File size: 8,536 Bytes
b876688
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from torch import nn
from torch.nn import functional as F
from attention import SelfAttention, CrossAttention

class TimeEmbedding(nn.Module):
    def __init__(self, n_embd):
        super().__init__()
        self.linear_1 = nn.Linear(n_embd, 4 * n_embd)
        self.linear_2 = nn.Linear(4 * n_embd, 4 * n_embd)

    def forward(self, x):
        x = F.silu(self.linear_1(x))
        return self.linear_2(x)

class SqueezeExcitation(nn.Module):
    def __init__(self, channels, reduction=16):
        super().__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.fc = nn.Sequential(
            nn.Linear(channels, channels // reduction, bias=False),
            nn.ReLU(inplace=True),
            nn.Linear(channels // reduction, channels, bias=False),
            nn.Sigmoid()
        )

    def forward(self, x):
        b, c, _, _ = x.size()
        y = self.avg_pool(x).view(b, c)
        y = self.fc(y).view(b, c, 1, 1)
        return x * y.expand_as(x)

class UNET_ResidualBlock(nn.Module):
    def __init__(self, in_channels, out_channels, n_time=1280, use_se=False):
        super().__init__()
        self.groupnorm_feature = nn.GroupNorm(32, in_channels)
        self.conv_feature = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
        self.linear_time = nn.Linear(n_time, out_channels)
        self.groupnorm_merged = nn.GroupNorm(32, out_channels)
        self.conv_merged = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
        self.residual_layer = nn.Identity() if in_channels == out_channels else nn.Conv2d(in_channels, out_channels, kernel_size=1, padding=0)
        
        # Add Squeeze-Excitation blocks only if use_se is True
        self.use_se = use_se
        if use_se:
            self.se1 = SqueezeExcitation(out_channels)
            self.se2 = SqueezeExcitation(out_channels)
    
    def forward(self, feature, time):
        residue = feature
        feature = F.silu(self.groupnorm_feature(feature))
        feature = self.conv_feature(feature)
        if self.use_se:
            feature = self.se1(feature)  # Apply SE after first conv
        
        time = self.linear_time(F.silu(time))
        merged = feature + time.unsqueeze(-1).unsqueeze(-1)
        merged = F.silu(self.groupnorm_merged(merged))
        merged = self.conv_merged(merged)
        if self.use_se:
            merged = self.se2(merged)  # Apply SE after second conv
        
        return merged + self.residual_layer(residue)

class UNET_AttentionBlock(nn.Module):
    def __init__(self, n_head: int, n_embd: int, d_context=768):
        super().__init__()
        channels = n_head * n_embd
        self.groupnorm = nn.GroupNorm(32, channels, eps=1e-6)
        self.conv_input = nn.Conv2d(channels, channels, kernel_size=1, padding=0)
        self.layernorm_1 = nn.LayerNorm(channels)
        self.attention_1 = SelfAttention(n_head, channels, in_proj_bias=False)
        self.layernorm_2 = nn.LayerNorm(channels)
        self.attention_2 = CrossAttention(n_head, channels, d_context, in_proj_bias=False)
        self.layernorm_3 = nn.LayerNorm(channels)
        self.linear_geglu_1 = nn.Linear(channels, 4 * channels * 2)
        self.linear_geglu_2 = nn.Linear(4 * channels, channels)
        self.conv_output = nn.Conv2d(channels, channels, kernel_size=1, padding=0)
    
    def forward(self, x, context):
        residue_long = x
        x = self.conv_input(self.groupnorm(x))
        n, c, h, w = x.shape
        x = x.view((n, c, h * w)).transpose(-1, -2)
        residue_short = x
        x = self.attention_1(self.layernorm_1(x)) + residue_short
        residue_short = x
        x = self.attention_2(self.layernorm_2(x), context) + residue_short
        residue_short = x
        x, gate = self.linear_geglu_1(self.layernorm_3(x)).chunk(2, dim=-1)
        x = self.linear_geglu_2(x * F.gelu(gate)) + residue_short
        x = x.transpose(-1, -2).view((n, c, h, w))
        return self.conv_output(x) + residue_long

class Upsample(nn.Module):
    def __init__(self, channels):
        super().__init__()
        self.conv = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
    
    def forward(self, x):
        return self.conv(F.interpolate(x, scale_factor=2, mode='nearest'))

class SwitchSequential(nn.Sequential):
    def forward(self, x, context, time):
        for layer in self:
            if isinstance(layer, UNET_AttentionBlock):
                x = layer(x, context)
            elif isinstance(layer, UNET_ResidualBlock):
                x = layer(x, time)
            else:
                x = layer(x)
        return x

class UNET(nn.Module):
    def __init__(self, use_se=False):
        super().__init__()
        self.encoders = nn.ModuleList([
            SwitchSequential(nn.Conv2d(4, 320, kernel_size=3, padding=1)),
            SwitchSequential(UNET_ResidualBlock(320, 320, use_se=use_se), UNET_AttentionBlock(8, 40)),
            SwitchSequential(UNET_ResidualBlock(320, 320, use_se=use_se), UNET_AttentionBlock(8, 40)),
            SwitchSequential(nn.Conv2d(320, 320, kernel_size=3, stride=2, padding=1)),
            SwitchSequential(UNET_ResidualBlock(320, 640, use_se=use_se), UNET_AttentionBlock(8, 80)),
            SwitchSequential(UNET_ResidualBlock(640, 640, use_se=use_se), UNET_AttentionBlock(8, 80)),
            SwitchSequential(nn.Conv2d(640, 640, kernel_size=3, stride=2, padding=1)),
            SwitchSequential(UNET_ResidualBlock(640, 1280, use_se=use_se), UNET_AttentionBlock(8, 160)),
            SwitchSequential(UNET_ResidualBlock(1280, 1280, use_se=use_se), UNET_AttentionBlock(8, 160)),
            SwitchSequential(nn.Conv2d(1280, 1280, kernel_size=3, stride=2, padding=1)),
            SwitchSequential(UNET_ResidualBlock(1280, 1280, use_se=use_se)),
            SwitchSequential(UNET_ResidualBlock(1280, 1280, use_se=use_se)),
        ])

        self.bottleneck = SwitchSequential(
            UNET_ResidualBlock(1280, 1280, use_se=use_se),
            UNET_AttentionBlock(8, 160),
            UNET_ResidualBlock(1280, 1280, use_se=use_se),
        )
        
        self.decoders = nn.ModuleList([
            SwitchSequential(UNET_ResidualBlock(2560, 1280, use_se=use_se)),
            SwitchSequential(UNET_ResidualBlock(2560, 1280, use_se=use_se)),
            SwitchSequential(UNET_ResidualBlock(2560, 1280, use_se=use_se), Upsample(1280)),
            SwitchSequential(UNET_ResidualBlock(2560, 1280, use_se=use_se), UNET_AttentionBlock(8, 160)),
            SwitchSequential(UNET_ResidualBlock(2560, 1280, use_se=use_se), UNET_AttentionBlock(8, 160)),
            SwitchSequential(UNET_ResidualBlock(1920, 1280, use_se=use_se), UNET_AttentionBlock(8, 160), Upsample(1280)),
            SwitchSequential(UNET_ResidualBlock(1920, 640, use_se=use_se), UNET_AttentionBlock(8, 80)),
            SwitchSequential(UNET_ResidualBlock(1280, 640, use_se=use_se), UNET_AttentionBlock(8, 80)),
            SwitchSequential(UNET_ResidualBlock(960, 640, use_se=use_se), UNET_AttentionBlock(8, 80), Upsample(640)),
            SwitchSequential(UNET_ResidualBlock(960, 320, use_se=use_se), UNET_AttentionBlock(8, 40)),
            SwitchSequential(UNET_ResidualBlock(640, 320, use_se=use_se), UNET_AttentionBlock(8, 40)),
            SwitchSequential(UNET_ResidualBlock(640, 320, use_se=use_se), UNET_AttentionBlock(8, 40)),
        ])

    def forward(self, x, context, time):
        skip_connections = []
        for layers in self.encoders:
            x = layers(x, context, time)
            skip_connections.append(x)

        x = self.bottleneck(x, context, time)

        for layers in self.decoders:
            x = torch.cat((x, skip_connections.pop()), dim=1)
            x = layers(x, context, time)
        
        return x

class UNET_OutputLayer(nn.Module):
    def __init__(self, in_channels, out_channels):
        super().__init__()
        self.groupnorm = nn.GroupNorm(32, in_channels)
        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
    
    def forward(self, x):
        x = F.silu(self.groupnorm(x))
        return self.conv(x)

class Diffusion(nn.Module):
    def __init__(self, use_se=False):
        super().__init__()
        self.time_embedding = TimeEmbedding(320)
        self.unet = UNET(use_se=use_se)
        self.final = UNET_OutputLayer(320, 4)
    
    def forward(self, latent, context, time):
        time = self.time_embedding(time)
        output = self.unet(latent, context, time)
        return self.final(output)