-
Notifications
You must be signed in to change notification settings - Fork 0
/
ResAttUNet.py
215 lines (160 loc) · 7.17 KB
/
ResAttUNet.py
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
import torch
from torch import nn
import torch.nn.functional as F
import math
class Swish(nn.Module):
def forward(self, x):
return x * torch.sigmoid(x)
class TimeEmbedding(nn.Module):
def __init__(self, n_channels: int):
super().__init__()
self.n_channels = n_channels
self.lin1 = nn.Linear(self.n_channels // 4, self.n_channels)
self.act = Swish()
self.lin2 = nn.Linear(self.n_channels, self.n_channels)
def forward(self, t: torch.Tensor):
half_dim = self.n_channels // 8
emb = math.log(10_000) / (half_dim - 1)
emb = torch.exp(torch.arange(half_dim, device=t.device) * -emb)
emb = t[:, None] * emb[None, :]
emb = torch.cat((emb.sin(), emb.cos()), dim=1)
emb = self.act(self.lin1(emb))
emb = self.lin2(emb)
return emb
class AttentionBlock(nn.Module):
def __init__(self, n_channels: int, n_heads: int = 1, d_k: int = None, n_groups: int = 32):
super().__init__()
if d_k is None:
d_k = n_channels
self.norm = nn.GroupNorm(n_groups, n_channels)
self.projection = nn.Linear(n_channels, n_heads * d_k * 3)
self.output = nn.Linear(n_heads * d_k, n_channels)
self.scale = d_k ** -0.5
self.n_heads = n_heads
self.d_k = d_k
def forward(self, x: torch.Tensor):
batch_size, n_channels, height, width = x.shape
x = x.view(batch_size, n_channels, -1).permute(0, 2, 1)
qkv = self.projection(x).view(batch_size, -1, self.n_heads, 3 * self.d_k)
q, k, v = torch.chunk(qkv, 3, dim=-1)
attn = torch.einsum('bihd,bjhd->bijh', q, k) * self.scale
attn = attn.softmax(dim=1)
res = torch.einsum('bijh,bjhd->bihd', attn, v)
res = res.view(batch_size, -1, self.n_heads * self.d_k)
res = self.output(res)
res = res + x
res = res.permute(0, 2, 1).view(batch_size, n_channels, height, width)
return res
class ConvBlock(nn.Module):
def __init__(self, in_channels, out_channels, time_channels):
super(ConvBlock, self).__init__()
self.norm1 = nn.GroupNorm(32, in_channels)
self.act1 = Swish()
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
#self.bn1 = nn.BatchNorm2d(out_channels)
self.norm2 = nn.GroupNorm(32, out_channels)
self.act2 = Swish()
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
#self.bn2 = nn.BatchNorm2d(out_channels)
self.time_emb = nn.Linear(time_channels, out_channels)
def forward(self, x, t):
x = self.conv1(self.act1(self.norm1(x)))
x = x + self.time_emb(t)[:, :, None, None]
x = self.conv2(self.act2(self.norm2(x)))
return x
class ResidualBlock(nn.Module):
def __init__(self, in_channels, out_channels, time_channels):
super(ResidualBlock, self).__init__()
self.conv_block = ConvBlock(in_channels, out_channels, time_channels)
self.shortcut = nn.Sequential()
if in_channels != out_channels:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False),
)
else:
self.shortcut = nn.Sequential(
nn.Identity()
)
def forward(self, x, t):
return self.conv_block(x,t) + self.shortcut(x)
class EncoderBlock(nn.Module):
def __init__(self, in_channels, out_channels, time_channels, has_attn: bool):
super(EncoderBlock, self).__init__()
self.res_block = ResidualBlock(in_channels, out_channels, time_channels)
if has_attn:
self.attn = AttentionBlock(out_channels)
else:
self.attn = nn.Identity()
self.res_block2 = ResidualBlock(out_channels, out_channels, time_channels)
if has_attn:
self.attn2 = AttentionBlock(out_channels)
else:
self.attn2 = nn.Identity()
#self.pool = nn.MaxPool2d(2)
self.conv = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=2, padding=1)
def forward(self, x, t):
x = self.res_block(x,t)
x = self.attn(x)
x = self.res_block2(x,t)
x = self.attn2(x)
#p = self.pool(x)
p = self.conv(x)
return x, p
class DecoderBlock(nn.Module):
def __init__(self, in_channels, middle_channels, out_channels, time_channels, has_attn: bool):
super(DecoderBlock, self).__init__()
#self.up = nn.ConvTranspose2d(in_channels, middle_channels, kernel_size=2, stride=2)
self.up = nn.ConvTranspose2d(in_channels, middle_channels, kernel_size=4, stride=2, padding=1)
self.res_block = ResidualBlock(middle_channels + out_channels, out_channels, time_channels)
if has_attn:
self.attn = AttentionBlock(out_channels)
else:
self.attn = nn.Identity()
self.res_block2 = ResidualBlock(out_channels, out_channels, time_channels)
if has_attn:
self.attn2 = AttentionBlock(out_channels)
else:
self.attn2 = nn.Identity()
def forward(self, x, skip, t):
x = self.up(x)
x = torch.cat([x, skip], dim=1)
x = self.res_block(x,t)
x = self.attn(x)
x = self.res_block2(x,t)
x = self.attn2(x)
return x
class ResidualAttentionUNet(nn.Module):
def __init__(self):
super(ResidualAttentionUNet, self).__init__()
self.time_emb = TimeEmbedding(64 * 4)
self.start = nn.Conv2d(3, 32, kernel_size=3, padding=1) # 32
self.encoder1 = EncoderBlock(32, 32, 64 * 4, False) # 32,32
self.encoder2 = EncoderBlock(32, 64, 64 * 4, False) # 32,64
self.encoder3 = EncoderBlock(64, 128, 64 * 4, True) # 64,128
self.encoder4 = EncoderBlock(128, 512, 64 * 4, True) #128,512
self.bridge = ResidualBlock(512, 512, 64 * 4) #512,512
self.att = AttentionBlock(512)
self.bridge2 = ResidualBlock(512, 512, 64 * 4) #512,512
self.decoder1 = DecoderBlock(512, 512, 512, 64 * 4, True) # 512
self.decoder2 = DecoderBlock(512, 128, 128, 64 * 4, True)
self.decoder3 = DecoderBlock(128, 64, 64, 64 * 4, False)
self.decoder4 = DecoderBlock(64, 32, 32, 64 * 4, False)
self.norm = nn.GroupNorm(8, 32)
self.act = Swish()
self.final = nn.Conv2d(32, 3, kernel_size=1)
def forward(self, x, t):
t = self.time_emb(t)
x = self.start(x) # relu is inside the residual block
s1, p1 = self.encoder1(x,t)
s2, p2 = self.encoder2(p1,t)
s3, p3 = self.encoder3(p2,t)
s4, p4 = self.encoder4(p3,t)
b = self.bridge(p4,t)
b = self.att(b)
b = self.bridge2(b,t)
d1 = self.decoder1(b, s4, t)
d2 = self.decoder2(d1, s3, t)
d3 = self.decoder3(d2, s2, t)
d4 = self.decoder4(d3, s1, t)
output = self.final(self.act(self.norm(d4)))
return output