File size: 5,678 Bytes
a04730e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import torch
from torch import nn

from src.models.autoencoder.encoder import Encoder
from src.models.autoencoder.decoder import Decoder
from src.models.autoencoder.distributions import DiagonalGaussianDistribution


class AutoencoderKL(nn.Module):
    """
    VAE / AutoencoderKL wrapper
        posterior = vae.encode(x)
        z = posterior.sample()
        x_recon = vae.decode(z)

    Or directly:

        x_recon, posterior, z = vae(x)

    Input image range:
        x in [-1, 1]

    Output image range:
        x_recon in [-1, 1]
    """

    def __init__(
        self,
        in_channels: int = 3,
        out_channels: int = 3,
        latent_channels: int = 8,
        base_channels: int = 128,
        channel_multipliers: list[int] | tuple[int, ...] = (1, 2, 4, 4),
        num_res_blocks: int = 3,
        dropout: float = 0.0,
        use_attention: bool = True,
        attention_heads: int = 4,
        scaling_factor: float = 1.0,
        attention_resolutions: tuple[int, ...] = (32,),
    ):
        super().__init__()

        self.in_channels = in_channels
        self.out_channels = out_channels
        self.latent_channels = latent_channels
        self.base_channels = base_channels
        self.channel_multipliers = list(channel_multipliers)
        self.num_res_blocks = num_res_blocks
        self.scaling_factor = scaling_factor

        self.encoder = Encoder(
            in_channels=in_channels,
            latent_channels=latent_channels,
            base_channels=base_channels,
            channel_multipliers=channel_multipliers,
            num_res_blocks=num_res_blocks,
            dropout=dropout,
            use_attention=use_attention,
            attention_heads=attention_heads,
            attention_resolutions= attention_resolutions
        )

        self.decoder = Decoder(
            out_channels=out_channels,
            latent_channels=latent_channels,
            base_channels=base_channels,
            channel_multipliers=channel_multipliers,
            num_res_blocks=num_res_blocks,
            dropout=dropout,
            use_attention=use_attention,
            attention_heads=attention_heads,
        )

    def encode(
        self,
        x: torch.Tensor,
        deterministic: bool = False,
    ) -> DiagonalGaussianDistribution:
        """
        Encode image into posterior distribution.
            deterministic:
                If True, posterior.sample() will return mean only.
        Returns:
            DiagonalGaussianDistribution.
        """
        moments = self.encoder(x)
        posterior = DiagonalGaussianDistribution(
            moments=moments,
            deterministic=deterministic,
        )
        return posterior

    def decode(
        self,
        z: torch.Tensor,
        unscale: bool = True,
    ) -> torch.Tensor:
        """
        Decode latent into image
            z:
                Latent tensor [B, latent_channels, H/8, W/8].

            unscale:
                If True, divide by scaling_factor before decoding.

        Returns:
            Reconstructed image in [-1, 1].
        """
        if unscale:
            z = z / self.scaling_factor

        return self.decoder(z)

    def forward(
        self,
        x: torch.Tensor,
        sample_posterior: bool = True,
    ) -> tuple[torch.Tensor, DiagonalGaussianDistribution, torch.Tensor]:
        """
        Full VAE forward pass.

        Args:
            x:
                Image tensor [B, 3, H, W], normalized to [-1, 1].

            sample_posterior:
                If True:
                    z = posterior.sample()
                If False:
                    z = posterior.mode()

        Returns:
            x_recon:
                Reconstructed image [B, 3, H, W].

            posterior:
                DiagonalGaussianDistribution object.

            z:
                Latent tensor before scaling.
        """
        posterior = self.encode(x)

        if sample_posterior:
            z = posterior.sample()
        else:
            z = posterior.mode()

        x_recon = self.decode(z, unscale=False)

        return x_recon, posterior, z

    @torch.no_grad()
    def reconstruct(
        self,
        x: torch.Tensor,
        sample_posterior: bool = False,
    ) -> torch.Tensor:
        """
        Convenience method for inference reconstruction.

        By default, uses posterior mode for stable reconstructions.
        """
        x_recon, _, _ = self.forward(
            x,
            sample_posterior=sample_posterior,
        )
        return x_recon

    @torch.no_grad()
    def encode_to_latent(
        self,
        x: torch.Tensor,
        sample_posterior: bool = False,
        scale: bool = True,
    ) -> torch.Tensor:
        """
        Encode image into latent tensor for latent diffusion training.

        Usually for latent caching, use:

            sample_posterior=False

        because the posterior mean is deterministic and stable.

        If scale=True:

            z_scaled = z * scaling_factor

        Stable Diffusion-style LDMs often scale latents before diffusion.
        """
        posterior = self.encode(x)

        if sample_posterior:
            z = posterior.sample()
        else:
            z = posterior.mode()

        if scale:
            z = z * self.scaling_factor

        return z

    @torch.no_grad()
    def decode_from_latent(
        self,
        z: torch.Tensor,
        unscale: bool = True,
    ) -> torch.Tensor:
        """
        Decode latent tensor produced by diffusion model.
        """
        return self.decode(z, unscale=unscale)