File size: 8,944 Bytes
14ce5a9 |
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 |
import torch
import torch.nn as nn
from einops import rearrange
import torch.nn.functional as F
def swish(x):
# swish
return x * torch.sigmoid(x)
class ResBlock(nn.Module):
def __init__(
self,
in_filters,
out_filters,
use_conv_shortcut=False,
use_agn=False,
) -> None:
super().__init__()
self.in_filters = in_filters
self.out_filters = out_filters
self.use_conv_shortcut = use_conv_shortcut
self.use_agn = use_agn
if not use_agn: ## agn is GroupNorm likewise skip it if has agn before
self.norm1 = nn.GroupNorm(32, in_filters, eps=1e-6)
self.norm2 = nn.GroupNorm(32, out_filters, eps=1e-6)
self.conv1 = nn.Conv2d(
in_filters, out_filters, kernel_size=(3, 3), padding=1, bias=False
)
self.conv2 = nn.Conv2d(
out_filters, out_filters, kernel_size=(3, 3), padding=1, bias=False
)
if in_filters != out_filters:
if self.use_conv_shortcut:
self.conv_shortcut = nn.Conv2d(
in_filters, out_filters, kernel_size=(3, 3), padding=1, bias=False
)
else:
self.nin_shortcut = nn.Conv2d(
in_filters, out_filters, kernel_size=(1, 1), padding=0, bias=False
)
def forward(self, x, **kwargs):
residual = x
if not self.use_agn:
x = self.norm1(x)
x = swish(x)
x = self.conv1(x)
x = self.norm2(x)
x = swish(x)
x = self.conv2(x)
if self.in_filters != self.out_filters:
if self.use_conv_shortcut:
residual = self.conv_shortcut(residual)
else:
residual = self.nin_shortcut(residual)
return x + residual
class Encoder(nn.Module):
def __init__(
self,
*,
ch,
out_ch,
in_channels,
num_res_blocks,
z_channels,
ch_mult=(1, 2, 2, 4),
resolution,
double_z=False,
):
super().__init__()
self.in_channels = in_channels
self.z_channels = z_channels
self.resolution = resolution
self.num_res_blocks = num_res_blocks
self.num_blocks = len(ch_mult)
self.conv_in = nn.Conv2d(
in_channels, ch, kernel_size=(3, 3), padding=1, bias=False
)
## construct the model
self.down = nn.ModuleList()
in_ch_mult = (1,) + tuple(ch_mult)
for i_level in range(self.num_blocks):
block = nn.ModuleList()
block_in = ch * in_ch_mult[i_level] # [1, 1, 2, 2, 4]
block_out = ch * ch_mult[i_level] # [1, 2, 2, 4]
for _ in range(self.num_res_blocks):
block.append(ResBlock(block_in, block_out))
block_in = block_out
down = nn.Module()
down.block = block
if i_level < self.num_blocks - 1:
down.downsample = nn.Conv2d(
block_out, block_out, kernel_size=(3, 3), stride=(2, 2), padding=1
)
self.down.append(down)
### mid
self.mid_block = nn.ModuleList()
for res_idx in range(self.num_res_blocks):
self.mid_block.append(ResBlock(block_in, block_in))
### end
self.norm_out = nn.GroupNorm(32, block_out, eps=1e-6)
self.conv_out = nn.Conv2d(block_out, z_channels, kernel_size=(1, 1))
def forward(self, x):
## down
x = self.conv_in(x)
for i_level in range(self.num_blocks):
for i_block in range(self.num_res_blocks):
x = self.down[i_level].block[i_block](x)
if i_level < self.num_blocks - 1:
x = self.down[i_level].downsample(x)
## mid
for res in range(self.num_res_blocks):
x = self.mid_block[res](x)
x = self.norm_out(x)
x = swish(x)
x = self.conv_out(x)
return x
class Decoder(nn.Module):
def __init__(
self,
*,
ch,
out_ch,
in_channels,
num_res_blocks,
z_channels,
ch_mult=(1, 2, 2, 4),
resolution,
double_z=False,
) -> None:
super().__init__()
self.ch = ch
self.num_blocks = len(ch_mult)
self.num_res_blocks = num_res_blocks
self.resolution = resolution
self.in_channels = in_channels
block_in = ch * ch_mult[self.num_blocks - 1]
self.conv_in = nn.Conv2d(
z_channels, block_in, kernel_size=(3, 3), padding=1, bias=True
)
self.mid_block = nn.ModuleList()
for res_idx in range(self.num_res_blocks):
self.mid_block.append(ResBlock(block_in, block_in))
self.up = nn.ModuleList()
self.adaptive = nn.ModuleList()
for i_level in reversed(range(self.num_blocks)):
block = nn.ModuleList()
block_out = ch * ch_mult[i_level]
self.adaptive.insert(0, AdaptiveGroupNorm(z_channels, block_in))
for i_block in range(self.num_res_blocks):
# if i_block == 0:
# block.append(ResBlock(block_in, block_out, use_agn=True))
# else:
block.append(ResBlock(block_in, block_out))
block_in = block_out
up = nn.Module()
up.block = block
if i_level > 0:
up.upsample = Upsampler(block_in)
self.up.insert(0, up)
self.norm_out = nn.GroupNorm(32, block_in, eps=1e-6)
self.conv_out = nn.Conv2d(block_in, out_ch, kernel_size=(3, 3), padding=1)
def forward(self, z):
style = z.clone() # for adaptive groupnorm
z = self.conv_in(z)
## mid
for res in range(self.num_res_blocks):
z = self.mid_block[res](z)
## upsample
for i_level in reversed(range(self.num_blocks)):
### pass in each resblock first adaGN
z = self.adaptive[i_level](z, style)
for i_block in range(self.num_res_blocks):
z = self.up[i_level].block[i_block](z)
if i_level > 0:
z = self.up[i_level].upsample(z)
z = self.norm_out(z)
z = swish(z)
z = self.conv_out(z)
return z
def depth_to_space(x: torch.Tensor, block_size: int) -> torch.Tensor:
"""Depth-to-Space DCR mode (depth-column-row) core implementation.
Args:
x (torch.Tensor): input tensor. The channels-first (*CHW) layout is supported.
block_size (int): block side size
"""
# check inputs
if x.dim() < 3:
raise ValueError(
f"Expecting a channels-first (*CHW) tensor of at least 3 dimensions"
)
c, h, w = x.shape[-3:]
s = block_size**2
if c % s != 0:
raise ValueError(
f"Expecting a channels-first (*CHW) tensor with C divisible by {s}, but got C={c} channels"
)
outer_dims = x.shape[:-3]
# splitting two additional dimensions from the channel dimension
x = x.view(-1, block_size, block_size, c // s, h, w)
# putting the two new dimensions along H and W
x = x.permute(0, 3, 4, 1, 5, 2)
# merging the two new dimensions with H and W
x = x.contiguous().view(*outer_dims, c // s, h * block_size, w * block_size)
return x
class Upsampler(nn.Module):
def __init__(self, dim, dim_out=None):
super().__init__()
dim_out = dim * 4
self.conv1 = nn.Conv2d(dim, dim_out, (3, 3), padding=1)
self.depth2space = depth_to_space
def forward(self, x):
"""
input_image: [B C H W]
"""
out = self.conv1(x)
out = self.depth2space(out, block_size=2)
return out
class AdaptiveGroupNorm(nn.Module):
def __init__(self, z_channel, in_filters, num_groups=32, eps=1e-6):
super().__init__()
self.gn = nn.GroupNorm(
num_groups=32, num_channels=in_filters, eps=eps, affine=False
)
# self.lin = nn.Linear(z_channels, in_filters * 2)
self.gamma = nn.Linear(z_channel, in_filters)
self.beta = nn.Linear(z_channel, in_filters)
self.eps = eps
def forward(self, x, quantizer):
B, C, _, _ = x.shape
# quantizer = F.adaptive_avg_pool2d(quantizer, (1, 1))
### calcuate var for scale
scale = rearrange(quantizer, "b c h w -> b c (h w)")
scale = scale.var(dim=-1) + self.eps # not unbias
scale = scale.sqrt()
scale = self.gamma(scale).view(B, C, 1, 1)
### calculate mean for bias
bias = rearrange(quantizer, "b c h w -> b c (h w)")
bias = bias.mean(dim=-1)
bias = self.beta(bias).view(B, C, 1, 1)
x = self.gn(x)
x = scale * x + bias
return x
|