File size: 16,506 Bytes
f12ab4c |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 |
# SPDX-FileCopyrightText: Copyright (c) 2021-2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: LicenseRef-NvidiaProprietary
#
# NVIDIA CORPORATION, its affiliates and licensors retain all intellectual
# property and proprietary rights in and to this material, related
# documentation and any modifications thereto. Any use, reproduction,
# disclosure or distribution of this material and related documentation
# without an express license agreement from NVIDIA CORPORATION or
# its affiliates is strictly prohibited.
"""Custom PyTorch ops for efficient resampling of 2D images."""
import os
import numpy as np
import torch
from .. import custom_ops
from .. import misc
from . import conv2d_gradfix
#----------------------------------------------------------------------------
_plugin = None
def _init():
global _plugin
if _plugin is None:
_plugin = custom_ops.get_plugin(
module_name='upfirdn2d_plugin',
sources=['upfirdn2d.cpp', 'upfirdn2d.cu'],
headers=['upfirdn2d.h'],
source_dir=os.path.dirname(__file__),
extra_cuda_cflags=['--use_fast_math'],
)
return True
def _parse_scaling(scaling):
if isinstance(scaling, int):
scaling = [scaling, scaling]
assert isinstance(scaling, (list, tuple))
assert all(isinstance(x, int) for x in scaling)
sx, sy = scaling
assert sx >= 1 and sy >= 1
return sx, sy
def _parse_padding(padding):
if isinstance(padding, int):
padding = [padding, padding]
assert isinstance(padding, (list, tuple))
assert all(isinstance(x, int) for x in padding)
if len(padding) == 2:
padx, pady = padding
padding = [padx, padx, pady, pady]
padx0, padx1, pady0, pady1 = padding
return padx0, padx1, pady0, pady1
def _get_filter_size(f):
if f is None:
return 1, 1
assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
fw = f.shape[-1]
fh = f.shape[0]
with misc.suppress_tracer_warnings():
fw = int(fw)
fh = int(fh)
misc.assert_shape(f, [fh, fw][:f.ndim])
assert fw >= 1 and fh >= 1
return fw, fh
#----------------------------------------------------------------------------
def setup_filter(f, device=torch.device('cpu'), normalize=True, flip_filter=False, gain=1, separable=None):
r"""Convenience function to setup 2D FIR filter for `upfirdn2d()`.
Args:
f: Torch tensor, numpy array, or python list of the shape
`[filter_height, filter_width]` (non-separable),
`[filter_taps]` (separable),
`[]` (impulse), or
`None` (identity).
device: Result device (default: cpu).
normalize: Normalize the filter so that it retains the magnitude
for constant input signal (DC)? (default: True).
flip_filter: Flip the filter? (default: False).
gain: Overall scaling factor for signal magnitude (default: 1).
separable: Return a separable filter? (default: select automatically).
Returns:
Float32 tensor of the shape
`[filter_height, filter_width]` (non-separable) or
`[filter_taps]` (separable).
"""
# Validate.
if f is None:
f = 1
f = torch.as_tensor(f, dtype=torch.float32)
assert f.ndim in [0, 1, 2]
assert f.numel() > 0
if f.ndim == 0:
f = f[np.newaxis]
# Separable?
if separable is None:
separable = (f.ndim == 1 and f.numel() >= 8)
if f.ndim == 1 and not separable:
f = f.ger(f)
assert f.ndim == (1 if separable else 2)
# Apply normalize, flip, gain, and device.
if normalize:
f /= f.sum()
if flip_filter:
f = f.flip(list(range(f.ndim)))
f = f * (gain ** (f.ndim / 2))
f = f.to(device=device)
return f
#----------------------------------------------------------------------------
def upfirdn2d(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1, impl='cuda'):
r"""Pad, upsample, filter, and downsample a batch of 2D images.
Performs the following sequence of operations for each channel:
1. Upsample the image by inserting N-1 zeros after each pixel (`up`).
2. Pad the image with the specified number of zeros on each side (`padding`).
Negative padding corresponds to cropping the image.
3. Convolve the image with the specified 2D FIR filter (`f`), shrinking it
so that the footprint of all output pixels lies within the input image.
4. Downsample the image by keeping every Nth pixel (`down`).
This sequence of operations bears close resemblance to scipy.signal.upfirdn().
The fused op is considerably more efficient than performing the same calculation
using standard PyTorch ops. It supports gradients of arbitrary order.
Args:
x: Float32/float64/float16 input tensor of the shape
`[batch_size, num_channels, in_height, in_width]`.
f: Float32 FIR filter of the shape
`[filter_height, filter_width]` (non-separable),
`[filter_taps]` (separable), or
`None` (identity).
up: Integer upsampling factor. Can be a single int or a list/tuple
`[x, y]` (default: 1).
down: Integer downsampling factor. Can be a single int or a list/tuple
`[x, y]` (default: 1).
padding: Padding with respect to the upsampled image. Can be a single number
or a list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
(default: 0).
flip_filter: False = convolution, True = correlation (default: False).
gain: Overall scaling factor for signal magnitude (default: 1).
impl: Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
Returns:
Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
"""
assert isinstance(x, torch.Tensor)
assert impl in ['ref', 'cuda']
if impl == 'cuda' and x.device.type == 'cuda' and _init():
return _upfirdn2d_cuda(up=up, down=down, padding=padding, flip_filter=flip_filter, gain=gain).apply(x, f)
return _upfirdn2d_ref(x, f, up=up, down=down, padding=padding, flip_filter=flip_filter, gain=gain)
#----------------------------------------------------------------------------
@misc.profiled_function
def _upfirdn2d_ref(x, f, up=1, down=1, padding=0, flip_filter=False, gain=1):
"""Slow reference implementation of `upfirdn2d()` using standard PyTorch ops.
"""
# Validate arguments.
assert isinstance(x, torch.Tensor) and x.ndim == 4
if f is None:
f = torch.ones([1, 1], dtype=torch.float32, device=x.device)
assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
assert f.dtype == torch.float32 and not f.requires_grad
batch_size, num_channels, in_height, in_width = x.shape
upx, upy = _parse_scaling(up)
downx, downy = _parse_scaling(down)
padx0, padx1, pady0, pady1 = _parse_padding(padding)
# Check that upsampled buffer is not smaller than the filter.
upW = in_width * upx + padx0 + padx1
upH = in_height * upy + pady0 + pady1
assert upW >= f.shape[-1] and upH >= f.shape[0]
# Upsample by inserting zeros.
x = x.reshape([batch_size, num_channels, in_height, 1, in_width, 1])
x = torch.nn.functional.pad(x, [0, upx - 1, 0, 0, 0, upy - 1])
x = x.reshape([batch_size, num_channels, in_height * upy, in_width * upx])
# Pad or crop.
x = torch.nn.functional.pad(x, [max(padx0, 0), max(padx1, 0), max(pady0, 0), max(pady1, 0)])
x = x[:, :, max(-pady0, 0) : x.shape[2] - max(-pady1, 0), max(-padx0, 0) : x.shape[3] - max(-padx1, 0)]
# Setup filter.
f = f * (gain ** (f.ndim / 2))
f = f.to(x.dtype)
if not flip_filter:
f = f.flip(list(range(f.ndim)))
# Convolve with the filter.
f = f[np.newaxis, np.newaxis].repeat([num_channels, 1] + [1] * f.ndim)
if f.ndim == 4:
x = conv2d_gradfix.conv2d(input=x, weight=f, groups=num_channels)
else:
x = conv2d_gradfix.conv2d(input=x, weight=f.unsqueeze(2), groups=num_channels)
x = conv2d_gradfix.conv2d(input=x, weight=f.unsqueeze(3), groups=num_channels)
# Downsample by throwing away pixels.
x = x[:, :, ::downy, ::downx]
return x
#----------------------------------------------------------------------------
_upfirdn2d_cuda_cache = dict()
def _upfirdn2d_cuda(up=1, down=1, padding=0, flip_filter=False, gain=1):
"""Fast CUDA implementation of `upfirdn2d()` using custom ops.
"""
# Parse arguments.
upx, upy = _parse_scaling(up)
downx, downy = _parse_scaling(down)
padx0, padx1, pady0, pady1 = _parse_padding(padding)
# Lookup from cache.
key = (upx, upy, downx, downy, padx0, padx1, pady0, pady1, flip_filter, gain)
if key in _upfirdn2d_cuda_cache:
return _upfirdn2d_cuda_cache[key]
# Forward op.
class Upfirdn2dCuda(torch.autograd.Function):
@staticmethod
def forward(ctx, x, f): # pylint: disable=arguments-differ
assert isinstance(x, torch.Tensor) and x.ndim == 4
if f is None:
f = torch.ones([1, 1], dtype=torch.float32, device=x.device)
if f.ndim == 1 and f.shape[0] == 1:
f = f.square().unsqueeze(0) # Convert separable-1 into full-1x1.
assert isinstance(f, torch.Tensor) and f.ndim in [1, 2]
y = x
if f.ndim == 2:
y = _plugin.upfirdn2d(y, f, upx, upy, downx, downy, padx0, padx1, pady0, pady1, flip_filter, gain)
else:
y = _plugin.upfirdn2d(y, f.unsqueeze(0), upx, 1, downx, 1, padx0, padx1, 0, 0, flip_filter, 1.0)
y = _plugin.upfirdn2d(y, f.unsqueeze(1), 1, upy, 1, downy, 0, 0, pady0, pady1, flip_filter, gain)
ctx.save_for_backward(f)
ctx.x_shape = x.shape
return y
@staticmethod
def backward(ctx, dy): # pylint: disable=arguments-differ
f, = ctx.saved_tensors
_, _, ih, iw = ctx.x_shape
_, _, oh, ow = dy.shape
fw, fh = _get_filter_size(f)
p = [
fw - padx0 - 1,
iw * upx - ow * downx + padx0 - upx + 1,
fh - pady0 - 1,
ih * upy - oh * downy + pady0 - upy + 1,
]
dx = None
df = None
if ctx.needs_input_grad[0]:
dx = _upfirdn2d_cuda(up=down, down=up, padding=p, flip_filter=(not flip_filter), gain=gain).apply(dy, f)
assert not ctx.needs_input_grad[1]
return dx, df
# Add to cache.
_upfirdn2d_cuda_cache[key] = Upfirdn2dCuda
return Upfirdn2dCuda
#----------------------------------------------------------------------------
def filter2d(x, f, padding=0, flip_filter=False, gain=1, impl='cuda'):
r"""Filter a batch of 2D images using the given 2D FIR filter.
By default, the result is padded so that its shape matches the input.
User-specified padding is applied on top of that, with negative values
indicating cropping. Pixels outside the image are assumed to be zero.
Args:
x: Float32/float64/float16 input tensor of the shape
`[batch_size, num_channels, in_height, in_width]`.
f: Float32 FIR filter of the shape
`[filter_height, filter_width]` (non-separable),
`[filter_taps]` (separable), or
`None` (identity).
padding: Padding with respect to the output. Can be a single number or a
list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
(default: 0).
flip_filter: False = convolution, True = correlation (default: False).
gain: Overall scaling factor for signal magnitude (default: 1).
impl: Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
Returns:
Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
"""
padx0, padx1, pady0, pady1 = _parse_padding(padding)
fw, fh = _get_filter_size(f)
p = [
padx0 + fw // 2,
padx1 + (fw - 1) // 2,
pady0 + fh // 2,
pady1 + (fh - 1) // 2,
]
return upfirdn2d(x, f, padding=p, flip_filter=flip_filter, gain=gain, impl=impl)
#----------------------------------------------------------------------------
def upsample2d(x, f, up=2, padding=0, flip_filter=False, gain=1, impl='cuda'):
r"""Upsample a batch of 2D images using the given 2D FIR filter.
By default, the result is padded so that its shape is a multiple of the input.
User-specified padding is applied on top of that, with negative values
indicating cropping. Pixels outside the image are assumed to be zero.
Args:
x: Float32/float64/float16 input tensor of the shape
`[batch_size, num_channels, in_height, in_width]`.
f: Float32 FIR filter of the shape
`[filter_height, filter_width]` (non-separable),
`[filter_taps]` (separable), or
`None` (identity).
up: Integer upsampling factor. Can be a single int or a list/tuple
`[x, y]` (default: 1).
padding: Padding with respect to the output. Can be a single number or a
list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
(default: 0).
flip_filter: False = convolution, True = correlation (default: False).
gain: Overall scaling factor for signal magnitude (default: 1).
impl: Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
Returns:
Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
"""
upx, upy = _parse_scaling(up)
padx0, padx1, pady0, pady1 = _parse_padding(padding)
fw, fh = _get_filter_size(f)
p = [
padx0 + (fw + upx - 1) // 2,
padx1 + (fw - upx) // 2,
pady0 + (fh + upy - 1) // 2,
pady1 + (fh - upy) // 2,
]
return upfirdn2d(x, f, up=up, padding=p, flip_filter=flip_filter, gain=gain*upx*upy, impl=impl)
#----------------------------------------------------------------------------
def downsample2d(x, f, down=2, padding=0, flip_filter=False, gain=1, impl='cuda'):
r"""Downsample a batch of 2D images using the given 2D FIR filter.
By default, the result is padded so that its shape is a fraction of the input.
User-specified padding is applied on top of that, with negative values
indicating cropping. Pixels outside the image are assumed to be zero.
Args:
x: Float32/float64/float16 input tensor of the shape
`[batch_size, num_channels, in_height, in_width]`.
f: Float32 FIR filter of the shape
`[filter_height, filter_width]` (non-separable),
`[filter_taps]` (separable), or
`None` (identity).
down: Integer downsampling factor. Can be a single int or a list/tuple
`[x, y]` (default: 1).
padding: Padding with respect to the input. Can be a single number or a
list/tuple `[x, y]` or `[x_before, x_after, y_before, y_after]`
(default: 0).
flip_filter: False = convolution, True = correlation (default: False).
gain: Overall scaling factor for signal magnitude (default: 1).
impl: Implementation to use. Can be `'ref'` or `'cuda'` (default: `'cuda'`).
Returns:
Tensor of the shape `[batch_size, num_channels, out_height, out_width]`.
"""
downx, downy = _parse_scaling(down)
padx0, padx1, pady0, pady1 = _parse_padding(padding)
fw, fh = _get_filter_size(f)
p = [
padx0 + (fw - downx + 1) // 2,
padx1 + (fw - downx) // 2,
pady0 + (fh - downy + 1) // 2,
pady1 + (fh - downy) // 2,
]
return upfirdn2d(x, f, down=down, padding=p, flip_filter=flip_filter, gain=gain, impl=impl)
#----------------------------------------------------------------------------
|