这是一个基于 U-Net 的模型,用于预测噪声。
U-Net 是从模型图中的 U 形中获取它的名字。它通过逐步降低(减半)要素图分辨率,然后提高分辨率来处理给定的图像。每种分辨率都有直通连接。
此实现包含对原始 U-Net(残差块、多头注意)的大量修改,还添加了时间步长嵌入。
24import math
25from typing import Optional, Tuple, Union, List
26
27import torch
28from torch import nn
29
30from labml_helpers.module import Module
33class Swish(Module):
40 def forward(self, x):
41 return x * torch.sigmoid(x)
44class TimeEmbedding(nn.Module):
n_channels
是嵌入中的维数49 def __init__(self, n_channels: int):
53 super().__init__()
54 self.n_channels = n_channels
第一个线性层
56 self.lin1 = nn.Linear(self.n_channels // 4, self.n_channels)
激活
58 self.act = Swish()
第二个线性层
60 self.lin2 = nn.Linear(self.n_channels, self.n_channels)
62 def forward(self, t: torch.Tensor):
72 half_dim = self.n_channels // 8
73 emb = math.log(10_000) / (half_dim - 1)
74 emb = torch.exp(torch.arange(half_dim, device=t.device) * -emb)
75 emb = t[:, None] * emb[None, :]
76 emb = torch.cat((emb.sin(), emb.cos()), dim=1)
使用 MLP 进行转型
79 emb = self.act(self.lin1(emb))
80 emb = self.lin2(emb)
83 return emb
86class ResidualBlock(Module):
in_channels
是输入通道的数量out_channels
是输入通道的数量time_channels
是时间步 () 嵌入中的通道数n_groups
是用于组标准化的组数dropout
是辍学率94 def __init__(self, in_channels: int, out_channels: int, time_channels: int,
95 n_groups: int = 32, dropout: float = 0.1):
103 super().__init__()
组归一化和第一个卷积层
105 self.norm1 = nn.GroupNorm(n_groups, in_channels)
106 self.act1 = Swish()
107 self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=(3, 3), padding=(1, 1))
组归一化和第二个卷积层
110 self.norm2 = nn.GroupNorm(n_groups, out_channels)
111 self.act2 = Swish()
112 self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=(3, 3), padding=(1, 1))
如果输入通道的数量不等于输出通道的数量,我们必须投影快捷方式连接
116 if in_channels != out_channels:
117 self.shortcut = nn.Conv2d(in_channels, out_channels, kernel_size=(1, 1))
118 else:
119 self.shortcut = nn.Identity()
用于时间嵌入的线性层
122 self.time_emb = nn.Linear(time_channels, out_channels)
123 self.time_act = Swish()
124
125 self.dropout = nn.Dropout(dropout)
x
有形状[batch_size, in_channels, height, width]
t
有形状[batch_size, time_channels]
127 def forward(self, x: torch.Tensor, t: torch.Tensor):
第一个卷积层
133 h = self.conv1(self.act1(self.norm1(x)))
添加时间嵌入
135 h += self.time_emb(self.time_act(t))[:, :, None, None]
第二个卷积层
137 h = self.conv2(self.dropout(self.act2(self.norm2(h))))
添加快捷方式连接并返回
140 return h + self.shortcut(x)
150 def __init__(self, n_channels: int, n_heads: int = 1, d_k: int = None, n_groups: int = 32):
157 super().__init__()
默认d_k
160 if d_k is None:
161 d_k = n_channels
归一化层
163 self.norm = nn.GroupNorm(n_groups, n_channels)
查询、键和值的投影
165 self.projection = nn.Linear(n_channels, n_heads * d_k * 3)
用于最终变换的线性层
167 self.output = nn.Linear(n_heads * d_k, n_channels)
缩放点产品注意力
169 self.scale = d_k ** -0.5
171 self.n_heads = n_heads
172 self.d_k = d_k
x
有形状[batch_size, in_channels, height, width]
t
有形状[batch_size, time_channels]
174 def forward(self, x: torch.Tensor, t: Optional[torch.Tensor] = None):
t
未使用,但它保留在参数中,因为要与注意层函数签名匹配ResidualBlock
。
181 _ = t
塑造身材
183 batch_size, n_channels, height, width = x.shape
改x
成形状[batch_size, seq, n_channels]
185 x = x.view(batch_size, n_channels, -1).permute(0, 2, 1)
获取查询、键和值(串联)并将其调整为[batch_size, seq, n_heads, 3 * d_k]
187 qkv = self.projection(x).view(batch_size, -1, self.n_heads, 3 * self.d_k)
拆分查询、键和值。他们每个人都会有形状[batch_size, seq, n_heads, d_k]
189 q, k, v = torch.chunk(qkv, 3, dim=-1)
计算缩放的点积
191 attn = torch.einsum('bihd,bjhd->bijh', q, k) * self.scale
顺序维度上的 Softmax
193 attn = attn.softmax(dim=2)
乘以值
195 res = torch.einsum('bijh,bjhd->bihd', attn, v)
重塑为[batch_size, seq, n_heads * d_k]
197 res = res.view(batch_size, -1, self.n_heads * self.d_k)
变换为[batch_size, seq, n_channels]
199 res = self.output(res)
添加跳过连接
202 res += x
改成形状[batch_size, in_channels, height, width]
205 res = res.permute(0, 2, 1).view(batch_size, n_channels, height, width)
208 return res
211class DownBlock(Module):
218 def __init__(self, in_channels: int, out_channels: int, time_channels: int, has_attn: bool):
219 super().__init__()
220 self.res = ResidualBlock(in_channels, out_channels, time_channels)
221 if has_attn:
222 self.attn = AttentionBlock(out_channels)
223 else:
224 self.attn = nn.Identity()
226 def forward(self, x: torch.Tensor, t: torch.Tensor):
227 x = self.res(x, t)
228 x = self.attn(x)
229 return x
232class UpBlock(Module):
239 def __init__(self, in_channels: int, out_channels: int, time_channels: int, has_attn: bool):
240 super().__init__()
输入之in_channels + out_channels
所以有,是因为我们将 U-Net 前半部分相同分辨率的输出连接起来
243 self.res = ResidualBlock(in_channels + out_channels, out_channels, time_channels)
244 if has_attn:
245 self.attn = AttentionBlock(out_channels)
246 else:
247 self.attn = nn.Identity()
249 def forward(self, x: torch.Tensor, t: torch.Tensor):
250 x = self.res(x, t)
251 x = self.attn(x)
252 return x
255class MiddleBlock(Module):
263 def __init__(self, n_channels: int, time_channels: int):
264 super().__init__()
265 self.res1 = ResidualBlock(n_channels, n_channels, time_channels)
266 self.attn = AttentionBlock(n_channels)
267 self.res2 = ResidualBlock(n_channels, n_channels, time_channels)
269 def forward(self, x: torch.Tensor, t: torch.Tensor):
270 x = self.res1(x, t)
271 x = self.attn(x)
272 x = self.res2(x, t)
273 return x
276class Upsample(nn.Module):
281 def __init__(self, n_channels):
282 super().__init__()
283 self.conv = nn.ConvTranspose2d(n_channels, n_channels, (4, 4), (2, 2), (1, 1))
285 def forward(self, x: torch.Tensor, t: torch.Tensor):
t
未使用,但它保留在参数中,因为要与注意层函数签名匹配ResidualBlock
。
288 _ = t
289 return self.conv(x)
292class Downsample(nn.Module):
297 def __init__(self, n_channels):
298 super().__init__()
299 self.conv = nn.Conv2d(n_channels, n_channels, (3, 3), (2, 2), (1, 1))
301 def forward(self, x: torch.Tensor, t: torch.Tensor):
t
未使用,但它保留在参数中,因为要与注意层函数签名匹配ResidualBlock
。
304 _ = t
305 return self.conv(x)
308class UNet(Module):
image_channels
是图像中的通道数。对于 RGB。n_channels
是初始特征图中我们将图像转换为的通道数ch_mults
是每种分辨率下的通道编号列表。频道的数量是ch_mults[i] * n_channels
is_attn
是一个布尔值列表,用于指示是否在每个分辨率下使用注意力n_blocks
是每种分辨UpDownBlocks
率的数字313 def __init__(self, image_channels: int = 3, n_channels: int = 64,
314 ch_mults: Union[Tuple[int, ...], List[int]] = (1, 2, 2, 4),
315 is_attn: Union[Tuple[bool, ...], List[bool]] = (False, False, True, True),
316 n_blocks: int = 2):
324 super().__init__()
分辨率数量
327 n_resolutions = len(ch_mults)
将图像投影到要素地图中
330 self.image_proj = nn.Conv2d(image_channels, n_channels, kernel_size=(3, 3), padding=(1, 1))
时间嵌入层。时间嵌入有n_channels * 4
频道
333 self.time_emb = TimeEmbedding(n_channels * 4)
336 down = []
频道数量
338 out_channels = in_channels = n_channels
对于每种分辨率
340 for i in range(n_resolutions):
此分辨率下的输出声道数
342 out_channels = in_channels * ch_mults[i]
添加n_blocks
344 for _ in range(n_blocks):
345 down.append(DownBlock(in_channels, out_channels, n_channels * 4, is_attn[i]))
346 in_channels = out_channels
除最后一个分辨率之外的所有分辨率都向下采样
348 if i < n_resolutions - 1:
349 down.append(Downsample(in_channels))
组合这组模块
352 self.down = nn.ModuleList(down)
中间方块
355 self.middle = MiddleBlock(out_channels, n_channels * 4, )
358 up = []
频道数量
360 in_channels = out_channels
对于每种分辨率
362 for i in reversed(range(n_resolutions)):
n_blocks
以相同的分辨率
364 out_channels = in_channels
365 for _ in range(n_blocks):
366 up.append(UpBlock(in_channels, out_channels, n_channels * 4, is_attn[i]))
减少信道数量的最终区块
368 out_channels = in_channels // ch_mults[i]
369 up.append(UpBlock(in_channels, out_channels, n_channels * 4, is_attn[i]))
370 in_channels = out_channels
除最后一个以外的所有分辨率向上采样
372 if i > 0:
373 up.append(Upsample(in_channels))
组合这组模块
376 self.up = nn.ModuleList(up)
最终归一化和卷积层
379 self.norm = nn.GroupNorm(8, n_channels)
380 self.act = Swish()
381 self.final = nn.Conv2d(in_channels, image_channels, kernel_size=(3, 3), padding=(1, 1))
x
有形状[batch_size, in_channels, height, width]
t
有形状[batch_size]
383 def forward(self, x: torch.Tensor, t: torch.Tensor):
获取时间步长嵌入
390 t = self.time_emb(t)
获取图像投影
393 x = self.image_proj(x)
h
将以每种分辨率存储输出以进行跳过连接
396 h = [x]
U-Net 的上半年
398 for m in self.down:
399 x = m(x, t)
400 h.append(x)
中间(底部)
403 x = self.middle(x, t)
U-Net 的下半场
406 for m in self.up:
407 if isinstance(m, Upsample):
408 x = m(x, t)
409 else:
从 U-Net 的前半部分获取跳过连接并连接
411 s = h.pop()
412 x = torch.cat((x, s), dim=1)
414 x = m(x, t)
最终归一化和卷积
417 return self.final(self.act(self.norm(x)))