File size: 9,056 Bytes
33719d5 |
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 |
import functools
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.nn.utils import spectral_norm
### single layers
def conv2d(*args, **kwargs):
return spectral_norm(nn.Conv2d(*args, **kwargs))
def convTranspose2d(*args, **kwargs):
return spectral_norm(nn.ConvTranspose2d(*args, **kwargs))
def embedding(*args, **kwargs):
return spectral_norm(nn.Embedding(*args, **kwargs))
def linear(*args, **kwargs):
return spectral_norm(nn.Linear(*args, **kwargs))
def NormLayer(c, mode='batch'):
if mode == 'group':
return nn.GroupNorm(c//2, c)
elif mode == 'batch':
return nn.BatchNorm2d(c)
### Activations
class GLU(nn.Module):
def forward(self, x):
nc = x.size(1)
assert nc % 2 == 0, 'channels dont divide 2!'
nc = int(nc/2)
return x[:, :nc] * torch.sigmoid(x[:, nc:])
class Swish(nn.Module):
def forward(self, feat):
return feat * torch.sigmoid(feat)
### Upblocks
class InitLayer(nn.Module):
def __init__(self, nz, channel, sz=4):
super().__init__()
self.init = nn.Sequential(
convTranspose2d(nz, channel*2, sz, 1, 0, bias=False),
NormLayer(channel*2),
GLU(),
)
def forward(self, noise):
noise = noise.view(noise.shape[0], -1, 1, 1)
return self.init(noise)
def UpBlockSmall(in_planes, out_planes):
block = nn.Sequential(
nn.Upsample(scale_factor=2, mode='nearest'),
conv2d(in_planes, out_planes*2, 3, 1, 1, bias=False),
NormLayer(out_planes*2), GLU())
return block
class UpBlockSmallCond(nn.Module):
def __init__(self, in_planes, out_planes, z_dim):
super().__init__()
self.in_planes = in_planes
self.out_planes = out_planes
self.up = nn.Upsample(scale_factor=2, mode='nearest')
self.conv = conv2d(in_planes, out_planes*2, 3, 1, 1, bias=False)
which_bn = functools.partial(CCBN, which_linear=linear, input_size=z_dim)
self.bn = which_bn(2*out_planes)
self.act = GLU()
def forward(self, x, c):
x = self.up(x)
x = self.conv(x)
x = self.bn(x, c)
x = self.act(x)
return x
def UpBlockBig(in_planes, out_planes):
block = nn.Sequential(
nn.Upsample(scale_factor=2, mode='nearest'),
conv2d(in_planes, out_planes*2, 3, 1, 1, bias=False),
NoiseInjection(),
NormLayer(out_planes*2), GLU(),
conv2d(out_planes, out_planes*2, 3, 1, 1, bias=False),
NoiseInjection(),
NormLayer(out_planes*2), GLU()
)
return block
class UpBlockBigCond(nn.Module):
def __init__(self, in_planes, out_planes, z_dim):
super().__init__()
self.in_planes = in_planes
self.out_planes = out_planes
self.up = nn.Upsample(scale_factor=2, mode='nearest')
self.conv1 = conv2d(in_planes, out_planes*2, 3, 1, 1, bias=False)
self.conv2 = conv2d(out_planes, out_planes*2, 3, 1, 1, bias=False)
which_bn = functools.partial(CCBN, which_linear=linear, input_size=z_dim)
self.bn1 = which_bn(2*out_planes)
self.bn2 = which_bn(2*out_planes)
self.act = GLU()
self.noise = NoiseInjection()
def forward(self, x, c):
# block 1
x = self.up(x)
x = self.conv1(x)
x = self.noise(x)
x = self.bn1(x, c)
x = self.act(x)
# block 2
x = self.conv2(x)
x = self.noise(x)
x = self.bn2(x, c)
x = self.act(x)
return x
class SEBlock(nn.Module):
def __init__(self, ch_in, ch_out):
super().__init__()
self.main = nn.Sequential(
nn.AdaptiveAvgPool2d(4),
conv2d(ch_in, ch_out, 4, 1, 0, bias=False),
Swish(),
conv2d(ch_out, ch_out, 1, 1, 0, bias=False),
nn.Sigmoid(),
)
def forward(self, feat_small, feat_big):
return feat_big * self.main(feat_small)
### Downblocks
class SeparableConv2d(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, bias=False):
super(SeparableConv2d, self).__init__()
self.depthwise = conv2d(in_channels, in_channels, kernel_size=kernel_size,
groups=in_channels, bias=bias, padding=1)
self.pointwise = conv2d(in_channels, out_channels,
kernel_size=1, bias=bias)
def forward(self, x):
out = self.depthwise(x)
out = self.pointwise(out)
return out
class DownBlock(nn.Module):
def __init__(self, in_planes, out_planes, separable=False):
super().__init__()
if not separable:
self.main = nn.Sequential(
conv2d(in_planes, out_planes, 4, 2, 1),
NormLayer(out_planes),
nn.LeakyReLU(0.2, inplace=True),
)
else:
self.main = nn.Sequential(
SeparableConv2d(in_planes, out_planes, 3),
NormLayer(out_planes),
nn.LeakyReLU(0.2, inplace=True),
nn.AvgPool2d(2, 2),
)
def forward(self, feat):
return self.main(feat)
class DownBlockPatch(nn.Module):
def __init__(self, in_planes, out_planes, separable=False):
super().__init__()
self.main = nn.Sequential(
DownBlock(in_planes, out_planes, separable),
conv2d(out_planes, out_planes, 1, 1, 0, bias=False),
NormLayer(out_planes),
nn.LeakyReLU(0.2, inplace=True),
)
def forward(self, feat):
return self.main(feat)
### CSM
class ResidualConvUnit(nn.Module):
def __init__(self, cin, activation, bn):
super().__init__()
self.conv = nn.Conv2d(cin, cin, kernel_size=3, stride=1, padding=1, bias=True)
self.skip_add = nn.quantized.FloatFunctional()
def forward(self, x):
return self.skip_add.add(self.conv(x), x)
class FeatureFusionBlock(nn.Module):
def __init__(self, features, activation, deconv=False, bn=False, expand=False, align_corners=True, lowest=False):
super().__init__()
self.deconv = deconv
self.align_corners = align_corners
self.expand = expand
out_features = features
if self.expand==True:
out_features = features//2
self.out_conv = nn.Conv2d(features, out_features, kernel_size=1, stride=1, padding=0, bias=True, groups=1)
self.skip_add = nn.quantized.FloatFunctional()
def forward(self, *xs):
output = xs[0]
if len(xs) == 2:
output = self.skip_add.add(output, xs[1])
output = nn.functional.interpolate(
output, scale_factor=2, mode="bilinear", align_corners=self.align_corners
)
output = self.out_conv(output)
return output
### Misc
class NoiseInjection(nn.Module):
def __init__(self):
super().__init__()
self.weight = nn.Parameter(torch.zeros(1), requires_grad=True)
def forward(self, feat, noise=None):
if noise is None:
batch, _, height, width = feat.shape
noise = torch.randn(batch, 1, height, width).to(feat.device)
return feat + self.weight * noise
class CCBN(nn.Module):
''' conditional batchnorm '''
def __init__(self, output_size, input_size, which_linear, eps=1e-5, momentum=0.1):
super().__init__()
self.output_size, self.input_size = output_size, input_size
# Prepare gain and bias layers
self.gain = which_linear(input_size, output_size)
self.bias = which_linear(input_size, output_size)
# epsilon to avoid dividing by 0
self.eps = eps
# Momentum
self.momentum = momentum
self.register_buffer('stored_mean', torch.zeros(output_size))
self.register_buffer('stored_var', torch.ones(output_size))
def forward(self, x, y):
# Calculate class-conditional gains and biases
gain = (1 + self.gain(y)).view(y.size(0), -1, 1, 1)
bias = self.bias(y).view(y.size(0), -1, 1, 1)
out = F.batch_norm(x, self.stored_mean, self.stored_var, None, None,
self.training, 0.1, self.eps)
return out * gain + bias
class Interpolate(nn.Module):
"""Interpolation module."""
def __init__(self, size, mode='bilinear', align_corners=False):
"""Init.
Args:
scale_factor (float): scaling
mode (str): interpolation mode
"""
super(Interpolate, self).__init__()
self.interp = nn.functional.interpolate
self.size = size
self.mode = mode
self.align_corners = align_corners
def forward(self, x):
"""Forward pass.
Args:
x (tensor): input
Returns:
tensor: interpolated data
"""
x = self.interp(
x,
size=self.size,
mode=self.mode,
align_corners=self.align_corners,
)
return x
|