Spaces:
Runtime error
Runtime error
File size: 43,239 Bytes
a6c9602 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 |
ο»Ώ# Copyright (c) SenseTime Research. All rights reserved.
# Copyright (c) 2021, 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.
import numpy as np
import torch
from torch_utils import misc
from torch_utils import persistence
from torch_utils.ops import conv2d_resample
from torch_utils.ops import upfirdn2d
from torch_utils.ops import bias_act
from torch_utils.ops import fma
# ----------------------------------------------------------------------------
@misc.profiled_function
def normalize_2nd_moment(x, dim=1, eps=1e-8):
return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()
# ----------------------------------------------------------------------------
@misc.profiled_function
def modulated_conv2d(
# Input tensor of shape [batch_size, in_channels, in_height, in_width].
x,
# Weight tensor of shape [out_channels, in_channels, kernel_height, kernel_width].
weight,
# Modulation coefficients of shape [batch_size, in_channels].
styles,
noise=None, # Optional noise tensor to add to the output activations.
up=1, # Integer upsampling factor.
down=1, # Integer downsampling factor.
padding=0, # Padding with respect to the upsampled image.
# Low-pass filter to apply when resampling activations. Must be prepared beforehand by calling upfirdn2d.setup_filter().
resample_filter=None,
demodulate=True, # Apply weight demodulation?
# False = convolution, True = correlation (matches torch.nn.functional.conv2d).
flip_weight=True,
# Perform modulation, convolution, and demodulation as a single fused operation?
fused_modconv=True,
):
batch_size = x.shape[0]
out_channels, in_channels, kh, kw = weight.shape
misc.assert_shape(weight, [out_channels, in_channels, kh, kw]) # [OIkk]
misc.assert_shape(x, [batch_size, in_channels, None, None]) # [NIHW]
misc.assert_shape(styles, [batch_size, in_channels]) # [NI]
# Pre-normalize inputs to avoid FP16 overflow.
if x.dtype == torch.float16 and demodulate:
weight = weight * (1 / np.sqrt(in_channels * kh * kw) /
weight.norm(float('inf'), dim=[1, 2, 3], keepdim=True)) # max_Ikk
styles = styles / \
styles.norm(float('inf'), dim=1, keepdim=True) # max_I
# Calculate per-sample weights and demodulation coefficients.
w = None
dcoefs = None
if demodulate or fused_modconv:
w = weight.unsqueeze(0) # [NOIkk]
w = w * styles.reshape(batch_size, 1, -1, 1, 1) # [NOIkk]
if demodulate:
dcoefs = (w.square().sum(dim=[2, 3, 4]) + 1e-8).rsqrt() # [NO]
if demodulate and fused_modconv:
w = w * dcoefs.reshape(batch_size, -1, 1, 1, 1) # [NOIkk]
# Execute by scaling the activations before and after the convolution.
if not fused_modconv:
x = x * styles.to(x.dtype).reshape(batch_size, -1, 1, 1)
x = conv2d_resample.conv2d_resample(x=x, w=weight.to(
x.dtype), f=resample_filter, up=up, down=down, padding=padding, flip_weight=flip_weight)
if demodulate and noise is not None:
x = fma.fma(x, dcoefs.to(x.dtype).reshape(
batch_size, -1, 1, 1), noise.to(x.dtype))
elif demodulate:
x = x * dcoefs.to(x.dtype).reshape(batch_size, -1, 1, 1)
elif noise is not None:
x = x.add_(noise.to(x.dtype))
return x
# Execute as one fused op using grouped convolution.
with misc.suppress_tracer_warnings(): # this value will be treated as a constant
batch_size = int(batch_size)
misc.assert_shape(x, [batch_size, in_channels, None, None])
x = x.reshape(1, -1, *x.shape[2:])
w = w.reshape(-1, in_channels, kh, kw)
x = conv2d_resample.conv2d_resample(x=x, w=w.to(
x.dtype), f=resample_filter, up=up, down=down, padding=padding, groups=batch_size, flip_weight=flip_weight)
x = x.reshape(batch_size, -1, *x.shape[2:])
if noise is not None:
x = x.add_(noise)
return x
# ----------------------------------------------------------------------------
@persistence.persistent_class
class FullyConnectedLayer(torch.nn.Module):
def __init__(self,
in_features, # Number of input features.
out_features, # Number of output features.
bias=True, # Apply additive bias before the activation function?
# Activation function: 'relu', 'lrelu', etc.
activation='linear',
lr_multiplier=1, # Learning rate multiplier.
bias_init=0, # Initial value for the additive bias.
):
super().__init__()
self.activation = activation
self.weight = torch.nn.Parameter(torch.randn(
[out_features, in_features]) / lr_multiplier)
self.bias = torch.nn.Parameter(torch.full(
[out_features], np.float32(bias_init))) if bias else None
self.weight_gain = lr_multiplier / np.sqrt(in_features)
self.bias_gain = lr_multiplier
def forward(self, x):
w = self.weight.to(x.dtype) * self.weight_gain
b = self.bias
if b is not None:
b = b.to(x.dtype)
if self.bias_gain != 1:
b = b * self.bias_gain
if self.activation == 'linear' and b is not None:
x = torch.addmm(b.unsqueeze(0), x, w.t())
else:
x = x.matmul(w.t())
x = bias_act.bias_act(x, b, act=self.activation)
return x
# ----------------------------------------------------------------------------
@persistence.persistent_class
class Conv2dLayer(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
# Width and height of the convolution kernel.
kernel_size,
bias=True, # Apply additive bias before the activation function?
# Activation function: 'relu', 'lrelu', etc.
activation='linear',
up=1, # Integer upsampling factor.
down=1, # Integer downsampling factor.
# Low-pass filter to apply when resampling activations.
resample_filter=[1, 3, 3, 1],
# Clamp the output to +-X, None = disable clamping.
conv_clamp=None,
channels_last=False, # Expect the input to have memory_format=channels_last?
trainable=True, # Update the weights of this layer during training?
):
super().__init__()
self.activation = activation
self.up = up
self.down = down
self.conv_clamp = conv_clamp
self.register_buffer(
'resample_filter', upfirdn2d.setup_filter(resample_filter))
self.padding = kernel_size // 2
self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
self.act_gain = bias_act.activation_funcs[activation].def_gain
memory_format = torch.channels_last if channels_last else torch.contiguous_format
weight = torch.randn([out_channels, in_channels, kernel_size, kernel_size]).to(
memory_format=memory_format)
bias = torch.zeros([out_channels]) if bias else None
if trainable:
self.weight = torch.nn.Parameter(weight)
self.bias = torch.nn.Parameter(bias) if bias is not None else None
else:
self.register_buffer('weight', weight)
if bias is not None:
self.register_buffer('bias', bias)
else:
self.bias = None
def forward(self, x, gain=1):
w = self.weight * self.weight_gain
b = self.bias.to(x.dtype) if self.bias is not None else None
flip_weight = (self.up == 1) # slightly faster
x = conv2d_resample.conv2d_resample(x=x, w=w.to(
x.dtype), f=self.resample_filter, up=self.up, down=self.down, padding=self.padding, flip_weight=flip_weight)
act_gain = self.act_gain * gain
act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
x = bias_act.bias_act(x, b, act=self.activation,
gain=act_gain, clamp=act_clamp)
return x
# ----------------------------------------------------------------------------
@persistence.persistent_class
class MappingNetwork(torch.nn.Module):
def __init__(self,
# Input latent (Z) dimensionality, 0 = no latent.
z_dim,
# Conditioning label (C) dimensionality, 0 = no label.
c_dim,
# Intermediate latent (W) dimensionality.
w_dim,
# Number of intermediate latents to output, None = do not broadcast.
num_ws,
num_layers=8, # Number of mapping layers.
# Label embedding dimensionality, None = same as w_dim.
embed_features=None,
# Number of intermediate features in the mapping layers, None = same as w_dim.
layer_features=None,
# Activation function: 'relu', 'lrelu', etc.
activation='lrelu',
# Learning rate multiplier for the mapping layers.
lr_multiplier=0.01,
# Decay for tracking the moving average of W during training, None = do not track.
w_avg_beta=0.995,
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.num_ws = num_ws
self.num_layers = num_layers
self.w_avg_beta = w_avg_beta
if embed_features is None:
embed_features = w_dim
if c_dim == 0:
embed_features = 0
if layer_features is None:
layer_features = w_dim
features_list = [z_dim + embed_features] + \
[layer_features] * (num_layers - 1) + [w_dim]
if c_dim > 0:
self.embed = FullyConnectedLayer(c_dim, embed_features)
for idx in range(num_layers):
in_features = features_list[idx]
out_features = features_list[idx + 1]
layer = FullyConnectedLayer(
in_features, out_features, activation=activation, lr_multiplier=lr_multiplier)
setattr(self, f'fc{idx}', layer)
if num_ws is not None and w_avg_beta is not None:
self.register_buffer('w_avg', torch.zeros([w_dim]))
def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, skip_w_avg_update=False):
# Embed, normalize, and concat inputs.
x = None
with torch.autograd.profiler.record_function('input'):
if self.z_dim > 0:
misc.assert_shape(z, [None, self.z_dim])
x = normalize_2nd_moment(z.to(torch.float32))
if self.c_dim > 0:
misc.assert_shape(c, [None, self.c_dim])
y = normalize_2nd_moment(self.embed(c.to(torch.float32)))
x = torch.cat([x, y], dim=1) if x is not None else y
# Main layers.
for idx in range(self.num_layers):
layer = getattr(self, f'fc{idx}')
x = layer(x)
# Update moving average of W.
if self.w_avg_beta is not None and self.training and not skip_w_avg_update:
with torch.autograd.profiler.record_function('update_w_avg'):
self.w_avg.copy_(x.detach().mean(
dim=0).lerp(self.w_avg, self.w_avg_beta))
# Broadcast.
if self.num_ws is not None:
with torch.autograd.profiler.record_function('broadcast'):
x = x.unsqueeze(1).repeat([1, self.num_ws, 1])
# Apply truncation.
if truncation_psi != 1:
with torch.autograd.profiler.record_function('truncate'):
assert self.w_avg_beta is not None
if self.num_ws is None or truncation_cutoff is None:
x = self.w_avg.lerp(x, truncation_psi)
else:
x[:, :truncation_cutoff] = self.w_avg.lerp(
x[:, :truncation_cutoff], truncation_psi)
return x
# ----------------------------------------------------------------------------
@persistence.persistent_class
class SynthesisLayer(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
out_channels, # Number of output channels.
# Intermediate latent (W) dimensionality.
w_dim,
resolution, # Resolution of this layer.
kernel_size=3, # Convolution kernel size.
up=1, # Integer upsampling factor.
use_noise=True, # Enable noise input?
# Activation function: 'relu', 'lrelu', etc.
activation='lrelu',
# Low-pass filter to apply when resampling activations.
resample_filter=[1, 3, 3, 1],
# Clamp the output of convolution layers to +-X, None = disable clamping.
conv_clamp=None,
channels_last=False, # Use channels_last format for the weights?
square=False, # default if for rectangle images
):
super().__init__()
self.resolution = resolution
self.up = up
self.use_noise = use_noise
self.activation = activation
self.conv_clamp = conv_clamp
self.register_buffer(
'resample_filter', upfirdn2d.setup_filter(resample_filter))
self.padding = kernel_size // 2
self.act_gain = bias_act.activation_funcs[activation].def_gain
self.square = square
self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1)
memory_format = torch.channels_last if channels_last else torch.contiguous_format
self.weight = torch.nn.Parameter(torch.randn(
[out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format))
if use_noise:
if self.square:
self.register_buffer(
'noise_const', torch.randn([resolution, resolution]))
else:
self.register_buffer('noise_const', torch.randn(
[resolution, resolution // 2]))
self.noise_strength = torch.nn.Parameter(torch.zeros([]))
self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
def forward(self, x, w, noise_mode='random', fused_modconv=True, gain=1):
assert noise_mode in ['random', 'const', 'none']
in_resolution = self.resolution // self.up
if self.square:
misc.assert_shape(
x, [None, self.weight.shape[1], in_resolution, in_resolution])
else:
misc.assert_shape(
x, [None, self.weight.shape[1], in_resolution, in_resolution // 2])
styles = self.affine(w)
noise = None
if self.use_noise and noise_mode == 'random':
if self.square:
noise = torch.randn(
[x.shape[0], 1, self.resolution, self.resolution], device=x.device) * self.noise_strength
else:
noise = torch.randn(
[x.shape[0], 1, self.resolution, self.resolution // 2], device=x.device) * self.noise_strength
if self.use_noise and noise_mode == 'const':
noise = self.noise_const * self.noise_strength
flip_weight = (self.up == 1) # slightly faster
x = modulated_conv2d(x=x, weight=self.weight, styles=styles, noise=noise, up=self.up,
padding=self.padding, resample_filter=self.resample_filter, flip_weight=flip_weight, fused_modconv=fused_modconv)
act_gain = self.act_gain * gain
act_clamp = self.conv_clamp * gain if self.conv_clamp is not None else None
x = bias_act.bias_act(x, self.bias.to(
x.dtype), act=self.activation, gain=act_gain, clamp=act_clamp)
return x
# ----------------------------------------------------------------------------
@persistence.persistent_class
class ToRGBLayer(torch.nn.Module):
def __init__(self, in_channels, out_channels, w_dim, kernel_size=1, conv_clamp=None, channels_last=False):
super().__init__()
self.conv_clamp = conv_clamp
self.affine = FullyConnectedLayer(w_dim, in_channels, bias_init=1)
memory_format = torch.channels_last if channels_last else torch.contiguous_format
self.weight = torch.nn.Parameter(torch.randn(
[out_channels, in_channels, kernel_size, kernel_size]).to(memory_format=memory_format))
self.bias = torch.nn.Parameter(torch.zeros([out_channels]))
self.weight_gain = 1 / np.sqrt(in_channels * (kernel_size ** 2))
def forward(self, x, w, fused_modconv=True):
styles = self.affine(w) * self.weight_gain
x = modulated_conv2d(x=x, weight=self.weight, styles=styles,
demodulate=False, fused_modconv=fused_modconv)
x = bias_act.bias_act(x, self.bias.to(x.dtype), clamp=self.conv_clamp)
return x
# ----------------------------------------------------------------------------
@persistence.persistent_class
class SynthesisBlock(torch.nn.Module):
def __init__(self,
# Number of input channels, 0 = first block.
in_channels,
# Number of output channels.
out_channels,
# Intermediate latent (W) dimensionality.
w_dim,
# Resolution of this block.
resolution,
# Number of output color channels.
img_channels,
is_last, # Is this the last block?
# Architecture: 'orig', 'skip', 'resnet'.
architecture='skip',
# Low-pass filter to apply when resampling activations.
resample_filter=[1, 3, 3, 1],
# Clamp the output of convolution layers to +-X, None = disable clamping.
conv_clamp=None,
use_fp16=False, # Use FP16 for this block?
fp16_channels_last=False, # Use channels-last memory format with FP16?
square=False, # default is for rectangle images
# Arguments for SynthesisLayer.
**layer_kwargs,
):
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.w_dim = w_dim
self.resolution = resolution
self.img_channels = img_channels
self.is_last = is_last
self.architecture = architecture
self.use_fp16 = use_fp16
self.channels_last = (use_fp16 and fp16_channels_last)
self.register_buffer(
'resample_filter', upfirdn2d.setup_filter(resample_filter))
self.num_conv = 0
self.num_torgb = 0
self.square = square
if in_channels == 0:
if self.square:
self.const = torch.nn.Parameter(torch.randn(
[out_channels, resolution, resolution]))
else: # rectangle
self.const = torch.nn.Parameter(torch.randn(
[out_channels, resolution, resolution // 2]))
if in_channels != 0:
self.conv0 = SynthesisLayer(in_channels, out_channels, w_dim=w_dim, resolution=resolution, up=2,
resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last, square=square, **layer_kwargs)
self.num_conv += 1
self.conv1 = SynthesisLayer(out_channels, out_channels, w_dim=w_dim, resolution=resolution,
conv_clamp=conv_clamp, channels_last=self.channels_last, square=square, **layer_kwargs)
self.num_conv += 1
if is_last or architecture == 'skip':
self.torgb = ToRGBLayer(out_channels, img_channels, w_dim=w_dim,
conv_clamp=conv_clamp, channels_last=self.channels_last)
self.num_torgb += 1
if in_channels != 0 and architecture == 'resnet':
self.skip = Conv2dLayer(in_channels, out_channels, kernel_size=1, bias=False, up=2,
resample_filter=resample_filter, channels_last=self.channels_last)
def forward(self, x, img, ws, force_fp32=False, fused_modconv=None, **layer_kwargs):
misc.assert_shape(
ws, [None, self.num_conv + self.num_torgb, self.w_dim])
w_iter = iter(ws.unbind(dim=1))
dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32
memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format
if fused_modconv is None:
with misc.suppress_tracer_warnings(): # this value will be treated as a constant
fused_modconv = (not self.training) and (
dtype == torch.float32 or int(x.shape[0]) == 1)
# Input.
if self.in_channels == 0:
x = self.const.to(dtype=dtype, memory_format=memory_format)
x = x.unsqueeze(0).repeat([ws.shape[0], 1, 1, 1])
else:
if self.square:
misc.assert_shape(
x, [None, self.in_channels, self.resolution // 2, self.resolution // 2])
else: # rectangle
misc.assert_shape(
x, [None, self.in_channels, self.resolution // 2, self.resolution // 4])
x = x.to(dtype=dtype, memory_format=memory_format)
# Main layers.
if self.in_channels == 0:
x = self.conv1(x, next(w_iter),
fused_modconv=fused_modconv, **layer_kwargs)
elif self.architecture == 'resnet':
y = self.skip(x, gain=np.sqrt(0.5))
x = self.conv0(x, next(w_iter),
fused_modconv=fused_modconv, **layer_kwargs)
x = self.conv1(x, next(w_iter), fused_modconv=fused_modconv,
gain=np.sqrt(0.5), **layer_kwargs)
x = y.add_(x)
else:
x = self.conv0(x, next(w_iter),
fused_modconv=fused_modconv, **layer_kwargs)
x = self.conv1(x, next(w_iter),
fused_modconv=fused_modconv, **layer_kwargs)
# ToRGB.
if img is not None:
if self.square:
misc.assert_shape(
img, [None, self.img_channels, self.resolution // 2, self.resolution // 2])
else:
misc.assert_shape(
img, [None, self.img_channels, self.resolution // 2, self.resolution // 4])
img = upfirdn2d.upsample2d(img, self.resample_filter)
if self.is_last or self.architecture == 'skip':
y = self.torgb(x, next(w_iter), fused_modconv=fused_modconv)
y = y.to(dtype=torch.float32,
memory_format=torch.contiguous_format)
img = img.add_(y) if img is not None else y
assert x.dtype == dtype
assert img is None or img.dtype == torch.float32
return x, img
# ----------------------------------------------------------------------------
@persistence.persistent_class
class SynthesisNetwork(torch.nn.Module):
def __init__(self,
# Intermediate latent (W) dimensionality.
w_dim,
img_resolution, # Output image resolution.
img_channels, # Number of color channels.
square,
# Overall multiplier for the number of channels.
channel_base=32768,
# Maximum number of channels in any layer.
channel_max=512,
# Use FP16 for the N highest resolutions.
num_fp16_res=0,
**block_kwargs, # Arguments for SynthesisBlock.
):
assert img_resolution >= 4 and img_resolution & (
img_resolution - 1) == 0
super().__init__()
self.w_dim = w_dim
self.img_resolution = img_resolution
self.img_resolution_log2 = int(np.log2(img_resolution))
self.img_channels = img_channels
self.square = square
self.block_resolutions = [
2 ** i for i in range(2, self.img_resolution_log2 + 1)]
channels_dict = {res: min(channel_base // res, channel_max)
for res in self.block_resolutions}
fp16_resolution = max(
2 ** (self.img_resolution_log2 + 1 - num_fp16_res), 8)
self.num_ws = 0
for res in self.block_resolutions:
in_channels = channels_dict[res // 2] if res > 4 else 0
out_channels = channels_dict[res]
use_fp16 = (res >= fp16_resolution)
is_last = (res == self.img_resolution)
block = SynthesisBlock(in_channels, out_channels, w_dim=w_dim, resolution=res,
img_channels=img_channels, is_last=is_last, use_fp16=use_fp16, square=square, **block_kwargs)
self.num_ws += block.num_conv
if is_last:
self.num_ws += block.num_torgb
setattr(self, f'b{res}', block)
def forward(self, ws, return_feature=False, **block_kwargs):
block_ws = []
features = []
with torch.autograd.profiler.record_function('split_ws'):
misc.assert_shape(ws, [None, self.num_ws, self.w_dim])
ws = ws.to(torch.float32)
w_idx = 0
for res in self.block_resolutions:
block = getattr(self, f'b{res}')
block_ws.append(
ws.narrow(1, w_idx, block.num_conv + block.num_torgb))
w_idx += block.num_conv
x = img = None
for res, cur_ws in zip(self.block_resolutions, block_ws):
block = getattr(self, f'b{res}')
x, img = block(x, img, cur_ws, **block_kwargs)
features.append(x)
if return_feature:
return img, features
else:
return img
# ----------------------------------------------------------------------------
@persistence.persistent_class
class Generator(torch.nn.Module):
def __init__(self,
z_dim, # Input latent (Z) dimensionality.
# Conditioning label (C) dimensionality.
c_dim,
# Intermediate latent (W) dimensionality.
w_dim,
img_resolution, # Output resolution.
square,
img_channels, # Number of output color channels.
mapping_kwargs={}, # Arguments for MappingNetwork.
synthesis_kwargs={}, # Arguments for SynthesisNetwork.
padding=False
):
super().__init__()
self.z_dim = z_dim
self.c_dim = c_dim
self.w_dim = w_dim
self.square = square
self.img_resolution = img_resolution
self.img_channels = img_channels
self.padding = padding
self.synthesis = SynthesisNetwork(
w_dim=w_dim, img_resolution=img_resolution, img_channels=img_channels, square=square, **synthesis_kwargs)
self.num_ws = self.synthesis.num_ws
self.mapping = MappingNetwork(
z_dim=z_dim, c_dim=c_dim, w_dim=w_dim, num_ws=self.num_ws, **mapping_kwargs)
def forward(self, z, c, truncation_psi=1, truncation_cutoff=None, input_is_w=False, return_feature=False, **synthesis_kwargs):
if input_is_w:
ws = z
if ws.dim() == 2:
ws = ws.unsqueeze(1).repeat([1, self.mapping.num_ws, 1])
else:
ws = self.mapping(z, c, truncation_psi=truncation_psi,
truncation_cutoff=truncation_cutoff)
img = self.synthesis(
ws, return_feature=return_feature, **synthesis_kwargs)
if return_feature:
img, feature = img
if self.padding:
pad = (img.size(2) - img.size(3)) // 2
img = torch.nn.functional.pad(img, (pad, pad), "constant", 1)
if return_feature:
for i, feat in enumerate(feature):
pad = (feat.size(2) - feat.size(3)) // 2
feature[i] = torch.nn.functional.pad(
feat, (pad, pad), "constant", 0)
if return_feature:
return img, feature
else:
return img
# ----------------------------------------------------------------------------
@persistence.persistent_class
class DiscriminatorBlock(torch.nn.Module):
def __init__(self,
# Number of input channels, 0 = first block.
in_channels,
# Number of intermediate channels.
tmp_channels,
# Number of output channels.
out_channels,
# Resolution of this block.
resolution,
# Number of input color channels.
img_channels,
# Index of the first layer.
first_layer_idx,
# Architecture: 'orig', 'skip', 'resnet'.
architecture='resnet',
# Activation function: 'relu', 'lrelu', etc.
activation='lrelu',
# Low-pass filter to apply when resampling activations.
resample_filter=[1, 3, 3, 1],
# Clamp the output of convolution layers to +-X, None = disable clamping.
conv_clamp=None,
use_fp16=False, # Use FP16 for this block?
fp16_channels_last=False, # Use channels-last memory format with FP16?
# Freeze-D: Number of layers to freeze.
freeze_layers=0,
square=False,
):
assert in_channels in [0, tmp_channels]
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.resolution = resolution
self.img_channels = img_channels
self.first_layer_idx = first_layer_idx
self.architecture = architecture
self.use_fp16 = use_fp16
self.channels_last = (use_fp16 and fp16_channels_last)
self.register_buffer(
'resample_filter', upfirdn2d.setup_filter(resample_filter))
self.square = square
self.num_layers = 0
def trainable_gen():
while True:
layer_idx = self.first_layer_idx + self.num_layers
trainable = (layer_idx >= freeze_layers)
self.num_layers += 1
yield trainable
trainable_iter = trainable_gen()
if in_channels == 0 or architecture == 'skip':
self.fromrgb = Conv2dLayer(img_channels, tmp_channels, kernel_size=1, activation=activation,
trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last)
self.conv0 = Conv2dLayer(tmp_channels, tmp_channels, kernel_size=3, activation=activation,
trainable=next(trainable_iter), conv_clamp=conv_clamp, channels_last=self.channels_last)
self.conv1 = Conv2dLayer(tmp_channels, out_channels, kernel_size=3, activation=activation, down=2,
trainable=next(trainable_iter), resample_filter=resample_filter, conv_clamp=conv_clamp, channels_last=self.channels_last)
if architecture == 'resnet':
self.skip = Conv2dLayer(tmp_channels, out_channels, kernel_size=1, bias=False, down=2,
trainable=next(trainable_iter), resample_filter=resample_filter, channels_last=self.channels_last)
def forward(self, x, img, force_fp32=False):
dtype = torch.float16 if self.use_fp16 and not force_fp32 else torch.float32
memory_format = torch.channels_last if self.channels_last and not force_fp32 else torch.contiguous_format
# Input.
if x is not None:
if self.square:
misc.assert_shape(
x, [None, self.in_channels, self.resolution, self.resolution])
else:
misc.assert_shape(
x, [None, self.in_channels, self.resolution, self.resolution // 2])
x = x.to(dtype=dtype, memory_format=memory_format)
# FromRGB.
if self.in_channels == 0 or self.architecture == 'skip':
if self.square:
misc.assert_shape(
img, [None, self.img_channels, self.resolution, self.resolution])
else:
misc.assert_shape(
img, [None, self.img_channels, self.resolution, self.resolution // 2])
img = img.to(dtype=dtype, memory_format=memory_format)
y = self.fromrgb(img)
x = x + y if x is not None else y
img = upfirdn2d.downsample2d(
img, self.resample_filter) if self.architecture == 'skip' else None
# Main layers.
if self.architecture == 'resnet':
y = self.skip(x, gain=np.sqrt(0.5))
x = self.conv0(x)
x = self.conv1(x, gain=np.sqrt(0.5))
x = y.add_(x)
else:
x = self.conv0(x)
x = self.conv1(x)
assert x.dtype == dtype
return x, img
# ----------------------------------------------------------------------------
@persistence.persistent_class
class MinibatchStdLayer(torch.nn.Module):
def __init__(self, group_size, num_channels=1):
super().__init__()
self.group_size = group_size
self.num_channels = num_channels
def forward(self, x):
N, C, H, W = x.shape
with misc.suppress_tracer_warnings(): # as_tensor results are registered as constants
G = torch.min(torch.as_tensor(self.group_size), torch.as_tensor(
N)) if self.group_size is not None else N
F = self.num_channels
c = C // F
# [GnFcHW] Split minibatch N into n groups of size G, and channels C into F groups of size c.
y = x.reshape(G, -1, F, c, H, W)
# [GnFcHW] Subtract mean over group.
y = y - y.mean(dim=0)
# [nFcHW] Calc variance over group.
y = y.square().mean(dim=0)
y = (y + 1e-8).sqrt() # [nFcHW] Calc stddev over group.
# [nF] Take average over channels and pixels.
y = y.mean(dim=[2, 3, 4])
y = y.reshape(-1, F, 1, 1) # [nF11] Add missing dimensions.
# [NFHW] Replicate over group and pixels.
y = y.repeat(G, 1, H, W)
# [NCHW] Append to input as new channels.
x = torch.cat([x, y], dim=1)
return x
# ----------------------------------------------------------------------------
@persistence.persistent_class
class DiscriminatorEpilogue(torch.nn.Module):
def __init__(self,
in_channels, # Number of input channels.
# Dimensionality of mapped conditioning label, 0 = no label.
cmap_dim,
resolution, # Resolution of this block.
# Number of input color channels.
img_channels,
# Architecture: 'orig', 'skip', 'resnet'.
architecture='resnet',
# Group size for the minibatch standard deviation layer, None = entire minibatch.
mbstd_group_size=4,
# Number of features for the minibatch standard deviation layer, 0 = disable.
mbstd_num_channels=1,
# Activation function: 'relu', 'lrelu', etc.
activation='lrelu',
# Clamp the output of convolution layers to +-X, None = disable clamping.
conv_clamp=None,
square=False,
):
assert architecture in ['orig', 'skip', 'resnet']
super().__init__()
self.in_channels = in_channels
self.cmap_dim = cmap_dim
self.resolution = resolution
self.img_channels = img_channels
self.architecture = architecture
self.square = square
if architecture == 'skip':
self.fromrgb = Conv2dLayer(
img_channels, in_channels, kernel_size=1, activation=activation)
self.mbstd = MinibatchStdLayer(
group_size=mbstd_group_size, num_channels=mbstd_num_channels) if mbstd_num_channels > 0 else None
self.conv = Conv2dLayer(in_channels + mbstd_num_channels, in_channels,
kernel_size=3, activation=activation, conv_clamp=conv_clamp)
if self.square:
self.fc = FullyConnectedLayer(
in_channels * (resolution ** 2), in_channels, activation=activation)
else:
self.fc = FullyConnectedLayer(
in_channels * (resolution ** 2 // 2), in_channels, activation=activation)
self.out = FullyConnectedLayer(
in_channels, 1 if cmap_dim == 0 else cmap_dim)
def forward(self, x, img, cmap, force_fp32=False):
if self.square:
misc.assert_shape(x, [None, self.in_channels,
self.resolution, self.resolution])
else:
misc.assert_shape(
x, [None, self.in_channels, self.resolution, self.resolution // 2]) # [NCHW]
_ = force_fp32 # unused
dtype = torch.float32
memory_format = torch.contiguous_format
# FromRGB.
x = x.to(dtype=dtype, memory_format=memory_format)
if self.architecture == 'skip':
if self.square:
misc.assert_shape(
img, [None, self.img_channels, self.resolution, self.resolution])
else:
misc.assert_shape(
img, [None, self.img_channels, self.resolution, self.resolution // 2])
img = img.to(dtype=dtype, memory_format=memory_format)
x = x + self.fromrgb(img)
# Main layers.
if self.mbstd is not None:
x = self.mbstd(x)
x = self.conv(x)
x = self.fc(x.flatten(1))
x = self.out(x)
# Conditioning.
if self.cmap_dim > 0:
misc.assert_shape(cmap, [None, self.cmap_dim])
x = (x * cmap).sum(dim=1, keepdim=True) * \
(1 / np.sqrt(self.cmap_dim))
assert x.dtype == dtype
return x
# ----------------------------------------------------------------------------
@persistence.persistent_class
class Discriminator(torch.nn.Module):
def __init__(self,
# Conditioning label (C) dimensionality.
c_dim,
img_resolution, # Input resolution.
# Number of input color channels.
img_channels,
# Architecture: 'orig', 'skip', 'resnet'.
architecture='resnet',
# Overall multiplier for the number of channels.
channel_base=32768,
# Maximum number of channels in any layer.
channel_max=512,
# Use FP16 for the N highest resolutions.
num_fp16_res=0,
# Clamp the output of convolution layers to +-X, None = disable clamping.
conv_clamp=None,
# Dimensionality of mapped conditioning label, None = default.
cmap_dim=None,
square=False, # default for rectangle images
block_kwargs={}, # Arguments for DiscriminatorBlock.
mapping_kwargs={}, # Arguments for MappingNetwork.
# Arguments for DiscriminatorEpilogue.
epilogue_kwargs={},
):
super().__init__()
self.c_dim = c_dim
self.img_resolution = img_resolution
self.img_resolution_log2 = int(np.log2(img_resolution))
self.img_channels = img_channels
self.square = square
self.block_resolutions = [
2 ** i for i in range(self.img_resolution_log2, 2, -1)]
channels_dict = {res: min(channel_base // res, channel_max)
for res in self.block_resolutions + [4]}
fp16_resolution = max(
2 ** (self.img_resolution_log2 + 1 - num_fp16_res), 8)
if cmap_dim is None:
cmap_dim = channels_dict[4]
if c_dim == 0:
cmap_dim = 0
common_kwargs = dict(img_channels=img_channels,
architecture=architecture, conv_clamp=conv_clamp)
cur_layer_idx = 0
for res in self.block_resolutions:
in_channels = channels_dict[res] if res < img_resolution else 0
tmp_channels = channels_dict[res]
out_channels = channels_dict[res // 2]
use_fp16 = (res >= fp16_resolution)
block = DiscriminatorBlock(in_channels, tmp_channels, out_channels, resolution=res,
first_layer_idx=cur_layer_idx, use_fp16=use_fp16, square=square, **block_kwargs, **common_kwargs)
setattr(self, f'b{res}', block)
cur_layer_idx += block.num_layers
if c_dim > 0:
self.mapping = MappingNetwork(
z_dim=0, c_dim=c_dim, w_dim=cmap_dim, num_ws=None, w_avg_beta=None, **mapping_kwargs)
self.b4 = DiscriminatorEpilogue(
channels_dict[4], cmap_dim=cmap_dim, resolution=4, square=square, **epilogue_kwargs, **common_kwargs)
def forward(self, img, c, **block_kwargs):
x = None
for res in self.block_resolutions:
block = getattr(self, f'b{res}')
x, img = block(x, img, **block_kwargs)
cmap = None
if self.c_dim > 0:
cmap = self.mapping(None, c)
x = self.b4(x, img, cmap)
return x
# ----------------------------------------------------------------------------
|