Spaces:
Running on Zero
Running on Zero
File size: 9,612 Bytes
0d99394 | 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 | import torch
import torch.nn as nn
import torch.nn.functional as F
# ---------------------------
# 1) 基础Block: 2D卷积 + InstanceNorm + ReLU
# ---------------------------
class SameBlock2d(nn.Module):
"""
不改变分辨率的卷积层:Conv2d + InstanceNorm2d + ReLU
"""
def __init__(self, in_channels, out_channels, kernel_size=3, padding=1):
super(SameBlock2d, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding)
self.inorm = nn.InstanceNorm2d(out_channels, affine=True)
self.relu = nn.ReLU(inplace=True)
def forward(self, x):
x = self.conv(x)
x = self.inorm(x)
x = self.relu(x)
return x
class DownBlock2d(nn.Module):
"""
下采样层:Conv2d(stride=2) + InstanceNorm2d + ReLU,分辨率减半
"""
def __init__(self, in_channels, out_channels, kernel_size=3, padding=1):
super(DownBlock2d, self).__init__()
self.conv = nn.Conv2d(in_channels, out_channels,
kernel_size=kernel_size,
stride=2,
padding=padding)
self.inorm = nn.InstanceNorm2d(out_channels, affine=True)
self.relu = nn.ReLU(inplace=True)
def forward(self, x):
x = self.conv(x)
x = self.inorm(x)
x = self.relu(x)
return x
class ResBlock3d(nn.Module):
"""
Res block, preserve spatial resolution.
"""
def __init__(self, in_features, kernel_size, padding):
super(ResBlock3d, self).__init__()
self.conv1 = nn.Conv3d(in_channels=in_features, out_channels=in_features, kernel_size=kernel_size, padding=padding)
self.conv2 = nn.Conv3d(in_channels=in_features, out_channels=in_features, kernel_size=kernel_size, padding=padding)
self.norm1 = nn.InstanceNorm3d(in_features, affine=True)
self.norm2 = nn.InstanceNorm3d(in_features, affine=True)
def forward(self, x):
out = self.norm1(x)
out = F.relu(out)
out = self.conv1(out)
out = self.norm2(out)
out = F.relu(out)
out = self.conv2(out)
out = out + x
return out
# ---------------------------
# 2) ID Encoder: 更深更宽
# 输出 shape: B x 512 x 16 x 16
# ---------------------------
class IdentityEncoder(nn.Module):
def __init__(self,
image_channel=3,
base_channels=32,
max_channels =128,
num_down_blocks=4):
super(IdentityEncoder, self).__init__()
# 1) 初始SameBlock
self.initial = SameBlock2d(image_channel, base_channels)
# 2) 下采样
down_blocks = []
in_ch = base_channels
for i in range(num_down_blocks):
out_ch = min(max_channels, in_ch * 2)
down_blocks.append(DownBlock2d(in_ch, out_ch))
in_ch = out_ch
self.down_blocks = nn.ModuleList(down_blocks)
# 3) 最终 1x1 卷积,通道固定为 max_channels
self.final_conv = nn.Conv2d(in_ch, max_channels, kernel_size=1, stride=1)
def forward(self, x):
"""
x shape: B x 3 x 256 x 256
output shape: B x 512 x 16 x 16
"""
x = self.initial(x) # -> B x 64 x 256 x 256
for block in self.down_blocks:
x = block(x) # 最终 -> B x 128 x 16 x 16
x = self.final_conv(x) # 保持 B x 128 x 16 x 16
return x
# ---------------------------
# 3) Attribute Encoder: 更小,下采样较少
# 输出 shape: B x 256 x 64 x 64
# ---------------------------
class AttributeEncoder(nn.Module):
def __init__(self,
image_channel=3,
base_channels=32,
max_channels=128,
num_down_blocks=2):
super(AttributeEncoder, self).__init__()
# 1) 初始 SameBlock
self.initial = SameBlock2d(image_channel, base_channels)
# 2) 下采样
down_blocks = []
in_ch = base_channels
for i in range(num_down_blocks):
out_ch = min(max_channels, in_ch * 2)
down_blocks.append(DownBlock2d(in_ch, out_ch))
in_ch = out_ch
self.down_blocks = nn.ModuleList(down_blocks)
# 3) 最终卷积,得到输出通道=64
self.final_conv = nn.Conv2d(in_ch, max_channels, kernel_size=1, stride=1)
def forward(self, x):
"""
x shape: B x 3 x 256 x 256
output shape: B x 256 x 64 x 64
"""
x = self.initial(x) # -> B x 32 x 256 x 256
for block in self.down_blocks:
x = block(x) # -> B x 128 x 64 x 64 (after 2 downblocks)
x = self.final_conv(x) # -> B x 256 x 64 x 64
return x
# ---------------------------
# 4) Decoder:
# - 包含: 上采样(对ID特征), concat, 若干卷积, reshape
# - 最终输出: B x 32 x 16 x 64 x 64 (即通道=512, reshape为(32,16))
# ---------------------------
class Decoder(nn.Module):
def __init__(self,
id_in_channels=128, # ID Encoder输出: 512
attr_in_channels=128, # Attribute Encoder输出: 256
out_channels=512, # Decoder最终输出通道
reshape_channel=32,
reshape_depth=16,
num_resblocks=4):
super(Decoder, self).__init__()
# 先定义上采样方法 (这里简单用最近邻插值,你也可以换成转置卷积等)
# self.upsample_id = nn.Upsample(scale_factor=4, mode='nearest')
self.upsample_id = nn.ConvTranspose2d(
in_channels=id_in_channels,
out_channels=id_in_channels, # 保持通道数不变
kernel_size=4, # 卷积核大小
stride=4, # 步长,scale_factor=4
padding=0, # 无填充
output_padding=0 # 无额外填充
)
# scale_factor=4: 16 -> 64
# 拼接后的通道: 128 + 128 =256
in_channels = id_in_channels + attr_in_channels
# 几个卷积层,用 InstanceNorm 提升容量
self.conv1 = nn.Conv2d(in_channels, 128, kernel_size=3, padding=1)
self.inorm1 = nn.InstanceNorm2d(128, affine=True)
self.relu1 = nn.ReLU(inplace=True)
self.conv2 = nn.Conv2d(128, 256, kernel_size=3, padding=1)
self.inorm2 = nn.InstanceNorm2d(256, affine=True)
self.relu2 = nn.ReLU(inplace=True)
# 最终再做一个 conv, 输出 out_channels=512
self.conv3 = nn.Conv2d(256, out_channels, kernel_size=3, padding=1)
self.inorm3 = nn.InstanceNorm2d(out_channels, affine=True)
self.relu3 = nn.ReLU(inplace=True)
self.reshape_channel = reshape_channel
self.reshape_depth = reshape_depth
self.resblocks_3d = torch.nn.Sequential()
for i in range(num_resblocks):
self.resblocks_3d.add_module('3dr' + str(i), ResBlock3d(reshape_channel, kernel_size=3, padding=1))
def forward(self, id_feat, attr_feat):
"""
id_feat: B x 512 x 16 x 16
attr_feat: B x 256 x 64 x 64
目标:
1) upsample id_feat -> B x 512 x 64 x 64
2) concat -> B x 768 x 64 x 64
3) conv -> B x 512 x 64 x 64
4) reshape -> B x 32 x 16 x 64 x 64
"""
# 1) 上采样 ID特征: 16 -> 64
id_up = self.upsample_id(id_feat) # B x 512 x 64 x 64
# 2) 拼接
x = torch.cat([id_up, attr_feat], dim=1) # B x 768 x 64 x 64
# 3) 经过若干卷积层
x = self.relu1(self.inorm1(self.conv1(x))) # -> B x 512 x 64 x 64
x = self.relu2(self.inorm2(self.conv2(x))) # -> B x 512 x 64 x 64
x = self.relu3(self.inorm3(self.conv3(x))) # -> B x 512 x 64 x 64
# 4) reshape -> 3D
b, c, h, w = x.shape # c=512, h=64, w=64
assert c == self.reshape_channel * self.reshape_depth, (
f"通道数{c}与reshape要求{self.reshape_channel}*{self.reshape_depth}不符!"
)
out_3d = x.view(b, self.reshape_channel, self.reshape_depth, h, w)
# => B x 32 x 16 x 64 x 64
out_3d = self.resblocks_3d(out_3d) # -> B x 32 x 16 x 64 x 64
return out_3d
# ---------------------------
# 5) 整合: DualEncoderDecoder
# ---------------------------
class DualEncoderDecoder(nn.Module):
def __init__(self):
super(DualEncoderDecoder, self).__init__()
self.id_encoder = IdentityEncoder(
image_channel=3,
base_channels=64, # 初始通道
max_channels=512,
num_down_blocks=4 # 共 4 次下采样 => 16x16
)
self.attr_encoder = AttributeEncoder(
image_channel=3,
base_channels=32,
max_channels=256,
num_down_blocks=2 # 2 次下采样 => 64x64
)
self.decoder = Decoder(
id_in_channels=512,
attr_in_channels=256,
out_channels=512, # 最终要 reshape 成 32*16=512
reshape_channel=32,
reshape_depth=16
)
def forward(self, id_img, attr_img):
"""
id_img: B x 3 x 256 x 256
attr_img: B x 3 x 256 x 256
最终输出: B x 32 x 16 x 64 x 64
"""
id_feat = self.id_encoder(id_img) # => B x 512 x 16 x 16
attr_feat = self.attr_encoder(attr_img) # => B x 256 x 64 x 64
out_3d = self.decoder(id_feat, attr_feat) # => B x 32 x 16 x 64 x 64
return out_3d |