File size: 4,437 Bytes
d6d864b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Text-conditional U-Net for diffusion."""
import torch
import torch.nn as nn
import math
import config


class TimeEmbedding(nn.Module):
    """Sinusoidal time embedding."""

    def __init__(self, dim):
        super().__init__()
        self.dim = dim

    def forward(self, t):
        half_dim = self.dim // 2
        emb = math.log(10000) / (half_dim - 1)
        emb = torch.exp(torch.arange(half_dim, device=t.device) * -emb)
        emb = t[:, None] * emb[None, :]
        return torch.cat([emb.sin(), emb.cos()], dim=1)


class ResBlock(nn.Module):
    """Residual block with time and text conditioning."""

    def __init__(self, in_ch, out_ch, time_dim, text_dim=None):
        super().__init__()
        self.time_mlp = nn.Linear(time_dim, out_ch)
        self.text_mlp = nn.Linear(text_dim, out_ch) if text_dim else None

        self.conv1 = nn.Conv2d(in_ch, out_ch, 3, padding=1)
        self.conv2 = nn.Conv2d(out_ch, out_ch, 3, padding=1)
        self.norm1 = nn.GroupNorm(min(8, in_ch), in_ch)
        self.norm2 = nn.GroupNorm(min(8, out_ch), out_ch)
        self.act = nn.SiLU()

        self.skip = nn.Conv2d(in_ch, out_ch, 1) if in_ch != out_ch else nn.Identity()

    def forward(self, x, t_emb, text_emb=None):
        h = self.act(self.norm1(x))
        h = self.conv1(h)

        # Add time embedding
        h = h + self.time_mlp(t_emb)[:, :, None, None]

        # Add text embedding
        if self.text_mlp is not None and text_emb is not None:
            h = h + self.text_mlp(text_emb)[:, :, None, None]

        h = self.act(self.norm2(h))
        h = self.conv2(h)

        return h + self.skip(x)


class TextConditionedUNet(nn.Module):
    """U-Net with CLIP text conditioning."""

    def __init__(self, text_dim=512):
        super().__init__()
        self.text_dim = text_dim

        self.time_emb = TimeEmbedding(config.TIME_DIM)
        self.time_mlp = nn.Sequential(
            nn.Linear(config.TIME_DIM, config.TIME_DIM),
            nn.SiLU(),
            nn.Linear(config.TIME_DIM, config.TIME_DIM)
        )

        self.text_proj = nn.Sequential(
            nn.Linear(text_dim, text_dim),
            nn.SiLU(),
            nn.Linear(text_dim, text_dim)
        )

        # Down path
        self.down1 = ResBlock(1, config.CHANNELS, config.TIME_DIM, text_dim)
        self.down2 = ResBlock(config.CHANNELS, config.CHANNELS * 2, config.TIME_DIM, text_dim)
        self.down3 = ResBlock(config.CHANNELS * 2, config.CHANNELS * 4, config.TIME_DIM, text_dim)

        # Middle
        self.mid = ResBlock(config.CHANNELS * 4, config.CHANNELS * 4, config.TIME_DIM, text_dim)

        # Up path
        self.up3 = ResBlock(config.CHANNELS * 8, config.CHANNELS * 2, config.TIME_DIM, text_dim)
        self.up2 = ResBlock(config.CHANNELS * 4, config.CHANNELS, config.TIME_DIM, text_dim)
        self.up1 = ResBlock(config.CHANNELS * 2, config.CHANNELS, config.TIME_DIM, text_dim)

        # Output
        self.out = nn.Conv2d(config.CHANNELS, 1, 1)

        # Pooling/Upsampling
        self.pool = nn.MaxPool2d(2)
        self.upsample = nn.Upsample(scale_factor=2, mode='nearest')

    def forward(self, x, t, text_emb):
        """
        Args:
            x: [B, 1, H, W] noisy images
            t: [B] timesteps
            text_emb: [B, text_dim] CLIP text embeddings
        """
        # Embeddings
        t_emb = self.time_mlp(self.time_emb(t))
        text_emb = self.text_proj(text_emb)

        # Down
        h1 = self.down1(x, t_emb, text_emb)
        h2 = self.down2(self.pool(h1), t_emb, text_emb)
        h3 = self.down3(self.pool(h2), t_emb, text_emb)

        # Middle
        h = self.mid(self.pool(h3), t_emb, text_emb)

        # Up
        h = self.up3(torch.cat([self.upsample(h), h3], dim=1), t_emb, text_emb)
        h = self.up2(torch.cat([self.upsample(h), h2], dim=1), t_emb, text_emb)
        h = self.up1(torch.cat([self.upsample(h), h1], dim=1), t_emb, text_emb)

        return self.out(h)


if __name__ == "__main__":
    # Test model
    print("Testing Text-Conditioned U-Net...")
    model = TextConditionedUNet(text_dim=512)

    # Test forward pass
    batch_size = 2
    x = torch.randn(batch_size, 1, 64, 64)
    t = torch.randint(0, 1000, (batch_size,))
    text_emb = torch.randn(batch_size, 512)

    out = model(x, t, text_emb)
    print(f"Input shape: {x.shape}")
    print(f"Output shape: {out.shape}")
    print(f"✅ Model test passed!")