Spaces:
Paused
Paused
File size: 16,722 Bytes
f1ea451 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 |
import math
import torch
from torch import nn
from torch.nn import functional as F
import numpy as np
def fused_leaky_relu(input, bias, negative_slope=0.2, scale=2 ** 0.5):
return F.leaky_relu(input + bias, negative_slope) * scale
class FusedLeakyReLU(nn.Module):
def __init__(self, channel, negative_slope=0.2, scale=2 ** 0.5):
super().__init__()
self.bias = nn.Parameter(torch.zeros(1, channel, 1, 1))
self.negative_slope = negative_slope
self.scale = scale
def forward(self, input):
out = fused_leaky_relu(input, self.bias, self.negative_slope, self.scale)
return out
def upfirdn2d_native(input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1):
_, minor, in_h, in_w = input.shape
kernel_h, kernel_w = kernel.shape
out = input.view(-1, minor, in_h, 1, in_w, 1)
out = F.pad(out, [0, up_x - 1, 0, 0, 0, up_y - 1, 0, 0])
out = out.view(-1, minor, in_h * up_y, in_w * up_x)
out = F.pad(out, [max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)])
out = out[:, :, max(-pad_y0, 0): out.shape[2] - max(-pad_y1, 0),
max(-pad_x0, 0): out.shape[3] - max(-pad_x1, 0), ]
out = out.reshape([-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1])
w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
out = F.conv2d(out, w)
out = out.reshape(-1, minor, in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1, )
return out[:, :, ::down_y, ::down_x]
def upfirdn2d(input, kernel, up=1, down=1, pad=(0, 0)):
return upfirdn2d_native(input, kernel, up, up, down, down, pad[0], pad[1], pad[0], pad[1])
class PixelNorm(nn.Module):
def __init__(self):
super().__init__()
def forward(self, input):
return input * torch.rsqrt(torch.mean(input ** 2, dim=1, keepdim=True) + 1e-8)
class MotionPixelNorm(nn.Module):
def __init__(self):
super().__init__()
def forward(self, input):
return input * torch.rsqrt(torch.mean(input ** 2, dim=2, keepdim=True) + 1e-8)
def make_kernel(k):
k = torch.tensor(k, dtype=torch.float32)
if k.ndim == 1:
k = k[None, :] * k[:, None]
k /= k.sum()
return k
class Upsample(nn.Module):
def __init__(self, kernel, factor=2):
super().__init__()
self.factor = factor
kernel = make_kernel(kernel) * (factor ** 2)
self.register_buffer('kernel', kernel)
p = kernel.shape[0] - factor
pad0 = (p + 1) // 2 + factor - 1
pad1 = p // 2
self.pad = (pad0, pad1)
def forward(self, input):
return upfirdn2d(input, self.kernel, up=self.factor, down=1, pad=self.pad)
class Downsample(nn.Module):
def __init__(self, kernel, factor=2):
super().__init__()
self.factor = factor
kernel = make_kernel(kernel)
self.register_buffer('kernel', kernel)
p = kernel.shape[0] - factor
pad0 = (p + 1) // 2
pad1 = p // 2
self.pad = (pad0, pad1)
def forward(self, input):
return upfirdn2d(input, self.kernel, up=1, down=self.factor, pad=self.pad)
class Blur(nn.Module):
def __init__(self, kernel, pad, upsample_factor=1):
super().__init__()
kernel = make_kernel(kernel)
if upsample_factor > 1:
kernel = kernel * (upsample_factor ** 2)
self.register_buffer('kernel', kernel)
self.pad = pad
def forward(self, input):
return upfirdn2d(input, self.kernel, pad=self.pad)
class EqualConv2d(nn.Module):
def __init__(self, in_channel, out_channel, kernel_size, stride=1, padding=0, bias=True):
super().__init__()
self.weight = nn.Parameter(torch.randn(out_channel, in_channel, kernel_size, kernel_size))
self.scale = 1 / math.sqrt(in_channel * kernel_size ** 2)
self.stride = stride
self.padding = padding
if bias:
self.bias = nn.Parameter(torch.zeros(out_channel))
else:
self.bias = None
def forward(self, input):
return F.conv2d(input, self.weight * self.scale, bias=self.bias, stride=self.stride, padding=self.padding, )
def __repr__(self):
return (
f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]},'
f' {self.weight.shape[2]}, stride={self.stride}, padding={self.padding})'
)
class EqualLinear(nn.Module):
def __init__(self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None):
super().__init__()
self.weight = nn.Parameter(torch.randn(out_dim, in_dim).div_(lr_mul))
if bias:
self.bias = nn.Parameter(torch.zeros(out_dim).fill_(bias_init))
else:
self.bias = None
self.activation = activation
self.scale = (1 / math.sqrt(in_dim)) * lr_mul
self.lr_mul = lr_mul
def forward(self, input):
if self.activation:
out = F.linear(input, self.weight * self.scale)
out = fused_leaky_relu(out, self.bias * self.lr_mul)
else:
out = F.linear(input, self.weight * self.scale, bias=self.bias * self.lr_mul)
return out
def __repr__(self):
return (f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]})')
class ScaledLeakyReLU(nn.Module):
def __init__(self, negative_slope=0.2):
super().__init__()
self.negative_slope = negative_slope
def forward(self, input):
return F.leaky_relu(input, negative_slope=self.negative_slope)
class ModulatedConv2d(nn.Module):
def __init__(self, in_channel, out_channel, kernel_size, style_dim, demodulate=True, upsample=False,
downsample=False, blur_kernel=[1, 3, 3, 1], ):
super().__init__()
self.eps = 1e-8
self.kernel_size = kernel_size
self.in_channel = in_channel
self.out_channel = out_channel
self.upsample = upsample
self.downsample = downsample
if upsample:
factor = 2
p = (len(blur_kernel) - factor) - (kernel_size - 1)
pad0 = (p + 1) // 2 + factor - 1
pad1 = p // 2 + 1
self.blur = Blur(blur_kernel, pad=(pad0, pad1), upsample_factor=factor)
if downsample:
factor = 2
p = (len(blur_kernel) - factor) + (kernel_size - 1)
pad0 = (p + 1) // 2
pad1 = p // 2
self.blur = Blur(blur_kernel, pad=(pad0, pad1))
fan_in = in_channel * kernel_size ** 2
self.scale = 1 / math.sqrt(fan_in)
self.padding = kernel_size // 2
self.weight = nn.Parameter(torch.randn(1, out_channel, in_channel, kernel_size, kernel_size))
self.modulation = EqualLinear(style_dim, in_channel, bias_init=1)
self.demodulate = demodulate
def __repr__(self):
return (
f'{self.__class__.__name__}({self.in_channel}, {self.out_channel}, {self.kernel_size}, '
f'upsample={self.upsample}, downsample={self.downsample})'
)
def forward(self, input, style):
batch, in_channel, height, width = input.shape
style = self.modulation(style).view(batch, 1, in_channel, 1, 1)
weight = self.scale * self.weight * style
if self.demodulate:
demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + 1e-8)
weight = weight * demod.view(batch, self.out_channel, 1, 1, 1)
weight = weight.view(batch * self.out_channel, in_channel, self.kernel_size, self.kernel_size)
if self.upsample:
input = input.view(1, batch * in_channel, height, width)
weight = weight.view(batch, self.out_channel, in_channel, self.kernel_size, self.kernel_size)
weight = weight.transpose(1, 2).reshape(batch * in_channel, self.out_channel, self.kernel_size,
self.kernel_size)
out = F.conv_transpose2d(input, weight, padding=0, stride=2, groups=batch)
_, _, height, width = out.shape
out = out.view(batch, self.out_channel, height, width)
out = self.blur(out)
elif self.downsample:
input = self.blur(input)
_, _, height, width = input.shape
input = input.view(1, batch * in_channel, height, width)
out = F.conv2d(input, weight, padding=0, stride=2, groups=batch)
_, _, height, width = out.shape
out = out.view(batch, self.out_channel, height, width)
else:
input = input.view(1, batch * in_channel, height, width)
out = F.conv2d(input, weight, padding=self.padding, groups=batch)
_, _, height, width = out.shape
out = out.view(batch, self.out_channel, height, width)
return out
class NoiseInjection(nn.Module):
def __init__(self):
super().__init__()
self.weight = nn.Parameter(torch.zeros(1))
def forward(self, image, noise=None):
if noise is None:
return image
else:
return image + self.weight * noise
class ConstantInput(nn.Module):
def __init__(self, channel, size=4):
super().__init__()
self.input = nn.Parameter(torch.randn(1, channel, size, size))
def forward(self, input):
batch = input.shape[0]
out = self.input.repeat(batch, 1, 1, 1)
return out
class StyledConv(nn.Module):
def __init__(self, in_channel, out_channel, kernel_size, style_dim, upsample=False, blur_kernel=[1, 3, 3, 1],
demodulate=True):
super().__init__()
self.conv = ModulatedConv2d(
in_channel,
out_channel,
kernel_size,
style_dim,
upsample=upsample,
blur_kernel=blur_kernel,
demodulate=demodulate,
)
self.noise = NoiseInjection()
self.activate = FusedLeakyReLU(out_channel)
def forward(self, input, style, noise=None):
out = self.conv(input, style)
out = self.noise(out, noise=noise)
out = self.activate(out)
return out
class ConvLayer(nn.Sequential):
def __init__(
self,
in_channel,
out_channel,
kernel_size,
downsample=False,
blur_kernel=[1, 3, 3, 1],
bias=True,
activate=True,
):
layers = []
if downsample:
factor = 2
p = (len(blur_kernel) - factor) + (kernel_size - 1)
pad0 = (p + 1) // 2
pad1 = p // 2
layers.append(Blur(blur_kernel, pad=(pad0, pad1)))
stride = 2
self.padding = 0
else:
stride = 1
self.padding = kernel_size // 2
layers.append(EqualConv2d(in_channel, out_channel, kernel_size, padding=self.padding, stride=stride,
bias=bias and not activate))
if activate:
if bias:
layers.append(FusedLeakyReLU(out_channel))
else:
layers.append(ScaledLeakyReLU(0.2))
super().__init__(*layers)
class ToRGB(nn.Module):
def __init__(self, in_channel, style_dim, upsample=True, blur_kernel=[1, 3, 3, 1]):
super().__init__()
if upsample:
self.upsample = Upsample(blur_kernel)
self.conv = ConvLayer(in_channel, 3, 1)
self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
def forward(self, input, skip=None):
out = self.conv(input)
out = out + self.bias
if skip is not None:
skip = self.upsample(skip)
out = out + skip
return out
class ToFlow(nn.Module):
def __init__(self, in_channel, style_dim, upsample=True, blur_kernel=[1, 3, 3, 1]):
super().__init__()
if upsample:
self.upsample = Upsample(blur_kernel)
self.style_dim = style_dim
self.in_channel = in_channel
self.conv = ModulatedConv2d(in_channel, 3, 1, style_dim, demodulate=False)
self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
def forward(self, input, style, feat, skip=None): # input 是来自上一层的 feature, style 是 512 的 condition, feat 是来自于 unet 的跳层
out = self.conv(input, style)
out = out + self.bias
# warping
xs = np.linspace(-1, 1, input.size(2))
xs = np.meshgrid(xs, xs)
xs = np.stack(xs, 2)
xs = torch.tensor(xs, requires_grad=False).float().unsqueeze(0).repeat(input.size(0), 1, 1, 1).to(input.device)
# import pdb;pdb.set_trace()
if skip is not None:
skip = self.upsample(skip)
out = out + skip
sampler = torch.tanh(out[:, 0:2, :, :])
mask = torch.sigmoid(out[:, 2:3, :, :])
flow = sampler.permute(0, 2, 3, 1) + xs # xs在这里相当于一个 location 的位置
feat_warp = F.grid_sample(feat, flow) * mask
# import pdb;pdb.set_trace()
return feat_warp, feat_warp + input * (1.0 - mask), out
class Direction(nn.Module):
def __init__(self, motion_dim):
super(Direction, self).__init__()
self.weight = nn.Parameter(torch.randn(512, motion_dim))
def forward(self, input):
# input: (bs*t) x 512
weight = self.weight + 1e-8
Q, R = torch.qr(weight) # get eignvector, orthogonal [n1, n2, n3, n4]
if input is None:
return Q
else:
input_diag = torch.diag_embed(input) # alpha, diagonal matrix
out = torch.matmul(input_diag, Q.T)
out = torch.sum(out, dim=1)
return out
class Synthesis(nn.Module):
def __init__(self, size, style_dim, motion_dim, blur_kernel=[1, 3, 3, 1], channel_multiplier=1):
super(Synthesis, self).__init__()
self.size = size
self.style_dim = style_dim
self.motion_dim = motion_dim
self.direction = Direction(motion_dim) # Linear Motion Decomposition (LMD) from LIA
self.channels = {
4: 512,
8: 512,
16: 512,
32: 512,
64: 256 * channel_multiplier,
128: 128 * channel_multiplier,
256: 64 * channel_multiplier,
512: 32 * channel_multiplier,
1024: 16 * channel_multiplier,
}
self.input = ConstantInput(self.channels[4])
self.conv1 = StyledConv(self.channels[4], self.channels[4], 3, style_dim, blur_kernel=blur_kernel)
self.to_rgb1 = ToRGB(self.channels[4], style_dim, upsample=False)
self.log_size = int(math.log(size, 2))
self.num_layers = (self.log_size - 2) * 2 + 1
self.convs = nn.ModuleList()
self.upsamples = nn.ModuleList()
self.to_rgbs = nn.ModuleList()
self.to_flows = nn.ModuleList()
in_channel = self.channels[4]
for i in range(3, self.log_size + 1):
out_channel = self.channels[2 ** i]
self.convs.append(StyledConv(in_channel, out_channel, 3, style_dim, upsample=True,
blur_kernel=blur_kernel))
self.convs.append(StyledConv(out_channel, out_channel, 3, style_dim, blur_kernel=blur_kernel))
self.to_rgbs.append(ToRGB(out_channel, style_dim))
self.to_flows.append(ToFlow(out_channel, style_dim))
in_channel = out_channel
self.n_latent = self.log_size * 2 - 2
def forward(self, source_before_decoupling, target_motion, feats):
directions = self.direction(target_motion)
latent = source_before_decoupling + directions # wa + directions
inject_index = self.n_latent
latent = latent.unsqueeze(1).repeat(1, inject_index, 1)
out = self.input(latent)
out = self.conv1(out, latent[:, 0])
i = 1
for conv1, conv2, to_rgb, to_flow, feat in zip(self.convs[::2], self.convs[1::2], self.to_rgbs,
self.to_flows, feats):
out = conv1(out, latent[:, i])
out = conv2(out, latent[:, i + 1])
if out.size(2) == 8:
out_warp, out, skip_flow = to_flow(out, latent[:, i + 2], feat)
skip = to_rgb(out_warp)
else:
out_warp, out, skip_flow = to_flow(out, latent[:, i + 2], feat, skip_flow)
skip = to_rgb(out_warp, skip)
i += 2
img = skip
return img
|