pre-segment, but we _only_ want to do this
# if there is no pre- or a post-segment. If we have one of those, then
# the normal sorting rules will handle this case correctly.
if pre is None and post is None and dev is not None:
pre = -Infinity
# Versions without a pre-release (except as noted above) should sort after
# those with one.
elif pre is None:
pre = Infinity
# Versions without a post-segment should sort before those with one.
if post is None:
post = -Infinity
# Versions without a development segment should sort after those with one.
if dev is None:
dev = Infinity
if local is None:
# Versions without a local segment should sort before those with one.
local = -Infinity
else:
# Versions with a local segment need that segment parsed to implement
# the sorting rules in PEP440.
# - Alphanumeric segments sort before numeric segments
# - Alphanumeric segments sort lexicographically
# - Numeric segments sort numerically
# - Shorter versions sort before longer versions when the prefixes
# match exactly
local = tuple(
(i, "") if isinstance(i, int) else (-Infinity, i)
for i in local
)
return epoch, release, pre, post, dev, local
@@ -0,0 +1,41 @@
from abc import ABC
__all__ = ["SparseABC", "issparse"]
class SparseABC(ABC):
pass
def issparse(x):
"""Is `x` of a sparse array or sparse matrix type?
Parameters
----------
x
object to check for being a sparse array or sparse matrix
Returns
-------
bool
True if `x` is a sparse array or a sparse matrix, False otherwise
Notes
-----
Use `isinstance(x, sp.sparse.sparray)` to check between an array or matrix.
Use `a.format` to check the sparse format, e.g. `a.format == 'csr'`.
Examples
--------
>>> import numpy as np
>>> from scipy.sparse import csr_array, csr_matrix, issparse
>>> issparse(csr_matrix([[5]]))
True
>>> issparse(csr_array([[5]]))
True
>>> issparse(np.array([[5]]))
False
>>> issparse(5)
False
"""
return isinstance(x, SparseABC)
@@ -0,0 +1,373 @@
"""
Generic test utilities.
"""
import inspect
import os
import re
import shutil
import subprocess
import sys
import sysconfig
import threading
from importlib.util import module_from_spec, spec_from_file_location
import numpy as np
import scipy
try:
# Need type: ignore[import-untyped] for mypy >= 1.6
import cython # type: ignore[import-untyped]
from Cython.Compiler.Version import ( # type: ignore[import-untyped]
version as cython_version,
)
except ImportError:
cython = None
else:
from scipy._lib import _pep440
required_version = '3.0.8'
if _pep440.parse(cython_version) < _pep440.Version(required_version):
# too old or wrong cython, skip Cython API tests
cython = None
__all__ = ['PytestTester', 'check_free_memory', '_TestPythranFunc', 'IS_MUSL']
IS_MUSL = False
# alternate way is
# from packaging.tags import sys_tags
# _tags = list(sys_tags())
# if 'musllinux' in _tags[0].platform:
_v = sysconfig.get_config_var('HOST_GNU_TYPE') or ''
if 'musl' in _v:
IS_MUSL = True
IS_EDITABLE = 'editable' in scipy.__path__[0]
class FPUModeChangeWarning(RuntimeWarning):
"""Warning about FPU mode change"""
pass
class PytestTester:
"""
Run tests for this namespace
``scipy.test()`` runs tests for all of SciPy, with the default settings.
When used from a submodule (e.g., ``scipy.cluster.test()``, only the tests
for that namespace are run.
Parameters
----------
label : {'fast', 'full'}, optional
Whether to run only the fast tests, or also those marked as slow.
Default is 'fast'.
verbose : int, optional
Test output verbosity. Default is 1.
extra_argv : list, optional
Arguments to pass through to Pytest.
doctests : bool, optional
Whether to run doctests or not. Default is False.
coverage : bool, optional
Whether to run tests with code coverage measurements enabled.
Default is False.
tests : list of str, optional
List of module names to run tests for. By default, uses the module
from which the ``test`` function is called.
parallel : int, optional
Run tests in parallel with pytest-xdist, if number given is larger than
1. Default is 1.
"""
def __init__(self, module_name):
self.module_name = module_name
def __call__(self, label="fast", verbose=1, extra_argv=None, doctests=False,
coverage=False, tests=None, parallel=None):
import pytest
module = sys.modules[self.module_name]
module_path = os.path.abspath(module.__path__[0])
pytest_args = ['--showlocals', '--tb=short']
if extra_argv is None:
extra_argv = []
pytest_args += extra_argv
if any(arg == "-m" or arg == "--markers" for arg in extra_argv):
# Likely conflict with default --mode=fast
raise ValueError("Must specify -m before --")
if verbose and int(verbose) > 1:
pytest_args += ["-" + "v"*(int(verbose)-1)]
if coverage:
pytest_args += ["--cov=" + module_path]
if label == "fast":
pytest_args += ["-m", "not slow"]
elif label != "full":
pytest_args += ["-m", label]
if tests is None:
tests = [self.module_name]
if parallel is not None and parallel > 1:
if _pytest_has_xdist():
pytest_args += ['-n', str(parallel)]
else:
import warnings
warnings.warn('Could not run tests in parallel because '
'pytest-xdist plugin is not available.',
stacklevel=2)
pytest_args += ['--pyargs'] + list(tests)
try:
code = pytest.main(pytest_args)
except SystemExit as exc:
code = exc.code
return (code == 0)
class _TestPythranFunc:
'''
These are situations that can be tested in our pythran tests:
- A function with multiple array arguments and then
other positional and keyword arguments.
- A function with array-like keywords (e.g. `def somefunc(x0, x1=None)`.
Note: list/tuple input is not yet tested!
`self.arguments`: A dictionary which key is the index of the argument,
value is tuple(array value, all supported dtypes)
`self.partialfunc`: A function used to freeze some non-array argument
that of no interests in the original function
'''
ALL_INTEGER = [np.int8, np.int16, np.int32, np.int64, np.intc, np.intp]
ALL_FLOAT = [np.float32, np.float64]
ALL_COMPLEX = [np.complex64, np.complex128]
def setup_method(self):
self.arguments = {}
self.partialfunc = None
self.expected = None
def get_optional_args(self, func):
# get optional arguments with its default value,
# used for testing keywords
signature = inspect.signature(func)
optional_args = {}
for k, v in signature.parameters.items():
if v.default is not inspect.Parameter.empty:
optional_args[k] = v.default
return optional_args
def get_max_dtype_list_length(self):
# get the max supported dtypes list length in all arguments
max_len = 0
for arg_idx in self.arguments:
cur_len = len(self.arguments[arg_idx][1])
if cur_len > max_len:
max_len = cur_len
return max_len
def get_dtype(self, dtype_list, dtype_idx):
# get the dtype from dtype_list via index
# if the index is out of range, then return the last dtype
if dtype_idx > len(dtype_list)-1:
return dtype_list[-1]
else:
return dtype_list[dtype_idx]
def test_all_dtypes(self):
for type_idx in range(self.get_max_dtype_list_length()):
args_array = []
for arg_idx in self.arguments:
new_dtype = self.get_dtype(self.arguments[arg_idx][1],
type_idx)
args_array.append(self.arguments[arg_idx][0].astype(new_dtype))
self.pythranfunc(*args_array)
def test_views(self):
args_array = []
for arg_idx in self.arguments:
args_array.append(self.arguments[arg_idx][0][::-1][::-1])
self.pythranfunc(*args_array)
def test_strided(self):
args_array = []
for arg_idx in self.arguments:
args_array.append(np.repeat(self.arguments[arg_idx][0],
2, axis=0)[::2])
self.pythranfunc(*args_array)
def _pytest_has_xdist():
"""
Check if the pytest-xdist plugin is installed, providing parallel tests
"""
# Check xdist exists without importing, otherwise pytests emits warnings
from importlib.util import find_spec
return find_spec('xdist') is not None
def check_free_memory(free_mb):
"""
Check *free_mb* of memory is available, otherwise do pytest.skip
"""
import pytest
try:
mem_free = _parse_size(os.environ['SCIPY_AVAILABLE_MEM'])
msg = '{} MB memory required, but environment SCIPY_AVAILABLE_MEM={}'.format(
free_mb, os.environ['SCIPY_AVAILABLE_MEM'])
except KeyError:
mem_free = _get_mem_available()
if mem_free is None:
pytest.skip("Could not determine available memory; set SCIPY_AVAILABLE_MEM "
"variable to free memory in MB to run the test.")
msg = f'{free_mb} MB memory required, but {mem_free/1e6} MB available'
if mem_free < free_mb * 1e6:
pytest.skip(msg)
def _parse_size(size_str):
suffixes = {'': 1e6,
'b': 1.0,
'k': 1e3, 'M': 1e6, 'G': 1e9, 'T': 1e12,
'kb': 1e3, 'Mb': 1e6, 'Gb': 1e9, 'Tb': 1e12,
'kib': 1024.0, 'Mib': 1024.0**2, 'Gib': 1024.0**3, 'Tib': 1024.0**4}
m = re.match(r'^\s*(\d+)\s*({})\s*$'.format('|'.join(suffixes.keys())),
size_str,
re.I)
if not m or m.group(2) not in suffixes:
raise ValueError("Invalid size string")
return float(m.group(1)) * suffixes[m.group(2)]
def _get_mem_available():
"""
Get information about memory available, not counting swap.
"""
try:
import psutil
return psutil.virtual_memory().available
except (ImportError, AttributeError):
pass
if sys.platform.startswith('linux'):
info = {}
with open('/proc/meminfo') as f:
for line in f:
p = line.split()
info[p[0].strip(':').lower()] = float(p[1]) * 1e3
if 'memavailable' in info:
# Linux >= 3.14
return info['memavailable']
else:
return info['memfree'] + info['cached']
return None
def _test_cython_extension(tmp_path, srcdir):
"""
Helper function to test building and importing Cython modules that
make use of the Cython APIs for BLAS, LAPACK, optimize, and special.
"""
import pytest
try:
subprocess.check_call(["meson", "--version"])
except FileNotFoundError:
pytest.skip("No usable 'meson' found")
# Make safe for being called by multiple threads within one test
tmp_path = tmp_path / str(threading.get_ident())
# build the examples in a temporary directory
mod_name = os.path.split(srcdir)[1]
shutil.copytree(srcdir, tmp_path / mod_name)
build_dir = tmp_path / mod_name / 'tests' / '_cython_examples'
target_dir = build_dir / 'build'
os.makedirs(target_dir, exist_ok=True)
# Ensure we use the correct Python interpreter even when `meson` is
# installed in a different Python environment (see numpy#24956)
native_file = str(build_dir / 'interpreter-native-file.ini')
with open(native_file, 'w') as f:
f.write("[binaries]\n")
f.write(f"python = '{sys.executable}'")
if sys.platform == "win32":
subprocess.check_call(["meson", "setup",
"--buildtype=release",
"--native-file", native_file,
"--vsenv", str(build_dir)],
cwd=target_dir,
)
else:
subprocess.check_call(["meson", "setup",
"--native-file", native_file, str(build_dir)],
cwd=target_dir
)
subprocess.check_call(["meson", "compile", "-vv"], cwd=target_dir)
# import without adding the directory to sys.path
suffix = sysconfig.get_config_var('EXT_SUFFIX')
def load(modname):
so = (target_dir / modname).with_suffix(suffix)
spec = spec_from_file_location(modname, so)
mod = module_from_spec(spec)
spec.loader.exec_module(mod)
return mod
# test that the module can be imported
return load("extending"), load("extending_cpp")
def _run_concurrent_barrier(n_workers, fn, *args, **kwargs):
"""
Run a given function concurrently across a given number of threads.
This is equivalent to using a ThreadPoolExecutor, but using the threading
primitives instead. This function ensures that the closure passed by
parameter gets called concurrently by setting up a barrier before it gets
called before any of the threads.
Arguments
---------
n_workers: int
Number of concurrent threads to spawn.
fn: callable
Function closure to execute concurrently. Its first argument will
be the thread id.
*args: tuple
Variable number of positional arguments to pass to the function.
**kwargs: dict
Keyword arguments to pass to the function.
"""
barrier = threading.Barrier(n_workers)
def closure(i, *args, **kwargs):
barrier.wait()
fn(i, *args, **kwargs)
workers = []
for i in range(0, n_workers):
workers.append(threading.Thread(
target=closure,
args=(i,) + args, kwargs=kwargs))
for worker in workers:
worker.start()
for worker in workers:
worker.join()
@@ -0,0 +1,58 @@
import threading
import scipy._lib.decorator
__all__ = ['ReentrancyError', 'ReentrancyLock', 'non_reentrant']
class ReentrancyError(RuntimeError):
pass
class ReentrancyLock:
"""
Threading lock that raises an exception for reentrant calls.
Calls from different threads are serialized, and nested calls from the
same thread result to an error.
The object can be used as a context manager or to decorate functions
via the decorate() method.
"""
def __init__(self, err_msg):
self._rlock = threading.RLock()
self._entered = False
self._err_msg = err_msg
def __enter__(self):
self._rlock.acquire()
if self._entered:
self._rlock.release()
raise ReentrancyError(self._err_msg)
self._entered = True
def __exit__(self, type, value, traceback):
self._entered = False
self._rlock.release()
def decorate(self, func):
def caller(func, *a, **kw):
with self:
return func(*a, **kw)
return scipy._lib.decorator.decorate(func, caller)
def non_reentrant(err_msg=None):
"""
Decorate a function with a threading lock and prevent reentrant calls.
"""
def decorator(func):
msg = err_msg
if msg is None:
msg = f"{func.__name__} is not re-entrant"
lock = ReentrancyLock(msg)
return lock.decorate(func)
return decorator
@@ -0,0 +1,86 @@
''' Contexts for *with* statement providing temporary directories
'''
import os
from contextlib import contextmanager
from shutil import rmtree
from tempfile import mkdtemp
@contextmanager
def tempdir():
"""Create and return a temporary directory. This has the same
behavior as mkdtemp but can be used as a context manager.
Upon exiting the context, the directory and everything contained
in it are removed.
Examples
--------
>>> import os
>>> with tempdir() as tmpdir:
... fname = os.path.join(tmpdir, 'example_file.txt')
... with open(fname, 'wt') as fobj:
... _ = fobj.write('a string\\n')
>>> os.path.exists(tmpdir)
False
"""
d = mkdtemp()
yield d
rmtree(d)
@contextmanager
def in_tempdir():
''' Create, return, and change directory to a temporary directory
Examples
--------
>>> import os
>>> my_cwd = os.getcwd()
>>> with in_tempdir() as tmpdir:
... _ = open('test.txt', 'wt').write('some text')
... assert os.path.isfile('test.txt')
... assert os.path.isfile(os.path.join(tmpdir, 'test.txt'))
>>> os.path.exists(tmpdir)
False
>>> os.getcwd() == my_cwd
True
'''
pwd = os.getcwd()
d = mkdtemp()
os.chdir(d)
yield d
os.chdir(pwd)
rmtree(d)
@contextmanager
def in_dir(dir=None):
""" Change directory to given directory for duration of ``with`` block
Useful when you want to use `in_tempdir` for the final test, but
you are still debugging. For example, you may want to do this in the end:
>>> with in_tempdir() as tmpdir:
... # do something complicated which might break
... pass
But, indeed, the complicated thing does break, and meanwhile, the
``in_tempdir`` context manager wiped out the directory with the
temporary files that you wanted for debugging. So, while debugging, you
replace with something like:
>>> with in_dir() as tmpdir: # Use working directory by default
... # do something complicated which might break
... pass
You can then look at the temporary file outputs to debug what is happening,
fix, and finally replace ``in_dir`` with ``in_tempdir`` again.
"""
cwd = os.getcwd()
if dir is None:
yield cwd
return
os.chdir(dir)
yield dir
os.chdir(cwd)
@@ -0,0 +1,29 @@
BSD 3-Clause License
Copyright (c) 2018, Quansight-Labs
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright notice, this
list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
* Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
@@ -0,0 +1,116 @@
"""
.. note:
If you are looking for overrides for NumPy-specific methods, see the
documentation for :obj:`unumpy`. This page explains how to write
back-ends and multimethods.
``uarray`` is built around a back-end protocol, and overridable multimethods.
It is necessary to define multimethods for back-ends to be able to override them.
See the documentation of :obj:`generate_multimethod` on how to write multimethods.
Let's start with the simplest:
``__ua_domain__`` defines the back-end *domain*. The domain consists of period-
separated string consisting of the modules you extend plus the submodule. For
example, if a submodule ``module2.submodule`` extends ``module1``
(i.e., it exposes dispatchables marked as types available in ``module1``),
then the domain string should be ``"module1.module2.submodule"``.
For the purpose of this demonstration, we'll be creating an object and setting
its attributes directly. However, note that you can use a module or your own type
as a backend as well.
>>> class Backend: pass
>>> be = Backend()
>>> be.__ua_domain__ = "ua_examples"
It might be useful at this point to sidetrack to the documentation of
:obj:`generate_multimethod` to find out how to generate a multimethod
overridable by :obj:`uarray`. Needless to say, writing a backend and
creating multimethods are mostly orthogonal activities, and knowing
one doesn't necessarily require knowledge of the other, although it
is certainly helpful. We expect core API designers/specifiers to write the
multimethods, and implementors to override them. But, as is often the case,
similar people write both.
Without further ado, here's an example multimethod:
>>> import uarray as ua
>>> from uarray import Dispatchable
>>> def override_me(a, b):
... return Dispatchable(a, int),
>>> def override_replacer(args, kwargs, dispatchables):
... return (dispatchables[0], args[1]), {}
>>> overridden_me = ua.generate_multimethod(
... override_me, override_replacer, "ua_examples"
... )
Next comes the part about overriding the multimethod. This requires
the ``__ua_function__`` protocol, and the ``__ua_convert__``
protocol. The ``__ua_function__`` protocol has the signature
``(method, args, kwargs)`` where ``method`` is the passed
multimethod, ``args``/``kwargs`` specify the arguments and ``dispatchables``
is the list of converted dispatchables passed in.
>>> def __ua_function__(method, args, kwargs):
... return method.__name__, args, kwargs
>>> be.__ua_function__ = __ua_function__
The other protocol of interest is the ``__ua_convert__`` protocol. It has the
signature ``(dispatchables, coerce)``. When ``coerce`` is ``False``, conversion
between the formats should ideally be an ``O(1)`` operation, but it means that
no memory copying should be involved, only views of the existing data.
>>> def __ua_convert__(dispatchables, coerce):
... for d in dispatchables:
... if d.type is int:
... if coerce and d.coercible:
... yield str(d.value)
... else:
... yield d.value
>>> be.__ua_convert__ = __ua_convert__
Now that we have defined the backend, the next thing to do is to call the multimethod.
>>> with ua.set_backend(be):
... overridden_me(1, "2")
('override_me', (1, '2'), {})
Note that the marked type has no effect on the actual type of the passed object.
We can also coerce the type of the input.
>>> with ua.set_backend(be, coerce=True):
... overridden_me(1, "2")
... overridden_me(1.0, "2")
('override_me', ('1', '2'), {})
('override_me', ('1.0', '2'), {})
Another feature is that if you remove ``__ua_convert__``, the arguments are not
converted at all and it's up to the backend to handle that.
>>> del be.__ua_convert__
>>> with ua.set_backend(be):
... overridden_me(1, "2")
('override_me', (1, '2'), {})
You also have the option to return ``NotImplemented``, in which case processing moves on
to the next back-end, which in this case, doesn't exist. The same applies to
``__ua_convert__``.
>>> be.__ua_function__ = lambda *a, **kw: NotImplemented
>>> with ua.set_backend(be):
... overridden_me(1, "2")
Traceback (most recent call last):
...
uarray.BackendNotImplementedError: ...
The last possibility is if we don't have ``__ua_convert__``, in which case the job is
left up to ``__ua_function__``, but putting things back into arrays after conversion
will not be possible.
"""
from ._backend import *
__version__ = '0.8.8.dev0+aa94c5a4.scipy'
@@ -0,0 +1,707 @@
import typing
import types
import inspect
import functools
from . import _uarray
import copyreg
import pickle
import contextlib
import threading
from ._uarray import ( # type: ignore
BackendNotImplementedError,
_Function,
_SkipBackendContext,
_SetBackendContext,
_BackendState,
)
__all__ = [
"set_backend",
"set_global_backend",
"skip_backend",
"register_backend",
"determine_backend",
"determine_backend_multi",
"clear_backends",
"create_multimethod",
"generate_multimethod",
"_Function",
"BackendNotImplementedError",
"Dispatchable",
"wrap_single_convertor",
"wrap_single_convertor_instance",
"all_of_type",
"mark_as",
"set_state",
"get_state",
"reset_state",
"_BackendState",
"_SkipBackendContext",
"_SetBackendContext",
]
ArgumentExtractorType = typing.Callable[..., tuple["Dispatchable", ...]]
ArgumentReplacerType = typing.Callable[
[tuple, dict, tuple], tuple[tuple, dict]
]
def unpickle_function(mod_name, qname, self_):
import importlib
try:
module = importlib.import_module(mod_name)
qname = qname.split(".")
func = module
for q in qname:
func = getattr(func, q)
if self_ is not None:
func = types.MethodType(func, self_)
return func
except (ImportError, AttributeError) as e:
from pickle import UnpicklingError
raise UnpicklingError from e
def pickle_function(func):
mod_name = getattr(func, "__module__", None)
qname = getattr(func, "__qualname__", None)
self_ = getattr(func, "__self__", None)
try:
test = unpickle_function(mod_name, qname, self_)
except pickle.UnpicklingError:
test = None
if test is not func:
raise pickle.PicklingError(
f"Can't pickle {func}: it's not the same object as {test}"
)
return unpickle_function, (mod_name, qname, self_)
def pickle_state(state):
return _uarray._BackendState._unpickle, state._pickle()
def pickle_set_backend_context(ctx):
return _SetBackendContext, ctx._pickle()
def pickle_skip_backend_context(ctx):
return _SkipBackendContext, ctx._pickle()
copyreg.pickle(_Function, pickle_function)
copyreg.pickle(_uarray._BackendState, pickle_state)
copyreg.pickle(_SetBackendContext, pickle_set_backend_context)
copyreg.pickle(_SkipBackendContext, pickle_skip_backend_context)
def get_state():
"""
Returns an opaque object containing the current state of all the backends.
Can be used for synchronization between threads/processes.
See Also
--------
set_state
Sets the state returned by this function.
"""
return _uarray.get_state()
@contextlib.contextmanager
def reset_state():
"""
Returns a context manager that resets all state once exited.
See Also
--------
set_state
Context manager that sets the backend state.
get_state
Gets a state to be set by this context manager.
"""
with set_state(get_state()):
yield
@contextlib.contextmanager
def set_state(state):
"""
A context manager that sets the state of the backends to one returned by :obj:`get_state`.
See Also
--------
get_state
Gets a state to be set by this context manager.
""" # noqa: E501
old_state = get_state()
_uarray.set_state(state)
try:
yield
finally:
_uarray.set_state(old_state, True)
def create_multimethod(*args, **kwargs):
"""
Creates a decorator for generating multimethods.
This function creates a decorator that can be used with an argument
extractor in order to generate a multimethod. Other than for the
argument extractor, all arguments are passed on to
:obj:`generate_multimethod`.
See Also
--------
generate_multimethod
Generates a multimethod.
"""
def wrapper(a):
return generate_multimethod(a, *args, **kwargs)
return wrapper
def generate_multimethod(
argument_extractor: ArgumentExtractorType,
argument_replacer: ArgumentReplacerType,
domain: str,
default: typing.Callable | None = None,
):
"""
Generates a multimethod.
Parameters
----------
argument_extractor : ArgumentExtractorType
A callable which extracts the dispatchable arguments. Extracted arguments
should be marked by the :obj:`Dispatchable` class. It has the same signature
as the desired multimethod.
argument_replacer : ArgumentReplacerType
A callable with the signature (args, kwargs, dispatchables), which should also
return an (args, kwargs) pair with the dispatchables replaced inside the
args/kwargs.
domain : str
A string value indicating the domain of this multimethod.
default: Optional[Callable], optional
The default implementation of this multimethod, where ``None`` (the default)
specifies there is no default implementation.
Examples
--------
In this example, ``a`` is to be dispatched over, so we return it, while marking it
as an ``int``.
The trailing comma is needed because the args have to be returned as an iterable.
>>> def override_me(a, b):
... return Dispatchable(a, int),
Next, we define the argument replacer that replaces the dispatchables inside
args/kwargs with the supplied ones.
>>> def override_replacer(args, kwargs, dispatchables):
... return (dispatchables[0], args[1]), {}
Next, we define the multimethod.
>>> overridden_me = generate_multimethod(
... override_me, override_replacer, "ua_examples"
... )
Notice that there's no default implementation, unless you supply one.
>>> overridden_me(1, "a")
Traceback (most recent call last):
...
uarray.BackendNotImplementedError: ...
>>> overridden_me2 = generate_multimethod(
... override_me, override_replacer, "ua_examples", default=lambda x, y: (x, y)
... )
>>> overridden_me2(1, "a")
(1, 'a')
See Also
--------
uarray
See the module documentation for how to override the method by creating
backends.
"""
kw_defaults, arg_defaults, opts = get_defaults(argument_extractor)
ua_func = _Function(
argument_extractor,
argument_replacer,
domain,
arg_defaults,
kw_defaults,
default,
)
return functools.update_wrapper(ua_func, argument_extractor)
def set_backend(backend, coerce=False, only=False):
"""
A context manager that sets the preferred backend.
Parameters
----------
backend
The backend to set.
coerce
Whether or not to coerce to a specific backend's types. Implies ``only``.
only
Whether or not this should be the last backend to try.
See Also
--------
skip_backend: A context manager that allows skipping of backends.
set_global_backend: Set a single, global backend for a domain.
"""
tid = threading.get_native_id()
try:
return backend.__ua_cache__[tid, "set", coerce, only]
except AttributeError:
backend.__ua_cache__ = {}
except KeyError:
pass
ctx = _SetBackendContext(backend, coerce, only)
backend.__ua_cache__[tid, "set", coerce, only] = ctx
return ctx
def skip_backend(backend):
"""
A context manager that allows one to skip a given backend from processing
entirely. This allows one to use another backend's code in a library that
is also a consumer of the same backend.
Parameters
----------
backend
The backend to skip.
See Also
--------
set_backend: A context manager that allows setting of backends.
set_global_backend: Set a single, global backend for a domain.
"""
tid = threading.get_native_id()
try:
return backend.__ua_cache__[tid, "skip"]
except AttributeError:
backend.__ua_cache__ = {}
except KeyError:
pass
ctx = _SkipBackendContext(backend)
backend.__ua_cache__[tid, "skip"] = ctx
return ctx
def get_defaults(f):
sig = inspect.signature(f)
kw_defaults = {}
arg_defaults = []
opts = set()
for k, v in sig.parameters.items():
if v.default is not inspect.Parameter.empty:
kw_defaults[k] = v.default
if v.kind in (
inspect.Parameter.POSITIONAL_ONLY,
inspect.Parameter.POSITIONAL_OR_KEYWORD,
):
arg_defaults.append(v.default)
opts.add(k)
return kw_defaults, tuple(arg_defaults), opts
def set_global_backend(backend, coerce=False, only=False, *, try_last=False):
"""
This utility method replaces the default backend for permanent use. It
will be tried in the list of backends automatically, unless the
``only`` flag is set on a backend. This will be the first tried
backend outside the :obj:`set_backend` context manager.
Note that this method is not thread-safe.
.. warning::
We caution library authors against using this function in
their code. We do *not* support this use-case. This function
is meant to be used only by users themselves, or by a reference
implementation, if one exists.
Parameters
----------
backend
The backend to register.
coerce : bool
Whether to coerce input types when trying this backend.
only : bool
If ``True``, no more backends will be tried if this fails.
Implied by ``coerce=True``.
try_last : bool
If ``True``, the global backend is tried after registered backends.
See Also
--------
set_backend: A context manager that allows setting of backends.
skip_backend: A context manager that allows skipping of backends.
"""
_uarray.set_global_backend(backend, coerce, only, try_last)
def register_backend(backend):
"""
This utility method sets registers backend for permanent use. It
will be tried in the list of backends automatically, unless the
``only`` flag is set on a backend.
Note that this method is not thread-safe.
Parameters
----------
backend
The backend to register.
"""
_uarray.register_backend(backend)
def clear_backends(domain, registered=True, globals=False):
"""
This utility method clears registered backends.
.. warning::
We caution library authors against using this function in
their code. We do *not* support this use-case. This function
is meant to be used only by users themselves.
.. warning::
Do NOT use this method inside a multimethod call, or the
program is likely to crash.
Parameters
----------
domain : Optional[str]
The domain for which to de-register backends. ``None`` means
de-register for all domains.
registered : bool
Whether or not to clear registered backends. See :obj:`register_backend`.
globals : bool
Whether or not to clear global backends. See :obj:`set_global_backend`.
See Also
--------
register_backend : Register a backend globally.
set_global_backend : Set a global backend.
"""
_uarray.clear_backends(domain, registered, globals)
class Dispatchable:
"""
A utility class which marks an argument with a specific dispatch type.
Attributes
----------
value
The value of the Dispatchable.
type
The type of the Dispatchable.
Examples
--------
>>> x = Dispatchable(1, str)
>>> x
<Dispatchable: type=<class 'str'>, value=1>
See Also
--------
all_of_type
Marks all unmarked parameters of a function.
mark_as
Allows one to create a utility function to mark as a given type.
"""
def __init__(self, value, dispatch_type, coercible=True):
self.value = value
self.type = dispatch_type
self.coercible = coercible
def __getitem__(self, index):
return (self.type, self.value)[index]
def __str__(self):
return f"<{type(self).__name__}: type={self.type!r}, value={self.value!r}>"
__repr__ = __str__
def mark_as(dispatch_type):
"""
Creates a utility function to mark something as a specific type.
Examples
--------
>>> mark_int = mark_as(int)
>>> mark_int(1)
<Dispatchable: type=<class 'int'>, value=1>
"""
return functools.partial(Dispatchable, dispatch_type=dispatch_type)
def all_of_type(arg_type):
"""
Marks all unmarked arguments as a given type.
Examples
--------
>>> @all_of_type(str)
... def f(a, b):
... return a, Dispatchable(b, int)
>>> f('a', 1)
(<Dispatchable: type=<class 'str'>, value='a'>,
<Dispatchable: type=<class 'int'>, value=1>)
"""
def outer(func):
@functools.wraps(func)
def inner(*args, **kwargs):
extracted_args = func(*args, **kwargs)
return tuple(
Dispatchable(arg, arg_type)
if not isinstance(arg, Dispatchable)
else arg
for arg in extracted_args
)
return inner
return outer
def wrap_single_convertor(convert_single):
"""
Wraps a ``__ua_convert__`` defined for a single element to all elements.
If any of them return ``NotImplemented``, the operation is assumed to be
undefined.
Accepts a signature of (value, type, coerce).
"""
@functools.wraps(convert_single)
def __ua_convert__(dispatchables, coerce):
converted = []
for d in dispatchables:
c = convert_single(d.value, d.type, coerce and d.coercible)
if c is NotImplemented:
return NotImplemented
converted.append(c)
return converted
return __ua_convert__
def wrap_single_convertor_instance(convert_single):
"""
Wraps a ``__ua_convert__`` defined for a single element to all elements.
If any of them return ``NotImplemented``, the operation is assumed to be
undefined.
Accepts a signature of (value, type, coerce).
"""
@functools.wraps(convert_single)
def __ua_convert__(self, dispatchables, coerce):
converted = []
for d in dispatchables:
c = convert_single(self, d.value, d.type, coerce and d.coercible)
if c is NotImplemented:
return NotImplemented
converted.append(c)
return converted
return __ua_convert__
def determine_backend(value, dispatch_type, *, domain, only=True, coerce=False):
"""Set the backend to the first active backend that supports ``value``
This is useful for functions that call multimethods without any dispatchable
arguments. You can use :func:`determine_backend` to ensure the same backend
is used everywhere in a block of multimethod calls.
Parameters
----------
value
The value being tested
dispatch_type
The dispatch type associated with ``value``, aka
":ref:`marking <MarkingGlossary>`".
domain: string
The domain to query for backends and set.
coerce: bool
Whether or not to allow coercion to the backend's types. Implies ``only``.
only: bool
Whether or not this should be the last backend to try.
See Also
--------
set_backend: For when you know which backend to set
Notes
-----
Support is determined by the ``__ua_convert__`` protocol. Backends not
supporting the type must return ``NotImplemented`` from their
``__ua_convert__`` if they don't support input of that type.
Examples
--------
Suppose we have two backends ``BackendA`` and ``BackendB`` each supporting
different types, ``TypeA`` and ``TypeB``. Neither supporting the other type:
>>> with ua.set_backend(ex.BackendA):
... ex.call_multimethod(ex.TypeB(), ex.TypeB())
Traceback (most recent call last):
...
uarray.BackendNotImplementedError: ...
Now consider a multimethod that creates a new object of ``TypeA``, or
``TypeB`` depending on the active backend.
>>> with ua.set_backend(ex.BackendA), ua.set_backend(ex.BackendB):
... res = ex.creation_multimethod()
... ex.call_multimethod(res, ex.TypeA())
Traceback (most recent call last):
...
uarray.BackendNotImplementedError: ...
``res`` is an object of ``TypeB`` because ``BackendB`` is set in the
innermost with statement. So, ``call_multimethod`` fails since the types
don't match.
Instead, we need to first find a backend suitable for all of our objects.
>>> with ua.set_backend(ex.BackendA), ua.set_backend(ex.BackendB):
... x = ex.TypeA()
... with ua.determine_backend(x, "mark", domain="ua_examples"):
... res = ex.creation_multimethod()
... ex.call_multimethod(res, x)
TypeA
"""
dispatchables = (Dispatchable(value, dispatch_type, coerce),)
backend = _uarray.determine_backend(domain, dispatchables, coerce)
return set_backend(backend, coerce=coerce, only=only)
def determine_backend_multi(
dispatchables, *, domain, only=True, coerce=False, **kwargs
):
"""Set a backend supporting all ``dispatchables``
This is useful for functions that call multimethods without any dispatchable
arguments. You can use :func:`determine_backend_multi` to ensure the same
backend is used everywhere in a block of multimethod calls involving
multiple arrays.
Parameters
----------
dispatchables: Sequence[Union[uarray.Dispatchable, Any]]
The dispatchables that must be supported
domain: string
The domain to query for backends and set.
coerce: bool
Whether or not to allow coercion to the backend's types. Implies ``only``.
only: bool
Whether or not this should be the last backend to try.
dispatch_type: Optional[Any]
The default dispatch type associated with ``dispatchables``, aka
":ref:`marking <MarkingGlossary>`".
See Also
--------
determine_backend: For a single dispatch value
set_backend: For when you know which backend to set
Notes
-----
Support is determined by the ``__ua_convert__`` protocol. Backends not
supporting the type must return ``NotImplemented`` from their
``__ua_convert__`` if they don't support input of that type.
Examples
--------
:func:`determine_backend` allows the backend to be set from a single
object. :func:`determine_backend_multi` allows multiple objects to be
checked simultaneously for support in the backend. Suppose we have a
``BackendAB`` which supports ``TypeA`` and ``TypeB`` in the same call,
and a ``BackendBC`` that doesn't support ``TypeA``.
>>> with ua.set_backend(ex.BackendAB), ua.set_backend(ex.BackendBC):
... a, b = ex.TypeA(), ex.TypeB()
... with ua.determine_backend_multi(
... [ua.Dispatchable(a, "mark"), ua.Dispatchable(b, "mark")],
... domain="ua_examples"
... ):
... res = ex.creation_multimethod()
... ex.call_multimethod(res, a, b)
TypeA
This won't call ``BackendBC`` because it doesn't support ``TypeA``.
We can also use leave out the ``ua.Dispatchable`` if we specify the
default ``dispatch_type`` for the ``dispatchables`` argument.
>>> with ua.set_backend(ex.BackendAB), ua.set_backend(ex.BackendBC):
... a, b = ex.TypeA(), ex.TypeB()
... with ua.determine_backend_multi(
... [a, b], dispatch_type="mark", domain="ua_examples"
... ):
... res = ex.creation_multimethod()
... ex.call_multimethod(res, a, b)
TypeA
"""
if "dispatch_type" in kwargs:
disp_type = kwargs.pop("dispatch_type")
dispatchables = tuple(
d if isinstance(d, Dispatchable) else Dispatchable(d, disp_type)
for d in dispatchables
)
else:
dispatchables = tuple(dispatchables)
if not all(isinstance(d, Dispatchable) for d in dispatchables):
raise TypeError("dispatchables must be instances of uarray.Dispatchable")
if len(kwargs) != 0:
raise TypeError(f"Received unexpected keyword arguments: {kwargs}")
backend = _uarray.determine_backend(domain, dispatchables, coerce)
return set_backend(backend, coerce=coerce, only=only)
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,22 @@
"""
NumPy Array API compatibility library
This is a small wrapper around NumPy, CuPy, JAX, sparse and others that are
compatible with the Array API standard https://data-apis.org/array-api/latest/.
See also NEP 47 https://numpy.org/neps/nep-0047-array-api-standard.html.
Unlike array_api_strict, this is not a strict minimal implementation of the
Array API, but rather just an extension of the main NumPy namespace with
changes needed to be compliant with the Array API. See
https://numpy.org/doc/stable/reference/array_api.html for a full list of
changes. In particular, unlike array_api_strict, this package does not use a
separate Array object, but rather just uses numpy.ndarray directly.
Library authors using the Array API may wish to test against array_api_strict
to ensure they are not using functionality outside of the standard, but prefer
this implementation for the default when working with NumPy arrays.
"""
__version__ = '1.12.0'
from .common import * # noqa: F401, F403
@@ -0,0 +1,59 @@
"""
Internal helpers
"""
from collections.abc import Callable
from functools import wraps
from inspect import signature
from types import ModuleType
from typing import TypeVar
_T = TypeVar("_T")
def get_xp(xp: ModuleType) -> Callable[[Callable[..., _T]], Callable[..., _T]]:
"""
Decorator to automatically replace xp with the corresponding array module.
Use like
import numpy as np
@get_xp(np)
def func(x, /, xp, kwarg=None):
return xp.func(x, kwarg=kwarg)
Note that xp must be a keyword argument and come after all non-keyword
arguments.
"""
def inner(f: Callable[..., _T], /) -> Callable[..., _T]:
@wraps(f)
def wrapped_f(*args: object, **kwargs: object) -> object:
return f(*args, xp=xp, **kwargs)
sig = signature(f)
new_sig = sig.replace(
parameters=[par for i, par in sig.parameters.items() if i != "xp"]
)
if wrapped_f.__doc__ is None:
wrapped_f.__doc__ = f"""\
Array API compatibility wrapper for {f.__name__}.
See the corresponding documentation in NumPy/CuPy and/or the array API
specification for more details.
"""
wrapped_f.__signature__ = new_sig # pyright: ignore[reportAttributeAccessIssue]
return wrapped_f # pyright: ignore[reportReturnType]
return inner
__all__ = ["get_xp"]
def __dir__() -> list[str]:
return __all__
@@ -0,0 +1 @@
from ._helpers import * # noqa: F403
@@ -0,0 +1,727 @@
"""
These are functions that are just aliases of existing functions in NumPy.
"""
from __future__ import annotations
import inspect
from typing import TYPE_CHECKING, Any, NamedTuple, Optional, Sequence, cast
from ._helpers import _check_device, array_namespace
from ._helpers import device as _get_device
from ._helpers import is_cupy_namespace as _is_cupy_namespace
from ._typing import Array, Device, DType, Namespace
if TYPE_CHECKING:
# TODO: import from typing (requires Python >=3.13)
from typing_extensions import TypeIs
# These functions are modified from the NumPy versions.
# Creation functions add the device keyword (which does nothing for NumPy and Dask)
def arange(
start: float,
/,
stop: float | None = None,
step: float = 1,
*,
xp: Namespace,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.arange(start, stop=stop, step=step, dtype=dtype, **kwargs)
def empty(
shape: int | tuple[int, ...],
xp: Namespace,
*,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.empty(shape, dtype=dtype, **kwargs)
def empty_like(
x: Array,
/,
xp: Namespace,
*,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.empty_like(x, dtype=dtype, **kwargs)
def eye(
n_rows: int,
n_cols: int | None = None,
/,
*,
xp: Namespace,
k: int = 0,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.eye(n_rows, M=n_cols, k=k, dtype=dtype, **kwargs)
def full(
shape: int | tuple[int, ...],
fill_value: complex,
xp: Namespace,
*,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.full(shape, fill_value, dtype=dtype, **kwargs)
def full_like(
x: Array,
/,
fill_value: complex,
*,
xp: Namespace,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.full_like(x, fill_value, dtype=dtype, **kwargs)
def linspace(
start: float,
stop: float,
/,
num: int,
*,
xp: Namespace,
dtype: DType | None = None,
device: Device | None = None,
endpoint: bool = True,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.linspace(start, stop, num, dtype=dtype, endpoint=endpoint, **kwargs)
def ones(
shape: int | tuple[int, ...],
xp: Namespace,
*,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.ones(shape, dtype=dtype, **kwargs)
def ones_like(
x: Array,
/,
xp: Namespace,
*,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.ones_like(x, dtype=dtype, **kwargs)
def zeros(
shape: int | tuple[int, ...],
xp: Namespace,
*,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.zeros(shape, dtype=dtype, **kwargs)
def zeros_like(
x: Array,
/,
xp: Namespace,
*,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
_check_device(xp, device)
return xp.zeros_like(x, dtype=dtype, **kwargs)
# np.unique() is split into four functions in the array API:
# unique_all, unique_counts, unique_inverse, and unique_values (this is done
# to remove polymorphic return types).
# The functions here return namedtuples (np.unique() returns a normal
# tuple).
# Note that these named tuples aren't actually part of the standard namespace,
# but I don't see any issue with exporting the names here regardless.
class UniqueAllResult(NamedTuple):
values: Array
indices: Array
inverse_indices: Array
counts: Array
class UniqueCountsResult(NamedTuple):
values: Array
counts: Array
class UniqueInverseResult(NamedTuple):
values: Array
inverse_indices: Array
def _unique_kwargs(xp: Namespace) -> dict[str, bool]:
# Older versions of NumPy and CuPy do not have equal_nan. Rather than
# trying to parse version numbers, just check if equal_nan is in the
# signature.
s = inspect.signature(xp.unique)
if "equal_nan" in s.parameters:
return {"equal_nan": False}
return {}
def unique_all(x: Array, /, xp: Namespace) -> UniqueAllResult:
kwargs = _unique_kwargs(xp)
values, indices, inverse_indices, counts = xp.unique(
x,
return_counts=True,
return_index=True,
return_inverse=True,
**kwargs,
)
# np.unique() flattens inverse indices, but they need to share x's shape
# See https://github.com/numpy/numpy/issues/20638
inverse_indices = inverse_indices.reshape(x.shape)
return UniqueAllResult(
values,
indices,
inverse_indices,
counts,
)
def unique_counts(x: Array, /, xp: Namespace) -> UniqueCountsResult:
kwargs = _unique_kwargs(xp)
res = xp.unique(
x, return_counts=True, return_index=False, return_inverse=False, **kwargs
)
return UniqueCountsResult(*res)
def unique_inverse(x: Array, /, xp: Namespace) -> UniqueInverseResult:
kwargs = _unique_kwargs(xp)
values, inverse_indices = xp.unique(
x,
return_counts=False,
return_index=False,
return_inverse=True,
**kwargs,
)
# xp.unique() flattens inverse indices, but they need to share x's shape
# See https://github.com/numpy/numpy/issues/20638
inverse_indices = inverse_indices.reshape(x.shape)
return UniqueInverseResult(values, inverse_indices)
def unique_values(x: Array, /, xp: Namespace) -> Array:
kwargs = _unique_kwargs(xp)
return xp.unique(
x,
return_counts=False,
return_index=False,
return_inverse=False,
**kwargs,
)
# These functions have different keyword argument names
def std(
x: Array,
/,
xp: Namespace,
*,
axis: int | tuple[int, ...] | None = None,
correction: float = 0.0, # correction instead of ddof
keepdims: bool = False,
**kwargs: object,
) -> Array:
return xp.std(x, axis=axis, ddof=correction, keepdims=keepdims, **kwargs)
def var(
x: Array,
/,
xp: Namespace,
*,
axis: int | tuple[int, ...] | None = None,
correction: float = 0.0, # correction instead of ddof
keepdims: bool = False,
**kwargs: object,
) -> Array:
return xp.var(x, axis=axis, ddof=correction, keepdims=keepdims, **kwargs)
# cumulative_sum is renamed from cumsum, and adds the include_initial keyword
# argument
def cumulative_sum(
x: Array,
/,
xp: Namespace,
*,
axis: int | None = None,
dtype: DType | None = None,
include_initial: bool = False,
**kwargs: object,
) -> Array:
wrapped_xp = array_namespace(x)
# TODO: The standard is not clear about what should happen when x.ndim == 0.
if axis is None:
if x.ndim > 1:
raise ValueError(
"axis must be specified in cumulative_sum for more than one dimension"
)
axis = 0
res = xp.cumsum(x, axis=axis, dtype=dtype, **kwargs)
# np.cumsum does not support include_initial
if include_initial:
initial_shape = list(x.shape)
initial_shape[axis] = 1
res = xp.concatenate(
[
wrapped_xp.zeros(
shape=initial_shape, dtype=res.dtype, device=_get_device(res)
),
res,
],
axis=axis,
)
return res
def cumulative_prod(
x: Array,
/,
xp: Namespace,
*,
axis: int | None = None,
dtype: DType | None = None,
include_initial: bool = False,
**kwargs: object,
) -> Array:
wrapped_xp = array_namespace(x)
if axis is None:
if x.ndim > 1:
raise ValueError(
"axis must be specified in cumulative_prod for more than one dimension"
)
axis = 0
res = xp.cumprod(x, axis=axis, dtype=dtype, **kwargs)
# np.cumprod does not support include_initial
if include_initial:
initial_shape = list(x.shape)
initial_shape[axis] = 1
res = xp.concatenate(
[
wrapped_xp.ones(
shape=initial_shape, dtype=res.dtype, device=_get_device(res)
),
res,
],
axis=axis,
)
return res
# The min and max argument names in clip are different and not optional in numpy, and type
# promotion behavior is different.
def clip(
x: Array,
/,
min: float | Array | None = None,
max: float | Array | None = None,
*,
xp: Namespace,
# TODO: np.clip has other ufunc kwargs
out: Array | None = None,
) -> Array:
def _isscalar(a: object) -> TypeIs[int | float | None]:
return isinstance(a, (int, float, type(None)))
min_shape = () if _isscalar(min) else min.shape
max_shape = () if _isscalar(max) else max.shape
wrapped_xp = array_namespace(x)
result_shape = xp.broadcast_shapes(x.shape, min_shape, max_shape)
# np.clip does type promotion but the array API clip requires that the
# output have the same dtype as x. We do this instead of just downcasting
# the result of xp.clip() to handle some corner cases better (e.g.,
# avoiding uint64 -> float64 promotion).
# Note: cases where min or max overflow (integer) or round (float) in the
# wrong direction when downcasting to x.dtype are unspecified. This code
# just does whatever NumPy does when it downcasts in the assignment, but
# other behavior could be preferred, especially for integers. For example,
# this code produces:
# >>> clip(asarray(0, dtype=int8), asarray(128, dtype=int16), None)
# -128
# but an answer of 0 might be preferred. See
# https://github.com/numpy/numpy/issues/24976 for more discussion on this issue.
# At least handle the case of Python integers correctly (see
# https://github.com/numpy/numpy/pull/26892).
if wrapped_xp.isdtype(x.dtype, "integral"):
if type(min) is int and min <= wrapped_xp.iinfo(x.dtype).min:
min = None
if type(max) is int and max >= wrapped_xp.iinfo(x.dtype).max:
max = None
dev = _get_device(x)
if out is None:
out = wrapped_xp.empty(result_shape, dtype=x.dtype, device=dev)
assert out is not None # workaround for a type-narrowing issue in pyright
out[()] = x
if min is not None:
a = wrapped_xp.asarray(min, dtype=x.dtype, device=dev)
a = xp.broadcast_to(a, result_shape)
ia = (out < a) | xp.isnan(a)
out[ia] = a[ia]
if max is not None:
b = wrapped_xp.asarray(max, dtype=x.dtype, device=dev)
b = xp.broadcast_to(b, result_shape)
ib = (out > b) | xp.isnan(b)
out[ib] = b[ib]
# Return a scalar for 0-D
return out[()]
# Unlike transpose(), the axes argument to permute_dims() is required.
def permute_dims(x: Array, /, axes: tuple[int, ...], xp: Namespace) -> Array:
return xp.transpose(x, axes)
# np.reshape calls the keyword argument 'newshape' instead of 'shape'
def reshape(
x: Array,
/,
shape: tuple[int, ...],
xp: Namespace,
*,
copy: Optional[bool] = None,
**kwargs: object,
) -> Array:
if copy is True:
x = x.copy()
elif copy is False:
y = x.view()
y.shape = shape
return y
return xp.reshape(x, shape, **kwargs)
# The descending keyword is new in sort and argsort, and 'kind' replaced with
# 'stable'
def argsort(
x: Array,
/,
xp: Namespace,
*,
axis: int = -1,
descending: bool = False,
stable: bool = True,
**kwargs: object,
) -> Array:
# Note: this keyword argument is different, and the default is different.
# We set it in kwargs like this because numpy.sort uses kind='quicksort'
# as the default whereas cupy.sort uses kind=None.
if stable:
kwargs["kind"] = "stable"
if not descending:
res = xp.argsort(x, axis=axis, **kwargs)
else:
# As NumPy has no native descending sort, we imitate it here. Note that
# simply flipping the results of xp.argsort(x, ...) would not
# respect the relative order like it would in native descending sorts.
res = xp.flip(
xp.argsort(xp.flip(x, axis=axis), axis=axis, **kwargs),
axis=axis,
)
# Rely on flip()/argsort() to validate axis
normalised_axis = axis if axis >= 0 else x.ndim + axis
max_i = x.shape[normalised_axis] - 1
res = max_i - res
return res
def sort(
x: Array,
/,
xp: Namespace,
*,
axis: int = -1,
descending: bool = False,
stable: bool = True,
**kwargs: object,
) -> Array:
# Note: this keyword argument is different, and the default is different.
# We set it in kwargs like this because numpy.sort uses kind='quicksort'
# as the default whereas cupy.sort uses kind=None.
if stable:
kwargs["kind"] = "stable"
res = xp.sort(x, axis=axis, **kwargs)
if descending:
res = xp.flip(res, axis=axis)
return res
# nonzero should error for zero-dimensional arrays
def nonzero(x: Array, /, xp: Namespace, **kwargs: object) -> tuple[Array, ...]:
if x.ndim == 0:
raise ValueError("nonzero() does not support zero-dimensional arrays")
return xp.nonzero(x, **kwargs)
# ceil, floor, and trunc return integers for integer inputs
def ceil(x: Array, /, xp: Namespace, **kwargs: object) -> Array:
if xp.issubdtype(x.dtype, xp.integer):
return x
return xp.ceil(x, **kwargs)
def floor(x: Array, /, xp: Namespace, **kwargs: object) -> Array:
if xp.issubdtype(x.dtype, xp.integer):
return x
return xp.floor(x, **kwargs)
def trunc(x: Array, /, xp: Namespace, **kwargs: object) -> Array:
if xp.issubdtype(x.dtype, xp.integer):
return x
return xp.trunc(x, **kwargs)
# linear algebra functions
def matmul(x1: Array, x2: Array, /, xp: Namespace, **kwargs: object) -> Array:
return xp.matmul(x1, x2, **kwargs)
# Unlike transpose, matrix_transpose only transposes the last two axes.
def matrix_transpose(x: Array, /, xp: Namespace) -> Array:
if x.ndim < 2:
raise ValueError("x must be at least 2-dimensional for matrix_transpose")
return xp.swapaxes(x, -1, -2)
def tensordot(
x1: Array,
x2: Array,
/,
xp: Namespace,
*,
axes: int | tuple[Sequence[int], Sequence[int]] = 2,
**kwargs: object,
) -> Array:
return xp.tensordot(x1, x2, axes=axes, **kwargs)
def vecdot(x1: Array, x2: Array, /, xp: Namespace, *, axis: int = -1) -> Array:
if x1.shape[axis] != x2.shape[axis]:
raise ValueError("x1 and x2 must have the same size along the given axis")
if hasattr(xp, "broadcast_tensors"):
_broadcast = xp.broadcast_tensors
else:
_broadcast = xp.broadcast_arrays
x1_ = xp.moveaxis(x1, axis, -1)
x2_ = xp.moveaxis(x2, axis, -1)
x1_, x2_ = _broadcast(x1_, x2_)
res = xp.conj(x1_[..., None, :]) @ x2_[..., None]
return res[..., 0, 0]
# isdtype is a new function in the 2022.12 array API specification.
def isdtype(
dtype: DType,
kind: DType | str | tuple[DType | str, ...],
xp: Namespace,
*,
_tuple: bool = True, # Disallow nested tuples
) -> bool:
"""
Returns a boolean indicating whether a provided dtype is of a specified data type ``kind``.
Note that outside of this function, this compat library does not yet fully
support complex numbers.
See
https://data-apis.org/array-api/latest/API_specification/generated/array_api.isdtype.html
for more details
"""
if isinstance(kind, tuple) and _tuple:
return any(
isdtype(dtype, k, xp, _tuple=False)
for k in cast("tuple[DType | str, ...]", kind)
)
elif isinstance(kind, str):
if kind == "bool":
return dtype == xp.bool_
elif kind == "signed integer":
return xp.issubdtype(dtype, xp.signedinteger)
elif kind == "unsigned integer":
return xp.issubdtype(dtype, xp.unsignedinteger)
elif kind == "integral":
return xp.issubdtype(dtype, xp.integer)
elif kind == "real floating":
return xp.issubdtype(dtype, xp.floating)
elif kind == "complex floating":
return xp.issubdtype(dtype, xp.complexfloating)
elif kind == "numeric":
return xp.issubdtype(dtype, xp.number)
else:
raise ValueError(f"Unrecognized data type kind: {kind!r}")
else:
# This will allow things that aren't required by the spec, like
# isdtype(np.float64, float) or isdtype(np.int64, 'l'). Should we be
# more strict here to match the type annotation? Note that the
# array_api_strict implementation will be very strict.
return dtype == kind
# unstack is a new function in the 2023.12 array API standard
def unstack(x: Array, /, xp: Namespace, *, axis: int = 0) -> tuple[Array, ...]:
if x.ndim == 0:
raise ValueError("Input array must be at least 1-d.")
return tuple(xp.moveaxis(x, axis, 0))
# numpy 1.26 does not use the standard definition for sign on complex numbers
def sign(x: Array, /, xp: Namespace, **kwargs: object) -> Array:
if isdtype(x.dtype, "complex floating", xp=xp):
out = (x / xp.abs(x, **kwargs))[...]
# sign(0) = 0 but the above formula would give nan
out[x == 0j] = 0j
else:
out = xp.sign(x, **kwargs)
# CuPy sign() does not propagate nans. See
# https://github.com/data-apis/array-api-compat/issues/136
if _is_cupy_namespace(xp) and isdtype(x.dtype, "real floating", xp=xp):
out[xp.isnan(x)] = xp.nan
return out[()]
def finfo(type_: DType | Array, /, xp: Namespace) -> Any:
# It is surprisingly difficult to recognize a dtype apart from an array.
# np.int64 is not the same as np.asarray(1).dtype!
try:
return xp.finfo(type_)
except (ValueError, TypeError):
return xp.finfo(type_.dtype)
def iinfo(type_: DType | Array, /, xp: Namespace) -> Any:
try:
return xp.iinfo(type_)
except (ValueError, TypeError):
return xp.iinfo(type_.dtype)
__all__ = [
"arange",
"empty",
"empty_like",
"eye",
"full",
"full_like",
"linspace",
"ones",
"ones_like",
"zeros",
"zeros_like",
"UniqueAllResult",
"UniqueCountsResult",
"UniqueInverseResult",
"unique_all",
"unique_counts",
"unique_inverse",
"unique_values",
"std",
"var",
"cumulative_sum",
"cumulative_prod",
"clip",
"permute_dims",
"reshape",
"argsort",
"sort",
"nonzero",
"ceil",
"floor",
"trunc",
"matmul",
"matrix_transpose",
"tensordot",
"vecdot",
"isdtype",
"unstack",
"sign",
"finfo",
"iinfo",
]
_all_ignore = ["inspect", "array_namespace", "NamedTuple"]
def __dir__() -> list[str]:
return __all__
@@ -0,0 +1,213 @@
from __future__ import annotations
from collections.abc import Sequence
from typing import Literal, TypeAlias
from ._typing import Array, Device, DType, Namespace
_Norm: TypeAlias = Literal["backward", "ortho", "forward"]
# Note: NumPy fft functions improperly upcast float32 and complex64 to
# complex128, which is why we require wrapping them all here.
def fft(
x: Array,
/,
xp: Namespace,
*,
n: int | None = None,
axis: int = -1,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.fft(x, n=n, axis=axis, norm=norm)
if x.dtype in [xp.float32, xp.complex64]:
return res.astype(xp.complex64)
return res
def ifft(
x: Array,
/,
xp: Namespace,
*,
n: int | None = None,
axis: int = -1,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.ifft(x, n=n, axis=axis, norm=norm)
if x.dtype in [xp.float32, xp.complex64]:
return res.astype(xp.complex64)
return res
def fftn(
x: Array,
/,
xp: Namespace,
*,
s: Sequence[int] | None = None,
axes: Sequence[int] | None = None,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.fftn(x, s=s, axes=axes, norm=norm)
if x.dtype in [xp.float32, xp.complex64]:
return res.astype(xp.complex64)
return res
def ifftn(
x: Array,
/,
xp: Namespace,
*,
s: Sequence[int] | None = None,
axes: Sequence[int] | None = None,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.ifftn(x, s=s, axes=axes, norm=norm)
if x.dtype in [xp.float32, xp.complex64]:
return res.astype(xp.complex64)
return res
def rfft(
x: Array,
/,
xp: Namespace,
*,
n: int | None = None,
axis: int = -1,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.rfft(x, n=n, axis=axis, norm=norm)
if x.dtype == xp.float32:
return res.astype(xp.complex64)
return res
def irfft(
x: Array,
/,
xp: Namespace,
*,
n: int | None = None,
axis: int = -1,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.irfft(x, n=n, axis=axis, norm=norm)
if x.dtype == xp.complex64:
return res.astype(xp.float32)
return res
def rfftn(
x: Array,
/,
xp: Namespace,
*,
s: Sequence[int] | None = None,
axes: Sequence[int] | None = None,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.rfftn(x, s=s, axes=axes, norm=norm)
if x.dtype == xp.float32:
return res.astype(xp.complex64)
return res
def irfftn(
x: Array,
/,
xp: Namespace,
*,
s: Sequence[int] | None = None,
axes: Sequence[int] | None = None,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.irfftn(x, s=s, axes=axes, norm=norm)
if x.dtype == xp.complex64:
return res.astype(xp.float32)
return res
def hfft(
x: Array,
/,
xp: Namespace,
*,
n: int | None = None,
axis: int = -1,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.hfft(x, n=n, axis=axis, norm=norm)
if x.dtype in [xp.float32, xp.complex64]:
return res.astype(xp.float32)
return res
def ihfft(
x: Array,
/,
xp: Namespace,
*,
n: int | None = None,
axis: int = -1,
norm: _Norm = "backward",
) -> Array:
res = xp.fft.ihfft(x, n=n, axis=axis, norm=norm)
if x.dtype in [xp.float32, xp.complex64]:
return res.astype(xp.complex64)
return res
def fftfreq(
n: int,
/,
xp: Namespace,
*,
d: float = 1.0,
dtype: DType | None = None,
device: Device | None = None,
) -> Array:
if device not in ["cpu", None]:
raise ValueError(f"Unsupported device {device!r}")
res = xp.fft.fftfreq(n, d=d)
if dtype is not None:
return res.astype(dtype)
return res
def rfftfreq(
n: int,
/,
xp: Namespace,
*,
d: float = 1.0,
dtype: DType | None = None,
device: Device | None = None,
) -> Array:
if device not in ["cpu", None]:
raise ValueError(f"Unsupported device {device!r}")
res = xp.fft.rfftfreq(n, d=d)
if dtype is not None:
return res.astype(dtype)
return res
def fftshift(
x: Array, /, xp: Namespace, *, axes: int | Sequence[int] | None = None
) -> Array:
return xp.fft.fftshift(x, axes=axes)
def ifftshift(
x: Array, /, xp: Namespace, *, axes: int | Sequence[int] | None = None
) -> Array:
return xp.fft.ifftshift(x, axes=axes)
__all__ = [
"fft",
"ifft",
"fftn",
"ifftn",
"rfft",
"irfft",
"rfftn",
"irfftn",
"hfft",
"ihfft",
"fftfreq",
"rfftfreq",
"fftshift",
"ifftshift",
]
def __dir__() -> list[str]:
return __all__
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,232 @@
from __future__ import annotations
import math
from typing import Literal, NamedTuple, cast
import numpy as np
if np.__version__[0] == "2":
from numpy.lib.array_utils import normalize_axis_tuple
else:
from numpy.core.numeric import normalize_axis_tuple
from .._internal import get_xp
from ._aliases import isdtype, matmul, matrix_transpose, tensordot, vecdot
from ._typing import Array, DType, JustFloat, JustInt, Namespace
# These are in the main NumPy namespace but not in numpy.linalg
def cross(
x1: Array,
x2: Array,
/,
xp: Namespace,
*,
axis: int = -1,
**kwargs: object,
) -> Array:
return xp.cross(x1, x2, axis=axis, **kwargs)
def outer(x1: Array, x2: Array, /, xp: Namespace, **kwargs: object) -> Array:
return xp.outer(x1, x2, **kwargs)
class EighResult(NamedTuple):
eigenvalues: Array
eigenvectors: Array
class QRResult(NamedTuple):
Q: Array
R: Array
class SlogdetResult(NamedTuple):
sign: Array
logabsdet: Array
class SVDResult(NamedTuple):
U: Array
S: Array
Vh: Array
# These functions are the same as their NumPy counterparts except they return
# a namedtuple.
def eigh(x: Array, /, xp: Namespace, **kwargs: object) -> EighResult:
return EighResult(*xp.linalg.eigh(x, **kwargs))
def qr(
x: Array,
/,
xp: Namespace,
*,
mode: Literal["reduced", "complete"] = "reduced",
**kwargs: object,
) -> QRResult:
return QRResult(*xp.linalg.qr(x, mode=mode, **kwargs))
def slogdet(x: Array, /, xp: Namespace, **kwargs: object) -> SlogdetResult:
return SlogdetResult(*xp.linalg.slogdet(x, **kwargs))
def svd(
x: Array,
/,
xp: Namespace,
*,
full_matrices: bool = True,
**kwargs: object,
) -> SVDResult:
return SVDResult(*xp.linalg.svd(x, full_matrices=full_matrices, **kwargs))
# These functions have additional keyword arguments
# The upper keyword argument is new from NumPy
def cholesky(
x: Array,
/,
xp: Namespace,
*,
upper: bool = False,
**kwargs: object,
) -> Array:
L = xp.linalg.cholesky(x, **kwargs)
if upper:
U = get_xp(xp)(matrix_transpose)(L)
if get_xp(xp)(isdtype)(U.dtype, 'complex floating'):
U = xp.conj(U) # pyright: ignore[reportConstantRedefinition]
return U
return L
# The rtol keyword argument of matrix_rank() and pinv() is new from NumPy.
# Note that it has a different semantic meaning from tol and rcond.
def matrix_rank(
x: Array,
/,
xp: Namespace,
*,
rtol: float | Array | None = None,
**kwargs: object,
) -> Array:
# this is different from xp.linalg.matrix_rank, which supports 1
# dimensional arrays.
if x.ndim < 2:
raise xp.linalg.LinAlgError("1-dimensional array given. Array must be at least two-dimensional")
S: Array = get_xp(xp)(svdvals)(x, **kwargs)
if rtol is None:
tol = S.max(axis=-1, keepdims=True) * max(x.shape[-2:]) * xp.finfo(S.dtype).eps
else:
# this is different from xp.linalg.matrix_rank, which does not
# multiply the tolerance by the largest singular value.
tol = S.max(axis=-1, keepdims=True)*xp.asarray(rtol)[..., xp.newaxis]
return xp.count_nonzero(S > tol, axis=-1)
def pinv(
x: Array,
/,
xp: Namespace,
*,
rtol: float | Array | None = None,
**kwargs: object,
) -> Array:
# this is different from xp.linalg.pinv, which does not multiply the
# default tolerance by max(M, N).
if rtol is None:
rtol = max(x.shape[-2:]) * xp.finfo(x.dtype).eps
return xp.linalg.pinv(x, rcond=rtol, **kwargs)
# These functions are new in the array API spec
def matrix_norm(
x: Array,
/,
xp: Namespace,
*,
keepdims: bool = False,
ord: Literal[1, 2, -1, -2] | JustFloat | Literal["fro", "nuc"] | None = "fro",
) -> Array:
return xp.linalg.norm(x, axis=(-2, -1), keepdims=keepdims, ord=ord)
# svdvals is not in NumPy (but it is in SciPy). It is equivalent to
# xp.linalg.svd(compute_uv=False).
def svdvals(x: Array, /, xp: Namespace) -> Array | tuple[Array, ...]:
return xp.linalg.svd(x, compute_uv=False)
def vector_norm(
x: Array,
/,
xp: Namespace,
*,
axis: int | tuple[int, ...] | None = None,
keepdims: bool = False,
ord: JustInt | JustFloat = 2,
) -> Array:
# xp.linalg.norm tries to do a matrix norm whenever axis is a 2-tuple or
# when axis=None and the input is 2-D, so to force a vector norm, we make
# it so the input is 1-D (for axis=None), or reshape so that norm is done
# on a single dimension.
if axis is None:
# Note: xp.linalg.norm() doesn't handle 0-D arrays
_x = x.ravel()
_axis = 0
elif isinstance(axis, tuple):
# Note: The axis argument supports any number of axes, whereas
# xp.linalg.norm() only supports a single axis for vector norm.
normalized_axis = cast(
"tuple[int, ...]",
normalize_axis_tuple(axis, x.ndim), # pyright: ignore[reportCallIssue]
)
rest = tuple(i for i in range(x.ndim) if i not in normalized_axis)
newshape = axis + rest
_x = xp.transpose(x, newshape).reshape(
(math.prod([x.shape[i] for i in axis]), *[x.shape[i] for i in rest]))
_axis = 0
else:
_x = x
_axis = axis
res = xp.linalg.norm(_x, axis=_axis, ord=ord)
if keepdims:
# We can't reuse xp.linalg.norm(keepdims) because of the reshape hacks
# above to avoid matrix norm logic.
shape = list(x.shape)
_axis = cast(
"tuple[int, ...]",
normalize_axis_tuple( # pyright: ignore[reportCallIssue]
range(x.ndim) if axis is None else axis,
x.ndim,
),
)
for i in _axis:
shape[i] = 1
res = xp.reshape(res, tuple(shape))
return res
# xp.diagonal and xp.trace operate on the first two axes whereas these
# operates on the last two
def diagonal(x: Array, /, xp: Namespace, *, offset: int = 0, **kwargs: object) -> Array:
return xp.diagonal(x, offset=offset, axis1=-2, axis2=-1, **kwargs)
def trace(
x: Array,
/,
xp: Namespace,
*,
offset: int = 0,
dtype: DType | None = None,
**kwargs: object,
) -> Array:
return xp.asarray(
xp.trace(x, offset=offset, dtype=dtype, axis1=-2, axis2=-1, **kwargs)
)
__all__ = ['cross', 'matmul', 'outer', 'tensordot', 'EighResult',
'QRResult', 'SlogdetResult', 'SVDResult', 'eigh', 'qr', 'slogdet',
'svd', 'cholesky', 'matrix_rank', 'pinv', 'matrix_norm',
'matrix_transpose', 'svdvals', 'vecdot', 'vector_norm', 'diagonal',
'trace']
_all_ignore = ['math', 'normalize_axis_tuple', 'get_xp', 'np', 'isdtype']
def __dir__() -> list[str]:
return __all__
@@ -0,0 +1,192 @@
from __future__ import annotations
from collections.abc import Mapping
from types import ModuleType as Namespace
from typing import (
TYPE_CHECKING,
Literal,
Protocol,
TypeAlias,
TypedDict,
TypeVar,
final,
)
if TYPE_CHECKING:
from _typeshed import Incomplete
SupportsBufferProtocol: TypeAlias = Incomplete
Array: TypeAlias = Incomplete
Device: TypeAlias = Incomplete
DType: TypeAlias = Incomplete
else:
SupportsBufferProtocol = object
Array = object
Device = object
DType = object
_T_co = TypeVar("_T_co", covariant=True)
# These "Just" types are equivalent to the `Just` type from the `optype` library,
# apart from them not being `@runtime_checkable`.
# - docs: https://github.com/jorenham/optype/blob/master/README.md#just
# - code: https://github.com/jorenham/optype/blob/master/optype/_core/_just.py
@final
class JustInt(Protocol):
@property
def __class__(self, /) -> type[int]: ...
@__class__.setter
def __class__(self, value: type[int], /) -> None: ... # pyright: ignore[reportIncompatibleMethodOverride]
@final
class JustFloat(Protocol):
@property
def __class__(self, /) -> type[float]: ...
@__class__.setter
def __class__(self, value: type[float], /) -> None: ... # pyright: ignore[reportIncompatibleMethodOverride]
@final
class JustComplex(Protocol):
@property
def __class__(self, /) -> type[complex]: ...
@__class__.setter
def __class__(self, value: type[complex], /) -> None: ... # pyright: ignore[reportIncompatibleMethodOverride]
#
class NestedSequence(Protocol[_T_co]):
def __getitem__(self, key: int, /) -> _T_co | NestedSequence[_T_co]: ...
def __len__(self, /) -> int: ...
class SupportsArrayNamespace(Protocol[_T_co]):
def __array_namespace__(self, /, *, api_version: str | None) -> _T_co: ...
class HasShape(Protocol[_T_co]):
@property
def shape(self, /) -> _T_co: ...
# Return type of `__array_namespace_info__.default_dtypes`
Capabilities = TypedDict(
"Capabilities",
{
"boolean indexing": bool,
"data-dependent shapes": bool,
"max dimensions": int,
},
)
# Return type of `__array_namespace_info__.default_dtypes`
DefaultDTypes = TypedDict(
"DefaultDTypes",
{
"real floating": DType,
"complex floating": DType,
"integral": DType,
"indexing": DType,
},
)
_DTypeKind: TypeAlias = Literal[
"bool",
"signed integer",
"unsigned integer",
"integral",
"real floating",
"complex floating",
"numeric",
]
# Type of the `kind` parameter in `__array_namespace_info__.dtypes`
DTypeKind: TypeAlias = _DTypeKind | tuple[_DTypeKind, ...]
# `__array_namespace_info__.dtypes(kind="bool")`
class DTypesBool(TypedDict):
bool: DType
# `__array_namespace_info__.dtypes(kind="signed integer")`
class DTypesSigned(TypedDict):
int8: DType
int16: DType
int32: DType
int64: DType
# `__array_namespace_info__.dtypes(kind="unsigned integer")`
class DTypesUnsigned(TypedDict):
uint8: DType
uint16: DType
uint32: DType
uint64: DType
# `__array_namespace_info__.dtypes(kind="integral")`
class DTypesIntegral(DTypesSigned, DTypesUnsigned):
pass
# `__array_namespace_info__.dtypes(kind="real floating")`
class DTypesReal(TypedDict):
float32: DType
float64: DType
# `__array_namespace_info__.dtypes(kind="complex floating")`
class DTypesComplex(TypedDict):
complex64: DType
complex128: DType
# `__array_namespace_info__.dtypes(kind="numeric")`
class DTypesNumeric(DTypesIntegral, DTypesReal, DTypesComplex):
pass
# `__array_namespace_info__.dtypes(kind=None)` (default)
class DTypesAll(DTypesBool, DTypesNumeric):
pass
# `__array_namespace_info__.dtypes(kind=?)` (fallback)
DTypesAny: TypeAlias = Mapping[str, DType]
__all__ = [
"Array",
"Capabilities",
"DType",
"DTypeKind",
"DTypesAny",
"DTypesAll",
"DTypesBool",
"DTypesNumeric",
"DTypesIntegral",
"DTypesSigned",
"DTypesUnsigned",
"DTypesReal",
"DTypesComplex",
"DefaultDTypes",
"Device",
"HasShape",
"Namespace",
"JustInt",
"JustFloat",
"JustComplex",
"NestedSequence",
"SupportsArrayNamespace",
"SupportsBufferProtocol",
]
def __dir__() -> list[str]:
return __all__
@@ -0,0 +1,13 @@
from cupy import * # noqa: F403
# from cupy import * doesn't overwrite these builtin names
from cupy import abs, max, min, round # noqa: F401
# These imports may overwrite names from the import * above.
from ._aliases import * # noqa: F403
# See the comment in the numpy __init__.py
__import__(__package__ + '.linalg')
__import__(__package__ + '.fft')
__array_api_version__ = '2024.12'
@@ -0,0 +1,156 @@
from __future__ import annotations
from typing import Optional
import cupy as cp
from ..common import _aliases, _helpers
from ..common._typing import NestedSequence, SupportsBufferProtocol
from .._internal import get_xp
from ._info import __array_namespace_info__
from ._typing import Array, Device, DType
bool = cp.bool_
# Basic renames
acos = cp.arccos
acosh = cp.arccosh
asin = cp.arcsin
asinh = cp.arcsinh
atan = cp.arctan
atan2 = cp.arctan2
atanh = cp.arctanh
bitwise_left_shift = cp.left_shift
bitwise_invert = cp.invert
bitwise_right_shift = cp.right_shift
concat = cp.concatenate
pow = cp.power
arange = get_xp(cp)(_aliases.arange)
empty = get_xp(cp)(_aliases.empty)
empty_like = get_xp(cp)(_aliases.empty_like)
eye = get_xp(cp)(_aliases.eye)
full = get_xp(cp)(_aliases.full)
full_like = get_xp(cp)(_aliases.full_like)
linspace = get_xp(cp)(_aliases.linspace)
ones = get_xp(cp)(_aliases.ones)
ones_like = get_xp(cp)(_aliases.ones_like)
zeros = get_xp(cp)(_aliases.zeros)
zeros_like = get_xp(cp)(_aliases.zeros_like)
UniqueAllResult = get_xp(cp)(_aliases.UniqueAllResult)
UniqueCountsResult = get_xp(cp)(_aliases.UniqueCountsResult)
UniqueInverseResult = get_xp(cp)(_aliases.UniqueInverseResult)
unique_all = get_xp(cp)(_aliases.unique_all)
unique_counts = get_xp(cp)(_aliases.unique_counts)
unique_inverse = get_xp(cp)(_aliases.unique_inverse)
unique_values = get_xp(cp)(_aliases.unique_values)
std = get_xp(cp)(_aliases.std)
var = get_xp(cp)(_aliases.var)
cumulative_sum = get_xp(cp)(_aliases.cumulative_sum)
cumulative_prod = get_xp(cp)(_aliases.cumulative_prod)
clip = get_xp(cp)(_aliases.clip)
permute_dims = get_xp(cp)(_aliases.permute_dims)
reshape = get_xp(cp)(_aliases.reshape)
argsort = get_xp(cp)(_aliases.argsort)
sort = get_xp(cp)(_aliases.sort)
nonzero = get_xp(cp)(_aliases.nonzero)
ceil = get_xp(cp)(_aliases.ceil)
floor = get_xp(cp)(_aliases.floor)
trunc = get_xp(cp)(_aliases.trunc)
matmul = get_xp(cp)(_aliases.matmul)
matrix_transpose = get_xp(cp)(_aliases.matrix_transpose)
tensordot = get_xp(cp)(_aliases.tensordot)
sign = get_xp(cp)(_aliases.sign)
finfo = get_xp(cp)(_aliases.finfo)
iinfo = get_xp(cp)(_aliases.iinfo)
# asarray also adds the copy keyword, which is not present in numpy 1.0.
def asarray(
obj: (
Array
| bool | int | float | complex
| NestedSequence[bool | int | float | complex]
| SupportsBufferProtocol
),
/,
*,
dtype: Optional[DType] = None,
device: Optional[Device] = None,
copy: Optional[bool] = None,
**kwargs,
) -> Array:
"""
Array API compatibility wrapper for asarray().
See the corresponding documentation in the array library and/or the array API
specification for more details.
"""
with cp.cuda.Device(device):
if copy is None:
return cp.asarray(obj, dtype=dtype, **kwargs)
else:
res = cp.array(obj, dtype=dtype, copy=copy, **kwargs)
if not copy and res is not obj:
raise ValueError("Unable to avoid copy while creating an array as requested")
return res
def astype(
x: Array,
dtype: DType,
/,
*,
copy: bool = True,
device: Optional[Device] = None,
) -> Array:
if device is None:
return x.astype(dtype=dtype, copy=copy)
out = _helpers.to_device(x.astype(dtype=dtype, copy=False), device)
return out.copy() if copy and out is x else out
# cupy.count_nonzero does not have keepdims
def count_nonzero(
x: Array,
axis=None,
keepdims=False
) -> Array:
result = cp.count_nonzero(x, axis)
if keepdims:
if axis is None:
return cp.reshape(result, [1]*x.ndim)
return cp.expand_dims(result, axis)
return result
# take_along_axis: axis defaults to -1 but in cupy (and numpy) axis is a required arg
def take_along_axis(x: Array, indices: Array, /, *, axis: int = -1):
return cp.take_along_axis(x, indices, axis=axis)
# These functions are completely new here. If the library already has them
# (i.e., numpy 2.0), use the library version instead of our wrapper.
if hasattr(cp, 'vecdot'):
vecdot = cp.vecdot
else:
vecdot = get_xp(cp)(_aliases.vecdot)
if hasattr(cp, 'isdtype'):
isdtype = cp.isdtype
else:
isdtype = get_xp(cp)(_aliases.isdtype)
if hasattr(cp, 'unstack'):
unstack = cp.unstack
else:
unstack = get_xp(cp)(_aliases.unstack)
__all__ = _aliases.__all__ + ['__array_namespace_info__', 'asarray', 'astype',
'acos', 'acosh', 'asin', 'asinh', 'atan',
'atan2', 'atanh', 'bitwise_left_shift',
'bitwise_invert', 'bitwise_right_shift',
'bool', 'concat', 'count_nonzero', 'pow', 'sign',
'take_along_axis']
_all_ignore = ['cp', 'get_xp']
@@ -0,0 +1,336 @@
"""
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.
"""
from cupy import (
dtype,
cuda,
bool_ as bool,
intp,
int8,
int16,
int32,
int64,
uint8,
uint16,
uint32,
uint64,
float32,
float64,
complex64,
complex128,
)
class __array_namespace_info__:
"""
Get the array API inspection namespace for CuPy.
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 CuPy.
Examples
--------
>>> info = xp.__array_namespace_info__()
>>> info.default_dtypes()
{'real floating': cupy.float64,
'complex floating': cupy.complex128,
'integral': cupy.int64,
'indexing': cupy.int64}
"""
__module__ = 'cupy'
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 CuPy.
- **"data-dependent shapes"**: boolean indicating whether an array
library supports data-dependent output shapes. Always ``True`` for
CuPy.
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 CuPy 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 CuPy arrays.
Examples
--------
>>> info = xp.__array_namespace_info__()
>>> info.default_device()
Device(0)
Notes
-----
This method returns the static default device when CuPy is initialized.
However, the *current* device used by creation functions (``empty`` etc.)
can be changed globally or with a context manager.
See Also
--------
https://github.com/data-apis/array-api/issues/835
"""
return cuda.Device(0)
def default_dtypes(self, *, device=None):
"""
The default data types used for new CuPy arrays.
For CuPy, this always returns the following dictionary:
- **"real floating"**: ``cupy.float64``
- **"complex floating"**: ``cupy.complex128``
- **"integral"**: ``cupy.intp``
- **"indexing"**: ``cupy.intp``
Parameters
----------
device : str, optional
The device to get the default data types for.
Returns
-------
dtypes : dict
A dictionary describing the default data types used for new CuPy
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': cupy.float64,
'complex floating': cupy.complex128,
'integral': cupy.int64,
'indexing': cupy.int64}
"""
# TODO: Does this depend on device?
return {
"real floating": dtype(float64),
"complex floating": dtype(complex128),
"integral": dtype(intp),
"indexing": dtype(intp),
}
def dtypes(self, *, device=None, kind=None):
"""
The array API data types supported by CuPy.
Note that this function only returns data types that are defined by
the array API.
Parameters
----------
device : str, optional
The device to get the data types for.
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
CuPy 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': cupy.int8,
'int16': cupy.int16,
'int32': cupy.int32,
'int64': cupy.int64}
"""
# TODO: Does this depend on device?
if kind is None:
return {
"bool": dtype(bool),
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
"float32": dtype(float32),
"float64": dtype(float64),
"complex64": dtype(complex64),
"complex128": dtype(complex128),
}
if kind == "bool":
return {"bool": bool}
if kind == "signed integer":
return {
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
}
if kind == "unsigned integer":
return {
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
}
if kind == "integral":
return {
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
}
if kind == "real floating":
return {
"float32": dtype(float32),
"float64": dtype(float64),
}
if kind == "complex floating":
return {
"complex64": dtype(complex64),
"complex128": dtype(complex128),
}
if kind == "numeric":
return {
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
"float32": dtype(float32),
"float64": dtype(float64),
"complex64": dtype(complex64),
"complex128": dtype(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}")
def devices(self):
"""
The devices supported by CuPy.
Returns
-------
devices : list[Device]
The devices supported by CuPy.
See Also
--------
__array_namespace_info__.capabilities,
__array_namespace_info__.default_device,
__array_namespace_info__.default_dtypes,
__array_namespace_info__.dtypes
"""
return [cuda.Device(i) for i in range(cuda.runtime.getDeviceCount())]
@@ -0,0 +1,31 @@
from __future__ import annotations
__all__ = ["Array", "DType", "Device"]
_all_ignore = ["cp"]
from typing import TYPE_CHECKING
import cupy as cp
from cupy import ndarray as Array
from cupy.cuda.device import Device
if TYPE_CHECKING:
# NumPy 1.x on Python 3.10 fails to parse np.dtype[]
DType = cp.dtype[
cp.intp
| cp.int8
| cp.int16
| cp.int32
| cp.int64
| cp.uint8
| cp.uint16
| cp.uint32
| cp.uint64
| cp.float32
| cp.float64
| cp.complex64
| cp.complex128
| cp.bool_
]
else:
DType = cp.dtype
@@ -0,0 +1,36 @@
from cupy.fft import * # noqa: F403
# cupy.fft doesn't have __all__. If it is added, replace this with
#
# from cupy.fft import __all__ as linalg_all
_n = {}
exec('from cupy.fft import *', _n)
del _n['__builtins__']
fft_all = list(_n)
del _n
from ..common import _fft
from .._internal import get_xp
import cupy as cp
fft = get_xp(cp)(_fft.fft)
ifft = get_xp(cp)(_fft.ifft)
fftn = get_xp(cp)(_fft.fftn)
ifftn = get_xp(cp)(_fft.ifftn)
rfft = get_xp(cp)(_fft.rfft)
irfft = get_xp(cp)(_fft.irfft)
rfftn = get_xp(cp)(_fft.rfftn)
irfftn = get_xp(cp)(_fft.irfftn)
hfft = get_xp(cp)(_fft.hfft)
ihfft = get_xp(cp)(_fft.ihfft)
fftfreq = get_xp(cp)(_fft.fftfreq)
rfftfreq = get_xp(cp)(_fft.rfftfreq)
fftshift = get_xp(cp)(_fft.fftshift)
ifftshift = get_xp(cp)(_fft.ifftshift)
__all__ = fft_all + _fft.__all__
del get_xp
del cp
del fft_all
del _fft
@@ -0,0 +1,49 @@
from cupy.linalg import * # noqa: F403
# cupy.linalg doesn't have __all__. If it is added, replace this with
#
# from cupy.linalg import __all__ as linalg_all
_n = {}
exec('from cupy.linalg import *', _n)
del _n['__builtins__']
linalg_all = list(_n)
del _n
from ..common import _linalg
from .._internal import get_xp
import cupy as cp
# These functions are in both the main and linalg namespaces
from ._aliases import matmul, matrix_transpose, tensordot, vecdot # noqa: F401
cross = get_xp(cp)(_linalg.cross)
outer = get_xp(cp)(_linalg.outer)
EighResult = _linalg.EighResult
QRResult = _linalg.QRResult
SlogdetResult = _linalg.SlogdetResult
SVDResult = _linalg.SVDResult
eigh = get_xp(cp)(_linalg.eigh)
qr = get_xp(cp)(_linalg.qr)
slogdet = get_xp(cp)(_linalg.slogdet)
svd = get_xp(cp)(_linalg.svd)
cholesky = get_xp(cp)(_linalg.cholesky)
matrix_rank = get_xp(cp)(_linalg.matrix_rank)
pinv = get_xp(cp)(_linalg.pinv)
matrix_norm = get_xp(cp)(_linalg.matrix_norm)
svdvals = get_xp(cp)(_linalg.svdvals)
diagonal = get_xp(cp)(_linalg.diagonal)
trace = get_xp(cp)(_linalg.trace)
# These functions are completely new here. If the library already has them
# (i.e., numpy 2.0), use the library version instead of our wrapper.
if hasattr(cp.linalg, 'vector_norm'):
vector_norm = cp.linalg.vector_norm
else:
vector_norm = get_xp(cp)(_linalg.vector_norm)
__all__ = linalg_all + _linalg.__all__
del get_xp
del cp
del linalg_all
del _linalg
@@ -0,0 +1,12 @@
from typing import Final
from dask.array import * # noqa: F403
# These imports may overwrite names from the import * above.
from ._aliases import * # noqa: F403
__array_api_version__: Final = "2024.12"
# See the comment in the numpy __init__.py
__import__(__package__ + '.linalg')
__import__(__package__ + '.fft')
@@ -0,0 +1,376 @@
# pyright: reportPrivateUsage=false
# pyright: reportUnknownArgumentType=false
# pyright: reportUnknownMemberType=false
# pyright: reportUnknownVariableType=false
from __future__ import annotations
from builtins import bool as py_bool
from collections.abc import Callable
from typing import TYPE_CHECKING, Any
if TYPE_CHECKING:
from typing_extensions import TypeIs
import dask.array as da
import numpy as np
from numpy import bool_ as bool
from numpy import (
can_cast,
complex64,
complex128,
float32,
float64,
int8,
int16,
int32,
int64,
result_type,
uint8,
uint16,
uint32,
uint64,
)
from ..._internal import get_xp
from ...common import _aliases, _helpers, array_namespace
from ...common._typing import (
Array,
Device,
DType,
NestedSequence,
SupportsBufferProtocol,
)
from ._info import __array_namespace_info__
isdtype = get_xp(np)(_aliases.isdtype)
unstack = get_xp(da)(_aliases.unstack)
# da.astype doesn't respect copy=True
def astype(
x: Array,
dtype: DType,
/,
*,
copy: py_bool = True,
device: Device | None = None,
) -> Array:
"""
Array API compatibility wrapper for astype().
See the corresponding documentation in the array library and/or the array API
specification for more details.
"""
# TODO: respect device keyword?
_helpers._check_device(da, device)
if not copy and dtype == x.dtype:
return x
x = x.astype(dtype)
return x.copy() if copy else x
# Common aliases
# This arange func is modified from the common one to
# not pass stop/step as keyword arguments, which will cause
# an error with dask
def arange(
start: float,
/,
stop: float | None = None,
step: float = 1,
*,
dtype: DType | None = None,
device: Device | None = None,
**kwargs: object,
) -> Array:
"""
Array API compatibility wrapper for arange().
See the corresponding documentation in the array library and/or the array API
specification for more details.
"""
# TODO: respect device keyword?
_helpers._check_device(da, device)
args: list[Any] = [start]
if stop is not None:
args.append(stop)
else:
# stop is None, so start is actually stop
# prepend the default value for start which is 0
args.insert(0, 0)
args.append(step)
return da.arange(*args, dtype=dtype, **kwargs)
eye = get_xp(da)(_aliases.eye)
linspace = get_xp(da)(_aliases.linspace)
UniqueAllResult = get_xp(da)(_aliases.UniqueAllResult)
UniqueCountsResult = get_xp(da)(_aliases.UniqueCountsResult)
UniqueInverseResult = get_xp(da)(_aliases.UniqueInverseResult)
unique_all = get_xp(da)(_aliases.unique_all)
unique_counts = get_xp(da)(_aliases.unique_counts)
unique_inverse = get_xp(da)(_aliases.unique_inverse)
unique_values = get_xp(da)(_aliases.unique_values)
permute_dims = get_xp(da)(_aliases.permute_dims)
std = get_xp(da)(_aliases.std)
var = get_xp(da)(_aliases.var)
cumulative_sum = get_xp(da)(_aliases.cumulative_sum)
cumulative_prod = get_xp(da)(_aliases.cumulative_prod)
empty = get_xp(da)(_aliases.empty)
empty_like = get_xp(da)(_aliases.empty_like)
full = get_xp(da)(_aliases.full)
full_like = get_xp(da)(_aliases.full_like)
ones = get_xp(da)(_aliases.ones)
ones_like = get_xp(da)(_aliases.ones_like)
zeros = get_xp(da)(_aliases.zeros)
zeros_like = get_xp(da)(_aliases.zeros_like)
reshape = get_xp(da)(_aliases.reshape)
matrix_transpose = get_xp(da)(_aliases.matrix_transpose)
vecdot = get_xp(da)(_aliases.vecdot)
nonzero = get_xp(da)(_aliases.nonzero)
ceil = get_xp(np)(_aliases.ceil)
floor = get_xp(np)(_aliases.floor)
trunc = get_xp(np)(_aliases.trunc)
matmul = get_xp(np)(_aliases.matmul)
tensordot = get_xp(np)(_aliases.tensordot)
sign = get_xp(np)(_aliases.sign)
finfo = get_xp(np)(_aliases.finfo)
iinfo = get_xp(np)(_aliases.iinfo)
# asarray also adds the copy keyword, which is not present in numpy 1.0.
def asarray(
obj: complex | NestedSequence[complex] | Array | SupportsBufferProtocol,
/,
*,
dtype: DType | None = None,
device: Device | None = None,
copy: py_bool | None = None,
**kwargs: object,
) -> Array:
"""
Array API compatibility wrapper for asarray().
See the corresponding documentation in the array library and/or the array API
specification for more details.
"""
# TODO: respect device keyword?
_helpers._check_device(da, device)
if isinstance(obj, da.Array):
if dtype is not None and dtype != obj.dtype:
if copy is False:
raise ValueError("Unable to avoid copy when changing dtype")
obj = obj.astype(dtype)
return obj.copy() if copy else obj # pyright: ignore[reportAttributeAccessIssue]
if copy is False:
raise ValueError(
"Unable to avoid copy when converting a non-dask object to dask"
)
# copy=None to be uniform across dask < 2024.12 and >= 2024.12
# see https://github.com/dask/dask/pull/11524/
obj = np.array(obj, dtype=dtype, copy=True)
return da.from_array(obj)
# Element wise aliases
from dask.array import arccos as acos
from dask.array import arccosh as acosh
from dask.array import arcsin as asin
from dask.array import arcsinh as asinh
from dask.array import arctan as atan
from dask.array import arctan2 as atan2
from dask.array import arctanh as atanh
# Other
from dask.array import concatenate as concat
from dask.array import invert as bitwise_invert
from dask.array import left_shift as bitwise_left_shift
from dask.array import power as pow
from dask.array import right_shift as bitwise_right_shift
# dask.array.clip does not work unless all three arguments are provided.
# Furthermore, the masking workaround in common._aliases.clip cannot work with
# dask (meaning uint64 promoting to float64 is going to just be unfixed for
# now).
def clip(
x: Array,
/,
min: float | Array | None = None,
max: float | Array | None = None,
) -> Array:
"""
Array API compatibility wrapper for clip().
See the corresponding documentation in the array library and/or the array API
specification for more details.
"""
def _isscalar(a: float | Array | None, /) -> TypeIs[float | None]:
return a is None or isinstance(a, (int, float))
min_shape = () if _isscalar(min) else min.shape
max_shape = () if _isscalar(max) else max.shape
# TODO: This won't handle dask unknown shapes
result_shape = np.broadcast_shapes(x.shape, min_shape, max_shape)
if min is not None:
min = da.broadcast_to(da.asarray(min), result_shape)
if max is not None:
max = da.broadcast_to(da.asarray(max), result_shape)
if min is None and max is None:
return da.positive(x)
if min is None:
return astype(da.minimum(x, max), x.dtype)
if max is None:
return astype(da.maximum(x, min), x.dtype)
return astype(da.minimum(da.maximum(x, min), max), x.dtype)
def _ensure_single_chunk(x: Array, axis: int) -> tuple[Array, Callable[[Array], Array]]:
"""
Make sure that Array is not broken into multiple chunks along axis.
Returns
-------
x : Array
The input Array with a single chunk along axis.
restore : Callable[Array, Array]
function to apply to the output to rechunk it back into reasonable chunks
"""
if axis < 0:
axis += x.ndim
if x.numblocks[axis] < 2:
return x, lambda x: x
# Break chunks on other axes in an attempt to keep chunk size low
x = x.rechunk({i: -1 if i == axis else "auto" for i in range(x.ndim)})
# Rather than reconstructing the original chunks, which can be a
# very expensive affair, just break down oversized chunks without
# incurring in any transfers over the network.
# This has the downside of a risk of overchunking if the array is
# then used in operations against other arrays that match the
# original chunking pattern.
return x, lambda x: x.rechunk()
def sort(
x: Array,
/,
*,
axis: int = -1,
descending: py_bool = False,
stable: py_bool = True,
) -> Array:
"""
Array API compatibility layer around the lack of sort() in Dask.
Warnings
--------
This function temporarily rechunks the array along `axis` to a single chunk.
This can be extremely inefficient and can lead to out-of-memory errors.
See the corresponding documentation in the array library and/or the array API
specification for more details.
"""
x, restore = _ensure_single_chunk(x, axis)
meta_xp = array_namespace(x._meta)
x = da.map_blocks(
meta_xp.sort,
x,
axis=axis,
meta=x._meta,
dtype=x.dtype,
descending=descending,
stable=stable,
)
return restore(x)
def argsort(
x: Array,
/,
*,
axis: int = -1,
descending: py_bool = False,
stable: py_bool = True,
) -> Array:
"""
Array API compatibility layer around the lack of argsort() in Dask.
See the corresponding documentation in the array library and/or the array API
specification for more details.
Warnings
--------
This function temporarily rechunks the array along `axis` into a single chunk.
This can be extremely inefficient and can lead to out-of-memory errors.
"""
x, restore = _ensure_single_chunk(x, axis)
meta_xp = array_namespace(x._meta)
dtype = meta_xp.argsort(x._meta).dtype
meta = meta_xp.astype(x._meta, dtype)
x = da.map_blocks(
meta_xp.argsort,
x,
axis=axis,
meta=meta,
dtype=dtype,
descending=descending,
stable=stable,
)
return restore(x)
# dask.array.count_nonzero does not have keepdims
def count_nonzero(
x: Array,
axis: int | None = None,
keepdims: py_bool = False,
) -> Array:
result = da.count_nonzero(x, axis)
if keepdims:
if axis is None:
return da.reshape(result, [1] * x.ndim)
return da.expand_dims(result, axis)
return result
__all__ = [
"__array_namespace_info__",
"count_nonzero",
"bool",
"int8", "int16", "int32", "int64",
"uint8", "uint16", "uint32", "uint64",
"float32", "float64",
"complex64", "complex128",
"asarray", "astype", "can_cast", "result_type",
"pow",
"concat",
"acos", "acosh", "asin", "asinh", "atan", "atan2", "atanh",
"bitwise_left_shift", "bitwise_right_shift", "bitwise_invert",
] # fmt: skip
__all__ += _aliases.__all__
_all_ignore = ["array_namespace", "get_xp", "da", "np"]
def __dir__() -> list[str]:
return __all__
@@ -0,0 +1,416 @@
"""
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.
"""
# pyright: reportPrivateUsage=false
from __future__ import annotations
from typing import Literal as L
from typing import TypeAlias, overload
from numpy import bool_ as bool
from numpy import (
complex64,
complex128,
dtype,
float32,
float64,
int8,
int16,
int32,
int64,
intp,
uint8,
uint16,
uint32,
uint64,
)
from ...common._helpers import _DASK_DEVICE, _dask_device
from ...common._typing import (
Capabilities,
DefaultDTypes,
DType,
DTypeKind,
DTypesAll,
DTypesAny,
DTypesBool,
DTypesComplex,
DTypesIntegral,
DTypesNumeric,
DTypesReal,
DTypesSigned,
DTypesUnsigned,
)
_Device: TypeAlias = L["cpu"] | _dask_device
class __array_namespace_info__:
"""
Get the array API inspection namespace for Dask.
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 Dask.
Examples
--------
>>> info = xp.__array_namespace_info__()
>>> info.default_dtypes()
{'real floating': dask.float64,
'complex floating': dask.complex128,
'integral': dask.int64,
'indexing': dask.int64}
"""
__module__ = "dask.array"
def capabilities(self) -> Capabilities:
"""
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.
Dask support boolean indexing as long as both the index
and the indexed arrays have known shapes.
Note however that the output .shape and .size properties
will contain a non-compliant math.nan instead of None.
- **"data-dependent shapes"**: boolean indicating whether an array
library supports data-dependent output shapes.
Dask implements unique_values et.al.
Note however that the output .shape and .size properties
will contain a non-compliant math.nan instead of None.
- **"max dimensions"**: integer indicating the maximum number of
dimensions supported by the array library.
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) -> L["cpu"]:
"""
The default device used for new Dask arrays.
For Dask, this always returns ``'cpu'``.
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 Dask arrays.
Examples
--------
>>> info = xp.__array_namespace_info__()
>>> info.default_device()
'cpu'
"""
return "cpu"
def default_dtypes(self, /, *, device: _Device | None = None) -> DefaultDTypes:
"""
The default data types used for new Dask arrays.
For Dask, this always returns the following dictionary:
- **"real floating"**: ``numpy.float64``
- **"complex floating"**: ``numpy.complex128``
- **"integral"**: ``numpy.intp``
- **"indexing"**: ``numpy.intp``
Parameters
----------
device : str, optional
The device to get the default data types for.
Returns
-------
dtypes : dict
A dictionary describing the default data types used for new Dask
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': dask.float64,
'complex floating': dask.complex128,
'integral': dask.int64,
'indexing': dask.int64}
"""
if device not in ["cpu", _DASK_DEVICE, None]:
raise ValueError(
f'Device not understood. Only "cpu" or _DASK_DEVICE is allowed, '
f"but received: {device!r}"
)
return {
"real floating": dtype(float64),
"complex floating": dtype(complex128),
"integral": dtype(intp),
"indexing": dtype(intp),
}
@overload
def dtypes(
self, /, *, device: _Device | None = None, kind: None = None
) -> DTypesAll: ...
@overload
def dtypes(
self, /, *, device: _Device | None = None, kind: L["bool"]
) -> DTypesBool: ...
@overload
def dtypes(
self, /, *, device: _Device | None = None, kind: L["signed integer"]
) -> DTypesSigned: ...
@overload
def dtypes(
self, /, *, device: _Device | None = None, kind: L["unsigned integer"]
) -> DTypesUnsigned: ...
@overload
def dtypes(
self, /, *, device: _Device | None = None, kind: L["integral"]
) -> DTypesIntegral: ...
@overload
def dtypes(
self, /, *, device: _Device | None = None, kind: L["real floating"]
) -> DTypesReal: ...
@overload
def dtypes(
self, /, *, device: _Device | None = None, kind: L["complex floating"]
) -> DTypesComplex: ...
@overload
def dtypes(
self, /, *, device: _Device | None = None, kind: L["numeric"]
) -> DTypesNumeric: ...
def dtypes(
self, /, *, device: _Device | None = None, kind: DTypeKind | None = None
) -> DTypesAny:
"""
The array API data types supported by Dask.
Note that this function only returns data types that are defined by
the array API.
Parameters
----------
device : str, optional
The device to get the data types for.
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
Dask 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': dask.int8,
'int16': dask.int16,
'int32': dask.int32,
'int64': dask.int64}
"""
if device not in ["cpu", _DASK_DEVICE, None]:
raise ValueError(
'Device not understood. Only "cpu" or _DASK_DEVICE is allowed, but received:'
f" {device}"
)
if kind is None:
return {
"bool": dtype(bool),
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
"float32": dtype(float32),
"float64": dtype(float64),
"complex64": dtype(complex64),
"complex128": dtype(complex128),
}
if kind == "bool":
return {"bool": bool}
if kind == "signed integer":
return {
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
}
if kind == "unsigned integer":
return {
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
}
if kind == "integral":
return {
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
}
if kind == "real floating":
return {
"float32": dtype(float32),
"float64": dtype(float64),
}
if kind == "complex floating":
return {
"complex64": dtype(complex64),
"complex128": dtype(complex128),
}
if kind == "numeric":
return {
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
"float32": dtype(float32),
"float64": dtype(float64),
"complex64": dtype(complex64),
"complex128": dtype(complex128),
}
if isinstance(kind, tuple): # type: ignore[reportUnnecessaryIsinstanceCall]
res: dict[str, DType] = {}
for k in kind:
res.update(self.dtypes(kind=k))
return res
raise ValueError(f"unsupported kind: {kind!r}")
def devices(self) -> list[_Device]:
"""
The devices supported by Dask.
For Dask, this always returns ``['cpu', DASK_DEVICE]``.
Returns
-------
devices : list[Device]
The devices supported by Dask.
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()
['cpu', DASK_DEVICE]
"""
return ["cpu", _DASK_DEVICE]
@@ -0,0 +1,21 @@
from dask.array.fft import * # noqa: F403
# dask.array.fft doesn't have __all__. If it is added, replace this with
#
# from dask.array.fft import __all__ as linalg_all
_n = {}
exec('from dask.array.fft import *', _n)
for k in ("__builtins__", "Sequence", "annotations", "warnings"):
_n.pop(k, None)
fft_all = list(_n)
del _n, k
from ...common import _fft
from ..._internal import get_xp
import dask.array as da
fftfreq = get_xp(da)(_fft.fftfreq)
rfftfreq = get_xp(da)(_fft.rfftfreq)
__all__ = fft_all + ["fftfreq", "rfftfreq"]
_all_ignore = ["da", "fft_all", "get_xp", "warnings"]
@@ -0,0 +1,72 @@
from __future__ import annotations
from typing import Literal
import dask.array as da
# The `matmul` and `tensordot` functions are in both the main and linalg namespaces
from dask.array import matmul, outer, tensordot
# Exports
from dask.array.linalg import * # noqa: F403
from ..._internal import get_xp
from ...common import _linalg
from ...common._typing import Array as _Array
from ._aliases import matrix_transpose, vecdot
# dask.array.linalg doesn't have __all__. If it is added, replace this with
#
# from dask.array.linalg import __all__ as linalg_all
_n = {}
exec('from dask.array.linalg import *', _n)
for k in ('__builtins__', 'annotations', 'operator', 'warnings', 'Array'):
_n.pop(k, None)
linalg_all = list(_n)
del _n, k
EighResult = _linalg.EighResult
QRResult = _linalg.QRResult
SlogdetResult = _linalg.SlogdetResult
SVDResult = _linalg.SVDResult
# TODO: use the QR wrapper once dask
# supports the mode keyword on QR
# https://github.com/dask/dask/issues/10388
#qr = get_xp(da)(_linalg.qr)
def qr(
x: _Array,
mode: Literal["reduced", "complete"] = "reduced",
**kwargs: object,
) -> QRResult:
if mode != "reduced":
raise ValueError("dask arrays only support using mode='reduced'")
return QRResult(*da.linalg.qr(x, **kwargs))
trace = get_xp(da)(_linalg.trace)
cholesky = get_xp(da)(_linalg.cholesky)
matrix_rank = get_xp(da)(_linalg.matrix_rank)
matrix_norm = get_xp(da)(_linalg.matrix_norm)
# Wrap the svd functions to not pass full_matrices to dask
# when full_matrices=False (as that is the default behavior for dask),
# and dask doesn't have the full_matrices keyword
def svd(x: _Array, full_matrices: bool = True, **kwargs) -> SVDResult:
if full_matrices:
raise ValueError("full_matrics=True is not supported by dask.")
return da.linalg.svd(x, coerce_signs=False, **kwargs)
def svdvals(x: _Array) -> _Array:
# TODO: can't avoid computing U or V for dask
_, s, _ = svd(x)
return s
vector_norm = get_xp(da)(_linalg.vector_norm)
diagonal = get_xp(da)(_linalg.diagonal)
__all__ = linalg_all + ["trace", "outer", "matmul", "tensordot",
"matrix_transpose", "vecdot", "EighResult",
"QRResult", "SlogdetResult", "SVDResult", "qr",
"cholesky", "matrix_rank", "matrix_norm", "svdvals",
"vector_norm", "diagonal"]
_all_ignore = ['get_xp', 'da', 'linalg_all', 'warnings']
@@ -0,0 +1,28 @@
# ruff: noqa: PLC0414
from typing import Final
from numpy import * # noqa: F403 # pyright: ignore[reportWildcardImportFromLibrary]
# from numpy import * doesn't overwrite these builtin names
from numpy import abs as abs
from numpy import max as max
from numpy import min as min
from numpy import round as round
# These imports may overwrite names from the import * above.
from ._aliases import * # noqa: F403
# Don't know why, but we have to do an absolute import to import linalg. If we
# instead do
#
# from . import linalg
#
# It doesn't overwrite np.linalg from above. The import is generated
# dynamically so that the library can be vendored.
__import__(__package__ + ".linalg")
__import__(__package__ + ".fft")
from .linalg import matrix_transpose, vecdot # type: ignore[no-redef] # noqa: F401
__array_api_version__: Final = "2024.12"
@@ -0,0 +1,190 @@
# pyright: reportPrivateUsage=false
from __future__ import annotations
from builtins import bool as py_bool
from typing import TYPE_CHECKING, Any, Literal, TypeAlias, cast
import numpy as np
from .._internal import get_xp
from ..common import _aliases, _helpers
from ..common._typing import NestedSequence, SupportsBufferProtocol
from ._info import __array_namespace_info__
from ._typing import Array, Device, DType
if TYPE_CHECKING:
from typing_extensions import Buffer, TypeIs
# The values of the `_CopyMode` enum can be either `False`, `True`, or `2`:
# https://github.com/numpy/numpy/blob/5a8a6a79d9c2fff8f07dcab5d41e14f8508d673f/numpy/_globals.pyi#L7-L10
_Copy: TypeAlias = py_bool | Literal[2] | np._CopyMode
bool = np.bool_
# Basic renames
acos = np.arccos
acosh = np.arccosh
asin = np.arcsin
asinh = np.arcsinh
atan = np.arctan
atan2 = np.arctan2
atanh = np.arctanh
bitwise_left_shift = np.left_shift
bitwise_invert = np.invert
bitwise_right_shift = np.right_shift
concat = np.concatenate
pow = np.power
arange = get_xp(np)(_aliases.arange)
empty = get_xp(np)(_aliases.empty)
empty_like = get_xp(np)(_aliases.empty_like)
eye = get_xp(np)(_aliases.eye)
full = get_xp(np)(_aliases.full)
full_like = get_xp(np)(_aliases.full_like)
linspace = get_xp(np)(_aliases.linspace)
ones = get_xp(np)(_aliases.ones)
ones_like = get_xp(np)(_aliases.ones_like)
zeros = get_xp(np)(_aliases.zeros)
zeros_like = get_xp(np)(_aliases.zeros_like)
UniqueAllResult = get_xp(np)(_aliases.UniqueAllResult)
UniqueCountsResult = get_xp(np)(_aliases.UniqueCountsResult)
UniqueInverseResult = get_xp(np)(_aliases.UniqueInverseResult)
unique_all = get_xp(np)(_aliases.unique_all)
unique_counts = get_xp(np)(_aliases.unique_counts)
unique_inverse = get_xp(np)(_aliases.unique_inverse)
unique_values = get_xp(np)(_aliases.unique_values)
std = get_xp(np)(_aliases.std)
var = get_xp(np)(_aliases.var)
cumulative_sum = get_xp(np)(_aliases.cumulative_sum)
cumulative_prod = get_xp(np)(_aliases.cumulative_prod)
clip = get_xp(np)(_aliases.clip)
permute_dims = get_xp(np)(_aliases.permute_dims)
reshape = get_xp(np)(_aliases.reshape)
argsort = get_xp(np)(_aliases.argsort)
sort = get_xp(np)(_aliases.sort)
nonzero = get_xp(np)(_aliases.nonzero)
ceil = get_xp(np)(_aliases.ceil)
floor = get_xp(np)(_aliases.floor)
trunc = get_xp(np)(_aliases.trunc)
matmul = get_xp(np)(_aliases.matmul)
matrix_transpose = get_xp(np)(_aliases.matrix_transpose)
tensordot = get_xp(np)(_aliases.tensordot)
sign = get_xp(np)(_aliases.sign)
finfo = get_xp(np)(_aliases.finfo)
iinfo = get_xp(np)(_aliases.iinfo)
def _supports_buffer_protocol(obj: object) -> TypeIs[Buffer]: # pyright: ignore[reportUnusedFunction]
try:
memoryview(obj) # pyright: ignore[reportArgumentType]
except TypeError:
return False
return True
# asarray also adds the copy keyword, which is not present in numpy 1.0.
# asarray() is different enough between numpy, cupy, and dask, the logic
# complicated enough that it's easier to define it separately for each module
# rather than trying to combine everything into one function in common/
def asarray(
obj: Array | complex | NestedSequence[complex] | SupportsBufferProtocol,
/,
*,
dtype: DType | None = None,
device: Device | None = None,
copy: _Copy | None = None,
**kwargs: Any,
) -> Array:
"""
Array API compatibility wrapper for asarray().
See the corresponding documentation in the array library and/or the array API
specification for more details.
"""
_helpers._check_device(np, device)
if copy is None:
copy = np._CopyMode.IF_NEEDED
elif copy is False:
copy = np._CopyMode.NEVER
elif copy is True:
copy = np._CopyMode.ALWAYS
return np.array(obj, copy=copy, dtype=dtype, **kwargs) # pyright: ignore
def astype(
x: Array,
dtype: DType,
/,
*,
copy: py_bool = True,
device: Device | None = None,
) -> Array:
_helpers._check_device(np, device)
return x.astype(dtype=dtype, copy=copy)
# count_nonzero returns a python int for axis=None and keepdims=False
# https://github.com/numpy/numpy/issues/17562
def count_nonzero(
x: Array,
axis: int | tuple[int, ...] | None = None,
keepdims: py_bool = False,
) -> Array:
# NOTE: this is currently incorrectly typed in numpy, but will be fixed in
# numpy 2.2.5 and 2.3.0: https://github.com/numpy/numpy/pull/28750
result = cast("Any", np.count_nonzero(x, axis=axis, keepdims=keepdims)) # pyright: ignore[reportArgumentType, reportCallIssue]
if axis is None and not keepdims:
return np.asarray(result)
return result
# take_along_axis: axis defaults to -1 but in numpy axis is a required arg
def take_along_axis(x: Array, indices: Array, /, *, axis: int = -1):
return np.take_along_axis(x, indices, axis=axis)
# These functions are completely new here. If the library already has them
# (i.e., numpy 2.0), use the library version instead of our wrapper.
if hasattr(np, "vecdot"):
vecdot = np.vecdot
else:
vecdot = get_xp(np)(_aliases.vecdot)
if hasattr(np, "isdtype"):
isdtype = np.isdtype
else:
isdtype = get_xp(np)(_aliases.isdtype)
if hasattr(np, "unstack"):
unstack = np.unstack
else:
unstack = get_xp(np)(_aliases.unstack)
__all__ = [
"__array_namespace_info__",
"asarray",
"astype",
"acos",
"acosh",
"asin",
"asinh",
"atan",
"atan2",
"atanh",
"bitwise_left_shift",
"bitwise_invert",
"bitwise_right_shift",
"bool",
"concat",
"count_nonzero",
"pow",
"take_along_axis"
]
__all__ += _aliases.__all__
_all_ignore = ["np", "get_xp"]
def __dir__() -> list[str]:
return __all__
@@ -0,0 +1,366 @@
"""
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.
"""
from __future__ import annotations
from numpy import bool_ as bool
from numpy import (
complex64,
complex128,
dtype,
float32,
float64,
int8,
int16,
int32,
int64,
intp,
uint8,
uint16,
uint32,
uint64,
)
from ._typing import Device, DType
class __array_namespace_info__:
"""
Get the array API inspection namespace for NumPy.
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 NumPy.
Examples
--------
>>> info = np.__array_namespace_info__()
>>> info.default_dtypes()
{'real floating': numpy.float64,
'complex floating': numpy.complex128,
'integral': numpy.int64,
'indexing': numpy.int64}
"""
__module__ = 'numpy'
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 NumPy.
- **"data-dependent shapes"**: boolean indicating whether an array
library supports data-dependent output shapes. Always ``True`` for
NumPy.
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 = np.__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 NumPy arrays.
For NumPy, this always returns ``'cpu'``.
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 NumPy arrays.
Examples
--------
>>> info = np.__array_namespace_info__()
>>> info.default_device()
'cpu'
"""
return "cpu"
def default_dtypes(
self,
*,
device: Device | None = None,
) -> dict[str, dtype[intp | float64 | complex128]]:
"""
The default data types used for new NumPy arrays.
For NumPy, this always returns the following dictionary:
- **"real floating"**: ``numpy.float64``
- **"complex floating"**: ``numpy.complex128``
- **"integral"**: ``numpy.intp``
- **"indexing"**: ``numpy.intp``
Parameters
----------
device : str, optional
The device to get the default data types for. For NumPy, only
``'cpu'`` is allowed.
Returns
-------
dtypes : dict
A dictionary describing the default data types used for new NumPy
arrays.
See Also
--------
__array_namespace_info__.capabilities,
__array_namespace_info__.default_device,
__array_namespace_info__.dtypes,
__array_namespace_info__.devices
Examples
--------
>>> info = np.__array_namespace_info__()
>>> info.default_dtypes()
{'real floating': numpy.float64,
'complex floating': numpy.complex128,
'integral': numpy.int64,
'indexing': numpy.int64}
"""
if device not in ["cpu", None]:
raise ValueError(
'Device not understood. Only "cpu" is allowed, but received:'
f' {device}'
)
return {
"real floating": dtype(float64),
"complex floating": dtype(complex128),
"integral": dtype(intp),
"indexing": dtype(intp),
}
def dtypes(
self,
*,
device: Device | None = None,
kind: str | tuple[str, ...] | None = None,
) -> dict[str, DType]:
"""
The array API data types supported by NumPy.
Note that this function only returns data types that are defined by
the array API.
Parameters
----------
device : str, optional
The device to get the data types for. For NumPy, only ``'cpu'`` is
allowed.
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
NumPy data types.
See Also
--------
__array_namespace_info__.capabilities,
__array_namespace_info__.default_device,
__array_namespace_info__.default_dtypes,
__array_namespace_info__.devices
Examples
--------
>>> info = np.__array_namespace_info__()
>>> info.dtypes(kind='signed integer')
{'int8': numpy.int8,
'int16': numpy.int16,
'int32': numpy.int32,
'int64': numpy.int64}
"""
if device not in ["cpu", None]:
raise ValueError(
'Device not understood. Only "cpu" is allowed, but received:'
f' {device}'
)
if kind is None:
return {
"bool": dtype(bool),
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
"float32": dtype(float32),
"float64": dtype(float64),
"complex64": dtype(complex64),
"complex128": dtype(complex128),
}
if kind == "bool":
return {"bool": dtype(bool)}
if kind == "signed integer":
return {
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
}
if kind == "unsigned integer":
return {
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
}
if kind == "integral":
return {
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
}
if kind == "real floating":
return {
"float32": dtype(float32),
"float64": dtype(float64),
}
if kind == "complex floating":
return {
"complex64": dtype(complex64),
"complex128": dtype(complex128),
}
if kind == "numeric":
return {
"int8": dtype(int8),
"int16": dtype(int16),
"int32": dtype(int32),
"int64": dtype(int64),
"uint8": dtype(uint8),
"uint16": dtype(uint16),
"uint32": dtype(uint32),
"uint64": dtype(uint64),
"float32": dtype(float32),
"float64": dtype(float64),
"complex64": dtype(complex64),
"complex128": dtype(complex128),
}
if isinstance(kind, tuple):
res: dict[str, DType] = {}
for k in kind:
res.update(self.dtypes(kind=k))
return res
raise ValueError(f"unsupported kind: {kind!r}")
def devices(self) -> list[Device]:
"""
The devices supported by NumPy.
For NumPy, this always returns ``['cpu']``.
Returns
-------
devices : list[Device]
The devices supported by NumPy.
See Also
--------
__array_namespace_info__.capabilities,
__array_namespace_info__.default_device,
__array_namespace_info__.default_dtypes,
__array_namespace_info__.dtypes
Examples
--------
>>> info = np.__array_namespace_info__()
>>> info.devices()
['cpu']
"""
return ["cpu"]
__all__ = ["__array_namespace_info__"]
def __dir__() -> list[str]:
return __all__
@@ -0,0 +1,30 @@
from __future__ import annotations
from typing import TYPE_CHECKING, Any, Literal, TypeAlias
import numpy as np
Device: TypeAlias = Literal["cpu"]
if TYPE_CHECKING:
# NumPy 1.x on Python 3.10 fails to parse np.dtype[]
DType: TypeAlias = np.dtype[
np.bool_
| np.integer[Any]
| np.float32
| np.float64
| np.complex64
| np.complex128
]
Array: TypeAlias = np.ndarray[Any, DType]
else:
DType: TypeAlias = np.dtype
Array: TypeAlias = np.ndarray
__all__ = ["Array", "DType", "Device"]
_all_ignore = ["np"]
def __dir__() -> list[str]:
return __all__
@@ -0,0 +1,35 @@
import numpy as np
from numpy.fft import __all__ as fft_all
from numpy.fft import fft2, ifft2, irfft2, rfft2
from .._internal import get_xp
from ..common import _fft
fft = get_xp(np)(_fft.fft)
ifft = get_xp(np)(_fft.ifft)
fftn = get_xp(np)(_fft.fftn)
ifftn = get_xp(np)(_fft.ifftn)
rfft = get_xp(np)(_fft.rfft)
irfft = get_xp(np)(_fft.irfft)
rfftn = get_xp(np)(_fft.rfftn)
irfftn = get_xp(np)(_fft.irfftn)
hfft = get_xp(np)(_fft.hfft)
ihfft = get_xp(np)(_fft.ihfft)
fftfreq = get_xp(np)(_fft.fftfreq)
rfftfreq = get_xp(np)(_fft.rfftfreq)
fftshift = get_xp(np)(_fft.fftshift)
ifftshift = get_xp(np)(_fft.ifftshift)
__all__ = ["rfft2", "irfft2", "fft2", "ifft2"]
__all__ += _fft.__all__
def __dir__() -> list[str]:
return __all__
del get_xp
del np
del fft_all
del _fft
@@ -0,0 +1,143 @@
# pyright: reportAttributeAccessIssue=false
# pyright: reportUnknownArgumentType=false
# pyright: reportUnknownMemberType=false
# pyright: reportUnknownVariableType=false
from __future__ import annotations
import numpy as np
# intersection of `np.linalg.__all__` on numpy 1.22 and 2.2, minus `_linalg.__all__`
from numpy.linalg import (
LinAlgError,
cond,
det,
eig,
eigvals,
eigvalsh,
inv,
lstsq,
matrix_power,
multi_dot,
norm,
tensorinv,
tensorsolve,
)
from .._internal import get_xp
from ..common import _linalg
# These functions are in both the main and linalg namespaces
from ._aliases import matmul, matrix_transpose, tensordot, vecdot # noqa: F401
from ._typing import Array
cross = get_xp(np)(_linalg.cross)
outer = get_xp(np)(_linalg.outer)
EighResult = _linalg.EighResult
QRResult = _linalg.QRResult
SlogdetResult = _linalg.SlogdetResult
SVDResult = _linalg.SVDResult
eigh = get_xp(np)(_linalg.eigh)
qr = get_xp(np)(_linalg.qr)
slogdet = get_xp(np)(_linalg.slogdet)
svd = get_xp(np)(_linalg.svd)
cholesky = get_xp(np)(_linalg.cholesky)
matrix_rank = get_xp(np)(_linalg.matrix_rank)
pinv = get_xp(np)(_linalg.pinv)
matrix_norm = get_xp(np)(_linalg.matrix_norm)
svdvals = get_xp(np)(_linalg.svdvals)
diagonal = get_xp(np)(_linalg.diagonal)
trace = get_xp(np)(_linalg.trace)
# Note: unlike np.linalg.solve, the array API solve() only accepts x2 as a
# vector when it is exactly 1-dimensional. All other cases treat x2 as a stack
# of matrices. The np.linalg.solve behavior of allowing stacks of both
# matrices and vectors is ambiguous c.f.
# https://github.com/numpy/numpy/issues/15349 and
# https://github.com/data-apis/array-api/issues/285.
# To workaround this, the below is the code from np.linalg.solve except
# only calling solve1 in the exactly 1D case.
# This code is here instead of in common because it is numpy specific. Also
# note that CuPy's solve() does not currently support broadcasting (see
# https://github.com/cupy/cupy/blob/main/cupy/cublas.py#L43).
def solve(x1: Array, x2: Array, /) -> Array:
try:
from numpy.linalg._linalg import (
_assert_stacked_2d,
_assert_stacked_square,
_commonType,
_makearray,
_raise_linalgerror_singular,
isComplexType,
)
except ImportError:
from numpy.linalg.linalg import (
_assert_stacked_2d,
_assert_stacked_square,
_commonType,
_makearray,
_raise_linalgerror_singular,
isComplexType,
)
from numpy.linalg import _umath_linalg
x1, _ = _makearray(x1)
_assert_stacked_2d(x1)
_assert_stacked_square(x1)
x2, wrap = _makearray(x2)
t, result_t = _commonType(x1, x2)
# This part is different from np.linalg.solve
gufunc: np.ufunc
if x2.ndim == 1:
gufunc = _umath_linalg.solve1
else:
gufunc = _umath_linalg.solve
# This does nothing currently but is left in because it will be relevant
# when complex dtype support is added to the spec in 2022.
signature = "DD->D" if isComplexType(t) else "dd->d"
with np.errstate(
call=_raise_linalgerror_singular,
invalid="call",
over="ignore",
divide="ignore",
under="ignore",
):
r: Array = gufunc(x1, x2, signature=signature)
return wrap(r.astype(result_t, copy=False))
# These functions are completely new here. If the library already has them
# (i.e., numpy 2.0), use the library version instead of our wrapper.
if hasattr(np.linalg, "vector_norm"):
vector_norm = np.linalg.vector_norm
else:
vector_norm = get_xp(np)(_linalg.vector_norm)
__all__ = [
"LinAlgError",
"cond",
"det",
"eig",
"eigvals",
"eigvalsh",
"inv",
"lstsq",
"matrix_power",
"multi_dot",
"norm",
"tensorinv",
"tensorsolve",
]
__all__ += _linalg.__all__
__all__ += ["solve", "vector_norm"]
def __dir__() -> list[str]:
return __all__
@@ -0,0 +1,22 @@
from torch import * # noqa: F403
# Several names are not included in the above import *
import torch
for n in dir(torch):
if (n.startswith('_')
or n.endswith('_')
or 'cuda' in n
or 'cpu' in n
or 'backward' in n):
continue
exec(f"{n} = torch.{n}")
del n
# These imports may overwrite names from the import * above.
from ._aliases import * # noqa: F403
# See the comment in the numpy __init__.py
__import__(__package__ + '.linalg')
__import__(__package__ + '.fft')
__array_api_version__ = '2024.12'
@@ -0,0 +1,855 @@
from __future__ import annotations
from functools import reduce as _reduce, wraps as _wraps
from builtins import all as _builtin_all, any as _builtin_any
from typing import Any, List, Optional, Sequence, Tuple, Union, Literal
import torch
from .._internal import get_xp
from ..common import _aliases
from ..common._typing import NestedSequence, SupportsBufferProtocol
from ._info import __array_namespace_info__
from ._typing import Array, Device, DType
_int_dtypes = {
torch.uint8,
torch.int8,
torch.int16,
torch.int32,
torch.int64,
}
try:
# torch >=2.3
_int_dtypes |= {torch.uint16, torch.uint32, torch.uint64}
except AttributeError:
pass
_array_api_dtypes = {
torch.bool,
*_int_dtypes,
torch.float32,
torch.float64,
torch.complex64,
torch.complex128,
}
_promotion_table = {
# ints
(torch.int8, torch.int16): torch.int16,
(torch.int8, torch.int32): torch.int32,
(torch.int8, torch.int64): torch.int64,
(torch.int16, torch.int32): torch.int32,
(torch.int16, torch.int64): torch.int64,
(torch.int32, torch.int64): torch.int64,
# ints and uints (mixed sign)
(torch.uint8, torch.int8): torch.int16,
(torch.uint8, torch.int16): torch.int16,
(torch.uint8, torch.int32): torch.int32,
(torch.uint8, torch.int64): torch.int64,
# floats
(torch.float32, torch.float64): torch.float64,
# complexes
(torch.complex64, torch.complex128): torch.complex128,
# Mixed float and complex
(torch.float32, torch.complex64): torch.complex64,
(torch.float32, torch.complex128): torch.complex128,
(torch.float64, torch.complex64): torch.complex128,
(torch.float64, torch.complex128): torch.complex128,
}
_promotion_table.update({(b, a): c for (a, b), c in _promotion_table.items()})
_promotion_table.update({(a, a): a for a in _array_api_dtypes})
def _two_arg(f):
@_wraps(f)
def _f(x1, x2, /, **kwargs):
x1, x2 = _fix_promotion(x1, x2)
return f(x1, x2, **kwargs)
if _f.__doc__ is None:
_f.__doc__ = f"""\
Array API compatibility wrapper for torch.{f.__name__}.
See the corresponding PyTorch documentation and/or the array API specification
for more details.
"""
return _f
def _fix_promotion(x1, x2, only_scalar=True):
if not isinstance(x1, torch.Tensor) or not isinstance(x2, torch.Tensor):
return x1, x2
if x1.dtype not in _array_api_dtypes or x2.dtype not in _array_api_dtypes:
return x1, x2
# If an argument is 0-D pytorch downcasts the other argument
if not only_scalar or x1.shape == ():
dtype = result_type(x1, x2)
x2 = x2.to(dtype)
if not only_scalar or x2.shape == ():
dtype = result_type(x1, x2)
x1 = x1.to(dtype)
return x1, x2
_py_scalars = (bool, int, float, complex)
def result_type(
*arrays_and_dtypes: Array | DType | bool | int | float | complex
) -> DType:
num = len(arrays_and_dtypes)
if num == 0:
raise ValueError("At least one array or dtype must be provided")
elif num == 1:
x = arrays_and_dtypes[0]
if isinstance(x, torch.dtype):
return x
return x.dtype
if num == 2:
x, y = arrays_and_dtypes
return _result_type(x, y)
else:
# sort scalars so that they are treated last
scalars, others = [], []
for x in arrays_and_dtypes:
if isinstance(x, _py_scalars):
scalars.append(x)
else:
others.append(x)
if not others:
raise ValueError("At least one array or dtype must be provided")
# combine left-to-right
return _reduce(_result_type, others + scalars)
def _result_type(
x: Array | DType | bool | int | float | complex,
y: Array | DType | bool | int | float | complex,
) -> DType:
if not (isinstance(x, _py_scalars) or isinstance(y, _py_scalars)):
xdt = x if isinstance(x, torch.dtype) else x.dtype
ydt = y if isinstance(y, torch.dtype) else y.dtype
try:
return _promotion_table[xdt, ydt]
except KeyError:
pass
# This doesn't result_type(dtype, dtype) for non-array API dtypes
# because torch.result_type only accepts tensors. This does however, allow
# cross-kind promotion.
x = torch.tensor([], dtype=x) if isinstance(x, torch.dtype) else x
y = torch.tensor([], dtype=y) if isinstance(y, torch.dtype) else y
return torch.result_type(x, y)
def can_cast(from_: Union[DType, Array], to: DType, /) -> bool:
if not isinstance(from_, torch.dtype):
from_ = from_.dtype
return torch.can_cast(from_, to)
# Basic renames
bitwise_invert = torch.bitwise_not
newaxis = None
# torch.conj sets the conjugation bit, which breaks conversion to other
# libraries. See https://github.com/data-apis/array-api-compat/issues/173
conj = torch.conj_physical
# Two-arg elementwise functions
# These require a wrapper to do the correct type promotion on 0-D tensors
add = _two_arg(torch.add)
atan2 = _two_arg(torch.atan2)
bitwise_and = _two_arg(torch.bitwise_and)
bitwise_left_shift = _two_arg(torch.bitwise_left_shift)
bitwise_or = _two_arg(torch.bitwise_or)
bitwise_right_shift = _two_arg(torch.bitwise_right_shift)
bitwise_xor = _two_arg(torch.bitwise_xor)
copysign = _two_arg(torch.copysign)
divide = _two_arg(torch.divide)
# Also a rename. torch.equal does not broadcast
equal = _two_arg(torch.eq)
floor_divide = _two_arg(torch.floor_divide)
greater = _two_arg(torch.greater)
greater_equal = _two_arg(torch.greater_equal)
hypot = _two_arg(torch.hypot)
less = _two_arg(torch.less)
less_equal = _two_arg(torch.less_equal)
logaddexp = _two_arg(torch.logaddexp)
# logical functions are not included here because they only accept bool in the
# spec, so type promotion is irrelevant.
maximum = _two_arg(torch.maximum)
minimum = _two_arg(torch.minimum)
multiply = _two_arg(torch.multiply)
not_equal = _two_arg(torch.not_equal)
pow = _two_arg(torch.pow)
remainder = _two_arg(torch.remainder)
subtract = _two_arg(torch.subtract)
def asarray(
obj: (
Array
| bool | int | float | complex
| NestedSequence[bool | int | float | complex]
| SupportsBufferProtocol
),
/,
*,
dtype: DType | None = None,
device: Device | None = None,
copy: bool | None = None,
**kwargs: Any,
) -> Array:
# torch.asarray does not respect input->output device propagation
# https://github.com/pytorch/pytorch/issues/150199
if device is None and isinstance(obj, torch.Tensor):
device = obj.device
return torch.asarray(obj, dtype=dtype, device=device, copy=copy, **kwargs)
# These wrappers are mostly based on the fact that pytorch uses 'dim' instead
# of 'axis'.
# torch.min and torch.max return a tuple and don't support multiple axes https://github.com/pytorch/pytorch/issues/58745
def max(x: Array, /, *, axis: Optional[Union[int, Tuple[int, ...]]] = None, keepdims: bool = False) -> Array:
# https://github.com/pytorch/pytorch/issues/29137
if axis == ():
return torch.clone(x)
return torch.amax(x, axis, keepdims=keepdims)
def min(x: Array, /, *, axis: Optional[Union[int, Tuple[int, ...]]] = None, keepdims: bool = False) -> Array:
# https://github.com/pytorch/pytorch/issues/29137
if axis == ():
return torch.clone(x)
return torch.amin(x, axis, keepdims=keepdims)
clip = get_xp(torch)(_aliases.clip)
unstack = get_xp(torch)(_aliases.unstack)
cumulative_sum = get_xp(torch)(_aliases.cumulative_sum)
cumulative_prod = get_xp(torch)(_aliases.cumulative_prod)
finfo = get_xp(torch)(_aliases.finfo)
iinfo = get_xp(torch)(_aliases.iinfo)
# torch.sort also returns a tuple
# https://github.com/pytorch/pytorch/issues/70921
def sort(x: Array, /, *, axis: int = -1, descending: bool = False, stable: bool = True, **kwargs) -> Array:
return torch.sort(x, dim=axis, descending=descending, stable=stable, **kwargs).values
def _normalize_axes(axis, ndim):
axes = []
if ndim == 0 and axis:
# Better error message in this case
raise IndexError(f"Dimension out of range: {axis[0]}")
lower, upper = -ndim, ndim - 1
for a in axis:
if a < lower or a > upper:
# Match torch error message (e.g., from sum())
raise IndexError(f"Dimension out of range (expected to be in range of [{lower}, {upper}], but got {a}")
if a < 0:
a = a + ndim
if a in axes:
# Use IndexError instead of RuntimeError, and "axis" instead of "dim"
raise IndexError(f"Axis {a} appears multiple times in the list of axes")
axes.append(a)
return sorted(axes)
def _axis_none_keepdims(x, ndim, keepdims):
# Apply keepdims when axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
# Note that this is only valid for the axis=None case.
if keepdims:
for i in range(ndim):
x = torch.unsqueeze(x, 0)
return x
def _reduce_multiple_axes(f, x, axis, keepdims=False, **kwargs):
# Some reductions don't support multiple axes
# (https://github.com/pytorch/pytorch/issues/56586).
axes = _normalize_axes(axis, x.ndim)
for a in reversed(axes):
x = torch.movedim(x, a, -1)
x = torch.flatten(x, -len(axes))
out = f(x, -1, **kwargs)
if keepdims:
for a in axes:
out = torch.unsqueeze(out, a)
return out
def _sum_prod_no_axis(x: Array, dtype: DType | None) -> Array:
"""
Implements `sum(..., axis=())` and `prod(..., axis=())`.
Works around https://github.com/pytorch/pytorch/issues/29137
"""
if dtype is not None:
return x.clone() if dtype == x.dtype else x.to(dtype)
# We can't upcast uint8 according to the spec because there is no
# torch.uint64, so at least upcast to int64 which is what prod does
# when axis=None.
if x.dtype in (torch.uint8, torch.int8, torch.int16, torch.int32):
return x.to(torch.int64)
return x.clone()
def prod(x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
dtype: Optional[DType] = None,
keepdims: bool = False,
**kwargs) -> Array:
if axis == ():
return _sum_prod_no_axis(x, dtype)
# torch.prod doesn't support multiple axes
# (https://github.com/pytorch/pytorch/issues/56586).
if isinstance(axis, tuple):
return _reduce_multiple_axes(torch.prod, x, axis, keepdims=keepdims, dtype=dtype, **kwargs)
if axis is None:
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.prod(x, dtype=dtype, **kwargs)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res
return torch.prod(x, axis, dtype=dtype, keepdims=keepdims, **kwargs)
def sum(x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
dtype: Optional[DType] = None,
keepdims: bool = False,
**kwargs) -> Array:
if axis == ():
return _sum_prod_no_axis(x, dtype)
if axis is None:
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.sum(x, dtype=dtype, **kwargs)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res
return torch.sum(x, axis, dtype=dtype, keepdims=keepdims, **kwargs)
def any(x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
keepdims: bool = False,
**kwargs) -> Array:
if axis == ():
return x.to(torch.bool)
# torch.any doesn't support multiple axes
# (https://github.com/pytorch/pytorch/issues/56586).
if isinstance(axis, tuple):
res = _reduce_multiple_axes(torch.any, x, axis, keepdims=keepdims, **kwargs)
return res.to(torch.bool)
if axis is None:
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.any(x, **kwargs)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res.to(torch.bool)
# torch.any doesn't return bool for uint8
return torch.any(x, axis, keepdims=keepdims).to(torch.bool)
def all(x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
keepdims: bool = False,
**kwargs) -> Array:
if axis == ():
return x.to(torch.bool)
# torch.all doesn't support multiple axes
# (https://github.com/pytorch/pytorch/issues/56586).
if isinstance(axis, tuple):
res = _reduce_multiple_axes(torch.all, x, axis, keepdims=keepdims, **kwargs)
return res.to(torch.bool)
if axis is None:
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.all(x, **kwargs)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res.to(torch.bool)
# torch.all doesn't return bool for uint8
return torch.all(x, axis, keepdims=keepdims).to(torch.bool)
def mean(x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
keepdims: bool = False,
**kwargs) -> Array:
# https://github.com/pytorch/pytorch/issues/29137
if axis == ():
return torch.clone(x)
if axis is None:
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.mean(x, **kwargs)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res
return torch.mean(x, axis, keepdims=keepdims, **kwargs)
def std(x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
correction: Union[int, float] = 0.0,
keepdims: bool = False,
**kwargs) -> Array:
# Note, float correction is not supported
# https://github.com/pytorch/pytorch/issues/61492. We don't try to
# implement it here for now.
if isinstance(correction, float):
_correction = int(correction)
if correction != _correction:
raise NotImplementedError("float correction in torch std() is not yet supported")
else:
_correction = correction
# https://github.com/pytorch/pytorch/issues/29137
if axis == ():
return torch.zeros_like(x)
if isinstance(axis, int):
axis = (axis,)
if axis is None:
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.std(x, tuple(range(x.ndim)), correction=_correction, **kwargs)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res
return torch.std(x, axis, correction=_correction, keepdims=keepdims, **kwargs)
def var(x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
correction: Union[int, float] = 0.0,
keepdims: bool = False,
**kwargs) -> Array:
# Note, float correction is not supported
# https://github.com/pytorch/pytorch/issues/61492. We don't try to
# implement it here for now.
# if isinstance(correction, float):
# correction = int(correction)
# https://github.com/pytorch/pytorch/issues/29137
if axis == ():
return torch.zeros_like(x)
if isinstance(axis, int):
axis = (axis,)
if axis is None:
# torch doesn't support keepdims with axis=None
# (https://github.com/pytorch/pytorch/issues/71209)
res = torch.var(x, tuple(range(x.ndim)), correction=correction, **kwargs)
res = _axis_none_keepdims(res, x.ndim, keepdims)
return res
return torch.var(x, axis, correction=correction, keepdims=keepdims, **kwargs)
# torch.concat doesn't support dim=None
# https://github.com/pytorch/pytorch/issues/70925
def concat(arrays: Union[Tuple[Array, ...], List[Array]],
/,
*,
axis: Optional[int] = 0,
**kwargs) -> Array:
if axis is None:
arrays = tuple(ar.flatten() for ar in arrays)
axis = 0
return torch.concat(arrays, axis, **kwargs)
# torch.squeeze only accepts int dim and doesn't require it
# https://github.com/pytorch/pytorch/issues/70924. Support for tuple dim was
# added at https://github.com/pytorch/pytorch/pull/89017.
def squeeze(x: Array, /, axis: Union[int, Tuple[int, ...]]) -> Array:
if isinstance(axis, int):
axis = (axis,)
for a in axis:
if x.shape[a] != 1:
raise ValueError("squeezed dimensions must be equal to 1")
axes = _normalize_axes(axis, x.ndim)
# Remove this once pytorch 1.14 is released with the above PR #89017.
sequence = [a - i for i, a in enumerate(axes)]
for a in sequence:
x = torch.squeeze(x, a)
return x
# torch.broadcast_to uses size instead of shape
def broadcast_to(x: Array, /, shape: Tuple[int, ...], **kwargs) -> Array:
return torch.broadcast_to(x, shape, **kwargs)
# torch.permute uses dims instead of axes
def permute_dims(x: Array, /, axes: Tuple[int, ...]) -> Array:
return torch.permute(x, axes)
# The axis parameter doesn't work for flip() and roll()
# https://github.com/pytorch/pytorch/issues/71210. Also torch.flip() doesn't
# accept axis=None
def flip(x: Array, /, *, axis: Optional[Union[int, Tuple[int, ...]]] = None, **kwargs) -> Array:
if axis is None:
axis = tuple(range(x.ndim))
# torch.flip doesn't accept dim as an int but the method does
# https://github.com/pytorch/pytorch/issues/18095
return x.flip(axis, **kwargs)
def roll(x: Array, /, shift: Union[int, Tuple[int, ...]], *, axis: Optional[Union[int, Tuple[int, ...]]] = None, **kwargs) -> Array:
return torch.roll(x, shift, axis, **kwargs)
def nonzero(x: Array, /, **kwargs) -> Tuple[Array, ...]:
if x.ndim == 0:
raise ValueError("nonzero() does not support zero-dimensional arrays")
return torch.nonzero(x, as_tuple=True, **kwargs)
# torch uses `dim` instead of `axis`
def diff(
x: Array,
/,
*,
axis: int = -1,
n: int = 1,
prepend: Optional[Array] = None,
append: Optional[Array] = None,
) -> Array:
return torch.diff(x, dim=axis, n=n, prepend=prepend, append=append)
# torch uses `dim` instead of `axis`, does not have keepdims
def count_nonzero(
x: Array,
/,
*,
axis: Optional[Union[int, Tuple[int, ...]]] = None,
keepdims: bool = False,
) -> Array:
result = torch.count_nonzero(x, dim=axis)
if keepdims:
if isinstance(axis, int):
return result.unsqueeze(axis)
elif isinstance(axis, tuple):
n_axis = [x.ndim + ax if ax < 0 else ax for ax in axis]
sh = [1 if i in n_axis else x.shape[i] for i in range(x.ndim)]
return torch.reshape(result, sh)
return _axis_none_keepdims(result, x.ndim, keepdims)
else:
return result
# "repeat" is torch.repeat_interleave; also the dim argument
def repeat(x: Array, repeats: int | Array, /, *, axis: int | None = None) -> Array:
return torch.repeat_interleave(x, repeats, axis)
def where(
condition: Array,
x1: Array | bool | int | float | complex,
x2: Array | bool | int | float | complex,
/,
) -> Array:
x1, x2 = _fix_promotion(x1, x2)
return torch.where(condition, x1, x2)
# torch.reshape doesn't have the copy keyword
def reshape(x: Array,
/,
shape: Tuple[int, ...],
*,
copy: Optional[bool] = None,
**kwargs) -> Array:
if copy is not None:
raise NotImplementedError("torch.reshape doesn't yet support the copy keyword")
return torch.reshape(x, shape, **kwargs)
# torch.arange doesn't support returning empty arrays
# (https://github.com/pytorch/pytorch/issues/70915), and doesn't support some
# keyword argument combinations
# (https://github.com/pytorch/pytorch/issues/70914)
def arange(start: Union[int, float],
/,
stop: Optional[Union[int, float]] = None,
step: Union[int, float] = 1,
*,
dtype: Optional[DType] = None,
device: Optional[Device] = None,
**kwargs) -> Array:
if stop is None:
start, stop = 0, start
if step > 0 and stop <= start or step < 0 and stop >= start:
if dtype is None:
if _builtin_all(isinstance(i, int) for i in [start, stop, step]):
dtype = torch.int64
else:
dtype = torch.float32
return torch.empty(0, dtype=dtype, device=device, **kwargs)
return torch.arange(start, stop, step, dtype=dtype, device=device, **kwargs)
# torch.eye does not accept None as a default for the second argument and
# doesn't support off-diagonals (https://github.com/pytorch/pytorch/issues/70910)
def eye(n_rows: int,
n_cols: Optional[int] = None,
/,
*,
k: int = 0,
dtype: Optional[DType] = None,
device: Optional[Device] = None,
**kwargs) -> Array:
if n_cols is None:
n_cols = n_rows
z = torch.zeros(n_rows, n_cols, dtype=dtype, device=device, **kwargs)
if abs(k) <= n_rows + n_cols:
z.diagonal(k).fill_(1)
return z
# torch.linspace doesn't have the endpoint parameter
def linspace(start: Union[int, float],
stop: Union[int, float],
/,
num: int,
*,
dtype: Optional[DType] = None,
device: Optional[Device] = None,
endpoint: bool = True,
**kwargs) -> Array:
if not endpoint:
return torch.linspace(start, stop, num+1, dtype=dtype, device=device, **kwargs)[:-1]
return torch.linspace(start, stop, num, dtype=dtype, device=device, **kwargs)
# torch.full does not accept an int size
# https://github.com/pytorch/pytorch/issues/70906
def full(shape: Union[int, Tuple[int, ...]],
fill_value: bool | int | float | complex,
*,
dtype: Optional[DType] = None,
device: Optional[Device] = None,
**kwargs) -> Array:
if isinstance(shape, int):
shape = (shape,)
return torch.full(shape, fill_value, dtype=dtype, device=device, **kwargs)
# ones, zeros, and empty do not accept shape as a keyword argument
def ones(shape: Union[int, Tuple[int, ...]],
*,
dtype: Optional[DType] = None,
device: Optional[Device] = None,
**kwargs) -> Array:
return torch.ones(shape, dtype=dtype, device=device, **kwargs)
def zeros(shape: Union[int, Tuple[int, ...]],
*,
dtype: Optional[DType] = None,
device: Optional[Device] = None,
**kwargs) -> Array:
return torch.zeros(shape, dtype=dtype, device=device, **kwargs)
def empty(shape: Union[int, Tuple[int, ...]],
*,
dtype: Optional[DType] = None,
device: Optional[Device] = None,
**kwargs) -> Array:
return torch.empty(shape, dtype=dtype, device=device, **kwargs)
# tril and triu do not call the keyword argument k
def tril(x: Array, /, *, k: int = 0) -> Array:
return torch.tril(x, k)
def triu(x: Array, /, *, k: int = 0) -> Array:
return torch.triu(x, k)
# Functions that aren't in torch https://github.com/pytorch/pytorch/issues/58742
def expand_dims(x: Array, /, *, axis: int = 0) -> Array:
return torch.unsqueeze(x, axis)
def astype(
x: Array,
dtype: DType,
/,
*,
copy: bool = True,
device: Optional[Device] = None,
) -> Array:
if device is not None:
return x.to(device, dtype=dtype, copy=copy)
return x.to(dtype=dtype, copy=copy)
def broadcast_arrays(*arrays: Array) -> List[Array]:
shape = torch.broadcast_shapes(*[a.shape for a in arrays])
return [torch.broadcast_to(a, shape) for a in arrays]
# Note that these named tuples aren't actually part of the standard namespace,
# but I don't see any issue with exporting the names here regardless.
from ..common._aliases import (UniqueAllResult, UniqueCountsResult,
UniqueInverseResult)
# https://github.com/pytorch/pytorch/issues/70920
def unique_all(x: Array) -> UniqueAllResult:
# torch.unique doesn't support returning indices.
# https://github.com/pytorch/pytorch/issues/36748. The workaround
# suggested in that issue doesn't actually function correctly (it relies
# on non-deterministic behavior of scatter()).
raise NotImplementedError("unique_all() not yet implemented for pytorch (see https://github.com/pytorch/pytorch/issues/36748)")
# values, inverse_indices, counts = torch.unique(x, return_counts=True, return_inverse=True)
# # torch.unique incorrectly gives a 0 count for nan values.
# # https://github.com/pytorch/pytorch/issues/94106
# counts[torch.isnan(values)] = 1
# return UniqueAllResult(values, indices, inverse_indices, counts)
def unique_counts(x: Array) -> UniqueCountsResult:
values, counts = torch.unique(x, return_counts=True)
# torch.unique incorrectly gives a 0 count for nan values.
# https://github.com/pytorch/pytorch/issues/94106
counts[torch.isnan(values)] = 1
return UniqueCountsResult(values, counts)
def unique_inverse(x: Array) -> UniqueInverseResult:
values, inverse = torch.unique(x, return_inverse=True)
return UniqueInverseResult(values, inverse)
def unique_values(x: Array) -> Array:
return torch.unique(x)
def matmul(x1: Array, x2: Array, /, **kwargs) -> Array:
# torch.matmul doesn't type promote (but differently from _fix_promotion)
x1, x2 = _fix_promotion(x1, x2, only_scalar=False)
return torch.matmul(x1, x2, **kwargs)
matrix_transpose = get_xp(torch)(_aliases.matrix_transpose)
_vecdot = get_xp(torch)(_aliases.vecdot)
def vecdot(x1: Array, x2: Array, /, *, axis: int = -1) -> Array:
x1, x2 = _fix_promotion(x1, x2, only_scalar=False)
return _vecdot(x1, x2, axis=axis)
# torch.tensordot uses dims instead of axes
def tensordot(
x1: Array,
x2: Array,
/,
*,
axes: Union[int, Tuple[Sequence[int], Sequence[int]]] = 2,
**kwargs,
) -> Array:
# Note: torch.tensordot fails with integer dtypes when there is only 1
# element in the axis (https://github.com/pytorch/pytorch/issues/84530).
x1, x2 = _fix_promotion(x1, x2, only_scalar=False)
return torch.tensordot(x1, x2, dims=axes, **kwargs)
def isdtype(
dtype: DType, kind: Union[DType, str, Tuple[Union[DType, str], ...]],
*, _tuple=True, # Disallow nested tuples
) -> bool:
"""
Returns a boolean indicating whether a provided dtype is of a specified data type ``kind``.
Note that outside of this function, this compat library does not yet fully
support complex numbers.
See
https://data-apis.org/array-api/latest/API_specification/generated/array_api.isdtype.html
for more details
"""
if isinstance(kind, tuple) and _tuple:
return _builtin_any(isdtype(dtype, k, _tuple=False) for k in kind)
elif isinstance(kind, str):
if kind == 'bool':
return dtype == torch.bool
elif kind == 'signed integer':
return dtype in _int_dtypes and dtype.is_signed
elif kind == 'unsigned integer':
return dtype in _int_dtypes and not dtype.is_signed
elif kind == 'integral':
return dtype in _int_dtypes
elif kind == 'real floating':
return dtype.is_floating_point
elif kind == 'complex floating':
return dtype.is_complex
elif kind == 'numeric':
return isdtype(dtype, ('integral', 'real floating', 'complex floating'))
else:
raise ValueError(f"Unrecognized data type kind: {kind!r}")
else:
return dtype == kind
def take(x: Array, indices: Array, /, *, axis: Optional[int] = None, **kwargs) -> Array:
if axis is None:
if x.ndim != 1:
raise ValueError("axis must be specified when ndim > 1")
axis = 0
return torch.index_select(x, axis, indices, **kwargs)
def take_along_axis(x: Array, indices: Array, /, *, axis: int = -1) -> Array:
return torch.take_along_dim(x, indices, dim=axis)
def sign(x: Array, /) -> Array:
# torch sign() does not support complex numbers and does not propagate
# nans. See https://github.com/data-apis/array-api-compat/issues/136
if x.dtype.is_complex:
out = x/torch.abs(x)
# sign(0) = 0 but the above formula would give nan
out[x == 0+0j] = 0+0j
return out
else:
out = torch.sign(x)
if x.dtype.is_floating_point:
out[torch.isnan(x)] = torch.nan
return out
def meshgrid(*arrays: Array, indexing: Literal['xy', 'ij'] = 'xy') -> List[Array]:
# enforce the default of 'xy'
# TODO: is the return type a list or a tuple
return list(torch.meshgrid(*arrays, indexing='xy'))
__all__ = ['__array_namespace_info__', 'asarray', 'result_type', 'can_cast',
'permute_dims', 'bitwise_invert', 'newaxis', 'conj', 'add',
'atan2', 'bitwise_and', 'bitwise_left_shift', 'bitwise_or',
'bitwise_right_shift', 'bitwise_xor', 'copysign', 'count_nonzero',
'diff', 'divide',
'equal', 'floor_divide', 'greater', 'greater_equal', 'hypot',
'less', 'less_equal', 'logaddexp', 'maximum', 'minimum',
'multiply', 'not_equal', 'pow', 'remainder', 'subtract', 'max',
'min', 'clip', 'unstack', 'cumulative_sum', 'cumulative_prod', 'sort', 'prod', 'sum',
'any', 'all', 'mean', 'std', 'var', 'concat', 'squeeze',
'broadcast_to', 'flip', 'roll', 'nonzero', 'where', 'reshape',
'arange', 'eye', 'linspace', 'full', 'ones', 'zeros', 'empty',
'tril', 'triu', 'expand_dims', 'astype', 'broadcast_arrays',
'UniqueAllResult', 'UniqueCountsResult', 'UniqueInverseResult',
'unique_all', 'unique_counts', 'unique_inverse', 'unique_values',
'matmul', 'matrix_transpose', 'vecdot', 'tensordot', 'isdtype',
'take', 'take_along_axis', 'sign', 'finfo', 'iinfo', 'repeat', 'meshgrid']
_all_ignore = ['torch', 'get_xp']
@@ -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
@@ -0,0 +1,3 @@
__all__ = ["Array", "Device", "DType"]
from torch import device as Device, dtype as DType, Tensor as Array
@@ -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']
@@ -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__
@@ -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",
]
@@ -0,0 +1,171 @@
"""Delegation to existing implementations for Public API Functions."""
from collections.abc import Sequence
from types import ModuleType
from typing import Literal
from ._lib import _funcs
from ._lib._utils._compat import (
array_namespace,
is_cupy_namespace,
is_dask_namespace,
is_jax_namespace,
is_numpy_namespace,
is_pydata_sparse_namespace,
is_torch_namespace,
)
from ._lib._utils._helpers import asarrays
from ._lib._utils._typing import Array
__all__ = ["isclose", "pad"]
def isclose(
a: Array | complex,
b: Array | complex,
*,
rtol: float = 1e-05,
atol: float = 1e-08,
equal_nan: bool = False,
xp: ModuleType | None = None,
) -> Array:
"""
Return a boolean array where two arrays are element-wise equal within a tolerance.
The tolerance values are positive, typically very small numbers. The relative
difference ``(rtol * abs(b))`` and the absolute difference `atol` are added together
to compare against the absolute difference between `a` and `b`.
NaNs are treated as equal if they are in the same place and if ``equal_nan=True``.
Infs are treated as equal if they are in the same place and of the same sign in both
arrays.
Parameters
----------
a, b : Array | int | float | complex | bool
Input objects to compare. At least one must be an array.
rtol : array_like, optional
The relative tolerance parameter (see Notes).
atol : array_like, optional
The absolute tolerance parameter (see Notes).
equal_nan : bool, optional
Whether to compare NaN's as equal. If True, NaN's in `a` will be considered
equal to NaN's in `b` in the output array.
xp : array_namespace, optional
The standard-compatible namespace for `a` and `b`. Default: infer.
Returns
-------
Array
A boolean array of shape broadcasted from `a` and `b`, containing ``True`` where
`a` is close to `b`, and ``False`` otherwise.
Warnings
--------
The default `atol` is not appropriate for comparing numbers with magnitudes much
smaller than one (see notes).
See Also
--------
math.isclose : Similar function in stdlib for Python scalars.
Notes
-----
For finite values, `isclose` uses the following equation to test whether two
floating point values are equivalent::
absolute(a - b) <= (atol + rtol * absolute(b))
Unlike the built-in `math.isclose`,
the above equation is not symmetric in `a` and `b`,
so that ``isclose(a, b)`` might be different from ``isclose(b, a)`` in some rare
cases.
The default value of `atol` is not appropriate when the reference value `b` has
magnitude smaller than one. For example, it is unlikely that ``a = 1e-9`` and
``b = 2e-9`` should be considered "close", yet ``isclose(1e-9, 2e-9)`` is ``True``
with default settings. Be sure to select `atol` for the use case at hand, especially
for defining the threshold below which a non-zero value in `a` will be considered
"close" to a very small or zero value in `b`.
The comparison of `a` and `b` uses standard broadcasting, which means that `a` and
`b` need not have the same shape in order for ``isclose(a, b)`` to evaluate to
``True``.
`isclose` is not defined for non-numeric data types.
``bool`` is considered a numeric data-type for this purpose.
"""
xp = array_namespace(a, b) if xp is None else xp
if (
is_numpy_namespace(xp)
or is_cupy_namespace(xp)
or is_dask_namespace(xp)
or is_jax_namespace(xp)
):
return xp.isclose(a, b, rtol=rtol, atol=atol, equal_nan=equal_nan)
if is_torch_namespace(xp):
a, b = asarrays(a, b, xp=xp) # Array API 2024.12 support
return xp.isclose(a, b, rtol=rtol, atol=atol, equal_nan=equal_nan)
return _funcs.isclose(a, b, rtol=rtol, atol=atol, equal_nan=equal_nan, xp=xp)
def pad(
x: Array,
pad_width: int | tuple[int, int] | Sequence[tuple[int, int]],
mode: Literal["constant"] = "constant",
*,
constant_values: complex = 0,
xp: ModuleType | None = None,
) -> Array:
"""
Pad the input array.
Parameters
----------
x : array
Input array.
pad_width : int or tuple of ints or sequence of pairs of ints
Pad the input array with this many elements from each side.
If a sequence of tuples, ``[(before_0, after_0), ... (before_N, after_N)]``,
each pair applies to the corresponding axis of ``x``.
A single tuple, ``(before, after)``, is equivalent to a list of ``x.ndim``
copies of this tuple.
mode : str, optional
Only "constant" mode is currently supported, which pads with
the value passed to `constant_values`.
constant_values : python scalar, optional
Use this value to pad the input. Default is zero.
xp : array_namespace, optional
The standard-compatible namespace for `x`. Default: infer.
Returns
-------
array
The input array,
padded with ``pad_width`` elements equal to ``constant_values``.
"""
xp = array_namespace(x) if xp is None else xp
if mode != "constant":
msg = "Only `'constant'` mode is currently supported"
raise NotImplementedError(msg)
if (
is_numpy_namespace(xp)
or is_cupy_namespace(xp)
or is_jax_namespace(xp)
or is_pydata_sparse_namespace(xp)
):
return xp.pad(x, pad_width, mode, constant_values=constant_values)
# https://github.com/pytorch/pytorch/blob/cf76c05b4dc629ac989d1fb8e789d4fac04a095a/torch/_numpy/_funcs_impl.py#L2045-L2056
if is_torch_namespace(xp):
pad_width = xp.asarray(pad_width)
pad_width = xp.broadcast_to(pad_width, (x.ndim, 2))
pad_width = xp.flip(pad_width, axis=(0,)).flatten()
return xp.nn.functional.pad(x, tuple(pad_width), value=constant_values) # type: ignore[arg-type] # pyright: ignore[reportArgumentType]
return _funcs.pad(x, pad_width, constant_values=constant_values, xp=xp)
@@ -0,0 +1 @@
"""Internals of array-api-extra."""
@@ -0,0 +1,463 @@
"""Update operations for read-only arrays."""
from __future__ import annotations
import operator
from collections.abc import Callable
from enum import Enum
from types import ModuleType
from typing import TYPE_CHECKING, ClassVar, cast
from ._utils import _compat
from ._utils._compat import (
array_namespace,
is_dask_array,
is_jax_array,
is_torch_array,
is_writeable_array,
)
from ._utils._helpers import meta_namespace
from ._utils._typing import Array, SetIndex
if TYPE_CHECKING: # pragma: no cover
# TODO import from typing (requires Python >=3.11)
from typing_extensions import Self
class _AtOp(Enum):
"""Operations for use in `xpx.at`."""
SET = "set"
ADD = "add"
SUBTRACT = "subtract"
MULTIPLY = "multiply"
DIVIDE = "divide"
POWER = "power"
MIN = "min"
MAX = "max"
# @override from Python 3.12
def __str__(self) -> str: # type: ignore[explicit-override] # pyright: ignore[reportImplicitOverride]
"""
Return string representation (useful for pytest logs).
Returns
-------
str
The operation's name.
"""
return self.value
class Undef(Enum):
"""Sentinel for undefined values."""
UNDEF = 0
_undef = Undef.UNDEF
class at: # pylint: disable=invalid-name # numpydoc ignore=PR02
"""
Update operations for read-only arrays.
This implements ``jax.numpy.ndarray.at`` for all writeable
backends (those that support ``__setitem__``) and routes
to the ``.at[]`` method for JAX arrays.
Parameters
----------
x : array
Input array.
idx : index, optional
Only `array API standard compliant indices
<https://data-apis.org/array-api/latest/API_specification/indexing.html>`_
are supported.
You may use two alternate syntaxes::
>>> import array_api_extra as xpx
>>> xpx.at(x, idx).set(value) # or add(value), etc.
>>> xpx.at(x)[idx].set(value)
copy : bool, optional
None (default)
The array parameter *may* be modified in place if it is
possible and beneficial for performance.
You should not reuse it after calling this function.
True
Ensure that the inputs are not modified.
False
Ensure that the update operation writes back to the input.
Raise ``ValueError`` if a copy cannot be avoided.
xp : array_namespace, optional
The standard-compatible namespace for `x`. Default: infer.
Returns
-------
Updated input array.
Warnings
--------
(a) When you omit the ``copy`` parameter, you should never reuse the parameter
array later on; ideally, you should reassign it immediately::
>>> import array_api_extra as xpx
>>> x = xpx.at(x, 0).set(2)
The above best practice pattern ensures that the behaviour won't change depending
on whether ``x`` is writeable or not, as the original ``x`` object is dereferenced
as soon as ``xpx.at`` returns; this way there is no risk to accidentally update it
twice.
On the reverse, the anti-pattern below must be avoided, as it will result in
different behaviour on read-only versus writeable arrays::
>>> x = xp.asarray([0, 0, 0])
>>> y = xpx.at(x, 0).set(2)
>>> z = xpx.at(x, 1).set(3)
In the above example, both calls to ``xpx.at`` update ``x`` in place *if possible*.
This causes the behaviour to diverge depending on whether ``x`` is writeable or not:
- If ``x`` is writeable, then after the snippet above you'll have
``x == y == z == [2, 3, 0]``
- If ``x`` is read-only, then you'll end up with
``x == [0, 0, 0]``, ``y == [2, 0, 0]`` and ``z == [0, 3, 0]``.
The correct pattern to use if you want diverging outputs from the same input is
to enforce copies::
>>> x = xp.asarray([0, 0, 0])
>>> y = xpx.at(x, 0).set(2, copy=True) # Never updates x
>>> z = xpx.at(x, 1).set(3) # May or may not update x in place
>>> del x # avoid accidental reuse of x as we don't know its state anymore
(b) The array API standard does not support integer array indices.
The behaviour of update methods when the index is an array of integers is
undefined and will vary between backends; this is particularly true when the
index contains multiple occurrences of the same index, e.g.::
>>> import numpy as np
>>> import jax.numpy as jnp
>>> import array_api_extra as xpx
>>> xpx.at(np.asarray([123]), np.asarray([0, 0])).add(1)
array([124])
>>> xpx.at(jnp.asarray([123]), jnp.asarray([0, 0])).add(1)
Array([125], dtype=int32)
See Also
--------
jax.numpy.ndarray.at : Equivalent array method in JAX.
Notes
-----
`sparse <https://sparse.pydata.org/>`_, as well as read-only arrays from libraries
not explicitly covered by ``array-api-compat``, are not supported by update
methods.
Boolean masks are supported on Dask and jitted JAX arrays exclusively
when `idx` has the same shape as `x` and `y` is 0-dimensional.
Note that this support is not available in JAX's native
``x.at[mask].set(y)``.
This pattern::
>>> mask = m(x)
>>> x[mask] = f(x[mask])
Can't be replaced by `at`, as it won't work on Dask and JAX inside jax.jit::
>>> mask = m(x)
>>> x = xpx.at(x, mask).set(f(x[mask]) # Crash on Dask and jax.jit
You should instead use::
>>> x = xp.where(m(x), f(x), x)
Examples
--------
Given either of these equivalent expressions::
>>> import array_api_extra as xpx
>>> x = xpx.at(x)[1].add(2)
>>> x = xpx.at(x, 1).add(2)
If x is a JAX array, they are the same as::
>>> x = x.at[1].add(2)
If x is a read-only NumPy array, they are the same as::
>>> x = x.copy()
>>> x[1] += 2
For other known backends, they are the same as::
>>> x[1] += 2
"""
_x: Array
_idx: SetIndex | Undef
__slots__: ClassVar[tuple[str, ...]] = ("_idx", "_x")
def __init__(
self, x: Array, idx: SetIndex | Undef = _undef, /
) -> None: # numpydoc ignore=GL08
self._x = x
self._idx = idx
def __getitem__(self, idx: SetIndex, /) -> Self: # numpydoc ignore=PR01,RT01
"""
Allow for the alternate syntax ``at(x)[start:stop:step]``.
It looks prettier than ``at(x, slice(start, stop, step))``
and feels more intuitive coming from the JAX documentation.
"""
if self._idx is not _undef:
msg = "Index has already been set"
raise ValueError(msg)
return type(self)(self._x, idx)
def _op(
self,
at_op: _AtOp,
in_place_op: Callable[[Array, Array | complex], Array] | None,
out_of_place_op: Callable[[Array, Array], Array] | None,
y: Array | complex,
/,
copy: bool | None,
xp: ModuleType | None,
) -> Array:
"""
Implement all update operations.
Parameters
----------
at_op : _AtOp
Method of JAX's Array.at[].
in_place_op : Callable[[Array, Array | complex], Array] | None
In-place operation to apply on mutable backends::
x[idx] = in_place_op(x[idx], y)
If None::
x[idx] = y
out_of_place_op : Callable[[Array, Array], Array] | None
Out-of-place operation to apply when idx is a boolean mask and the backend
doesn't support in-place updates::
x = xp.where(idx, out_of_place_op(x, y), x)
If None::
x = xp.where(idx, y, x)
y : array or complex
Right-hand side of the operation.
copy : bool or None
Whether to copy the input array. See the class docstring for details.
xp : array_namespace, optional
The array namespace for the input array. Default: infer.
Returns
-------
Array
Updated `x`.
"""
from ._funcs import apply_where # pylint: disable=cyclic-import
x, idx = self._x, self._idx
xp = array_namespace(x, y) if xp is None else xp
if isinstance(idx, Undef):
msg = (
"Index has not been set.\n"
"Usage: either\n"
" at(x, idx).set(value)\n"
"or\n"
" at(x)[idx].set(value)\n"
"(same for all other methods)."
)
raise ValueError(msg)
if copy not in (True, False, None):
msg = f"copy must be True, False, or None; got {copy!r}"
raise ValueError(msg)
writeable = None if copy else is_writeable_array(x)
# JAX inside jax.jit doesn't support in-place updates with boolean
# masks; Dask exclusively supports __setitem__ but not iops.
# We can handle the common special case of 0-dimensional y
# with where(idx, y, x) instead.
if (
(is_dask_array(idx) or is_jax_array(idx))
and idx.dtype == xp.bool
and idx.shape == x.shape
):
y_xp = xp.asarray(y, dtype=x.dtype, device=_compat.device(x))
if y_xp.ndim == 0:
if out_of_place_op: # add(), subtract(), ...
# suppress inf warnings on Dask
out = apply_where(
idx, (x, y_xp), out_of_place_op, fill_value=x, xp=xp
)
# Undo int->float promotion on JAX after _AtOp.DIVIDE
out = xp.astype(out, x.dtype, copy=False)
else: # set()
out = xp.where(idx, y_xp, x)
if copy is False:
x[()] = out
return x
return out
# else: this will work on eager JAX and crash on jax.jit and Dask
if copy or (copy is None and not writeable):
if is_jax_array(x):
# Use JAX's at[]
func = cast(
Callable[[Array | complex], Array],
getattr(x.at[idx], at_op.value), # type: ignore[attr-defined] # pyright: ignore[reportAttributeAccessIssue,reportUnknownArgumentType]
)
out = func(y)
# Undo int->float promotion on JAX after _AtOp.DIVIDE
return xp.astype(out, x.dtype, copy=False)
# Emulate at[] behaviour for non-JAX arrays
# with a copy followed by an update
x = xp.asarray(x, copy=True)
# A copy of a read-only numpy array is writeable
# Note: this assumes that a copy of a writeable array is writeable
assert not writeable
writeable = None
if writeable is None:
writeable = is_writeable_array(x)
if not writeable:
# sparse crashes here
msg = f"Can't update read-only array {x}"
raise ValueError(msg)
# Work around bug in PyTorch where __setitem__ doesn't
# always support mismatched dtypes
# https://github.com/pytorch/pytorch/issues/150017
if is_torch_array(y):
y = xp.astype(y, x.dtype, copy=False)
# Backends without boolean indexing (other than JAX) crash here
if in_place_op: # add(), subtract(), ...
x[idx] = in_place_op(x[idx], y)
else: # set()
x[idx] = y
return x
def set(
self,
y: Array | complex,
/,
copy: bool | None = None,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""Apply ``x[idx] = y`` and return the update array."""
return self._op(_AtOp.SET, None, None, y, copy=copy, xp=xp)
def add(
self,
y: Array | complex,
/,
copy: bool | None = None,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""Apply ``x[idx] += y`` and return the updated array."""
# Note for this and all other methods based on _iop:
# operator.iadd and operator.add subtly differ in behaviour, as
# only iadd will trigger exceptions when y has an incompatible dtype.
return self._op(_AtOp.ADD, operator.iadd, operator.add, y, copy=copy, xp=xp)
def subtract(
self,
y: Array | complex,
/,
copy: bool | None = None,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""Apply ``x[idx] -= y`` and return the updated array."""
return self._op(
_AtOp.SUBTRACT, operator.isub, operator.sub, y, copy=copy, xp=xp
)
def multiply(
self,
y: Array | complex,
/,
copy: bool | None = None,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""Apply ``x[idx] *= y`` and return the updated array."""
return self._op(
_AtOp.MULTIPLY, operator.imul, operator.mul, y, copy=copy, xp=xp
)
def divide(
self,
y: Array | complex,
/,
copy: bool | None = None,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""Apply ``x[idx] /= y`` and return the updated array."""
return self._op(
_AtOp.DIVIDE, operator.itruediv, operator.truediv, y, copy=copy, xp=xp
)
def power(
self,
y: Array | complex,
/,
copy: bool | None = None,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""Apply ``x[idx] **= y`` and return the updated array."""
return self._op(_AtOp.POWER, operator.ipow, operator.pow, y, copy=copy, xp=xp)
def min(
self,
y: Array | complex,
/,
copy: bool | None = None,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""Apply ``x[idx] = minimum(x[idx], y)`` and return the updated array."""
# On Dask, this function runs on the chunks, so we need to determine the
# namespace that Dask is wrapping.
# Note that da.minimum _incidentally_ works on NumPy, CuPy, and sparse
# thanks to all these meta-namespaces implementing the __array_ufunc__
# interface, but there's no guarantee that it will work for other
# wrapped libraries in the future.
xp = array_namespace(self._x) if xp is None else xp
mxp = meta_namespace(self._x, xp=xp)
y = xp.asarray(y)
return self._op(_AtOp.MIN, mxp.minimum, mxp.minimum, y, copy=copy, xp=xp)
def max(
self,
y: Array | complex,
/,
copy: bool | None = None,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""Apply ``x[idx] = maximum(x[idx], y)`` and return the updated array."""
# See note on min()
xp = array_namespace(self._x) if xp is None else xp
mxp = meta_namespace(self._x, xp=xp)
y = xp.asarray(y)
return self._op(_AtOp.MAX, mxp.maximum, mxp.maximum, y, copy=copy, xp=xp)
@@ -0,0 +1,46 @@
"""Backends against which array-api-extra runs its tests."""
from __future__ import annotations
from enum import Enum
__all__ = ["Backend"]
class Backend(Enum): # numpydoc ignore=PR02
"""
All array library backends explicitly tested by array-api-extra.
Parameters
----------
value : str
Tag of the backend's module, in the format ``<namespace>[:<extra tag>]``.
"""
# Use :<tag> to prevent Enum from deduplicating items with the same value
ARRAY_API_STRICT = "array_api_strict"
ARRAY_API_STRICTEST = "array_api_strict:strictest"
NUMPY = "numpy"
NUMPY_READONLY = "numpy:readonly"
CUPY = "cupy"
TORCH = "torch"
TORCH_GPU = "torch:gpu"
DASK = "dask.array"
SPARSE = "sparse"
JAX = "jax.numpy"
JAX_GPU = "jax.numpy:gpu"
def __str__(self) -> str: # type: ignore[explicit-override] # pyright: ignore[reportImplicitOverride] # numpydoc ignore=RT01
"""Pretty-print parameterized test names."""
return (
self.name.lower().replace("_gpu", ":gpu").replace("_readonly", ":readonly")
)
@property
def modname(self) -> str: # numpydoc ignore=RT01
"""Module name to be imported."""
return self.value.split(":")[0]
def like(self, *others: Backend) -> bool: # numpydoc ignore=PR01,RT01
"""Check if this backend uses the same module as others."""
return any(self.modname == other.modname for other in others)
@@ -0,0 +1,937 @@
"""Array-agnostic implementations for the public API."""
import math
import warnings
from collections.abc import Callable, Sequence
from types import ModuleType, NoneType
from typing import cast, overload
from ._at import at
from ._utils import _compat, _helpers
from ._utils._compat import array_namespace, is_dask_namespace, is_jax_array
from ._utils._helpers import (
asarrays,
capabilities,
eager_shape,
meta_namespace,
ndindex,
)
from ._utils._typing import Array
__all__ = [
"apply_where",
"atleast_nd",
"broadcast_shapes",
"cov",
"create_diagonal",
"expand_dims",
"kron",
"nunique",
"pad",
"setdiff1d",
"sinc",
]
@overload
def apply_where( # type: ignore[explicit-any,decorated-any] # numpydoc ignore=GL08
cond: Array,
args: Array | tuple[Array, ...],
f1: Callable[..., Array],
f2: Callable[..., Array],
/,
*,
xp: ModuleType | None = None,
) -> Array: ...
@overload
def apply_where( # type: ignore[explicit-any,decorated-any] # numpydoc ignore=GL08
cond: Array,
args: Array | tuple[Array, ...],
f1: Callable[..., Array],
/,
*,
fill_value: Array | complex,
xp: ModuleType | None = None,
) -> Array: ...
def apply_where( # type: ignore[explicit-any] # numpydoc ignore=PR01,PR02
cond: Array,
args: Array | tuple[Array, ...],
f1: Callable[..., Array],
f2: Callable[..., Array] | None = None,
/,
*,
fill_value: Array | complex | None = None,
xp: ModuleType | None = None,
) -> Array:
"""
Run one of two elementwise functions depending on a condition.
Equivalent to ``f1(*args) if cond else fill_value`` performed elementwise
when `fill_value` is defined, otherwise to ``f1(*args) if cond else f2(*args)``.
Parameters
----------
cond : array
The condition, expressed as a boolean array.
args : Array or tuple of Arrays
Argument(s) to `f1` (and `f2`). Must be broadcastable with `cond`.
f1 : callable
Elementwise function of `args`, returning a single array.
Where `cond` is True, output will be ``f1(arg0[cond], arg1[cond], ...)``.
f2 : callable, optional
Elementwise function of `args`, returning a single array.
Where `cond` is False, output will be ``f2(arg0[cond], arg1[cond], ...)``.
Mutually exclusive with `fill_value`.
fill_value : Array or scalar, optional
If provided, value with which to fill output array where `cond` is False.
It does not need to be scalar; it needs however to be broadcastable with
`cond` and `args`.
Mutually exclusive with `f2`. You must provide one or the other.
xp : array_namespace, optional
The standard-compatible namespace for `cond` and `args`. Default: infer.
Returns
-------
Array
An array with elements from the output of `f1` where `cond` is True and either
the output of `f2` or `fill_value` where `cond` is False. The returned array has
data type determined by type promotion rules between the output of `f1` and
either `fill_value` or the output of `f2`.
Notes
-----
``xp.where(cond, f1(*args), f2(*args))`` requires explicitly evaluating `f1` even
when `cond` is False, and `f2` when cond is True. This function evaluates each
function only for their matching condition, if the backend allows for it.
On Dask, `f1` and `f2` are applied to the individual chunks and should use functions
from the namespace of the chunks.
Examples
--------
>>> import array_api_strict as xp
>>> import array_api_extra as xpx
>>> a = xp.asarray([5, 4, 3])
>>> b = xp.asarray([0, 2, 2])
>>> def f(a, b):
... return a // b
>>> xpx.apply_where(b != 0, (a, b), f, fill_value=xp.nan)
array([ nan, 2., 1.])
"""
# Parse and normalize arguments
if (f2 is None) == (fill_value is None):
msg = "Exactly one of `fill_value` or `f2` must be given."
raise TypeError(msg)
args_ = list(args) if isinstance(args, tuple) else [args]
del args
xp = array_namespace(cond, fill_value, *args_) if xp is None else xp
if isinstance(fill_value, int | float | complex | NoneType):
cond, *args_ = xp.broadcast_arrays(cond, *args_)
else:
cond, fill_value, *args_ = xp.broadcast_arrays(cond, fill_value, *args_)
if is_dask_namespace(xp):
meta_xp = meta_namespace(cond, fill_value, *args_, xp=xp)
# map_blocks doesn't descend into tuples of Arrays
return xp.map_blocks(_apply_where, cond, f1, f2, fill_value, *args_, xp=meta_xp)
return _apply_where(cond, f1, f2, fill_value, *args_, xp=xp)
def _apply_where( # type: ignore[explicit-any] # numpydoc ignore=PR01,RT01
cond: Array,
f1: Callable[..., Array],
f2: Callable[..., Array] | None,
fill_value: Array | int | float | complex | bool | None,
*args: Array,
xp: ModuleType,
) -> Array:
"""Helper of `apply_where`. On Dask, this runs on a single chunk."""
if not capabilities(xp)["boolean indexing"]:
# jax.jit does not support assignment by boolean mask
return xp.where(cond, f1(*args), f2(*args) if f2 is not None else fill_value)
temp1 = f1(*(arr[cond] for arr in args))
if f2 is None:
dtype = xp.result_type(temp1, fill_value)
if isinstance(fill_value, int | float | complex):
out = xp.full_like(cond, dtype=dtype, fill_value=fill_value)
else:
out = xp.astype(fill_value, dtype, copy=True)
else:
ncond = ~cond
temp2 = f2(*(arr[ncond] for arr in args))
dtype = xp.result_type(temp1, temp2)
out = xp.empty_like(cond, dtype=dtype)
out = at(out, ncond).set(temp2)
return at(out, cond).set(temp1)
def atleast_nd(x: Array, /, *, ndim: int, xp: ModuleType | None = None) -> Array:
"""
Recursively expand the dimension of an array to at least `ndim`.
Parameters
----------
x : array
Input array.
ndim : int
The minimum number of dimensions for the result.
xp : array_namespace, optional
The standard-compatible namespace for `x`. Default: infer.
Returns
-------
array
An array with ``res.ndim`` >= `ndim`.
If ``x.ndim`` >= `ndim`, `x` is returned.
If ``x.ndim`` < `ndim`, `x` is expanded by prepending new axes
until ``res.ndim`` equals `ndim`.
Examples
--------
>>> import array_api_strict as xp
>>> import array_api_extra as xpx
>>> x = xp.asarray([1])
>>> xpx.atleast_nd(x, ndim=3, xp=xp)
Array([[[1]]], dtype=array_api_strict.int64)
>>> x = xp.asarray([[[1, 2],
... [3, 4]]])
>>> xpx.atleast_nd(x, ndim=1, xp=xp) is x
True
"""
if xp is None:
xp = array_namespace(x)
if x.ndim < ndim:
x = xp.expand_dims(x, axis=0)
x = atleast_nd(x, ndim=ndim, xp=xp)
return x
# `float` in signature to accept `math.nan` for Dask.
# `int`s are still accepted as `float` is a superclass of `int` in typing
def broadcast_shapes(*shapes: tuple[float | None, ...]) -> tuple[int | None, ...]:
"""
Compute the shape of the broadcasted arrays.
Duplicates :func:`numpy.broadcast_shapes`, with additional support for
None and NaN sizes.
This is equivalent to ``xp.broadcast_arrays(arr1, arr2, ...)[0].shape``
without needing to worry about the backend potentially deep copying
the arrays.
Parameters
----------
*shapes : tuple[int | None, ...]
Shapes of the arrays to broadcast.
Returns
-------
tuple[int | None, ...]
The shape of the broadcasted arrays.
See Also
--------
numpy.broadcast_shapes : Equivalent NumPy function.
array_api.broadcast_arrays : Function to broadcast actual arrays.
Notes
-----
This function accepts the Array API's ``None`` for unknown sizes,
as well as Dask's non-standard ``math.nan``.
Regardless of input, the output always contains ``None`` for unknown sizes.
Examples
--------
>>> import array_api_extra as xpx
>>> xpx.broadcast_shapes((2, 3), (2, 1))
(2, 3)
>>> xpx.broadcast_shapes((4, 2, 3), (2, 1), (1, 3))
(4, 2, 3)
"""
if not shapes:
return () # Match NumPy output
ndim = max(len(shape) for shape in shapes)
out: list[int | None] = []
for axis in range(-ndim, 0):
sizes = {shape[axis] for shape in shapes if axis >= -len(shape)}
# Dask uses NaN for unknown shape, which predates the Array API spec for None
none_size = None in sizes or math.nan in sizes
sizes -= {1, None, math.nan}
if len(sizes) > 1:
msg = (
"shape mismatch: objects cannot be broadcast to a single shape: "
f"{shapes}."
)
raise ValueError(msg)
out.append(None if none_size else cast(int, sizes.pop()) if sizes else 1)
return tuple(out)
def cov(m: Array, /, *, xp: ModuleType | None = None) -> Array:
"""
Estimate a covariance matrix.
Covariance indicates the level to which two variables vary together.
If we examine N-dimensional samples, :math:`X = [x_1, x_2, ... x_N]^T`,
then the covariance matrix element :math:`C_{ij}` is the covariance of
:math:`x_i` and :math:`x_j`. The element :math:`C_{ii}` is the variance
of :math:`x_i`.
This provides a subset of the functionality of ``numpy.cov``.
Parameters
----------
m : array
A 1-D or 2-D array containing multiple variables and observations.
Each row of `m` represents a variable, and each column a single
observation of all those variables.
xp : array_namespace, optional
The standard-compatible namespace for `m`. Default: infer.
Returns
-------
array
The covariance matrix of the variables.
Examples
--------
>>> import array_api_strict as xp
>>> import array_api_extra as xpx
Consider two variables, :math:`x_0` and :math:`x_1`, which
correlate perfectly, but in opposite directions:
>>> x = xp.asarray([[0, 2], [1, 1], [2, 0]]).T
>>> x
Array([[0, 1, 2],
[2, 1, 0]], dtype=array_api_strict.int64)
Note how :math:`x_0` increases while :math:`x_1` decreases. The covariance
matrix shows this clearly:
>>> xpx.cov(x, xp=xp)
Array([[ 1., -1.],
[-1., 1.]], dtype=array_api_strict.float64)
Note that element :math:`C_{0,1}`, which shows the correlation between
:math:`x_0` and :math:`x_1`, is negative.
Further, note how `x` and `y` are combined:
>>> x = xp.asarray([-2.1, -1, 4.3])
>>> y = xp.asarray([3, 1.1, 0.12])
>>> X = xp.stack((x, y), axis=0)
>>> xpx.cov(X, xp=xp)
Array([[11.71 , -4.286 ],
[-4.286 , 2.14413333]], dtype=array_api_strict.float64)
>>> xpx.cov(x, xp=xp)
Array(11.71, dtype=array_api_strict.float64)
>>> xpx.cov(y, xp=xp)
Array(2.14413333, dtype=array_api_strict.float64)
"""
if xp is None:
xp = array_namespace(m)
m = xp.asarray(m, copy=True)
dtype = (
xp.float64 if xp.isdtype(m.dtype, "integral") else xp.result_type(m, xp.float64)
)
m = atleast_nd(m, ndim=2, xp=xp)
m = xp.astype(m, dtype)
avg = _helpers.mean(m, axis=1, xp=xp)
m_shape = eager_shape(m)
fact = m_shape[1] - 1
if fact <= 0:
warnings.warn("Degrees of freedom <= 0 for slice", RuntimeWarning, stacklevel=2)
fact = 0
m -= avg[:, None]
m_transpose = m.T
if xp.isdtype(m_transpose.dtype, "complex floating"):
m_transpose = xp.conj(m_transpose)
c = m @ m_transpose
c /= fact
axes = tuple(axis for axis, length in enumerate(c.shape) if length == 1)
return xp.squeeze(c, axis=axes)
def create_diagonal(
x: Array, /, *, offset: int = 0, xp: ModuleType | None = None
) -> Array:
"""
Construct a diagonal array.
Parameters
----------
x : array
An array having shape ``(*batch_dims, k)``.
offset : int, optional
Offset from the leading diagonal (default is ``0``).
Use positive ints for diagonals above the leading diagonal,
and negative ints for diagonals below the leading diagonal.
xp : array_namespace, optional
The standard-compatible namespace for `x`. Default: infer.
Returns
-------
array
An array having shape ``(*batch_dims, k+abs(offset), k+abs(offset))`` with `x`
on the diagonal (offset by `offset`).
Examples
--------
>>> import array_api_strict as xp
>>> import array_api_extra as xpx
>>> x = xp.asarray([2, 4, 8])
>>> xpx.create_diagonal(x, xp=xp)
Array([[2, 0, 0],
[0, 4, 0],
[0, 0, 8]], dtype=array_api_strict.int64)
>>> xpx.create_diagonal(x, offset=-2, xp=xp)
Array([[0, 0, 0, 0, 0],
[0, 0, 0, 0, 0],
[2, 0, 0, 0, 0],
[0, 4, 0, 0, 0],
[0, 0, 8, 0, 0]], dtype=array_api_strict.int64)
"""
if xp is None:
xp = array_namespace(x)
if x.ndim == 0:
err_msg = "`x` must be at least 1-dimensional."
raise ValueError(err_msg)
x_shape = eager_shape(x)
batch_dims = x_shape[:-1]
n = x_shape[-1] + abs(offset)
diag = xp.zeros((*batch_dims, n**2), dtype=x.dtype, device=_compat.device(x))
target_slice = slice(
offset if offset >= 0 else abs(offset) * n,
min(n * (n - offset), diag.shape[-1]),
n + 1,
)
for index in ndindex(*batch_dims):
diag = at(diag)[(*index, target_slice)].set(x[(*index, slice(None))])
return xp.reshape(diag, (*batch_dims, n, n))
def expand_dims(
a: Array, /, *, axis: int | tuple[int, ...] = (0,), xp: ModuleType | None = None
) -> Array:
"""
Expand the shape of an array.
Insert (a) new axis/axes that will appear at the position(s) specified by
`axis` in the expanded array shape.
This is ``xp.expand_dims`` for `axis` an int *or a tuple of ints*.
Roughly equivalent to ``numpy.expand_dims`` for NumPy arrays.
Parameters
----------
a : array
Array to have its shape expanded.
axis : int or tuple of ints, optional
Position(s) in the expanded axes where the new axis (or axes) is/are placed.
If multiple positions are provided, they should be unique (note that a position
given by a positive index could also be referred to by a negative index -
that will also result in an error).
Default: ``(0,)``.
xp : array_namespace, optional
The standard-compatible namespace for `a`. Default: infer.
Returns
-------
array
`a` with an expanded shape.
Examples
--------
>>> import array_api_strict as xp
>>> import array_api_extra as xpx
>>> x = xp.asarray([1, 2])
>>> x.shape
(2,)
The following is equivalent to ``x[xp.newaxis, :]`` or ``x[xp.newaxis]``:
>>> y = xpx.expand_dims(x, axis=0, xp=xp)
>>> y
Array([[1, 2]], dtype=array_api_strict.int64)
>>> y.shape
(1, 2)
The following is equivalent to ``x[:, xp.newaxis]``:
>>> y = xpx.expand_dims(x, axis=1, xp=xp)
>>> y
Array([[1],
[2]], dtype=array_api_strict.int64)
>>> y.shape
(2, 1)
``axis`` may also be a tuple:
>>> y = xpx.expand_dims(x, axis=(0, 1), xp=xp)
>>> y
Array([[[1, 2]]], dtype=array_api_strict.int64)
>>> y = xpx.expand_dims(x, axis=(2, 0), xp=xp)
>>> y
Array([[[1],
[2]]], dtype=array_api_strict.int64)
"""
if xp is None:
xp = array_namespace(a)
if not isinstance(axis, tuple):
axis = (axis,)
ndim = a.ndim + len(axis)
if axis != () and (min(axis) < -ndim or max(axis) >= ndim):
err_msg = (
f"a provided axis position is out of bounds for array of dimension {a.ndim}"
)
raise IndexError(err_msg)
axis = tuple(dim % ndim for dim in axis)
if len(set(axis)) != len(axis):
err_msg = "Duplicate dimensions specified in `axis`."
raise ValueError(err_msg)
for i in sorted(axis):
a = xp.expand_dims(a, axis=i)
return a
def isclose(
a: Array | complex,
b: Array | complex,
*,
rtol: float = 1e-05,
atol: float = 1e-08,
equal_nan: bool = False,
xp: ModuleType,
) -> Array: # numpydoc ignore=PR01,RT01
"""See docstring in array_api_extra._delegation."""
a, b = asarrays(a, b, xp=xp)
a_inexact = xp.isdtype(a.dtype, ("real floating", "complex floating"))
b_inexact = xp.isdtype(b.dtype, ("real floating", "complex floating"))
if a_inexact or b_inexact:
# prevent warnings on NumPy and Dask on inf - inf
mxp = meta_namespace(a, b, xp=xp)
out = apply_where(
xp.isinf(a) | xp.isinf(b),
(a, b),
lambda a, b: mxp.isinf(a) & mxp.isinf(b) & (mxp.sign(a) == mxp.sign(b)), # pyright: ignore[reportUnknownArgumentType]
# Note: inf <= inf is True!
lambda a, b: mxp.abs(a - b) <= (atol + rtol * mxp.abs(b)), # pyright: ignore[reportUnknownArgumentType]
xp=xp,
)
if equal_nan:
out = xp.where(xp.isnan(a) & xp.isnan(b), True, out)
return out
if xp.isdtype(a.dtype, "bool") or xp.isdtype(b.dtype, "bool"):
if atol >= 1 or rtol >= 1:
return xp.ones_like(a == b)
return a == b
# integer types
atol = int(atol)
if rtol == 0:
return xp.abs(a - b) <= atol
# Don't rely on OverflowError, as it is not guaranteed by the Array API.
nrtol = int(1.0 / rtol)
if nrtol > xp.iinfo(b.dtype).max:
# rtol * max_int < 1, so it's inconsequential
return xp.abs(a - b) <= atol
return xp.abs(a - b) <= (atol + xp.abs(b) // nrtol)
def kron(
a: Array | complex,
b: Array | complex,
/,
*,
xp: ModuleType | None = None,
) -> Array:
"""
Kronecker product of two arrays.
Computes the Kronecker product, a composite array made of blocks of the
second array scaled by the first.
Equivalent to ``numpy.kron`` for NumPy arrays.
Parameters
----------
a, b : Array | int | float | complex
Input arrays or scalars. At least one must be an array.
xp : array_namespace, optional
The standard-compatible namespace for `a` and `b`. Default: infer.
Returns
-------
array
The Kronecker product of `a` and `b`.
Notes
-----
The function assumes that the number of dimensions of `a` and `b`
are the same, if necessary prepending the smallest with ones.
If ``a.shape = (r0,r1,..,rN)`` and ``b.shape = (s0,s1,...,sN)``,
the Kronecker product has shape ``(r0*s0, r1*s1, ..., rN*SN)``.
The elements are products of elements from `a` and `b`, organized
explicitly by::
kron(a,b)[k0,k1,...,kN] = a[i0,i1,...,iN] * b[j0,j1,...,jN]
where::
kt = it * st + jt, t = 0,...,N
In the common 2-D case (N=1), the block structure can be visualized::
[[ a[0,0]*b, a[0,1]*b, ... , a[0,-1]*b ],
[ ... ... ],
[ a[-1,0]*b, a[-1,1]*b, ... , a[-1,-1]*b ]]
Examples
--------
>>> import array_api_strict as xp
>>> import array_api_extra as xpx
>>> xpx.kron(xp.asarray([1, 10, 100]), xp.asarray([5, 6, 7]), xp=xp)
Array([ 5, 6, 7, 50, 60, 70, 500,
600, 700], dtype=array_api_strict.int64)
>>> xpx.kron(xp.asarray([5, 6, 7]), xp.asarray([1, 10, 100]), xp=xp)
Array([ 5, 50, 500, 6, 60, 600, 7,
70, 700], dtype=array_api_strict.int64)
>>> xpx.kron(xp.eye(2), xp.ones((2, 2)), xp=xp)
Array([[1., 1., 0., 0.],
[1., 1., 0., 0.],
[0., 0., 1., 1.],
[0., 0., 1., 1.]], dtype=array_api_strict.float64)
>>> a = xp.reshape(xp.arange(100), (2, 5, 2, 5))
>>> b = xp.reshape(xp.arange(24), (2, 3, 4))
>>> c = xpx.kron(a, b, xp=xp)
>>> c.shape
(2, 10, 6, 20)
>>> I = (1, 3, 0, 2)
>>> J = (0, 2, 1)
>>> J1 = (0,) + J # extend to ndim=4
>>> S1 = (1,) + b.shape
>>> K = tuple(xp.asarray(I) * xp.asarray(S1) + xp.asarray(J1))
>>> c[K] == a[I]*b[J]
Array(True, dtype=array_api_strict.bool)
"""
if xp is None:
xp = array_namespace(a, b)
a, b = asarrays(a, b, xp=xp)
singletons = (1,) * (b.ndim - a.ndim)
a = cast(Array, xp.broadcast_to(a, singletons + a.shape))
nd_b, nd_a = b.ndim, a.ndim
nd_max = max(nd_b, nd_a)
if nd_a == 0 or nd_b == 0:
return xp.multiply(a, b)
a_shape = eager_shape(a)
b_shape = eager_shape(b)
# Equalise the shapes by prepending smaller one with 1s
a_shape = (1,) * max(0, nd_b - nd_a) + a_shape
b_shape = (1,) * max(0, nd_a - nd_b) + b_shape
# Insert empty dimensions
a_arr = expand_dims(a, axis=tuple(range(nd_b - nd_a)), xp=xp)
b_arr = expand_dims(b, axis=tuple(range(nd_a - nd_b)), xp=xp)
# Compute the product
a_arr = expand_dims(a_arr, axis=tuple(range(1, nd_max * 2, 2)), xp=xp)
b_arr = expand_dims(b_arr, axis=tuple(range(0, nd_max * 2, 2)), xp=xp)
result = xp.multiply(a_arr, b_arr)
# Reshape back and return
res_shape = tuple(a_s * b_s for a_s, b_s in zip(a_shape, b_shape, strict=True))
return xp.reshape(result, res_shape)
def nunique(x: Array, /, *, xp: ModuleType | None = None) -> Array:
"""
Count the number of unique elements in an array.
Compatible with JAX and Dask, whose laziness would be otherwise
problematic.
Parameters
----------
x : Array
Input array.
xp : array_namespace, optional
The standard-compatible namespace for `x`. Default: infer.
Returns
-------
array: 0-dimensional integer array
The number of unique elements in `x`. It can be lazy.
"""
if xp is None:
xp = array_namespace(x)
if is_jax_array(x):
# size= is JAX-specific
# https://github.com/data-apis/array-api/issues/883
_, counts = xp.unique_counts(x, size=_compat.size(x))
return (counts > 0).sum()
# There are 3 general use cases:
# 1. backend has unique_counts and it returns an array with known shape
# 2. backend has unique_counts and it returns a None-sized array;
# e.g. Dask, ndonnx
# 3. backend does not have unique_counts; e.g. wrapped JAX
if capabilities(xp)["data-dependent shapes"]:
# xp has unique_counts; O(n) complexity
_, counts = xp.unique_counts(x)
n = _compat.size(counts)
if n is None:
return xp.sum(xp.ones_like(counts))
return xp.asarray(n, device=_compat.device(x))
# xp does not have unique_counts; O(n*logn) complexity
x = xp.reshape(x, (-1,))
x = xp.sort(x)
mask = x != xp.roll(x, -1)
default_int = xp.__array_namespace_info__().default_dtypes(
device=_compat.device(x)
)["integral"]
return xp.maximum(
# Special cases:
# - array is size 0
# - array has all elements equal to each other
xp.astype(xp.any(~mask), default_int),
xp.sum(xp.astype(mask, default_int)),
)
def pad(
x: Array,
pad_width: int | tuple[int, int] | Sequence[tuple[int, int]],
*,
constant_values: complex = 0,
xp: ModuleType,
) -> Array: # numpydoc ignore=PR01,RT01
"""See docstring in `array_api_extra._delegation.py`."""
# make pad_width a list of length-2 tuples of ints
if isinstance(pad_width, int):
pad_width_seq = [(pad_width, pad_width)] * x.ndim
elif (
isinstance(pad_width, tuple)
and len(pad_width) == 2
and all(isinstance(i, int) for i in pad_width)
):
pad_width_seq = [cast(tuple[int, int], pad_width)] * x.ndim
else:
pad_width_seq = cast(list[tuple[int, int]], list(pad_width))
# https://github.com/python/typeshed/issues/13376
slices: list[slice] = [] # type: ignore[explicit-any]
newshape: list[int] = []
for ax, w_tpl in enumerate(pad_width_seq):
if len(w_tpl) != 2:
msg = f"expect a 2-tuple (before, after), got {w_tpl}."
raise ValueError(msg)
sh = eager_shape(x)[ax]
if w_tpl[0] == 0 and w_tpl[1] == 0:
sl = slice(None, None, None)
else:
start, stop = w_tpl
stop = None if stop == 0 else -stop
sl = slice(start, stop, None)
sh += w_tpl[0] + w_tpl[1]
newshape.append(sh)
slices.append(sl)
padded = xp.full(
tuple(newshape),
fill_value=constant_values,
dtype=x.dtype,
device=_compat.device(x),
)
return at(padded, tuple(slices)).set(x)
def setdiff1d(
x1: Array | complex,
x2: Array | complex,
/,
*,
assume_unique: bool = False,
xp: ModuleType | None = None,
) -> Array:
"""
Find the set difference of two arrays.
Return the unique values in `x1` that are not in `x2`.
Parameters
----------
x1 : array | int | float | complex | bool
Input array.
x2 : array
Input comparison array.
assume_unique : bool
If ``True``, the input arrays are both assumed to be unique, which
can speed up the calculation. Default is ``False``.
xp : array_namespace, optional
The standard-compatible namespace for `x1` and `x2`. Default: infer.
Returns
-------
array
1D array of values in `x1` that are not in `x2`. The result
is sorted when `assume_unique` is ``False``, but otherwise only sorted
if the input is sorted.
Examples
--------
>>> import array_api_strict as xp
>>> import array_api_extra as xpx
>>> x1 = xp.asarray([1, 2, 3, 2, 4, 1])
>>> x2 = xp.asarray([3, 4, 5, 6])
>>> xpx.setdiff1d(x1, x2, xp=xp)
Array([1, 2], dtype=array_api_strict.int64)
"""
if xp is None:
xp = array_namespace(x1, x2)
# https://github.com/microsoft/pyright/issues/10103
x1_, x2_ = asarrays(x1, x2, xp=xp)
if assume_unique:
x1_ = xp.reshape(x1_, (-1,))
x2_ = xp.reshape(x2_, (-1,))
else:
x1_ = xp.unique_values(x1_)
x2_ = xp.unique_values(x2_)
return x1_[_helpers.in1d(x1_, x2_, assume_unique=True, invert=True, xp=xp)]
def sinc(x: Array, /, *, xp: ModuleType | None = None) -> Array:
r"""
Return the normalized sinc function.
The sinc function is equal to :math:`\sin(\pi x)/(\pi x)` for any argument
:math:`x\ne 0`. ``sinc(0)`` takes the limit value 1, making ``sinc`` not
only everywhere continuous but also infinitely differentiable.
.. note::
Note the normalization factor of ``pi`` used in the definition.
This is the most commonly used definition in signal processing.
Use ``sinc(x / xp.pi)`` to obtain the unnormalized sinc function
:math:`\sin(x)/x` that is more common in mathematics.
Parameters
----------
x : array
Array (possibly multi-dimensional) of values for which to calculate
``sinc(x)``. Must have a real floating point dtype.
xp : array_namespace, optional
The standard-compatible namespace for `x`. Default: infer.
Returns
-------
array
``sinc(x)`` calculated elementwise, which has the same shape as the input.
Notes
-----
The name sinc is short for "sine cardinal" or "sinus cardinalis".
The sinc function is used in various signal processing applications,
including in anti-aliasing, in the construction of a Lanczos resampling
filter, and in interpolation.
For bandlimited interpolation of discrete-time signals, the ideal
interpolation kernel is proportional to the sinc function.
References
----------
#. Weisstein, Eric W. "Sinc Function." From MathWorld--A Wolfram Web
Resource. https://mathworld.wolfram.com/SincFunction.html
#. Wikipedia, "Sinc function",
https://en.wikipedia.org/wiki/Sinc_function
Examples
--------
>>> import array_api_strict as xp
>>> import array_api_extra as xpx
>>> x = xp.linspace(-4, 4, 41)
>>> xpx.sinc(x, xp=xp)
Array([-3.89817183e-17, -4.92362781e-02,
-8.40918587e-02, -8.90384387e-02,
-5.84680802e-02, 3.89817183e-17,
6.68206631e-02, 1.16434881e-01,
1.26137788e-01, 8.50444803e-02,
-3.89817183e-17, -1.03943254e-01,
-1.89206682e-01, -2.16236208e-01,
-1.55914881e-01, 3.89817183e-17,
2.33872321e-01, 5.04551152e-01,
7.56826729e-01, 9.35489284e-01,
1.00000000e+00, 9.35489284e-01,
7.56826729e-01, 5.04551152e-01,
2.33872321e-01, 3.89817183e-17,
-1.55914881e-01, -2.16236208e-01,
-1.89206682e-01, -1.03943254e-01,
-3.89817183e-17, 8.50444803e-02,
1.26137788e-01, 1.16434881e-01,
6.68206631e-02, 3.89817183e-17,
-5.84680802e-02, -8.90384387e-02,
-8.40918587e-02, -4.92362781e-02,
-3.89817183e-17], dtype=array_api_strict.float64)
"""
if xp is None:
xp = array_namespace(x)
if not xp.isdtype(x.dtype, "real floating"):
err_msg = "`x` must have a real floating data type."
raise ValueError(err_msg)
# no scalars in `where` - array-api#807
y = xp.pi * xp.where(
xp.astype(x, xp.bool),
x,
xp.asarray(xp.finfo(x.dtype).eps, dtype=x.dtype, device=_compat.device(x)),
)
return xp.sin(y) / y
@@ -0,0 +1,357 @@
"""Public API Functions."""
from __future__ import annotations
import math
from collections.abc import Callable, Sequence
from functools import partial, wraps
from types import ModuleType
from typing import TYPE_CHECKING, Any, ParamSpec, TypeAlias, cast, overload
from ._funcs import broadcast_shapes
from ._utils import _compat
from ._utils._compat import (
array_namespace,
is_dask_namespace,
is_jax_namespace,
)
from ._utils._helpers import is_python_scalar
from ._utils._typing import Array, DType
if TYPE_CHECKING: # pragma: no cover
import numpy as np
from numpy.typing import ArrayLike
NumPyObject: TypeAlias = np.ndarray[Any, Any] | np.generic # type: ignore[explicit-any]
else:
# Sphinx hack
NumPyObject = Any
P = ParamSpec("P")
@overload
def lazy_apply( # type: ignore[decorated-any, valid-type]
func: Callable[P, Array | ArrayLike],
*args: Array | complex | None,
shape: tuple[int | None, ...] | None = None,
dtype: DType | None = None,
as_numpy: bool = False,
xp: ModuleType | None = None,
**kwargs: P.kwargs, # pyright: ignore[reportGeneralTypeIssues]
) -> Array: ... # numpydoc ignore=GL08
@overload
def lazy_apply( # type: ignore[decorated-any, valid-type]
func: Callable[P, Sequence[Array | ArrayLike]],
*args: Array | complex | None,
shape: Sequence[tuple[int | None, ...]],
dtype: Sequence[DType] | None = None,
as_numpy: bool = False,
xp: ModuleType | None = None,
**kwargs: P.kwargs, # pyright: ignore[reportGeneralTypeIssues]
) -> tuple[Array, ...]: ... # numpydoc ignore=GL08
def lazy_apply( # type: ignore[valid-type] # numpydoc ignore=GL07,SA04
func: Callable[P, Array | ArrayLike | Sequence[Array | ArrayLike]],
*args: Array | complex | None,
shape: tuple[int | None, ...] | Sequence[tuple[int | None, ...]] | None = None,
dtype: DType | Sequence[DType] | None = None,
as_numpy: bool = False,
xp: ModuleType | None = None,
**kwargs: P.kwargs, # pyright: ignore[reportGeneralTypeIssues]
) -> Array | tuple[Array, ...]:
"""
Lazily apply an eager function.
If the backend of the input arrays is lazy, e.g. Dask or jitted JAX, the execution
of the function is delayed until the graph is materialized; if it's eager, the
function is executed immediately.
Parameters
----------
func : callable
The function to apply.
It must accept one or more array API compliant arrays as positional arguments.
If `as_numpy=True`, inputs are converted to NumPy before they are passed to
`func`.
It must return either a single array-like or a sequence of array-likes.
`func` must be a pure function, i.e. without side effects, as depending on the
backend it may be executed more than once or never.
*args : Array | int | float | complex | bool | None
One or more Array API compliant arrays, Python scalars, or None's.
If `as_numpy=True`, you need to be able to apply :func:`numpy.asarray` to
non-None args to convert them to NumPy; read notes below about specific
backends.
shape : tuple[int | None, ...] | Sequence[tuple[int | None, ...]], optional
Output shape or sequence of output shapes, one for each output of `func`.
Default: assume single output and broadcast shapes of the input arrays.
dtype : DType | Sequence[DType], optional
Output dtype or sequence of output dtypes, one for each output of `func`.
dtype(s) must belong to the same array namespace as the input arrays.
Default: infer the result type(s) from the input arrays.
as_numpy : bool, optional
If True, convert the input arrays to NumPy before passing them to `func`.
This is particularly useful to make NumPy-only functions, e.g. written in Cython
or Numba, work transparently with array API-compliant arrays.
Default: False.
xp : array_namespace, optional
The standard-compatible namespace for `args`. Default: infer.
**kwargs : Any, optional
Additional keyword arguments to pass verbatim to `func`.
They cannot contain Array objects.
Returns
-------
Array | tuple[Array, ...]
The result(s) of `func` applied to the input arrays, wrapped in the same
array namespace as the inputs.
If shape is omitted or a single `tuple[int | None, ...]`, return a single array.
Otherwise, return a tuple of arrays.
Notes
-----
JAX
This allows applying eager functions to jitted JAX arrays, which are lazy.
The function won't be applied until the JAX array is materialized.
When running inside ``jax.jit``, `shape` must be fully known, i.e. it cannot
contain any `None` elements.
.. warning::
`func` must never raise inside ``jax.jit``, as the resulting behavior is
undefined.
Using this with `as_numpy=False` is particularly useful to apply non-jittable
JAX functions to arrays on GPU devices.
If ``as_numpy=True``, the :doc:`jax:transfer_guard` may prevent arrays on a GPU
device from being transferred back to CPU. This is treated as an implicit
transfer.
PyTorch, CuPy
If ``as_numpy=True``, these backends raise by default if you attempt to convert
arrays on a GPU device to NumPy.
Sparse
If ``as_numpy=True``, by default sparse prevents implicit densification through
:func:`numpy.asarray`. `This safety mechanism can be disabled
<https://sparse.pydata.org/en/stable/operations.html#package-configuration>`_.
Dask
This allows applying eager functions to Dask arrays.
The Dask graph won't be computed until the user calls ``compute()`` or
``persist()`` down the line.
The function name will be prominently visible on the user-facing Dask
dashboard and on Prometheus metrics, so it is recommended for it to be
meaningful.
`lazy_apply` doesn't know if `func` reduces along any axes; also, shape
changes are non-trivial in chunked Dask arrays. For these reasons, all inputs
will be rechunked into a single chunk.
.. warning::
The whole operation needs to fit in memory all at once on a single worker.
The outputs will also be returned as a single chunk and you should consider
rechunking them into smaller chunks afterwards.
If you want to distribute the calculation across multiple workers, you
should use :func:`dask.array.map_blocks`, :func:`dask.array.map_overlap`,
:func:`dask.array.blockwise`, or a native Dask wrapper instead of
`lazy_apply`.
Dask wrapping around other backends
If ``as_numpy=False``, `func` will receive in input eager arrays of the meta
namespace, as defined by the ``._meta`` attribute of the input Dask arrays.
The outputs of `func` will be wrapped by the meta namespace, and then wrapped
again by Dask.
Raises
------
ValueError
When ``xp=jax.numpy``, the output `shape` is unknown (it contains ``None`` on
one or more axes) and this function was called inside ``jax.jit``.
RuntimeError
When ``xp=sparse`` and auto-densification is disabled.
Exception (backend-specific)
When the backend disallows implicit device to host transfers and the input
arrays are on a non-CPU device, e.g. on GPU.
See Also
--------
jax.transfer_guard
jax.pure_callback
dask.array.map_blocks
dask.array.map_overlap
dask.array.blockwise
"""
args_not_none = [arg for arg in args if arg is not None]
array_args = [arg for arg in args_not_none if not is_python_scalar(arg)]
if not array_args:
msg = "Must have at least one argument array"
raise ValueError(msg)
if xp is None:
xp = array_namespace(*args)
# Normalize and validate shape and dtype
shapes: list[tuple[int | None, ...]]
dtypes: list[DType]
multi_output = False
if shape is None:
shapes = [broadcast_shapes(*(arg.shape for arg in array_args))]
elif all(isinstance(s, int | None) for s in shape):
# Do not test for shape to be a tuple
# https://github.com/data-apis/array-api/issues/891#issuecomment-2637430522
shapes = [cast(tuple[int | None, ...], shape)]
else:
shapes = list(shape) # type: ignore[arg-type] # pyright: ignore[reportAssignmentType]
multi_output = True
if dtype is None:
dtypes = [xp.result_type(*args_not_none)] * len(shapes)
elif multi_output:
if not isinstance(dtype, Sequence):
msg = "Got multiple shapes but only one dtype"
raise ValueError(msg)
dtypes = list(dtype) # pyright: ignore[reportUnknownArgumentType]
else:
if isinstance(dtype, Sequence):
msg = "Got single shape but multiple dtypes"
raise ValueError(msg)
dtypes = [dtype]
if len(shapes) != len(dtypes):
msg = f"Got {len(shapes)} shapes and {len(dtypes)} dtypes"
raise ValueError(msg)
del shape
del dtype
# End of shape and dtype parsing
# Backend-specific branches
if is_dask_namespace(xp):
import dask
metas: list[Array] = [arg._meta for arg in array_args] # pylint: disable=protected-access # pyright: ignore[reportAttributeAccessIssue]
meta_xp = array_namespace(*metas)
wrapped = dask.delayed( # type: ignore[attr-defined] # pyright: ignore[reportPrivateImportUsage]
_lazy_apply_wrapper(func, as_numpy, multi_output, meta_xp),
pure=True,
)
# This finalizes each arg, which is the same as arg.rechunk(-1).
# Please read docstring above for why we're not using
# dask.array.map_blocks or dask.array.blockwise!
delayed_out = wrapped(*args, **kwargs)
out = tuple(
xp.from_delayed(
delayed_out[i], # pyright: ignore[reportIndexIssue]
# Dask's unknown shapes diverge from the Array API specification
shape=tuple(math.nan if s is None else s for s in shape),
dtype=dtype,
meta=metas[0],
)
for i, (shape, dtype) in enumerate(zip(shapes, dtypes, strict=True))
)
elif is_jax_namespace(xp) and _is_jax_jit_enabled(xp):
# Delay calling func with jax.pure_callback, which will forward to func eager
# JAX arrays. Do not use jax.pure_callback when running outside of the JIT,
# as it does not support raising exceptions:
# https://github.com/jax-ml/jax/issues/26102
import jax
if any(None in shape for shape in shapes):
msg = "Output shape must be fully known when running inside jax.jit"
raise ValueError(msg)
# Shield kwargs from being coerced into JAX arrays.
# jax.pure_callback calls jax.jit under the hood, but without the chance of
# passing static_argnames / static_argnums.
wrapped = _lazy_apply_wrapper(
partial(func, **kwargs), as_numpy, multi_output, xp
)
# suppress unused-ignore to run mypy in -e lint as well as -e dev
out = cast( # type: ignore[bad-cast,unused-ignore]
tuple[Array, ...],
jax.pure_callback(
wrapped,
tuple(
jax.ShapeDtypeStruct(shape, dtype) # pyright: ignore[reportUnknownArgumentType]
for shape, dtype in zip(shapes, dtypes, strict=True)
),
*args,
),
)
else:
# Eager backends, including non-jitted JAX
wrapped = _lazy_apply_wrapper(func, as_numpy, multi_output, xp)
out = wrapped(*args, **kwargs)
return out if multi_output else out[0]
def _is_jax_jit_enabled(xp: ModuleType) -> bool: # numpydoc ignore=PR01,RT01
"""Return True if this function is being called inside ``jax.jit``."""
import jax # pylint: disable=import-outside-toplevel
x = xp.asarray(False)
try:
return bool(x)
except jax.errors.TracerBoolConversionError:
return True
def _lazy_apply_wrapper( # type: ignore[explicit-any] # numpydoc ignore=PR01,RT01
func: Callable[..., Array | ArrayLike | Sequence[Array | ArrayLike]],
as_numpy: bool,
multi_output: bool,
xp: ModuleType,
) -> Callable[..., tuple[Array, ...]]:
"""
Helper of `lazy_apply`.
Given a function that accepts one or more arrays as positional arguments and returns
a single array-like or a sequence of array-likes, return a function that accepts the
same number of Array API arrays and always returns a tuple of Array API array.
Any keyword arguments are passed through verbatim to the wrapped function.
"""
# On Dask, @wraps causes the graph key to contain the wrapped function's name
@wraps(func)
def wrapper( # type: ignore[decorated-any,explicit-any]
*args: Array | complex | None, **kwargs: Any
) -> tuple[Array, ...]: # numpydoc ignore=GL08
args_list = []
device = None
for arg in args:
if arg is not None and not is_python_scalar(arg):
if device is None:
device = _compat.device(arg)
if as_numpy:
import numpy as np
arg = cast(Array, np.asarray(arg)) # type: ignore[bad-cast] # noqa: PLW2901
args_list.append(arg)
assert device is not None
out = func(*args_list, **kwargs)
if multi_output:
assert isinstance(out, Sequence)
return tuple(xp.asarray(o, device=device) for o in out)
return (xp.asarray(out, device=device),)
return wrapper
@@ -0,0 +1,278 @@
"""
Testing utilities.
Note that this is private API; don't expect it to be stable.
See also ..testing for public testing utilities.
"""
from __future__ import annotations
import math
from types import ModuleType
from typing import Any, cast
import numpy as np
import pytest
from ._utils._compat import (
array_namespace,
is_array_api_strict_namespace,
is_cupy_namespace,
is_dask_namespace,
is_jax_namespace,
is_numpy_namespace,
is_pydata_sparse_namespace,
is_torch_namespace,
to_device,
)
from ._utils._typing import Array, Device
__all__ = ["as_numpy_array", "xp_assert_close", "xp_assert_equal", "xp_assert_less"]
def _check_ns_shape_dtype(
actual: Array,
desired: Array,
check_dtype: bool,
check_shape: bool,
check_scalar: bool,
) -> ModuleType: # numpydoc ignore=RT03
"""
Assert that namespace, shape and dtype of the two arrays match.
Parameters
----------
actual : Array
The array produced by the tested function.
desired : Array
The expected array (typically hardcoded).
check_dtype, check_shape : bool, default: True
Whether to check agreement between actual and desired dtypes and shapes
check_scalar : bool, default: False
NumPy only: whether to check agreement between actual and desired types -
0d array vs scalar.
Returns
-------
Arrays namespace.
"""
actual_xp = array_namespace(actual) # Raises on scalars and lists
desired_xp = array_namespace(desired)
msg = f"namespaces do not match: {actual_xp} != f{desired_xp}"
assert actual_xp == desired_xp, msg
if check_shape:
actual_shape = actual.shape
desired_shape = desired.shape
if is_dask_namespace(desired_xp):
# Dask uses nan instead of None for unknown shapes
if any(math.isnan(i) for i in cast(tuple[float, ...], actual_shape)):
actual_shape = actual.compute().shape # type: ignore[attr-defined] # pyright: ignore[reportAttributeAccessIssue]
if any(math.isnan(i) for i in cast(tuple[float, ...], desired_shape)):
desired_shape = desired.compute().shape # type: ignore[attr-defined] # pyright: ignore[reportAttributeAccessIssue]
msg = f"shapes do not match: {actual_shape} != f{desired_shape}"
assert actual_shape == desired_shape, msg
if check_dtype:
msg = f"dtypes do not match: {actual.dtype} != {desired.dtype}"
assert actual.dtype == desired.dtype, msg
if is_numpy_namespace(actual_xp) and check_scalar:
# only NumPy distinguishes between scalars and arrays; we do if check_scalar.
_msg = (
"array-ness does not match:\n Actual: "
f"{type(actual)}\n Desired: {type(desired)}"
)
assert np.isscalar(actual) == np.isscalar(desired), _msg
return desired_xp
def as_numpy_array(array: Array, *, xp: ModuleType) -> np.typing.NDArray[Any]: # type: ignore[explicit-any]
"""
Convert array to NumPy, bypassing GPU-CPU transfer guards and densification guards.
"""
if is_cupy_namespace(xp):
return xp.asnumpy(array)
if is_pydata_sparse_namespace(xp):
return array.todense() # type: ignore[attr-defined] # pyright: ignore[reportAttributeAccessIssue]
if is_torch_namespace(xp):
array = to_device(array, "cpu")
if is_array_api_strict_namespace(xp):
cpu: Device = xp.Device("CPU_DEVICE")
array = to_device(array, cpu)
if is_jax_namespace(xp):
import jax
# Note: only needed if the transfer guard is enabled
cpu = cast(Device, jax.devices("cpu")[0])
array = to_device(array, cpu)
return np.asarray(array)
def xp_assert_equal(
actual: Array,
desired: Array,
*,
err_msg: str = "",
check_dtype: bool = True,
check_shape: bool = True,
check_scalar: bool = False,
) -> None:
"""
Array-API compatible version of `np.testing.assert_array_equal`.
Parameters
----------
actual : Array
The array produced by the tested function.
desired : Array
The expected array (typically hardcoded).
err_msg : str, optional
Error message to display on failure.
check_dtype, check_shape : bool, default: True
Whether to check agreement between actual and desired dtypes and shapes
check_scalar : bool, default: False
NumPy only: whether to check agreement between actual and desired types -
0d array vs scalar.
See Also
--------
xp_assert_close : Similar function for inexact equality checks.
numpy.testing.assert_array_equal : Similar function for NumPy arrays.
"""
xp = _check_ns_shape_dtype(actual, desired, check_dtype, check_shape, check_scalar)
actual_np = as_numpy_array(actual, xp=xp)
desired_np = as_numpy_array(desired, xp=xp)
np.testing.assert_array_equal(actual_np, desired_np, err_msg=err_msg)
def xp_assert_less(
x: Array,
y: Array,
*,
err_msg: str = "",
check_dtype: bool = True,
check_shape: bool = True,
check_scalar: bool = False,
) -> None:
"""
Array-API compatible version of `np.testing.assert_array_less`.
Parameters
----------
x, y : Array
The arrays to compare according to ``x < y`` (elementwise).
err_msg : str, optional
Error message to display on failure.
check_dtype, check_shape : bool, default: True
Whether to check agreement between actual and desired dtypes and shapes
check_scalar : bool, default: False
NumPy only: whether to check agreement between actual and desired types -
0d array vs scalar.
See Also
--------
xp_assert_close : Similar function for inexact equality checks.
numpy.testing.assert_array_equal : Similar function for NumPy arrays.
"""
xp = _check_ns_shape_dtype(x, y, check_dtype, check_shape, check_scalar)
x_np = as_numpy_array(x, xp=xp)
y_np = as_numpy_array(y, xp=xp)
np.testing.assert_array_less(x_np, y_np, err_msg=err_msg)
def xp_assert_close(
actual: Array,
desired: Array,
*,
rtol: float | None = None,
atol: float = 0,
err_msg: str = "",
check_dtype: bool = True,
check_shape: bool = True,
check_scalar: bool = False,
) -> None:
"""
Array-API compatible version of `np.testing.assert_allclose`.
Parameters
----------
actual : Array
The array produced by the tested function.
desired : Array
The expected array (typically hardcoded).
rtol : float, optional
Relative tolerance. Default: dtype-dependent.
atol : float, optional
Absolute tolerance. Default: 0.
err_msg : str, optional
Error message to display on failure.
check_dtype, check_shape : bool, default: True
Whether to check agreement between actual and desired dtypes and shapes
check_scalar : bool, default: False
NumPy only: whether to check agreement between actual and desired types -
0d array vs scalar.
See Also
--------
xp_assert_equal : Similar function for exact equality checks.
isclose : Public function for checking closeness.
numpy.testing.assert_allclose : Similar function for NumPy arrays.
Notes
-----
The default `atol` and `rtol` differ from `xp.all(xpx.isclose(a, b))`.
"""
xp = _check_ns_shape_dtype(actual, desired, check_dtype, check_shape, check_scalar)
if rtol is None:
if xp.isdtype(actual.dtype, ("real floating", "complex floating")):
# multiplier of 4 is used as for `np.float64` this puts the default `rtol`
# roughly half way between sqrt(eps) and the default for
# `numpy.testing.assert_allclose`, 1e-7
rtol = xp.finfo(actual.dtype).eps ** 0.5 * 4
else:
rtol = 1e-7
actual_np = as_numpy_array(actual, xp=xp)
desired_np = as_numpy_array(desired, xp=xp)
np.testing.assert_allclose( # pyright: ignore[reportCallIssue]
actual_np,
desired_np,
rtol=rtol, # pyright: ignore[reportArgumentType]
atol=atol,
err_msg=err_msg,
)
def xfail(
request: pytest.FixtureRequest, *, reason: str, strict: bool | None = None
) -> None:
"""
XFAIL the currently running test.
Unlike ``pytest.xfail``, allow rest of test to execute instead of immediately
halting it, so that it may result in a XPASS.
xref https://github.com/pandas-dev/pandas/issues/38902
Parameters
----------
request : pytest.FixtureRequest
``request`` argument of the test function.
reason : str
Reason for the expected failure.
strict: bool, optional
If True, the test will be marked as failed if it passes.
If False, the test will be marked as passed if it fails.
Default: ``xfail_strict`` value in ``pyproject.toml``, or False if absent.
"""
if strict is not None:
marker = pytest.mark.xfail(reason=reason, strict=strict)
else:
marker = pytest.mark.xfail(reason=reason)
request.node.add_marker(marker)
@@ -0,0 +1 @@
"""Modules housing private utility functions."""
@@ -0,0 +1,74 @@
"""Acquire helpers from array-api-compat."""
# Allow packages that vendor both `array-api-extra` and
# `array-api-compat` to override the import location
# pylint: disable=duplicate-code
try:
from ...._array_api_compat_vendor import (
array_namespace,
device,
is_array_api_obj,
is_array_api_strict_namespace,
is_cupy_array,
is_cupy_namespace,
is_dask_array,
is_dask_namespace,
is_jax_array,
is_jax_namespace,
is_lazy_array,
is_numpy_array,
is_numpy_namespace,
is_pydata_sparse_array,
is_pydata_sparse_namespace,
is_torch_array,
is_torch_namespace,
is_writeable_array,
size,
to_device,
)
except ImportError:
from array_api_compat import (
array_namespace,
device,
is_array_api_obj,
is_array_api_strict_namespace,
is_cupy_array,
is_cupy_namespace,
is_dask_array,
is_dask_namespace,
is_jax_array,
is_jax_namespace,
is_lazy_array,
is_numpy_array,
is_numpy_namespace,
is_pydata_sparse_array,
is_pydata_sparse_namespace,
is_torch_array,
is_torch_namespace,
is_writeable_array,
size,
to_device,
)
__all__ = [
"array_namespace",
"device",
"is_array_api_obj",
"is_array_api_strict_namespace",
"is_cupy_array",
"is_cupy_namespace",
"is_dask_array",
"is_dask_namespace",
"is_jax_array",
"is_jax_namespace",
"is_lazy_array",
"is_numpy_array",
"is_numpy_namespace",
"is_pydata_sparse_array",
"is_pydata_sparse_namespace",
"is_torch_array",
"is_torch_namespace",
"is_writeable_array",
"size",
"to_device",
]
@@ -0,0 +1,45 @@
"""Static type stubs for `_compat.py`."""
# https://github.com/scikit-learn/scikit-learn/pull/27910#issuecomment-2568023972
from __future__ import annotations
from types import ModuleType
from typing import Any, TypeGuard
# TODO import from typing (requires Python >=3.13)
from typing_extensions import TypeIs
from ._typing import Array, Device
# pylint: disable=missing-class-docstring,unused-argument
def array_namespace(
*xs: Array | complex | None,
api_version: str | None = None,
use_compat: bool | None = None,
) -> ModuleType: ...
def device(x: Array, /) -> Device: ...
def is_array_api_obj(x: object, /) -> TypeIs[Array]: ...
def is_array_api_strict_namespace(xp: ModuleType, /) -> bool: ...
def is_cupy_namespace(xp: ModuleType, /) -> bool: ...
def is_dask_namespace(xp: ModuleType, /) -> bool: ...
def is_jax_namespace(xp: ModuleType, /) -> bool: ...
def is_numpy_namespace(xp: ModuleType, /) -> bool: ...
def is_pydata_sparse_namespace(xp: ModuleType, /) -> bool: ...
def is_torch_namespace(xp: ModuleType, /) -> bool: ...
def is_cupy_array(x: object, /) -> TypeGuard[Array]: ...
def is_dask_array(x: object, /) -> TypeGuard[Array]: ...
def is_jax_array(x: object, /) -> TypeGuard[Array]: ...
def is_numpy_array(x: object, /) -> TypeGuard[Array]: ...
def is_pydata_sparse_array(x: object, /) -> TypeGuard[Array]: ...
def is_torch_array(x: object, /) -> TypeGuard[Array]: ...
def is_lazy_array(x: object, /) -> TypeGuard[Array]: ...
def is_writeable_array(x: object, /) -> TypeGuard[Array]: ...
def size(x: Array, /) -> int | None: ...
def to_device( # type: ignore[explicit-any]
x: Array,
device: Device, # pylint: disable=redefined-outer-name
/,
*,
stream: int | Any | None = None,
) -> Array: ...
@@ -0,0 +1,559 @@
"""Helper functions used by `array_api_extra/_funcs.py`."""
from __future__ import annotations
import io
import math
import pickle
import types
from collections.abc import Callable, Generator, Iterable
from functools import wraps
from types import ModuleType
from typing import (
TYPE_CHECKING,
Any,
ClassVar,
Generic,
Literal,
ParamSpec,
TypeAlias,
TypeVar,
cast,
)
from . import _compat
from ._compat import (
array_namespace,
is_array_api_obj,
is_dask_namespace,
is_jax_namespace,
is_numpy_array,
is_pydata_sparse_namespace,
)
from ._typing import Array
if TYPE_CHECKING: # pragma: no cover
# TODO import from typing (requires Python >=3.12 and >=3.13)
from typing_extensions import TypeIs, override
else:
def override(func):
return func
P = ParamSpec("P")
T = TypeVar("T")
__all__ = [
"asarrays",
"capabilities",
"eager_shape",
"in1d",
"is_python_scalar",
"jax_autojit",
"mean",
"meta_namespace",
"pickle_flatten",
"pickle_unflatten",
]
def in1d(
x1: Array,
x2: Array,
/,
*,
assume_unique: bool = False,
invert: bool = False,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""
Check whether each element of an array is also present in a second array.
Returns a boolean array the same length as `x1` that is True
where an element of `x1` is in `x2` and False otherwise.
This function has been adapted using the original implementation
present in numpy:
https://github.com/numpy/numpy/blob/v1.26.0/numpy/lib/arraysetops.py#L524-L758
"""
if xp is None:
xp = array_namespace(x1, x2)
x1_shape = eager_shape(x1)
x2_shape = eager_shape(x2)
# This code is run to make the code significantly faster
if x2_shape[0] < 10 * x1_shape[0] ** 0.145 and isinstance(x2, Iterable):
if invert:
mask = xp.ones(x1_shape[0], dtype=xp.bool, device=_compat.device(x1))
for a in x2:
mask &= x1 != a
else:
mask = xp.zeros(x1_shape[0], dtype=xp.bool, device=_compat.device(x1))
for a in x2:
mask |= x1 == a
return mask
rev_idx = xp.empty(0) # placeholder
if not assume_unique:
x1, rev_idx = xp.unique_inverse(x1)
x2 = xp.unique_values(x2)
ar = xp.concat((x1, x2))
device_ = _compat.device(ar)
# We need this to be a stable sort.
order = xp.argsort(ar, stable=True)
reverse_order = xp.argsort(order, stable=True)
sar = xp.take(ar, order, axis=0)
ar_size = _compat.size(sar)
assert ar_size is not None, "xp.unique*() on lazy backends raises"
if ar_size >= 1:
bool_ar = sar[1:] != sar[:-1] if invert else sar[1:] == sar[:-1]
else:
bool_ar = xp.asarray([False]) if invert else xp.asarray([True])
flag = xp.concat((bool_ar, xp.asarray([invert], device=device_)))
ret = xp.take(flag, reverse_order, axis=0)
if assume_unique:
return ret[: x1.shape[0]]
return xp.take(ret, rev_idx, axis=0)
def mean(
x: Array,
/,
*,
axis: int | tuple[int, ...] | None = None,
keepdims: bool = False,
xp: ModuleType | None = None,
) -> Array: # numpydoc ignore=PR01,RT01
"""
Complex mean, https://github.com/data-apis/array-api/issues/846.
"""
if xp is None:
xp = array_namespace(x)
if xp.isdtype(x.dtype, "complex floating"):
x_real = xp.real(x)
x_imag = xp.imag(x)
mean_real = xp.mean(x_real, axis=axis, keepdims=keepdims)
mean_imag = xp.mean(x_imag, axis=axis, keepdims=keepdims)
return mean_real + (mean_imag * xp.asarray(1j))
return xp.mean(x, axis=axis, keepdims=keepdims)
def is_python_scalar(x: object) -> TypeIs[complex]: # numpydoc ignore=PR01,RT01
"""Return True if `x` is a Python scalar, False otherwise."""
# isinstance(x, float) returns True for np.float64
# isinstance(x, complex) returns True for np.complex128
# bool is a subclass of int
return isinstance(x, int | float | complex) and not is_numpy_array(x)
def asarrays(
a: Array | complex,
b: Array | complex,
xp: ModuleType,
) -> tuple[Array, Array]:
"""
Ensure both `a` and `b` are arrays.
If `b` is a python scalar, it is converted to the same dtype as `a`, and vice versa.
Behavior is not specified when mixing a Python ``float`` and an array with an
integer data type; this may give ``float32``, ``float64``, or raise an exception.
Behavior is implementation-specific.
Similarly, behavior is not specified when mixing a Python ``complex`` and an array
with a real-valued data type; this may give ``complex64``, ``complex128``, or raise
an exception. Behavior is implementation-specific.
Parameters
----------
a, b : Array | int | float | complex | bool
Input arrays or scalars. At least one must be an array.
xp : array_namespace, optional
The standard-compatible namespace for `x`. Default: infer.
Returns
-------
Array, Array
The input arrays, possibly converted to arrays if they were scalars.
See Also
--------
mixing-arrays-with-python-scalars : Array API specification for the behavior.
"""
a_scalar = is_python_scalar(a)
b_scalar = is_python_scalar(b)
if not a_scalar and not b_scalar:
# This includes misc. malformed input e.g. str
return a, b # type: ignore[return-value]
swap = False
if a_scalar:
swap = True
b, a = a, b
if is_array_api_obj(a):
# a is an Array API object
# b is a int | float | complex | bool
xa = a
# https://data-apis.org/array-api/draft/API_specification/type_promotion.html#mixing-arrays-with-python-scalars
same_dtype = {
bool: "bool",
int: ("integral", "real floating", "complex floating"),
float: ("real floating", "complex floating"),
complex: "complex floating",
}
kind = same_dtype[type(cast(complex, b))] # type: ignore[index]
if xp.isdtype(a.dtype, kind):
xb = xp.asarray(b, dtype=a.dtype)
else:
# Undefined behaviour. Let the function deal with it, if it can.
xb = xp.asarray(b)
else:
# Neither a nor b are Array API objects.
# Note: we can only reach this point when one explicitly passes
# xp=xp to the calling function; otherwise we fail earlier on
# array_namespace(a, b).
xa, xb = xp.asarray(a), xp.asarray(b)
return (xb, xa) if swap else (xa, xb)
def ndindex(*x: int) -> Generator[tuple[int, ...]]:
"""
Generate all N-dimensional indices for a given array shape.
Given the shape of an array, an ndindex instance iterates over the N-dimensional
index of the array. At each iteration a tuple of indices is returned, the last
dimension is iterated over first.
This has an identical API to numpy.ndindex.
Parameters
----------
*x : int
The shape of the array.
"""
if not x:
yield ()
return
for i in ndindex(*x[:-1]):
for j in range(x[-1]):
yield *i, j
def eager_shape(x: Array, /) -> tuple[int, ...]:
"""
Return shape of an array. Raise if shape is not fully defined.
Parameters
----------
x : Array
Input array.
Returns
-------
tuple[int, ...]
Shape of the array.
"""
shape = x.shape
# Dask arrays uses non-standard NaN instead of None
if any(s is None or math.isnan(s) for s in shape):
msg = "Unsupported lazy shape"
raise TypeError(msg)
return cast(tuple[int, ...], shape)
def meta_namespace(
*arrays: Array | complex | None, xp: ModuleType | None = None
) -> ModuleType:
"""
Get the namespace of Dask chunks.
On all other backends, just return the namespace of the arrays.
Parameters
----------
*arrays : Array | int | float | complex | bool | None
Input arrays.
xp : array_namespace, optional
The standard-compatible namespace for the input arrays. Default: infer.
Returns
-------
array_namespace
If xp is Dask, the namespace of the Dask chunks;
otherwise, the namespace of the arrays.
"""
xp = array_namespace(*arrays) if xp is None else xp
if not is_dask_namespace(xp):
return xp
# Quietly skip scalars and None's
metas = [cast(Array | None, getattr(a, "_meta", None)) for a in arrays]
return array_namespace(*metas)
def capabilities(xp: ModuleType) -> dict[str, int]:
"""
Return patched ``xp.__array_namespace_info__().capabilities()``.
TODO this helper should be eventually removed once all the special cases
it handles are fixed in the respective backends.
Parameters
----------
xp : array_namespace
The standard-compatible namespace.
Returns
-------
dict
Capabilities of the namespace.
"""
if is_pydata_sparse_namespace(xp):
# No __array_namespace_info__(); no indexing by sparse arrays
return {"boolean indexing": False, "data-dependent shapes": True}
out = xp.__array_namespace_info__().capabilities()
if is_jax_namespace(xp) and out["boolean indexing"]:
# FIXME https://github.com/jax-ml/jax/issues/27418
# Fixed in jax >=0.6.0
out = out.copy()
out["boolean indexing"] = False
return out
_BASIC_PICKLED_TYPES = frozenset((
bool, int, float, complex, str, bytes, bytearray,
list, tuple, dict, set, frozenset, range, slice,
types.NoneType, types.EllipsisType,
)) # fmt: skip
_BASIC_REST_TYPES = frozenset((
type, types.BuiltinFunctionType, types.FunctionType, types.ModuleType
)) # fmt: skip
FlattenRest: TypeAlias = tuple[object, ...]
def pickle_flatten(
obj: object, cls: type[T] | tuple[type[T], ...]
) -> tuple[list[T], FlattenRest]:
"""
Use the pickle machinery to extract objects out of an arbitrary container.
Unlike regular ``pickle.dumps``, this function always succeeds.
Parameters
----------
obj : object
The object to pickle.
cls : type | tuple[type, ...]
One or multiple classes to extract from the object.
The instances of these classes inside ``obj`` will not be pickled.
Returns
-------
instances : list[cls]
All instances of ``cls`` found inside ``obj`` (not pickled).
rest
Opaque object containing the pickled bytes plus all other objects where
``__reduce__`` / ``__reduce_ex__`` is either not implemented or raised.
These are unpickleable objects, types, modules, and functions.
This object is *typically* hashable save for fairly exotic objects
that are neither pickleable nor hashable.
This object is pickleable if everything except ``instances`` was pickleable
in the input object.
See Also
--------
pickle_unflatten : Reverse function.
Examples
--------
>>> class A:
... def __repr__(self):
... return "<A>"
>>> class NS:
... def __repr__(self):
... return "<NS>"
... def __reduce__(self):
... assert False, "not serializable"
>>> obj = {1: A(), 2: [A(), NS(), A()]}
>>> instances, rest = pickle_flatten(obj, A)
>>> instances
[<A>, <A>, <A>]
>>> pickle_unflatten(instances, rest)
{1: <A>, 2: [<A>, <NS>, <A>]}
This can be also used to swap inner objects; the only constraint is that
the number of objects in and out must be the same:
>>> pickle_unflatten(["foo", "bar", "baz"], rest)
{1: "foo", 2: ["bar", <NS>, "baz"]}
"""
instances: list[T] = []
rest: list[object] = []
class Pickler(pickle.Pickler): # numpydoc ignore=GL08
"""
Use the `pickle.Pickler.persistent_id` hook to extract objects.
"""
@override
def persistent_id(self, obj: object) -> Literal[0, 1, None]: # pyright: ignore[reportIncompatibleMethodOverride] # numpydoc ignore=GL08
if isinstance(obj, cls):
instances.append(obj) # type: ignore[arg-type]
return 0
typ_ = type(obj)
if typ_ in _BASIC_PICKLED_TYPES: # No subclasses!
# If obj is a collection, recursively descend inside it
return None
if typ_ in _BASIC_REST_TYPES:
rest.append(obj)
return 1
try:
# Note: a class that defines __slots__ without defining __getstate__
# cannot be pickled with __reduce__(), but can with __reduce_ex__(5)
_ = obj.__reduce_ex__(pickle.HIGHEST_PROTOCOL)
except Exception: # pylint: disable=broad-exception-caught
rest.append(obj)
return 1
# Object can be pickled. Let the Pickler recursively descend inside it.
return None
f = io.BytesIO()
p = Pickler(f, protocol=pickle.HIGHEST_PROTOCOL)
p.dump(obj)
return instances, (f.getvalue(), *rest)
def pickle_unflatten(instances: Iterable[object], rest: FlattenRest) -> Any: # type: ignore[explicit-any]
"""
Reverse of ``pickle_flatten``.
Parameters
----------
instances : Iterable
Inner objects to be reinserted into the flattened container.
rest : FlattenRest
Extra bits, as returned by ``pickle_flatten``.
Returns
-------
object
The outer object originally passed to ``pickle_flatten`` after a
pickle->unpickle round-trip.
See Also
--------
pickle_flatten : Serializing function.
pickle.loads : Standard unpickle function.
Notes
-----
The `instances` iterable must yield at least the same number of elements as the ones
returned by ``pickle_without``, but the elements do not need to be the same objects
or even the same types of objects. Excess elements, if any, will be left untouched.
"""
iters = iter(instances), iter(rest)
pik = cast(bytes, next(iters[1]))
class Unpickler(pickle.Unpickler): # numpydoc ignore=GL08
"""Mirror of the overridden Pickler in pickle_flatten."""
@override
def persistent_load(self, pid: Literal[0, 1]) -> object: # pyright: ignore[reportIncompatibleMethodOverride] # numpydoc ignore=GL08
try:
return next(iters[pid])
except StopIteration as e:
msg = "Not enough objects to unpickle"
raise ValueError(msg) from e
f = io.BytesIO(pik)
return Unpickler(f).load()
class _AutoJITWrapper(Generic[T]): # numpydoc ignore=PR01
"""
Helper of :func:`jax_autojit`.
Wrap arbitrary inputs and outputs of the jitted function and
convert them to/from PyTrees.
"""
obj: T
_registered: ClassVar[bool] = False
__slots__: tuple[str, ...] = ("obj",)
def __init__(self, obj: T) -> None: # numpydoc ignore=GL08
self._register()
self.obj = obj
@classmethod
def _register(cls): # numpydoc ignore=SS06
"""
Register upon first use instead of at import time, to avoid
globally importing JAX.
"""
if not cls._registered:
import jax
jax.tree_util.register_pytree_node(
cls,
lambda obj: pickle_flatten(obj, jax.Array), # pyright: ignore[reportUnknownArgumentType]
lambda aux_data, children: pickle_unflatten(children, aux_data), # pyright: ignore[reportUnknownArgumentType]
)
cls._registered = True
def jax_autojit(
func: Callable[P, T],
) -> Callable[P, T]: # numpydoc ignore=PR01,RT01,SS03
"""
Wrap `func` with ``jax.jit``, with the following differences:
- Python scalar arguments and return values are not automatically converted to
``jax.Array`` objects.
- All non-array arguments are automatically treated as static.
Unlike ``jax.jit``, static arguments must be either hashable or serializable with
``pickle``.
- Unlike ``jax.jit``, non-array arguments and return values are not limited to
tuple/list/dict, but can be any object serializable with ``pickle``.
- Automatically descend into non-array arguments and find ``jax.Array`` objects
inside them, then rebuild the arguments when entering `func`, swapping the JAX
concrete arrays with tracer objects.
- Automatically descend into non-array return values and find ``jax.Array`` objects
inside them, then rebuild them downstream of exiting the JIT, swapping the JAX
tracer objects with concrete arrays.
See Also
--------
jax.jit : JAX JIT compilation function.
"""
import jax
@jax.jit # type: ignore[misc] # pyright: ignore[reportUntypedFunctionDecorator]
def inner( # type: ignore[decorated-any,explicit-any] # numpydoc ignore=GL08
wargs: _AutoJITWrapper[Any],
) -> _AutoJITWrapper[T]:
args, kwargs = wargs.obj
res = func(*args, **kwargs) # pyright: ignore[reportCallIssue]
return _AutoJITWrapper(res)
@wraps(func)
def outer(*args: P.args, **kwargs: P.kwargs) -> T: # numpydoc ignore=GL08
wargs = _AutoJITWrapper((args, kwargs))
return inner(wargs).obj
return outer
@@ -0,0 +1,10 @@
# numpydoc ignore=GL08
# pylint: disable=missing-module-docstring,duplicate-code
Array = object
DType = object
Device = object
GetIndex = object
SetIndex = object
__all__ = ["Array", "DType", "Device", "GetIndex", "SetIndex"]
@@ -0,0 +1,105 @@
"""Static typing helpers."""
from __future__ import annotations
from types import EllipsisType
from typing import Protocol, TypeAlias
# TODO import from typing (requires Python >=3.12)
from typing_extensions import override
# TODO: use array-api-typing once it is available
class Array(Protocol): # pylint: disable=missing-class-docstring
# Unary operations
def __abs__(self) -> Array: ...
def __pos__(self) -> Array: ...
def __neg__(self) -> Array: ...
def __invert__(self) -> Array: ...
# Binary operations
def __add__(self, other: Array | complex, /) -> Array: ...
def __sub__(self, other: Array | complex, /) -> Array: ...
def __mul__(self, other: Array | complex, /) -> Array: ...
def __truediv__(self, other: Array | complex, /) -> Array: ...
def __floordiv__(self, other: Array | complex, /) -> Array: ...
def __mod__(self, other: Array | complex, /) -> Array: ...
def __pow__(self, other: Array | complex, /) -> Array: ...
def __matmul__(self, other: Array, /) -> Array: ...
def __and__(self, other: Array | int, /) -> Array: ...
def __or__(self, other: Array | int, /) -> Array: ...
def __xor__(self, other: Array | int, /) -> Array: ...
def __lshift__(self, other: Array | int, /) -> Array: ...
def __rshift__(self, other: Array | int, /) -> Array: ...
def __lt__(self, other: Array | complex, /) -> Array: ...
def __le__(self, other: Array | complex, /) -> Array: ...
def __gt__(self, other: Array | complex, /) -> Array: ...
def __ge__(self, other: Array | complex, /) -> Array: ...
@override
def __eq__(self, other: Array | complex, /) -> Array: ... # type: ignore[override] # pyright: ignore[reportIncompatibleMethodOverride]
@override
def __ne__(self, other: Array | complex, /) -> Array: ... # type: ignore[override] # pyright: ignore[reportIncompatibleMethodOverride]
# Reflected operations
def __radd__(self, other: Array | complex, /) -> Array: ...
def __rsub__(self, other: Array | complex, /) -> Array: ...
def __rmul__(self, other: Array | complex, /) -> Array: ...
def __rtruediv__(self, other: Array | complex, /) -> Array: ...
def __rfloordiv__(self, other: Array | complex, /) -> Array: ...
def __rmod__(self, other: Array | complex, /) -> Array: ...
def __rpow__(self, other: Array | complex, /) -> Array: ...
def __rmatmul__(self, other: Array, /) -> Array: ...
def __rand__(self, other: Array | int, /) -> Array: ...
def __ror__(self, other: Array | int, /) -> Array: ...
def __rxor__(self, other: Array | int, /) -> Array: ...
def __rlshift__(self, other: Array | int, /) -> Array: ...
def __rrshift__(self, other: Array | int, /) -> Array: ...
# Attributes
@property
def dtype(self) -> DType: ...
@property
def device(self) -> Device: ...
@property
def mT(self) -> Array: ... # pylint: disable=invalid-name
@property
def ndim(self) -> int: ...
@property
def shape(self) -> tuple[int | None, ...]: ...
@property
def size(self) -> int | None: ...
@property
def T(self) -> Array: ... # pylint: disable=invalid-name
# Collection operations (note: an Array does not have to be Sized or Iterable)
def __getitem__(self, key: GetIndex, /) -> Array: ...
def __setitem__(self, key: SetIndex, value: Array | complex, /) -> None: ...
# Materialization methods (may raise on lazy arrays)
def __bool__(self) -> bool: ...
def __complex__(self) -> complex: ...
def __float__(self) -> float: ...
def __index__(self) -> int: ...
def __int__(self) -> int: ...
# Misc methods (frequently not implemented in Arrays wrapped by array-api-compat)
# def __array_namespace__(*, api_version: str | None) -> ModuleType: ...
# def __dlpack__(
# *,
# stream: int | Any | None = None,
# max_version: tuple[int, int] | None = None,
# dl_device: tuple[int, int] | None = None, # tuple[Enum, int]
# copy: bool | None = None,
# ) -> Any: ...
# def __dlpack_device__() -> tuple[int, int]: ... # tuple[Enum, int]
# def to_device(device: Device, /, *, stream: int | Any | None = None) -> Array: ...
class DType(Protocol): # pylint: disable=missing-class-docstring
pass
class Device(Protocol): # pylint: disable=missing-class-docstring
pass
SetIndex: TypeAlias = ( # type: ignore[explicit-any]
int | slice | EllipsisType | Array | tuple[int | slice | EllipsisType | Array, ...]
)
GetIndex: TypeAlias = ( # type: ignore[explicit-any]
SetIndex | None | tuple[int | slice | EllipsisType | None | Array, ...]
)
__all__ = ["Array", "DType", "Device", "GetIndex", "SetIndex"]
@@ -0,0 +1,359 @@
"""
Public testing utilities.
See also _lib._testing for additional private testing utilities.
"""
from __future__ import annotations
import contextlib
import enum
import warnings
from collections.abc import Callable, Iterator, Sequence
from functools import wraps
from types import ModuleType
from typing import TYPE_CHECKING, Any, ParamSpec, TypeVar, cast
from ._lib._utils._compat import is_dask_namespace, is_jax_namespace
from ._lib._utils._helpers import jax_autojit, pickle_flatten, pickle_unflatten
__all__ = ["lazy_xp_function", "patch_lazy_xp_functions"]
if TYPE_CHECKING: # pragma: no cover
# TODO import override from typing (requires Python >=3.12)
import pytest
from dask.typing import Graph, Key, SchedulerGetCallable
from typing_extensions import override
else:
# Sphinx hacks
SchedulerGetCallable = object
def override(func):
return func
P = ParamSpec("P")
T = TypeVar("T")
_ufuncs_tags: dict[object, dict[str, Any]] = {} # type: ignore[explicit-any]
class Deprecated(enum.Enum):
"""Unique type for deprecated parameters."""
DEPRECATED = 1
DEPRECATED = Deprecated.DEPRECATED
def lazy_xp_function( # type: ignore[explicit-any]
func: Callable[..., Any],
*,
allow_dask_compute: bool | int = False,
jax_jit: bool = True,
static_argnums: Deprecated = DEPRECATED,
static_argnames: Deprecated = DEPRECATED,
) -> None: # numpydoc ignore=GL07
"""
Tag a function to be tested on lazy backends.
Tag a function so that when any tests are executed with ``xp=jax.numpy`` the
function is replaced with a jitted version of itself, and when it is executed with
``xp=dask.array`` the function will raise if it attempts to materialize the graph.
This will be later expanded to provide test coverage for other lazy backends.
In order for the tag to be effective, the test or a fixture must call
:func:`patch_lazy_xp_functions`.
Parameters
----------
func : callable
Function to be tested.
allow_dask_compute : bool | int, optional
Whether `func` is allowed to internally materialize the Dask graph, or maximum
number of times it is allowed to do so. This is typically triggered by
``bool()``, ``float()``, or ``np.asarray()``.
Set to 1 if you are aware that `func` converts the input parameters to NumPy and
want to let it do so at least for the time being, knowing that it is going to be
extremely detrimental for performance.
If a test needs values higher than 1 to pass, it is a canary that the conversion
to NumPy/bool/float is happening multiple times, which translates to multiple
computations of the whole graph. Short of making the function fully lazy, you
should at least add explicit calls to ``np.asarray()`` early in the function.
*Note:* the counter of `allow_dask_compute` resets after each call to `func`, so
a test function that invokes `func` multiple times should still work with this
parameter set to 1.
Set to True to allow `func` to materialize the graph an unlimited number
of times.
Default: False, meaning that `func` must be fully lazy and never materialize the
graph.
jax_jit : bool, optional
Set to True to replace `func` with a smart variant of ``jax.jit(func)`` after
calling the :func:`patch_lazy_xp_functions` test helper with ``xp=jax.numpy``.
Set to False if `func` is only compatible with eager (non-jitted) JAX.
Unlike with vanilla ``jax.jit``, all arguments and return types that are not JAX
arrays are treated as static; the function can accept and return arbitrary
wrappers around JAX arrays. This difference is because, in real life, most users
won't wrap the function directly with ``jax.jit`` but rather they will use it
within their own code, which is itself then wrapped by ``jax.jit``, and
internally consume the function's outputs.
In other words, the pattern that is being tested is::
>>> @jax.jit
... def user_func(x):
... y = user_prepares_inputs(x)
... z = func(y, some_static_arg=True)
... return user_consumes(z)
Default: True.
static_argnums :
Deprecated; ignored
static_argnames :
Deprecated; ignored
See Also
--------
patch_lazy_xp_functions : Companion function to call from the test or fixture.
jax.jit : JAX function to compile a function for performance.
Examples
--------
In ``test_mymodule.py``::
from array_api_extra.testing import lazy_xp_function from mymodule import myfunc
lazy_xp_function(myfunc)
def test_myfunc(xp):
a = xp.asarray([1, 2])
# When xp=jax.numpy, this is similar to `b = jax.jit(myfunc)(a)`
# When xp=dask.array, crash on compute() or persist()
b = myfunc(a)
Notes
-----
In order for this tag to be effective, the test function must be imported into the
test module globals without its namespace; alternatively its namespace must be
declared in a ``lazy_xp_modules`` list in the test module globals.
Example 1::
from mymodule import myfunc
lazy_xp_function(myfunc)
def test_myfunc(xp):
x = myfunc(xp.asarray([1, 2]))
Example 2::
import mymodule
lazy_xp_modules = [mymodule]
lazy_xp_function(mymodule.myfunc)
def test_myfunc(xp):
x = mymodule.myfunc(xp.asarray([1, 2]))
A test function can circumvent this monkey-patching system by using a namespace
outside of the two above patterns. You need to sanitize your code to make sure this
only happens intentionally.
Example 1::
import mymodule
from mymodule import myfunc
lazy_xp_function(myfunc)
def test_myfunc(xp):
a = xp.asarray([1, 2])
b = myfunc(a) # This is wrapped when xp=jax.numpy or xp=dask.array
c = mymodule.myfunc(a) # This is not
Example 2::
import mymodule
class naked:
myfunc = mymodule.myfunc
lazy_xp_modules = [mymodule]
lazy_xp_function(mymodule.myfunc)
def test_myfunc(xp):
a = xp.asarray([1, 2])
b = mymodule.myfunc(a) # This is wrapped when xp=jax.numpy or xp=dask.array
c = naked.myfunc(a) # This is not
"""
if static_argnums is not DEPRECATED or static_argnames is not DEPRECATED:
warnings.warn(
(
"The `static_argnums` and `static_argnames` parameters are deprecated "
"and ignored. They will be removed in a future version."
),
DeprecationWarning,
stacklevel=2,
)
tags = {
"allow_dask_compute": allow_dask_compute,
"jax_jit": jax_jit,
}
try:
func._lazy_xp_function = tags # type: ignore[attr-defined] # pylint: disable=protected-access # pyright: ignore[reportFunctionMemberAccess]
except AttributeError: # @cython.vectorize
_ufuncs_tags[func] = tags
def patch_lazy_xp_functions(
request: pytest.FixtureRequest, monkeypatch: pytest.MonkeyPatch, *, xp: ModuleType
) -> None:
"""
Test lazy execution of functions tagged with :func:`lazy_xp_function`.
If ``xp==jax.numpy``, search for all functions which have been tagged with
:func:`lazy_xp_function` in the globals of the module that defines the current test,
as well as in the ``lazy_xp_modules`` list in the globals of the same module,
and wrap them with :func:`jax.jit`. Unwrap them at the end of the test.
If ``xp==dask.array``, wrap the functions with a decorator that disables
``compute()`` and ``persist()`` and ensures that exceptions and warnings are raised
eagerly.
This function should be typically called by your library's `xp` fixture that runs
tests on multiple backends::
@pytest.fixture(params=[numpy, array_api_strict, jax.numpy, dask.array])
def xp(request, monkeypatch):
patch_lazy_xp_functions(request, monkeypatch, xp=request.param)
return request.param
but it can be otherwise be called by the test itself too.
Parameters
----------
request : pytest.FixtureRequest
Pytest fixture, as acquired by the test itself or by one of its fixtures.
monkeypatch : pytest.MonkeyPatch
Pytest fixture, as acquired by the test itself or by one of its fixtures.
xp : array_namespace
Array namespace to be tested.
See Also
--------
lazy_xp_function : Tag a function to be tested on lazy backends.
pytest.FixtureRequest : `request` test function parameter.
"""
mod = cast(ModuleType, request.module)
mods = [mod, *cast(list[ModuleType], getattr(mod, "lazy_xp_modules", []))]
def iter_tagged() -> ( # type: ignore[explicit-any]
Iterator[tuple[ModuleType, str, Callable[..., Any], dict[str, Any]]]
):
for mod in mods:
for name, func in mod.__dict__.items():
tags: dict[str, Any] | None = None # type: ignore[explicit-any]
with contextlib.suppress(AttributeError):
tags = func._lazy_xp_function # pylint: disable=protected-access
if tags is None:
with contextlib.suppress(KeyError, TypeError):
tags = _ufuncs_tags[func]
if tags is not None:
yield mod, name, func, tags
if is_dask_namespace(xp):
for mod, name, func, tags in iter_tagged():
n = tags["allow_dask_compute"]
if n is True:
n = 1_000_000
elif n is False:
n = 0
wrapped = _dask_wrap(func, n)
monkeypatch.setattr(mod, name, wrapped)
elif is_jax_namespace(xp):
for mod, name, func, tags in iter_tagged():
if tags["jax_jit"]:
wrapped = jax_autojit(func)
monkeypatch.setattr(mod, name, wrapped)
class CountingDaskScheduler(SchedulerGetCallable):
"""
Dask scheduler that counts how many times `dask.compute` is called.
If the number of times exceeds 'max_count', it raises an error.
This is a wrapper around Dask's own 'synchronous' scheduler.
Parameters
----------
max_count : int
Maximum number of allowed calls to `dask.compute`.
msg : str
Assertion to raise when the count exceeds `max_count`.
"""
count: int
max_count: int
msg: str
def __init__(self, max_count: int, msg: str): # numpydoc ignore=GL08
self.count = 0
self.max_count = max_count
self.msg = msg
@override
def __call__(self, dsk: Graph, keys: Sequence[Key] | Key, **kwargs: Any) -> Any: # type: ignore[decorated-any,explicit-any] # numpydoc ignore=GL08
import dask
self.count += 1
# This should yield a nice traceback to the
# offending line in the user's code
assert self.count <= self.max_count, self.msg
return dask.get(dsk, keys, **kwargs) # type: ignore[attr-defined,no-untyped-call] # pyright: ignore[reportPrivateImportUsage]
def _dask_wrap(
func: Callable[P, T], n: int
) -> Callable[P, T]: # numpydoc ignore=PR01,RT01
"""
Wrap `func` to raise if it attempts to call `dask.compute` more than `n` times.
After the function returns, materialize the graph in order to re-raise exceptions.
"""
import dask
import dask.array as da
func_name = getattr(func, "__name__", str(func))
n_str = f"only up to {n}" if n else "no"
msg = (
f"Called `dask.compute()` or `dask.persist()` {n + 1} times, "
f"but {n_str} calls are allowed. Set "
f"`lazy_xp_function({func_name}, allow_dask_compute={n + 1})` "
"to allow for more (but note that this will harm performance). "
)
@wraps(func)
def wrapper(*args: P.args, **kwargs: P.kwargs) -> T: # numpydoc ignore=GL08
scheduler = CountingDaskScheduler(n, msg)
with dask.config.set({"scheduler": scheduler}): # pyright: ignore[reportPrivateImportUsage]
out = func(*args, **kwargs)
# Block until the graph materializes and reraise exceptions. This allows
# `pytest.raises` and `pytest.warns` to work as expected. Note that this would
# not work on scheduler='distributed', as it would not block.
arrays, rest = pickle_flatten(out, da.Array)
arrays = dask.persist(arrays, scheduler="threads")[0] # type: ignore[attr-defined,no-untyped-call,func-returns-value,index] # pyright: ignore[reportPrivateImportUsage]
return pickle_unflatten(arrays, rest) # pyright: ignore[reportUnknownArgumentType]
return wrapper
@@ -0,0 +1,20 @@
from .main import minimize
from .utils import show_versions
# PEP0440 compatible formatted version, see:
# https://www.python.org/dev/peps/pep-0440/
#
# Final release markers:
# X.Y.0 # For first release after an increment in Y
# X.Y.Z # For bugfix releases
#
# Admissible pre-release markers:
# X.YaN # Alpha release
# X.YbN # Beta release
# X.YrcN # Release Candidate
#
# Dev branch marker is: 'X.Y.dev' or 'X.Y.devN' where N is an integer.
# 'X.Y.dev0' is the canonical version of 'X.Y.dev'.
__version__ = "1.1.2"
__all__ = ["minimize", "show_versions"]
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,132 @@
import sys
from enum import Enum
import numpy as np
# Exit status.
class ExitStatus(Enum):
"""
Exit statuses.
"""
RADIUS_SUCCESS = 0
TARGET_SUCCESS = 1
FIXED_SUCCESS = 2
CALLBACK_SUCCESS = 3
FEASIBLE_SUCCESS = 4
MAX_EVAL_WARNING = 5
MAX_ITER_WARNING = 6
INFEASIBLE_ERROR = -1
LINALG_ERROR = -2
class Options(str, Enum):
"""
Options.
"""
DEBUG = "debug"
FEASIBILITY_TOL = "feasibility_tol"
FILTER_SIZE = "filter_size"
HISTORY_SIZE = "history_size"
MAX_EVAL = "maxfev"
MAX_ITER = "maxiter"
NPT = "nb_points"
RHOBEG = "radius_init"
RHOEND = "radius_final"
SCALE = "scale"
STORE_HISTORY = "store_history"
TARGET = "target"
VERBOSE = "disp"
class Constants(str, Enum):
"""
Constants.
"""
DECREASE_RADIUS_FACTOR = "decrease_radius_factor"
INCREASE_RADIUS_FACTOR = "increase_radius_factor"
INCREASE_RADIUS_THRESHOLD = "increase_radius_threshold"
DECREASE_RADIUS_THRESHOLD = "decrease_radius_threshold"
DECREASE_RESOLUTION_FACTOR = "decrease_resolution_factor"
LARGE_RESOLUTION_THRESHOLD = "large_resolution_threshold"
MODERATE_RESOLUTION_THRESHOLD = "moderate_resolution_threshold"
LOW_RATIO = "low_ratio"
HIGH_RATIO = "high_ratio"
VERY_LOW_RATIO = "very_low_ratio"
PENALTY_INCREASE_THRESHOLD = "penalty_increase_threshold"
PENALTY_INCREASE_FACTOR = "penalty_increase_factor"
SHORT_STEP_THRESHOLD = "short_step_threshold"
LOW_RADIUS_FACTOR = "low_radius_factor"
BYRD_OMOJOKUN_FACTOR = "byrd_omojokun_factor"
THRESHOLD_RATIO_CONSTRAINTS = "threshold_ratio_constraints"
LARGE_SHIFT_FACTOR = "large_shift_factor"
LARGE_GRADIENT_FACTOR = "large_gradient_factor"
RESOLUTION_FACTOR = "resolution_factor"
IMPROVE_TCG = "improve_tcg"
# Default options.
DEFAULT_OPTIONS = {
Options.DEBUG.value: False,
Options.FEASIBILITY_TOL.value: np.sqrt(np.finfo(float).eps),
Options.FILTER_SIZE.value: sys.maxsize,
Options.HISTORY_SIZE.value: sys.maxsize,
Options.MAX_EVAL.value: lambda n: 500 * n,
Options.MAX_ITER.value: lambda n: 1000 * n,
Options.NPT.value: lambda n: 2 * n + 1,
Options.RHOBEG.value: 1.0,
Options.RHOEND.value: 1e-6,
Options.SCALE.value: False,
Options.STORE_HISTORY.value: False,
Options.TARGET.value: -np.inf,
Options.VERBOSE.value: False,
}
# Default constants.
DEFAULT_CONSTANTS = {
Constants.DECREASE_RADIUS_FACTOR.value: 0.5,
Constants.INCREASE_RADIUS_FACTOR.value: np.sqrt(2.0),
Constants.INCREASE_RADIUS_THRESHOLD.value: 2.0,
Constants.DECREASE_RADIUS_THRESHOLD.value: 1.4,
Constants.DECREASE_RESOLUTION_FACTOR.value: 0.1,
Constants.LARGE_RESOLUTION_THRESHOLD.value: 250.0,
Constants.MODERATE_RESOLUTION_THRESHOLD.value: 16.0,
Constants.LOW_RATIO.value: 0.1,
Constants.HIGH_RATIO.value: 0.7,
Constants.VERY_LOW_RATIO.value: 0.01,
Constants.PENALTY_INCREASE_THRESHOLD.value: 1.5,
Constants.PENALTY_INCREASE_FACTOR.value: 2.0,
Constants.SHORT_STEP_THRESHOLD.value: 0.5,
Constants.LOW_RADIUS_FACTOR.value: 0.1,
Constants.BYRD_OMOJOKUN_FACTOR.value: 0.8,
Constants.THRESHOLD_RATIO_CONSTRAINTS.value: 2.0,
Constants.LARGE_SHIFT_FACTOR.value: 10.0,
Constants.LARGE_GRADIENT_FACTOR.value: 10.0,
Constants.RESOLUTION_FACTOR.value: 2.0,
Constants.IMPROVE_TCG.value: True,
}
# Printing options.
PRINT_OPTIONS = {
"threshold": 6,
"edgeitems": 2,
"linewidth": sys.maxsize,
"formatter": {
"float_kind": lambda x: np.format_float_scientific(
x,
precision=3,
unique=False,
pad_left=2,
)
},
}
# Constants.
BARRIER = 2.0 ** min(
100,
np.finfo(float).maxexp // 2,
-np.finfo(float).minexp // 2,
)
@@ -0,0 +1,14 @@
from .geometry import cauchy_geometry, spider_geometry
from .optim import (
tangential_byrd_omojokun,
constrained_tangential_byrd_omojokun,
normal_byrd_omojokun,
)
__all__ = [
"cauchy_geometry",
"spider_geometry",
"tangential_byrd_omojokun",
"constrained_tangential_byrd_omojokun",
"normal_byrd_omojokun",
]
@@ -0,0 +1,387 @@
import inspect
import numpy as np
from ..utils import get_arrays_tol
TINY = np.finfo(float).tiny
def cauchy_geometry(const, grad, curv, xl, xu, delta, debug):
r"""
Maximize approximately the absolute value of a quadratic function subject
to bound constraints in a trust region.
This function solves approximately
.. math::
\max_{s \in \mathbb{R}^n} \quad \bigg\lvert c + g^{\mathsf{T}} s +
\frac{1}{2} s^{\mathsf{T}} H s \bigg\rvert \quad \text{s.t.} \quad
\left\{ \begin{array}{l}
l \le s \le u,\\
\lVert s \rVert \le \Delta,
\end{array} \right.
by maximizing the objective function along the constrained Cauchy
direction.
Parameters
----------
const : float
Constant :math:`c` as shown above.
grad : `numpy.ndarray`, shape (n,)
Gradient :math:`g` as shown above.
curv : callable
Curvature of :math:`H` along any vector.
``curv(s) -> float``
returns :math:`s^{\mathsf{T}} H s`.
xl : `numpy.ndarray`, shape (n,)
Lower bounds :math:`l` as shown above.
xu : `numpy.ndarray`, shape (n,)
Upper bounds :math:`u` as shown above.
delta : float
Trust-region radius :math:`\Delta` as shown above.
debug : bool
Whether to make debugging tests during the execution.
Returns
-------
`numpy.ndarray`, shape (n,)
Approximate solution :math:`s`.
Notes
-----
This function is described as the first alternative in Section 6.5 of [1]_.
It is assumed that the origin is feasible with respect to the bound
constraints and that `delta` is finite and positive.
References
----------
.. [1] T. M. Ragonneau. *Model-Based Derivative-Free Optimization Methods
and Software*. PhD thesis, Department of Applied Mathematics, The Hong
Kong Polytechnic University, Hong Kong, China, 2022. URL:
https://theses.lib.polyu.edu.hk/handle/200/12294.
"""
if debug:
assert isinstance(const, float)
assert isinstance(grad, np.ndarray) and grad.ndim == 1
assert inspect.signature(curv).bind(grad)
assert isinstance(xl, np.ndarray) and xl.shape == grad.shape
assert isinstance(xu, np.ndarray) and xu.shape == grad.shape
assert isinstance(delta, float)
assert isinstance(debug, bool)
tol = get_arrays_tol(xl, xu)
assert np.all(xl <= tol)
assert np.all(xu >= -tol)
assert np.isfinite(delta) and delta > 0.0
xl = np.minimum(xl, 0.0)
xu = np.maximum(xu, 0.0)
# To maximize the absolute value of a quadratic function, we maximize the
# function itself or its negative, and we choose the solution that provides
# the largest function value.
step1, q_val1 = _cauchy_geom(const, grad, curv, xl, xu, delta, debug)
step2, q_val2 = _cauchy_geom(
-const,
-grad,
lambda x: -curv(x),
xl,
xu,
delta,
debug,
)
step = step1 if abs(q_val1) >= abs(q_val2) else step2
if debug:
assert np.all(xl <= step)
assert np.all(step <= xu)
assert np.linalg.norm(step) < 1.1 * delta
return step
def spider_geometry(const, grad, curv, xpt, xl, xu, delta, debug):
r"""
Maximize approximately the absolute value of a quadratic function subject
to bound constraints in a trust region.
This function solves approximately
.. math::
\max_{s \in \mathbb{R}^n} \quad \bigg\lvert c + g^{\mathsf{T}} s +
\frac{1}{2} s^{\mathsf{T}} H s \bigg\rvert \quad \text{s.t.} \quad
\left\{ \begin{array}{l}
l \le s \le u,\\
\lVert s \rVert \le \Delta,
\end{array} \right.
by maximizing the objective function along given straight lines.
Parameters
----------
const : float
Constant :math:`c` as shown above.
grad : `numpy.ndarray`, shape (n,)
Gradient :math:`g` as shown above.
curv : callable
Curvature of :math:`H` along any vector.
``curv(s) -> float``
returns :math:`s^{\mathsf{T}} H s`.
xpt : `numpy.ndarray`, shape (n, npt)
Points defining the straight lines. The straight lines considered are
the ones passing through the origin and the points in `xpt`.
xl : `numpy.ndarray`, shape (n,)
Lower bounds :math:`l` as shown above.
xu : `numpy.ndarray`, shape (n,)
Upper bounds :math:`u` as shown above.
delta : float
Trust-region radius :math:`\Delta` as shown above.
debug : bool
Whether to make debugging tests during the execution.
Returns
-------
`numpy.ndarray`, shape (n,)
Approximate solution :math:`s`.
Notes
-----
This function is described as the second alternative in Section 6.5 of
[1]_. It is assumed that the origin is feasible with respect to the bound
constraints and that `delta` is finite and positive.
References
----------
.. [1] T. M. Ragonneau. *Model-Based Derivative-Free Optimization Methods
and Software*. PhD thesis, Department of Applied Mathematics, The Hong
Kong Polytechnic University, Hong Kong, China, 2022. URL:
https://theses.lib.polyu.edu.hk/handle/200/12294.
"""
if debug:
assert isinstance(const, float)
assert isinstance(grad, np.ndarray) and grad.ndim == 1
assert inspect.signature(curv).bind(grad)
assert (
isinstance(xpt, np.ndarray)
and xpt.ndim == 2
and xpt.shape[0] == grad.size
)
assert isinstance(xl, np.ndarray) and xl.shape == grad.shape
assert isinstance(xu, np.ndarray) and xu.shape == grad.shape
assert isinstance(delta, float)
assert isinstance(debug, bool)
tol = get_arrays_tol(xl, xu)
assert np.all(xl <= tol)
assert np.all(xu >= -tol)
assert np.isfinite(delta) and delta > 0.0
xl = np.minimum(xl, 0.0)
xu = np.maximum(xu, 0.0)
# Iterate through the straight lines.
step = np.zeros_like(grad)
q_val = const
s_norm = np.linalg.norm(xpt, axis=0)
# Set alpha_xl to the step size for the lower-bound constraint and
# alpha_xu to the step size for the upper-bound constraint.
# xl.shape = (N,)
# xpt.shape = (N, M)
# i_xl_pos.shape = (M, N)
i_xl_pos = (xl > -np.inf) & (xpt.T > -TINY * xl)
i_xl_neg = (xl > -np.inf) & (xpt.T < TINY * xl)
i_xu_pos = (xu < np.inf) & (xpt.T > TINY * xu)
i_xu_neg = (xu < np.inf) & (xpt.T < -TINY * xu)
# (M, N)
alpha_xl_pos = np.atleast_2d(
np.broadcast_to(xl, i_xl_pos.shape)[i_xl_pos] / xpt.T[i_xl_pos]
)
# (M,)
alpha_xl_pos = np.max(alpha_xl_pos, axis=1, initial=-np.inf)
# make sure it's (M,)
alpha_xl_pos = np.broadcast_to(np.atleast_1d(alpha_xl_pos), xpt.shape[1])
alpha_xl_neg = np.atleast_2d(
np.broadcast_to(xl, i_xl_neg.shape)[i_xl_neg] / xpt.T[i_xl_neg]
)
alpha_xl_neg = np.max(alpha_xl_neg, axis=1, initial=np.inf)
alpha_xl_neg = np.broadcast_to(np.atleast_1d(alpha_xl_neg), xpt.shape[1])
alpha_xu_neg = np.atleast_2d(
np.broadcast_to(xu, i_xu_neg.shape)[i_xu_neg] / xpt.T[i_xu_neg]
)
alpha_xu_neg = np.max(alpha_xu_neg, axis=1, initial=-np.inf)
alpha_xu_neg = np.broadcast_to(np.atleast_1d(alpha_xu_neg), xpt.shape[1])
alpha_xu_pos = np.atleast_2d(
np.broadcast_to(xu, i_xu_pos.shape)[i_xu_pos] / xpt.T[i_xu_pos]
)
alpha_xu_pos = np.max(alpha_xu_pos, axis=1, initial=np.inf)
alpha_xu_pos = np.broadcast_to(np.atleast_1d(alpha_xu_pos), xpt.shape[1])
for k in range(xpt.shape[1]):
# Set alpha_tr to the step size for the trust-region constraint.
if s_norm[k] > TINY * delta:
alpha_tr = max(delta / s_norm[k], 0.0)
else:
# The current straight line is basically zero.
continue
alpha_bd_pos = max(min(alpha_xu_pos[k], alpha_xl_neg[k]), 0.0)
alpha_bd_neg = min(max(alpha_xl_pos[k], alpha_xu_neg[k]), 0.0)
# Set alpha_quad_pos and alpha_quad_neg to the step size to the extrema
# of the quadratic function along the positive and negative directions.
grad_step = grad @ xpt[:, k]
curv_step = curv(xpt[:, k])
if (
grad_step >= 0.0
and curv_step < -TINY * grad_step
or grad_step <= 0.0
and curv_step > -TINY * grad_step
):
alpha_quad_pos = max(-grad_step / curv_step, 0.0)
else:
alpha_quad_pos = np.inf
if (
grad_step >= 0.0
and curv_step > TINY * grad_step
or grad_step <= 0.0
and curv_step < TINY * grad_step
):
alpha_quad_neg = min(-grad_step / curv_step, 0.0)
else:
alpha_quad_neg = -np.inf
# Select the step that provides the largest value of the objective
# function if it improves the current best. The best positive step is
# either the one that reaches the constraints or the one that reaches
# the extremum of the objective function along the current direction
# (only possible if the resulting step is feasible). We test both, and
# we perform similar calculations along the negative step.
# N.B.: we select the largest possible step among all the ones that
# maximize the objective function. This is to avoid returning the zero
# step in some extreme cases.
alpha_pos = min(alpha_tr, alpha_bd_pos)
alpha_neg = max(-alpha_tr, alpha_bd_neg)
q_val_pos = (
const + alpha_pos * grad_step + 0.5 * alpha_pos**2.0 * curv_step
)
q_val_neg = (
const + alpha_neg * grad_step + 0.5 * alpha_neg**2.0 * curv_step
)
if alpha_quad_pos < alpha_pos:
q_val_quad_pos = (
const
+ alpha_quad_pos * grad_step
+ 0.5 * alpha_quad_pos**2.0 * curv_step
)
if abs(q_val_quad_pos) > abs(q_val_pos):
alpha_pos = alpha_quad_pos
q_val_pos = q_val_quad_pos
if alpha_quad_neg > alpha_neg:
q_val_quad_neg = (
const
+ alpha_quad_neg * grad_step
+ 0.5 * alpha_quad_neg**2.0 * curv_step
)
if abs(q_val_quad_neg) > abs(q_val_neg):
alpha_neg = alpha_quad_neg
q_val_neg = q_val_quad_neg
if abs(q_val_pos) >= abs(q_val_neg) and abs(q_val_pos) > abs(q_val):
step = np.clip(alpha_pos * xpt[:, k], xl, xu)
q_val = q_val_pos
elif abs(q_val_neg) > abs(q_val_pos) and abs(q_val_neg) > abs(q_val):
step = np.clip(alpha_neg * xpt[:, k], xl, xu)
q_val = q_val_neg
if debug:
assert np.all(xl <= step)
assert np.all(step <= xu)
assert np.linalg.norm(step) < 1.1 * delta
return step
def _cauchy_geom(const, grad, curv, xl, xu, delta, debug):
"""
Same as `bound_constrained_cauchy_step` without the absolute value.
"""
# Calculate the initial active set.
fixed_xl = (xl < 0.0) & (grad > 0.0)
fixed_xu = (xu > 0.0) & (grad < 0.0)
# Calculate the Cauchy step.
cauchy_step = np.zeros_like(grad)
cauchy_step[fixed_xl] = xl[fixed_xl]
cauchy_step[fixed_xu] = xu[fixed_xu]
if np.linalg.norm(cauchy_step) > delta:
working = fixed_xl | fixed_xu
while True:
# Calculate the Cauchy step for the directions in the working set.
g_norm = np.linalg.norm(grad[working])
delta_reduced = np.sqrt(
delta**2.0 - cauchy_step[~working] @ cauchy_step[~working]
)
if g_norm > TINY * abs(delta_reduced):
mu = max(delta_reduced / g_norm, 0.0)
else:
break
cauchy_step[working] = mu * grad[working]
# Update the working set.
fixed_xl = working & (cauchy_step < xl)
fixed_xu = working & (cauchy_step > xu)
if not np.any(fixed_xl) and not np.any(fixed_xu):
# Stop the calculations as the Cauchy step is now feasible.
break
cauchy_step[fixed_xl] = xl[fixed_xl]
cauchy_step[fixed_xu] = xu[fixed_xu]
working = working & ~(fixed_xl | fixed_xu)
# Calculate the step that maximizes the quadratic along the Cauchy step.
grad_step = grad @ cauchy_step
if grad_step >= 0.0:
# Set alpha_tr to the step size for the trust-region constraint.
s_norm = np.linalg.norm(cauchy_step)
if s_norm > TINY * delta:
alpha_tr = max(delta / s_norm, 0.0)
else:
# The Cauchy step is basically zero.
alpha_tr = 0.0
# Set alpha_quad to the step size for the maximization problem.
curv_step = curv(cauchy_step)
if curv_step < -TINY * grad_step:
alpha_quad = max(-grad_step / curv_step, 0.0)
else:
alpha_quad = np.inf
# Set alpha_bd to the step size for the bound constraints.
i_xl = (xl > -np.inf) & (cauchy_step < TINY * xl)
i_xu = (xu < np.inf) & (cauchy_step > TINY * xu)
alpha_xl = np.min(xl[i_xl] / cauchy_step[i_xl], initial=np.inf)
alpha_xu = np.min(xu[i_xu] / cauchy_step[i_xu], initial=np.inf)
alpha_bd = min(alpha_xl, alpha_xu)
# Calculate the solution and the corresponding function value.
alpha = min(alpha_tr, alpha_quad, alpha_bd)
step = np.clip(alpha * cauchy_step, xl, xu)
q_val = const + alpha * grad_step + 0.5 * alpha**2.0 * curv_step
else:
# This case is never reached in exact arithmetic. It prevents this
# function to return a step that decreases the objective function.
step = np.zeros_like(grad)
q_val = const
if debug:
assert np.all(xl <= step)
assert np.all(step <= xu)
assert np.linalg.norm(step) < 1.1 * delta
return step, q_val
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,18 @@
from .exceptions import (
MaxEvalError,
TargetSuccess,
CallbackSuccess,
FeasibleSuccess,
)
from .math import get_arrays_tol, exact_1d_array
from .versions import show_versions
__all__ = [
"MaxEvalError",
"TargetSuccess",
"CallbackSuccess",
"FeasibleSuccess",
"get_arrays_tol",
"exact_1d_array",
"show_versions",
]
@@ -0,0 +1,22 @@
class MaxEvalError(Exception):
"""
Exception raised when the maximum number of evaluations is reached.
"""
class TargetSuccess(Exception):
"""
Exception raised when the target value is reached.
"""
class CallbackSuccess(StopIteration):
"""
Exception raised when the callback function raises a ``StopIteration``.
"""
class FeasibleSuccess(Exception):
"""
Exception raised when a feasible point of a feasible problem is found.
"""
@@ -0,0 +1,77 @@
import numpy as np
EPS = np.finfo(float).eps
def get_arrays_tol(*arrays):
"""
Get a relative tolerance for a set of arrays.
Parameters
----------
*arrays: tuple
Set of `numpy.ndarray` to get the tolerance for.
Returns
-------
float
Relative tolerance for the set of arrays.
Raises
------
ValueError
If no array is provided.
"""
if len(arrays) == 0:
raise ValueError("At least one array must be provided.")
size = max(array.size for array in arrays)
weight = max(
np.max(np.abs(array[np.isfinite(array)]), initial=1.0)
for array in arrays
)
return 10.0 * EPS * max(size, 1.0) * weight
def exact_1d_array(x, message):
"""
Preprocess a 1-dimensional array.
Parameters
----------
x : array_like
Array to be preprocessed.
message : str
Error message if `x` cannot be interpreter as a 1-dimensional array.
Returns
-------
`numpy.ndarray`
Preprocessed array.
"""
x = np.atleast_1d(np.squeeze(x)).astype(float)
if x.ndim != 1:
raise ValueError(message)
return x
def exact_2d_array(x, message):
"""
Preprocess a 2-dimensional array.
Parameters
----------
x : array_like
Array to be preprocessed.
message : str
Error message if `x` cannot be interpreter as a 2-dimensional array.
Returns
-------
`numpy.ndarray`
Preprocessed array.
"""
x = np.atleast_2d(x).astype(float)
if x.ndim != 2:
raise ValueError(message)
return x
@@ -0,0 +1,67 @@
import os
import platform
import sys
from importlib.metadata import PackageNotFoundError, version
def _get_sys_info():
"""
Get useful system information.
Returns
-------
dict
Useful system information.
"""
return {
"python": sys.version.replace(os.linesep, " "),
"executable": sys.executable,
"machine": platform.platform(),
}
def _get_deps_info():
"""
Get the versions of the dependencies.
Returns
-------
dict
Versions of the dependencies.
"""
deps = ["cobyqa", "numpy", "scipy", "setuptools", "pip"]
deps_info = {}
for module in deps:
try:
deps_info[module] = version(module)
except PackageNotFoundError:
deps_info[module] = None
return deps_info
def show_versions():
"""
Display useful system and dependencies information.
When reporting issues, please include this information.
"""
print("System settings")
print("---------------")
sys_info = _get_sys_info()
print(
"\n".join(
f"{k:>{max(map(len, sys_info.keys())) + 1}}: {v}"
for k, v in sys_info.items()
)
)
print()
print("Python dependencies")
print("-------------------")
deps_info = _get_deps_info()
print(
"\n".join(
f"{k:>{max(map(len, deps_info.keys())) + 1}}: {v}"
for k, v in deps_info.items()
)
)
@@ -0,0 +1,399 @@
# ######################### LICENSE ############################ #
# Copyright (c) 2005-2015, Michele Simionato
# All rights reserved.
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are
# met:
# Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
# Redistributions in bytecode form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in
# the documentation and/or other materials provided with the
# distribution.
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
# HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
# INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
# BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
# OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
# ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR
# TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
# USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
# DAMAGE.
"""
Decorator module, see https://pypi.python.org/pypi/decorator
for the documentation.
"""
import re
import sys
import inspect
import operator
import itertools
import collections
from inspect import getfullargspec
__version__ = '4.0.5'
def get_init(cls):
return cls.__init__
# getargspec has been deprecated in Python 3.5
ArgSpec = collections.namedtuple(
'ArgSpec', 'args varargs varkw defaults')
def getargspec(f):
"""A replacement for inspect.getargspec"""
spec = getfullargspec(f)
return ArgSpec(spec.args, spec.varargs, spec.varkw, spec.defaults)
DEF = re.compile(r'\s*def\s*([_\w][_\w\d]*)\s*\(')
# basic functionality
class FunctionMaker:
"""
An object with the ability to create functions with a given signature.
It has attributes name, doc, module, signature, defaults, dict, and
methods update and make.
"""
# Atomic get-and-increment provided by the GIL
_compile_count = itertools.count()
def __init__(self, func=None, name=None, signature=None,
defaults=None, doc=None, module=None, funcdict=None):
self.shortsignature = signature
if func:
# func can be a class or a callable, but not an instance method
self.name = func.__name__
if self.name == '<lambda>': # small hack for lambda functions
self.name = '_lambda_'
self.doc = func.__doc__
self.module = func.__module__
if inspect.isfunction(func):
argspec = getfullargspec(func)
self.annotations = getattr(func, '__annotations__', {})
for a in ('args', 'varargs', 'varkw', 'defaults', 'kwonlyargs',
'kwonlydefaults'):
setattr(self, a, getattr(argspec, a))
for i, arg in enumerate(self.args):
setattr(self, f'arg{i}', arg)
allargs = list(self.args)
allshortargs = list(self.args)
if self.varargs:
allargs.append('*' + self.varargs)
allshortargs.append('*' + self.varargs)
elif self.kwonlyargs:
allargs.append('*') # single star syntax
for a in self.kwonlyargs:
allargs.append(f'{a}=None')
allshortargs.append(f'{a}={a}')
if self.varkw:
allargs.append('**' + self.varkw)
allshortargs.append('**' + self.varkw)
self.signature = ', '.join(allargs)
self.shortsignature = ', '.join(allshortargs)
self.dict = func.__dict__.copy()
# func=None happens when decorating a caller
if name:
self.name = name
if signature is not None:
self.signature = signature
if defaults:
self.defaults = defaults
if doc:
self.doc = doc
if module:
self.module = module
if funcdict:
self.dict = funcdict
# check existence required attributes
assert hasattr(self, 'name')
if not hasattr(self, 'signature'):
raise TypeError(f'You are decorating a non-function: {func}')
def update(self, func, **kw):
"Update the signature of func with the data in self"
func.__name__ = self.name
func.__doc__ = getattr(self, 'doc', None)
func.__dict__ = getattr(self, 'dict', {})
func.__defaults__ = getattr(self, 'defaults', ())
func.__kwdefaults__ = getattr(self, 'kwonlydefaults', None)
func.__annotations__ = getattr(self, 'annotations', None)
try:
frame = sys._getframe(3)
except AttributeError: # for IronPython and similar implementations
callermodule = '?'
else:
callermodule = frame.f_globals.get('__name__', '?')
func.__module__ = getattr(self, 'module', callermodule)
func.__dict__.update(kw)
def make(self, src_templ, evaldict=None, addsource=False, **attrs):
"Make a new function from a given template and update the signature"
src = src_templ % vars(self) # expand name and signature
evaldict = evaldict or {}
mo = DEF.match(src)
if mo is None:
raise SyntaxError(f'not a valid function template\n{src}')
name = mo.group(1) # extract the function name
names = set([name] + [arg.strip(' *') for arg in
self.shortsignature.split(',')])
for n in names:
if n in ('_func_', '_call_'):
raise NameError(f'{n} is overridden in\n{src}')
if not src.endswith('\n'): # add a newline just for safety
src += '\n' # this is needed in old versions of Python
# Ensure each generated function has a unique filename for profilers
# (such as cProfile) that depend on the tuple of (<filename>,
# <definition line>, <function name>) being unique.
filename = f'<decorator-gen-{next(self._compile_count)}>'
try:
code = compile(src, filename, 'single')
exec(code, evaldict)
except: # noqa: E722
print('Error in generated code:', file=sys.stderr)
print(src, file=sys.stderr)
raise
func = evaldict[name]
if addsource:
attrs['__source__'] = src
self.update(func, **attrs)
return func
@classmethod
def create(cls, obj, body, evaldict, defaults=None,
doc=None, module=None, addsource=True, **attrs):
"""
Create a function from the strings name, signature, and body.
evaldict is the evaluation dictionary. If addsource is true, an
attribute __source__ is added to the result. The attributes attrs
are added, if any.
"""
if isinstance(obj, str): # "name(signature)"
name, rest = obj.strip().split('(', 1)
signature = rest[:-1] # strip a right parens
func = None
else: # a function
name = None
signature = None
func = obj
self = cls(func, name, signature, defaults, doc, module)
ibody = '\n'.join(' ' + line for line in body.splitlines())
return self.make('def %(name)s(%(signature)s):\n' + ibody,
evaldict, addsource, **attrs)
def decorate(func, caller):
"""
decorate(func, caller) decorates a function using a caller.
"""
evaldict = func.__globals__.copy()
evaldict['_call_'] = caller
evaldict['_func_'] = func
fun = FunctionMaker.create(
func, "return _call_(_func_, %(shortsignature)s)",
evaldict, __wrapped__=func)
if hasattr(func, '__qualname__'):
fun.__qualname__ = func.__qualname__
return fun
def decorator(caller, _func=None):
"""decorator(caller) converts a caller function into a decorator"""
if _func is not None: # return a decorated function
# this is obsolete behavior; you should use decorate instead
return decorate(_func, caller)
# else return a decorator function
if inspect.isclass(caller):
name = caller.__name__.lower()
callerfunc = get_init(caller)
doc = (f'decorator({caller.__name__}) converts functions/generators into '
f'factories of {caller.__name__} objects')
elif inspect.isfunction(caller):
if caller.__name__ == '<lambda>':
name = '_lambda_'
else:
name = caller.__name__
callerfunc = caller
doc = caller.__doc__
else: # assume caller is an object with a __call__ method
name = caller.__class__.__name__.lower()
callerfunc = caller.__call__.__func__
doc = caller.__call__.__doc__
evaldict = callerfunc.__globals__.copy()
evaldict['_call_'] = caller
evaldict['_decorate_'] = decorate
return FunctionMaker.create(
f'{name}(func)', 'return _decorate_(func, _call_)',
evaldict, doc=doc, module=caller.__module__,
__wrapped__=caller)
# ####################### contextmanager ####################### #
try: # Python >= 3.2
from contextlib import _GeneratorContextManager
except ImportError: # Python >= 2.5
from contextlib import GeneratorContextManager as _GeneratorContextManager
class ContextManager(_GeneratorContextManager):
def __call__(self, func):
"""Context manager decorator"""
return FunctionMaker.create(
func, "with _self_: return _func_(%(shortsignature)s)",
dict(_self_=self, _func_=func), __wrapped__=func)
init = getfullargspec(_GeneratorContextManager.__init__)
n_args = len(init.args)
if n_args == 2 and not init.varargs: # (self, genobj) Python 2.7
def __init__(self, g, *a, **k):
return _GeneratorContextManager.__init__(self, g(*a, **k))
ContextManager.__init__ = __init__
elif n_args == 2 and init.varargs: # (self, gen, *a, **k) Python 3.4
pass
elif n_args == 4: # (self, gen, args, kwds) Python 3.5
def __init__(self, g, *a, **k):
return _GeneratorContextManager.__init__(self, g, a, k)
ContextManager.__init__ = __init__
contextmanager = decorator(ContextManager)
# ############################ dispatch_on ############################ #
def append(a, vancestors):
"""
Append ``a`` to the list of the virtual ancestors, unless it is already
included.
"""
add = True
for j, va in enumerate(vancestors):
if issubclass(va, a):
add = False
break
if issubclass(a, va):
vancestors[j] = a
add = False
if add:
vancestors.append(a)
# inspired from simplegeneric by P.J. Eby and functools.singledispatch
def dispatch_on(*dispatch_args):
"""
Factory of decorators turning a function into a generic function
dispatching on the given arguments.
"""
assert dispatch_args, 'No dispatch args passed'
dispatch_str = f"({', '.join(dispatch_args)},)"
def check(arguments, wrong=operator.ne, msg=''):
"""Make sure one passes the expected number of arguments"""
if wrong(len(arguments), len(dispatch_args)):
raise TypeError(f'Expected {len(dispatch_args)} arguments, '
'got {len(arguments)}{msg}')
def gen_func_dec(func):
"""Decorator turning a function into a generic function"""
# first check the dispatch arguments
argset = set(getfullargspec(func).args)
if not set(dispatch_args) <= argset:
raise NameError(f'Unknown dispatch arguments {dispatch_str}')
typemap = {}
def vancestors(*types):
"""
Get a list of sets of virtual ancestors for the given types
"""
check(types)
ras = [[] for _ in range(len(dispatch_args))]
for types_ in typemap:
for t, type_, ra in zip(types, types_, ras):
if issubclass(t, type_) and type_ not in t.__mro__:
append(type_, ra)
return [set(ra) for ra in ras]
def ancestors(*types):
"""
Get a list of virtual MROs, one for each type
"""
check(types)
lists = []
for t, vas in zip(types, vancestors(*types)):
n_vas = len(vas)
if n_vas > 1:
raise RuntimeError(
f'Ambiguous dispatch for {t}: {vas}')
elif n_vas == 1:
va, = vas
mro = type('t', (t, va), {}).__mro__[1:]
else:
mro = t.__mro__
lists.append(mro[:-1]) # discard t and object
return lists
def register(*types):
"""
Decorator to register an implementation for the given types
"""
check(types)
def dec(f):
check(getfullargspec(f).args, operator.lt, ' in ' + f.__name__)
typemap[types] = f
return f
return dec
def dispatch_info(*types):
"""
An utility to introspect the dispatch algorithm
"""
check(types)
lst = [tuple(a.__name__ for a in anc)
for anc in itertools.product(*ancestors(*types))]
return lst
def _dispatch(dispatch_args, *args, **kw):
types = tuple(type(arg) for arg in dispatch_args)
try: # fast path
f = typemap[types]
except KeyError:
pass
else:
return f(*args, **kw)
combinations = itertools.product(*ancestors(*types))
next(combinations) # the first one has been already tried
for types_ in combinations:
f = typemap.get(types_)
if f is not None:
return f(*args, **kw)
# else call the default implementation
return func(*args, **kw)
return FunctionMaker.create(
func, f'return _f_({dispatch_str}, %%(shortsignature)s)',
dict(_f_=_dispatch), register=register, default=func,
typemap=typemap, vancestors=vancestors, ancestors=ancestors,
dispatch_info=dispatch_info, __wrapped__=func)
gen_func_dec.__name__ = 'dispatch_on' + dispatch_str
return gen_func_dec
@@ -0,0 +1,274 @@
from inspect import Parameter, signature
import functools
import warnings
from importlib import import_module
from scipy._lib._docscrape import FunctionDoc
__all__ = ["_deprecated"]
# Object to use as default value for arguments to be deprecated. This should
# be used over 'None' as the user could parse 'None' as a positional argument
_NoValue = object()
def _sub_module_deprecation(*, sub_package, module, private_modules, all,
attribute, correct_module=None, dep_version="1.16.0"):
"""Helper function for deprecating modules that are public but were
intended to be private.
Parameters
----------
sub_package : str
Subpackage the module belongs to eg. stats
module : str
Public but intended private module to deprecate
private_modules : list
Private replacement(s) for `module`; should contain the
content of ``all``, possibly spread over several modules.
all : list
``__all__`` belonging to `module`
attribute : str
The attribute in `module` being accessed
correct_module : str, optional
Module in `sub_package` that `attribute` should be imported from.
Default is that `attribute` should be imported from ``scipy.sub_package``.
dep_version : str, optional
Version in which deprecated attributes will be removed.
"""
if correct_module is not None:
correct_import = f"scipy.{sub_package}.{correct_module}"
else:
correct_import = f"scipy.{sub_package}"
if attribute not in all:
raise AttributeError(
f"`scipy.{sub_package}.{module}` has no attribute `{attribute}`; "
f"furthermore, `scipy.{sub_package}.{module}` is deprecated "
f"and will be removed in SciPy 2.0.0."
)
attr = getattr(import_module(correct_import), attribute, None)
if attr is not None:
message = (
f"Please import `{attribute}` from the `{correct_import}` namespace; "
f"the `scipy.{sub_package}.{module}` namespace is deprecated "
f"and will be removed in SciPy 2.0.0."
)
else:
message = (
f"`scipy.{sub_package}.{module}.{attribute}` is deprecated along with "
f"the `scipy.{sub_package}.{module}` namespace. "
f"`scipy.{sub_package}.{module}.{attribute}` will be removed "
f"in SciPy {dep_version}, and the `scipy.{sub_package}.{module}` namespace "
f"will be removed in SciPy 2.0.0."
)
warnings.warn(message, category=DeprecationWarning, stacklevel=3)
for module in private_modules:
try:
return getattr(import_module(f"scipy.{sub_package}.{module}"), attribute)
except AttributeError as e:
# still raise an error if the attribute isn't in any of the expected
# private modules
if module == private_modules[-1]:
raise e
continue
def _deprecated(msg, stacklevel=2):
"""Deprecate a function by emitting a warning on use."""
def wrap(fun):
if isinstance(fun, type):
warnings.warn(
f"Trying to deprecate class {fun!r}",
category=RuntimeWarning, stacklevel=2)
return fun
@functools.wraps(fun)
def call(*args, **kwargs):
warnings.warn(msg, category=DeprecationWarning,
stacklevel=stacklevel)
return fun(*args, **kwargs)
call.__doc__ = fun.__doc__
return call
return wrap
class _DeprecationHelperStr:
"""
Helper class used by deprecate_cython_api
"""
def __init__(self, content, message):
self._content = content
self._message = message
def __hash__(self):
return hash(self._content)
def __eq__(self, other):
res = (self._content == other)
if res:
warnings.warn(self._message, category=DeprecationWarning,
stacklevel=2)
return res
def deprecate_cython_api(module, routine_name, new_name=None, message=None):
"""
Deprecate an exported cdef function in a public Cython API module.
Only functions can be deprecated; typedefs etc. cannot.
Parameters
----------
module : module
Public Cython API module (e.g. scipy.linalg.cython_blas).
routine_name : str
Name of the routine to deprecate. May also be a fused-type
routine (in which case its all specializations are deprecated).
new_name : str
New name to include in the deprecation warning message
message : str
Additional text in the deprecation warning message
Examples
--------
Usually, this function would be used in the top-level of the
module ``.pyx`` file:
>>> from scipy._lib.deprecation import deprecate_cython_api
>>> import scipy.linalg.cython_blas as mod
>>> deprecate_cython_api(mod, "dgemm", "dgemm_new",
... message="Deprecated in Scipy 1.5.0")
>>> del deprecate_cython_api, mod
After this, Cython modules that use the deprecated function emit a
deprecation warning when they are imported.
"""
old_name = f"{module.__name__}.{routine_name}"
if new_name is None:
depdoc = f"`{old_name}` is deprecated!"
else:
depdoc = f"`{old_name}` is deprecated, use `{new_name}` instead!"
if message is not None:
depdoc += "\n" + message
d = module.__pyx_capi__
# Check if the function is a fused-type function with a mangled name
j = 0
has_fused = False
while True:
fused_name = f"__pyx_fuse_{j}{routine_name}"
if fused_name in d:
has_fused = True
d[_DeprecationHelperStr(fused_name, depdoc)] = d.pop(fused_name)
j += 1
else:
break
# If not, apply deprecation to the named routine
if not has_fused:
d[_DeprecationHelperStr(routine_name, depdoc)] = d.pop(routine_name)
# taken from scikit-learn, see
# https://github.com/scikit-learn/scikit-learn/blob/1.3.0/sklearn/utils/validation.py#L38
def _deprecate_positional_args(func=None, *, version=None,
deprecated_args=None, custom_message=""):
"""Decorator for methods that issues warnings for positional arguments.
Using the keyword-only argument syntax in pep 3102, arguments after the
* will issue a warning when passed as a positional argument.
Parameters
----------
func : callable, default=None
Function to check arguments on.
version : callable, default=None
The version when positional arguments will result in error.
deprecated_args : set of str, optional
Arguments to deprecate - whether passed by position or keyword.
custom_message : str, optional
Custom message to add to deprecation warning and documentation.
"""
if version is None:
msg = "Need to specify a version where signature will be changed"
raise ValueError(msg)
deprecated_args = set() if deprecated_args is None else set(deprecated_args)
def _inner_deprecate_positional_args(f):
sig = signature(f)
kwonly_args = []
all_args = []
for name, param in sig.parameters.items():
if param.kind == Parameter.POSITIONAL_OR_KEYWORD:
all_args.append(name)
elif param.kind == Parameter.KEYWORD_ONLY:
kwonly_args.append(name)
def warn_deprecated_args(kwargs):
intersection = deprecated_args.intersection(kwargs)
if intersection:
message = (f"Arguments {intersection} are deprecated, whether passed "
"by position or keyword. They will be removed in SciPy "
f"{version}. ")
message += custom_message
warnings.warn(message, category=DeprecationWarning, stacklevel=3)
@functools.wraps(f)
def inner_f(*args, **kwargs):
extra_args = len(args) - len(all_args)
if extra_args <= 0:
warn_deprecated_args(kwargs)
return f(*args, **kwargs)
# extra_args > 0
kwonly_extra_args = set(kwonly_args[:extra_args]) - deprecated_args
args_msg = ", ".join(kwonly_extra_args)
warnings.warn(
(
f"You are passing as positional arguments: {args_msg}. "
"Please change your invocation to use keyword arguments. "
f"From SciPy {version}, passing these as positional "
"arguments will result in an error."
),
DeprecationWarning,
stacklevel=2,
)
kwargs.update(zip(sig.parameters, args))
warn_deprecated_args(kwargs)
return f(**kwargs)
doc = FunctionDoc(inner_f)
kwonly_extra_args = set(kwonly_args) - deprecated_args
admonition = f"""
.. deprecated:: {version}
Use of argument(s) ``{kwonly_extra_args}`` by position is deprecated; beginning in
SciPy {version}, these will be keyword-only. """
if deprecated_args:
admonition += (f"Argument(s) ``{deprecated_args}`` are deprecated, whether "
"passed by position or keyword; they will be removed in "
f"SciPy {version}. ")
admonition += custom_message
doc['Extended Summary'] += [admonition]
doc = str(doc).split("\n", 1)[1] # remove signature
inner_f.__doc__ = str(doc)
return inner_f
if func is not None:
return _inner_deprecate_positional_args(func)
return _inner_deprecate_positional_args
@@ -0,0 +1,366 @@
"""Utilities to allow inserting docstring fragments for common
parameters into function and method docstrings."""
from collections.abc import Callable, Iterable, Mapping
from typing import Protocol, TypeVar
import sys
__all__ = [
"docformat",
"inherit_docstring_from",
"indentcount_lines",
"filldoc",
"unindent_dict",
"unindent_string",
"extend_notes_in_docstring",
"replace_notes_in_docstring",
"doc_replace",
]
_F = TypeVar("_F", bound=Callable[..., object])
class Decorator(Protocol):
"""A decorator of a function."""
def __call__(self, func: _F, /) -> _F: ...
def docformat(docstring: str, docdict: Mapping[str, str] | None = None) -> str:
"""Fill a function docstring from variables in dictionary.
Adapt the indent of the inserted docs
Parameters
----------
docstring : str
A docstring from a function, possibly with dict formatting strings.
docdict : dict[str, str], optional
A dictionary with keys that match the dict formatting strings
and values that are docstring fragments to be inserted. The
indentation of the inserted docstrings is set to match the
minimum indentation of the ``docstring`` by adding this
indentation to all lines of the inserted string, except the
first.
Returns
-------
docstring : str
string with requested ``docdict`` strings inserted.
Examples
--------
>>> docformat(' Test string with %(value)s', {'value':'inserted value'})
' Test string with inserted value'
>>> docstring = 'First line\\n Second line\\n %(value)s'
>>> inserted_string = "indented\\nstring"
>>> docdict = {'value': inserted_string}
>>> docformat(docstring, docdict)
'First line\\n Second line\\n indented\\n string'
"""
if not docstring:
return docstring
if docdict is None:
docdict = {}
if not docdict:
return docstring
lines = docstring.expandtabs().splitlines()
# Find the minimum indent of the main docstring, after first line
if len(lines) < 2:
icount = 0
else:
icount = indentcount_lines(lines[1:])
indent = " " * icount
# Insert this indent to dictionary docstrings
indented = {}
for name, dstr in docdict.items():
lines = dstr.expandtabs().splitlines()
indented[name] = ("\n" + indent).join(lines)
return docstring % indented
def inherit_docstring_from(cls: object) -> Decorator:
"""This decorator modifies the decorated function's docstring by
replacing occurrences of '%(super)s' with the docstring of the
method of the same name from the class `cls`.
If the decorated method has no docstring, it is simply given the
docstring of `cls`s method.
Parameters
----------
cls : type or object
A class with a method with the same name as the decorated method.
The docstring of the method in this class replaces '%(super)s' in the
docstring of the decorated method.
Returns
-------
decfunc : function
The decorator function that modifies the __doc__ attribute
of its argument.
Examples
--------
In the following, the docstring for Bar.func created using the
docstring of `Foo.func`.
>>> class Foo:
... def func(self):
... '''Do something useful.'''
... return
...
>>> class Bar(Foo):
... @inherit_docstring_from(Foo)
... def func(self):
... '''%(super)s
... Do it fast.
... '''
... return
...
>>> b = Bar()
>>> b.func.__doc__
'Do something useful.\n Do it fast.\n '
"""
def _doc(func: _F) -> _F:
cls_docstring = getattr(cls, func.__name__).__doc__
func_docstring = func.__doc__
if func_docstring is None:
func.__doc__ = cls_docstring
else:
new_docstring = func_docstring % dict(super=cls_docstring)
func.__doc__ = new_docstring
return func
return _doc
def extend_notes_in_docstring(cls: object, notes: str) -> Decorator:
"""This decorator replaces the decorated function's docstring
with the docstring from corresponding method in `cls`.
It extends the 'Notes' section of that docstring to include
the given `notes`.
Parameters
----------
cls : type or object
A class with a method with the same name as the decorated method.
The docstring of the method in this class replaces the docstring of the
decorated method.
notes : str
Additional notes to append to the 'Notes' section of the docstring.
Returns
-------
decfunc : function
The decorator function that modifies the __doc__ attribute
of its argument.
"""
def _doc(func: _F) -> _F:
cls_docstring = getattr(cls, func.__name__).__doc__
# If python is called with -OO option,
# there is no docstring
if cls_docstring is None:
return func
end_of_notes = cls_docstring.find(" References\n")
if end_of_notes == -1:
end_of_notes = cls_docstring.find(" Examples\n")
if end_of_notes == -1:
end_of_notes = len(cls_docstring)
func.__doc__ = (
cls_docstring[:end_of_notes] + notes + cls_docstring[end_of_notes:]
)
return func
return _doc
def replace_notes_in_docstring(cls: object, notes: str) -> Decorator:
"""This decorator replaces the decorated function's docstring
with the docstring from corresponding method in `cls`.
It replaces the 'Notes' section of that docstring with
the given `notes`.
Parameters
----------
cls : type or object
A class with a method with the same name as the decorated method.
The docstring of the method in this class replaces the docstring of the
decorated method.
notes : str
The notes to replace the existing 'Notes' section with.
Returns
-------
decfunc : function
The decorator function that modifies the __doc__ attribute
of its argument.
"""
def _doc(func: _F) -> _F:
cls_docstring = getattr(cls, func.__name__).__doc__
notes_header = " Notes\n -----\n"
# If python is called with -OO option,
# there is no docstring
if cls_docstring is None:
return func
start_of_notes = cls_docstring.find(notes_header)
end_of_notes = cls_docstring.find(" References\n")
if end_of_notes == -1:
end_of_notes = cls_docstring.find(" Examples\n")
if end_of_notes == -1:
end_of_notes = len(cls_docstring)
func.__doc__ = (
cls_docstring[: start_of_notes + len(notes_header)]
+ notes
+ cls_docstring[end_of_notes:]
)
return func
return _doc
def indentcount_lines(lines: Iterable[str]) -> int:
"""Minimum indent for all lines in line list
Parameters
----------
lines : Iterable[str]
The lines to find the minimum indent of.
Returns
-------
indent : int
The minimum indent.
Examples
--------
>>> lines = [' one', ' two', ' three']
>>> indentcount_lines(lines)
1
>>> lines = []
>>> indentcount_lines(lines)
0
>>> lines = [' one']
>>> indentcount_lines(lines)
1
>>> indentcount_lines([' '])
0
"""
indentno = sys.maxsize
for line in lines:
stripped = line.lstrip()
if stripped:
indentno = min(indentno, len(line) - len(stripped))
if indentno == sys.maxsize:
return 0
return indentno
def filldoc(docdict: Mapping[str, str], unindent_params: bool = True) -> Decorator:
"""Return docstring decorator using docdict variable dictionary.
Parameters
----------
docdict : dict[str, str]
A dictionary containing name, docstring fragment pairs.
unindent_params : bool, optional
If True, strip common indentation from all parameters in docdict.
Default is False.
Returns
-------
decfunc : function
The decorator function that applies dictionary to its
argument's __doc__ attribute.
"""
if unindent_params:
docdict = unindent_dict(docdict)
def decorate(func: _F) -> _F:
# __doc__ may be None for optimized Python (-OO)
doc = func.__doc__ or ""
func.__doc__ = docformat(doc, docdict)
return func
return decorate
def unindent_dict(docdict: Mapping[str, str]) -> dict[str, str]:
"""Unindent all strings in a docdict.
Parameters
----------
docdict : dict[str, str]
A dictionary with string values to unindent.
Returns
-------
docdict : dict[str, str]
The `docdict` dictionary but each of its string values are unindented.
"""
can_dict: dict[str, str] = {}
for name, dstr in docdict.items():
can_dict[name] = unindent_string(dstr)
return can_dict
def unindent_string(docstring: str) -> str:
"""Set docstring to minimum indent for all lines, including first.
Parameters
----------
docstring : str
The input docstring to unindent.
Returns
-------
docstring : str
The unindented docstring.
Examples
--------
>>> unindent_string(' two')
'two'
>>> unindent_string(' two\\n three')
'two\\n three'
"""
lines = docstring.expandtabs().splitlines()
icount = indentcount_lines(lines)
if icount == 0:
return docstring
return "\n".join([line[icount:] for line in lines])
def doc_replace(obj: object, oldval: str, newval: str) -> Decorator:
"""Decorator to take the docstring from obj, with oldval replaced by newval
Equivalent to ``func.__doc__ = obj.__doc__.replace(oldval, newval)``
Parameters
----------
obj : object
A class or object whose docstring will be used as the basis for the
replacement operation.
oldval : str
The string to search for in the docstring.
newval : str
The string to replace `oldval` with in the docstring.
Returns
-------
decfunc : function
A decorator function that replaces occurrences of `oldval` with `newval`
in the docstring of the decorated function.
"""
# __doc__ may be None for optimized Python (-OO)
doc = (obj.__doc__ or "").replace(oldval, newval)
def inner(func: _F) -> _F:
func.__doc__ = doc
return func
return inner
@@ -0,0 +1,212 @@
# Bounds may appear unused in this file but we need to import it to make it available to the user
from scipy.optimize import NonlinearConstraint, LinearConstraint, Bounds
from .common._nonlinear_constraints import process_nl_constraints
from .common._linear_constraints import (
combine_multiple_linear_constraints,
separate_LC_into_eq_and_ineq,
)
from .common._bounds import process_bounds
from enum import Enum
from .common._project import _project
from .common.linalg import get_arrays_tol
from .cobyla.cobyla import cobyla
import numpy as np
from collections.abc import Iterable
class ConstraintType(Enum):
LINEAR_OBJECT = 5
NONLINEAR_OBJECT = 10
LINEAR_DICT = 15
NONLINEAR_DICT = 20
def get_constraint_type(constraint):
if isinstance(constraint, dict) and ("A" in constraint) and ("lb" in constraint) and ("ub" in constraint):
return ConstraintType.LINEAR_DICT
elif isinstance(constraint, dict) and ("fun" in constraint) and ("lb" in constraint) and ("ub" in constraint):
return ConstraintType.NONLINEAR_DICT
elif hasattr(constraint, "A") and hasattr(constraint, "lb") and hasattr(constraint, "ub"):
return ConstraintType.LINEAR_OBJECT
elif hasattr(constraint, "fun") and hasattr(constraint, "lb") and hasattr(constraint, "ub"):
return ConstraintType.NONLINEAR_OBJECT
else:
raise ValueError(f"Constraint type {type(constraint)} not recognized")
def process_constraints(constraints):
# First throw it back if it's an empty tuple
if not constraints:
return None, None
# Next figure out if it's a list of constraints or a single constraint
# If it's a single constraint, make it a list, and then the remaining logic
# doesn't have to change
if not isinstance(constraints, Iterable):
constraints = [constraints]
# Separate out the linear and nonlinear constraints
linear_constraints = []
nonlinear_constraints = []
for constraint in constraints:
constraint_type = get_constraint_type(constraint)
if constraint_type is ConstraintType.LINEAR_OBJECT:
linear_constraints.append(constraint)
elif constraint_type is ConstraintType.NONLINEAR_OBJECT:
nonlinear_constraints.append(constraint)
elif constraint_type == ConstraintType.LINEAR_DICT:
linear_constraints.append(LinearConstraint(constraint["A"], constraint["lb"], constraint["ub"]))
elif constraint_type == ConstraintType.NONLINEAR_DICT:
nonlinear_constraints.append(NonlinearConstraint(constraint["fun"], constraint["lb"], constraint["ub"]))
else:
raise ValueError("Constraint type not recognized")
if len(nonlinear_constraints) > 0:
nonlinear_constraint_function = process_nl_constraints(nonlinear_constraints)
else:
nonlinear_constraint_function = None
# Determine if we have multiple linear constraints, just 1, or none, and process accordingly
if len(linear_constraints) > 1:
linear_constraint = combine_multiple_linear_constraints(linear_constraints)
elif len(linear_constraints) == 1:
linear_constraint = linear_constraints[0]
else:
linear_constraint = None
return linear_constraint, nonlinear_constraint_function
def minimize(fun, x0, args=(), method=None, bounds=None, constraints=(), callback=None, options=None):
linear_constraint, nonlinear_constraint_function = process_constraints(constraints)
options = {'quiet': True} if options is None else options
quiet = options.get("quiet", True)
if method is None:
if nonlinear_constraint_function is not None:
if not quiet: print("Nonlinear constraints detected, applying COBYLA")
method = "cobyla"
elif linear_constraint is not None:
if not quiet: print("Linear constraints detected without nonlinear constraints, applying LINCOA")
method = "lincoa"
elif bounds is not None:
if not quiet: print("Bounds without linear or nonlinear constraints detected, applying BOBYQA")
method = "bobyqa"
else:
if not quiet: print("No bounds or constraints detected, applying NEWUOA")
method = "newuoa"
else:
# Raise some errors if methods were called with inappropriate options
method = method.lower()
if method not in ('newuoa', 'uobyqa', 'bobyqa', 'cobyla', 'lincoa'):
raise ValueError(f"Method must be one of NEWUOA, UOBYQA, BOBYQA, COBYLA, or LINCOA, not '{method}'")
if method != "cobyla" and nonlinear_constraint_function is not None:
raise ValueError("Nonlinear constraints were provided for an algorithm that cannot handle them")
if method not in ("cobyla", "lincoa") and linear_constraint is not None:
raise ValueError("Linear constraints were provided for an algorithm that cannot handle them")
if method not in ("cobyla", "bobyqa", "lincoa") and bounds is not None:
raise ValueError("Bounds were provided for an algorithm that cannot handle them")
# Try to get the length of x0. If we can't that likely means it's a scalar, and
# in that case we turn it into an array and wrap the original function so that it
# can accept an array and return a scalar.
try:
lenx0 = len(x0)
except TypeError:
x0 = np.array([x0])
original_scalar_fun = fun
def scalar_fun(x):
return original_scalar_fun(x[0], *args)
fun = scalar_fun
lenx0 = 1
lb, ub = process_bounds(bounds, lenx0)
# Check which variables are fixed and eliminate them from the problem.
# Save the indices and values so that we can call the original function with
# an array of the appropriate size, and so that we can add the fixed values to the
# result when COBYLA returns.
tol = get_arrays_tol(lb, ub)
_fixed_idx = (
(lb <= ub)
& (np.abs(lb - ub) < tol)
)
if any(_fixed_idx):
_fixed_values = 0.5 * (
lb[_fixed_idx] + ub[_fixed_idx]
)
_fixed_values = np.clip(
_fixed_values,
lb[_fixed_idx],
ub[_fixed_idx],
)
x0 = x0[~_fixed_idx]
lb = lb[~_fixed_idx]
ub = ub[~_fixed_idx]
original_fun = fun
def fixed_fun(x):
newx = np.zeros(lenx0)
newx[_fixed_idx] = _fixed_values
newx[~_fixed_idx] = x
return original_fun(newx, *args)
fun = fixed_fun
# Project x0 onto the feasible set
if nonlinear_constraint_function is None:
result = _project(x0, lb, ub, {"linear": linear_constraint, "nonlinear": None})
x0 = result.x
if linear_constraint is not None:
A_eq, b_eq, A_ineq, b_ineq = separate_LC_into_eq_and_ineq(linear_constraint)
else:
A_eq, b_eq, A_ineq, b_ineq = None, None, None, None
if nonlinear_constraint_function is not None:
# If there is a nonlinear constraint function, we will call COBYLA, which needs the number of nonlinear
# constraints (m_nlcon). In order to get this number we need to evaluate the constraint function at x0.
# The constraint value at x0 (nlconstr0) is not discarded but passed down to the Fortran backend, as its
# evaluation is assumed to be expensive. We also evaluate the objective function at x0 and pass the result
# (f0) down to the Fortran backend, which expects nlconstr0 and f0 to be provided in sync.
def calcfc(x):
f = fun(x, *args)
nlconstr = nonlinear_constraint_function(x)
return f, nlconstr
else:
def calcfc(x):
f = fun(x, *args)
constr = np.zeros(0)
return f, constr
f0, nlconstr0 = calcfc(x0)
if 'quiet' in options:
del options['quiet']
if 'maxfev' in options:
options['maxfun'] = options['maxfev']
del options['maxfev']
result = cobyla(
calcfc,
len(nlconstr0),
x0,
A_ineq,
b_ineq,
A_eq,
b_eq,
lb,
ub,
f0=f0,
nlconstr0=nlconstr0,
callback=callback,
**options
)
if any(_fixed_idx):
newx = np.zeros(lenx0)
newx[_fixed_idx] = _fixed_values
newx[~_fixed_idx] = result.x
result.x = newx
return result
@@ -0,0 +1,559 @@
'''
This module provides Powell's COBYLA algorithm.
Translated from Zaikun Zhang's modern-Fortran reference implementation in PRIMA.
Dedicated to late Professor M. J. D. Powell FRS (1936--2015).
Python translation by Nickolai Belakovski.
N.B.:
1. The modern-Fortran reference implementation in PRIMA contains bug fixes and improvements over the
original Fortran 77 implementation by Powell. Consequently, the PRIMA implementation behaves differently
from the original Fortran 77 implementation by Powell. Therefore, it is important to point out that
you are using PRIMA rather than the original solvers if you want your results to be reproducible.
2. Compared to Powell's Fortran 77 implementation, the modern-Fortran implementation and hence any
faithful translation like this one generally produce better solutions with fewer function evaluations,
making them preferable for applications with expensive function evaluations. However, if function
evaluations are not the dominant cost in your application, the Fortran 77 solvers are likely to be
faster, as they are more efficient in terms of memory usage and flops thanks to the careful and
ingenious (but unmaintained and unmaintainable) implementation by Powell.
See the PRIMA documentation (www.libprima.net) for more information.
'''
from ..common.evaluate import evaluate, moderatex, moderatef, moderatec
from ..common.consts import (EPS, RHOBEG_DEFAULT, RHOEND_DEFAULT, CTOL_DEFAULT,
CWEIGHT_DEFAULT, FTARGET_DEFAULT, IPRINT_DEFAULT,
MAXFUN_DIM_DEFAULT, DEBUGGING, BOUNDMAX,
ETA1_DEFAULT, ETA2_DEFAULT, GAMMA1_DEFAULT,
GAMMA2_DEFAULT)
from ..common.preproc import preproc
from ..common.present import present
from ..common.linalg import matprod
from .cobylb import cobylb
import numpy as np
from dataclasses import dataclass
from copy import copy
@dataclass
class COBYLAResult:
x: np.ndarray
f: float
constr: np.ndarray
cstrv: float
nf: int
xhist: np.ndarray | None
fhist: np.ndarray | None
chist: np.ndarray | None
conhist: np.ndarray | None
info: int
def cobyla(calcfc, m_nlcon, x, Aineq=None, bineq=None, Aeq=None, beq=None,
xl=None, xu=None, f0=None, nlconstr0=None, rhobeg=None, rhoend=None,
ftarget=FTARGET_DEFAULT, ctol=CTOL_DEFAULT, cweight=CWEIGHT_DEFAULT,
maxfun=None, iprint=IPRINT_DEFAULT, eta1=None, eta2=None,
gamma1=GAMMA1_DEFAULT, gamma2=GAMMA2_DEFAULT, maxhist=None, maxfilt=2000,
callback=None):
"""
Among all the arguments, only CALCFC, M_NLCON, and X are obligatory. The others are
OPTIONAL and you can neglect them unless you are familiar with the algorithm. Any
unspecified optional input will take the default value detailed below. For
instance, we may invoke the solver as follows.
# First define CALCFC, M_NLCON, and X, and then do the following.
result = cobyla(calcfc, m_nlcon, x)
or
# First define CALCFC, M_NLCON, X, Aineq, and Bineq, and then do the following.
result = cobyla(calcfc, m_nlcon, x, Aineq=Aineq, bineq=bineq, rhobeg=1.0e0,
rhoend=1.0e-6)
####################################################################################
# IMPORTANT NOTICE: The user must set M_NLCON correctly to the number of nonlinear
# constraints, namely the size of NLCONSTR introduced below. Set it to 0 if there
# is no nonlinear constraint.
####################################################################################
See examples/cobyla/cobyla_example.py for a concrete example.
A detailed introduction to the arguments is as follows.
####################################################################################
# INPUTS
####################################################################################
CALCFC
Input, function.
f, nlconstr = CALCFC(X) should evaluate the objective function and nonlinear
constraints at the given vector X; it should return a tuple consisting of the
objective function value and the nonlinear constraint value. It must be provided
by the user, and its definition must conform to the following interface:
#-------------------------------------------------------------------------#
def calcfc(x):
f = 0.0
nlconstr = np.zeros(m_nlcon)
return f, nlconstr
#-------------------------------------------------------------------------#
M_NLCON
Input, scalar.
M_NLCON must be set to the number of nonlinear constraints, namely the size of
NLCONSTR(X).
N.B.:
1. Why don't we define M_NLCON as optional and default it to 0 when it is absent?
This is because we need to allocate memory for CONSTR_LOC using M_NLCON. To
ensure that the size of CONSTR_LOC is correct, we require the user to specify
M_NLCON explicitly.
X
Input, vector.
As an input, X should be an N-dimensional vector that contains the starting
point, N being the dimension of the problem.
Aineq, Bineq
Input, matrix of size [Mineq, N] and vector of size Mineq unless they are both
empty, default: None and None.
Aineq and Bineq represent the linear inequality constraints: Aineq*X <= Bineq.
Aeq, Beq
Input, matrix of size [Meq, N] and vector of size Meq unless they are both
empty, default: None and None.
Aeq and Beq represent the linear equality constraints: Aeq*X = Beq.
XL, XU
Input, vectors of size N unless they are both None, default: None and None.
XL is the lower bound for X. If XL is None, X has no
lower bound. Any entry of XL that is NaN or below -BOUNDMAX will be taken as
-BOUNDMAX, which effectively means there is no lower bound for the corresponding
entry of X. The value of BOUNDMAX is 0.25*HUGE(X), which is about 8.6E37 for
single precision and 4.5E307 for double precision. XU is similar.
F0
Input, scalar.
F0, if present, should be set to the objective function value of the starting X.
NLCONSTR0
Input, vector.
NLCONSTR0, if present, should be set to the nonlinear constraint value at the
starting X; in addition, SIZE(NLCONSTR0) must be M_NLCON, or the solver will
abort.
RHOBEG, RHOEND
Inputs, scalars, default: RHOBEG = 1, RHOEND = 10^-6. RHOBEG and RHOEND must be
set to the initial and final values of a trust-region radius, both being positive
and RHOEND <= RHOBEG. Typically RHOBEG should be about one tenth of the greatest
expected change to a variable, and RHOEND should indicate the accuracy that is
required in the final values of the variables.
FTARGET
Input, scalar, default: -Inf.
FTARGET is the target function value. The algorithm will terminate when a
feasible point with a function value <= FTARGET is found.
CTOL
Input, scalar, default: sqrt(machine epsilon).
CTOL is the tolerance of constraint violation. X is considered feasible if
CSTRV(X) <= CTOL.
N.B.:
1. CTOL is absolute, not relative.
2. CTOL is used only when selecting the returned X. It does not affect the
iterations of the algorithm.
CWEIGHT
Input, scalar, default: CWEIGHT_DFT defined in common/consts.py.
CWEIGHT is the weight that the constraint violation takes in the selection of the
returned X.
MAXFUN
Input, integer scalar, default: MAXFUN_DIM_DFT*N with MAXFUN_DIM_DFT defined in
common/consts.py. MAXFUN is the maximal number of calls of CALCFC.
IPRINT
Input, integer scalar, default: 0.
The value of IPRINT should be set to 0, 1, -1, 2, -2, 3, or -3, which controls
how much information will be printed during the computation:
0: there will be no printing;
1: a message will be printed to the screen at the return, showing the best vector
of variables found and its objective function value;
2: in addition to 1, each new value of RHO is printed to the screen, with the
best vector of variables so far and its objective function value; each new
value of CPEN is also printed;
3: in addition to 2, each function evaluation with its variables will be printed
to the screen; -1, -2, -3: the same information as 1, 2, 3 will be printed,
not to the screen but to a file named COBYLA_output.txt; the file will be
created if it does not exist; the new output will be appended to the end of
this file if it already exists.
Note that IPRINT = +/-3 can be costly in terms of time and/or space.
ETA1, ETA2, GAMMA1, GAMMA2
Input, scalars, default: ETA1 = 0.1, ETA2 = 0.7, GAMMA1 = 0.5, and GAMMA2 = 2.
ETA1, ETA2, GAMMA1, and GAMMA2 are parameters in the updating scheme of the
trust-region radius detailed in the subroutine TRRAD in trustregion.py. Roughly
speaking, the trust-region radius is contracted by a factor of GAMMA1 when the
reduction ratio is below ETA1, and enlarged by a factor of GAMMA2 when the
reduction ratio is above ETA2. It is required that 0 < ETA1 <= ETA2 < 1 and
0 < GAMMA1 < 1 < GAMMA2. Normally, ETA1 <= 0.25. It is NOT advised to set
ETA1 >= 0.5.
MAXFILT
Input, scalar.
MAXFILT is a nonnegative integer indicating the maximal length of the filter used
for selecting the returned solution; default: MAXFILT_DFT (a value lower than
MIN_MAXFILT is not recommended);
see common/consts.py for the definitions of MAXFILT_DFT and MIN_MAXFILT.
CALLBACK
Input, function to report progress and optionally request termination.
####################################################################################
# OUTPUTS
####################################################################################
The output is a single data structure, COBYLAResult, with the following fields:
X
Output, vector.
As an output, X will be set to an approximate minimizer.
F
Output, scalar.
F will be set to the objective function value of X at exit.
CONSTR
Output, vector.
CONSTR will be set to the constraint value of X at exit.
CSTRV
Output, scalar.
CSTRV will be set to the constraint violation of X at exit, i.e.,
max([0, XL - X, X - XU, Aineq*X - Bineq, ABS(Aeq*X -Beq), NLCONSTR(X)]).
NF
Output, scalar.
NF will be set to the number of calls of CALCFC at exit.
XHIST, FHIST, CHIST, CONHIST, MAXHIST
XHIST: Output, rank 2 array;
FHIST: Output, rank 1 array;
CHIST: Output, rank 1 array;
CONHIST: Output, rank 2 array;
MAXHIST: Input, scalar, default: MAXFUN
XHIST, if present, will output the history of iterates; FHIST, if present, will
output the history function values; CHIST, if present, will output the history of
constraint violations; CONHIST, if present, will output the history of constraint
values; MAXHIST should be a nonnegative integer, and XHIST/FHIST/CHIST/CONHIST
will output only the history of the last MAXHIST iterations.
Therefore, MAXHIST= 0 means XHIST/FHIST/CONHIST/CHIST will output
nothing, while setting MAXHIST = MAXFUN requests XHIST/FHIST/CHIST/CONHIST to
output all the history. If XHIST is present, its size at exit will be
(N, min(NF, MAXHIST)); if FHIST is present, its size at exit will be
min(NF, MAXHIST); if CHIST is present, its size at exit will be min(NF, MAXHIST);
if CONHIST is present, its size at exit will be (M, min(NF, MAXHIST)).
IMPORTANT NOTICE:
Setting MAXHIST to a large value can be costly in terms of memory for large
problems.
MAXHIST will be reset to a smaller value if the memory needed exceeds MAXHISTMEM
defined in common/consts.py
Use *HIST with caution!!! (N.B.: the algorithm is NOT designed for large
problems).
INFO
Output, scalar.
INFO is the exit flag. It will be set to one of the following values defined in
common/infos.py:
SMALL_TR_RADIUS: the lower bound for the trust region radius is reached;
FTARGET_ACHIEVED: the target function value is reached;
MAXFUN_REACHED: the objective function has been evaluated MAXFUN times;
MAXTR_REACHED: the trust region iteration has been performed MAXTR times (MAXTR = 2*MAXFUN);
NAN_INF_X: NaN or Inf occurs in X;
DAMAGING_ROUNDING: rounding errors are becoming damaging.
#--------------------------------------------------------------------------#
The following case(s) should NEVER occur unless there is a bug.
NAN_INF_F: the objective function returns NaN or +Inf;
NAN_INF_MODEL: NaN or Inf occurs in the model;
TRSUBP_FAILED: a trust region step failed to reduce the model
#--------------------------------------------------------------------------#
"""
# Local variables
solver = "COBYLA"
srname = "COBYLA"
# Sizes
mineq = len(bineq) if present(bineq) else 0
meq = len(beq) if present(beq) else 0
mxl = sum(xl > -BOUNDMAX) if present(xl) else 0
mxu = sum(xu < BOUNDMAX) if present(xu) else 0
mmm = mxu + mxl + 2*meq + mineq + m_nlcon
num_vars = len(x)
# Preconditions
if DEBUGGING:
assert m_nlcon >= 0, f'{srname} M_NLCON >= 0'
assert num_vars >= 1, f'{srname} N >= 1'
assert present(Aineq) == present(bineq), \
f'{srname} Aineq and Bineq are both present or both absent'
if (present(Aineq)):
assert Aineq.shape == (mineq, num_vars), f'{srname} SIZE(Aineq) == [Mineq, N]'
assert present(Aeq) == present(beq), \
f'{srname} Aeq and Beq are both present or both absent'
if (present(Aeq)):
assert Aeq.shape == (meq, num_vars), f'{srname} SIZE(Aeq) == [Meq, N]'
if (present(xl)):
assert len(xl) == num_vars, f'{srname} SIZE(XL) == N'
if (present(xu)):
assert len(xu) == num_vars, f'{srname} SIZE(XU) == N'
# N.B.: If NLCONSTR0 is present, then F0 must be present, and we assume that
# F(X0) = F0 even if F0 is NaN; if NLCONSTR0 is absent, then F0 must be either
# absent or NaN, both of which will be interpreted as F(X0) is not provided.
if present(nlconstr0):
assert present(f0), f'{srname} If NLCONSTR0 is present, then F0 is present'
if present(f0):
assert np.isnan(f0) or present(nlconstr0), \
f'{srname} If F0 is present and not NaN, then NLCONSTR0 is present'
# Exit if the size of NLCONSTR0 is inconsistent with M_NLCON.
if present(nlconstr0):
assert np.size(nlconstr0) == m_nlcon
# Read the inputs.
if xl is not None:
xl = copy(xl)
xl[np.isnan(xl)] = -BOUNDMAX
xl[xl < -BOUNDMAX] = -BOUNDMAX
if xu is not None:
xu = copy(xu)
xu[np.isnan(xu)] = BOUNDMAX
xu[xu > BOUNDMAX] = BOUNDMAX
# Wrap the linear and bound constraints into a single constraint: AMAT@X <= BVEC.
amat, bvec = get_lincon(Aeq, Aineq, beq, bineq, xl, xu)
# Create constraint vector
constr = np.zeros(mmm)
# Set [F_LOC, CONSTR_LOC] to [F(X0), CONSTR(X0)] after evaluating the latter if
# needed. In this way, COBYLB only needs one interface.
# N.B.: Due to the preconditions above, there are two possibilities for F0 and
# NLCONSTR0.
# If NLCONSTR0 is present, then F0 must be present, and we assume that F(X0) = F0
# even if F0 is NaN.
# If NLCONSTR0 is absent, then F0 must be either absent or NaN, both of which will
# be interpreted as F(X0) is not provided and we have to evaluate F(X0) and
# NLCONSTR(X0) now.
if (present(f0) and present(nlconstr0) and all(np.isfinite(x))):
f = moderatef(f0)
if amat is not None:
constr[:mmm - m_nlcon] = moderatec(matprod(amat, x) - bvec)
constr[mmm - m_nlcon:] = moderatec(nlconstr0)
else:
x = moderatex(x)
f, constr = evaluate(calcfc, x, m_nlcon, amat, bvec)
constr[:mmm - m_nlcon] = moderatec(constr[:mmm - m_nlcon])
# N.B.: Do NOT call FMSG, SAVEHIST, or SAVEFILT for the function/constraint evaluation at X0.
# They will be called during the initialization, which will read the function/constraint at X0.
cstrv = max(np.append(0, constr))
# If RHOBEG is present, use it; otherwise, RHOBEG takes the default value for
# RHOBEG, taking the value of RHOEND into account. Note that RHOEND is considered
# only if it is present and it is VALID (i.e., finite and positive). The other
# inputs are read similarly.
if present(rhobeg):
rhobeg = rhobeg
elif present(rhoend) and np.isfinite(rhoend) and rhoend > 0:
rhobeg = max(10 * rhoend, RHOBEG_DEFAULT)
else:
rhobeg = RHOBEG_DEFAULT
if present(rhoend):
rhoend = rhoend
elif rhobeg > 0:
rhoend = max(EPS, min(RHOEND_DEFAULT/RHOBEG_DEFAULT * rhobeg, RHOEND_DEFAULT))
else:
rhoend = RHOEND_DEFAULT
maxfun = maxfun if present(maxfun) else MAXFUN_DIM_DEFAULT * num_vars
if present(eta1):
eta1 = eta1
elif present(eta2) and 0 < eta2 < 1:
eta1 = max(EPS, eta2 / 7)
else:
eta1 = ETA1_DEFAULT
if present(eta2):
eta2 = eta2
elif 0 < eta1 < 1:
eta2 = (eta1 + 2) / 3
else:
eta2 = ETA2_DEFAULT
maxhist = (
maxhist
if present(maxhist)
else max(maxfun, num_vars + 2, MAXFUN_DIM_DEFAULT * num_vars)
)
# Preprocess the inputs in case some of them are invalid. It does nothing if all
# inputs are valid.
(
iprint,
maxfun,
maxhist,
ftarget,
rhobeg,
rhoend,
npt, # Unused in COBYLA
maxfilt,
ctol,
cweight,
eta1,
eta2,
gamma1,
gamma2,
_x0, # Unused in COBYLA
) = preproc(
solver,
num_vars,
iprint,
maxfun,
maxhist,
ftarget,
rhobeg,
rhoend,
num_constraints=mmm,
maxfilt=maxfilt,
ctol=ctol,
cweight=cweight,
eta1=eta1,
eta2=eta2,
gamma1=gamma1,
gamma2=gamma2,
is_constrained=(mmm > 0),
)
# Further revise MAXHIST according to MAXHISTMEM, and allocate memory for the history.
# In MATLAB/Python/Julia/R implementation, we should simply set MAXHIST = MAXFUN and initialize
# CHIST = NaN(1, MAXFUN), CONHIST = NaN(M, MAXFUN), FHIST = NaN(1, MAXFUN), XHIST = NaN(N, MAXFUN)
# if they are requested; replace MAXFUN with 0 for the history that is not requested.
# prehist(maxhist, num_vars, present(xhist), xhist_loc, present(fhist), fhist_loc, &
# & present(chist), chist_loc, m, present(conhist), conhist_loc)
# call cobylb, which performs the real calculations
x, f, constr, cstrv, nf, xhist, fhist, chist, conhist, info = cobylb(
calcfc,
iprint,
maxfilt,
maxfun,
amat,
bvec,
ctol,
cweight,
eta1,
eta2,
ftarget,
gamma1,
gamma2,
rhobeg,
rhoend,
constr,
f,
x,
maxhist,
callback
)
return COBYLAResult(x, f, constr, cstrv, nf, xhist, fhist, chist, conhist, info)
def get_lincon(Aeq=None, Aineq=None, beq=None, bineq=None, xl=None, xu=None):
"""
This subroutine wraps the linear and bound constraints into a single constraint:
AMAT*X <= BVEC.
N.B.:
LINCOA normalizes the linear constraints so that each constraint has a gradient
of norm 1. However, COBYLA does not do this.
"""
# Sizes
if Aeq is not None:
num_vars = Aeq.shape[1]
elif Aineq is not None:
num_vars = Aineq.shape[1]
elif xl is not None:
num_vars = len(xl)
elif xu is not None:
num_vars = len(xu)
else:
return None, None
# Preconditions
if DEBUGGING:
assert Aineq is None or Aineq.shape == (len(bineq), num_vars)
assert Aeq is None or Aeq.shape == (len(beq), num_vars)
assert (xl is None or xu is None) or len(xl) == len(xu) == num_vars
#====================#
# Calculation starts #
#====================#
# Define the indices of the nontrivial bound constraints.
ixl = np.where(xl > -BOUNDMAX)[0] if xl is not None else None
ixu = np.where(xu < BOUNDMAX)[0] if xu is not None else None
# Wrap the linear constraints.
# The bound constraint XL <= X <= XU is handled as two constraints:
# -X <= -XL, X <= XU.
# The equality constraint Aeq*X = Beq is handled as two constraints:
# -Aeq*X <= -Beq, Aeq*X <= Beq.
# N.B.:
# 1. The treatment of the equality constraints is naive. One may choose to
# eliminate them instead.
idmat = np.eye(num_vars)
amat = np.vstack([
-idmat[ixl, :] if ixl is not None else np.empty((0, num_vars)),
idmat[ixu, :] if ixu is not None else np.empty((0, num_vars)),
-Aeq if Aeq is not None else np.empty((0, num_vars)),
Aeq if Aeq is not None else np.empty((0, num_vars)),
Aineq if Aineq is not None else np.empty((0, num_vars))
])
bvec = np.hstack([
-xl[ixl] if ixl is not None else np.empty(0),
xu[ixu] if ixu is not None else np.empty(0),
-beq if beq is not None else np.empty(0),
beq if beq is not None else np.empty(0),
bineq if bineq is not None else np.empty(0)
])
amat = amat if amat.shape[0] > 0 else None
bvec = bvec if bvec.shape[0] > 0 else None
#==================#
# Calculation ends #
#==================#
# Postconditions
if DEBUGGING:
assert (amat is None and bvec is None) or amat.shape == (len(bvec), num_vars)
return amat, bvec
@@ -0,0 +1,714 @@
'''
This module performs the major calculations of COBYLA.
Translated from Zaikun Zhang's modern-Fortran reference implementation in PRIMA.
Dedicated to late Professor M. J. D. Powell FRS (1936--2015).
Python translation by Nickolai Belakovski.
'''
import numpy as np
from ..common.checkbreak import checkbreak_con
from ..common.consts import REALMAX, EPS, DEBUGGING, MIN_MAXFILT
from ..common.infos import INFO_DEFAULT, MAXTR_REACHED, DAMAGING_ROUNDING, \
SMALL_TR_RADIUS, CALLBACK_TERMINATE
from ..common.evaluate import evaluate
from ..common.history import savehist
from ..common.linalg import isinv, matprod, inprod, norm, primasum, primapow2
from ..common.message import fmsg, retmsg, rhomsg
from ..common.ratio import redrat
from ..common.redrho import redrho
from ..common.selectx import savefilt, selectx
from .update import updatepole, findpole, updatexfc
from .geometry import setdrop_tr, geostep
from .trustregion import trstlp, trrad
from .initialize import initxfc, initfilt
def cobylb(calcfc, iprint, maxfilt, maxfun, amat, bvec, ctol, cweight, eta1, eta2,
ftarget, gamma1, gamma2, rhobeg, rhoend, constr, f, x, maxhist, callback):
'''
This subroutine performs the actual computations of COBYLA.
'''
# Outputs
xhist = []
fhist = []
chist = []
conhist = []
# Local variables
solver = 'COBYLA'
A = np.zeros((np.size(x), np.size(constr))) # A contains the approximate gradient for the constraints
distsq = np.zeros(np.size(x) + 1)
# CPENMIN is the minimum of the penalty parameter CPEN for the L-infinity
# constraint violation in the merit function. Note that CPENMIN = 0 in Powell's
# implementation, which allows CPEN to be 0. Here, we take CPENMIN > 0 so that CPEN
# is always positive. This avoids the situation where PREREM becomes 0 when
# PREREF = 0 = CPEN. It brings two advantages as follows.
# 1. If the trust-region subproblem solver works correctly and the trust-region
# center is not optimal for the subproblem, then PREREM > 0 is guaranteed. This
# is because, in theory, PREREC >= 0 and MAX(PREREC, PREREF) > 0, and the
# definition of CPEN in GETCPEN ensures that PREREM > 0.
# 2. There is no need to revise ACTREM and PREREM when CPEN = 0 and F = FVAL(N+1)
# as in lines 312--314 of Powell's cobylb.f code. Powell's code revises ACTREM
# to CVAL(N + 1) - CSTRV and PREREM to PREREC in this case, which is crucial for
# feasibility problems.
cpenmin = EPS
# Sizes
m_lcon = np.size(bvec) if bvec is not None else 0
num_constraints = np.size(constr)
m_nlcon = num_constraints - m_lcon
num_vars = np.size(x)
# Preconditions
if DEBUGGING:
assert abs(iprint) <= 3
assert num_constraints >= m_lcon and m_lcon >= 0
assert num_vars >= 1
assert maxfun >= num_vars + 2
assert rhobeg >= rhoend and rhoend > 0
assert all(np.isfinite(x))
assert 0 <= eta1 <= eta2 < 1
assert 0 < gamma1 < 1 < gamma2
assert 0 <= ctol
assert 0 <= cweight
assert 0 <= maxhist <= maxfun
assert amat is None or np.shape(amat) == (m_lcon, num_vars)
assert min(MIN_MAXFILT, maxfun) <= maxfilt <= maxfun
#====================#
# Calculation starts #
#====================#
# Initialize SIM, FVAL, CONMAT, and CVAL, together with the history.
# After the initialization, SIM[:, NUM_VARS] holds the vertex of the initial
# simplex with the smallest function value (regardless of the constraint
# violation), and SIM[:, :NUM_VARS] holds the displacements from the other vertices
# to SIM[:, NUM_VARS]. FVAL, CONMAT, and CVAL hold the function values, constraint
# values, and constraint violations on the vertices in the order corresponding to
# SIM.
evaluated, conmat, cval, sim, simi, fval, nf, subinfo = initxfc(calcfc, iprint,
maxfun, constr, amat, bvec, ctol, f, ftarget, rhobeg, x,
xhist, fhist, chist, conhist, maxhist)
# Initialize the filter, including xfilt, ffilt, confilt, cfilt, and nfilt.
# N.B.: The filter is used only when selecting which iterate to return. It does not
# interfere with the iterations. COBYLA is NOT a filter method but a trust-region
# method based on an L-infinity merit function. Powell's implementation does not
# use a filter to select the iterate, possibly returning a suboptimal iterate.
cfilt = np.zeros(np.minimum(np.maximum(maxfilt, 1), maxfun))
confilt = np.zeros((np.size(constr), np.size(cfilt)))
ffilt = np.zeros(np.size(cfilt))
xfilt = np.zeros((np.size(x), np.size(cfilt)))
nfilt = initfilt(conmat, ctol, cweight, cval, fval, sim, evaluated, cfilt, confilt,
ffilt, xfilt)
# Check whether to return due to abnormal cases that may occur during the initialization.
if subinfo != INFO_DEFAULT:
info = subinfo
# Return the best calculated values of the variables
# N.B: Selectx and findpole choose X by different standards, one cannot replace the other
kopt = selectx(ffilt[:nfilt], cfilt[:nfilt], cweight, ctol)
x = xfilt[:, kopt]
f = ffilt[kopt]
constr = confilt[:, kopt]
cstrv = cfilt[kopt]
# print a return message according to IPRINT.
retmsg(solver, info, iprint, nf, f, x, cstrv, constr)
# Postconditions
if DEBUGGING:
assert nf <= maxfun
assert np.size(x) == num_vars and not any(np.isnan(x))
assert not (np.isnan(f) or np.isposinf(f))
# assert np.size(xhist, 0) == n and np.size(xhist, 1) == maxxhist
# assert not any(np.isnan(xhist(:, 1:min(nf, maxxhist))))
# The last calculated X can be Inf (finite + finite can be Inf numerically).
# assert np.size(fhist) == maxfhist
# assert not any(np.isnan(fhist(1:min(nf, maxfhist))) or np.isposinf(fhist(1:min(nf, maxfhist))))
# assert np.size(conhist, 0) == m and np.size(conhist, 1) == maxconhist
# assert not any(np.isnan(conhist(:, 1:min(nf, maxconhist))) or np.isneginf(conhist(:, 1:min(nf, maxconhist))))
# assert np.size(chist) == maxchist
# assert not any(chist(1:min(nf, maxchist)) < 0 or np.isnan(chist(1:min(nf, maxchist))) or np.isposinf(chist(1:min(nf, maxchist))))
# nhist = minval([nf, maxfhist, maxchist])
# assert not any(isbetter(fhist(1:nhist), chist(1:nhist), f, cstrv, ctol))
return x, f, constr, cstrv, nf, xhist, fhist, chist, conhist, info
# Set some more initial values.
# We must initialize shortd, ratio, and jdrop_tr because these get defined on
# branches that are not guaranteed to be executed, but their values are used later.
# Our initialization of CPEN differs from Powell's in two ways. First, we use the
# ratio defined in (13) of Powell's COBYLA paper to initialize CPEN. Second, we
# impose CPEN >= CPENMIN > 0. Powell's code simply initializes CPEN to 0.
rho = rhobeg
delta = rhobeg
cpen = np.maximum(cpenmin, np.minimum(1.0E3, fcratio(conmat, fval))) # Powell's code: CPEN = ZERO
shortd = False
ratio = -1
jdrop_tr = 0
# If DELTA <= GAMMA3*RHO after an update, we set DELTA to RHO. GAMMA3 must be less
# than GAMMA2. The reason is as follows. Imagine a very successful step with
# DNORM = the un-updated DELTA = RHO. The TRRAD will update DELTA to GAMMA2*RHO.
# If GAMMA3 >= GAMMA2, then DELTA will be reset to RHO, which is not reasonable as
# D is very successful. See paragraph two of Sec 5.2.5 in T. M. Ragonneau's thesis:
# "Model-Based Derivative-Free Optimization Methods and Software." According to
# test on 20230613, for COBYLA, this Powellful updating scheme of DELTA works
# slightly better than setting directly DELTA = max(NEW_DELTA, RHO).
gamma3 = np.maximum(1, np.minimum(0.75 * gamma2, 1.5))
# MAXTR is the maximal number of trust region iterations. Each trust-region
# iteration takes 1 or 2 function evaluations unless the trust-region step is short
# or the trust-region subproblem solver fails but the geometry step is not invoked.
# Thus the following MAXTR is unlikely to be reached.
maxtr = 10 * maxfun
info = MAXTR_REACHED
# Begin the iterative procedure
# After solving a trust-region subproblem, we use three boolean variables to
# control the workflow.
# SHORTD - Is the trust-region trial step too short to invoke # a function
# evaluation?
# IMPROVE_GEO - Will we improve the model after the trust-region iteration? If yes,
# a geometry step will be taken, corresponding to the "Branch (Delta)"
# in the COBYLA paper.
# REDUCE_RHO - Will we reduce rho after the trust-region iteration?
# COBYLA never sets IMPROVE_GEO and REDUCE_RHO to True simultaneously.
for tr in range(maxtr):
# Increase the penalty parameter CPEN, if needed, so that
# PREREM = PREREF + CPEN * PREREC > 0.
# This is the first (out of two) update of CPEN, where CPEN increases or
# remains the same.
# N.B.: CPEN and the merit function PHI = FVAL + CPEN*CVAL are used in three
# places only.
# 1. In FINDPOLE/UPDATEPOLE, deciding the optimal vertex of the current simplex.
# 2. After the trust-region trial step, calculating the reduction ratio.
# 3. In GEOSTEP, deciding the direction of the geometry step.
# They do not appear explicitly in the trust-region subproblem, though the
# trust-region center (i.e. the current optimal vertex) is defined by them.
cpen = getcpen(amat, bvec, conmat, cpen, cval, delta, fval, rho, sim, simi)
# Switch the best vertex of the current simplex to SIM[:, NUM_VARS].
conmat, cval, fval, sim, simi, subinfo = updatepole(cpen, conmat, cval, fval,
sim, simi)
# Check whether to exit due to damaging rounding in UPDATEPOLE.
if subinfo == DAMAGING_ROUNDING:
info = subinfo
break # Better action to take? Geometry step, or simply continue?
# Does the interpolation set have adequate geometry? It affects improve_geo and
# reduce_rho.
adequate_geo = all(primasum(primapow2(sim[:, :num_vars]), axis=0) <= 4 * primapow2(delta))
# Calculate the linear approximations to the objective and constraint functions.
# N.B.: TRSTLP accesses A mostly by columns, so it is more reasonable to save A
# instead of A^T.
# Zaikun 2023108: According to a test on 2023108, calculating G and
# A(:, M_LCON+1:M) by solving the linear systems SIM^T*G = FVAL(1:N)-FVAL(N+1)
# and SIM^T*A = CONMAT(:, 1:N)-CONMAT(:, N+1) does not seem to improve or worsen
# the performance of COBYLA in terms of the number of function evaluations. The
# system was solved by SOLVE in LINALG_MOD based on a QR factorization of SIM
# (not necessarily a good algorithm). No preconditioning or scaling was used.
g = matprod((fval[:num_vars] - fval[num_vars]), simi)
A[:, :m_lcon] = amat.T if amat is not None else amat
A[:, m_lcon:] = matprod((conmat[m_lcon:, :num_vars] -
np.tile(conmat[m_lcon:, num_vars], (num_vars, 1)).T), simi).T
# Calculate the trust-region trial step d. Note that d does NOT depend on cpen.
d = trstlp(A, -conmat[:, num_vars], delta, g)
dnorm = min(delta, norm(d))
# Is the trust-region trial step short? N.B.: we compare DNORM with RHO, not
# DELTA. Powell's code especially defines SHORTD by SHORTD = (DNORM < 0.5 *
# RHO). In our tests 1/10 seems to work better than 1/2 or 1/4, especially for
# linearly constrained problems. Note that LINCOA has a slightly more
# sophisticated way of defining SHORTD, taking into account whether D causes a
# change to the active set. Should we try the same here?
shortd = (dnorm <= 0.1 * rho)
# Predict the change to F (PREREF) and to the constraint violation (PREREC) due
# to D. We have the following in precise arithmetic. They may fail to hold due
# to rounding errors.
# 1. B[:NUM_CONSTRAINTS] = -CONMAT[:, NUM_VARS] and hence
# np.max(np.append(B[:NUM_CONSTRAINTS] - D@A[:, :NUM_CONSTRAINTS], 0)) is the
# L-infinity violation of the linearized constraints corresponding to D. When
# D=0, the violation is np.max(np.append(B[:NUM_CONSTRAINTS], 0)) =
# CVAL[NUM_VARS]. PREREC is the reduction of this violation achieved by D,
# which is nonnegative in theory; PREREC = 0 iff B[:NUM_CONSTRAINTS] <= 0, i.e.
# the trust-region center satisfies the linearized constraints.
# 2. PREREF may be negative or 0, but it is positive when PREREC = 0 and shortd
# is False
# 3. Due to 2, in theory, max(PREREC, PREREF) > 0 if shortd is False.
preref = -inprod(d, g) # Can be negative
prerec = cval[num_vars] - np.max(np.append(0, conmat[:, num_vars] + matprod(d, A)))
# Evaluate PREREM, which is the predicted reduction in the merit function.
# In theory, PREREM >= 0 and it is 0 iff CPEN = 0 = PREREF. This may not be true
# numerically.
prerem = preref + cpen * prerec
trfail = not (prerem > 1.0E-6 * min(cpen, 1) * rho)
if shortd or trfail:
# Reduce DELTA if D is short or if D fails to render PREREM > 0. The latter
# can only happen due to rounding errors. This seems quite important for
# performance
delta *= 0.1
if delta <= gamma3 * rho:
delta = rho # set delta to rho when it is close to or below
else:
# Calculate the next value of the objective and constraint functions.
# If X is close to one of the points in the interpolation set, then we do
# not evaluate the objective and constraints at X, assuming them to have
# the values at the closest point.
# N.B.: If this happens, do NOT include X into the filter, as F and CONSTR
# are inaccurate.
x = sim[:, num_vars] + d
distsq[num_vars] = primasum(primapow2(x - sim[:, num_vars]))
distsq[:num_vars] = primasum(primapow2(x.reshape(num_vars, 1) -
(sim[:, num_vars].reshape(num_vars, 1) + sim[:, :num_vars])), axis=0)
j = np.argmin(distsq)
if distsq[j] <= primapow2(1e-4 * rhoend):
f = fval[j]
constr = conmat[:, j]
cstrv = cval[j]
else:
# Evaluate the objective and constraints at X, taking care of possible
# inf/nan values.
f, constr = evaluate(calcfc, x, m_nlcon, amat, bvec)
cstrv = np.max(np.append(0, constr))
nf += 1
# Save X, F, CONSTR, CSTRV into the history.
savehist(maxhist, x, xhist, f, fhist, cstrv, chist, constr, conhist)
# Save X, F, CONSTR, CSTRV into the filter.
nfilt, cfilt, ffilt, xfilt, confilt = savefilt(cstrv, ctol, cweight, f,
x, nfilt, cfilt, ffilt,
xfilt, constr, confilt)
# Print a message about the function/constraint evaluation according to
# iprint
fmsg(solver, 'Trust region', iprint, nf, delta, f, x, cstrv, constr)
# Evaluate ACTREM, which is the actual reduction in the merit function
actrem = (fval[num_vars] + cpen * cval[num_vars]) - (f + cpen * cstrv)
# Calculate the reduction ratio by redrat, which hands inf/nan carefully
ratio = redrat(actrem, prerem, eta1)
# Update DELTA. After this, DELTA < DNORM may hold.
# N.B.:
# 1. Powell's code uses RHO as the trust-region radius and updates it as
# follows.
# Reduce RHO to GAMMA1*RHO if ADEQUATE_GEO is TRUE and either SHORTD is
# TRUE or RATIO < ETA1, and then revise RHO to RHOEND if its new value is
# not more than GAMMA3*RHOEND; RHO remains unchanged in all other cases;
# in particular, RHO is never increased.
# 2. Our implementation uses DELTA as the trust-region radius, while using
# RHO as a lower bound for DELTA. DELTA is updated in a way that is
# typical for trust-region methods, and it is revised to RHO if its new
# value is not more than GAMMA3*RHO. RHO reflects the current resolution
# of the algorithm; its update is essentially the same as the update of
# RHO in Powell's code (see the definition of REDUCE_RHO below). Our
# implementation aligns with UOBYQA/NEWUOA/BOBYQA/LINCOA and improves the
# performance of COBYLA.
# 3. The same as Powell's code, we do not reduce RHO unless ADEQUATE_GEO is
# TRUE. This is also how Powell updated RHO in
# UOBYQA/NEWUOA/BOBYQA/LINCOA. What about we also use ADEQUATE_GEO ==
# TRUE as a prerequisite for reducing DELTA? The argument would be that
# the bad (small) value of RATIO may be because of a bad geometry (and
# hence a bad model) rather than an improperly large DELTA, and it might
# be good to try improving the geometry first without reducing DELTA.
# However, according to a test on 20230206, it does not improve the
# performance if we skip the update of DELTA when ADEQUATE_GEO is FALSE
# and RATIO < 0.1. Therefore, we choose to update DELTA without checking
# ADEQUATE_GEO.
delta = trrad(delta, dnorm, eta1, eta2, gamma1, gamma2, ratio)
if delta <= gamma3*rho:
delta = rho # Set delta to rho when it is close to or below.
# Is the newly generated X better than the current best point?
ximproved = actrem > 0 # If ACTREM is NaN, then XIMPROVED should and will be False
# Set JDROP_TR to the index of the vertex to be replaced with X. JDROP_TR = 0 means there
# is no good point to replace, and X will not be included into the simplex; in this case,
# the geometry of the simplex likely needs improvement, which will be handled below.
jdrop_tr = setdrop_tr(ximproved, d, delta, rho, sim, simi)
# Update SIM, SIMI, FVAL, CONMAT, and CVAL so that SIM[:, JDROP_TR] is replaced with D.
# UPDATEXFC does nothing if JDROP_TR is None, as the algorithm decides to discard X.
sim, simi, fval, conmat, cval, subinfo = updatexfc(jdrop_tr, constr, cpen, cstrv, d, f, conmat, cval, fval, sim, simi)
# Check whether to break due to damaging rounding in UPDATEXFC
if subinfo == DAMAGING_ROUNDING:
info = subinfo
break # Better action to take? Geometry step, or a RESCUE as in BOBYQA?
# Check whether to break due to maxfun, ftarget, etc.
subinfo = checkbreak_con(maxfun, nf, cstrv, ctol, f, ftarget, x)
if subinfo != INFO_DEFAULT:
info = subinfo
break
# End of if SHORTD or TRFAIL. The normal trust-region calculation ends.
# Before the next trust-region iteration, we possibly improve the geometry of the simplex or
# reduce RHO according to IMPROVE_GEO and REDUCE_RHO. Now we decide these indicators.
# N.B.: We must ensure that the algorithm does not set IMPROVE_GEO = True at infinitely many
# consecutive iterations without moving SIM[:, NUM_VARS] or reducing RHO. Otherwise, the algorithm
# will get stuck in repetitive invocations of GEOSTEP. This is ensured by the following facts:
# 1. If an iteration sets IMPROVE_GEO to True, it must also reduce DELTA or set DELTA to RHO.
# 2. If SIM[:, NUM_VARS] and RHO remain unchanged, then ADEQUATE_GEO will become True after at
# most NUM_VARS invocations of GEOSTEP.
# BAD_TRSTEP: Is the last trust-region step bad?
bad_trstep = shortd or trfail or ratio <= 0 or jdrop_tr is None
# IMPROVE_GEO: Should we take a geometry step to improve the geometry of the interpolation set?
improve_geo = bad_trstep and not adequate_geo
# REDUCE_RHO: Should we enhance the resolution by reducing rho?
reduce_rho = bad_trstep and adequate_geo and max(delta, dnorm) <= rho
# COBYLA never sets IMPROVE_GEO and REDUCE_RHO to True simultaneously.
# assert not (IMPROVE_GEO and REDUCE_RHO), 'IMPROVE_GEO or REDUCE_RHO are not both TRUE, COBYLA'
# If SHORTD or TRFAIL is True, then either IMPROVE_GEO or REDUCE_RHO is True unless ADEQUATE_GEO
# is True and max(DELTA, DNORM) > RHO.
# assert not (shortd or trfail) or (improve_geo or reduce_rho or (adequate_geo and max(delta, dnorm) > rho)), \
# 'If SHORTD or TRFAIL is TRUE, then either IMPROVE_GEO or REDUCE_RHO is TRUE unless ADEQUATE_GEO is TRUE and MAX(DELTA, DNORM) > RHO'
# Comments on BAD_TRSTEP:
# 1. Powell's definition of BAD_TRSTEP is as follows. The one used above seems to work better,
# especially for linearly constrained problems due to the factor TENTH (= ETA1).
# !bad_trstep = (shortd .or. actrem <= 0 .or. actrem < TENTH * prerem .or. jdrop_tr == 0)
# Besides, Powell did not check PREREM > 0 in BAD_TRSTEP, which is reasonable to do but has
# little impact upon the performance.
# 2. NEWUOA/BOBYQA/LINCOA would define BAD_TRSTEP, IMPROVE_GEO, and REDUCE_RHO as follows. Two
# different thresholds are used in BAD_TRSTEP. It outperforms Powell's version.
# !bad_trstep = (shortd .or. trfail .or. ratio <= eta1 .or. jdrop_tr == 0)
# !improve_geo = bad_trstep .and. .not. adequate_geo
# !bad_trstep = (shortd .or. trfail .or. ratio <= 0 .or. jdrop_tr == 0)
# !reduce_rho = bad_trstep .and. adequate_geo .and. max(delta, dnorm) <= rho
# 3. Theoretically, JDROP_TR > 0 when ACTREM > 0 (guaranteed by RATIO > 0). However, in Powell's
# implementation, JDROP_TR may be 0 even RATIO > 0 due to NaN. The modernized code has rectified
# this in the function SETDROP_TR. After this rectification, we can indeed simplify the
# definition of BAD_TRSTEP by removing the condition JDROP_TR == 0. We retain it for robustness.
# Comments on REDUCE_RHO:
# When SHORTD is TRUE, UOBYQA/NEWUOA/BOBYQA/LINCOA all set REDUCE_RHO to TRUE if the recent
# models are sufficiently accurate according to certain criteria. See the paragraph around (37)
# in the UOBYQA paper and the discussions about Box 14 in the NEWUOA paper. This strategy is
# crucial for the performance of the solvers. However, as of 20221111, we have not managed to
# make it work in COBYLA. As in NEWUOA, we recorded the errors of the recent models, and set
# REDUCE_RHO to true if they are small (e.g., ALL(ABS(MODERR_REC) <= 0.1 * MAXVAL(ABS(A))*RHO) or
# ALL(ABS(MODERR_REC) <= RHO**2)) when SHORTD is TRUE. It made little impact on the performance.
# Since COBYLA never sets IMPROVE_GEO and REDUCE_RHO to TRUE simultaneously, the following
# two blocks are exchangeable: IF (IMPROVE_GEO) ... END IF and IF (REDUCE_RHO) ... END IF.
# Improve the geometry of the simplex by removing a point and adding a new one.
# If the current interpolation set has acceptable geometry, then we skip the geometry step.
# The code has a small difference from Powell's original code here: If the current geometry
# is acceptable, then we will continue with a new trust-region iteration; however, at the
# beginning of the iteration, CPEN may be updated, which may alter the pole point SIM(:, N+1)
# by UPDATEPOLE; the quality of the interpolation point depends on SIM(:, N + 1), meaning
# that the same interpolation set may have good or bad geometry with respect to different
# "poles"; if the geometry turns out bad with the new pole, the original COBYLA code will
# take a geometry step, but our code here will NOT do it but continue to take a trust-region
# step. The argument is this: even if the geometry step is not skipped in the first place, the
# geometry may turn out bad again after the pole is altered due to an update to CPEN; should
# we take another geometry step in that case? If no, why should we do it here? Indeed, this
# distinction makes no practical difference for CUTEst problems with at most 100 variables
# and 5000 constraints, while the algorithm framework is simplified.
if improve_geo and not all(primasum(primapow2(sim[:, :num_vars]), axis=0) <= 4 * primapow2(delta)):
# Before the geometry step, updatepole has been called either implicitly by UPDATEXFC or
# explicitly after CPEN is updated, so that SIM[:, :NUM_VARS] is the optimal vertex.
# Decide a vertex to drop from the simplex. It will be replaced with SIM[:, NUM_VARS] + D to
# improve the geometry of the simplex.
# N.B.:
# 1. COBYLA never sets JDROP_GEO = num_vars.
# 2. The following JDROP_GEO comes from UOBYQA/NEWUOA/BOBYQA/LINCOA.
# 3. In Powell's original algorithm, the geometry of the simplex is considered acceptable
# iff the distance between any vertex and the pole is at most 2.1*DELTA, and the distance
# between any vertex and the opposite face of the simplex is at least 0.25*DELTA, as
# specified in (14) of the COBYLA paper. Correspondingly, JDROP_GEO is set to the index of
# the vertex with the largest distance to the pole provided that the distance is larger than
# 2.1*DELTA, or the vertex with the smallest distance to the opposite face of the simplex,
# in which case the distance must be less than 0.25*DELTA, as the current simplex does not
# have acceptable geometry (see (15)--(16) of the COBYLA paper). Once JDROP_GEO is set, the
# algorithm replaces SIM(:, JDROP_GEO) with D specified in (17) of the COBYLA paper, which
# is orthogonal to the face opposite to SIM(:, JDROP_GEO) and has a length of 0.5*DELTA,
# intending to improve the geometry of the simplex as per (14).
# 4. Powell's geometry-improving procedure outlined above has an intrinsic flaw: it may lead
# to infinite cycling, as was observed in a test on 20240320. In this test, the geometry-
# improving point introduced in the previous iteration was replaced with the trust-region
# trial point in the current iteration, which was then replaced with the same geometry-
# improving point in the next iteration, and so on. In this process, the simplex alternated
# between two configurations, neither of which had acceptable geometry. Thus RHO was never
# reduced, leading to infinite cycling. (N.B.: Our implementation uses DELTA as the trust
# region radius, with RHO being its lower bound. When the infinite cycling occurred in this
# test, DELTA = RHO and it could not be reduced due to the requirement that DELTA >= RHO.)
jdrop_geo = np.argmax(primasum(primapow2(sim[:, :num_vars]), axis=0), axis=0)
# Calculate the geometry step D.
delbar = delta/2
d = geostep(jdrop_geo, amat, bvec, conmat, cpen, cval, delbar, fval, simi)
# Calculate the next value of the objective and constraint functions.
# If X is close to one of the points in the interpolation set, then we do not evaluate the
# objective and constraints at X, assuming them to have the values at the closest point.
# N.B.:
# 1. If this happens, do NOT include X into the filter, as F and CONSTR are inaccurate.
# 2. In precise arithmetic, the geometry improving step ensures that the distance between X
# and any interpolation point is at least DELBAR, yet X may be close to them due to
# rounding. In an experiment with single precision on 20240317, X = SIM(:, N+1) occurred.
x = sim[:, num_vars] + d
distsq[num_vars] = primasum(primapow2(x - sim[:, num_vars]))
distsq[:num_vars] = primasum(primapow2(x.reshape(num_vars, 1) -
(sim[:, num_vars].reshape(num_vars, 1) + sim[:, :num_vars])), axis=0)
j = np.argmin(distsq)
if distsq[j] <= primapow2(1e-4 * rhoend):
f = fval[j]
constr = conmat[:, j]
cstrv = cval[j]
else:
# Evaluate the objective and constraints at X, taking care of possible
# inf/nan values.
f, constr = evaluate(calcfc, x, m_nlcon, amat, bvec)
cstrv = np.max(np.append(0, constr))
nf += 1
# Save X, F, CONSTR, CSTRV into the history.
savehist(maxhist, x, xhist, f, fhist, cstrv, chist, constr, conhist)
# Save X, F, CONSTR, CSTRV into the filter.
nfilt, cfilt, ffilt, xfilt, confilt = savefilt(cstrv, ctol, cweight, f,
x, nfilt, cfilt, ffilt,
xfilt, constr, confilt)
# Print a message about the function/constraint evaluation according to iprint
fmsg(solver, 'Geometry', iprint, nf, delta, f, x, cstrv, constr)
# Update SIM, SIMI, FVAL, CONMAT, and CVAL so that SIM(:, JDROP_GEO) is replaced with D.
sim, simi, fval, conmat, cval, subinfo = updatexfc(jdrop_geo, constr, cpen, cstrv, d, f, conmat, cval, fval, sim, simi)
# Check whether to break due to damaging rounding in UPDATEXFC
if subinfo == DAMAGING_ROUNDING:
info = subinfo
break # Better action to take? Geometry step, or simply continue?
# Check whether to break due to maxfun, ftarget, etc.
subinfo = checkbreak_con(maxfun, nf, cstrv, ctol, f, ftarget, x)
if subinfo != INFO_DEFAULT:
info = subinfo
break
# end of if improve_geo. The procedure of improving the geometry ends.
# The calculations with the current RHO are complete. Enhance the resolution of the algorithm
# by reducing RHO; update DELTA and CPEN at the same time.
if reduce_rho:
if rho <= rhoend:
info = SMALL_TR_RADIUS
break
delta = max(0.5 * rho, redrho(rho, rhoend))
rho = redrho(rho, rhoend)
# THe second (out of two) updates of CPEN, where CPEN decreases or remains the same.
# Powell's code: cpen = min(cpen, fcratio(fval, conmat)), which may set CPEN to 0.
cpen = np.maximum(cpenmin, np.minimum(cpen, fcratio(conmat, fval)))
# Print a message about the reduction of rho according to iprint
rhomsg(solver, iprint, nf, fval[num_vars], rho, sim[:, num_vars], cval[num_vars], conmat[:, num_vars], cpen)
conmat, cval, fval, sim, simi, subinfo = updatepole(cpen, conmat, cval, fval, sim, simi)
# Check whether to break due to damaging rounding detected in updatepole
if subinfo == DAMAGING_ROUNDING:
info = subinfo
break # Better action to take? Geometry step, or simply continue?
# End of if reduce_rho. The procedure of reducing RHO ends.
# Report the current best value, and check if user asks for early termination.
if callback:
terminate = callback(sim[:, num_vars], fval[num_vars], nf, tr, cval[num_vars], conmat[:, num_vars])
if terminate:
info = CALLBACK_TERMINATE
break
# End of for loop. The iterative procedure ends
# Return from the calculation, after trying the last trust-region step if it has not been tried yet.
# Ensure that D has not been updated after SHORTD == TRUE occurred, or the code below is incorrect.
x = sim[:, num_vars] + d
if (info == SMALL_TR_RADIUS and
shortd and
norm(x - sim[:, num_vars]) > 1.0E-3 * rhoend and
nf < maxfun):
# Zaikun 20230615: UPDATEXFC or UPDATEPOLE is not called since the last trust-region step. Hence
# SIM[:, NUM_VARS] remains unchanged. Otherwise SIM[:, NUM_VARS] + D would not make sense.
f, constr = evaluate(calcfc, x, m_nlcon, amat, bvec)
cstrv = np.max(np.append(0, constr))
nf += 1
savehist(maxhist, x, xhist, f, fhist, cstrv, chist, constr, conhist)
nfilt, cfilt, ffilt, xfilt, confilt = savefilt(cstrv, ctol, cweight, f, x, nfilt, cfilt, ffilt, xfilt, constr, confilt)
# Zaikun 20230512: DELTA has been updated. RHO is only indicative here. TO BE IMPROVED.
fmsg(solver, 'Trust region', iprint, nf, rho, f, x, cstrv, constr)
# Return the best calculated values of the variables
# N.B.: SELECTX and FINDPOLE choose X by different standards, one cannot replace the other.
kopt = selectx(ffilt[:nfilt], cfilt[:nfilt], max(cpen, cweight), ctol)
x = xfilt[:, kopt]
f = ffilt[kopt]
constr = confilt[:, kopt]
cstrv = cfilt[kopt]
# Print a return message according to IPRINT.
retmsg(solver, info, iprint, nf, f, x, cstrv, constr)
return x, f, constr, cstrv, nf, xhist, fhist, chist, conhist, info
def getcpen(amat, bvec, conmat, cpen, cval, delta, fval, rho, sim, simi):
'''
This function gets the penalty parameter CPEN so that PREREM = PREREF + CPEN * PREREC > 0.
See the discussions around equation (9) of the COBYLA paper.
'''
# Even after nearly all of the pycutest problems were showing nearly bit for bit
# identical results between Python and the Fortran bindings, HS102 was still off by
# more than machine epsilon. It turned out to be due to the fact that getcpen was
# modifying fval, among other. It just goes to show that even when you're nearly
# perfect, you can still have non trivial bugs.
conmat = conmat.copy()
cval = cval.copy()
fval = fval.copy()
sim = sim.copy()
simi = simi.copy()
# Intermediate variables
A = np.zeros((np.size(sim, 0), np.size(conmat, 0)))
itol = 1
# Sizes
m_lcon = np.size(bvec) if bvec is not None else 0
num_constraints = np.size(conmat, 0)
num_vars = np.size(sim, 0)
# Preconditions
if DEBUGGING:
assert num_constraints >= 0
assert num_vars >= 1
assert cpen > 0
assert np.size(conmat, 0) == num_constraints and np.size(conmat, 1) == num_vars + 1
assert not (np.isnan(conmat) | np.isneginf(conmat)).any()
assert np.size(cval) == num_vars + 1 and \
not any(cval < 0 | np.isnan(cval) | np.isposinf(cval))
assert np.size(fval) == num_vars + 1 and not any(np.isnan(fval) | np.isposinf(fval))
assert np.size(sim, 0) == num_vars and np.size(sim, 1) == num_vars + 1
assert np.isfinite(sim).all()
assert all(np.max(abs(sim[:, :num_vars]), axis=0) > 0)
assert np.size(simi, 0) == num_vars and np.size(simi, 1) == num_vars
assert np.isfinite(simi).all()
assert isinv(sim[:, :num_vars], simi, itol)
assert delta >= rho and rho > 0
#====================#
# Calculation starts #
#====================#
# Initialize INFO which is needed in the postconditions
info = INFO_DEFAULT
# Increase CPEN if necessary to ensure PREREM > 0. Branch back for the next loop
# if this change alters the optimal vertex of the current simplex.
# Note the following:
# 1. In each loop, CPEN is changed only if PREREC > 0 > PREREF, in which case
# PREREM is guaranteed positive after the update. Note that PREREC >= 0 and
# max(PREREC, PREREF) > 0 in theory. If this holds numerically as well then CPEN
# is not changed only if PREREC = 0 or PREREF >= 0, in which case PREREM is
# currently positive, explaining why CPEN needs no update.
# 2. Even without an upper bound for the loop counter, the loop can occur at most
# NUM_VARS+1 times. This is because the update of CPEN does not decrease CPEN,
# and hence it can make vertex J (J <= NUM_VARS) become the new optimal vertex
# only if CVAL[J] is less than CVAL[NUM_VARS], which can happen at most NUM_VARS
# times. See the paragraph below (9) in the COBYLA paper. After the "correct"
# optimal vertex is found, one more loop is needed to calculate CPEN, and hence
# the loop can occur at most NUM_VARS+1 times.
for iter in range(num_vars + 1):
# Switch the best vertex of the current simplex to SIM[:, NUM_VARS]
conmat, cval, fval, sim, simi, info = updatepole(cpen, conmat, cval, fval, sim,
simi)
# Check whether to exit due to damaging rounding in UPDATEPOLE
if info == DAMAGING_ROUNDING:
break
# Calculate the linear approximations to the objective and constraint functions.
g = matprod(fval[:num_vars] - fval[num_vars], simi)
A[:, :m_lcon] = amat.T if amat is not None else amat
A[:, m_lcon:] = matprod((conmat[m_lcon:, :num_vars] -
np.tile(conmat[m_lcon:, num_vars], (num_vars, 1)).T), simi).T
# Calculate the trust-region trial step D. Note that D does NOT depend on CPEN.
d = trstlp(A, -conmat[:, num_vars], delta, g)
# Predict the change to F (PREREF) and to the constraint violation (PREREC) due
# to D.
preref = -inprod(d, g) # Can be negative
prerec = cval[num_vars] - np.max(np.append(0, conmat[:, num_vars] + matprod(d, A)))
# PREREC <= 0 or PREREF >=0 or either is NaN
if not (prerec > 0 and preref < 0):
break
# Powell's code defines BARMU = -PREREF / PREREC, and CPEN is increased to
# 2*BARMU if and only if it is currently less than 1.5*BARMU, a very
# "Powellful" scheme. In our implementation, however, we set CPEN directly to
# the maximum between its current value and 2*BARMU while handling possible
# overflow. The simplifies the scheme without worsening the performance of
# COBYLA.
cpen = max(cpen, min(-2 * preref / prerec, REALMAX))
if findpole(cpen, cval, fval) == num_vars:
break
#==================#
# Calculation ends #
#==================#
# Postconditions
if DEBUGGING:
assert cpen >= cpen and cpen > 0
assert preref + cpen * prerec > 0 or info == DAMAGING_ROUNDING or \
not (prerec >= 0 and np.maximum(prerec, preref) > 0) or not np.isfinite(preref)
return cpen
def fcratio(conmat, fval):
'''
This function calculates the ratio between the "typical change" of F and that of CONSTR.
See equations (12)-(13) in Section 3 of the COBYLA paper for the definition of the ratio.
'''
# Preconditions
if DEBUGGING:
assert np.size(fval) >= 1
assert np.size(conmat, 1) == np.size(fval)
#====================#
# Calculation starts #
#====================#
cmin = np.min(-conmat, axis=1)
cmax = np.max(-conmat, axis=1)
fmin = min(fval)
fmax = max(fval)
if any(cmin < 0.5 * cmax) and fmin < fmax:
denom = np.min(np.maximum(cmax, 0) - cmin, where=cmin < 0.5 * cmax, initial=np.inf)
# Powell mentioned the following alternative in section 4 of his COBYLA paper. According to a test
# on 20230610, it does not make much difference to the performance.
# denom = np.max(max(*cmax, 0) - cmin, mask=(cmin < 0.5 * cmax))
r = (fmax - fmin) / denom
else:
r = 0
#==================#
# Calculation ends #
#==================#
# Postconditions
if DEBUGGING:
assert r >= 0
return r
@@ -0,0 +1,226 @@
'''
This module contains subroutines concerning the geometry-improving of the interpolation set.
Translated from Zaikun Zhang's modern-Fortran reference implementation in PRIMA.
Dedicated to late Professor M. J. D. Powell FRS (1936--2015).
Python translation by Nickolai Belakovski.
'''
from ..common.consts import DEBUGGING
from ..common.linalg import isinv, matprod, inprod, norm, primasum, primapow2
import numpy as np
def setdrop_tr(ximproved, d, delta, rho, sim, simi):
'''
This function finds (the index) of a current interpolation point to be replaced with
the trust-region trial point. See (19)-(22) of the COBYLA paper.
N.B.:
1. If XIMPROVED == True, then JDROP > 0 so that D is included into XPT. Otherwise,
it is a bug.
2. COBYLA never sets JDROP = NUM_VARS
TODO: Check whether it improves the performance if JDROP = NUM_VARS is allowed when
XIMPROVED is True. Note that UPDATEXFC should be revised accordingly.
'''
# Local variables
itol = 0.1
# Sizes
num_vars = np.size(sim, 0)
# Preconditions
if DEBUGGING:
assert num_vars >= 1
assert np.size(d) == num_vars and all(np.isfinite(d))
assert delta >= rho and rho > 0
assert np.size(sim, 0) == num_vars and np.size(sim, 1) == num_vars + 1
assert np.isfinite(sim).all()
assert all(np.max(abs(sim[:, :num_vars]), axis=0) > 0)
assert np.size(simi, 0) == num_vars and np.size(simi, 1) == num_vars
assert np.isfinite(simi).all()
assert isinv(sim[:, :num_vars], simi, itol)
#====================#
# Calculation starts #
#====================#
# -------------------------------------------------------------------------------------------------- #
# The following code is Powell's scheme for defining JDROP.
# -------------------------------------------------------------------------------------------------- #
# ! JDROP = 0 by default. It cannot be removed, as JDROP may not be set below in some cases (e.g.,
# ! when XIMPROVED == FALSE, MAXVAL(ABS(SIMID)) <= 1, and MAXVAL(VETA) <= EDGMAX).
# jdrop = 0
#
# ! SIMID(J) is the value of the J-th Lagrange function at D. It is the counterpart of VLAG in UOBYQA
# ! and DEN in NEWUOA/BOBYQA/LINCOA, but it excludes the value of the (N+1)-th Lagrange function.
# simid = matprod(simi, d)
# if (any(abs(simid) > 1) .or. (ximproved .and. any(.not. is_nan(simid)))) then
# jdrop = int(maxloc(abs(simid), mask=(.not. is_nan(simid)), dim=1), kind(jdrop))
# !!MATLAB: [~, jdrop] = max(simid, [], 'omitnan');
# end if
#
# ! VETA(J) is the distance from the J-th vertex of the simplex to the best vertex, taking the trial
# ! point SIM(:, N+1) + D into account.
# if (ximproved) then
# veta = sqrt(sum((sim(:, 1:n) - spread(d, dim=2, ncopies=n))**2, dim=1))
# !!MATLAB: veta = sqrt(sum((sim(:, 1:n) - d).^2)); % d should be a column! Implicit expansion
# else
# veta = sqrt(sum(sim(:, 1:n)**2, dim=1))
# end if
#
# ! VSIG(J) (J=1, .., N) is the Euclidean distance from vertex J to the opposite face of the simplex.
# vsig = ONE / sqrt(sum(simi**2, dim=2))
# sigbar = abs(simid) * vsig
#
# ! The following JDROP will overwrite the previous one if its premise holds.
# mask = (veta > factor_delta * delta .and. (sigbar >= factor_alpha * delta .or. sigbar >= vsig))
# if (any(mask)) then
# jdrop = int(maxloc(veta, mask=mask, dim=1), kind(jdrop))
# !!MATLAB: etamax = max(veta(mask)); jdrop = find(mask & ~(veta < etamax), 1, 'first');
# end if
#
# ! Powell's code does not include the following instructions. With Powell's code, if SIMID consists
# ! of only NaN, then JDROP can be 0 even when XIMPROVED == TRUE (i.e., D reduces the merit function).
# ! With the following code, JDROP cannot be 0 when XIMPROVED == TRUE, unless VETA is all NaN, which
# ! should not happen if X0 does not contain NaN, the trust-region/geometry steps never contain NaN,
# ! and we exit once encountering an iterate containing Inf (due to overflow).
# if (ximproved .and. jdrop <= 0) then ! Write JDROP <= 0 instead of JDROP == 0 for robustness.
# jdrop = int(maxloc(veta, mask=(.not. is_nan(veta)), dim=1), kind(jdrop))
# !!MATLAB: [~, jdrop] = max(veta, [], 'omitnan');
# end if
# -------------------------------------------------------------------------------------------------- #
# Powell's scheme ends here.
# -------------------------------------------------------------------------------------------------- #
# The following definition of JDROP is inspired by SETDROP_TR in UOBYQA/NEWUOA/BOBYQA/LINCOA.
# It is simpler and works better than Powell's scheme. Note that we allow JDROP to be NUM_VARS+1 if
# XIMPROVED is True, whereas Powell's code does not.
# See also (4.1) of Scheinberg-Toint-2010: Self-Correcting Geometry in Model-Based Algorithms for
# Derivative-Free Unconstrained Optimization, which refers to the strategy here as the "combined
# distance/poisedness criteria".
# DISTSQ[j] is the square of the distance from the jth vertex of the simplex to get "best" point so
# far, taking the trial point SIM[:, NUM_VARS] + D into account.
distsq = np.zeros(np.size(sim, 1))
if ximproved:
distsq[:num_vars] = primasum(primapow2(sim[:, :num_vars] - np.tile(d, (num_vars, 1)).T), axis=0)
distsq[num_vars] = primasum(d*d)
else:
distsq[:num_vars] = primasum(primapow2(sim[:, :num_vars]), axis=0)
distsq[num_vars] = 0
weight = np.maximum(1, distsq / primapow2(np.maximum(rho, delta/10))) # Similar to Powell's NEWUOA code.
# Other possible definitions of weight. They work almost the same as the one above.
# weight = distsq # Similar to Powell's LINCOA code, but WRONG. See comments in LINCOA/geometry.f90.
# weight = max(1, max(25 * distsq / delta**2)) # Similar to Powell's BOBYQA code, works well.
# weight = max(1, max(10 * distsq / delta**2))
# weight = max(1, max(1e2 * distsq / delta**2))
# weight = max(1, max(distsq / rho**2)) ! Similar to Powell's UOBYQA
# If 0 <= j < NUM_VARS, SIMID[j] is the value of the jth Lagrange function at D; the value of the
# (NUM_VARS+1)th Lagrange function is 1 - sum(SIMID). [SIMID, 1 - sum(SIMID)] is the counterpart of
# VLAG in UOBYQA and DEN in NEWUOA/BOBYQA/LINCOA.
simid = matprod(simi, d)
score = weight * abs(np.array([*simid, 1 - primasum(simid)]))
# If XIMPROVED = False (D does not render a better X), set SCORE[NUM_VARS] = -1 to avoid JDROP = NUM_VARS.
if not ximproved:
score[num_vars] = -1
# score[j] is NaN implies SIMID[j] is NaN, but we want abs(SIMID) to be big. So we
# exclude such j.
score[np.isnan(score)] = -1
jdrop = None
# The following if statement works a bit better than
# `if any(score > 1) or (any(score > 0) and ximproved)` from Powell's UOBYQA and
# NEWUOA code.
if any(score > 0): # Powell's BOBYQA and LINCOA code.
jdrop = np.argmax(score)
if (ximproved and jdrop is None):
jdrop = np.argmax(distsq)
#==================#
# Calculation ends #
#==================#
# Postconditions
if DEBUGGING:
assert jdrop is None or (0 <= jdrop < num_vars + 1)
assert jdrop <= num_vars or ximproved
assert jdrop >= 0 or not ximproved
# JDROP >= 1 when XIMPROVED = TRUE unless NaN occurs in DISTSQ, which should not happen if the
# starting point does not contain NaN and the trust-region/geometry steps never contain NaN.
return jdrop
def geostep(jdrop, amat, bvec, conmat, cpen, cval, delbar, fval, simi):
'''
This function calculates a geometry step so that the geometry of the interpolation set is improved
when SIM[: JDROP_GEO] is replaced with SIM[:, NUM_VARS] + D. See (15)--(17) of the COBYLA paper.
'''
# Sizes
m_lcon = np.size(bvec, 0) if bvec is not None else 0
num_constraints = np.size(conmat, 0)
num_vars = np.size(simi, 0)
# Preconditions
if DEBUGGING:
assert num_constraints >= m_lcon >= 0
assert num_vars >= 1
assert delbar > 0
assert cpen > 0
assert np.size(simi, 0) == num_vars and np.size(simi, 1) == num_vars
assert np.isfinite(simi).all()
assert np.size(fval) == num_vars + 1 and not any(np.isnan(fval) | np.isposinf(fval))
assert np.size(conmat, 0) == num_constraints and np.size(conmat, 1) == num_vars + 1
assert not np.any(np.isnan(conmat) | np.isposinf(conmat))
assert np.size(cval) == num_vars + 1 and not any(cval < 0 | np.isnan(cval) | np.isposinf(cval))
assert 0 <= jdrop < num_vars
#====================#
# Calculation starts #
#====================#
# SIMI[JDROP, :] is a vector perpendicular to the face of the simplex to the opposite of vertex
# JDROP. Set D to the vector in this direction and with length DELBAR.
d = simi[jdrop, :]
d = delbar * (d / norm(d))
# The code below chooses the direction of D according to an approximation of the merit function.
# See (17) of the COBYLA paper and line 225 of Powell's cobylb.f.
# Calculate the coefficients of the linear approximations to the objective and constraint functions.
# N.B.: CONMAT and SIMI have been updated after the last trust-region step, but G and A have not.
# So we cannot pass G and A from outside.
g = matprod(fval[:num_vars] - fval[num_vars], simi)
A = np.zeros((num_vars, num_constraints))
A[:, :m_lcon] = amat.T if amat is not None else amat
A[:, m_lcon:] = matprod((conmat[m_lcon:, :num_vars] -
np.tile(conmat[m_lcon:, num_vars], (num_vars, 1)).T), simi).T
# CVPD and CVND are the predicted constraint violation of D and -D by the linear models.
cvpd = np.max(np.append(0, conmat[:, num_vars] + matprod(d, A)))
cvnd = np.max(np.append(0, conmat[:, num_vars] - matprod(d, A)))
if -inprod(d, g) + cpen * cvnd < inprod(d, g) + cpen * cvpd:
d *= -1
#==================#
# Calculation ends #
#==================#
# Postconditions
if DEBUGGING:
assert np.size(d) == num_vars and all(np.isfinite(d))
# In theory, ||S|| == DELBAR, which may be false due to rounding, but not too far.
# It is crucial to ensure that the geometry step is nonzero, which holds in theory.
assert 0.9 * delbar < np.linalg.norm(d) <= 1.1 * delbar
return d
@@ -0,0 +1,215 @@
'''
This module contains subroutines for initialization.
Translated from Zaikun Zhang's modern-Fortran reference implementation in PRIMA.
Dedicated to late Professor M. J. D. Powell FRS (1936--2015).
Python translation by Nickolai Belakovski.
'''
from ..common.checkbreak import checkbreak_con
from ..common.consts import DEBUGGING, REALMAX
from ..common.infos import INFO_DEFAULT
from ..common.evaluate import evaluate
from ..common.history import savehist
from ..common.linalg import inv
from ..common.message import fmsg
from ..common.selectx import savefilt
import numpy as np
def initxfc(calcfc, iprint, maxfun, constr0, amat, bvec, ctol, f0, ftarget, rhobeg, x0,
xhist, fhist, chist, conhist, maxhist):
'''
This subroutine does the initialization concerning X, function values, and
constraints.
'''
# Local variables
solver = 'COBYLA'
srname = "INITIALIZE"
# Sizes
num_constraints = np.size(constr0)
m_lcon = np.size(bvec) if bvec is not None else 0
m_nlcon = num_constraints - m_lcon
num_vars = np.size(x0)
# Preconditions
if DEBUGGING:
assert num_constraints >= 0, f'M >= 0 {srname}'
assert num_vars >= 1, f'N >= 1 {srname}'
assert abs(iprint) <= 3, f'IPRINT is 0, 1, -1, 2, -2, 3, or -3 {srname}'
# assert conmat.shape == (num_constraints , num_vars + 1), f'CONMAT.shape = [M, N+1] {srname}'
# assert cval.size == num_vars + 1, f'CVAL.size == N+1 {srname}'
# assert maxchist * (maxchist - maxhist) == 0, f'CHIST.shape == 0 or MAXHIST {srname}'
# assert conhist.shape[0] == num_constraints and maxconhist * (maxconhist - maxhist) == 0, 'CONHIST.shape[0] == num_constraints, SIZE(CONHIST, 2) == 0 or MAXHIST {srname)}'
# assert maxfhist * (maxfhist - maxhist) == 0, f'FHIST.shape == 0 or MAXHIST {srname}'
# assert xhist.shape[0] == num_vars and maxxhist * (maxxhist - maxhist) == 0, 'XHIST.shape[0] == N, SIZE(XHIST, 2) == 0 or MAXHIST {srname)}'
assert all(np.isfinite(x0)), f'X0 is finite {srname}'
assert rhobeg > 0, f'RHOBEG > 0 {srname}'
#====================#
# Calculation starts #
#====================#
# Initialize info to the default value. At return, a value different from this
# value will indicate an abnormal return
info = INFO_DEFAULT
# Initialize the simplex. It will be revised during the initialization.
sim = np.eye(num_vars, num_vars+1) * rhobeg
sim[:, num_vars] = x0
# Initialize the matrix simi. In most cases simi is overwritten, but not always.
simi = np.eye(num_vars) / rhobeg
# evaluated[j] = True iff the function/constraint of SIM[:, j] has been evaluated.
evaluated = np.zeros(num_vars+1, dtype=bool)
# Initialize fval
fval = np.zeros(num_vars+1) + REALMAX
cval = np.zeros(num_vars+1) + REALMAX
conmat = np.zeros((num_constraints, num_vars+1)) + REALMAX
for k in range(num_vars + 1):
x = sim[:, num_vars].copy()
# We will evaluate F corresponding to SIM(:, J).
if k == 0:
j = num_vars
f = f0
constr = constr0
else:
j = k - 1
x[j] += rhobeg
f, constr = evaluate(calcfc, x, m_nlcon, amat, bvec)
cstrv = np.max(np.append(0, constr))
# Print a message about the function/constraint evaluation according to IPRINT.
fmsg(solver, 'Initialization', iprint, k, rhobeg, f, x, cstrv, constr)
# Save X, F, CONSTR, CSTRV into the history.
savehist(maxhist, x, xhist, f, fhist, cstrv, chist, constr, conhist)
# Save F, CONSTR, and CSTRV to FVAL, CONMAT, and CVAL respectively.
evaluated[j] = True
fval[j] = f
conmat[:, j] = constr
cval[j] = cstrv
# Check whether to exit.
subinfo = checkbreak_con(maxfun, k, cstrv, ctol, f, ftarget, x)
if subinfo != INFO_DEFAULT:
info = subinfo
break
# Exchange the new vertex of the initial simplex with the optimal vertex if necessary.
# This is the ONLY part that is essentially non-parallel.
if j < num_vars and fval[j] < fval[num_vars]:
fval[j], fval[num_vars] = fval[num_vars], fval[j]
cval[j], cval[num_vars] = cval[num_vars], cval[j]
conmat[:, [j, num_vars]] = conmat[:, [num_vars, j]]
sim[:, num_vars] = x
sim[j, :j+1] = -rhobeg # SIM[:, :j+1] is lower triangular
nf = np.count_nonzero(evaluated)
if evaluated.all():
# Initialize SIMI to the inverse of SIM[:, :num_vars]
simi = inv(sim[:, :num_vars])
#==================#
# Calculation ends #
#==================#
# Postconditions
if DEBUGGING:
assert nf <= maxfun, f'NF <= MAXFUN {srname}'
assert evaluated.size == num_vars + 1, f'EVALUATED.size == Num_vars + 1 {srname}'
# assert chist.size == maxchist, f'CHIST.size == MAXCHIST {srname}'
# assert conhist.shape== (num_constraints, maxconhist), f'CONHIST.shape == [M, MAXCONHIST] {srname}'
assert conmat.shape == (num_constraints, num_vars + 1), f'CONMAT.shape = [M, N+1] {srname}'
assert not (np.isnan(conmat).any() or np.isneginf(conmat).any()), f'CONMAT does not contain NaN/-Inf {srname}'
assert cval.size == num_vars + 1 and not (any(cval < 0) or any(np.isnan(cval)) or any(np.isposinf(cval))), f'CVAL.shape == Num_vars+1 and CVAL does not contain negative values or NaN/+Inf {srname}'
# assert fhist.shape == maxfhist, f'FHIST.shape == MAXFHIST {srname}'
# assert maxfhist * (maxfhist - maxhist) == 0, f'FHIST.shape == 0 or MAXHIST {srname}'
assert fval.size == num_vars + 1 and not (any(np.isnan(fval)) or any(np.isposinf(fval))), f'FVAL.shape == Num_vars+1 and FVAL is not NaN/+Inf {srname}'
# assert xhist.shape == (num_vars, maxxhist), f'XHIST.shape == [N, MAXXHIST] {srname}'
assert sim.shape == (num_vars, num_vars + 1), f'SIM.shape == [N, N+1] {srname}'
assert np.isfinite(sim).all(), f'SIM is finite {srname}'
assert all(np.max(abs(sim[:, :num_vars]), axis=0) > 0), f'SIM(:, 1:N) has no zero column {srname}'
assert simi.shape == (num_vars, num_vars), f'SIMI.shape == [N, N] {srname}'
assert np.isfinite(simi).all(), f'SIMI is finite {srname}'
assert np.allclose(sim[:, :num_vars] @ simi, np.eye(num_vars), rtol=0.1, atol=0.1) or not all(evaluated), f'SIMI = SIM(:, 1:N)^{-1} {srname}'
return evaluated, conmat, cval, sim, simi, fval, nf, info
def initfilt(conmat, ctol, cweight, cval, fval, sim, evaluated, cfilt, confilt, ffilt, xfilt):
'''
This function initializes the filter (XFILT, etc) that will be used when selecting
x at the end of the solver.
N.B.:
1. Why not initialize the filters using XHIST, etc? Because the history is empty if
the user chooses not to output it.
2. We decouple INITXFC and INITFILT so that it is easier to parallelize the former
if needed.
'''
# Sizes
num_constraints = conmat.shape[0]
num_vars = sim.shape[0]
maxfilt = len(ffilt)
# Preconditions
if DEBUGGING:
assert num_constraints >= 0
assert num_vars >= 1
assert maxfilt >= 1
assert np.size(confilt, 0) == num_constraints and np.size(confilt, 1) == maxfilt
assert np.size(cfilt) == maxfilt
assert np.size(xfilt, 0) == num_vars and np.size(xfilt, 1) == maxfilt
assert np.size(ffilt) == maxfilt
assert np.size(conmat, 0) == num_constraints and np.size(conmat, 1) == num_vars + 1
assert not (np.isnan(conmat) | np.isneginf(conmat)).any()
assert np.size(cval) == num_vars + 1 and not any(cval < 0 | np.isnan(cval) | np.isposinf(cval))
assert np.size(fval) == num_vars + 1 and not any(np.isnan(fval) | np.isposinf(fval))
assert np.size(sim, 0) == num_vars and np.size(sim, 1) == num_vars + 1
assert np.isfinite(sim).all()
assert all(np.max(abs(sim[:, :num_vars]), axis=0) > 0)
assert np.size(evaluated) == num_vars + 1
#====================#
# Calculation starts #
#====================#
nfilt = 0
for i in range(num_vars+1):
if evaluated[i]:
if i < num_vars:
x = sim[:, i] + sim[:, num_vars]
else:
x = sim[:, i] # i == num_vars, i.e. the last column
nfilt, cfilt, ffilt, xfilt, confilt = savefilt(cval[i], ctol, cweight, fval[i], x, nfilt, cfilt, ffilt, xfilt, conmat[:, i], confilt)
#==================#
# Calculation ends #
#==================#
# Postconditions
if DEBUGGING:
assert nfilt <= maxfilt
assert np.size(confilt, 0) == num_constraints and np.size(confilt, 1) == maxfilt
assert not (np.isnan(confilt[:, :nfilt]) | np.isneginf(confilt[:, :nfilt])).any()
assert np.size(cfilt) == maxfilt
assert not any(cfilt[:nfilt] < 0 | np.isnan(cfilt[:nfilt]) | np.isposinf(cfilt[:nfilt]))
assert np.size(xfilt, 0) == num_vars and np.size(xfilt, 1) == maxfilt
assert not (np.isnan(xfilt[:, :nfilt])).any()
# The last calculated X can be Inf (finite + finite can be Inf numerically).
assert np.size(ffilt) == maxfilt
assert not any(np.isnan(ffilt[:nfilt]) | np.isposinf(ffilt[:nfilt]))
return nfilt
@@ -0,0 +1,492 @@
'''
This module provides subroutines concerning the trust-region calculations of COBYLA.
Translated from Zaikun Zhang's modern-Fortran reference implementation in PRIMA.
Dedicated to late Professor M. J. D. Powell FRS (1936--2015).
Python translation by Nickolai Belakovski.
'''
import numpy as np
import numpy.typing as npt
from ..common.consts import DEBUGGING, REALMIN, REALMAX, EPS
from ..common.powalg import qradd_Rdiag, qrexc_Rdiag
from ..common.linalg import isminor, matprod, inprod, lsqr, primasum
def trstlp(A, b, delta, g):
'''
This function calculated an n-component vector d by the following two stages. In the first
stage, d is set to the shortest vector that minimizes the greatest violation of the constraints
A.T @ D <= B, K = 1, 2, 3, ..., M,
subject to the Euclidean length of d being at most delta. If its length is strictly less than
delta, then the second stage uses the resultant freedom in d to minimize the objective function
G.T @ D
subject to no increase in any greatest constraint violation.
It is possible but rare that a degeneracy may prevent d from attaining the target length delta.
cviol is the largest constraint violation of the current d: max(max(A.T@D - b), 0)
icon is the index of a most violated constraint if cviol is positive.
nact is the number of constraints in the active set and iact[0], ..., iact[nact-1] are their indices,
while the remainder of the iact contains a permutation of the remaining constraint indicies.
N.B.: nact <= min(num_constraints, num_vars). Obviously nact <= num_constraints. In addition, the constraints
in iact[0, ..., nact-1] have linearly independent gradients (see the comments above the instruction
that delete a constraint from the active set to make room for the new active constraint with index iact[icon]);
it can also be seen from the update of nact: starting from 0, nact is incremented only if nact < n.
Further, Z is an orthogonal matrix whose first nact columns can be regarded as the result of
Gram-Schmidt applied to the active constraint gradients. For j = 0, 1, ..., nact-1, the number
zdota[j] is the scalar product of the jth column of Z with the gradient of the jth active
constraint. d is the current vector of variables and here the residuals of the active constraints
should be zero. Further, the active constraints have nonnegative Lagrange multipliers that are
held at the beginning of vmultc. The remainder of this vector holds the residuals of the inactive
constraints at d, the ordering of the components of vmultc being in agreement with the permutation
of the indices of the constraints that is in iact. All these residuals are nonnegative, which is
achieved by the shift cviol that makes the least residual zero.
N.B.:
0. In Powell's implementation, the constraints are A.T @ D >= B. In other words, the A and B in
our implementation are the negative of those in Powell's implementation.
1. The algorithm was NOT documented in the COBYLA paper. A note should be written to introduce it!
2. As a major part of the algorithm (see trstlp_sub), the code maintains and updates the QR
factorization of A[iact[:nact]], i.e. the gradients of all the active (linear) constraints. The
matrix Z is indeed Q, and the vector zdota is the diagonal of R. The factorization is updated by
Givens rotations when an index is added in or removed from iact.
3. There are probably better algorithms available for the trust-region linear programming problem.
'''
# Sizes
num_constraints = A.shape[1]
num_vars = A.shape[0]
# Preconditions
if DEBUGGING:
assert num_vars >= 1
assert num_constraints >= 0
assert np.size(g) == num_vars
assert np.size(b) == num_constraints
assert delta > 0
vmultc = np.zeros(num_constraints + 1)
iact = np.zeros(num_constraints + 1, dtype=int)
nact = 0
d = np.zeros(num_vars)
z = np.zeros((num_vars, num_vars))
# ==================
# Calculation starts
# ==================
# Form A_aug and B_aug. This allows the gradient of the objective function to be regarded as the
# gradient of a constraint in the second stage.
A_aug = np.hstack([A, g.reshape((num_vars, 1))])
b_aug = np.hstack([b, 0])
# Scale the problem if A contains large values. Otherwise floating point exceptions may occur.
# Note that the trust-region step is scale invariant.
for i in range(num_constraints+1): # Note that A_aug.shape[1] == num_constraints+1
if (maxval:=max(abs(A_aug[:, i]))) > 1e12:
modscal = max(2*REALMIN, 1/maxval)
A_aug[:, i] *= modscal
b_aug[i] *= modscal
# Stage 1: minimize the 1+infinity constraint violation of the linearized constraints.
iact[:num_constraints], nact, d, vmultc[:num_constraints], z = trstlp_sub(iact[:num_constraints], nact, 1, A_aug[:, :num_constraints], b_aug[:num_constraints], delta, d, vmultc[:num_constraints], z)
# Stage 2: minimize the linearized objective without increasing the 1_infinity constraint violation.
iact, nact, d, vmultc, z = trstlp_sub(iact, nact, 2, A_aug, b_aug, delta, d, vmultc, z)
# ================
# Calculation ends
# ================
# Postconditions
if DEBUGGING:
assert all(np.isfinite(d))
# Due to rounding, it may happen that ||D|| > DELTA, but ||D|| > 2*DELTA is highly improbable.
assert np.linalg.norm(d) <= 2 * delta
return d
def trstlp_sub(iact: npt.NDArray, nact: int, stage, A, b, delta, d, vmultc, z):
'''
This subroutine does the real calculations for trstlp, both stage 1 and stage 2.
Major differences between stage 1 and stage 2:
1. Initialization. Stage 2 inherits the values of some variables from stage 1, so they are
initialized in stage 1 but not in stage 2.
2. cviol. cviol is updated after at iteration in stage 1, while it remains a constant in stage2.
3. sdirn. See the definition of sdirn in the code for details.
4. optnew. The two stages have different objectives, so optnew is updated differently.
5. step. step <= cviol in stage 1.
'''
zdasav = np.zeros(z.shape[1])
vmultd = np.zeros(np.size(vmultc))
zdota = np.zeros(np.size(z, 1))
# Sizes
mcon = np.size(A, 1)
num_vars = np.size(A, 0)
# Preconditions
if DEBUGGING:
assert num_vars >= 1
assert stage == 1 or stage == 2
assert (mcon >= 0 and stage == 1) or (mcon >= 1 and stage == 2)
assert np.size(b) == mcon
assert np.size(iact) == mcon
assert np.size(vmultc) == mcon
assert np.size(d) == num_vars
assert np.size(z, 0) == num_vars and np.size(z, 1) == num_vars
assert delta > 0
if stage == 2:
assert all(np.isfinite(d)) and np.linalg.norm(d) <= 2 * delta
assert nact >= 0 and nact <= np.minimum(mcon, num_vars)
assert all(vmultc[:mcon]) >= 0
# N.B.: Stage 1 defines only VMULTC(1:M); VMULTC(M+1) is undefined!
# Initialize according to stage
if stage == 1:
iact = np.linspace(0, mcon-1, mcon, dtype=int)
nact = 0
d = np.zeros(num_vars)
cviol = np.max(np.append(0, -b))
vmultc = cviol + b
z = np.eye(num_vars)
if mcon == 0 or cviol <= 0:
# Check whether a quick return is possible. Make sure the in-outputs have been initialized.
return iact, nact, d, vmultc, z
if all(np.isnan(b)):
return iact, nact, d, vmultc, z
else:
icon = np.nanargmax(-b)
num_constraints = mcon
sdirn = np.zeros(len(d))
else:
if inprod(d, d) >= delta*delta:
# Check whether a quick return is possible.
return iact, nact, d, vmultc, z
iact[mcon-1] = mcon-1
vmultc[mcon-1] = 0
num_constraints = mcon - 1
icon = mcon - 1
# In Powell's code, stage 2 uses the zdota and cviol calculated by stage1. Here we recalculate
# them so that they need not be passed from stage 1 to 2, and hence the coupling is reduced.
cviol = np.max(np.append(0, matprod(d, A[:, :num_constraints]) - b[:num_constraints]))
zdota[:nact] = [inprod(z[:, k], A[:, iact[k]]) for k in range(nact)]
# More initialization
optold = REALMAX
nactold = nact
nfail = 0
# Zaikun 20211011: vmultd is computed from scratch at each iteration, but vmultc is inherited
# Powell's code can encounter infinite cycling, which did happen when testing the following CUTEst
# problems: DANWOODLS, GAUSS1LS, GAUSS2LS, GAUSS3LS, KOEBHELB, TAX13322, TAXR13322. Indeed, in all
# these cases, Inf/NaN appear in d due to extremely large values in A (up to 10^219). To resolve
# this, we set the maximal number of iterations to maxiter, and terminate if Inf/NaN occurs in d.
maxiter = np.minimum(10000, 100*max(num_constraints, num_vars))
for iter in range(maxiter):
if DEBUGGING:
assert all(vmultc >= 0)
if stage == 1:
optnew = cviol
else:
optnew = inprod(d, A[:, mcon-1])
# End the current stage of the calculation if 3 consecutive iterations have either failed to
# reduce the best calculated value of the objective function or to increase the number of active
# constraints since the best value was calculated. This strategy prevents cycling, but there is
# a remote possibility that it will cause premature termination.
if optnew < optold or nact > nactold:
nactold = nact
nfail = 0
else:
nfail += 1
optold = np.minimum(optold, optnew)
if nfail == 3:
break
# If icon exceeds nact, then we add the constraint with index iact[icon] to the active set.
if icon >= nact: # In Python this needs to be >= since Python is 0-indexed (in Fortran we have 1 > 0, in Python we need 0 >= 0)
zdasav[:nact] = zdota[:nact]
nactsav = nact
z, zdota, nact = qradd_Rdiag(A[:, iact[icon]], z, zdota, nact) # May update nact to nact+1
# Indeed it suffices to pass zdota[:min(num_vars, nact+1)] to qradd as follows:
# qradd(A[:, iact[icon]], z, zdota[:min(num_vars, nact+1)], nact)
if nact == nactsav + 1:
# N.B.: It is possible to index arrays using [nact, icon] when nact == icon.
# Zaikun 20211012: Why should vmultc[nact] = 0?
if nact != (icon + 1): # Need to add 1 to Python for 0 indexing
vmultc[[icon, nact-1]] = vmultc[nact-1], 0
iact[[icon, nact-1]] = iact[[nact-1, icon]]
else:
vmultc[nact-1] = 0
else:
# Zaikun 20211011:
# 1. VMULTD is calculated from scratch for the first time (out of 2) in one iteration.
# 2. Note that IACT has not been updated to replace IACT[NACT] with IACT[ICON]. Thus
# A[:, IACT[:NACT]] is the UNUPDATED version before QRADD (note Z[:, :NACT] remains the
# same before and after QRADD). Therefore if we supply ZDOTA to LSQR (as Rdiag) as
# Powell did, we should use the UNUPDATED version, namely ZDASAV.
# vmultd[:nact] = lsqr(A[:, iact[:nact]], A[:, iact[icon]], z[:, :nact], zdasav[:nact])
vmultd[:nact] = lsqr(A[:, iact[:nact]], A[:, iact[icon]], z[:, :nact], zdasav[:nact])
if not any(np.logical_and(vmultd[:nact] > 0, iact[:nact] <= num_constraints)):
# N.B.: This can be triggered by NACT == 0 (among other possibilities)! This is
# important, because NACT will be used as an index in the sequel.
break
# vmultd[NACT+1:mcon] is not used, but we have to initialize it in Fortran, or compilers
# complain about the where construct below (another solution: restrict where to 1:NACT).
vmultd[nact:mcon] = -1 # len(vmultd) == mcon
# Revise the Lagrange multipliers. The revision is not applicable to vmultc[nact:num_constraints].
fracmult = [vmultc[i]/vmultd[i] if vmultd[i] > 0 and iact[i] <= num_constraints else REALMAX for i in range(nact)]
# Only the places with vmultd > 0 and iact <= m is relevant below, if any.
frac = min(fracmult[:nact]) # fracmult[nact:mcon] may contain garbage
vmultc[:nact] = np.maximum(np.zeros(len(vmultc[:nact])), vmultc[:nact] - frac*vmultd[:nact])
# Reorder the active constraints so that the one to be replaced is at the end of the list.
# Exit if the new value of zdota[nact] is not acceptable. Powell's condition for the
# following If: not abs(zdota[nact]) > 0. Note that it is different from
# 'abs(zdota[nact]) <=0)' as zdota[nact] can be NaN.
# N.B.: We cannot arrive here with nact == 0, which should have triggered a break above
if np.isnan(zdota[nact - 1]) or abs(zdota[nact - 1]) <= EPS**2:
break
vmultc[[icon, nact - 1]] = 0, frac # vmultc[[icon, nact]] is valid as icon > nact
iact[[icon, nact - 1]] = iact[[nact - 1, icon]]
# end if nact == nactsav + 1
# In stage 2, ensure that the objective continues to be treated as the last active constraint.
# Zaikun 20211011, 20211111: Is it guaranteed for stage 2 that iact[nact-1] = mcon when
# iact[nact] != mcon??? If not, then how does the following procedure ensure that mcon is
# the last of iact[:nact]?
if stage == 2 and iact[nact - 1] != (mcon - 1):
if nact <= 1:
# We must exit, as nact-2 is used as an index below. Powell's code does not have this.
break
z, zdota[:nact] = qrexc_Rdiag(A[:, iact[:nact]], z, zdota[:nact], nact - 2) # We pass nact-2 in Python instead of nact-1
# Indeed, it suffices to pass Z[:, :nact] to qrexc as follows:
# z[:, :nact], zdota[:nact] = qrexc(A[:, iact[:nact]], z[:, :nact], zdota[:nact], nact - 1)
iact[[nact-2, nact-1]] = iact[[nact-1, nact-2]]
vmultc[[nact-2, nact-1]] = vmultc[[nact-1, nact-2]]
# Zaikun 20211117: It turns out that the last few lines do not guarantee iact[nact] == num_vars in
# stage 2; the following test cannot be passed. IS THIS A BUG?!
# assert iact[nact] == mcon or stage == 1, 'iact[nact] must == mcon in stage 2'
# Powell's code does not have the following. It avoids subsequent floating points exceptions.
if np.isnan(zdota[nact-1]) or abs(zdota[nact-1]) <= EPS**2:
break
# Set sdirn to the direction of the next change to the current vector of variables
# Usually during stage 1 the vector sdirn gives a search direction that reduces all the
# active constraint violations by one simultaneously.
if stage == 1:
sdirn -= ((inprod(sdirn, A[:, iact[nact-1]]) + 1)/zdota[nact-1])*z[:, nact-1]
else:
sdirn = -1/zdota[nact-1]*z[:, nact-1]
else: # icon < nact
# Delete the constraint with the index iact[icon] from the active set, which is done by
# reordering iact[icon:nact] into [iact[icon+1:nact], iact[icon]] and then reduce nact to
# nact - 1. In theory, icon > 0.
# assert icon > 0, "icon > 0 is required" # For Python I think this is irrelevant
z, zdota[:nact] = qrexc_Rdiag(A[:, iact[:nact]], z, zdota[:nact], icon) # qrexc does nothing if icon == nact
# Indeed, it suffices to pass Z[:, :nact] to qrexc as follows:
# z[:, :nact], zdota[:nact] = qrexc(A[:, iact[:nact]], z[:, :nact], zdota[:nact], icon)
iact[icon:nact] = [*iact[icon+1:nact], iact[icon]]
vmultc[icon:nact] = [*vmultc[icon+1:nact], vmultc[icon]]
nact -= 1
# Powell's code does not have the following. It avoids subsequent exceptions.
# Zaikun 20221212: In theory, nact > 0 in stage 2, as the objective function should always
# be considered as an "active constraint" --- more precisely, iact[nact] = mcon. However,
# looking at the code, I cannot see why in stage 2 nact must be positive after the reduction
# above. It did happen in stage 1 that nact became 0 after the reduction --- this is
# extremely rare, and it was never observed until 20221212, after almost one year of
# random tests. Maybe nact is theoretically positive even in stage 1?
if stage == 2 and nact < 0:
break # If this case ever occurs, we have to break, as nact is used as an index below.
if nact > 0:
if np.isnan(zdota[nact-1]) or abs(zdota[nact-1]) <= EPS**2:
break
# Set sdirn to the direction of the next change to the current vector of variables.
if stage == 1:
sdirn -= inprod(sdirn, z[:, nact]) * z[:, nact]
# sdirn is orthogonal to z[:, nact+1]
else:
sdirn = -1/zdota[nact-1] * z[:, nact-1]
# end if icon > nact
# Calculate the step to the trust region boundary or take the step that reduces cviol to 0.
# ----------------------------------------------------------------------------------------- #
# The following calculation of step is adopted from NEWUOA/BOBYQA/LINCOA. It seems to improve
# the performance of COBYLA. We also found that removing the precaution about underflows is
# beneficial to the overall performance of COBYLA --- the underflows are harmless anyway.
dd = delta*delta - inprod(d, d)
ss = inprod(sdirn, sdirn)
sd = inprod(sdirn, d)
if dd <= 0 or ss <= EPS * delta*delta or np.isnan(sd):
break
# sqrtd: square root of a discriminant. The max avoids sqrtd < abs(sd) due to underflow
sqrtd = max(np.sqrt(ss*dd + sd*sd), abs(sd), np.sqrt(ss * dd))
if sd > 0:
step = dd / (sqrtd + sd)
else:
step = (sqrtd - sd) / ss
# step < 0 should not happen. Step can be 0 or NaN when, e.g., sd or ss becomes inf
if step <= 0 or not np.isfinite(step):
break
# Powell's approach and comments are as follows.
# -------------------------------------------------- #
# The two statements below that include the factor eps prevent
# some harmless underflows that occurred in a test calculation
# (Zaikun: here, eps is the machine epsilon; Powell's original
# code used 1.0e-6, and Powell's code was written in single
# precision). Further, we skip the step if it could be 0 within
# a reasonable tolerance for computer rounding errors.
# !dd = delta*delta - sum(d**2, mask=(abs(d) >= EPS * delta))
# !ss = inprod(sdirn, sdirn)
# !if (dd <= 0) then
# ! exit
# !end if
# !sd = inprod(sdirn, d)
# !if (abs(sd) >= EPS * sqrt(ss * dd)) then
# ! step = dd / (sqrt(ss * dd + sd*sd) + sd)
# !else
# ! step = dd / (sqrt(ss * dd) + sd)
# !end if
# -------------------------------------------------- #
if stage == 1:
if isminor(cviol, step):
break
step = min(step, cviol)
# Set dnew to the new variables if step is the steplength, and reduce cviol to the corresponding
# maximum residual if stage 1 is being done
dnew = d + step * sdirn
if stage == 1:
cviol = np.max(np.append(0, matprod(dnew, A[:, iact[:nact]]) - b[iact[:nact]]))
# N.B.: cviol will be used when calculating vmultd[nact+1:mcon].
# Zaikun 20211011:
# 1. vmultd is computed from scratch for the second (out of 2) time in one iteration.
# 2. vmultd[:nact] and vmultd[nact:mcon] are calculated separately with no coupling.
# 3. vmultd will be calculated from scratch again in the next iteration.
# Set vmultd to the vmultc vector that would occur if d became dnew. A device is included to
# force vmultd[k] = 0 if deviations from this value can be attributed to computer rounding
# errors. First calculate the new Lagrange multipliers.
vmultd[:nact] = -lsqr(A[:, iact[:nact]], dnew, z[:, :nact], zdota[:nact])
if stage == 2:
vmultd[nact-1] = max(0, vmultd[nact-1]) # This seems never activated.
# Complete vmultd by finding the new constraint residuals. (Powell wrote "Complete vmultc ...")
cvshift = cviol - (matprod(dnew, A[:, iact]) - b[iact]) # Only cvshift[nact+1:mcon] is needed
cvsabs = matprod(abs(dnew), abs(A[:, iact])) + abs(b[iact]) + cviol
cvshift[isminor(cvshift, cvsabs)] = 0
vmultd[nact:mcon] = cvshift[nact:mcon]
# Calculate the fraction of the step from d to dnew that will be taken
fracmult = [vmultc[i]/(vmultc[i] - vmultd[i]) if vmultd[i] < 0 else REALMAX for i in range(len(vmultd))]
# Only the places with vmultd < 0 are relevant below, if any.
icon = np.argmin(np.append(1, fracmult)) - 1
frac = min(np.append(1, fracmult))
# Update d, vmultc, and cviol
dold = d
d = (1 - frac)*d + frac * dnew
vmultc = np.maximum(0, (1 - frac)*vmultc + frac*vmultd)
# Break in the case of inf/nan in d or vmultc.
if not (np.isfinite(primasum(abs(d))) and np.isfinite(primasum(abs(vmultc)))):
d = dold # Should we restore also iact, nact, vmultc, and z?
break
if stage == 1:
# cviol = (1 - frac) * cvold + frac * cviol # Powell's version
# In theory, cviol = np.max(np.append(d@A - b, 0)), yet the
# cviol updated as above can be quite different from this value if A has huge entries (e.g., > 1e20)
cviol = np.max(np.append(0, matprod(d, A) - b))
if icon < 0 or icon >= mcon:
# In Powell's code, the condition is icon == 0. Indeed, icon < 0 cannot hold unless
# fracmult contains only nan, which should not happen; icon >= mcon should never occur.
break
#==================#
# Calculation ends #
#==================#
# Postconditions
if DEBUGGING:
assert np.size(iact) == mcon
assert np.size(vmultc) == mcon
assert all(vmultc >= 0)
assert np.size(d) == num_vars
assert all(np.isfinite(d))
assert np.linalg.norm(d) <= 2 * delta
assert np.size(z, 0) == num_vars and np.size(z, 1) == num_vars
assert nact >= 0 and nact <= np.minimum(mcon, num_vars)
return iact, nact, d, vmultc, z
def trrad(delta_in, dnorm, eta1, eta2, gamma1, gamma2, ratio):
'''
This function updates the trust region radius according to RATIO and DNORM.
'''
# Preconditions
if DEBUGGING:
assert delta_in >= dnorm > 0
assert 0 <= eta1 <= eta2 < 1
assert 0 < gamma1 < 1 < gamma2
# By the definition of RATIO in ratio.f90, RATIO cannot be NaN unless the
# actual reduction is NaN, which should NOT happen due to the moderated extreme
# barrier.
assert not np.isnan(ratio)
#====================#
# Calculation starts #
#====================#
if ratio <= eta1:
delta = gamma1 * dnorm # Powell's UOBYQA/NEWUOA
# delta = gamma1 * delta_in # Powell's COBYLA/LINCOA
# delta = min(gamma1 * delta_in, dnorm) # Powell's BOBYQA
elif ratio <= eta2:
delta = max(gamma1 * delta_in, dnorm) # Powell's UOBYQA/NEWUOA/BOBYQA/LINCOA
else:
delta = max(gamma1 * delta_in, gamma2 * dnorm) # Powell's NEWUOA/BOBYQA
# delta = max(delta_in, gamma2 * dnorm) # Modified version. Works well for UOBYQA
# For noise-free CUTEst problems of <= 100 variables, Powell's version works slightly better
# than the modified one.
# delta = max(delta_in, 1.25*dnorm, dnorm + rho) # Powell's UOBYQA
# delta = min(max(gamma1 * delta_in, gamma2 * dnorm), gamma3 * delta_in) # Powell's LINCOA, gamma3 = np.sqrt(2)
# For noisy problems, the following may work better.
# if ratio <= eta1:
# delta = gamma1 * dnorm
# elseif ratio <= eta2: # Ensure DELTA >= DELTA_IN
# delta = delta_in
# else: # Ensure DELTA > DELTA_IN with a constant factor
# delta = max(delta_in * (1 + gamma2) / 2, gamma2 * dnorm)
#==================#
# Calculation ends #
#==================#
# Postconditions
if DEBUGGING:
assert delta > 0
return delta

Some files were not shown because too many files have changed in this diff Show More