Spaces:
Build error
Build error
File size: 13,662 Bytes
28c256d |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# Copyright (c) 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto. Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.
# source: https://github.com/NVlabs/stylegan3/blob/main/torch_utils/ops/conv2d_gradfix.py # noqa
"""Custom replacement for `torch.nn.functional.conv2d` that supports
arbitrarily high order gradients with zero performance penalty."""
import contextlib
import warnings
from typing import Dict, Optional, Tuple, Union
import torch
from mmengine.utils import digit_version
from mmengine.utils.dl_utils.parrots_wrapper import is_rocm_pytorch
enabled = True
weight_gradients_disabled = False
@contextlib.contextmanager
def no_weight_gradients(disable=True):
global weight_gradients_disabled
old = weight_gradients_disabled
if disable:
weight_gradients_disabled = True
yield
weight_gradients_disabled = old
def conv2d(input: torch.Tensor,
weight: torch.Tensor,
bias: Optional[torch.Tensor] = None,
stride: Union[int, Tuple[int, ...]] = 1,
padding: Union[int, Tuple[int, ...]] = 0,
dilation: Union[int, Tuple[int, ...]] = 1,
groups: int = 1):
flag = True
if digit_version(torch.__version__) >= digit_version('1.10.0'):
warnings.warn('Since '
'aten:cudnn_convolution_backward_weight is '
f'not supported in torch=={torch.__version__},'
' rolling back to `torch.nn.functional.conv2d`')
flag = False
if _should_use_custom_op(input) and flag:
return _conv2d_gradfix(
transpose=False,
weight_shape=weight.shape,
stride=stride,
padding=padding,
output_padding=0,
dilation=dilation,
groups=groups).apply(input, weight, bias)
return torch.nn.functional.conv2d(
input=input,
weight=weight,
bias=bias,
stride=stride,
padding=padding,
dilation=dilation,
groups=groups)
def conv_transpose2d(input: torch.Tensor,
weight: torch.Tensor,
bias: Optional[torch.Tensor] = None,
stride: Union[int, Tuple[int, ...]] = 1,
padding: Union[int, Tuple[int, ...]] = 0,
output_padding: Union[int, Tuple[int, ...]] = 0,
groups: int = 1,
dilation: Union[int, Tuple[int, ...]] = 1):
if _should_use_custom_op(input):
return _conv2d_gradfix(
transpose=True,
weight_shape=weight.shape,
stride=stride,
padding=padding,
output_padding=output_padding,
groups=groups,
dilation=dilation).apply(input, weight, bias)
return torch.nn.functional.conv_transpose2d(
input=input,
weight=weight,
bias=bias,
stride=stride,
padding=padding,
output_padding=output_padding,
groups=groups,
dilation=dilation)
def _should_use_custom_op(input):
assert isinstance(input, torch.Tensor)
if (not enabled) or (not torch.backends.cudnn.enabled):
return False
if input.device.type != 'cuda':
return False
return True
def _to_tuple(x, ndim):
xs = tuple(x) if isinstance(x, (tuple, list)) else (x, ) * ndim
assert len(xs) == ndim
assert all(isinstance(x, int) for x in xs)
return xs
_conv2d_gradfix_cache: Dict = dict()
_null_tensor = torch.empty([0])
def _conv2d_gradfix(
transpose: bool,
weight_shape: Tuple[int, ...],
stride: Union[int, Tuple[int, ...]],
padding: Union[int, Tuple[int, ...]],
output_padding: Union[int, Tuple[int, ...]],
dilation: Union[int, Tuple[int, ...]],
groups: int,
):
# Parse arguments.
ndim = 2
weight_shape = tuple(weight_shape)
stride = _to_tuple(stride, ndim)
padding = _to_tuple(padding, ndim)
output_padding = _to_tuple(output_padding, ndim)
dilation = _to_tuple(dilation, ndim)
# Lookup from cache.
key = (transpose, weight_shape, stride, padding, output_padding, dilation,
groups)
if key in _conv2d_gradfix_cache:
return _conv2d_gradfix_cache[key]
# Validate arguments.
assert groups >= 1
assert len(weight_shape) == ndim + 2
assert all(stride[i] >= 1 for i in range(ndim)) # type: ignore
assert all(padding[i] >= 0 for i in range(ndim)) # type: ignore
assert all(dilation[i] >= 0 for i in range(ndim)) # type: ignore
if not transpose:
assert all(output_padding[i] == 0 for i in range(ndim)) # type: ignore
else: # transpose
for i in range(ndim):
assert 0 <= output_padding[i] < max( # type: ignore
stride[i], # type: ignore
dilation[i]) # type: ignore
# Helpers.
common_kwargs = dict(
stride=stride, padding=padding, dilation=dilation, groups=groups)
def calc_output_padding(input_shape, output_shape):
if transpose:
return [0, 0]
return [
input_shape[i + 2] - (output_shape[i + 2] - 1) * stride[i] -
(1 - 2 * padding[i]) - dilation[i] * (weight_shape[i + 2] - 1)
for i in range(ndim)
]
# Forward & backward.
class Conv2d(torch.autograd.Function):
@staticmethod
def forward(ctx, input, weight, bias):
assert weight.shape == weight_shape
ctx.save_for_backward(
input if weight.requires_grad else _null_tensor,
weight if input.requires_grad else _null_tensor,
)
ctx.input_shape = input.shape
# Simple 1x1 convolution => cuBLAS (only on Volta, not on Ampere).
if weight_shape[2:] == stride == dilation == (
1, 1) and padding == (
0, 0) and torch.cuda.get_device_capability(
input.device) < (8, 0):
a = weight.reshape(groups, weight_shape[0] // groups,
weight_shape[1])
b = input.reshape(input.shape[0], groups,
input.shape[1] // groups, -1)
c = (a.transpose(1, 2) if transpose else a) @ b.permute(
1, 2, 0, 3).flatten(2)
c = c.reshape(-1, input.shape[0],
*input.shape[2:]).transpose(0, 1)
c = c if bias is None else c + bias.unsqueeze(0).unsqueeze(
2).unsqueeze(3)
return c.contiguous(
memory_format=(torch.channels_last if input.stride(1) ==
1 else torch.contiguous_format))
# General case => cuDNN.
if transpose:
return torch.nn.functional.conv_transpose2d(
input=input,
weight=weight,
bias=bias,
output_padding=output_padding,
**common_kwargs)
return torch.nn.functional.conv2d(
input=input, weight=weight, bias=bias, **common_kwargs)
@staticmethod
def backward(ctx, grad_output):
input, weight = ctx.saved_tensors
input_shape = ctx.input_shape
grad_input = None
grad_weight = None
grad_bias = None
if ctx.needs_input_grad[0]:
p = calc_output_padding(
input_shape=input_shape, output_shape=grad_output.shape)
op = _conv2d_gradfix(
transpose=(not transpose),
weight_shape=weight_shape,
output_padding=p,
**common_kwargs)
grad_input = op.apply(grad_output, weight, None)
assert grad_input.shape == input_shape
if ctx.needs_input_grad[1] and not weight_gradients_disabled:
grad_weight = Conv2dGradWeight.apply(grad_output, input)
assert grad_weight.shape == weight_shape
if ctx.needs_input_grad[2]:
grad_bias = grad_output.sum([0, 2, 3])
return grad_input, grad_weight, grad_bias
# Gradient with respect to the weights.
class Conv2dGradWeight(torch.autograd.Function):
@staticmethod
def forward(ctx, grad_output, input):
ctx.save_for_backward(
grad_output if input.requires_grad else _null_tensor,
input if grad_output.requires_grad else _null_tensor,
)
ctx.grad_output_shape = grad_output.shape
ctx.input_shape = input.shape
# Simple 1x1 convolution => cuBLAS (on both Volta and Ampere).
if weight_shape[2:] == stride == dilation == (
1, 1) and padding == (0, 0):
a = grad_output.reshape(grad_output.shape[0], groups,
grad_output.shape[1] // groups,
-1).permute(1, 2, 0, 3).flatten(2)
b = input.reshape(input.shape[0], groups,
input.shape[1] // groups,
-1).permute(1, 2, 0, 3).flatten(2)
c = (b @ a.transpose(1, 2) if transpose else
a @ b.transpose(1, 2)).reshape(weight_shape)
return c.contiguous(
memory_format=(torch.channels_last if input.stride(1) ==
1 else torch.contiguous_format))
# PyTorch consolidated convolution backward API in PR:
# https://github.com/pytorch/pytorch/commit/3dc3651e0ee3623f669c3a2c096408dbc476d122 # noqa: E501
# Enhance the code referring to the discussion:
# https://github.com/pytorch/pytorch/issues/74437
if digit_version(torch.__version__) >= digit_version('1.11.0'):
empty_weight = torch.tensor(
0.0, dtype=input.dtype,
device=input.device).expand(weight_shape)
output_padding = calc_output_padding(input.shape,
grad_output.shape)
return torch.ops.aten.convolution_backward(
grad_output,
input,
empty_weight,
None,
stride=stride,
dilation=dilation,
transposed=transpose,
padding=padding,
groups=groups,
output_padding=output_padding,
output_mask=[0, 1, 0])[1]
else:
if is_rocm_pytorch():
name = 'aten::miopen_convolution_transpose_backward_weight'
if not transpose:
name = 'aten::miopen_convolution_backward_weight'
flags = [
torch.backends.cudnn.benchmark,
torch.backends.cudnn.deterministic
]
else:
# General case => cuDNN.
name = ('aten::cudnn_convolution_transpose_backward_weight'
if transpose else
'aten::cudnn_convolution_backward_weight')
flags = [
torch.backends.cudnn.benchmark,
torch.backends.cudnn.deterministic,
torch.backends.cudnn.allow_tf32
]
return torch._C._jit_get_operation(name)(weight_shape,
grad_output, input,
padding, stride,
dilation, groups,
*flags)
@staticmethod
def backward(ctx, grad2_grad_weight):
grad_output, input = ctx.saved_tensors
grad_output_shape = ctx.grad_output_shape
input_shape = ctx.input_shape
grad2_grad_output = None
grad2_input = None
if ctx.needs_input_grad[0]:
grad2_grad_output = Conv2d.apply(input, grad2_grad_weight,
None)
assert grad2_grad_output.shape == grad_output_shape
if ctx.needs_input_grad[1]:
p = calc_output_padding(
input_shape=input_shape, output_shape=grad_output_shape)
op = _conv2d_gradfix(
transpose=(not transpose),
weight_shape=weight_shape,
output_padding=p,
**common_kwargs)
grad2_input = op.apply(grad_output, grad2_grad_weight, None)
assert grad2_input.shape == input_shape
return grad2_grad_output, grad2_input
_conv2d_gradfix_cache[key] = Conv2d
return Conv2d
|