File size: 6,543 Bytes
0a0f923
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# xray_generator/models/vae.py
import torch
import torch.nn as nn
import torch.nn.functional as F
from .unet import SelfAttention

class VAEEncoder(nn.Module):
    """Encoder for VAE with attention mechanisms."""
    def __init__(

        self,

        in_channels=1,

        latent_channels=4,

        hidden_dims=[64, 128, 256, 512],

        attention_resolutions=[32, 16]

    ):
        """Initialize VAE encoder."""
        super().__init__()
        
        # Input convolution
        self.conv_in = nn.Conv2d(in_channels, hidden_dims[0], 3, padding=1)
        
        # Downsampling blocks
        self.down_blocks = nn.ModuleList()
        
        # Create downsampling blocks
        for i in range(len(hidden_dims) - 1):
            in_dim = hidden_dims[i]
            out_dim = hidden_dims[i + 1]
            
            # Determine resolution
            resolution = 256 // (2 ** i)
            use_attention = resolution in attention_resolutions
            
            block = []
            
            # Add attention if needed
            if use_attention:
                block.append(SelfAttention(in_dim))
            
            # Convolution with GroupNorm and activation
            block.append(nn.Sequential(
                nn.GroupNorm(8, in_dim),
                nn.SiLU(),
                nn.Conv2d(in_dim, out_dim, 3, stride=2, padding=1)
            ))
            
            self.down_blocks.append(nn.Sequential(*block))
        
        # Final layers
        self.final = nn.Sequential(
            nn.GroupNorm(8, hidden_dims[-1]),
            nn.SiLU(),
            nn.Conv2d(hidden_dims[-1], latent_channels * 2, 3, padding=1)
        )
        
        # Initialize weights
        self.apply(self._init_weights)
    
    def _init_weights(self, m):
        """Initialize weights with Kaiming normal."""
        if isinstance(m, (nn.Conv2d, nn.Linear)):
            nn.init.kaiming_normal_(m.weight, nonlinearity='relu')
            if m.bias is not None:
                nn.init.zeros_(m.bias)
    
    def forward(self, x):
        """Forward pass through encoder."""
        # Initial convolution
        x = self.conv_in(x)
        
        # Downsampling
        for block in self.down_blocks:
            x = block(x)
        
        # Final layers
        x = self.final(x)
        
        # Split into mu and logvar
        mu, logvar = torch.chunk(x, 2, dim=1)
        
        return mu, logvar

class VAEDecoder(nn.Module):
    """Decoder for VAE with attention mechanisms."""
    def __init__(

        self,

        latent_channels=4,

        out_channels=1,

        hidden_dims=[512, 256, 128, 64],

        attention_resolutions=[16, 32]

    ):
        """Initialize VAE decoder."""
        super().__init__()
        
        # Input convolution
        self.conv_in = nn.Conv2d(latent_channels, hidden_dims[0], 3, padding=1)
        
        # Upsampling blocks
        self.up_blocks = nn.ModuleList()
        
        # Create upsampling blocks
        for i in range(len(hidden_dims) - 1):
            in_dim = hidden_dims[i]
            out_dim = hidden_dims[i + 1]
            
            # Determine resolution
            resolution = 16 * (2 ** i)  # Starting at 16x16 for latent space
            use_attention = resolution in attention_resolutions
            
            block = []
            
            # Add attention if needed
            if use_attention:
                block.append(SelfAttention(in_dim))
            
            # Add upsampling
            block.append(nn.Sequential(
                nn.GroupNorm(8, in_dim),
                nn.SiLU(),
                nn.ConvTranspose2d(in_dim, out_dim, 4, stride=2, padding=1)
            ))
            
            self.up_blocks.append(nn.Sequential(*block))
        
        # Final layers
        self.final = nn.Sequential(
            nn.GroupNorm(8, hidden_dims[-1]),
            nn.SiLU(),
            nn.Conv2d(hidden_dims[-1], out_channels, 3, padding=1)
        )
        
        # Initialize weights
        self.apply(self._init_weights)
    
    def _init_weights(self, m):
        """Initialize weights with Kaiming normal."""
        if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d, nn.Linear)):
            nn.init.kaiming_normal_(m.weight, nonlinearity='relu')
            if m.bias is not None:
                nn.init.zeros_(m.bias)
    
    def forward(self, x):
        """Forward pass through decoder."""
        # Initial convolution
        x = self.conv_in(x)
        
        # Upsampling
        for block in self.up_blocks:
            x = block(x)
        
        # Final layers
        x = self.final(x)
        
        return x

class MedicalVAE(nn.Module):
    """Complete VAE model for medical images."""
    def __init__(

        self,

        in_channels=1,

        out_channels=1,

        latent_channels=4,

        hidden_dims=[64, 128, 256, 512],

        attention_resolutions=[16, 32]

    ):
        """Initialize VAE."""
        super().__init__()
        
        # Create encoder and decoder
        self.encoder = VAEEncoder(
            in_channels=in_channels,
            latent_channels=latent_channels,
            hidden_dims=hidden_dims,
            attention_resolutions=attention_resolutions
        )
        
        self.decoder = VAEDecoder(
            latent_channels=latent_channels,
            out_channels=out_channels,
            hidden_dims=list(reversed(hidden_dims)),
            attention_resolutions=attention_resolutions
        )
        
        # Save parameters
        self.latent_channels = latent_channels
    
    def encode(self, x):
        """Encode input to latent space."""
        return self.encoder(x)
    
    def decode(self, z):
        """Decode from latent space."""
        return self.decoder(z)
    
    def reparameterize(self, mu, logvar):
        """Reparameterization trick."""
        std = torch.exp(0.5 * logvar)
        eps = torch.randn_like(std)
        return mu + eps * std
    
    def forward(self, x):
        """Forward pass through the VAE."""
        # Encode
        mu, logvar = self.encode(x)
        
        # Reparameterize
        z = self.reparameterize(mu, logvar)
        
        # Decode
        recon = self.decode(z)
        
        return recon, mu, logvar