Spaces:
Build error
Build error
File size: 14,533 Bytes
7eb6194 |
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 |
import torch.nn as nn
import torch.nn.functional as F
import torch as th
import datetime
import os
import time
import timeit
import copy
import numpy as np
from torch.nn import ModuleList
from torch.nn import Conv2d
from torch.nn import LeakyReLU
#PixelwiseNorm代替了BatchNorm
class PixelwiseNorm(th.nn.Module):
def __init__(self):
super(PixelwiseNorm, self).__init__()
def forward(self, x, alpha=1e-8):
"""
forward pass of the module
:param x: input activations volume
:param alpha: small number for numerical stability
:return: y => pixel normalized activations
"""
y = x.pow(2.).mean(dim=1, keepdim=True).add(alpha).sqrt() # [N1HW]
y = x / y # normalize the input x volume
return y
class MinibatchStdDev(th.nn.Module):
"""
Minibatch standard deviation layer for the discriminator
"""
def __init__(self):
"""
derived class constructor
"""
super().__init__()
def forward(self, x, alpha=1e-8):
"""
forward pass of the layer
:param x: input activation volume
:param alpha: small number for numerical stability
:return: y => x appended with standard deviation constant map
"""
batch_size, _, height, width = x.shape
# [B x C x H x W] Subtract mean over batch.
y = x - x.mean(dim=0, keepdim=True)
# [1 x C x H x W] Calc standard deviation over batch
y = th.sqrt(y.pow(2.).mean(dim=0, keepdim=False) + alpha)
# [1] Take average over feature_maps and pixels.
y = y.mean().view(1, 1, 1, 1)
# [B x 1 x H x W] Replicate over group and pixels.
y = y.repeat(batch_size, 1, height, width)
# [B x C x H x W] Append as new feature_map.
y = th.cat([x, y], 1)
# return the computed values:
return y
# ==========================================================
# Equalized learning rate blocks:
# extending Conv2D and Deconv2D layers for equalized learning rate logic
# ==========================================================
class _equalized_conv2d(th.nn.Module):
""" conv2d with the concept of equalized learning rate
Args:
:param c_in: input channels
:param c_out: output channels
:param k_size: kernel size (h, w) should be a tuple or a single integer
:param stride: stride for conv
:param pad: padding
:param bias: whether to use bias or not
"""
def __init__(self, c_in, c_out, k_size, stride=1, pad=0, bias=True):
""" constructor for the class """
from torch.nn.modules.utils import _pair
from numpy import sqrt, prod
super().__init__()
# define the weight and bias if to be used
self.weight = th.nn.Parameter(th.nn.init.normal_(
th.empty(c_out, c_in, *_pair(k_size))
))
self.use_bias = bias
self.stride = stride
self.pad = pad
if self.use_bias:
self.bias = th.nn.Parameter(th.FloatTensor(c_out).fill_(0))
fan_in = prod(_pair(k_size)) * c_in # value of fan_in
self.scale = sqrt(2) / sqrt(fan_in)
def forward(self, x):
"""
forward pass of the network
:param x: input
:return: y => output
"""
from torch.nn.functional import conv2d
return conv2d(input=x,
weight=self.weight * self.scale, # scale the weight on runtime
bias=self.bias if self.use_bias else None,
stride=self.stride,
padding=self.pad)
def extra_repr(self):
return ", ".join(map(str, self.weight.shape))
class _equalized_deconv2d(th.nn.Module):
""" Transpose convolution using the equalized learning rate
Args:
:param c_in: input channels
:param c_out: output channels
:param k_size: kernel size
:param stride: stride for convolution transpose
:param pad: padding
:param bias: whether to use bias or not
"""
def __init__(self, c_in, c_out, k_size, stride=1, pad=0, bias=True):
""" constructor for the class """
from torch.nn.modules.utils import _pair
from numpy import sqrt
super().__init__()
# define the weight and bias if to be used
self.weight = th.nn.Parameter(th.nn.init.normal_(
th.empty(c_in, c_out, *_pair(k_size))
))
self.use_bias = bias
self.stride = stride
self.pad = pad
if self.use_bias:
self.bias = th.nn.Parameter(th.FloatTensor(c_out).fill_(0))
fan_in = c_in # value of fan_in for deconv
self.scale = sqrt(2) / sqrt(fan_in)
def forward(self, x):
"""
forward pass of the layer
:param x: input
:return: y => output
"""
from torch.nn.functional import conv_transpose2d
return conv_transpose2d(input=x,
weight=self.weight * self.scale, # scale the weight on runtime
bias=self.bias if self.use_bias else None,
stride=self.stride,
padding=self.pad)
def extra_repr(self):
return ", ".join(map(str, self.weight.shape))
#basic block of the encoding part of the genarater
#编码器的基本卷积块
class conv_block(nn.Module):
"""
Convolution Block
with two convolution layers
"""
def __init__(self, in_ch, out_ch,use_eql=True):
super(conv_block, self).__init__()
if use_eql:
self.conv_1= _equalized_conv2d(in_ch, out_ch, (1, 1),
pad=0, bias=True)
self.conv_2 = _equalized_conv2d(out_ch, out_ch, (3, 3),
pad=1, bias=True)
self.conv_3 = _equalized_conv2d(out_ch, out_ch, (3, 3),
pad=1, bias=True)
else:
self.conv_1 = Conv2d(in_ch, out_ch, (3, 3),
padding=1, bias=True)
self.conv_2 = Conv2d(out_ch, out_ch, (3, 3),
padding=1, bias=True)
# pixel_wise feature normalizer:
self.pixNorm = PixelwiseNorm()
# leaky_relu:
self.lrelu = LeakyReLU(0.2)
def forward(self, x):
"""
forward pass of the block
:param x: input
:return: y => output
"""
from torch.nn.functional import interpolate
#y = interpolate(x, scale_factor=2)
y=self.conv_1(self.lrelu(self.pixNorm(x)))
residual=y
y=self.conv_2(self.lrelu(self.pixNorm(y)))
y=self.conv_3(self.lrelu(self.pixNorm(y)))
y=y+residual
return y
#basic up convolution block of the encoding part of the genarater
#编码器的基本卷积块
class up_conv(nn.Module):
"""
Up Convolution Block
"""
def __init__(self, in_ch, out_ch,use_eql=True):
super(up_conv, self).__init__()
if use_eql:
self.conv_1= _equalized_conv2d(in_ch, out_ch, (1, 1),
pad=0, bias=True)
self.conv_2 = _equalized_conv2d(out_ch, out_ch, (3, 3),
pad=1, bias=True)
self.conv_3 = _equalized_conv2d(out_ch, out_ch, (3, 3),
pad=1, bias=True)
else:
self.conv_1 = Conv2d(in_ch, out_ch, (3, 3),
padding=1, bias=True)
self.conv_2 = Conv2d(out_ch, out_ch, (3, 3),
padding=1, bias=True)
# pixel_wise feature normalizer:
self.pixNorm = PixelwiseNorm()
# leaky_relu:
self.lrelu = LeakyReLU(0.2)
def forward(self, x):
"""
forward pass of the block
:param x: input
:return: y => output
"""
from torch.nn.functional import interpolate
x = interpolate(x, scale_factor=2, mode="bilinear")
y=self.conv_1(self.lrelu(self.pixNorm(x)))
residual=y
y=self.conv_2(self.lrelu(self.pixNorm(y)))
y=self.conv_3(self.lrelu(self.pixNorm(y)))
y=y+residual
return y
#判别器的最后一层
class DisFinalBlock(th.nn.Module):
""" Final block for the Discriminator """
def __init__(self, in_channels, use_eql=True):
"""
constructor of the class
:param in_channels: number of input channels
:param use_eql: whether to use equalized learning rate
"""
from torch.nn import LeakyReLU
from torch.nn import Conv2d
super().__init__()
# declare the required modules for forward pass
self.batch_discriminator = MinibatchStdDev()
if use_eql:
self.conv_1 = _equalized_conv2d(in_channels + 1, in_channels, (3, 3),
pad=1, bias=True)
self.conv_2 = _equalized_conv2d(in_channels, in_channels, (4, 4),stride=2,pad=1,
bias=True)
# final layer emulates the fully connected layer
self.conv_3 = _equalized_conv2d(in_channels, 1, (1, 1), bias=True)
else:
# modules required:
self.conv_1 = Conv2d(in_channels + 1, in_channels, (3, 3), padding=1, bias=True)
self.conv_2 = Conv2d(in_channels, in_channels, (4, 4), bias=True)
# final conv layer emulates a fully connected layer
self.conv_3 = Conv2d(in_channels, 1, (1, 1), bias=True)
# leaky_relu:
self.lrelu = LeakyReLU(0.2)
def forward(self, x):
"""
forward pass of the FinalBlock
:param x: input
:return: y => output
"""
# minibatch_std_dev layer
y = self.batch_discriminator(x)
# define the computations
y = self.lrelu(self.conv_1(y))
y = self.lrelu(self.conv_2(y))
# fully connected layer
y = self.conv_3(y) # This layer has linear activation
# flatten the output raw discriminator scores
return y
#判别器基本卷积块
class DisGeneralConvBlock(th.nn.Module):
""" General block in the discriminator """
def __init__(self, in_channels, out_channels, use_eql=True):
"""
constructor of the class
:param in_channels: number of input channels
:param out_channels: number of output channels
:param use_eql: whether to use equalized learning rate
"""
from torch.nn import AvgPool2d, LeakyReLU
from torch.nn import Conv2d
super().__init__()
if use_eql:
self.conv_1 = _equalized_conv2d(in_channels, in_channels, (3, 3),
pad=1, bias=True)
self.conv_2 = _equalized_conv2d(in_channels, out_channels, (3, 3),
pad=1, bias=True)
else:
# convolutional modules
self.conv_1 = Conv2d(in_channels, in_channels, (3, 3),
padding=1, bias=True)
self.conv_2 = Conv2d(in_channels, out_channels, (3, 3),
padding=1, bias=True)
self.downSampler = AvgPool2d(2) # downsampler
# leaky_relu:
self.lrelu = LeakyReLU(0.2)
def forward(self, x):
"""
forward pass of the module
:param x: input
:return: y => output
"""
# define the computations
y = self.lrelu(self.conv_1(x))
y = self.lrelu(self.conv_2(y))
y = self.downSampler(y)
return y
class from_rgb(nn.Module):
"""
The RGB image is transformed into a multi-channel feature map to be concatenated with
the feature map with the same number of channels in the network
把RGB图转换为多通道特征图,以便与网络中相同通道数的特征图拼接
"""
def __init__(self, outchannels, use_eql=True):
super(from_rgb, self).__init__()
if use_eql:
self.conv_1 = _equalized_conv2d(3, outchannels, (1, 1), bias=True)
else:
self.conv_1 = nn.Conv2d(3, outchannels, (1, 1),bias=True)
# pixel_wise feature normalizer:
self.pixNorm = PixelwiseNorm()
# leaky_relu:
self.lrelu = LeakyReLU(0.2)
def forward(self, x):
"""
forward pass of the block
:param x: input
:return: y => output
"""
y = self.pixNorm(self.lrelu(self.conv_1(x)))
return y
class to_rgb(nn.Module):
"""
把多通道特征图转换为RGB三通道图,以便输入判别器
The multi-channel feature map is converted into RGB image for input to the discriminator
"""
def __init__(self, inchannels, use_eql=True):
super(to_rgb, self).__init__()
if use_eql:
self.conv_1 = _equalized_conv2d(inchannels, 3, (1, 1), bias=True)
else:
self.conv_1 = nn.Conv2d(inchannels, 3, (1, 1),bias=True)
def forward(self, x):
"""
forward pass of the block
:param x: input
:return: y => output
"""
y = self.conv_1(x)
return y
class Flatten(nn.Module):
def forward(self, x):
return x.view(x.size(0), -1)
class CCA(nn.Module):
"""
CCA Block
"""
def __init__(self, F_g, F_x):
super().__init__()
self.mlp_x = nn.Sequential(
Flatten(),
nn.Linear(F_x, F_x))
self.mlp_g = nn.Sequential(
Flatten(),
nn.Linear(F_g, F_x))
self.relu = nn.ReLU(inplace=True)
def forward(self, g, x):
# channel-wise attention
avg_pool_x = F.avg_pool2d( x, (x.size(2), x.size(3)), stride=(x.size(2), x.size(3)))
channel_att_x = self.mlp_x(avg_pool_x)
avg_pool_g = F.avg_pool2d( g, (g.size(2), g.size(3)), stride=(g.size(2), g.size(3)))
channel_att_g = self.mlp_g(avg_pool_g)
channel_att_sum = (channel_att_x + channel_att_g)/2.0
scale = th.sigmoid(channel_att_sum).unsqueeze(2).unsqueeze(3).expand_as(x)
x_after_channel = x * scale
out = self.relu(x_after_channel)
return out |