File size: 19,214 Bytes
baa26ad |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 |
# Copyright (c) 2020, NVIDIA CORPORATION. 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.
"""Custom TensorFlow ops for efficient resampling of 2D images."""
import os
import numpy as np
import tensorflow as tf
from .. import custom_ops
def _get_plugin():
return custom_ops.get_plugin(os.path.splitext(__file__)[0] + '.cu')
#----------------------------------------------------------------------------
def upfirdn_2d(x, k, upx=1, upy=1, downx=1, downy=1, padx0=0, padx1=0, pady0=0, pady1=0, impl='cuda'):
r"""Pad, upsample, FIR filter, and downsample a batch of 2D images.
Accepts a batch of 2D images of the shape `[majorDim, inH, inW, minorDim]`
and performs the following operations for each image, batched across
`majorDim` and `minorDim`:
1. Upsample the image by inserting the zeros after each pixel (`upx`, `upy`).
2. Pad the image with zeros by the specified number of pixels on each side
(`padx0`, `padx1`, `pady0`, `pady1`). Specifying a negative value
corresponds to cropping the image.
3. Convolve the image with the specified 2D FIR filter (`k`), shrinking the
image so that the footprint of all output pixels lies within the input image.
4. Downsample the image by throwing away pixels (`downx`, `downy`).
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 TensorFlow ops. It supports gradients of arbitrary order.
Args:
x: Input tensor of the shape `[majorDim, inH, inW, minorDim]`.
k: 2D FIR filter of the shape `[firH, firW]`.
upx: Integer upsampling factor along the X-axis (default: 1).
upy: Integer upsampling factor along the Y-axis (default: 1).
downx: Integer downsampling factor along the X-axis (default: 1).
downy: Integer downsampling factor along the Y-axis (default: 1).
padx0: Number of pixels to pad on the left side (default: 0).
padx1: Number of pixels to pad on the right side (default: 0).
pady0: Number of pixels to pad on the top side (default: 0).
pady1: Number of pixels to pad on the bottom side (default: 0).
impl: Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
Returns:
Tensor of the shape `[majorDim, outH, outW, minorDim]`, and same datatype as `x`.
"""
impl_dict = {
'ref': _upfirdn_2d_ref,
'cuda': _upfirdn_2d_cuda,
}
return impl_dict[impl](x=x, k=k, upx=upx, upy=upy, downx=downx, downy=downy, padx0=padx0, padx1=padx1, pady0=pady0, pady1=pady1)
#----------------------------------------------------------------------------
def _upfirdn_2d_ref(x, k, upx, upy, downx, downy, padx0, padx1, pady0, pady1):
"""Slow reference implementation of `upfirdn_2d()` using standard TensorFlow ops."""
x = tf.convert_to_tensor(x)
k = np.asarray(k, dtype=np.float32)
assert x.shape.rank == 4
inH = x.shape[1].value
inW = x.shape[2].value
minorDim = _shape(x, 3)
kernelH, kernelW = k.shape
assert inW >= 1 and inH >= 1
assert kernelW >= 1 and kernelH >= 1
assert isinstance(upx, int) and isinstance(upy, int)
assert isinstance(downx, int) and isinstance(downy, int)
assert isinstance(padx0, int) and isinstance(padx1, int)
assert isinstance(pady0, int) and isinstance(pady1, int)
# Upsample (insert zeros).
x = tf.reshape(x, [-1, inH, 1, inW, 1, minorDim])
x = tf.pad(x, [[0, 0], [0, 0], [0, upy - 1], [0, 0], [0, upx - 1], [0, 0]])
x = tf.reshape(x, [-1, inH * upy, inW * upx, minorDim])
# Pad (crop if negative).
x = tf.pad(x, [[0, 0], [max(pady0, 0), max(pady1, 0)], [max(padx0, 0), max(padx1, 0)], [0, 0]])
x = x[:, max(-pady0, 0) : x.shape[1].value - max(-pady1, 0), max(-padx0, 0) : x.shape[2].value - max(-padx1, 0), :]
# Convolve with filter.
x = tf.transpose(x, [0, 3, 1, 2])
x = tf.reshape(x, [-1, 1, inH * upy + pady0 + pady1, inW * upx + padx0 + padx1])
w = tf.constant(k[::-1, ::-1, np.newaxis, np.newaxis], dtype=x.dtype)
x = tf.nn.conv2d(x, w, strides=[1,1,1,1], padding='VALID', data_format='NCHW')
x = tf.reshape(x, [-1, minorDim, inH * upy + pady0 + pady1 - kernelH + 1, inW * upx + padx0 + padx1 - kernelW + 1])
x = tf.transpose(x, [0, 2, 3, 1])
# Downsample (throw away pixels).
return x[:, ::downy, ::downx, :]
#----------------------------------------------------------------------------
def _upfirdn_2d_cuda(x, k, upx, upy, downx, downy, padx0, padx1, pady0, pady1):
"""Fast CUDA implementation of `upfirdn_2d()` using custom ops."""
x = tf.convert_to_tensor(x)
k = np.asarray(k, dtype=np.float32)
majorDim, inH, inW, minorDim = x.shape.as_list()
kernelH, kernelW = k.shape
assert inW >= 1 and inH >= 1
assert kernelW >= 1 and kernelH >= 1
assert isinstance(upx, int) and isinstance(upy, int)
assert isinstance(downx, int) and isinstance(downy, int)
assert isinstance(padx0, int) and isinstance(padx1, int)
assert isinstance(pady0, int) and isinstance(pady1, int)
outW = (inW * upx + padx0 + padx1 - kernelW) // downx + 1
outH = (inH * upy + pady0 + pady1 - kernelH) // downy + 1
assert outW >= 1 and outH >= 1
cuda_op = _get_plugin().up_fir_dn2d
kc = tf.constant(k, dtype=x.dtype)
gkc = tf.constant(k[::-1, ::-1], dtype=x.dtype)
gpadx0 = kernelW - padx0 - 1
gpady0 = kernelH - pady0 - 1
gpadx1 = inW * upx - outW * downx + padx0 - upx + 1
gpady1 = inH * upy - outH * downy + pady0 - upy + 1
@tf.custom_gradient
def func(x):
y = cuda_op(x=x, k=kc, upx=int(upx), upy=int(upy), downx=int(downx), downy=int(downy), padx0=int(padx0), padx1=int(padx1), pady0=int(pady0), pady1=int(pady1))
y.set_shape([majorDim, outH, outW, minorDim])
@tf.custom_gradient
def grad(dy):
dx = cuda_op(x=dy, k=gkc, upx=int(downx), upy=int(downy), downx=int(upx), downy=int(upy), padx0=int(gpadx0), padx1=int(gpadx1), pady0=int(gpady0), pady1=int(gpady1))
dx.set_shape([majorDim, inH, inW, minorDim])
return dx, func
return y, grad
return func(x)
#----------------------------------------------------------------------------
def filter_2d(x, k, gain=1, padding=0, data_format='NCHW', impl='cuda'):
r"""Filter a batch of 2D images with the given FIR filter.
Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
and filters each image with the given filter. The filter is normalized so that
if the input pixels are constant, they will be scaled by the specified `gain`.
Pixels outside the image are assumed to be zero.
Args:
x: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
k: FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
gain: Scaling factor for signal magnitude (default: 1.0).
padding: Number of pixels to pad or crop the output on each side (default: 0).
data_format: `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
impl: Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
Returns:
Tensor of the same shape and datatype as `x`.
"""
assert isinstance(padding, int)
k = _FilterKernel(k=k, gain=gain)
assert k.w == k.h
pad0 = k.w // 2 + padding
pad1 = (k.w - 1) // 2 + padding
return _simple_upfirdn_2d(x, k, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
#----------------------------------------------------------------------------
def upsample_2d(x, k=None, factor=2, gain=1, padding=0, data_format='NCHW', impl='cuda'):
r"""Upsample a batch of 2D images with the given filter.
Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
and upsamples each image with the given filter. The filter is normalized so that
if the input pixels are constant, they will be scaled by the specified `gain`.
Pixels outside the image are assumed to be zero, and the filter is padded with
zeros so that its shape is a multiple of the upsampling factor.
Args:
x: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
k: FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
The default is `[1] * factor`, which corresponds to nearest-neighbor
upsampling.
factor: Integer upsampling factor (default: 2).
gain: Scaling factor for signal magnitude (default: 1.0).
padding: Number of pixels to pad or crop the output on each side (default: 0).
data_format: `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
impl: Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
Returns:
Tensor of the shape `[N, C, H * factor, W * factor]` or
`[N, H * factor, W * factor, C]`, and same datatype as `x`.
"""
assert isinstance(factor, int) and factor >= 1
assert isinstance(padding, int)
k = _FilterKernel(k if k is not None else [1] * factor, gain * (factor ** 2))
assert k.w == k.h
pad0 = (k.w + factor - 1) // 2 + padding
pad1 = (k.w - factor) // 2 + padding
return _simple_upfirdn_2d(x, k, up=factor, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
#----------------------------------------------------------------------------
def downsample_2d(x, k=None, factor=2, gain=1, padding=0, data_format='NCHW', impl='cuda'):
r"""Downsample a batch of 2D images with the given filter.
Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]`
and downsamples each image with the given filter. The filter is normalized so that
if the input pixels are constant, they will be scaled by the specified `gain`.
Pixels outside the image are assumed to be zero, and the filter is padded with
zeros so that its shape is a multiple of the downsampling factor.
Args:
x: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
k: FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
The default is `[1] * factor`, which corresponds to average pooling.
factor: Integer downsampling factor (default: 2).
gain: Scaling factor for signal magnitude (default: 1.0).
padding: Number of pixels to pad or crop the output on each side (default: 0).
data_format: `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
impl: Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
Returns:
Tensor of the shape `[N, C, H // factor, W // factor]` or
`[N, H // factor, W // factor, C]`, and same datatype as `x`.
"""
assert isinstance(factor, int) and factor >= 1
assert isinstance(padding, int)
k = _FilterKernel(k if k is not None else [1] * factor, gain)
assert k.w == k.h
pad0 = (k.w - factor + 1) // 2 + padding * factor
pad1 = (k.w - factor) // 2 + padding * factor
return _simple_upfirdn_2d(x, k, down=factor, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
#----------------------------------------------------------------------------
def upsample_conv_2d(x, w, k=None, factor=2, gain=1, padding=0, data_format='NCHW', impl='cuda'):
r"""Fused `upsample_2d()` followed by `tf.nn.conv2d()`.
Padding is performed only once at the beginning, not between the operations.
The fused op is considerably more efficient than performing the same calculation
using standard TensorFlow ops. It supports gradients of arbitrary order.
Args:
x: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
w: Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`.
Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`.
k: FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
The default is `[1] * factor`, which corresponds to nearest-neighbor
upsampling.
factor: Integer upsampling factor (default: 2).
gain: Scaling factor for signal magnitude (default: 1.0).
padding: Number of pixels to pad or crop the output on each side (default: 0).
data_format: `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
impl: Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
Returns:
Tensor of the shape `[N, C, H * factor, W * factor]` or
`[N, H * factor, W * factor, C]`, and same datatype as `x`.
"""
assert isinstance(factor, int) and factor >= 1
assert isinstance(padding, int)
# Check weight shape.
w = tf.convert_to_tensor(w)
ch, cw, _inC, _outC = w.shape.as_list()
inC = _shape(w, 2)
outC = _shape(w, 3)
assert cw == ch
# Fast path for 1x1 convolution.
if cw == 1 and ch == 1:
x = tf.nn.conv2d(x, w, data_format=data_format, strides=[1,1,1,1], padding='VALID')
x = upsample_2d(x, k, factor=factor, gain=gain, padding=padding, data_format=data_format, impl=impl)
return x
# Setup filter kernel.
k = _FilterKernel(k if k is not None else [1] * factor, gain * (factor ** 2))
assert k.w == k.h
# Determine data dimensions.
if data_format == 'NCHW':
stride = [1, 1, factor, factor]
output_shape = [_shape(x, 0), outC, (_shape(x, 2) - 1) * factor + ch, (_shape(x, 3) - 1) * factor + cw]
num_groups = _shape(x, 1) // inC
else:
stride = [1, factor, factor, 1]
output_shape = [_shape(x, 0), (_shape(x, 1) - 1) * factor + ch, (_shape(x, 2) - 1) * factor + cw, outC]
num_groups = _shape(x, 3) // inC
# Transpose weights.
w = tf.reshape(w, [ch, cw, inC, num_groups, -1])
w = tf.transpose(w[::-1, ::-1], [0, 1, 4, 3, 2])
w = tf.reshape(w, [ch, cw, -1, num_groups * inC])
# Execute.
x = tf.nn.conv2d_transpose(x, w, output_shape=output_shape, strides=stride, padding='VALID', data_format=data_format)
pad0 = (k.w + factor - cw) // 2 + padding
pad1 = (k.w - factor - cw + 3) // 2 + padding
return _simple_upfirdn_2d(x, k, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
#----------------------------------------------------------------------------
def conv_downsample_2d(x, w, k=None, factor=2, gain=1, padding=0, data_format='NCHW', impl='cuda'):
r"""Fused `tf.nn.conv2d()` followed by `downsample_2d()`.
Padding is performed only once at the beginning, not between the operations.
The fused op is considerably more efficient than performing the same calculation
using standard TensorFlow ops. It supports gradients of arbitrary order.
Args:
x: Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
w: Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`.
Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`.
k: FIR filter of the shape `[firH, firW]` or `[firN]` (separable).
The default is `[1] * factor`, which corresponds to average pooling.
factor: Integer downsampling factor (default: 2).
gain: Scaling factor for signal magnitude (default: 1.0).
padding: Number of pixels to pad or crop the output on each side (default: 0).
data_format: `'NCHW'` or `'NHWC'` (default: `'NCHW'`).
impl: Name of the implementation to use. Can be `"ref"` or `"cuda"` (default).
Returns:
Tensor of the shape `[N, C, H // factor, W // factor]` or
`[N, H // factor, W // factor, C]`, and same datatype as `x`.
"""
assert isinstance(factor, int) and factor >= 1
assert isinstance(padding, int)
# Check weight shape.
w = tf.convert_to_tensor(w)
ch, cw, _inC, _outC = w.shape.as_list()
assert cw == ch
# Fast path for 1x1 convolution.
if cw == 1 and ch == 1:
x = downsample_2d(x, k, factor=factor, gain=gain, padding=padding, data_format=data_format, impl=impl)
x = tf.nn.conv2d(x, w, data_format=data_format, strides=[1,1,1,1], padding='VALID')
return x
# Setup filter kernel.
k = _FilterKernel(k if k is not None else [1] * factor, gain)
assert k.w == k.h
# Determine stride.
if data_format == 'NCHW':
s = [1, 1, factor, factor]
else:
s = [1, factor, factor, 1]
# Execute.
pad0 = (k.w - factor + cw) // 2 + padding * factor
pad1 = (k.w - factor + cw - 1) // 2 + padding * factor
x = _simple_upfirdn_2d(x, k, pad0=pad0, pad1=pad1, data_format=data_format, impl=impl)
return tf.nn.conv2d(x, w, strides=s, padding='VALID', data_format=data_format)
#----------------------------------------------------------------------------
# Internal helpers.
class _FilterKernel:
def __init__(self, k, gain=1):
k = np.asarray(k, dtype=np.float32)
k /= np.sum(k)
# Separable.
if k.ndim == 1 and k.size >= 8:
self.w = k.size
self.h = k.size
self.kx = k[np.newaxis, :]
self.ky = k[:, np.newaxis] * gain
self.kxy = None
# Non-separable.
else:
if k.ndim == 1:
k = np.outer(k, k)
assert k.ndim == 2
self.w = k.shape[1]
self.h = k.shape[0]
self.kx = None
self.ky = None
self.kxy = k * gain
def _simple_upfirdn_2d(x, k, up=1, down=1, pad0=0, pad1=0, data_format='NCHW', impl='cuda'):
assert isinstance(k, _FilterKernel)
assert data_format in ['NCHW', 'NHWC']
assert x.shape.rank == 4
y = x
if data_format == 'NCHW':
y = tf.reshape(y, [-1, _shape(y, 2), _shape(y, 3), 1])
if k.kx is not None:
y = upfirdn_2d(y, k.kx, upx=up, downx=down, padx0=pad0, padx1=pad1, impl=impl)
if k.ky is not None:
y = upfirdn_2d(y, k.ky, upy=up, downy=down, pady0=pad0, pady1=pad1, impl=impl)
if k.kxy is not None:
y = upfirdn_2d(y, k.kxy, upx=up, upy=up, downx=down, downy=down, padx0=pad0, padx1=pad1, pady0=pad0, pady1=pad1, impl=impl)
if data_format == 'NCHW':
y = tf.reshape(y, [-1, _shape(x, 1), _shape(y, 1), _shape(y, 2)])
return y
def _shape(tf_expr, dim_idx):
if tf_expr.shape.rank is not None:
dim = tf_expr.shape[dim_idx].value
if dim is not None:
return dim
return tf.shape(tf_expr)[dim_idx]
#----------------------------------------------------------------------------
|