| 1 |
""" |
| 2 |
AUTORAY - backend agnostic array operations. |
| 3 |
|
| 4 |
|
| 5 |
Copyright 2019-2023 Johnnie Gray |
| 6 |
|
| 7 |
Licensed under the Apache License, Version 2.0 (the "License"); |
| 8 |
you may not use this file except in compliance with the License. |
| 9 |
You may obtain a copy of the License at |
| 10 |
|
| 11 |
http://www.apache.org/licenses/LICENSE-2.0 |
| 12 |
|
| 13 |
Unless required by applicable law or agreed to in writing, software |
| 14 |
distributed under the License is distributed on an "AS IS" BASIS, |
| 15 |
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 16 |
See the License for the specific language governing permissions and |
| 17 |
limitations under the License. |
| 18 |
""" |
| 19 |
|
| 20 |
import contextlib |
| 21 |
import functools |
| 22 |
import importlib |
| 23 |
import itertools |
| 24 |
import math |
| 25 |
import threading |
| 26 |
from collections import OrderedDict, defaultdict |
| 27 |
from inspect import signature |
| 28 |
|
| 29 |
|
| 30 |
def do(fn, *args, like=None, **kwargs): |
| 31 |
"""Do function named ``fn`` on ``(*args, **kwargs)``, peforming single |
| 32 |
dispatch to retrieve ``fn`` based on whichever library defines the class of |
| 33 |
the ``args[0]``, or the ``like`` keyword argument if specified. |
| 34 |
|
| 35 |
Examples |
| 36 |
-------- |
| 37 |
|
| 38 |
Works on numpy arrays: |
| 39 |
|
| 40 |
>>> import numpy as np |
| 41 |
>>> x_np = np.random.uniform(size=[5]) |
| 42 |
>>> y_np = do('sqrt', x_np) |
| 43 |
>>> y_np |
| 44 |
array([0.32464973, 0.90379787, 0.85037325, 0.88729814, 0.46768083]) |
| 45 |
|
| 46 |
>>> type(y_np) |
| 47 |
numpy.ndarray |
| 48 |
|
| 49 |
Works on cupy arrays: |
| 50 |
|
| 51 |
>>> import cupy as cp |
| 52 |
>>> x_cp = cp.random.uniform(size=[5]) |
| 53 |
>>> y_cp = do('sqrt', x_cp) |
| 54 |
>>> y_cp |
| 55 |
array([0.44541656, 0.88713113, 0.92626237, 0.64080557, 0.69620767]) |
| 56 |
|
| 57 |
>>> type(y_cp) |
| 58 |
cupy.core.core.ndarray |
| 59 |
|
| 60 |
Works on tensorflow arrays: |
| 61 |
|
| 62 |
>>> import tensorflow as tf |
| 63 |
>>> x_tf = tf.random.uniform(shape=[5]) |
| 64 |
>>> y_tf = do('sqrt', x_tf) |
| 65 |
>>> y_tf |
| 66 |
<tf.Tensor 'Sqrt_1:0' shape=(5,) dtype=float32> |
| 67 |
|
| 68 |
>>> type(y_tf) |
| 69 |
tensorflow.python.framework.ops.Tensor |
| 70 |
|
| 71 |
You get the idea. |
| 72 |
|
| 73 |
For functions that don't dispatch on the first argument you can use the |
| 74 |
``like`` keyword: |
| 75 |
|
| 76 |
>>> do('eye', 3, like=x_tf) |
| 77 |
<tf.Tensor: id=91, shape=(3, 3), dtype=float32> |
| 78 |
""" |
| 79 |
backend = _choose_backend(fn, args, kwargs, like=like) |
| 80 |
func = get_lib_fn(backend, fn) |
| 81 |
return func(*args, **kwargs) |
| 82 |
|
| 83 |
|
| 84 |
# ------------------------- efficiently dispatching ------------------------- # |
| 85 |
|
| 86 |
|
| 87 |
def _default_infer_from_sig(fn, *args, **kwargs): |
| 88 |
"""This is the default backend dispatcher, used if no global backend has |
| 89 |
been set. Hot swapping this function out as below avoids having to check |
| 90 |
manually for a global backend or worse, a thread aware global backend, on |
| 91 |
every call to ``do``. |
| 92 |
""" |
| 93 |
return _DISPATCHERS[fn](*args, **kwargs) |
| 94 |
|
| 95 |
|
| 96 |
_global_backend = None |
| 97 |
_inferrer_global = _default_infer_from_sig |
| 98 |
|
| 99 |
# this is the function that autoray uses when `do` is called without an |
| 100 |
# explicit like/backend argument. It is set to `_default_infer_from_sig` by |
| 101 |
# default, but can be set to `_always_the_same` if a global backend is set e.g. |
| 102 |
_infer_auto = _inferrer_global |
| 103 |
|
| 104 |
# if a thread that isn't the 'importing' thread tries to set a backend, this |
| 105 |
# by default turns on thread aware dispatching, but once such custom sub |
| 106 |
# backends have been reset, the global values above are used again. |
| 107 |
_global_backends_threadaware = {} |
| 108 |
_inferrers_threadaware = {} |
| 109 |
_importing_thrid = threading.get_ident() |
| 110 |
_backend_lock = threading.Lock() |
| 111 |
|
| 112 |
|
| 113 |
def _default_infer_from_sig_threadaware(fn, args, kwargs): |
| 114 |
# check for a thread aware inferrer, default to the global inferrer |
| 115 |
thrid = threading.get_ident() |
| 116 |
return _inferrers_threadaware.get(thrid, _inferrer_global)( |
| 117 |
fn, args, kwargs |
| 118 |
) |
| 119 |
|
| 120 |
|
| 121 |
def _always_the_same(fn, args, kwargs, backend): |
| 122 |
return backend |
| 123 |
|
| 124 |
|
| 125 |
def get_backend(get_globally="auto"): |
| 126 |
"""Return the universally set backend, if any. |
| 127 |
|
| 128 |
Parameters |
| 129 |
---------- |
| 130 |
get_globally : {"auto", False, True}, optional |
| 131 |
Which backend to return: |
| 132 |
|
| 133 |
- True: return the globally set backend, if any. |
| 134 |
- False: return the backend set for the current thread, if any. |
| 135 |
- "auto": return the globally set backend, if this thread is the thread |
| 136 |
that imported autoray. Otherwise return the backend set for the |
| 137 |
current thread, if any. |
| 138 |
|
| 139 |
Returns |
| 140 |
------- |
| 141 |
backend : str or None |
| 142 |
The name of the backend, or None if no backend is set. |
| 143 |
""" |
| 144 |
if get_globally == "auto": |
| 145 |
get_globally = threading.get_ident() == _importing_thrid |
| 146 |
|
| 147 |
if get_globally: |
| 148 |
backend = _global_backend |
| 149 |
else: |
| 150 |
thrid = threading.get_ident() |
| 151 |
backend = _global_backends_threadaware.get(thrid, None) |
| 152 |
|
| 153 |
return backend |
| 154 |
|
| 155 |
|
| 156 |
def set_backend(like, set_globally="auto"): |
| 157 |
"""Set a default global backend. The argument ``like`` can be an explicit |
| 158 |
backend name or an ``array``. |
| 159 |
|
| 160 |
Parameters |
| 161 |
---------- |
| 162 |
like : str or array |
| 163 |
The backend to set. If an array, the backend of the array's class will |
| 164 |
be set. |
| 165 |
set_globally : {"auto", False, True}, optional |
| 166 |
Whether to set the backend globally or for the current thread: |
| 167 |
|
| 168 |
- True: set the backend globally. |
| 169 |
- False: set the backend for the current thread. |
| 170 |
- "auto": set the backend globally if this thread is the thread that |
| 171 |
imported autoray. Otherwise set the backend for the current thread. |
| 172 |
|
| 173 |
Only one thread should ever call this function with |
| 174 |
``set_globally=True``, (by default this is importing thread). |
| 175 |
""" |
| 176 |
global _global_backend |
| 177 |
global _infer_auto |
| 178 |
global _inferrer_global |
| 179 |
|
| 180 |
if like is None: |
| 181 |
backend = None |
| 182 |
inferrer = _default_infer_from_sig |
| 183 |
elif isinstance(like, str): |
| 184 |
backend = like |
| 185 |
inferrer = functools.partial(_always_the_same, backend=backend) |
| 186 |
else: |
| 187 |
backend = _infer_class_backend_cached(like.__class__) |
| 188 |
inferrer = functools.partial(_always_the_same, backend=backend) |
| 189 |
|
| 190 |
if set_globally == "auto": |
| 191 |
set_globally = threading.get_ident() == _importing_thrid |
| 192 |
|
| 193 |
if set_globally: |
| 194 |
_global_backend = backend |
| 195 |
_inferrer_global = inferrer |
| 196 |
if not _inferrers_threadaware: |
| 197 |
# only revert the actual function if no subthread backends set |
| 198 |
_infer_auto = inferrer |
| 199 |
else: |
| 200 |
thrid = threading.get_ident() |
| 201 |
_backend_lock.acquire() |
| 202 |
if backend is None: |
| 203 |
_global_backends_threadaware.pop(thrid) |
| 204 |
_inferrers_threadaware.pop(thrid) |
| 205 |
else: |
| 206 |
_global_backends_threadaware[thrid] = backend |
| 207 |
_inferrers_threadaware[thrid] = inferrer |
| 208 |
|
| 209 |
if _inferrers_threadaware: |
| 210 |
# a subthread backend has been set, so we need to be thread aware |
| 211 |
_infer_auto = _default_infer_from_sig_threadaware |
| 212 |
else: |
| 213 |
# no subthread backend has been set anymore, so we can ignore |
| 214 |
# threads and just use the global inferrer |
| 215 |
_infer_auto = _inferrer_global |
| 216 |
_backend_lock.release() |
| 217 |
|
| 218 |
|
| 219 |
@contextlib.contextmanager |
| 220 |
def backend_like(like, set_globally="auto"): |
| 221 |
"""Context manager for setting a default backend. The argument ``like`` can |
| 222 |
be an explicit backend name or an ``array`` to infer it from. |
| 223 |
|
| 224 |
Parameters |
| 225 |
---------- |
| 226 |
like : str or array |
| 227 |
The backend to set. If an array, the backend of the array's class will |
| 228 |
be set. |
| 229 |
set_globally : {"auto", False, True}, optional |
| 230 |
Whether to set the backend globally or for the current thread: |
| 231 |
|
| 232 |
- True: set the backend globally. |
| 233 |
- False: set the backend for the current thread. |
| 234 |
- "auto": set the backend globally if this thread is the thread that |
| 235 |
imported autoray. Otherwise set the backend for the current thread. |
| 236 |
|
| 237 |
Only one thread should ever call this function with |
| 238 |
``set_globally=True``, (by default this is importing thread). |
| 239 |
""" |
| 240 |
if set_globally == "auto": |
| 241 |
set_globally = threading.get_ident() == _importing_thrid |
| 242 |
|
| 243 |
old_backend = get_backend(get_globally=set_globally) |
| 244 |
try: |
| 245 |
set_backend(like, set_globally) |
| 246 |
yield |
| 247 |
finally: |
| 248 |
set_backend(old_backend, set_globally) |
| 249 |
|
| 250 |
|
| 251 |
_CUSTOM_BACKENDS = {} |
| 252 |
|
| 253 |
|
| 254 |
def register_backend(cls, name): |
| 255 |
"""Register the name (and by default the module or submodule) of a custom |
| 256 |
array class. |
| 257 |
|
| 258 |
Parameters |
| 259 |
---------- |
| 260 |
cls : type |
| 261 |
The array class itself. |
| 262 |
name : str |
| 263 |
The name of the backend that should be used for this class. By default |
| 264 |
this wil be assumed to be the location of the relevant functions for |
| 265 |
this class, but this can be overridden. |
| 266 |
""" |
| 267 |
if not isinstance(cls, type): |
| 268 |
raise TypeError("The array class itself should be supplied.") |
| 269 |
|
| 270 |
global _CUSTOM_BACKENDS |
| 271 |
_CUSTOM_BACKENDS[cls] = name |
| 272 |
|
| 273 |
|
| 274 |
@functools.lru_cache(None) |
| 275 |
def _infer_class_backend_cached(cls): |
| 276 |
try: |
| 277 |
import numpy as _numpy |
| 278 |
|
| 279 |
if issubclass(cls, _numpy.ndarray): |
| 280 |
return "numpy" |
| 281 |
except ImportError: |
| 282 |
# numpy not installed |
| 283 |
pass |
| 284 |
|
| 285 |
if cls in _CUSTOM_BACKENDS: |
| 286 |
return _CUSTOM_BACKENDS[cls] |
| 287 |
|
| 288 |
lib = cls.__module__.split(".")[0] |
| 289 |
|
| 290 |
# check if lib should mapped entirely to another lib |
| 291 |
backend = _BACKEND_ALIASES.get(lib, lib) |
| 292 |
|
| 293 |
return backend |
| 294 |
|
| 295 |
|
| 296 |
def infer_backend(array): |
| 297 |
"""Get the name of the library that defined the class of ``array`` - unless |
| 298 |
``array`` is directly a subclass of ``numpy.ndarray``, in which case assume |
| 299 |
``numpy`` is the desired backend. |
| 300 |
""" |
| 301 |
return _infer_class_backend_cached(array.__class__) |
| 302 |
|
| 303 |
|
| 304 |
multi_class_priorities = { |
| 305 |
"builtins": -2, |
| 306 |
"numpy": -1, |
| 307 |
"autoray.lazy": 1, |
| 308 |
} |
| 309 |
|
| 310 |
|
| 311 |
@functools.lru_cache(None) |
| 312 |
def _infer_class_backend_multi_cached(classes): |
| 313 |
return max( |
| 314 |
map(_infer_class_backend_cached, classes), |
| 315 |
key=lambda n: multi_class_priorities.get(n, 0), |
| 316 |
) |
| 317 |
|
| 318 |
|
| 319 |
def infer_backend_multi(*arrays): |
| 320 |
"""Infer which backend should be used for a function that takes multiple |
| 321 |
arguments. This assigns a priority to each backend, and returns the backend |
| 322 |
with the highest priority. By default, the priority is: |
| 323 |
|
| 324 |
- ``builtins``: -2 |
| 325 |
- ``numpy``: -1 |
| 326 |
- other backends: 0 |
| 327 |
- ``autoray.lazy``: 1 |
| 328 |
|
| 329 |
I.e. when mixing with ``numpy``, other array libraries are preferred, when |
| 330 |
mixing with ``autoray.lazy``, ``autoray.lazy`` is preferred. This has quite |
| 331 |
low overhead due to caching. |
| 332 |
""" |
| 333 |
return _infer_class_backend_multi_cached( |
| 334 |
tuple(array.__class__ for array in arrays) |
| 335 |
) |
| 336 |
|
| 337 |
|
| 338 |
# the set of functions that create new arrays, with `dtype` and possibly |
| 339 |
# `device` kwargs, that should be inferred from the like argument |
| 340 |
_CREATION_ROUTINES = { |
| 341 |
"empty", |
| 342 |
"eye", |
| 343 |
"full", |
| 344 |
"identity", |
| 345 |
"ones", |
| 346 |
"zeros", |
| 347 |
# TODO: should these be included? |
| 348 |
# "arange", |
| 349 |
# "geomspace", |
| 350 |
# "linspace", |
| 351 |
# "logspace", |
| 352 |
} |
| 353 |
|
| 354 |
# cache for whether backends have a device attribute |
| 355 |
_CREATION_INJECT = {} |
| 356 |
|
| 357 |
|
| 358 |
def register_creation_routine( |
| 359 |
backend, fn, inject_dtype=True, inject_device=False |
| 360 |
): |
| 361 |
"""Register a function that creates a new array, with `dtype` and possibly |
| 362 |
`device` kwargs, that should be inferred from the like argument. This is |
| 363 |
not necessary for array creation routines that don't accept either. |
| 364 |
|
| 365 |
Parameters |
| 366 |
---------- |
| 367 |
backend : str |
| 368 |
The backend to register the function for. |
| 369 |
fn : str |
| 370 |
The name of the function to register. |
| 371 |
inject_dtype : bool, optional |
| 372 |
Whether to inject a `dtype` argument based on the `like` argument. |
| 373 |
inject_device : bool, optional |
| 374 |
Whether to inject a `device` argument based on the `like` argument. |
| 375 |
""" |
| 376 |
_CREATION_INJECT[backend, fn] = (inject_dtype, inject_device) |
| 377 |
|
| 378 |
|
| 379 |
def _choose_backend(fn, args, kwargs, like=None): |
| 380 |
"""Private function to choose a backend based on function name and |
| 381 |
signature, which passes args and kwargs by reference for performance and |
| 382 |
also to allow injection of dtype and device arguments for array creation |
| 383 |
routines. |
| 384 |
""" |
| 385 |
if like is None: |
| 386 |
# infer from function call (or global backend) |
| 387 |
return _infer_auto(fn, args, kwargs) |
| 388 |
elif isinstance(like, str): |
| 389 |
# explicit backend |
| 390 |
return like |
| 391 |
else: |
| 392 |
# explicit example array |
| 393 |
backend = _infer_class_backend_cached(like.__class__) |
| 394 |
|
| 395 |
# check if we should set some extra defaults based on the example array |
| 396 |
if fn in _CREATION_ROUTINES: |
| 397 |
try: |
| 398 |
inject_dtype, inject_device = _CREATION_INJECT[backend, fn] |
| 399 |
except KeyError: |
| 400 |
# default to just dtype (e.g. for numpy) |
| 401 |
inject_dtype = True |
| 402 |
inject_device = False |
| 403 |
_CREATION_INJECT[backend, fn] = (inject_dtype, inject_device) |
| 404 |
|
| 405 |
if inject_dtype: |
| 406 |
kwargs.setdefault("dtype", getattr(like, "dtype", type(like))) |
| 407 |
if inject_device: |
| 408 |
kwargs.setdefault("device", like.device) |
| 409 |
|
| 410 |
return backend |
| 411 |
|
| 412 |
|
| 413 |
def choose_backend(fn, *args, like=None, **kwargs): |
| 414 |
"""Choose a backend based on function name, arguments, and the ``like`` |
| 415 |
keyword argument. The default, if ``like`` is not specified, is to infer |
| 416 |
the backend from the function call, the default of which is simply to use |
| 417 |
the first argument, if no custom dispatcher is found. Otherwise the |
| 418 |
backend is chosen based on the ``like`` argument - which can be an explicit |
| 419 |
backend name or an arbitrary object. |
| 420 |
""" |
| 421 |
return _choose_backend(fn, args, kwargs, like=like) |
| 422 |
|
| 423 |
|
| 424 |
# ------------------- importing and caching the function -------------------- # |
| 425 |
|
| 426 |
# lookup for mapping entire lib to another |
| 427 |
_BACKEND_ALIASES = {} |
| 428 |
|
| 429 |
# global (non function specific) aliases |
| 430 |
_MODULE_ALIASES = {} |
| 431 |
|
| 432 |
# lookup for when functions are elsewhere than the expected location |
| 433 |
_SUBMODULE_ALIASES = {} |
| 434 |
|
| 435 |
# lookup for when functions are simply called something else |
| 436 |
_FUNC_ALIASES = {} |
| 437 |
|
| 438 |
# custom wrappers for when functions don't just have different location or |
| 439 |
# name. For example, when kwargs need to be translated or results modified |
| 440 |
_CUSTOM_WRAPPERS = {} |
| 441 |
|
| 442 |
# actual cache of funtions to use - this is populated lazily and can be used |
| 443 |
# to directly set an implementation of a function for a specific backend |
| 444 |
_FUNCS = {} |
| 445 |
|
| 446 |
# these are functions where a default implementation can be constructed |
| 447 |
# (composed of other functions), but this is only done lazily |
| 448 |
_COMPOSED_FUNCTION_GENERATORS = {} |
| 449 |
|
| 450 |
|
| 451 |
def import_lib_fn(backend, fn): |
| 452 |
# first check explicitly composed functions -> if the function hasn't been |
| 453 |
# called directly yet, it won't have been loaded into the cache, and needs |
| 454 |
# generating before e.g. the ``do`` verrsion will work |
| 455 |
if fn in _COMPOSED_FUNCTION_GENERATORS: |
| 456 |
return _COMPOSED_FUNCTION_GENERATORS[fn](backend) |
| 457 |
|
| 458 |
try: |
| 459 |
# submodule where function is found for backend, |
| 460 |
# e.g. ['tensorflow', trace'] -> 'tensorflow.linalg' |
| 461 |
try: |
| 462 |
full_location = _SUBMODULE_ALIASES[backend, fn] |
| 463 |
|
| 464 |
# if explicit submodule alias given, don't use prepended location |
| 465 |
# for example, ('torch', 'linalg.svd') -> torch.svd |
| 466 |
only_fn = fn.split(".")[-1] |
| 467 |
|
| 468 |
except KeyError: |
| 469 |
full_location = backend |
| 470 |
|
| 471 |
# move any prepended location into the full module path |
| 472 |
# e.g. 'fn=linalg.eigh' -> ['linalg', 'eigh'] |
| 473 |
split_fn = fn.split(".") |
| 474 |
full_location = ".".join([full_location] + split_fn[:-1]) |
| 475 |
only_fn = split_fn[-1] |
| 476 |
|
| 477 |
# try aliases for global (not function specific) modules and |
| 478 |
# submodules: |
| 479 |
# e.g. 'decimal' -> 'math' |
| 480 |
# e.g. 'cupy.scipy' -> 'cupyx.scipy' |
| 481 |
# we don't do this if the function location has been explicitly |
| 482 |
# give in _SUBMODULE_ALIASES, as that is already a full path |
| 483 |
for k, v in _MODULE_ALIASES.items(): |
| 484 |
if full_location[: len(k)] == k: |
| 485 |
full_location = full_location.replace(k, v, 1) |
| 486 |
break |
| 487 |
|
| 488 |
# cached lookup of custom name function might take |
| 489 |
# e.g. ['tensorflow', 'sum'] -> 'reduce_sum' |
| 490 |
fn_name = _FUNC_ALIASES.get((backend, fn), only_fn) |
| 491 |
|
| 492 |
# import the function into the cache |
| 493 |
try: |
| 494 |
lib = importlib.import_module(full_location) |
| 495 |
except ImportError: |
| 496 |
if "." in full_location: |
| 497 |
# sometimes libraries hack an attribute to look like submodule |
| 498 |
mod, *submods = full_location.split(".") |
| 499 |
lib = importlib.import_module(mod) |
| 500 |
# also need to handle nested submodules |
| 501 |
for submod in submods: |
| 502 |
lib = getattr(lib, submod) |
| 503 |
else: |
| 504 |
# failed to import library at all -> catch + raise ImportError |
| 505 |
raise AttributeError |
| 506 |
|
| 507 |
# check for a custom wrapper but default to identity |
| 508 |
wrapper = _CUSTOM_WRAPPERS.get((backend, fn), lambda fn: fn) |
| 509 |
|
| 510 |
# store the function! |
| 511 |
lib_fn = _FUNCS[backend, fn] = wrapper(getattr(lib, fn_name)) |
| 512 |
|
| 513 |
except AttributeError: |
| 514 |
# check if there is a backup function (e.g. for older library version) |
| 515 |
backend_alt = backend + "[alt]" |
| 516 |
if backend_alt in _MODULE_ALIASES: |
| 517 |
return import_lib_fn(backend_alt, fn) |
| 518 |
|
| 519 |
raise ImportError( |
| 520 |
f"autoray couldn't find function '{fn}' for " |
| 521 |
f"backend '{backend.replace('[alt]', '')}'." |
| 522 |
) |
| 523 |
|
| 524 |
return lib_fn |
| 525 |
|
| 526 |
|
| 527 |
def get_lib_fn(backend, fn): |
| 528 |
"""Cached retrieval of correct function for backend, all the logic for |
| 529 |
finding the correct funtion only runs the first time. |
| 530 |
|
| 531 |
Parameters |
| 532 |
---------- |
| 533 |
backend : str |
| 534 |
The module defining the array class to dispatch on. |
| 535 |
fn : str |
| 536 |
The function to retrieve. |
| 537 |
|
| 538 |
Returns |
| 539 |
------- |
| 540 |
callable |
| 541 |
""" |
| 542 |
|
| 543 |
try: |
| 544 |
lib_fn = _FUNCS[backend, fn] |
| 545 |
except KeyError: |
| 546 |
lib_fn = import_lib_fn(backend, fn) |
| 547 |
return lib_fn |
| 548 |
|
| 549 |
|
| 550 |
# --------------------------- register your own! ---------------------------- # |
| 551 |
|
| 552 |
|
| 553 |
def register_function(backend, name, fn, wrap=False): |
| 554 |
"""Directly provide your own function. |
| 555 |
|
| 556 |
Parameters |
| 557 |
---------- |
| 558 |
backend : str |
| 559 |
The name of the backend to register the function for. |
| 560 |
name : str |
| 561 |
Name of the function, e.g. `'sum'` or `'linalg.svd'`. |
| 562 |
fn : callable |
| 563 |
The function to register. |
| 564 |
wrap : bool, optional |
| 565 |
Whether to wrap the old function like ``fn(old_fn)`` rather than |
| 566 |
directly supply the entire new function. |
| 567 |
""" |
| 568 |
if wrap: |
| 569 |
old = get_lib_fn(backend, name) |
| 570 |
_FUNCS[backend, name] = fn(old) |
| 571 |
else: |
| 572 |
_FUNCS[backend, name] = fn |
| 573 |
|
| 574 |
|
| 575 |
# ------------------------------- tree utils -------------------------------- # |
| 576 |
|
| 577 |
TREE_MAP_REGISTRY = {} |
| 578 |
TREE_APPLY_REGISTRY = {} |
| 579 |
TREE_ITER_REGISTRY = {} |
| 580 |
|
| 581 |
|
| 582 |
def tree_register_container(cls, mapper, iterator, applier): |
| 583 |
"""Register a new container type for use with ``tree_map`` and |
| 584 |
``tree_apply``. |
| 585 |
|
| 586 |
Parameters |
| 587 |
---------- |
| 588 |
cls : type |
| 589 |
The container type to register. |
| 590 |
mapper : callable |
| 591 |
A function that takes ``f``, ``tree`` and ``is_leaf`` and returns a new |
| 592 |
tree of type ``cls`` with ``f`` applied to all leaves. |
| 593 |
applier : callable |
| 594 |
A function that takes ``f``, ``tree`` and ``is_leaf`` and applies ``f`` |
| 595 |
to all leaves in ``tree``. |
| 596 |
""" |
| 597 |
TREE_MAP_REGISTRY[cls] = mapper |
| 598 |
TREE_ITER_REGISTRY[cls] = iterator |
| 599 |
TREE_APPLY_REGISTRY[cls] = applier |
| 600 |
|
| 601 |
|
| 602 |
IS_CONTAINER_CACHE = {} |
| 603 |
|
| 604 |
|
| 605 |
def is_not_container(x): |
| 606 |
"""The default function to determine if an object is a leaf. This simply |
| 607 |
checks if the object is an instance of any of the registered container |
| 608 |
types. |
| 609 |
""" |
| 610 |
try: |
| 611 |
return IS_CONTAINER_CACHE[x.__class__] |
| 612 |
except KeyError: |
| 613 |
isleaf = not any(isinstance(x, cls) for cls in TREE_MAP_REGISTRY) |
| 614 |
IS_CONTAINER_CACHE[x.__class__] = isleaf |
| 615 |
return isleaf |
| 616 |
|
| 617 |
|
| 618 |
def is_array(x): |
| 619 |
"""An alternative leaf tester for addressing only arrays within trees.""" |
| 620 |
return hasattr(x, "shape") |
| 621 |
|
| 622 |
|
| 623 |
def identity(f, tree, is_leaf): |
| 624 |
return tree |
| 625 |
|
| 626 |
|
| 627 |
TREE_MAPPER_CACHE = {} |
| 628 |
|
| 629 |
|
| 630 |
def tree_map(f, tree, is_leaf=is_not_container): |
| 631 |
"""Map ``f`` over all leaves in ``tree``, returning a new pytree. |
| 632 |
|
| 633 |
Parameters |
| 634 |
---------- |
| 635 |
f : callable |
| 636 |
A function to apply to all leaves in ``tree``. |
| 637 |
tree : pytree |
| 638 |
A nested sequence of tuples, lists, dicts and other objects. |
| 639 |
is_leaf : callable |
| 640 |
A function to determine if an object is a leaf, ``f`` is only applied |
| 641 |
to objects for which ``is_leaf(x)`` returns ``True``. |
| 642 |
|
| 643 |
Returns |
| 644 |
------- |
| 645 |
pytree |
| 646 |
""" |
| 647 |
if is_leaf(tree): |
| 648 |
return f(tree) |
| 649 |
|
| 650 |
try: |
| 651 |
return TREE_MAPPER_CACHE[tree.__class__](f, tree, is_leaf) |
| 652 |
except KeyError: |
| 653 |
# reverse so later registered classes take precedence |
| 654 |
for cls, mapper in reversed(TREE_MAP_REGISTRY.items()): |
| 655 |
if isinstance(tree, cls): |
| 656 |
break |
| 657 |
else: |
| 658 |
# neither leaf nor container -> simply return it |
| 659 |
mapper = identity |
| 660 |
TREE_MAPPER_CACHE[tree.__class__] = mapper |
| 661 |
return mapper(f, tree, is_leaf) |
| 662 |
|
| 663 |
|
| 664 |
def empty(tree, is_leaf): |
| 665 |
return iter(()) |
| 666 |
|
| 667 |
|
| 668 |
TREE_ITER_CACHE = {} |
| 669 |
|
| 670 |
|
| 671 |
def tree_iter(tree, is_leaf=is_not_container): |
| 672 |
"""Iterate over all leaves in ``tree``. |
| 673 |
|
| 674 |
Parameters |
| 675 |
---------- |
| 676 |
f : callable |
| 677 |
A function to apply to all leaves in ``tree``. |
| 678 |
tree : pytree |
| 679 |
A nested sequence of tuples, lists, dicts and other objects. |
| 680 |
is_leaf : callable |
| 681 |
A function to determine if an object is a leaf, ``f`` is only applied |
| 682 |
to objects for which ``is_leaf(x)`` returns ``True``. |
| 683 |
""" |
| 684 |
if is_leaf(tree): |
| 685 |
yield tree |
| 686 |
return |
| 687 |
|
| 688 |
try: |
| 689 |
yield from TREE_ITER_CACHE[tree.__class__](tree, is_leaf) |
| 690 |
except KeyError: |
| 691 |
# reverse so later registered classes take precedence |
| 692 |
for cls, iterator in reversed(TREE_ITER_REGISTRY.items()): |
| 693 |
if isinstance(tree, cls): |
| 694 |
break |
| 695 |
else: |
| 696 |
# neither leaf nor container -> simply ignore it |
| 697 |
iterator = empty |
| 698 |
TREE_ITER_CACHE[tree.__class__] = iterator |
| 699 |
yield from iterator(tree, is_leaf) |
| 700 |
|
| 701 |
|
| 702 |
def nothing(f, tree, is_leaf): |
| 703 |
pass |
| 704 |
|
| 705 |
|
| 706 |
TREE_APPLIER_CACHE = {} |
| 707 |
|
| 708 |
|
| 709 |
def tree_apply(f, tree, is_leaf=is_not_container): |
| 710 |
"""Apply ``f`` to all leaves in ``tree``, no new pytree is built. |
| 711 |
|
| 712 |
Parameters |
| 713 |
---------- |
| 714 |
f : callable |
| 715 |
A function to apply to all leaves in ``tree``. |
| 716 |
tree : pytree |
| 717 |
A nested sequence of tuples, lists, dicts and other objects. |
| 718 |
is_leaf : callable |
| 719 |
A function to determine if an object is a leaf, ``f`` is only applied |
| 720 |
to objects for which ``is_leaf(x)`` returns ``True``. |
| 721 |
""" |
| 722 |
if is_leaf(tree): |
| 723 |
f(tree) |
| 724 |
return |
| 725 |
|
| 726 |
try: |
| 727 |
TREE_APPLIER_CACHE[tree.__class__](f, tree, is_leaf) |
| 728 |
except KeyError: |
| 729 |
# reverse so later registered classes take precedence |
| 730 |
for cls, applier in reversed(TREE_APPLY_REGISTRY.items()): |
| 731 |
if isinstance(tree, cls): |
| 732 |
break |
| 733 |
else: |
| 734 |
# neither leaf nor container -> simply ignore it |
| 735 |
applier = nothing |
| 736 |
TREE_APPLIER_CACHE[tree.__class__] = applier |
| 737 |
applier(f, tree, is_leaf) |
| 738 |
|
| 739 |
|
| 740 |
class Leaf: |
| 741 |
"""A singleton object to use as a placeholder in a pytree, for |
| 742 |
unflattening. |
| 743 |
""" |
| 744 |
|
| 745 |
__slots__ = () |
| 746 |
|
| 747 |
def __repr__(self): |
| 748 |
return "Leaf" |
| 749 |
|
| 750 |
|
| 751 |
LEAF = Leaf() |
| 752 |
|
| 753 |
|
| 754 |
def is_leaf_placeholder(x): |
| 755 |
# don't do `x is LEAF` to allow pickling / unpickling |
| 756 |
return x.__class__ is Leaf |
| 757 |
|
| 758 |
|
| 759 |
def tree_flatten(tree, is_leaf=is_not_container, get_ref=False): |
| 760 |
"""Flatten ``tree`` into a list of leaves. |
| 761 |
|
| 762 |
Parameters |
| 763 |
---------- |
| 764 |
tree : pytree |
| 765 |
A nested sequence of tuples, lists, dicts and other objects. |
| 766 |
is_leaf : callable |
| 767 |
A function to determine if an object is a leaf, only objects for which |
| 768 |
``is_leaf(x)`` returns ``True`` are returned in the flattened list. |
| 769 |
get_ref : bool |
| 770 |
If ``True``, a reference tree is also returned which can be used to |
| 771 |
reconstruct the original tree from a flattened list. |
| 772 |
|
| 773 |
Returns |
| 774 |
------- |
| 775 |
objs : list |
| 776 |
The flattened list of leaf objects. |
| 777 |
(ref_tree) : pytree |
| 778 |
If ``get_ref`` is ``True``, a reference tree, with leaves of ``Leaf``, |
| 779 |
is returned which can be used to reconstruct the original tree. |
| 780 |
""" |
| 781 |
objs = [] |
| 782 |
if get_ref: |
| 783 |
# return a new tree with Leaf leaves, as well as the flattened list |
| 784 |
|
| 785 |
def f(x): |
| 786 |
objs.append(x) |
| 787 |
return LEAF |
| 788 |
|
| 789 |
ref_tree = tree_map(f, tree, is_leaf) |
| 790 |
return objs, ref_tree |
| 791 |
else: |
| 792 |
tree_apply(objs.append, tree, is_leaf) |
| 793 |
return objs |
| 794 |
|
| 795 |
|
| 796 |
def tree_unflatten(objs, tree, is_leaf=is_leaf_placeholder): |
| 797 |
"""Unflatten ``objs`` into a pytree of the same structure as ``tree``. |
| 798 |
|
| 799 |
Parameters |
| 800 |
---------- |
| 801 |
objs : sequence |
| 802 |
A sequence of objects to be unflattened into a pytree. |
| 803 |
tree : pytree |
| 804 |
A nested sequence of tuples, lists, dicts and other objects, the objs |
| 805 |
will be inserted into a new pytree of the same structure. |
| 806 |
is_leaf : callable |
| 807 |
A function to determine if an object is a leaf, only objects for which |
| 808 |
``is_leaf(x)`` returns ``True`` will have the next item from ``objs`` |
| 809 |
inserted. By default checks for the ``Leaf`` object inserted by |
| 810 |
``tree_flatten(..., get_ref=True)``. |
| 811 |
|
| 812 |
Returns |
| 813 |
------- |
| 814 |
pytree |
| 815 |
""" |
| 816 |
objs = iter(objs) |
| 817 |
return tree_map(lambda _: next(objs), tree, is_leaf) |
| 818 |
|
| 819 |
|
| 820 |
def tree_map_tuple(f, tree, is_leaf): |
| 821 |
return tuple(tree_map(f, x, is_leaf) for x in tree) |
| 822 |
|
| 823 |
|
| 824 |
def tree_iter_tuple(tree, is_leaf): |
| 825 |
for x in tree: |
| 826 |
yield from tree_iter(x, is_leaf) |
| 827 |
|
| 828 |
|
| 829 |
def tree_apply_tuple(f, tree, is_leaf): |
| 830 |
for x in tree: |
| 831 |
tree_apply(f, x, is_leaf) |
| 832 |
|
| 833 |
|
| 834 |
tree_register_container( |
| 835 |
tuple, tree_map_tuple, tree_iter_tuple, tree_apply_tuple |
| 836 |
) |
| 837 |
|
| 838 |
|
| 839 |
def tree_map_list(f, tree, is_leaf): |
| 840 |
return [tree_map(f, x, is_leaf) for x in tree] |
| 841 |
|
| 842 |
|
| 843 |
def tree_iter_list(tree, is_leaf): |
| 844 |
for x in tree: |
| 845 |
yield from tree_iter(x, is_leaf) |
| 846 |
|
| 847 |
|
| 848 |
def tree_apply_list(f, tree, is_leaf): |
| 849 |
for x in tree: |
| 850 |
tree_apply(f, x, is_leaf) |
| 851 |
|
| 852 |
|
| 853 |
tree_register_container(list, tree_map_list, tree_iter_list, tree_apply_list) |
| 854 |
|
| 855 |
|
| 856 |
def tree_map_dict(f, tree, is_leaf): |
| 857 |
return {k: tree_map(f, v, is_leaf) for k, v in tree.items()} |
| 858 |
|
| 859 |
|
| 860 |
def tree_iter_dict(tree, is_leaf): |
| 861 |
for v in tree.values(): |
| 862 |
yield from tree_iter(v, is_leaf) |
| 863 |
|
| 864 |
|
| 865 |
def tree_apply_dict(f, tree, is_leaf): |
| 866 |
for v in tree.values(): |
| 867 |
tree_apply(f, v, is_leaf) |
| 868 |
|
| 869 |
|
| 870 |
tree_register_container(dict, tree_map_dict, tree_iter_dict, tree_apply_dict) |
| 871 |
|
| 872 |
|
| 873 |
# --------------------------- composed functions ---------------------------- # |
| 874 |
|
| 875 |
|
| 876 |
class Composed: |
| 877 |
"""Compose an ``autoray.do`` using function. See the main wrapper |
| 878 |
``compose``. |
| 879 |
""" |
| 880 |
|
| 881 |
def __init__(self, fn, name=None): |
| 882 |
self._default_fn = fn |
| 883 |
if name is None: |
| 884 |
name = fn.__name__ |
| 885 |
self._name = name |
| 886 |
self._supply_backend = "backend" in signature(fn).parameters |
| 887 |
|
| 888 |
# this registers the fact that when `get_lib_fn` is called, the |
| 889 |
# function can be created even if it doesn't exist for a specific |
| 890 |
# backend yet. |
| 891 |
_COMPOSED_FUNCTION_GENERATORS[self._name] = self.make_function |
| 892 |
|
| 893 |
def register(self, backend, fn=None): |
| 894 |
"""Register a different implementation for ``backend``.""" |
| 895 |
if fn is not None: |
| 896 |
register_function(backend, self._name, fn) |
| 897 |
else: |
| 898 |
# wrapper form |
| 899 |
def wrapper(fn): |
| 900 |
register_function(backend, self._name, fn) |
| 901 |
return fn |
| 902 |
|
| 903 |
return wrapper |
| 904 |
|
| 905 |
def make_function(self, backend): |
| 906 |
"""Make a new function for the specific ``backend``.""" |
| 907 |
if self._supply_backend: |
| 908 |
# make sure it inherits __name__ etc |
| 909 |
fn = functools.wraps(self._default_fn)( |
| 910 |
functools.partial(self._default_fn, backend=backend) |
| 911 |
) |
| 912 |
else: |
| 913 |
fn = self._default_fn |
| 914 |
self.register(backend, fn) |
| 915 |
return fn |
| 916 |
|
| 917 |
def __call__(self, *args, like=None, **kwargs): |
| 918 |
backend = _choose_backend(self._name, args, kwargs, like=like) |
| 919 |
# `get_lib_fn` will call `make_function` if the function doesn't exist |
| 920 |
fn = get_lib_fn(backend, self._name) |
| 921 |
return fn(*args, **kwargs) |
| 922 |
|
| 923 |
def __repr__(self): |
| 924 |
return f"Composed('{self._name}')" |
| 925 |
|
| 926 |
|
| 927 |
def compose(fn, *, name=None): |
| 928 |
"""Take a function consisting of multiple ``autoray.do`` calls and compose |
| 929 |
it into a new, single, named function, registered with ``autoray.do``. |
| 930 |
|
| 931 |
This creates a default implementation of this function for each new backend |
| 932 |
encountered without explicitly having to write each out, but also allows |
| 933 |
for specific implementations to be overridden for specific backends. |
| 934 |
|
| 935 |
If the function takes a ``backend`` argument, it will be supplied with the |
| 936 |
backend name, to save having to re-choose the backend. |
| 937 |
|
| 938 |
Specific implementations can be provided by calling the ``register`` method |
| 939 |
of the composed function, or it can itself be used like a decorator:: |
| 940 |
|
| 941 |
@compose |
| 942 |
def foo(x): |
| 943 |
... |
| 944 |
|
| 945 |
@foo.register("numpy") |
| 946 |
@numba.njit |
| 947 |
def foo_numba(x): |
| 948 |
... |
| 949 |
|
| 950 |
Parameters |
| 951 |
---------- |
| 952 |
fn : callable |
| 953 |
The funtion to compose, and its default implementation. |
| 954 |
name : str, optional |
| 955 |
The name of the composed function. If not provided, the name of the |
| 956 |
function will be used. |
| 957 |
""" |
| 958 |
if fn is None: |
| 959 |
return functools.partial(compose, name=name) |
| 960 |
return functools.wraps(fn)(Composed(fn, name)) |
| 961 |
|
| 962 |
|
| 963 |
# ---------------------- special top level functions ------------------------ # |
| 964 |
|
| 965 |
|
| 966 |
@compose |
| 967 |
def shape(x): |
| 968 |
"""Get the shape of an array as a tuple of int. This should be preferred |
| 969 |
to calling `x.shape` directly, as it: |
| 970 |
|
| 971 |
1. Allows customization (e.g. for torch and aesara which return |
| 972 |
different types for shape - use `@shape.register(backend)` to |
| 973 |
customize the behavior from this default implementation). |
| 974 |
2. Can be used on nested lists and tuples, without calling numpy. |
| 975 |
|
| 976 |
Parameters |
| 977 |
---------- |
| 978 |
x : array_like |
| 979 |
The array to get the shape of. It can be an arbitrary nested list or |
| 980 |
tuple of arrays and scalars, but is assumed not to be ragged. |
| 981 |
|
| 982 |
Returns |
| 983 |
------- |
| 984 |
shape : tuple of int |
| 985 |
The size of each dimension of the array. |
| 986 |
""" |
| 987 |
try: |
| 988 |
return x.shape |
| 989 |
except AttributeError: |
| 990 |
# want to handle builtins / nested stuff |
| 991 |
if isinstance(x, (list, tuple)): |
| 992 |
d = len(x) |
| 993 |
if d != 0: |
| 994 |
# NB: slightly different from np.shape, as we do not check for |
| 995 |
# ragged arrays, but that behavior is seemingly deprecated |
| 996 |
return (d,) + shape(x[0]) |
| 997 |
return (d,) |
| 998 |
return () |
| 999 |
|
| 1000 |
|
| 1001 |
@compose |
| 1002 |
def ndim(x): |
| 1003 |
"""Get the number of dimensions of an array. This should be preferred to |
| 1004 |
calling `x.ndim`, since not all backends implement that, and it can also be |
| 1005 |
called on nested lists and tuples. |
| 1006 |
|
| 1007 |
Parameters |
| 1008 |
---------- |
| 1009 |
x : array_like |
| 1010 |
The array to get the number of dimensions of. It can be an arbitrary |
| 1011 |
nested list or tuple of arrays and scalars. |
| 1012 |
|
| 1013 |
Returns |
| 1014 |
------- |
| 1015 |
ndim : int |
| 1016 |
""" |
| 1017 |
try: |
| 1018 |
return x.ndim |
| 1019 |
except AttributeError: |
| 1020 |
return len(shape(x)) |
| 1021 |
|
| 1022 |
|
| 1023 |
@compose |
| 1024 |
def size(x): |
| 1025 |
"""Get the size, or number of elements, of an array. This should be |
| 1026 |
preferred to calling `x.size`, since not all backends implement that, and |
| 1027 |
it can also be called on nested lists and tuples. |
| 1028 |
|
| 1029 |
Parameters |
| 1030 |
---------- |
| 1031 |
x : array_like |
| 1032 |
The array to get the size of. It can be an arbitrary nested list or |
| 1033 |
tuple of arrays and scalars. |
| 1034 |
|
| 1035 |
Returns |
| 1036 |
------- |
| 1037 |
size : int |
| 1038 |
""" |
| 1039 |
try: |
| 1040 |
return x.size |
| 1041 |
except AttributeError: |
| 1042 |
return math.prod(shape(x)) |
| 1043 |
|
| 1044 |
|
| 1045 |
def conj(x): |
| 1046 |
"""Array conjugate.""" |
| 1047 |
return do("conj", x) |
| 1048 |
|
| 1049 |
|
| 1050 |
def transpose(x, *args): |
| 1051 |
"""Array transpose.""" |
| 1052 |
return do("transpose", x, *args) |
| 1053 |
|
| 1054 |
|
| 1055 |
def dag(x): |
| 1056 |
"""Array Hermitian transpose.""" |
| 1057 |
try: |
| 1058 |
return x.H |
| 1059 |
except AttributeError: |
| 1060 |
backend = _infer_class_backend_cached(x.__class__) |
| 1061 |
return do("conj", do("transpose", x, like=backend), like=backend) |
| 1062 |
|
| 1063 |
|
| 1064 |
def real(x): |
| 1065 |
"""Array real part.""" |
| 1066 |
return do("real", x) |
| 1067 |
|
| 1068 |
|
| 1069 |
def imag(x): |
| 1070 |
"""Array imaginary part.""" |
| 1071 |
return do("imag", x) |
| 1072 |
|
| 1073 |
|
| 1074 |
def reshape(x, shape): |
| 1075 |
"""Array reshaped.""" |
| 1076 |
try: |
| 1077 |
return x.reshape(shape) |
| 1078 |
except AttributeError: |
| 1079 |
return do("reshape", x, shape) |
| 1080 |
|
| 1081 |
|
| 1082 |
def to_backend_dtype(dtype_name, like): |
| 1083 |
"""Turn string specifier ``dtype_name`` into dtype of backend ``like``.""" |
| 1084 |
if not isinstance(like, str): |
| 1085 |
like = _infer_class_backend_cached(like.__class__) |
| 1086 |
|
| 1087 |
try: |
| 1088 |
return get_lib_fn(like, dtype_name) |
| 1089 |
except ImportError: |
| 1090 |
return dtype_name |
| 1091 |
|
| 1092 |
|
| 1093 |
@compose |
| 1094 |
def get_dtype_name(x): |
| 1095 |
"""Find string specifier ``dtype_name`` of array ``x``.""" |
| 1096 |
dtype = x.dtype |
| 1097 |
try: |
| 1098 |
return dtype.name |
| 1099 |
except AttributeError: |
| 1100 |
return str(dtype) |
| 1101 |
|
| 1102 |
|
| 1103 |
_COMPLEX_DTYPES = {"complex64", "complex128"} |
| 1104 |
_DOUBLE_DTYPES = {"float64", "complex128"} |
| 1105 |
_DTYPE_MAP = { |
| 1106 |
(False, False): "float32", |
| 1107 |
(False, True): "float64", |
| 1108 |
(True, False): "complex64", |
| 1109 |
(True, True): "complex128", |
| 1110 |
} |
| 1111 |
|
| 1112 |
|
| 1113 |
def get_common_dtype(*arrays): |
| 1114 |
"""Compute the minimal dtype sufficient for ``arrays``.""" |
| 1115 |
dtypes = set(map(get_dtype_name, arrays)) |
| 1116 |
has_complex = not _COMPLEX_DTYPES.isdisjoint(dtypes) |
| 1117 |
has_double = not _DOUBLE_DTYPES.isdisjoint(dtypes) |
| 1118 |
return _DTYPE_MAP[has_complex, has_double] |
| 1119 |
|
| 1120 |
|
| 1121 |
def astype(x, dtype_name, **kwargs): |
| 1122 |
"""Cast array as type ``dtype_name`` - tries ``x.astype`` first.""" |
| 1123 |
dtype = to_backend_dtype(dtype_name, like=x) |
| 1124 |
try: |
| 1125 |
return x.astype(dtype, **kwargs) |
| 1126 |
except AttributeError: |
| 1127 |
return do("astype", x, dtype, **kwargs) |
| 1128 |
|
| 1129 |
|
| 1130 |
def to_numpy(x): |
| 1131 |
"""Get a numpy version of array ``x``.""" |
| 1132 |
return do("to_numpy", x) |
| 1133 |
|
| 1134 |
|
| 1135 |
# -------------------------- some common wrappers --------------------------- # |
| 1136 |
|
| 1137 |
|
| 1138 |
def svd_not_full_matrices_wrapper(fn): |
| 1139 |
@functools.wraps(fn) |
| 1140 |
def default_not_full_matrices(*args, **kwargs): |
| 1141 |
kwargs.setdefault("full_matrices", False) |
| 1142 |
return fn(*args, **kwargs) |
| 1143 |
|
| 1144 |
return default_not_full_matrices |
| 1145 |
|
| 1146 |
|
| 1147 |
def svd_sUV_to_UsVH_wrapper(fn): |
| 1148 |
@functools.wraps(fn) |
| 1149 |
def numpy_like(*args, **kwargs): |
| 1150 |
s, U, V = fn(*args, **kwargs) |
| 1151 |
return U, s, dag(V) |
| 1152 |
|
| 1153 |
return numpy_like |
| 1154 |
|
| 1155 |
|
| 1156 |
def svd_UsV_to_UsVH_wrapper(fn): |
| 1157 |
@functools.wraps(fn) |
| 1158 |
def numpy_like(*args, **kwargs): |
| 1159 |
U, s, V = fn(*args, **kwargs) |
| 1160 |
return U, s, dag(V) |
| 1161 |
|
| 1162 |
return numpy_like |
| 1163 |
|
| 1164 |
|
| 1165 |
def svd_manual_full_matrices_kwarg(fn): |
| 1166 |
@functools.wraps(fn) |
| 1167 |
def numpy_like(*args, full_matrices=False, **kwargs): |
| 1168 |
U, s, VH = fn(*args, **kwargs) |
| 1169 |
|
| 1170 |
if not full_matrices: |
| 1171 |
U, VH = U[:, : s.size], VH[: s.size, :] |
| 1172 |
|
| 1173 |
return U, s, VH |
| 1174 |
|
| 1175 |
return numpy_like |
| 1176 |
|
| 1177 |
|
| 1178 |
def qr_allow_fat(fn): |
| 1179 |
@functools.wraps(fn) |
| 1180 |
def numpy_like(a, **kwargs): |
| 1181 |
m, n = shape(a) |
| 1182 |
|
| 1183 |
if m >= n: |
| 1184 |
# square or thin |
| 1185 |
return fn(a, **kwargs) |
| 1186 |
|
| 1187 |
Q, R_sq = fn(a[:, :m]) |
| 1188 |
R_r = dag(Q) @ a[:, m:] |
| 1189 |
R = do("concatenate", (R_sq, R_r), axis=1, like=a) |
| 1190 |
|
| 1191 |
return Q, R |
| 1192 |
|
| 1193 |
return numpy_like |
| 1194 |
|
| 1195 |
|
| 1196 |
def tril_to_band_part(fn): |
| 1197 |
@functools.wraps(fn) |
| 1198 |
def numpy_like(x, k=0): |
| 1199 |
if k < 0: |
| 1200 |
raise ValueError( |
| 1201 |
"'k' must be positive to recreate 'numpy.tril' " |
| 1202 |
"behaviour with 'tensorflow.matrix_band_part'." |
| 1203 |
) |
| 1204 |
|
| 1205 |
return fn(x, -1, k) |
| 1206 |
|
| 1207 |
return numpy_like |
| 1208 |
|
| 1209 |
|
| 1210 |
def triu_to_band_part(fn): |
| 1211 |
@functools.wraps(fn) |
| 1212 |
def numpy_like(x, k=0): |
| 1213 |
if k > 0: |
| 1214 |
raise ValueError( |
| 1215 |
"'k' must be negative to recreate 'numpy.triu' " |
| 1216 |
"behaviour with 'tensorflow.matrix_band_part'." |
| 1217 |
) |
| 1218 |
|
| 1219 |
return fn(x, -k, -1) |
| 1220 |
|
| 1221 |
return numpy_like |
| 1222 |
|
| 1223 |
|
| 1224 |
def cholesky_lower(fn): |
| 1225 |
@functools.wraps(fn) |
| 1226 |
def cholesky_numpy_like(a): |
| 1227 |
return fn(a, lower=True) |
| 1228 |
|
| 1229 |
return cholesky_numpy_like |
| 1230 |
|
| 1231 |
|
| 1232 |
def binary_allow_1d_rhs_wrap(fn): |
| 1233 |
@functools.wraps(fn) |
| 1234 |
def allow_1d_rhs(a, b): |
| 1235 |
need_to_convert = ndim(a) != ndim(b) |
| 1236 |
if need_to_convert: |
| 1237 |
b = reshape(b, (*shape(b), 1)) |
| 1238 |
x = fn(a, b) |
| 1239 |
if need_to_convert: |
| 1240 |
x = reshape(x, shape(x)[:-1]) |
| 1241 |
return x |
| 1242 |
|
| 1243 |
return allow_1d_rhs |
| 1244 |
|
| 1245 |
|
| 1246 |
def scale_random_uniform_manually(fn): |
| 1247 |
@functools.wraps(fn) |
| 1248 |
def numpy_like(low=0.0, high=1.0, size=None, dtype=None, **kwargs): |
| 1249 |
if size is None: |
| 1250 |
size = () |
| 1251 |
|
| 1252 |
x = fn(size, **kwargs) |
| 1253 |
|
| 1254 |
if (low != 0.0) or (high != 1.0): |
| 1255 |
x = (high - low) * x + low |
| 1256 |
|
| 1257 |
if (dtype is not None) and get_dtype_name(x) != dtype: |
| 1258 |
x = astype(x, dtype) |
| 1259 |
return x |
| 1260 |
|
| 1261 |
return numpy_like |
| 1262 |
|
| 1263 |
|
| 1264 |
def scale_random_normal_manually(fn): |
| 1265 |
@functools.wraps(fn) |
| 1266 |
def numpy_like(loc=0.0, scale=1.0, size=None, dtype=None, **kwargs): |
| 1267 |
if size is None: |
| 1268 |
size = () |
| 1269 |
|
| 1270 |
x = fn(size, **kwargs) |
| 1271 |
|
| 1272 |
if (loc != 0.0) or (scale != 1.0): |
| 1273 |
x = scale * x + loc |
| 1274 |
|
| 1275 |
if (dtype is not None) and get_dtype_name(x) != dtype: |
| 1276 |
x = astype(x, dtype) |
| 1277 |
return x |
| 1278 |
|
| 1279 |
return numpy_like |
| 1280 |
|
| 1281 |
|
| 1282 |
def with_dtype_wrapper(fn): |
| 1283 |
"""Add ability to handle `dtype` keyword. |
| 1284 |
If not None, `dtype` should be specified as a string, otherwise conversion |
| 1285 |
will happen regardless. |
| 1286 |
""" |
| 1287 |
|
| 1288 |
@functools.wraps(fn) |
| 1289 |
def with_dtype(*args, dtype=None, **kwargs): |
| 1290 |
A = fn(*args, **kwargs) |
| 1291 |
if (dtype is not None) and (dtype != get_dtype_name(A)): |
| 1292 |
A = astype(A, dtype) |
| 1293 |
return A |
| 1294 |
|
| 1295 |
return with_dtype |
| 1296 |
|
| 1297 |
|
| 1298 |
def translate_wrapper(fn, translator): |
| 1299 |
"""Wrap a function to match the api of another according to a translation. |
| 1300 |
The ``translator`` entries in the form of an ordered dict should have |
| 1301 |
entries like: |
| 1302 |
|
| 1303 |
(desired_kwarg: (backend_kwarg, default_value)) |
| 1304 |
|
| 1305 |
with the order defining the args of the function. |
| 1306 |
""" |
| 1307 |
|
| 1308 |
@functools.wraps(fn) |
| 1309 |
def translated_function(*args, **kwargs): |
| 1310 |
new_kwargs = {} |
| 1311 |
translation = translator.copy() |
| 1312 |
|
| 1313 |
# convert args, pairing them off with kwargs |
| 1314 |
for arg_value in args: |
| 1315 |
new_arg_name = translation.popitem(last=False)[1][0] |
| 1316 |
new_kwargs[new_arg_name] = arg_value |
| 1317 |
|
| 1318 |
# convert kwargs - but only those in the translation |
| 1319 |
for key, value in kwargs.items(): |
| 1320 |
try: |
| 1321 |
new_kwargs[translation.pop(key)[0]] = value |
| 1322 |
except KeyError: |
| 1323 |
new_kwargs[key] = value |
| 1324 |
|
| 1325 |
# set remaining default kwargs |
| 1326 |
for opt in translation.values(): |
| 1327 |
if len(opt) == 2: |
| 1328 |
# backend_name, default_value |
| 1329 |
new_kwargs[opt[0]] = opt[1] |
| 1330 |
# else, no default value -> don't inject |
| 1331 |
|
| 1332 |
return fn(**new_kwargs) |
| 1333 |
|
| 1334 |
return translated_function |
| 1335 |
|
| 1336 |
|
| 1337 |
def make_translator(t): |
| 1338 |
return functools.partial(translate_wrapper, translator=OrderedDict(t)) |
| 1339 |
|
| 1340 |
|
| 1341 |
def complex_add_re_im(re, im): |
| 1342 |
return re + 1j * im |
| 1343 |
|
| 1344 |
|
| 1345 |
def allclose(x, y, rtol=1e-05, atol=1e-08): |
| 1346 |
return do("all", do("abs", x - y) <= atol + rtol * do("abs", y)) |
| 1347 |
|
| 1348 |
|
| 1349 |
def _handle_size_to_shape(size=None): |
| 1350 |
if size is None: |
| 1351 |
return () |
| 1352 |
try: |
| 1353 |
return tuple(size) |
| 1354 |
except TypeError: |
| 1355 |
return (size,) |
| 1356 |
|
| 1357 |
|
| 1358 |
# ----------------------------- Custom dispatchers -------------------------- # |
| 1359 |
|
| 1360 |
|
| 1361 |
def wrap_args_kwargs_from_raw(fn): |
| 1362 |
"""Take a function with signature ``(*args, **kwargs)`` and wrap it to |
| 1363 |
accept a single tuple of args and a dict of kwargs. |
| 1364 |
""" |
| 1365 |
|
| 1366 |
@functools.wraps(fn) |
| 1367 |
def wrapped(args, kwargs): |
| 1368 |
return fn(*args, **kwargs) |
| 1369 |
|
| 1370 |
return wrapped |
| 1371 |
|
| 1372 |
|
| 1373 |
def register_dispatch(fun, dispatcher, raw_signature=True): |
| 1374 |
"""Register a new dispatcher, a function that takes the arguments and |
| 1375 |
keyword arguments of a function and returns the backend to use, when the |
| 1376 |
backend is not explicitly given. |
| 1377 |
|
| 1378 |
This is useful in case the backend to be used by a function cannot be |
| 1379 |
inferred from the first argument. |
| 1380 |
|
| 1381 |
Parameters |
| 1382 |
---------- |
| 1383 |
fun : str |
| 1384 |
The name of the function to register the dispatcher for. |
| 1385 |
dispatcher : callable |
| 1386 |
The dispatcher function to use. This should take the arguments and |
| 1387 |
keyword arguments of the function and return the backend to use. |
| 1388 |
raw_signature : bool, optional |
| 1389 |
The ``dispatcher`` has signature ``(*args, **kwargs)`` if ``True``, |
| 1390 |
otherwise it has signature ``(args, kwargs)``. |
| 1391 |
""" |
| 1392 |
if raw_signature: |
| 1393 |
dispatcher = wrap_args_kwargs_from_raw(dispatcher) |
| 1394 |
|
| 1395 |
_DISPATCHERS[fun] = dispatcher |
| 1396 |
|
| 1397 |
|
| 1398 |
def default_dispatcher(args, kwargs): |
| 1399 |
"""Try to infer backend from first argument passed to function.""" |
| 1400 |
try: |
| 1401 |
return _infer_class_backend_cached(args[0].__class__) |
| 1402 |
except IndexError: |
| 1403 |
raise TypeError("No args to infer backend from.") |
| 1404 |
|
| 1405 |
|
| 1406 |
# lookup of custom dispatcher methods, for cases when backend cannot be |
| 1407 |
# inferred accurately from first argument. |
| 1408 |
_DISPATCHERS = defaultdict(lambda: default_dispatcher) |
| 1409 |
|
| 1410 |
|
| 1411 |
def join_array_dispatcher(args, kwargs): |
| 1412 |
"""Dispatcher for functions where first argument is a sequence.""" |
| 1413 |
try: |
| 1414 |
return _infer_class_backend_cached(args[0][0].__class__) |
| 1415 |
except (TypeError, ValueError): |
| 1416 |
# user passed an empty sequence, or something non-iterable |
| 1417 |
# try to infer backend from first argument as fallback |
| 1418 |
return _infer_class_backend_cached(args[0].__class__) |
| 1419 |
|
| 1420 |
|
| 1421 |
# List of functions listed in numpy API as array joining operations |
| 1422 |
register_dispatch("concatenate", join_array_dispatcher, raw_signature=False) |
| 1423 |
register_dispatch("stack", join_array_dispatcher, raw_signature=False) |
| 1424 |
register_dispatch("block", join_array_dispatcher, raw_signature=False) |
| 1425 |
register_dispatch("vstack", join_array_dispatcher, raw_signature=False) |
| 1426 |
register_dispatch("hstack", join_array_dispatcher, raw_signature=False) |
| 1427 |
register_dispatch("dstack", join_array_dispatcher, raw_signature=False) |
| 1428 |
register_dispatch("column_stack", join_array_dispatcher, raw_signature=False) |
| 1429 |
register_dispatch("row_stack", join_array_dispatcher, raw_signature=False) |
| 1430 |
|
| 1431 |
|
| 1432 |
def einsum_dispatcher(args, kwargs): |
| 1433 |
"""Dispatcher for handling einsum. |
| 1434 |
|
| 1435 |
einsum can be called with a str equation as the first argument, or with |
| 1436 |
'interleaved' inputs. This dispatcher handles both cases and also takes |
| 1437 |
into account all arrays. |
| 1438 |
""" |
| 1439 |
return infer_backend_multi(*args) |
| 1440 |
|
| 1441 |
|
| 1442 |
register_dispatch("einsum", einsum_dispatcher, raw_signature=False) |
| 1443 |
|
| 1444 |
|
| 1445 |
def binary_dispatcher(args, kwargs): |
| 1446 |
"""There are cases when we want to take into account both backends of two |
| 1447 |
arguments, e.g. a lazy variable and a constant array. |
| 1448 |
""" |
| 1449 |
return infer_backend_multi(*args[:2]) |
| 1450 |
|
| 1451 |
|
| 1452 |
register_dispatch("tensordot", binary_dispatcher, raw_signature=False) |
| 1453 |
register_dispatch("matmul", binary_dispatcher, raw_signature=False) |
| 1454 |
register_dispatch("multiply", binary_dispatcher, raw_signature=False) |
| 1455 |
register_dispatch("divide", binary_dispatcher, raw_signature=False) |
| 1456 |
register_dispatch("true_divide", binary_dispatcher, raw_signature=False) |
| 1457 |
register_dispatch("add", binary_dispatcher, raw_signature=False) |
| 1458 |
register_dispatch("subtract", binary_dispatcher, raw_signature=False) |
| 1459 |
|
| 1460 |
# TODO: register other binary functions? |
| 1461 |
|
| 1462 |
# --------------- object to act as drop-in replace for numpy ---------------- # |
| 1463 |
|
| 1464 |
|
| 1465 |
class InjectDtypeDevice: |
| 1466 |
"""Wrapper that injects defaultdtype and device arguments into a function""" |
| 1467 |
|
| 1468 |
__slots__ = ("_fn", "_device", "_dtype") |
| 1469 |
|
| 1470 |
def __init__(self, fn, device=None, dtype=None): |
| 1471 |
self._fn = fn |
| 1472 |
self._device = device |
| 1473 |
self._dtype = dtype |
| 1474 |
|
| 1475 |
def __call__(self, *args, **kwargs): |
| 1476 |
if self._device is not None and "device" not in kwargs: |
| 1477 |
kwargs["device"] = self._device |
| 1478 |
if self._dtype is not None and "dtype" not in kwargs: |
| 1479 |
kwargs["dtype"] = self._dtype |
| 1480 |
return self._fn(*args, **kwargs) |
| 1481 |
|
| 1482 |
def __repr__(self): |
| 1483 |
return ( |
| 1484 |
f"InjectDtypeDevice(fn={self._fn}, " |
| 1485 |
f"device={self._device}, " |
| 1486 |
f"dtype={self._dtype})" |
| 1487 |
) |
| 1488 |
|
| 1489 |
|
| 1490 |
class AutoNamespace: |
| 1491 |
"""Mimics a namespace, optionally for a specific backend, device, and |
| 1492 |
dtype, caching the lookup of functions, and injecting default device and |
| 1493 |
dtype arguments for certain creation routines. |
| 1494 |
|
| 1495 |
Parameters |
| 1496 |
---------- |
| 1497 |
like : array_like, str, or None |
| 1498 |
The backend to use, or an object to infer the backend from. If None, |
| 1499 |
the default behavior is to use `autoray.do` and auto dispatch backend |
| 1500 |
at function call time. If given, the functions are cached at first |
| 1501 |
call. |
| 1502 |
device : str, optional |
| 1503 |
The device to use for the backend. If None, it will be inferred from |
| 1504 |
the `like` paramater is that is array-like or set to None. |
| 1505 |
dtype : str, optional |
| 1506 |
The dtype to use for the backend. If None, it will be inferred from |
| 1507 |
the `like` parameter if that is array-like or set to None. |
| 1508 |
submodule : str, optional |
| 1509 |
This is used internally when nesting attribute lookups, e.g. |
| 1510 |
`xp.random.normal`, `xp.linalg.eigh`. |
| 1511 |
""" |
| 1512 |
|
| 1513 |
def __init__( |
| 1514 |
self, |
| 1515 |
like=None, |
| 1516 |
device=None, |
| 1517 |
dtype=None, |
| 1518 |
submodule=None, |
| 1519 |
): |
| 1520 |
if like is None: |
| 1521 |
# use autoray.do and auto dispatch |
| 1522 |
self._backend = None |
| 1523 |
elif isinstance(like, str): |
| 1524 |
# commit to a specific given backend |
| 1525 |
self._backend = like |
| 1526 |
else: |
| 1527 |
# commit to a specific backend inferred from array-like |
| 1528 |
self._backend = _infer_class_backend_cached(like.__class__) |
| 1529 |
|
| 1530 |
if device is None: |
| 1531 |
if like is not None: |
| 1532 |
if hasattr(like, "device"): |
| 1533 |
device = like.device |
| 1534 |
self._device = device |
| 1535 |
|
| 1536 |
if dtype is None: |
| 1537 |
if like is not None: |
| 1538 |
if hasattr(like, "dtype"): |
| 1539 |
dtype = like.dtype |
| 1540 |
self._dtype = dtype |
| 1541 |
self._submodule = submodule |
| 1542 |
|
| 1543 |
def _get_submodule(self, name): |
| 1544 |
new = object.__new__(type(self)) |
| 1545 |
new._backend = self._backend |
| 1546 |
new._device = self._device |
| 1547 |
new._dtype = self._dtype |
| 1548 |
new._submodule = name |
| 1549 |
return new |
| 1550 |
|
| 1551 |
def _get_fn(self, name): |
| 1552 |
if self._submodule is not None: |
| 1553 |
# prepend the submodule name |
| 1554 |
name = f"{self._submodule}.{name}" |
| 1555 |
|
| 1556 |
if name in ("random", "linalg"): |
| 1557 |
# note that other submodules can be accessed, these are just the |
| 1558 |
# ones with functions that we possibly want to translate |
| 1559 |
return self._get_submodule(name) |
| 1560 |
|
| 1561 |
if self._backend is None: |
| 1562 |
# use autoray.do and auto dispatch |
| 1563 |
return functools.partial(do, name) |
| 1564 |
|
| 1565 |
fn = get_lib_fn(self._backend, name) |
| 1566 |
|
| 1567 |
# possibly wrap for dtype and device injection |
| 1568 |
if name in _CREATION_ROUTINES: |
| 1569 |
inject_dtype, inject_device = _CREATION_INJECT.get( |
| 1570 |
(self._backend, fn), (True, False) |
| 1571 |
) |
| 1572 |
|
| 1573 |
if not inject_dtype: |
| 1574 |
# this is not a function accepts dtype |
| 1575 |
dtype_to_inject = None |
| 1576 |
else: |
| 1577 |
dtype_to_inject = self._dtype |
| 1578 |
|
| 1579 |
if not inject_device: |
| 1580 |
# this is not a function accepts device |
| 1581 |
device_to_inject = None |
| 1582 |
else: |
| 1583 |
device_to_inject = self._device |
| 1584 |
|
| 1585 |
if (dtype_to_inject is not None) or (device_to_inject is not None): |
| 1586 |
# only wrap if we actually inject something |
| 1587 |
fn = InjectDtypeDevice(fn, device_to_inject, dtype_to_inject) |
| 1588 |
|
| 1589 |
return fn |
| 1590 |
|
| 1591 |
def __getattribute__(self, name): |
| 1592 |
try: |
| 1593 |
return object.__getattribute__(self, name) |
| 1594 |
except AttributeError: |
| 1595 |
x = self._get_fn(name) |
| 1596 |
super().__getattribute__("__dict__")[name] = x |
| 1597 |
return x |
| 1598 |
|
| 1599 |
def __repr__(self): |
| 1600 |
return ( |
| 1601 |
f"{self.__class__.__name__}(" |
| 1602 |
f"backend={self._backend}, " |
| 1603 |
f"device={self._device}, " |
| 1604 |
f"dtype={self._dtype}, " |
| 1605 |
f"submodule={self._submodule}" |
| 1606 |
")" |
| 1607 |
) |
| 1608 |
|
| 1609 |
|
| 1610 |
numpy = AutoNamespace() |
| 1611 |
|
| 1612 |
|
| 1613 |
def get_namespace(like=None, device=None, dtype=None, submodule=None): |
| 1614 |
"""Get an automatic namespace object. |
| 1615 |
|
| 1616 |
If `like` is None, the namespace essentially provides an alternative syntax |
| 1617 |
to `do`, dispatching each function at calltime, and allowing the backend |
| 1618 |
and function implementations to be dynamically updated. |
| 1619 |
|
| 1620 |
If `like` is supplied however, the backend is eagerly dispatched and |
| 1621 |
functions are loaded and cached specifically for that backend. In this |
| 1622 |
case, default `device` and `dtype` can also be specified for various array |
| 1623 |
creation routines, or if `like` is an array, inferred from that. |
| 1624 |
|
| 1625 |
Parameters |
| 1626 |
---------- |
| 1627 |
like : array-like, str or None, optional |
| 1628 |
An array-like object to dispatch on, an explicit backend name, or None. |
| 1629 |
device : str or None, optional |
| 1630 |
The device to use for array creation, or None to infer from `like`. |
| 1631 |
dtype : str or None, optional |
| 1632 |
The data type to use for array creation, or None to infer from `like`. |
| 1633 |
|
| 1634 |
Returns |
| 1635 |
------- |
| 1636 |
AutoNamespace |
| 1637 |
An automatic namespace object. |
| 1638 |
""" |
| 1639 |
return AutoNamespace( |
| 1640 |
like=like, |
| 1641 |
device=device, |
| 1642 |
dtype=dtype, |
| 1643 |
submodule=submodule, |
| 1644 |
) |
| 1645 |
|
| 1646 |
|
| 1647 |
# --------------------------------------------------------------------------- # |
| 1648 |
# specific functions # |
| 1649 |
# --------------------------------------------------------------------------- # |
| 1650 |
|
| 1651 |
# ------------------------------ standard-lib ------------------------------- # |
| 1652 |
|
| 1653 |
_MODULE_ALIASES["decimal"] = "math" |
| 1654 |
_MODULE_ALIASES["builtins"] = "numpy" |
| 1655 |
|
| 1656 |
|
| 1657 |
_builtin_dtype_lookup = { |
| 1658 |
int: "int64", |
| 1659 |
float: "float64", |
| 1660 |
complex: "complex128", |
| 1661 |
} |
| 1662 |
|
| 1663 |
|
| 1664 |
@get_dtype_name.register("builtins") |
| 1665 |
def builtins_get_dtype_name(x): |
| 1666 |
return _builtin_dtype_lookup[x.__class__] |
| 1667 |
|
| 1668 |
|
| 1669 |
_FUNCS["builtins", "complex"] = complex |
| 1670 |
|
| 1671 |
# ---------------------------------- numpy ---------------------------------- # |
| 1672 |
|
| 1673 |
|
| 1674 |
def numpy_to_numpy(x): |
| 1675 |
return do("asarray", x, like="numpy") |
| 1676 |
|
| 1677 |
|
| 1678 |
_MODULE_ALIASES["numpy.scipy"] = "scipy" |
| 1679 |
_FUNCS["numpy", "to_numpy"] = numpy_to_numpy |
| 1680 |
_FUNCS["numpy", "complex"] = complex_add_re_im |
| 1681 |
_FUNCS["builtins", "to_numpy"] = numpy_to_numpy |
| 1682 |
_SUBMODULE_ALIASES["numpy", "linalg.lu"] = "scipy.linalg" |
| 1683 |
_SUBMODULE_ALIASES["numpy", "linalg.expm"] = "scipy.linalg" |
| 1684 |
_CUSTOM_WRAPPERS["numpy", "linalg.svd"] = svd_not_full_matrices_wrapper |
| 1685 |
_CUSTOM_WRAPPERS["numpy", "random.normal"] = with_dtype_wrapper |
| 1686 |
_CUSTOM_WRAPPERS["numpy", "random.uniform"] = with_dtype_wrapper |
| 1687 |
|
| 1688 |
# ---------------------------------- cupy ----------------------------------- # |
| 1689 |
|
| 1690 |
|
| 1691 |
def cupy_to_numpy(x): # pragma: no cover |
| 1692 |
return x.get() |
| 1693 |
|
| 1694 |
|
| 1695 |
_MODULE_ALIASES["cupy.scipy"] = "cupyx.scipy" |
| 1696 |
_FUNCS["cupy", "to_numpy"] = cupy_to_numpy |
| 1697 |
_FUNCS["cupy", "complex"] = complex_add_re_im |
| 1698 |
_CUSTOM_WRAPPERS["cupy", "linalg.svd"] = svd_not_full_matrices_wrapper |
| 1699 |
|
| 1700 |
# ----------------------------------- jax ----------------------------------- # |
| 1701 |
|
| 1702 |
|
| 1703 |
def jax_to_numpy(x): |
| 1704 |
return do("asarray", x, like="numpy") |
| 1705 |
|
| 1706 |
|
| 1707 |
class JaxDefaultRNG: |
| 1708 |
"""Stateful but deterministic random number generator for JAX following |
| 1709 |
numpy's Generator API, compatible with `jax.jit`. |
| 1710 |
""" |
| 1711 |
|
| 1712 |
def __init__(self, seed, **kwargs): |
| 1713 |
import jax |
| 1714 |
|
| 1715 |
self.jax = jax |
| 1716 |
self.key = jax.random.key(seed, **kwargs) |
| 1717 |
|
| 1718 |
def binomial(self, n, p, size=None, **kwargs): |
| 1719 |
self.key, subkey = self.jax.random.split(self.key) |
| 1720 |
shape = _handle_size_to_shape(size) |
| 1721 |
return self.jax.random.binomial( |
| 1722 |
subkey, n=n, p=p, shape=shape, **kwargs |
| 1723 |
) |
| 1724 |
|
| 1725 |
def choice(self, a, size=None, replace=True, p=None, axis=0, **kwargs): |
| 1726 |
self.key, subkey = self.jax.random.split(self.key) |
| 1727 |
shape = _handle_size_to_shape(size) |
| 1728 |
return self.jax.random.choice( |
| 1729 |
subkey, |
| 1730 |
a, |
| 1731 |
shape=shape, |
| 1732 |
replace=replace, |
| 1733 |
p=p, |
| 1734 |
axis=axis, |
| 1735 |
**kwargs, |
| 1736 |
) |
| 1737 |
|
| 1738 |
def exponential(self, scale=1.0, size=None, **kwargs): |
| 1739 |
self.key, subkey = self.jax.random.split(self.key) |
| 1740 |
shape = _handle_size_to_shape(size) |
| 1741 |
x = self.jax.random.exponential(subkey, shape=shape, **kwargs) |
| 1742 |
if scale != 1.0: |
| 1743 |
x *= scale |
| 1744 |
return x |
| 1745 |
|
| 1746 |
def gumbel(self, loc=0.0, scale=1.0, size=None, **kwargs): |
| 1747 |
self.key, subkey = self.jax.random.split(self.key) |
| 1748 |
shape = _handle_size_to_shape(size) |
| 1749 |
x = self.jax.random.gumbel(subkey, shape=shape, **kwargs) |
| 1750 |
if scale != 1.0: |
| 1751 |
x *= scale |
| 1752 |
if loc != 0.0: |
| 1753 |
x += loc |
| 1754 |
return x |
| 1755 |
|
| 1756 |
def integers(self, low, high=None, size=None, **kwargs): |
| 1757 |
self.key, subkey = self.jax.random.split(self.key) |
| 1758 |
shape = _handle_size_to_shape(size) |
| 1759 |
if high is None: |
| 1760 |
high = low |
| 1761 |
low = 0 |
| 1762 |
return self.jax.random.randint( |
| 1763 |
subkey, |
| 1764 |
shape=shape, |
| 1765 |
minval=low, |
| 1766 |
maxval=high, |
| 1767 |
**kwargs, |
| 1768 |
) |
| 1769 |
|
| 1770 |
def normal(self, loc=0.0, scale=1.0, size=None, **kwargs): |
| 1771 |
self.key, subkey = self.jax.random.split(self.key) |
| 1772 |
shape = _handle_size_to_shape(size) |
| 1773 |
x = self.jax.random.normal(subkey, shape=shape, **kwargs) |
| 1774 |
if scale != 1.0: |
| 1775 |
x *= scale |
| 1776 |
if loc != 0.0: |
| 1777 |
x += loc |
| 1778 |
return x |
| 1779 |
|
| 1780 |
def permutation(self, x, **kwargs): |
| 1781 |
self.key, subkey = self.jax.random.split(self.key) |
| 1782 |
return self.jax.random.permutation(subkey, x, **kwargs) |
| 1783 |
|
| 1784 |
def poisson(self, lam=1.0, size=None, **kwargs): |
| 1785 |
self.key, subkey = self.jax.random.split(self.key) |
| 1786 |
shape = _handle_size_to_shape(size) |
| 1787 |
x = self.jax.random.poisson(subkey, lam, shape=shape, **kwargs) |
| 1788 |
return x |
| 1789 |
|
| 1790 |
def random(self, size=None, **kwargs): |
| 1791 |
return self.uniform(size=size, **kwargs) |
| 1792 |
|
| 1793 |
def uniform(self, low=0.0, high=1.0, size=None, **kwargs): |
| 1794 |
self.key, subkey = self.jax.random.split(self.key) |
| 1795 |
shape = _handle_size_to_shape(size) |
| 1796 |
return self.jax.random.uniform( |
| 1797 |
subkey, shape=shape, minval=low, maxval=high, **kwargs |
| 1798 |
) |
| 1799 |
|
| 1800 |
|
| 1801 |
def jax_default_rng(seed, **kwargs): |
| 1802 |
if isinstance(seed, JaxDefaultRNG): |
| 1803 |
return seed |
| 1804 |
return JaxDefaultRNG(seed, **kwargs) |
| 1805 |
|
| 1806 |
|
| 1807 |
register_function("jax", "random.default_rng", jax_default_rng) |
| 1808 |
register_backend(JaxDefaultRNG, "jax") |
| 1809 |
|
| 1810 |
|
| 1811 |
_JAX_RANDOM_KEY = None |
| 1812 |
|
| 1813 |
|
| 1814 |
def jax_random_seed(seed=None): |
| 1815 |
from jax.random import PRNGKey |
| 1816 |
|
| 1817 |
global _JAX_RANDOM_KEY |
| 1818 |
if seed is None: |
| 1819 |
from random import SystemRandom |
| 1820 |
|
| 1821 |
seed = SystemRandom().randint(-(2**63), 2**63 - 1) # inclusive high |
| 1822 |
_JAX_RANDOM_KEY = PRNGKey(seed) |
| 1823 |
|
| 1824 |
|
| 1825 |
def jax_random_get_key(): |
| 1826 |
from jax.random import split |
| 1827 |
|
| 1828 |
global _JAX_RANDOM_KEY |
| 1829 |
if _JAX_RANDOM_KEY is None: |
| 1830 |
jax_random_seed() |
| 1831 |
_JAX_RANDOM_KEY, subkey = split(_JAX_RANDOM_KEY) |
| 1832 |
return subkey |
| 1833 |
|
| 1834 |
|
| 1835 |
def jax_random_uniform(low=0.0, high=1.0, size=None, **kwargs): |
| 1836 |
from jax.random import uniform |
| 1837 |
|
| 1838 |
if size is None: |
| 1839 |
size = () |
| 1840 |
return uniform( |
| 1841 |
jax_random_get_key(), shape=size, minval=low, maxval=high, **kwargs |
| 1842 |
) |
| 1843 |
|
| 1844 |
|
| 1845 |
def jax_random_normal(loc=0.0, scale=1.0, size=None, **kwargs): |
| 1846 |
from jax.random import normal |
| 1847 |
|
| 1848 |
if size is None: |
| 1849 |
size = () |
| 1850 |
x = normal(jax_random_get_key(), shape=size, **kwargs) |
| 1851 |
if scale != 1.0: |
| 1852 |
x *= scale |
| 1853 |
if loc != 0.0: |
| 1854 |
x += loc |
| 1855 |
return x |
| 1856 |
|
| 1857 |
|
| 1858 |
_BACKEND_ALIASES["jaxlib"] = "jax" |
| 1859 |
_MODULE_ALIASES["jax.scipy"] = "jax.scipy" |
| 1860 |
_MODULE_ALIASES["jax"] = "jax.numpy" |
| 1861 |
_SUBMODULE_ALIASES["jax", "complex"] = "jax.lax" |
| 1862 |
_SUBMODULE_ALIASES["jax", "linalg.expm"] = "jax.scipy.linalg" |
| 1863 |
_SUBMODULE_ALIASES["jax", "linalg.householder_product"] = "jax.lax.linalg" |
| 1864 |
# n.b. jax supports fat QR but not when computing gradients |
| 1865 |
_CUSTOM_WRAPPERS["jax", "linalg.qr"] = qr_allow_fat |
| 1866 |
_CUSTOM_WRAPPERS["jax", "linalg.svd"] = svd_not_full_matrices_wrapper |
| 1867 |
_FUNCS["jax", "to_numpy"] = jax_to_numpy |
| 1868 |
_FUNCS["jax", "random.seed"] = jax_random_seed |
| 1869 |
_FUNCS["jax", "random.uniform"] = jax_random_uniform |
| 1870 |
_FUNCS["jax", "random.normal"] = jax_random_normal |
| 1871 |
|
| 1872 |
|
| 1873 |
# --------------------------------- aesara ---------------------------------- # |
| 1874 |
|
| 1875 |
|
| 1876 |
@shape.register("aesara") |
| 1877 |
def aesara_shape(x): |
| 1878 |
return x.type.shape |
| 1879 |
|
| 1880 |
|
| 1881 |
_MODULE_ALIASES["aesara"] = "aesara.tensor" |
| 1882 |
_FUNCS["aesara", "shape"] = aesara_shape |
| 1883 |
|
| 1884 |
|
| 1885 |
# -------------------------------- autograd --------------------------------- # |
| 1886 |
|
| 1887 |
_MODULE_ALIASES["autograd"] = "autograd.numpy" |
| 1888 |
_CUSTOM_WRAPPERS["autograd", "linalg.svd"] = svd_not_full_matrices_wrapper |
| 1889 |
_FUNCS["autograd", "complex"] = complex_add_re_im |
| 1890 |
|
| 1891 |
|
| 1892 |
# ---------------------------------- dask ----------------------------------- # |
| 1893 |
|
| 1894 |
|
| 1895 |
def dask_to_numpy(x): |
| 1896 |
return x.compute() |
| 1897 |
|
| 1898 |
|
| 1899 |
def dask_eye_wrapper(fn): |
| 1900 |
# Make M work as positional argument |
| 1901 |
@functools.wraps(fn) |
| 1902 |
def numpy_like(N, M=None, **kwargs): |
| 1903 |
return fn(N, M=M, **kwargs) |
| 1904 |
|
| 1905 |
return numpy_like |
| 1906 |
|
| 1907 |
|
| 1908 |
_FUNCS["dask", "to_numpy"] = dask_to_numpy |
| 1909 |
_FUNCS["dask", "complex"] = complex_add_re_im |
| 1910 |
_FUNC_ALIASES["dask", "abs"] = "absolute" |
| 1911 |
_FUNC_ALIASES["dask", "identity"] = "eye" |
| 1912 |
_MODULE_ALIASES["dask"] = "dask.array" |
| 1913 |
_CUSTOM_WRAPPERS["dask", "linalg.svd"] = svd_manual_full_matrices_kwarg |
| 1914 |
_CUSTOM_WRAPPERS["dask", "linalg.cholesky"] = cholesky_lower |
| 1915 |
_CUSTOM_WRAPPERS["dask", "random.normal"] = with_dtype_wrapper |
| 1916 |
_CUSTOM_WRAPPERS["dask", "random.uniform"] = with_dtype_wrapper |
| 1917 |
_CUSTOM_WRAPPERS["dask", "eye"] = dask_eye_wrapper |
| 1918 |
|
| 1919 |
# ---------------------------------- mars ----------------------------------- # |
| 1920 |
|
| 1921 |
|
| 1922 |
def mars_to_numpy(x): |
| 1923 |
return x.to_numpy() |
| 1924 |
|
| 1925 |
|
| 1926 |
_FUNCS["mars", "to_numpy"] = mars_to_numpy |
| 1927 |
_FUNCS["mars", "complex"] = complex_add_re_im |
| 1928 |
_MODULE_ALIASES["mars"] = "mars.tensor" |
| 1929 |
_CUSTOM_WRAPPERS["mars", "linalg.cholesky"] = cholesky_lower |
| 1930 |
|
| 1931 |
|
| 1932 |
# ----------------------------------- ctf ----------------------------------- # |
| 1933 |
|
| 1934 |
|
| 1935 |
def ctf_array(x): |
| 1936 |
return do("astensor", x, like="ctf") |
| 1937 |
|
| 1938 |
|
| 1939 |
def ctf_to_numpy(x): |
| 1940 |
return x.to_nparray() |
| 1941 |
|
| 1942 |
|
| 1943 |
def ctf_count_nonzero(x): |
| 1944 |
return (x != 0).astype(int).sum() |
| 1945 |
|
| 1946 |
|
| 1947 |
@get_dtype_name.register("ctf") |
| 1948 |
def ctf_get_dtype_name(x): |
| 1949 |
return x.dtype.__name__ |
| 1950 |
|
| 1951 |
|
| 1952 |
_FUNCS["ctf", "array"] = ctf_array |
| 1953 |
_FUNCS["ctf", "complex"] = complex_add_re_im |
| 1954 |
_FUNCS["ctf", "allclose"] = allclose |
| 1955 |
_FUNCS["ctf", "to_numpy"] = ctf_to_numpy |
| 1956 |
_FUNCS["ctf", "count_nonzero"] = ctf_count_nonzero |
| 1957 |
|
| 1958 |
_SUBMODULE_ALIASES["ctf", "float32"] = "numpy" |
| 1959 |
_SUBMODULE_ALIASES["ctf", "float64"] = "numpy" |
| 1960 |
_SUBMODULE_ALIASES["ctf", "complex64"] = "numpy" |
| 1961 |
_SUBMODULE_ALIASES["ctf", "complex128"] = "numpy" |
| 1962 |
_SUBMODULE_ALIASES["ctf", "linalg.svd"] = "ctf" |
| 1963 |
_SUBMODULE_ALIASES["ctf", "linalg.eigh"] = "ctf" |
| 1964 |
_SUBMODULE_ALIASES["ctf", "linalg.qr"] = "ctf" |
| 1965 |
_SUBMODULE_ALIASES["ctf", "linalg.norm"] = "ctf" |
| 1966 |
|
| 1967 |
_FUNC_ALIASES["ctf", "random.uniform"] = "random" |
| 1968 |
|
| 1969 |
_CUSTOM_WRAPPERS["ctf", "random.uniform"] = scale_random_uniform_manually |
| 1970 |
|
| 1971 |
|
| 1972 |
# ------------------------------- sparse------------------------------------- # |
| 1973 |
|
| 1974 |
|
| 1975 |
def sparse_array(x): |
| 1976 |
return do("COO.from_numpy", x, like="sparse") |
| 1977 |
|
| 1978 |
|
| 1979 |
def sparse_to_numpy(x): |
| 1980 |
return x.todense() |
| 1981 |
|
| 1982 |
|
| 1983 |
def sparse_transpose(x, axes=None): |
| 1984 |
return x.transpose(axes) |
| 1985 |
|
| 1986 |
|
| 1987 |
def sparse_reshape(x, shape): |
| 1988 |
return x.reshape(shape) |
| 1989 |
|
| 1990 |
|
| 1991 |
def sparse_sum(x, axis=None, keepdims=False, dtype=None, out=None): |
| 1992 |
return x.sum(axis=axis, keepdims=keepdims, dtype=dtype, out=out) |
| 1993 |
|
| 1994 |
|
| 1995 |
def sparse_prod(x, axis=None, keepdims=False, dtype=None, out=None): |
| 1996 |
return x.prod(axis=axis, keepdims=keepdims, dtype=dtype, out=out) |
| 1997 |
|
| 1998 |
|
| 1999 |
def sparse_conj(x): |
| 2000 |
return x.conj() |
| 2001 |
|
| 2002 |
|
| 2003 |
def sparse_real(x): |
| 2004 |
return x.real |
| 2005 |
|
| 2006 |
|
| 2007 |
def sparse_imag(x): |
| 2008 |
return x.imag |
| 2009 |
|
| 2010 |
|
| 2011 |
def sparse_count_nonzero(x): |
| 2012 |
return x.nnz |
| 2013 |
|
| 2014 |
|
| 2015 |
def sparse_random_uniform(low=0.0, high=1.0, size=None, dtype=None, **kwargs): |
| 2016 |
def rvs(nnz): |
| 2017 |
return do( |
| 2018 |
"random.uniform", low, high, (nnz,), dtype=dtype, like="numpy" |
| 2019 |
) |
| 2020 |
|
| 2021 |
return do("random", size, data_rvs=rvs, **kwargs, like="sparse") |
| 2022 |
|
| 2023 |
|
| 2024 |
def sparse_random_normal(loc=0.0, scale=1.0, size=None, dtype=None, **kwargs): |
| 2025 |
def rvs(nnz): |
| 2026 |
return do( |
| 2027 |
"random.normal", loc, scale, (nnz,), dtype=dtype, like="numpy" |
| 2028 |
) |
| 2029 |
|
| 2030 |
return do("random", size, data_rvs=rvs, **kwargs, like="sparse") |
| 2031 |
|
| 2032 |
|
| 2033 |
_FUNCS["sparse", "array"] = sparse_array |
| 2034 |
_FUNCS["sparse", "to_numpy"] = sparse_to_numpy |
| 2035 |
_FUNCS["sparse", "transpose"] = sparse_transpose |
| 2036 |
_FUNCS["sparse", "reshape"] = sparse_reshape |
| 2037 |
_FUNCS["sparse", "sum"] = sparse_sum |
| 2038 |
_FUNCS["sparse", "prod"] = sparse_prod |
| 2039 |
_FUNCS["sparse", "conj"] = sparse_conj |
| 2040 |
_FUNCS["sparse", "real"] = sparse_real |
| 2041 |
_FUNCS["sparse", "real"] = sparse_real |
| 2042 |
_FUNCS["sparse", "imag"] = sparse_imag |
| 2043 |
_FUNCS["sparse", "complex"] = complex_add_re_im |
| 2044 |
_FUNCS["sparse", "count_nonzero"] = sparse_count_nonzero |
| 2045 |
_FUNCS["sparse", "random.uniform"] = sparse_random_uniform |
| 2046 |
_FUNCS["sparse", "random.normal"] = sparse_random_normal |
| 2047 |
|
| 2048 |
_FUNC_ALIASES["sparse", "identity"] = "eye" |
| 2049 |
|
| 2050 |
# sparse uses numpys __array_func__ interface |
| 2051 |
for f in ( |
| 2052 |
"log", |
| 2053 |
"log2", |
| 2054 |
"log10", |
| 2055 |
"exp", |
| 2056 |
"sqrt", |
| 2057 |
"sign", |
| 2058 |
"sin", |
| 2059 |
"cos", |
| 2060 |
"tan", |
| 2061 |
"arcsin", |
| 2062 |
"arccos", |
| 2063 |
"arctan", |
| 2064 |
"sinh", |
| 2065 |
"cosh", |
| 2066 |
"tanh", |
| 2067 |
"arcsinh", |
| 2068 |
"arccosh", |
| 2069 |
"arctanh", |
| 2070 |
"tensordot", |
| 2071 |
# NB put tensordot here, as sparse.tensordot can produce dense (numpy) |
| 2072 |
# arrays but errors when both inputs are dense - we want nested calls to |
| 2073 |
# tensordot to handle this |
| 2074 |
): |
| 2075 |
_SUBMODULE_ALIASES["sparse", f] = "numpy" |
| 2076 |
|
| 2077 |
|
| 2078 |
# ------------------------------- tensorflow -------------------------------- # |
| 2079 |
|
| 2080 |
|
| 2081 |
def tensorflow_to_numpy(x): |
| 2082 |
return x.numpy() |
| 2083 |
|
| 2084 |
|
| 2085 |
def tensorflow_indices(dimensions): |
| 2086 |
_meshgrid = get_lib_fn("tensorflow", "meshgrid") |
| 2087 |
_arange = get_lib_fn("tensorflow", "arange") |
| 2088 |
return _meshgrid(*map(_arange, dimensions), indexing="ij") |
| 2089 |
|
| 2090 |
|
| 2091 |
_MODULE_ALIASES["tensorflow.linalg"] = "tensorflow.linalg" |
| 2092 |
_MODULE_ALIASES["tensorflow.random"] = "tensorflow.random" |
| 2093 |
_MODULE_ALIASES["tensorflow"] = "tensorflow.experimental.numpy" |
| 2094 |
|
| 2095 |
_FUNCS["tensorflow", "to_numpy"] = tensorflow_to_numpy |
| 2096 |
_FUNCS["tensorflow", "indices"] = tensorflow_indices |
| 2097 |
|
| 2098 |
_FUNC_ALIASES["tensorflow", "astype"] = "cast" |
| 2099 |
_SUBMODULE_ALIASES["tensorflow", "cast"] = "tensorflow" |
| 2100 |
_SUBMODULE_ALIASES["tensorflow", "astype"] = "tensorflow" |
| 2101 |
_SUBMODULE_ALIASES["tensorflow", "complex"] = "tensorflow" |
| 2102 |
|
| 2103 |
_CUSTOM_WRAPPERS["tensorflow", "linalg.svd"] = svd_sUV_to_UsVH_wrapper |
| 2104 |
_CUSTOM_WRAPPERS["tensorflow", "linalg.solve"] = binary_allow_1d_rhs_wrap |
| 2105 |
_CUSTOM_WRAPPERS["tensorflow", "random.uniform"] = make_translator( |
| 2106 |
[ |
| 2107 |
("low", ("minval", 0.0)), |
| 2108 |
("high", ("maxval", 1.0)), |
| 2109 |
("size", ("shape", ())), |
| 2110 |
] |
| 2111 |
) |
| 2112 |
_CUSTOM_WRAPPERS["tensorflow", "random.normal"] = make_translator( |
| 2113 |
[ |
| 2114 |
("loc", ("mean", 0.0)), |
| 2115 |
("scale", ("stddev", 1.0)), |
| 2116 |
("size", ("shape", ())), |
| 2117 |
] |
| 2118 |
) |
| 2119 |
|
| 2120 |
|
| 2121 |
def tensorflow_pad_wrap(tf_pad): |
| 2122 |
def numpy_like(array, pad_width, mode="constant", constant_values=0): |
| 2123 |
if mode != "constant": |
| 2124 |
raise NotImplementedError |
| 2125 |
|
| 2126 |
try: |
| 2127 |
if len(pad_width) == 1: |
| 2128 |
pad_width = pad_width * ndim(array) |
| 2129 |
except TypeError: |
| 2130 |
pad_width = ((pad_width, pad_width),) * ndim(array) |
| 2131 |
|
| 2132 |
return tf_pad( |
| 2133 |
array, pad_width, mode="CONSTANT", constant_values=constant_values |
| 2134 |
) |
| 2135 |
|
| 2136 |
return numpy_like |
| 2137 |
|
| 2138 |
|
| 2139 |
_CUSTOM_WRAPPERS["tensorflow", "pad"] = tensorflow_pad_wrap |
| 2140 |
_SUBMODULE_ALIASES["tensorflow", "pad"] = "tensorflow" |
| 2141 |
|
| 2142 |
|
| 2143 |
register_creation_routine("tensorflow", "linspace", inject_dtype=False) |
| 2144 |
|
| 2145 |
|
| 2146 |
# ---------------------------------- torch ---------------------------------- # |
| 2147 |
|
| 2148 |
|
| 2149 |
@shape.register("torch") |
| 2150 |
def torch_shape(x): |
| 2151 |
# torch returns a Size object, we want tuple[int] |
| 2152 |
return tuple(map(int, x.shape)) |
| 2153 |
|
| 2154 |
|
| 2155 |
@size.register("torch") |
| 2156 |
def torch_size(x): |
| 2157 |
return x.numel() |
| 2158 |
|
| 2159 |
|
| 2160 |
def torch_to_numpy(x): |
| 2161 |
return x.detach().cpu().numpy() |
| 2162 |
|
| 2163 |
|
| 2164 |
def torch_copy(x): |
| 2165 |
return x.detach().clone() |
| 2166 |
|
| 2167 |
|
| 2168 |
def torch_transpose(x, axes=None): |
| 2169 |
if axes is None: |
| 2170 |
axes = reversed(range(0, x.ndimension())) |
| 2171 |
return x.permute(*axes) |
| 2172 |
|
| 2173 |
|
| 2174 |
def torch_count_nonzero(x): |
| 2175 |
return do("sum", x != 0, like="torch") |
| 2176 |
|
| 2177 |
|
| 2178 |
def torch_astype(x, dtype): |
| 2179 |
return x.to(dtype=to_backend_dtype(dtype, like=x)) |
| 2180 |
|
| 2181 |
|
| 2182 |
@functools.lru_cache(None) |
| 2183 |
def _torch_get_dtype_name(dtype): |
| 2184 |
return str(dtype).split(".")[-1] |
| 2185 |
|
| 2186 |
|
| 2187 |
@get_dtype_name.register("torch") |
| 2188 |
def torch_get_dtype_name(x): |
| 2189 |
return _torch_get_dtype_name(x.dtype) |
| 2190 |
|
| 2191 |
|
| 2192 |
def torch_real(x): |
| 2193 |
# torch doesn't support calling real on real arrays |
| 2194 |
try: |
| 2195 |
if x.is_complex(): |
| 2196 |
return x.real |
| 2197 |
except AttributeError: |
| 2198 |
pass |
| 2199 |
return x |
| 2200 |
|
| 2201 |
|
| 2202 |
def torch_imag(x): |
| 2203 |
# torch doesn't support calling imag on real arrays |
| 2204 |
try: |
| 2205 |
if x.is_complex(): |
| 2206 |
return x.imag |
| 2207 |
except AttributeError: |
| 2208 |
pass |
| 2209 |
return do("zeros_like", x) |
| 2210 |
|
| 2211 |
|
| 2212 |
def torch_linalg_solve_wrap(fn): |
| 2213 |
@binary_allow_1d_rhs_wrap |
| 2214 |
def numpy_like(a, b): |
| 2215 |
return fn(b, a)[0] |
| 2216 |
|
| 2217 |
return numpy_like |
| 2218 |
|
| 2219 |
|
| 2220 |
def torch_linalg_eigh(x): |
| 2221 |
return tuple(do("symeig", x, eigenvectors=True, like="torch")) |
| 2222 |
|
| 2223 |
|
| 2224 |
def torch_linalg_eigvalsh(x): |
| 2225 |
return do("symeig", x, eigenvectors=False, like="torch")[0] |
| 2226 |
|
| 2227 |
|
| 2228 |
def torch_tensordot_wrap(fn): |
| 2229 |
@functools.wraps(fn) |
| 2230 |
def numpy_like(a, b, axes=2): |
| 2231 |
return fn(a, b, dims=axes) |
| 2232 |
|
| 2233 |
return numpy_like |
| 2234 |
|
| 2235 |
|
| 2236 |
def torch_pad(array, pad_width, mode="constant", constant_values=0): |
| 2237 |
if mode != "constant": |
| 2238 |
raise NotImplementedError |
| 2239 |
|
| 2240 |
try: |
| 2241 |
# numpy takes pads like ((0, 0), (1, 1), ... (n-1, n-1)) |
| 2242 |
# torch takes pads like (n-1, n-1, n-2, n-2, n-3, n-3, ...) |
| 2243 |
pad = tuple(itertools.chain.from_iterable(pad_width))[::-1] |
| 2244 |
|
| 2245 |
# a single tuple was specified ((a, b),) - use for all axes |
| 2246 |
if len(pad) == 2: |
| 2247 |
pad = pad * array.ndimension() |
| 2248 |
|
| 2249 |
except TypeError: |
| 2250 |
# assume int |
| 2251 |
pad = (pad_width,) * 2 * array.ndimension() |
| 2252 |
|
| 2253 |
return do( |
| 2254 |
"nn.functional.pad", |
| 2255 |
array, |
| 2256 |
pad=pad, |
| 2257 |
mode=mode, |
| 2258 |
value=constant_values, |
| 2259 |
like="torch", |
| 2260 |
) |
| 2261 |
|
| 2262 |
|
| 2263 |
def torch_split_wrap(fn): |
| 2264 |
# for torch >=1.8 we can use tensor_split instead, but in current stable |
| 2265 |
# release this function has not been added |
| 2266 |
@functools.wraps(fn) |
| 2267 |
def numpy_like(ary, indices_or_sections, axis=0, **kwargs): |
| 2268 |
if isinstance(indices_or_sections, int): |
| 2269 |
split_size = shape(ary)[axis] // indices_or_sections |
| 2270 |
return fn(ary, split_size, dim=axis, **kwargs) |
| 2271 |
else: |
| 2272 |
# torch.split doesn't support empty splits |
| 2273 |
if len(indices_or_sections) == 0: |
| 2274 |
return (ary,) |
| 2275 |
|
| 2276 |
diff = do( |
| 2277 |
"diff", |
| 2278 |
indices_or_sections, |
| 2279 |
prepend=0, |
| 2280 |
append=shape(ary)[axis], |
| 2281 |
like="numpy", |
| 2282 |
) |
| 2283 |
diff = list(diff) |
| 2284 |
return fn(ary, diff, dim=axis) |
| 2285 |
|
| 2286 |
return numpy_like |
| 2287 |
|
| 2288 |
|
| 2289 |
def torch_zeros_ones_wrap(fn): |
| 2290 |
@functools.wraps(fn) |
| 2291 |
def numpy_like(shape, dtype=None, **kwargs): |
| 2292 |
if dtype is not None: |
| 2293 |
dtype = to_backend_dtype(dtype, like="torch") |
| 2294 |
return fn(shape, dtype=dtype, **kwargs) |
| 2295 |
|
| 2296 |
return numpy_like |
| 2297 |
|
| 2298 |
|
| 2299 |
def torch_eye_wrap(fn): |
| 2300 |
@functools.wraps(fn) |
| 2301 |
def numpy_like(N, M=None, dtype=None, **kwargs): |
| 2302 |
if dtype is not None: |
| 2303 |
dtype = to_backend_dtype(dtype, like="torch") |
| 2304 |
if M is not None: |
| 2305 |
return fn(N, m=M, dtype=dtype, **kwargs) |
| 2306 |
else: |
| 2307 |
return fn(N, dtype=dtype, **kwargs) |
| 2308 |
|
| 2309 |
return numpy_like |
| 2310 |
|
| 2311 |
|
| 2312 |
def torch_sort_wrap(fn): |
| 2313 |
@functools.wraps(fn) |
| 2314 |
def numpy_like(a, axis=-1): |
| 2315 |
return fn(a, dim=axis)[0] |
| 2316 |
|
| 2317 |
return numpy_like |
| 2318 |
|
| 2319 |
|
| 2320 |
def torch_indices(dimensions): |
| 2321 |
_meshgrid = get_lib_fn("torch", "meshgrid") |
| 2322 |
_arange = get_lib_fn("torch", "arange") |
| 2323 |
return _meshgrid(*map(_arange, dimensions), indexing="ij") |
| 2324 |
|
| 2325 |
|
| 2326 |
def torch_flip_wrap(torch_flip): |
| 2327 |
def numpy_like(x, axis=None): |
| 2328 |
if axis is None: |
| 2329 |
dims = tuple(range(x.ndimension())) |
| 2330 |
elif isinstance(axis, int): |
| 2331 |
dims = (axis,) |
| 2332 |
else: |
| 2333 |
# already tuple/list |
| 2334 |
dims = axis |
| 2335 |
return torch_flip(x, dims) |
| 2336 |
|
| 2337 |
return numpy_like |
| 2338 |
|
| 2339 |
|
| 2340 |
class TorchDefaultRNG: |
| 2341 |
def __init__(self, seed, device=None): |
| 2342 |
import torch |
| 2343 |
|
| 2344 |
self._torch = torch |
| 2345 |
self._generator = torch.Generator(device=device) |
| 2346 |
self._generator.manual_seed(seed) |
| 2347 |
|
| 2348 |
# def binomial(self, n, p, size=None, **kwargs): |
| 2349 |
# raise NotImplementedError() |
| 2350 |
|
| 2351 |
# def choice(self, a, size=None, replace=True, p=None, axis=0, **kwargs): |
| 2352 |
# raise NotImplementedError() |
| 2353 |
|
| 2354 |
# def exponential(self, scale=1.0, size=None, **kwargs): |
| 2355 |
# raise NotImplementedError() |
| 2356 |
|
| 2357 |
# def gumbel(self, loc=0.0, scale=1.0, size=None, **kwargs): |
| 2358 |
# raise NotImplementedError() |
| 2359 |
|
| 2360 |
def integers(self, low, high=None, size=None, **kwargs): |
| 2361 |
if high is None: |
| 2362 |
high = low |
| 2363 |
low = 0 |
| 2364 |
size = _handle_size_to_shape(size) |
| 2365 |
return self._torch.randint( |
| 2366 |
low, high, size, generator=self._generator, **kwargs |
| 2367 |
) |
| 2368 |
|
| 2369 |
# def poisson(self, lam=1.0, size=None, **kwargs): |
| 2370 |
# raise NotImplementedError() |
| 2371 |
|
| 2372 |
def normal(self, loc=0.0, scale=1.0, size=None, **kwargs): |
| 2373 |
size = _handle_size_to_shape(size) |
| 2374 |
x = self._torch.randn(size, generator=self._generator, **kwargs) |
| 2375 |
if scale != 1.0: |
| 2376 |
x = x * scale |
| 2377 |
if loc != 0.0: |
| 2378 |
x = x + loc |
| 2379 |
return x |
| 2380 |
|
| 2381 |
def random(self, size=None, **kwargs): |
| 2382 |
size = _handle_size_to_shape(size) |
| 2383 |
return self._torch.rand(size, generator=self._generator, **kwargs) |
| 2384 |
|
| 2385 |
def permutation(self, x, **kwargs): |
| 2386 |
if isinstance(x, int): |
| 2387 |
return self._torch.randperm(x, generator=self._generator, **kwargs) |
| 2388 |
|
| 2389 |
axis = kwargs.get("axis", 0) |
| 2390 |
n = x.shape[axis] |
| 2391 |
perm_indices = self._torch.randperm( |
| 2392 |
n, generator=self._generator, device=x.device |
| 2393 |
) |
| 2394 |
return self._torch.index_select(x, axis, perm_indices) |
| 2395 |
|
| 2396 |
def uniform(self, low=0.0, high=1.0, size=None, **kwargs): |
| 2397 |
size = _handle_size_to_shape(size) |
| 2398 |
x = self._torch.rand(size, generator=self._generator, **kwargs) |
| 2399 |
if low != 0.0 or high != 1.0: |
| 2400 |
x = x * (high - low) + low |
| 2401 |
return x |
| 2402 |
|
| 2403 |
|
| 2404 |
def torch_default_rng(seed, **kwargs): |
| 2405 |
if isinstance(seed, TorchDefaultRNG): |
| 2406 |
return seed |
| 2407 |
return TorchDefaultRNG(seed, **kwargs) |
| 2408 |
|
| 2409 |
|
| 2410 |
register_function("torch", "random.default_rng", torch_default_rng) |
| 2411 |
register_backend(TorchDefaultRNG, "torch") |
| 2412 |
|
| 2413 |
|
| 2414 |
_FUNCS["torch", "pad"] = torch_pad |
| 2415 |
_FUNCS["torch", "real"] = torch_real |
| 2416 |
_FUNCS["torch", "imag"] = torch_imag |
| 2417 |
_FUNCS["torch", "astype"] = torch_astype |
| 2418 |
_FUNCS["torch", "copy"] = torch_copy |
| 2419 |
_FUNCS["torch", "to_numpy"] = torch_to_numpy |
| 2420 |
_FUNCS["torch", "complex"] = complex_add_re_im |
| 2421 |
_FUNCS["torch", "transpose"] = torch_transpose |
| 2422 |
_FUNCS["torch", "count_nonzero"] = torch_count_nonzero |
| 2423 |
_FUNCS["torch", "indices"] = torch_indices |
| 2424 |
|
| 2425 |
_FUNC_ALIASES["torch", "array"] = "tensor" |
| 2426 |
_FUNC_ALIASES["torch", "asarray"] = "as_tensor" |
| 2427 |
_FUNC_ALIASES["torch", "clip"] = "clamp" |
| 2428 |
_FUNC_ALIASES["torch", "concatenate"] = "cat" |
| 2429 |
_FUNC_ALIASES["torch", "conjugate"] = "conj" |
| 2430 |
_FUNC_ALIASES["torch", "equal"] = "eq" |
| 2431 |
_FUNC_ALIASES["torch", "expand_dims"] = "unsqueeze" |
| 2432 |
_FUNC_ALIASES["torch", "identity"] = "eye" |
| 2433 |
_FUNC_ALIASES["torch", "linalg.expm"] = "matrix_exp" |
| 2434 |
_FUNC_ALIASES["torch", "max"] = "amax" |
| 2435 |
_FUNC_ALIASES["torch", "min"] = "amin" |
| 2436 |
_FUNC_ALIASES["torch", "power"] = "pow" |
| 2437 |
_FUNC_ALIASES["torch", "random.normal"] = "randn" |
| 2438 |
_FUNC_ALIASES["torch", "random.uniform"] = "rand" |
| 2439 |
_FUNC_ALIASES["torch", "scipy.linalg.expm"] = "matrix_exp" |
| 2440 |
_FUNC_ALIASES["torch", "split"] = "tensor_split" |
| 2441 |
_FUNC_ALIASES["torch", "take"] = "index_select" |
| 2442 |
_FUNC_ALIASES["torch", "take_along_axis"] = "take_along_dim" |
| 2443 |
|
| 2444 |
_SUBMODULE_ALIASES["torch", "linalg.expm"] = "torch" |
| 2445 |
_SUBMODULE_ALIASES["torch", "scipy.linalg.expm"] = "torch" |
| 2446 |
_SUBMODULE_ALIASES["torch", "random.normal"] = "torch" |
| 2447 |
_SUBMODULE_ALIASES["torch", "random.uniform"] = "torch" |
| 2448 |
|
| 2449 |
_CUSTOM_WRAPPERS["torch", "clip"] = make_translator( |
| 2450 |
[("a", ("input",)), ("a_min", ("min",)), ("a_max", ("max",))] |
| 2451 |
) |
| 2452 |
_CUSTOM_WRAPPERS["torch", "concatenate"] = make_translator( |
| 2453 |
[("arrays", ("tensors",)), ("axis", ("dim", 0))] |
| 2454 |
) |
| 2455 |
_CUSTOM_WRAPPERS["torch", "diagonal"] = make_translator( |
| 2456 |
[("a", ("input",)), ("axis1", ("dim1", 0)), ("axis2", ("dim2", 1))] |
| 2457 |
) |
| 2458 |
_CUSTOM_WRAPPERS["torch", "empty"] = make_translator([("shape", ("size",))]) |
| 2459 |
_CUSTOM_WRAPPERS["torch", "expand_dims"] = make_translator( |
| 2460 |
[("a", ("input",)), ("axis", ("dim",))] |
| 2461 |
) |
| 2462 |
_CUSTOM_WRAPPERS["torch", "eye"] = torch_eye_wrap |
| 2463 |
_CUSTOM_WRAPPERS["torch", "flip"] = torch_flip_wrap |
| 2464 |
_CUSTOM_WRAPPERS["torch", "linalg.svd"] = svd_not_full_matrices_wrapper |
| 2465 |
_CUSTOM_WRAPPERS["torch", "ones"] = torch_zeros_ones_wrap |
| 2466 |
_CUSTOM_WRAPPERS["torch", "random.normal"] = scale_random_normal_manually |
| 2467 |
_CUSTOM_WRAPPERS["torch", "random.uniform"] = scale_random_uniform_manually |
| 2468 |
_CUSTOM_WRAPPERS["torch", "sort"] = torch_sort_wrap |
| 2469 |
_CUSTOM_WRAPPERS["torch", "stack"] = make_translator( |
| 2470 |
[("arrays", ("tensors",)), ("axis", ("dim", 0))] |
| 2471 |
) |
| 2472 |
_CUSTOM_WRAPPERS["torch", "take"] = make_translator( |
| 2473 |
[("a", ("input",)), ("indices", ("index",)), ("axis", ("dim",))] |
| 2474 |
) |
| 2475 |
_CUSTOM_WRAPPERS["torch", "tensordot"] = torch_tensordot_wrap |
| 2476 |
_CUSTOM_WRAPPERS["torch", "tril"] = make_translator( |
| 2477 |
[("m", ("input",)), ("k", ("diagonal", 0))] |
| 2478 |
) |
| 2479 |
_CUSTOM_WRAPPERS["torch", "triu"] = make_translator( |
| 2480 |
[("m", ("input",)), ("k", ("diagonal", 0))] |
| 2481 |
) |
| 2482 |
_CUSTOM_WRAPPERS["torch", "zeros"] = torch_zeros_ones_wrap |
| 2483 |
_CUSTOM_WRAPPERS["torch", "take_along_axis"] = make_translator( |
| 2484 |
[("arr", ("input",)), ("indices", ("indices",)), ("axis", ("dim", -1))] |
| 2485 |
) |
| 2486 |
|
| 2487 |
_torch_reduce_translation = [ |
| 2488 |
("a", ("input",)), |
| 2489 |
("axis", ("dim",)), |
| 2490 |
("keepdims", ("keepdim",)), |
| 2491 |
] |
| 2492 |
for f in ("sum", "max", "min", "prod", "mean", "median", "std", "var"): |
| 2493 |
# TODO: search "keepdim" in torch docs to find more |
| 2494 |
_CUSTOM_WRAPPERS["torch", f] = make_translator(_torch_reduce_translation) |
| 2495 |
|
| 2496 |
# for older versions of torch, can provide some alternative implementations |
| 2497 |
_MODULE_ALIASES["torch[alt]"] = "torch" |
| 2498 |
|
| 2499 |
_FUNCS["torch[alt]", "linalg.eigh"] = torch_linalg_eigh |
| 2500 |
_FUNCS["torch[alt]", "linalg.eigvalsh"] = torch_linalg_eigvalsh |
| 2501 |
|
| 2502 |
_SUBMODULE_ALIASES["torch[alt]", "linalg.norm"] = "torch" |
| 2503 |
_SUBMODULE_ALIASES["torch[alt]", "linalg.qr"] = "torch" |
| 2504 |
_SUBMODULE_ALIASES["torch[alt]", "linalg.solve"] = "torch" |
| 2505 |
_SUBMODULE_ALIASES["torch[alt]", "linalg.svd"] = "torch" |
| 2506 |
|
| 2507 |
_CUSTOM_WRAPPERS["torch[alt]", "linalg.qr"] = qr_allow_fat |
| 2508 |
_CUSTOM_WRAPPERS["torch[alt]", "linalg.solve"] = torch_linalg_solve_wrap |
| 2509 |
_CUSTOM_WRAPPERS["torch[alt]", "linalg.svd"] = svd_UsV_to_UsVH_wrapper |
| 2510 |
_CUSTOM_WRAPPERS["torch[alt]", "split"] = torch_split_wrap |
| 2511 |
|
| 2512 |
for f in _CREATION_ROUTINES: |
| 2513 |
register_creation_routine("torch", f, inject_device=True) |
| 2514 |
|
| 2515 |
# ---------------------------------- mxnet ---------------------------------- # |
| 2516 |
|
| 2517 |
|
| 2518 |
def mxnet_to_numpy(x): |
| 2519 |
return x.asnumpy() |
| 2520 |
|
| 2521 |
|
| 2522 |
_MODULE_ALIASES["mxnet"] = "mxnet.numpy" |
| 2523 |
_FUNCS["mxnet", "to_numpy"] = mxnet_to_numpy |
| 2524 |
|
| 2525 |
|
| 2526 |
# --------------------------------- paddle ---------------------------------- # |
| 2527 |
|
| 2528 |
_paddle_dtype_name_conversion = { |
| 2529 |
"BOOL": "bool", |
| 2530 |
"INT8": "int8", |
| 2531 |
"INT16": "int16", |
| 2532 |
"INT32": "int32", |
| 2533 |
"INT64": "int64", |
| 2534 |
"FP16": "float16", |
| 2535 |
"FP32": "float32", |
| 2536 |
"FP64": "float64", |
| 2537 |
"COMPLEX64": "complex64", |
| 2538 |
"COMPLEX128": "complex128", |
| 2539 |
} |
| 2540 |
|
| 2541 |
|
| 2542 |
@get_dtype_name.register("paddle") |
| 2543 |
def paddle_get_dtype_name(x): |
| 2544 |
return _paddle_dtype_name_conversion[x.dtype.name] |
| 2545 |
|
| 2546 |
|
| 2547 |
@shape.register("paddle") |
| 2548 |
def paddle_shape(x): |
| 2549 |
# convert from list |
| 2550 |
return tuple(x.shape) |
| 2551 |
|
| 2552 |
|
| 2553 |
def paddle_to_numpy(x): |
| 2554 |
return x.numpy() |
| 2555 |
|
| 2556 |
|
| 2557 |
def paddle_transpose(a, axes=None): |
| 2558 |
if axes is None: |
| 2559 |
axes = tuple(range(a.ndim - 1, -1, -1)) |
| 2560 |
return a.transpose(perm=axes) |
| 2561 |
|
| 2562 |
|
| 2563 |
def paddle_real(x): |
| 2564 |
# paddle doesn't support calling real on real arrays |
| 2565 |
try: |
| 2566 |
if x.is_complex(): |
| 2567 |
return x.real() |
| 2568 |
except AttributeError: |
| 2569 |
pass |
| 2570 |
return x |
| 2571 |
|
| 2572 |
|
| 2573 |
def paddle_imag(x): |
| 2574 |
# paddle doesn't support calling imag on real arrays |
| 2575 |
try: |
| 2576 |
if x.is_complex(): |
| 2577 |
return x.imag() |
| 2578 |
except AttributeError: |
| 2579 |
pass |
| 2580 |
return do("zeros_like", x) |
| 2581 |
|
| 2582 |
|
| 2583 |
def paddle_indices(dimensions): |
| 2584 |
_meshgrid = get_lib_fn("paddle", "meshgrid") |
| 2585 |
_arange = get_lib_fn("paddle", "arange") |
| 2586 |
return _meshgrid(*map(_arange, dimensions), indexing="ij") |
| 2587 |
|
| 2588 |
|
| 2589 |
def paddle_ravel(x): |
| 2590 |
return x.reshape((-1,)) |
| 2591 |
|
| 2592 |
|
| 2593 |
def paddle_pad(array, pad_width, mode="constant", constant_values=0): |
| 2594 |
if mode != "constant": |
| 2595 |
raise NotImplementedError |
| 2596 |
|
| 2597 |
try: |
| 2598 |
# numpy takes pads like ((0, 0), (1, 1), ... (n-1, n-1)) |
| 2599 |
# paddle takes pads like (0, 0, 1, 1, 2, 2, ...) |
| 2600 |
pad = tuple(itertools.chain.from_iterable(pad_width)) |
| 2601 |
|
| 2602 |
# a single tuple was specified ((a, b),) - use for all axes |
| 2603 |
if len(pad) == 2: |
| 2604 |
pad = pad * array.ndim |
| 2605 |
|
| 2606 |
except TypeError: |
| 2607 |
# assume int |
| 2608 |
pad = (pad_width,) * 2 * array.ndim |
| 2609 |
|
| 2610 |
return do( |
| 2611 |
"nn.functional.pad", |
| 2612 |
array, |
| 2613 |
pad=pad, |
| 2614 |
mode=mode, |
| 2615 |
value=constant_values, |
| 2616 |
like="paddle", |
| 2617 |
) |
| 2618 |
|
| 2619 |
|
| 2620 |
def paddle_wrap_reduction(fn): |
| 2621 |
def numpy_like(*args, **kwargs): |
| 2622 |
keepdims = kwargs.pop("keepdims", None) |
| 2623 |
if keepdims is not None: |
| 2624 |
kwargs["keepdim"] = keepdims |
| 2625 |
return fn(*args, **kwargs) |
| 2626 |
|
| 2627 |
return numpy_like |
| 2628 |
|
| 2629 |
|
| 2630 |
def paddle_split_wrap(fn): |
| 2631 |
# paddle doesn't seem to have `tensor_split always` |
| 2632 |
|
| 2633 |
@functools.wraps(fn) |
| 2634 |
def numpy_like(ary, indices_or_sections, axis=0, **kwargs): |
| 2635 |
if isinstance(indices_or_sections, int): |
| 2636 |
return fn(ary, indices_or_sections, axis=axis, **kwargs) |
| 2637 |
else: |
| 2638 |
diff = do( |
| 2639 |
"diff", |
| 2640 |
indices_or_sections, |
| 2641 |
prepend=0, |
| 2642 |
append=shape(ary)[axis], |
| 2643 |
like="numpy", |
| 2644 |
) |
| 2645 |
diff = list(diff) |
| 2646 |
return fn(ary, diff, axis=axis) |
| 2647 |
|
| 2648 |
return numpy_like |
| 2649 |
|
| 2650 |
|
| 2651 |
_MODULE_ALIASES["paddle[alt]"] = "paddle" |
| 2652 |
|
| 2653 |
_FUNCS["paddle", "to_numpy"] = paddle_to_numpy |
| 2654 |
_FUNCS["paddle", "transpose"] = paddle_transpose |
| 2655 |
_FUNCS["paddle", "real"] = paddle_real |
| 2656 |
_FUNCS["paddle", "imag"] = paddle_imag |
| 2657 |
_FUNCS["paddle", "indices"] = paddle_indices |
| 2658 |
_FUNCS["paddle", "ravel"] = paddle_ravel |
| 2659 |
_FUNCS["paddle", "pad"] = paddle_pad |
| 2660 |
|
| 2661 |
_FUNC_ALIASES["paddle", "random.normal"] = "randn" |
| 2662 |
_FUNC_ALIASES["paddle", "random.uniform"] = "rand" |
| 2663 |
_FUNC_ALIASES["paddle", "asarray"] = "to_tensor" |
| 2664 |
_FUNC_ALIASES["paddle", "concatenate"] = "concat" |
| 2665 |
_FUNC_ALIASES["paddle", "power"] = "pow" |
| 2666 |
_FUNC_ALIASES["paddle", "identity"] = "eye" |
| 2667 |
_FUNC_ALIASES["paddle", "split"] = "tensor_split" |
| 2668 |
|
| 2669 |
_SUBMODULE_ALIASES["paddle", "random.normal"] = "paddle" |
| 2670 |
_SUBMODULE_ALIASES["paddle", "random.uniform"] = "paddle" |
| 2671 |
|
| 2672 |
_CUSTOM_WRAPPERS["paddle", "random.normal"] = scale_random_normal_manually |
| 2673 |
_CUSTOM_WRAPPERS["paddle", "random.uniform"] = scale_random_uniform_manually |
| 2674 |
_CUSTOM_WRAPPERS["paddle[alt]", "split"] = paddle_split_wrap |
| 2675 |
_CUSTOM_WRAPPERS["paddle", "tril"] = make_translator( |
| 2676 |
[("m", ("x",)), ("k", ("diagonal", 0))] |
| 2677 |
) |
| 2678 |
_CUSTOM_WRAPPERS["paddle", "triu"] = make_translator( |
| 2679 |
[("m", ("x",)), ("k", ("diagonal", 0))] |
| 2680 |
) |
| 2681 |
for f in ("sum", "max", "min", "prod", "mean", "std", "var"): |
| 2682 |
_CUSTOM_WRAPPERS["paddle", f] = paddle_wrap_reduction |
| 2683 |
|
| 2684 |
|
| 2685 |
# -------------------------------- pytensor --------------------------------- # |
| 2686 |
|
| 2687 |
|
| 2688 |
@shape.register("pytensor") |
| 2689 |
def pytensor_shape(x): |
| 2690 |
return x.type.shape |
| 2691 |
|
| 2692 |
|
| 2693 |
def pytensor_wrap_qr_with_shapes(fn): |
| 2694 |
import pytensor.tensor as pt |
| 2695 |
|
| 2696 |
@functools.wraps(fn) |
| 2697 |
def qr_shaped(x, **kwargs): |
| 2698 |
*b, m, n = x.type.shape |
| 2699 |
k = min(m, n) |
| 2700 |
q, r = fn(x, mode="economic", **kwargs) |
| 2701 |
q = pt.specify_shape(q, (*b, m, k)) |
| 2702 |
r = pt.specify_shape(r, (*b, k, n)) |
| 2703 |
return q, r |
| 2704 |
|
| 2705 |
return qr_shaped |
| 2706 |
|
| 2707 |
|
| 2708 |
def pytensor_wrap_svd_with_shapes(fn): |
| 2709 |
import pytensor.tensor as pt |
| 2710 |
|
| 2711 |
@functools.wraps(fn) |
| 2712 |
def svd_shaped(x, full_matrices=False, **kwargs): |
| 2713 |
*b, m, n = x.type.shape |
| 2714 |
|
| 2715 |
k = min(m, n) |
| 2716 |
if full_matrices: |
| 2717 |
u_shape = (*b, m, m) |
| 2718 |
vh_shape = (*b, n, n) |
| 2719 |
else: |
| 2720 |
u_shape = (*b, m, k) |
| 2721 |
vh_shape = (*b, k, n) |
| 2722 |
|
| 2723 |
u, s, vh = fn(x, full_matrices=full_matrices, **kwargs) |
| 2724 |
u = pt.specify_shape(u, u_shape) |
| 2725 |
s = pt.specify_shape(s, (*b, k)) |
| 2726 |
vh = pt.specify_shape(vh, vh_shape) |
| 2727 |
return u, s, vh |
| 2728 |
|
| 2729 |
return svd_shaped |
| 2730 |
|
| 2731 |
|
| 2732 |
_MODULE_ALIASES["pytensor"] = "pytensor.tensor" |
| 2733 |
_CUSTOM_WRAPPERS["pytensor", "linalg.qr"] = pytensor_wrap_qr_with_shapes |
| 2734 |
_CUSTOM_WRAPPERS["pytensor", "linalg.svd"] = pytensor_wrap_svd_with_shapes |