File size: 35,024 Bytes
275b5a1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 | """SWaG backbone adapted from the DiT architecture for 1-D waveforms.
The multi-scale waveform frontend and decoder follow the SWaG research code.
The transformer conditioning path contains diffusion timesteps only.
"""
from __future__ import annotations
import math
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from timm.models.vision_transformer import Attention, Mlp
def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
def sinusoidal_position_embedding(length: int, dimension: int) -> torch.Tensor:
if dimension % 2:
raise ValueError("Positional embedding dimension must be even")
positions = np.arange(length, dtype=np.float32)[:, None]
frequencies = np.exp(-math.log(10_000) * np.arange(dimension // 2) / (dimension // 2))
embedding = np.concatenate([np.sin(positions * frequencies), np.cos(positions * frequencies)], axis=1)
return torch.from_numpy(embedding.astype(np.float32)).unsqueeze(0)
class TimestepEmbedder(nn.Module):
def __init__(self, hidden_size: int, frequency_size: int = 256):
super().__init__()
self.frequency_size = frequency_size
self.mlp = nn.Sequential(
nn.Linear(frequency_size, hidden_size),
nn.SiLU(),
nn.Linear(hidden_size, hidden_size),
)
@staticmethod
def frequency_embedding(t: torch.Tensor, dimension: int, max_period: int = 10_000) -> torch.Tensor:
half = dimension // 2
frequencies = torch.exp(
-math.log(max_period) * torch.arange(half, dtype=torch.float32, device=t.device) / half
)
angles = t[:, None].float() * frequencies[None]
embedding = torch.cat([torch.cos(angles), torch.sin(angles)], dim=-1)
if dimension % 2:
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
return embedding
def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
return self.mlp(self.frequency_embedding(timesteps, self.frequency_size))
class ContinuousConditionEmbedder(nn.Module):
"""Embed one continuous scalar and provide a learned null value for CFG."""
def __init__(self, hidden_size: int, frequency_size: int = 256):
super().__init__()
self.frequency_size = frequency_size
self.mlp = nn.Sequential(
nn.Linear(frequency_size, hidden_size),
nn.SiLU(),
nn.Linear(hidden_size, hidden_size),
)
self.null_embedding = nn.Parameter(torch.zeros(hidden_size))
def forward(self, values: torch.Tensor, drop_mask: torch.Tensor) -> torch.Tensor:
embedded = self.mlp(TimestepEmbedder.frequency_embedding(values, self.frequency_size))
return torch.where(drop_mask[:, None], self.null_embedding[None].to(embedded.dtype), embedded)
class ConvBlock(nn.Module):
def __init__(self, in_channels: int, out_channels: int, kernel_size: int):
super().__init__()
self.block = nn.Sequential(
nn.Conv1d(in_channels, out_channels, kernel_size, padding=kernel_size // 2),
nn.GroupNorm(1, out_channels),
nn.GELU(),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.block(x)
class TokenBranch(nn.Module):
def __init__(self, in_channels: int, branch_channels: int, hidden_size: int, tokens: int, kernel: int):
super().__init__()
self.tokens = tokens
self.encoder = nn.Sequential(
ConvBlock(in_channels, branch_channels, kernel),
ConvBlock(branch_channels, branch_channels, kernel),
)
self.pool = nn.AdaptiveAvgPool1d(tokens)
self.projection = nn.Conv1d(branch_channels, hidden_size, 1)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.projection(self.pool(self.encoder(x))).transpose(1, 2)
class MultiScaleEncoder(nn.Module):
def __init__(
self,
in_channels: int,
hidden_size: int,
stem_channels: int,
small_channels: int,
mid_channels: int,
large_channels: int,
small_tokens: int,
mid_tokens: int,
large_tokens: int,
):
super().__init__()
self.stem = nn.Sequential(
ConvBlock(in_channels, stem_channels, 9),
ConvBlock(stem_channels, stem_channels, 7),
)
self.small = TokenBranch(stem_channels, small_channels, hidden_size, small_tokens, 7)
self.mid = TokenBranch(stem_channels, mid_channels, hidden_size, mid_tokens, 9)
self.large = TokenBranch(stem_channels, large_channels, hidden_size, large_tokens, 15)
def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
features = self.stem(x)
return self.small(features), self.mid(features), self.large(features)
class TokenPreprocessor(nn.Module):
def __init__(self, hidden_size: int):
super().__init__()
self.norms = nn.ModuleList([nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6) for _ in range(3)])
self.modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size))
def forward(self, small: torch.Tensor, mid: torch.Tensor, large: torch.Tensor, t: torch.Tensor):
shifts = self.modulation(t).chunk(6, dim=1)
return (
modulate(self.norms[0](small), shifts[0], shifts[1]),
modulate(self.norms[1](mid), shifts[2], shifts[3]),
modulate(self.norms[2](large), shifts[4], shifts[5]),
)
class MultiScaleDiTBlock(nn.Module):
def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float):
super().__init__()
self.norm_self = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.self_attention = Attention(hidden_size, num_heads=num_heads, qkv_bias=True)
self.norm_mid_query = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.norm_mid_context = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.mid_attention = nn.MultiheadAttention(hidden_size, num_heads, batch_first=True)
self.norm_large_query = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.norm_large_context = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.large_attention = nn.MultiheadAttention(hidden_size, num_heads, batch_first=True)
self.norm_mlp = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
self.mlp = Mlp(
in_features=hidden_size,
hidden_features=int(hidden_size * mlp_ratio),
act_layer=lambda: nn.GELU(approximate="tanh"),
drop=0,
)
self.modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 16 * hidden_size))
def forward(self, small: torch.Tensor, mid: torch.Tensor, large: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
values = self.modulation(t).chunk(16, dim=1)
small = small + values[2].unsqueeze(1) * self.self_attention(
modulate(self.norm_self(small), values[0], values[1])
)
query = modulate(self.norm_mid_query(small), values[3], values[4])
context = modulate(self.norm_mid_context(mid), values[5], values[6])
update, _ = self.mid_attention(query, context, context, need_weights=False)
small = small + values[7].unsqueeze(1) * update
query = modulate(self.norm_large_query(small), values[8], values[9])
context = modulate(self.norm_large_context(large), values[10], values[11])
update, _ = self.large_attention(query, context, context, need_weights=False)
small = small + values[12].unsqueeze(1) * update
return small + values[15].unsqueeze(1) * self.mlp(
modulate(self.norm_mlp(small), values[13], values[14])
)
class LoRALinearDelta(nn.Module):
def __init__(self, features: int, rank: int, alpha: float, dropout: float):
super().__init__()
self.scale = float(alpha) / int(rank)
self.dropout = nn.Dropout(float(dropout)) if dropout > 0 else nn.Identity()
self.down = nn.Linear(features, rank, bias=False)
self.up = nn.Linear(rank, features, bias=False)
nn.init.kaiming_uniform_(self.down.weight, a=math.sqrt(5))
nn.init.zeros_(self.up.weight)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.up(self.down(self.dropout(x))) * self.scale
class LoRAMultiheadAttention(nn.MultiheadAttention):
"""MultiheadAttention with additive LoRA while retaining base state keys."""
def __init__(self, embed_dim: int, num_heads: int, rank: int, alpha: float, dropout: float):
super().__init__(embed_dim, num_heads, batch_first=True)
self.lora_q = LoRALinearDelta(embed_dim, rank, alpha, dropout)
self.lora_k = LoRALinearDelta(embed_dim, rank, alpha, dropout)
self.lora_v = LoRALinearDelta(embed_dim, rank, alpha, dropout)
self.lora_out = LoRALinearDelta(embed_dim, rank, alpha, dropout)
def forward(self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, **kwargs):
dimension = self.embed_dim
q_weight, k_weight, v_weight = self.in_proj_weight.chunk(3, dim=0)
q_bias, k_bias, v_bias = self.in_proj_bias.chunk(3, dim=0)
q = F.linear(query, q_weight, q_bias) + self.lora_q(query)
k = F.linear(key, k_weight, k_bias) + self.lora_k(key)
v = F.linear(value, v_weight, v_bias) + self.lora_v(value)
batch, query_length, _ = q.shape
key_length = k.shape[1]
head_dimension = dimension // self.num_heads
q = q.view(batch, query_length, self.num_heads, head_dimension).transpose(1, 2)
k = k.view(batch, key_length, self.num_heads, head_dimension).transpose(1, 2)
v = v.view(batch, key_length, self.num_heads, head_dimension).transpose(1, 2)
attended = F.scaled_dot_product_attention(q, k, v, dropout_p=0.0)
attended = attended.transpose(1, 2).contiguous().view(batch, query_length, dimension)
return self.out_proj(attended) + self.lora_out(attended), None
class SWaG(nn.Module):
"""SWaG with the fixed multi-scale cross-attention waveform backbone."""
def __init__(
self,
in_channels: int = 3,
length: int = 6000,
hidden_size: int = 768,
depth: int = 12,
num_heads: int = 12,
mlp_ratio: float = 4.0,
frequency_embedding_size: int = 256,
small_token_count: int = 1024,
mid_token_count: int = 256,
large_token_count: int = 32,
stem_feature_channels: int = 64,
small_encoder_channels: int = 128,
mid_encoder_channels: int = 128,
large_encoder_channels: int = 128,
decoder_channels: int = 128,
skip_channels: int = 64,
fusion_channels: int = 128,
use_high_res_skip: bool = True,
learn_sigma: bool = True,
):
super().__init__()
if hidden_size % num_heads:
raise ValueError("hidden_size must be divisible by num_heads")
self.in_channels = in_channels
self.length = length
self.learn_sigma = learn_sigma
self.out_channels = in_channels * 2 if learn_sigma else in_channels
self.use_high_res_skip = use_high_res_skip
self.encoder = MultiScaleEncoder(
in_channels, hidden_size, stem_feature_channels,
small_encoder_channels, mid_encoder_channels, large_encoder_channels,
small_token_count, mid_token_count, large_token_count,
)
self.timestep_embedder = TimestepEmbedder(hidden_size, frequency_embedding_size)
self.register_buffer("small_position", sinusoidal_position_embedding(small_token_count, hidden_size), persistent=True)
self.register_buffer("mid_position", sinusoidal_position_embedding(mid_token_count, hidden_size), persistent=True)
self.register_buffer("large_position", sinusoidal_position_embedding(large_token_count, hidden_size), persistent=True)
self.token_preprocessor = TokenPreprocessor(hidden_size)
self.blocks = nn.ModuleList([MultiScaleDiTBlock(hidden_size, num_heads, mlp_ratio) for _ in range(depth)])
self.decoder_projection = nn.Conv1d(hidden_size, decoder_channels, 1)
self.decoder_refine = ConvBlock(decoder_channels, decoder_channels, 5)
self.skip_branch = (
nn.Sequential(ConvBlock(in_channels, skip_channels, 7), ConvBlock(skip_channels, skip_channels, 5))
if use_high_res_skip else None
)
fusion_input = decoder_channels + (skip_channels if use_high_res_skip else 0)
self.fusion = ConvBlock(fusion_input, fusion_channels, 5)
self.output_head = nn.Conv1d(fusion_channels, self.out_channels, 1)
self.initialize_weights()
def initialize_weights(self) -> None:
def initialize(module: nn.Module) -> None:
if isinstance(module, nn.Linear):
nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
self.apply(initialize)
nn.init.normal_(self.timestep_embedder.mlp[0].weight, std=0.02)
nn.init.normal_(self.timestep_embedder.mlp[2].weight, std=0.02)
nn.init.zeros_(self.token_preprocessor.modulation[-1].weight)
nn.init.zeros_(self.token_preprocessor.modulation[-1].bias)
for block in self.blocks:
nn.init.zeros_(block.modulation[-1].weight)
nn.init.zeros_(block.modulation[-1].bias)
nn.init.zeros_(self.output_head.weight)
nn.init.zeros_(self.output_head.bias)
def forward(self, x: torch.Tensor, timesteps: torch.Tensor) -> torch.Tensor:
if x.ndim != 3 or x.shape[1:] != (self.in_channels, self.length):
raise ValueError(f"Expected x shape [N, {self.in_channels}, {self.length}], got {tuple(x.shape)}")
if timesteps.ndim != 1 or timesteps.shape[0] != x.shape[0]:
raise ValueError(f"Expected timesteps shape [{x.shape[0]}], got {tuple(timesteps.shape)}")
timestep_embedding = self.timestep_embedder(timesteps)
small, mid, large = self.encoder(x)
small = small + self.small_position.to(dtype=small.dtype)
mid = mid + self.mid_position.to(dtype=mid.dtype)
large = large + self.large_position.to(dtype=large.dtype)
small, mid, large = self.token_preprocessor(small, mid, large, timestep_embedding)
for block in self.blocks:
small = block(small, mid, large, timestep_embedding)
decoded = self.decoder_projection(small.transpose(1, 2))
decoded = F.interpolate(decoded, size=self.length, mode="linear", align_corners=False)
decoded = self.decoder_refine(decoded)
if self.skip_branch is not None:
decoded = torch.cat([decoded, self.skip_branch(x)], dim=1)
return self.output_head(self.fusion(decoded))
class EmptyConditionSWaG(SWaG):
"""No-skip SWaG with eight reserved continuous condition slots.
The slots are passed as normalized scalar values. Their projection is
zero-initialized, so an all-zero condition vector leaves the base model
behavior unchanged while preserving a migration interface.
"""
def __init__(
self,
condition_slot_count: int = 8,
condition_normalization_length: float = 6000.0,
hidden_size: int = 768,
frequency_embedding_size: int = 256,
**kwargs,
):
super().__init__(
hidden_size=hidden_size,
frequency_embedding_size=frequency_embedding_size,
**kwargs,
)
if int(condition_slot_count) != 8:
raise ValueError("EmptyConditionSWaG requires exactly 8 condition slots")
self.condition_slot_count = 8
self.condition_normalization_length = float(condition_normalization_length)
hidden = int(hidden_size)
frequency_size = int(frequency_embedding_size)
self.condition_embedders = nn.ModuleList(
[ContinuousConditionEmbedder(hidden, frequency_size) for _ in range(self.condition_slot_count)]
)
self.condition_projection = nn.Sequential(nn.SiLU(), nn.Linear(hidden, hidden))
nn.init.zeros_(self.condition_projection[-1].weight)
nn.init.zeros_(self.condition_projection[-1].bias)
def forward(
self,
x: torch.Tensor,
timesteps: torch.Tensor,
conditions: torch.Tensor | None = None,
) -> torch.Tensor:
if conditions is None:
conditions = torch.zeros(
x.shape[0], self.condition_slot_count, device=x.device, dtype=torch.float32
)
if conditions.ndim != 2 or conditions.shape != (x.shape[0], self.condition_slot_count):
raise ValueError(
f"Expected conditions shape [{x.shape[0]}, {self.condition_slot_count}], "
f"got {tuple(conditions.shape)}"
)
if not torch.isfinite(conditions).all():
raise ValueError("Conditions contain non-finite values")
normalized = conditions.float() / self.condition_normalization_length
drop_mask = torch.zeros(x.shape[0], dtype=torch.bool, device=x.device)
encoded = sum(
embedder(normalized[:, index], drop_mask)
for index, embedder in enumerate(self.condition_embedders)
) / math.sqrt(self.condition_slot_count)
timestep_embedding = self.timestep_embedder(timesteps) + self.condition_projection(encoded)
small, mid, large = self.encoder(x)
small = small + self.small_position.to(dtype=small.dtype)
mid = mid + self.mid_position.to(dtype=mid.dtype)
large = large + self.large_position.to(dtype=large.dtype)
small, mid, large = self.token_preprocessor(small, mid, large, timestep_embedding)
for block in self.blocks:
small = block(small, mid, large, timestep_embedding)
decoded = self.decoder_projection(small.transpose(1, 2))
decoded = F.interpolate(decoded, size=self.length, mode="linear", align_corners=False)
decoded = self.decoder_refine(decoded)
return self.output_head(self.fusion(decoded))
class ConditionalSWaG(SWaG):
"""SWaG conditioned jointly on P- and S-arrival sample indices."""
def __init__(self, condition_dropout_prob: float = 0.1, **kwargs):
super().__init__(**kwargs)
if not 0.0 <= condition_dropout_prob <= 1.0:
raise ValueError("condition_dropout_prob must be in [0, 1]")
hidden_size = self.timestep_embedder.mlp[-1].out_features
frequency_size = self.timestep_embedder.frequency_size
self.condition_dropout_prob = float(condition_dropout_prob)
self.p_embedder = ContinuousConditionEmbedder(hidden_size, frequency_size)
self.s_embedder = ContinuousConditionEmbedder(hidden_size, frequency_size)
for embedder in (self.p_embedder, self.s_embedder):
nn.init.normal_(embedder.mlp[0].weight, std=0.02)
nn.init.normal_(embedder.mlp[2].weight, std=0.02)
def _drop_mask(self, labels: torch.Tensor, force_drop_mask: torch.Tensor | None) -> torch.Tensor:
if force_drop_mask is not None:
mask = force_drop_mask.to(device=labels.device, dtype=torch.bool)
if mask.shape != (labels.shape[0],):
raise ValueError(f"Expected force_drop_mask shape [{labels.shape[0]}], got {tuple(mask.shape)}")
return mask
if self.training and self.condition_dropout_prob > 0:
return torch.rand(labels.shape[0], device=labels.device) < self.condition_dropout_prob
return torch.zeros(labels.shape[0], dtype=torch.bool, device=labels.device)
def forward(
self,
x: torch.Tensor,
timesteps: torch.Tensor,
labels: torch.Tensor,
force_drop_mask: torch.Tensor | None = None,
) -> torch.Tensor:
if x.ndim != 3 or x.shape[1:] != (self.in_channels, self.length):
raise ValueError(f"Expected x shape [N, {self.in_channels}, {self.length}], got {tuple(x.shape)}")
if timesteps.ndim != 1 or timesteps.shape[0] != x.shape[0]:
raise ValueError(f"Expected timesteps shape [{x.shape[0]}], got {tuple(timesteps.shape)}")
if labels.shape != (x.shape[0], 2):
raise ValueError(f"Expected P/S labels shape [{x.shape[0]}, 2], got {tuple(labels.shape)}")
if not torch.isfinite(labels).all():
raise ValueError("P/S labels contain non-finite values")
drop_mask = self._drop_mask(labels, force_drop_mask)
condition = (self.p_embedder(labels[:, 0], drop_mask) + self.s_embedder(labels[:, 1], drop_mask)) / math.sqrt(2.0)
conditioning = self.timestep_embedder(timesteps) + condition
small, mid, large = self.encoder(x)
small = small + self.small_position.to(dtype=small.dtype)
mid = mid + self.mid_position.to(dtype=mid.dtype)
large = large + self.large_position.to(dtype=large.dtype)
small, mid, large = self.token_preprocessor(small, mid, large, conditioning)
for block in self.blocks:
small = block(small, mid, large, conditioning)
decoded = self.decoder_projection(small.transpose(1, 2))
decoded = F.interpolate(decoded, size=self.length, mode="linear", align_corners=False)
decoded = self.decoder_refine(decoded)
if self.skip_branch is not None:
decoded = torch.cat([decoded, self.skip_branch(x)], dim=1)
return self.output_head(self.fusion(decoded))
def forward_with_cfg(
self, x: torch.Tensor, timesteps: torch.Tensor, labels: torch.Tensor, cfg_scale: float
) -> torch.Tensor:
"""Run conditional and null passes and apply CFG to epsilon channels only."""
keep = torch.zeros(x.shape[0], dtype=torch.bool, device=x.device)
drop = torch.ones(x.shape[0], dtype=torch.bool, device=x.device)
conditional = self(x, timesteps, labels, force_drop_mask=keep)
unconditional = self(x, timesteps, labels, force_drop_mask=drop)
eps_c, rest_c = conditional[:, : self.in_channels], conditional[:, self.in_channels :]
eps_u = unconditional[:, : self.in_channels]
guided_eps = eps_u + float(cfg_scale) * (eps_c - eps_u)
return torch.cat([guided_eps, rest_c], dim=1)
class ConditionalLoRASWaG(ConditionalSWaG):
"""P/S-conditioned SWaG with normalized arrivals and cross-attention LoRA."""
def __init__(
self,
condition_normalization_length: float = 6000.0,
cross_attention_lora_rank: int = 8,
cross_attention_lora_alpha: float = 16.0,
cross_attention_lora_dropout: float = 0.05,
**kwargs,
):
super().__init__(**kwargs)
self.condition_normalization_length = float(condition_normalization_length)
for block in self.blocks:
for name in ("mid_attention", "large_attention"):
base = getattr(block, name)
adapted = LoRAMultiheadAttention(
base.embed_dim, base.num_heads, cross_attention_lora_rank,
cross_attention_lora_alpha, cross_attention_lora_dropout,
)
adapted.in_proj_weight.data.copy_(base.in_proj_weight.data)
adapted.in_proj_bias.data.copy_(base.in_proj_bias.data)
adapted.out_proj.load_state_dict(base.out_proj.state_dict())
setattr(block, name, adapted)
def forward(self, x, timesteps, labels, force_drop_mask=None):
normalized = labels / self.condition_normalization_length
return super().forward(x, timesteps, normalized, force_drop_mask=force_drop_mask)
class DualConditionLoRASWaG(ConditionalLoRASWaG):
"""LoRA SWaG with separate timestep and P/S adaLN modulation chains."""
def __init__(self, **kwargs):
super().__init__(**kwargs)
hidden = self.timestep_embedder.mlp[-1].out_features
self.token_preprocessor.condition_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden, 6 * hidden))
nn.init.zeros_(self.token_preprocessor.condition_modulation[-1].weight)
nn.init.zeros_(self.token_preprocessor.condition_modulation[-1].bias)
for block in self.blocks:
block.condition_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden, 16 * hidden))
nn.init.zeros_(block.condition_modulation[-1].weight)
nn.init.zeros_(block.condition_modulation[-1].bias)
def forward(self, x, timesteps, labels, force_drop_mask=None):
if x.ndim != 3 or x.shape[1:] != (self.in_channels, self.length):
raise ValueError(f"Expected x shape [N, {self.in_channels}, {self.length}], got {tuple(x.shape)}")
if labels.shape != (x.shape[0], 2):
raise ValueError(f"Expected labels shape [{x.shape[0]}, 2], got {tuple(labels.shape)}")
normalized = labels / self.condition_normalization_length
drop_mask = self._drop_mask(normalized, force_drop_mask)
condition = (self.p_embedder(normalized[:, 0], drop_mask) + self.s_embedder(normalized[:, 1], drop_mask)) / math.sqrt(2.0)
timestep = self.timestep_embedder(timesteps)
small, mid, large = self.encoder(x)
small = small + self.small_position.to(dtype=small.dtype)
mid = mid + self.mid_position.to(dtype=mid.dtype)
large = large + self.large_position.to(dtype=large.dtype)
time_values = self.token_preprocessor.modulation(timestep).chunk(6, dim=1)
cond_values = self.token_preprocessor.condition_modulation(condition).chunk(6, dim=1)
small, mid, large = (
modulate(self.token_preprocessor.norms[0](small), time_values[0] + cond_values[0], time_values[1] + cond_values[1]),
modulate(self.token_preprocessor.norms[1](mid), time_values[2] + cond_values[2], time_values[3] + cond_values[3]),
modulate(self.token_preprocessor.norms[2](large), time_values[4] + cond_values[4], time_values[5] + cond_values[5]),
)
for block in self.blocks:
time_values = block.modulation(timestep).chunk(16, dim=1)
cond_values = block.condition_modulation(condition).chunk(16, dim=1)
values = [time_values[i] + cond_values[i] for i in range(16)]
small = small + values[2].unsqueeze(1) * block.self_attention(modulate(block.norm_self(small), values[0], values[1]))
query = modulate(block.norm_mid_query(small), values[3], values[4])
context = modulate(block.norm_mid_context(mid), values[5], values[6])
update, _ = block.mid_attention(query, context, context, need_weights=False)
small = small + values[7].unsqueeze(1) * update
query = modulate(block.norm_large_query(small), values[8], values[9])
context = modulate(block.norm_large_context(large), values[10], values[11])
update, _ = block.large_attention(query, context, context, need_weights=False)
small = small + values[12].unsqueeze(1) * update
small = small + values[15].unsqueeze(1) * block.mlp(modulate(block.norm_mlp(small), values[13], values[14]))
decoded = self.decoder_projection(small.transpose(1, 2))
decoded = F.interpolate(decoded, size=self.length, mode="linear", align_corners=False)
decoded = self.decoder_refine(decoded)
if self.skip_branch is not None:
decoded = torch.cat([decoded, self.skip_branch(x)], dim=1)
return self.output_head(self.fusion(decoded))
class AdaLNSumSWaG(ConditionalSWaG):
"""Frozen-backbone SWaG with independent P/S modulation parameters added to timestep parameters."""
def __init__(self, condition_normalization_length: float = 6000.0, **kwargs):
super().__init__(**kwargs)
self.condition_normalization_length = float(condition_normalization_length)
hidden = self.timestep_embedder.mlp[-1].out_features
self.token_preprocessor.condition_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden, 6 * hidden))
nn.init.zeros_(self.token_preprocessor.condition_modulation[-1].weight)
nn.init.zeros_(self.token_preprocessor.condition_modulation[-1].bias)
for block in self.blocks:
block.condition_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden, 16 * hidden))
nn.init.zeros_(block.condition_modulation[-1].weight)
nn.init.zeros_(block.condition_modulation[-1].bias)
def _embeddings(self, timesteps, labels, force_drop_mask):
normalized = labels / self.condition_normalization_length
drop = self._drop_mask(normalized, force_drop_mask)
condition = (self.p_embedder(normalized[:, 0], drop) + self.s_embedder(normalized[:, 1], drop)) / math.sqrt(2.0)
return self.timestep_embedder(timesteps), condition
def _decode(self, small, x):
decoded = self.decoder_projection(small.transpose(1, 2))
decoded = self.decoder_refine(F.interpolate(decoded, size=self.length, mode="linear", align_corners=False))
if self.skip_branch is not None:
decoded = torch.cat([decoded, self.skip_branch(x)], dim=1)
return self.output_head(self.fusion(decoded))
def forward(self, x, timesteps, labels, force_drop_mask=None):
timestep, condition = self._embeddings(timesteps, labels, force_drop_mask)
small, mid, large = self.encoder(x)
small, mid, large = small + self.small_position, mid + self.mid_position, large + self.large_position
tv = self.token_preprocessor.modulation(timestep).chunk(6, 1)
cv = self.token_preprocessor.condition_modulation(condition).chunk(6, 1)
small, mid, large = tuple(
modulate(self.token_preprocessor.norms[i](token), tv[2*i] + cv[2*i], tv[2*i+1] + cv[2*i+1])
for i, token in enumerate((small, mid, large))
)
for block in self.blocks:
t = block.modulation(timestep).chunk(16, 1); c = block.condition_modulation(condition).chunk(16, 1)
v = [t[i] + c[i] for i in range(16)]
small = small + v[2].unsqueeze(1) * block.self_attention(modulate(block.norm_self(small), v[0], v[1]))
q = modulate(block.norm_mid_query(small), v[3], v[4]); ctx = modulate(block.norm_mid_context(mid), v[5], v[6])
update, _ = block.mid_attention(q, ctx, ctx, need_weights=False); small = small + v[7].unsqueeze(1) * update
q = modulate(block.norm_large_query(small), v[8], v[9]); ctx = modulate(block.norm_large_context(large), v[10], v[11])
update, _ = block.large_attention(q, ctx, ctx, need_weights=False); small = small + v[12].unsqueeze(1) * update
small = small + v[15].unsqueeze(1) * block.mlp(modulate(block.norm_mlp(small), v[13], v[14]))
return self._decode(small, x)
class AdaLNResidualSWaG(AdaLNSumSWaG):
"""SWaG with timestep and P/S modulation applied as separate residual updates."""
def __init__(self, **kwargs):
super().__init__(**kwargs)
hidden = self.timestep_embedder.mlp[-1].out_features
self.token_preprocessor.condition_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden, 9 * hidden))
nn.init.zeros_(self.token_preprocessor.condition_modulation[-1].weight)
nn.init.zeros_(self.token_preprocessor.condition_modulation[-1].bias)
def forward(self, x, timesteps, labels, force_drop_mask=None):
timestep, condition = self._embeddings(timesteps, labels, force_drop_mask)
small, mid, large = self.encoder(x)
small, mid, large = small + self.small_position, mid + self.mid_position, large + self.large_position
original = (small, mid, large)
tv = self.token_preprocessor.modulation(timestep).chunk(6, 1)
cv = self.token_preprocessor.condition_modulation(condition).chunk(9, 1)
tokens = []
for i, token in enumerate(original):
norm = self.token_preprocessor.norms[i](token)
time_token = modulate(norm, tv[2*i], tv[2*i+1])
cond_token = modulate(norm, cv[3*i], cv[3*i+1])
tokens.append(time_token + cv[3*i+2].unsqueeze(1) * cond_token)
small, mid, large = tokens
for block in self.blocks:
t = block.modulation(timestep).chunk(16, 1); c = block.condition_modulation(condition).chunk(16, 1)
norm = block.norm_self(small)
small = small + t[2].unsqueeze(1) * block.self_attention(modulate(norm, t[0], t[1]))
small = small + c[2].unsqueeze(1) * block.self_attention(modulate(norm, c[0], c[1]))
qt = modulate(block.norm_mid_query(small), t[3], t[4]); ct = modulate(block.norm_mid_context(mid), t[5], t[6])
qc = modulate(block.norm_mid_query(small), c[3], c[4]); cc = modulate(block.norm_mid_context(mid), c[5], c[6])
u, _ = block.mid_attention(qt, ct, ct, need_weights=False); small = small + t[7].unsqueeze(1) * u
u, _ = block.mid_attention(qc, cc, cc, need_weights=False); small = small + c[7].unsqueeze(1) * u
qt = modulate(block.norm_large_query(small), t[8], t[9]); ct = modulate(block.norm_large_context(large), t[10], t[11])
qc = modulate(block.norm_large_query(small), c[8], c[9]); cc = modulate(block.norm_large_context(large), c[10], c[11])
u, _ = block.large_attention(qt, ct, ct, need_weights=False); small = small + t[12].unsqueeze(1) * u
u, _ = block.large_attention(qc, cc, cc, need_weights=False); small = small + c[12].unsqueeze(1) * u
norm = block.norm_mlp(small)
small = small + t[15].unsqueeze(1) * block.mlp(modulate(norm, t[13], t[14]))
small = small + c[15].unsqueeze(1) * block.mlp(modulate(norm, c[13], c[14]))
return self._decode(small, x)
class AdaLNResidualLoRASWaG(AdaLNResidualSWaG):
"""Independent residual AdaLN condition chain plus cross-attention LoRA."""
def __init__(self, cross_attention_lora_rank=8, cross_attention_lora_alpha=16.0,
cross_attention_lora_dropout=0.05, **kwargs):
super().__init__(**kwargs)
for block in self.blocks:
for name in ("mid_attention", "large_attention"):
base = getattr(block, name)
adapted = LoRAMultiheadAttention(base.embed_dim, base.num_heads, cross_attention_lora_rank,
cross_attention_lora_alpha, cross_attention_lora_dropout)
adapted.in_proj_weight.data.copy_(base.in_proj_weight.data)
adapted.in_proj_bias.data.copy_(base.in_proj_bias.data)
adapted.out_proj.load_state_dict(base.out_proj.state_dict())
setattr(block, name, adapted)
|