File size: 8,901 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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
from __future__ import annotations

import torch
from torch import nn
import torch.nn.functional as F

from src.models.diffusion.timestep import TimestepEmbedding
from src.models.diffusion.blocks import (
    DownBlock,
    MiddleBlock,
    UpBlock,
    normalization,
)


class LatentDiffusionUNet(nn.Module):
    """
    Lightweight text-conditioned U-Net for latent diffusion.

    Input:
        z_t:
            noisy latent [B, in_channels, H, W]

        timesteps:
            diffusion timestep [B]

        context:
            CLIP token embeddings [B, seq_len, context_dim]

    Output:
        prediction [B, out_channels, H, W]
    """

    def __init__(
        self,
        in_channels: int = 8,
        out_channels: int = 8,
        latent_size: int = 32,
        base_channels: int = 128,
        channel_multipliers: list[int] | tuple[int, ...] = (1, 2, 3),
        num_res_blocks: int = 2,
        attention_resolutions: list[int] | tuple[int, ...] = (16, 8),
        context_dim: int = 768,
        num_heads: int = 4,
        head_dim: int = 32,
        transformer_depth: int = 1,
        dropout: float = 0.0,
        time_embedding_dim: int | None = None,
        use_middle_attention: bool = True,
    ):
        super().__init__()

        self.in_channels = in_channels
        self.out_channels = out_channels
        self.latent_size = latent_size
        self.base_channels = base_channels
        self.channel_multipliers = list(channel_multipliers)
        self.num_res_blocks = num_res_blocks
        self.attention_resolutions = set(attention_resolutions)
        self.context_dim = context_dim
        self.num_heads = num_heads
        self.head_dim = head_dim
        self.transformer_depth = transformer_depth
        self.dropout = dropout

        if time_embedding_dim is None:
            time_embedding_dim = base_channels * 4

        self.time_embedding_dim = time_embedding_dim

        self.time_embed = TimestepEmbedding(
            embedding_dim=base_channels,
            time_embed_dim=time_embedding_dim,
        )

        self.conv_in = nn.Conv2d(
            in_channels,
            base_channels,
            kernel_size=3,
            stride=1,
            padding=1,
        )

        channels_per_level = [
            base_channels * multiplier
            for multiplier in self.channel_multipliers
        ]

        # -------------------------
        # Down path
        # -------------------------
        self.down_blocks = nn.ModuleList()

        current_channels = base_channels
        current_resolution = latent_size

        self.down_block_channels: list[int] = []
        self.down_block_resolutions: list[int] = []

        for level, out_channels_level in enumerate(channels_per_level):
            is_last = level == len(channels_per_level) - 1

            use_attention = current_resolution in self.attention_resolutions

            block = DownBlock(
                in_channels=current_channels,
                out_channels=out_channels_level,
                time_embed_dim=time_embedding_dim,
                num_res_blocks=num_res_blocks,
                context_dim=context_dim,
                num_heads=num_heads,
                head_dim=head_dim,
                transformer_depth=transformer_depth,
                dropout=dropout,
                use_attention=use_attention,
                add_downsample=not is_last,
            )

            self.down_blocks.append(block)

            self.down_block_channels.append(out_channels_level)
            self.down_block_resolutions.append(current_resolution)

            current_channels = out_channels_level

            if not is_last:
                current_resolution //= 2

        # -------------------------
        # Middle
        # -------------------------
        self.middle = MiddleBlock(
            channels=current_channels,
            time_embed_dim=time_embedding_dim,
            context_dim=context_dim,
            num_heads=num_heads,
            head_dim=head_dim,
            transformer_depth=transformer_depth,
            dropout=dropout,
            use_attention=use_middle_attention,
        )

        # -------------------------
        # Up path
        # -------------------------
        self.up_blocks = nn.ModuleList()

        reversed_channels = list(reversed(self.down_block_channels))
        reversed_resolutions = list(reversed(self.down_block_resolutions))

        for level, (skip_channels, resolution) in enumerate(
            zip(reversed_channels, reversed_resolutions)
        ):
            is_last = level == len(reversed_channels) - 1

            out_channels_level = skip_channels
            use_attention = resolution in self.attention_resolutions

            block = UpBlock(
                in_channels=current_channels,
                skip_channels=skip_channels,
                out_channels=out_channels_level,
                time_embed_dim=time_embedding_dim,
                num_res_blocks=num_res_blocks,
                context_dim=context_dim,
                num_heads=num_heads,
                head_dim=head_dim,
                transformer_depth=transformer_depth,
                dropout=dropout,
                use_attention=use_attention,
                add_upsample=not is_last,
            )

            self.up_blocks.append(block)

            current_channels = out_channels_level

        self.norm_out = normalization(current_channels)

        self.conv_out = nn.Conv2d(
            current_channels,
            out_channels,
            kernel_size=3,
            stride=1,
            padding=1,
        )

        # Stable start: initially predicts near zero.
        nn.init.zeros_(self.conv_out.weight)
        nn.init.zeros_(self.conv_out.bias)

    def forward(
        self,
        z_t: torch.Tensor,
        timesteps: torch.Tensor,
        context: torch.Tensor,
        attention_mask: torch.Tensor | None = None,
    ) -> torch.Tensor:
        """
        Args:
            z_t:
                Noisy latent [B, C, H, W]

            timesteps:
                Diffusion timesteps [B]

            context:
                CLIP token embeddings [B, seq_len, context_dim]

            attention_mask:
                CLIP token mask [B, seq_len], optional

        Returns:
            model_output:
                [B, out_channels, H, W]
        """
        if z_t.ndim != 4:
            raise ValueError(f"z_t must be [B, C, H, W], got {z_t.shape}")

        if timesteps.ndim != 1:
            raise ValueError(f"timesteps must be [B], got {timesteps.shape}")

        time_emb = self.time_embed(timesteps)

        x = self.conv_in(z_t)

        skips: list[torch.Tensor] = []

        for block in self.down_blocks:
            x, block_skips = block(
                x=x,
                time_emb=time_emb,
                context=context,
                attention_mask=attention_mask,
            )

            skips.extend(block_skips)

        x = self.middle(
            x=x,
            time_emb=time_emb,
            context=context,
            attention_mask=attention_mask,
        )

        for block in self.up_blocks:
            x = block(
                x=x,
                skips=skips,
                time_emb=time_emb,
                context=context,
                attention_mask=attention_mask,
            )

        if len(skips) != 0:
            raise RuntimeError(
                f"Unused skip connections remain: {len(skips)}. "
                "Check U-Net block construction."
            )

        x = self.norm_out(x)
        x = F.silu(x)
        x = self.conv_out(x)

        return x


def count_parameters(
    model: nn.Module,
    trainable_only: bool = True,
) -> int:
    if trainable_only:
        return sum(p.numel() for p in model.parameters() if p.requires_grad)

    return sum(p.numel() for p in model.parameters())


def build_latent_diffusion_unet_from_config(cfg: dict) -> LatentDiffusionUNet:
    """
    Build U-Net from config dictionary.

    Expects:

        cfg["model"]
    """
    m = cfg["model"]

    return LatentDiffusionUNet(
        in_channels=int(m["in_channels"]),
        out_channels=int(m["out_channels"]),
        latent_size=int(m.get("latent_size", 32)),
        base_channels=int(m["base_channels"]),
        channel_multipliers=tuple(m["channel_multipliers"]),
        num_res_blocks=int(m["num_res_blocks"]),
        attention_resolutions=tuple(m.get("attention_resolutions", [16, 8])),
        context_dim=int(m["context_dim"]),
        num_heads=int(m["num_heads"]),
        head_dim=int(m["head_dim"]),
        transformer_depth=int(m.get("transformer_depth", 1)),
        dropout=float(m.get("dropout", 0.0)),
        time_embedding_dim=m.get("time_embedding_dim", None),
        use_middle_attention=bool(m.get("use_middle_attention", True)),
    )