File size: 12,198 Bytes
9c6594c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 |
"""
This module contains tensor creation utilities.
"""
import collections.abc
import functools
import math
import warnings
from typing import cast, Optional, Union
import torch
_INTEGRAL_TYPES = [
torch.uint8,
torch.int8,
torch.int16,
torch.int32,
torch.int64,
torch.uint16,
torch.uint32,
torch.uint64,
]
_FLOATING_TYPES = [torch.float16, torch.bfloat16, torch.float32, torch.float64]
_FLOATING_8BIT_TYPES = [
torch.float8_e4m3fn,
torch.float8_e5m2,
torch.float8_e4m3fnuz,
torch.float8_e5m2fnuz,
]
_COMPLEX_TYPES = [torch.complex32, torch.complex64, torch.complex128]
_BOOLEAN_OR_INTEGRAL_TYPES = [torch.bool, *_INTEGRAL_TYPES]
_FLOATING_OR_COMPLEX_TYPES = [*_FLOATING_TYPES, *_COMPLEX_TYPES]
def _uniform_random_(t: torch.Tensor, low: float, high: float) -> torch.Tensor:
# uniform_ requires to-from <= std::numeric_limits<scalar_t>::max()
# Work around this by scaling the range before and after the PRNG
if high - low >= torch.finfo(t.dtype).max:
return t.uniform_(low / 2, high / 2).mul_(2)
else:
return t.uniform_(low, high)
def make_tensor(
*shape: Union[int, torch.Size, list[int], tuple[int, ...]],
dtype: torch.dtype,
device: Union[str, torch.device],
low: Optional[float] = None,
high: Optional[float] = None,
requires_grad: bool = False,
noncontiguous: bool = False,
exclude_zero: bool = False,
memory_format: Optional[torch.memory_format] = None,
) -> torch.Tensor:
r"""Creates a tensor with the given :attr:`shape`, :attr:`device`, and :attr:`dtype`, and filled with
values uniformly drawn from ``[low, high)``.
If :attr:`low` or :attr:`high` are specified and are outside the range of the :attr:`dtype`'s representable
finite values then they are clamped to the lowest or highest representable finite value, respectively.
If ``None``, then the following table describes the default values for :attr:`low` and :attr:`high`,
which depend on :attr:`dtype`.
+---------------------------+------------+----------+
| ``dtype`` | ``low`` | ``high`` |
+===========================+============+==========+
| boolean type | ``0`` | ``2`` |
+---------------------------+------------+----------+
| unsigned integral type | ``0`` | ``10`` |
+---------------------------+------------+----------+
| signed integral types | ``-9`` | ``10`` |
+---------------------------+------------+----------+
| floating types | ``-9`` | ``9`` |
+---------------------------+------------+----------+
| complex types | ``-9`` | ``9`` |
+---------------------------+------------+----------+
Args:
shape (Tuple[int, ...]): Single integer or a sequence of integers defining the shape of the output tensor.
dtype (:class:`torch.dtype`): The data type of the returned tensor.
device (Union[str, torch.device]): The device of the returned tensor.
low (Optional[Number]): Sets the lower limit (inclusive) of the given range. If a number is provided it is
clamped to the least representable finite value of the given dtype. When ``None`` (default),
this value is determined based on the :attr:`dtype` (see the table above). Default: ``None``.
high (Optional[Number]): Sets the upper limit (exclusive) of the given range. If a number is provided it is
clamped to the greatest representable finite value of the given dtype. When ``None`` (default) this value
is determined based on the :attr:`dtype` (see the table above). Default: ``None``.
.. deprecated:: 2.1
Passing ``low==high`` to :func:`~torch.testing.make_tensor` for floating or complex types is deprecated
since 2.1 and will be removed in 2.3. Use :func:`torch.full` instead.
requires_grad (Optional[bool]): If autograd should record operations on the returned tensor. Default: ``False``.
noncontiguous (Optional[bool]): If `True`, the returned tensor will be noncontiguous. This argument is
ignored if the constructed tensor has fewer than two elements. Mutually exclusive with ``memory_format``.
exclude_zero (Optional[bool]): If ``True`` then zeros are replaced with the dtype's small positive value
depending on the :attr:`dtype`. For bool and integer types zero is replaced with one. For floating
point types it is replaced with the dtype's smallest positive normal number (the "tiny" value of the
:attr:`dtype`'s :func:`~torch.finfo` object), and for complex types it is replaced with a complex number
whose real and imaginary parts are both the smallest positive normal number representable by the complex
type. Default ``False``.
memory_format (Optional[torch.memory_format]): The memory format of the returned tensor. Mutually exclusive
with ``noncontiguous``.
Raises:
ValueError: If ``requires_grad=True`` is passed for integral `dtype`
ValueError: If ``low >= high``.
ValueError: If either :attr:`low` or :attr:`high` is ``nan``.
ValueError: If both :attr:`noncontiguous` and :attr:`memory_format` are passed.
TypeError: If :attr:`dtype` isn't supported by this function.
Examples:
>>> # xdoctest: +SKIP
>>> # xdoctest: +REQUIRES(env:TORCH_DOCTEST_CUDA)
>>> from torch.testing import make_tensor
>>> # Creates a float tensor with values in [-1, 1)
>>> make_tensor((3,), device='cpu', dtype=torch.float32, low=-1, high=1)
>>> # xdoctest: +SKIP
tensor([ 0.1205, 0.2282, -0.6380])
>>> # Creates a bool tensor on CUDA
>>> make_tensor((2, 2), device='cuda', dtype=torch.bool)
tensor([[False, False],
[False, True]], device='cuda:0')
"""
def modify_low_high(
low: Optional[float],
high: Optional[float],
*,
lowest_inclusive: float,
highest_exclusive: float,
default_low: float,
default_high: float,
) -> tuple[float, float]:
"""
Modifies (and raises ValueError when appropriate) low and high values given by the user (input_low, input_high)
if required.
"""
def clamp(a: float, l: float, h: float) -> float:
return min(max(a, l), h)
low = low if low is not None else default_low
high = high if high is not None else default_high
if any(isinstance(value, float) and math.isnan(value) for value in [low, high]):
raise ValueError(
f"`low` and `high` cannot be NaN, but got {low=} and {high=}"
)
elif low == high and dtype in _FLOATING_OR_COMPLEX_TYPES:
warnings.warn(
"Passing `low==high` to `torch.testing.make_tensor` for floating or complex types "
"is deprecated since 2.1 and will be removed in 2.3. "
"Use `torch.full(...)` instead.",
FutureWarning,
stacklevel=3,
)
elif low >= high:
raise ValueError(f"`low` must be less than `high`, but got {low} >= {high}")
elif high < lowest_inclusive or low >= highest_exclusive:
raise ValueError(
f"The value interval specified by `low` and `high` is [{low}, {high}), "
f"but {dtype} only supports [{lowest_inclusive}, {highest_exclusive})"
)
low = clamp(low, lowest_inclusive, highest_exclusive)
high = clamp(high, lowest_inclusive, highest_exclusive)
if dtype in _BOOLEAN_OR_INTEGRAL_TYPES:
# 1. `low` is ceiled to avoid creating values smaller than `low` and thus outside the specified interval
# 2. Following the same reasoning as for 1., `high` should be floored. However, the higher bound of
# `torch.randint` is exclusive, and thus we need to ceil here as well.
return math.ceil(low), math.ceil(high)
return low, high
if len(shape) == 1 and isinstance(shape[0], collections.abc.Sequence):
shape = shape[0] # type: ignore[assignment]
shape = cast(tuple[int, ...], tuple(shape))
if noncontiguous and memory_format is not None:
raise ValueError(
f"The parameters `noncontiguous` and `memory_format` are mutually exclusive, "
f"but got {noncontiguous=} and {memory_format=}"
)
if requires_grad and dtype in _BOOLEAN_OR_INTEGRAL_TYPES:
raise ValueError(
f"`requires_grad=True` is not supported for boolean and integral dtypes, but got {dtype=}"
)
noncontiguous = noncontiguous and functools.reduce(lambda x, y: x * y, shape, 1) > 1
if noncontiguous:
# Double the size of the shape in the last dimension, so that we have
# non-identical values when we make the non-contiguous operation.
shape = cast(tuple[int, ...], (*shape[:-1], 2 * shape[-1]))
if dtype is torch.bool:
low, high = cast(
tuple[int, int],
modify_low_high(
low,
high,
lowest_inclusive=0,
highest_exclusive=2,
default_low=0,
default_high=2,
),
)
result = torch.randint(low, high, shape, device=device, dtype=dtype)
elif dtype in _BOOLEAN_OR_INTEGRAL_TYPES:
low, high = cast(
tuple[int, int],
modify_low_high(
low,
high,
lowest_inclusive=torch.iinfo(dtype).min,
highest_exclusive=torch.iinfo(dtype).max
# In theory, `highest_exclusive` should always be the maximum value + 1. However, `torch.randint`
# internally converts the bounds to an int64 and would overflow. In other words: `torch.randint` cannot
# sample 2**63 - 1, i.e. the maximum value of `torch.int64` and we need to account for that here.
+ (1 if dtype is not torch.int64 else 0),
# This is incorrect for `torch.uint8`, but since we clamp to `lowest`, i.e. 0 for `torch.uint8`,
# _after_ we use the default value, we don't need to special case it here
default_low=-9,
default_high=10,
),
)
result = torch.randint(low, high, shape, device=device, dtype=dtype)
elif dtype in _FLOATING_OR_COMPLEX_TYPES:
low, high = modify_low_high(
low,
high,
lowest_inclusive=torch.finfo(dtype).min,
highest_exclusive=torch.finfo(dtype).max,
default_low=-9,
default_high=9,
)
result = torch.empty(shape, device=device, dtype=dtype)
_uniform_random_(
torch.view_as_real(result) if dtype in _COMPLEX_TYPES else result, low, high
)
elif dtype in _FLOATING_8BIT_TYPES:
low, high = modify_low_high(
low,
high,
lowest_inclusive=torch.finfo(dtype).min,
highest_exclusive=torch.finfo(dtype).max,
default_low=-9,
default_high=9,
)
result = torch.empty(shape, device=device, dtype=torch.float32)
_uniform_random_(result, low, high)
result = result.to(dtype)
else:
raise TypeError(
f"The requested dtype '{dtype}' is not supported by torch.testing.make_tensor()."
" To request support, file an issue at: https://github.com/pytorch/pytorch/issues"
)
if noncontiguous:
# Offset by 1 to also catch offsetting issues
result = result[..., 1::2]
elif memory_format is not None:
result = result.clone(memory_format=memory_format)
if exclude_zero:
result[result == 0] = (
1 if dtype in _BOOLEAN_OR_INTEGRAL_TYPES else torch.finfo(dtype).tiny
)
if dtype in _FLOATING_OR_COMPLEX_TYPES:
result.requires_grad = requires_grad
return result
|