File size: 16,108 Bytes
aa3d7f9 | 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 | from __future__ import annotations
import math
import torch
from torch import Tensor, nn
from torch.nn import functional as F
def masked_mean_prototype(per_image: Tensor, mask: Tensor) -> Tensor:
if per_image.ndim != 3 or mask.shape != per_image.shape[:2]:
raise ValueError("expected per_image [batch,references,width] and matching mask")
mask = mask.to(device=per_image.device, dtype=torch.bool)
counts = mask.sum(dim=1)
if torch.any(counts < 1) or torch.any(counts > 8):
raise ValueError("each reference set must contain 1-8 images")
weights = mask.unsqueeze(-1).to(per_image.dtype)
mean = (per_image * weights).sum(dim=1) / counts.unsqueeze(-1)
return F.normalize(mean, dim=-1)
class StyleMixerBlock(nn.Module):
def __init__(self, width: int, heads: int, dropout: float = 0.0) -> None:
super().__init__()
self.query_norm = nn.LayerNorm(width)
self.token_norm = nn.LayerNorm(width)
self.attention = nn.MultiheadAttention(width, heads, batch_first=True)
self.attention_dropout = nn.Dropout(dropout)
self.output_norm = nn.LayerNorm(width)
self.mlp = nn.Sequential(
nn.Linear(width, 4 * width),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(4 * width, width),
nn.Dropout(dropout),
)
def forward(self, query: Tensor, tokens: Tensor, token_mask: Tensor) -> Tensor:
value, _ = self.attention(
self.query_norm(query),
self.token_norm(tokens),
self.token_norm(tokens),
key_padding_mask=~token_mask,
need_weights=False,
)
query = query + self.attention_dropout(value)
return query + self.mlp(self.output_norm(query))
class UnifiedStyleEncoder(nn.Module):
"""Encode cached full/face vision tokens and optional Anima descriptors."""
def __init__(
self,
*,
external_dim: int = 1152,
external_tokens: int = 30,
anima_dim: int = 4096,
anima_tokens: int = 3,
width: int = 384,
blocks: int = 2,
heads: int = 6,
embedding_dim: int = 512,
use_anima: bool = True,
dropout: float = 0.0,
reference_grounding: bool = False,
functional_factors: bool = False,
) -> None:
super().__init__()
if (
blocks < 1
or width % heads
or (reference_grounding and embedding_dim % heads)
or not 0.0 <= dropout < 1.0
):
raise ValueError("blocks must be positive and width must be divisible by heads")
if functional_factors and embedding_dim != 512:
raise ValueError("functional factorization currently requires a 512-D embedding")
self.external_tokens = external_tokens
self.anima_tokens = anima_tokens if use_anima else 0
self.external_projector = nn.Sequential(
nn.LayerNorm(external_dim),
nn.Linear(external_dim, width),
)
self.anima_projector = (
nn.Sequential(nn.LayerNorm(anima_dim), nn.Linear(anima_dim, width))
if use_anima
else None
)
self.full_positions = nn.Parameter(torch.randn(external_tokens, width) * 0.02)
self.face_positions = nn.Parameter(torch.randn(external_tokens, width) * 0.02)
self.anima_positions = (
nn.Parameter(torch.randn(anima_tokens, width) * 0.02) if use_anima else None
)
self.reference_grounding = reference_grounding
self.functional_factors = functional_factors
query_count = 5 if functional_factors else 1
self.style_query = nn.Parameter(torch.randn(1, query_count, width) * 0.02)
self.blocks = nn.ModuleList(
[StyleMixerBlock(width, heads, dropout) for _ in range(blocks)]
)
if functional_factors:
self.factor_output = nn.Sequential(nn.LayerNorm(width), nn.Linear(width, 96))
self.shared_output = nn.Sequential(nn.LayerNorm(width), nn.Linear(width, 128))
self.output = nn.Sequential(nn.LayerNorm(512), nn.Linear(512, embedding_dim))
self.factor_decoders = nn.ModuleList([nn.Linear(96, external_dim) for _ in range(4)])
else:
self.output = nn.Sequential(nn.LayerNorm(width), nn.Linear(width, embedding_dim))
if reference_grounding:
self.consensus_query = nn.Parameter(torch.randn(1, 1, embedding_dim) * 0.02)
self.consensus_attention = nn.MultiheadAttention(
embedding_dim, heads, batch_first=True, dropout=dropout
)
self.consensus_condition = nn.Linear(embedding_dim, width)
self.consensus_gamma = nn.Parameter(torch.tensor(-2.9444))
def _read_tokens(
self, tokens: Tensor, token_mask: Tensor, conditioning: Tensor | None = None
) -> tuple[Tensor, Tensor | None]:
query = self.style_query.expand(tokens.shape[0], -1, -1)
if conditioning is not None:
query = query + self.consensus_condition(conditioning)[:, None]
for block in self.blocks:
query = block(query, tokens, token_mask)
if self.functional_factors:
factors = F.normalize(self.factor_output(query[:, :4]), dim=-1)
shared = self.shared_output(query[:, 4])
embedding = self.output(torch.cat((factors.flatten(1), shared), dim=-1))
return F.normalize(embedding, dim=-1), factors
return F.normalize(self.output(query[:, 0]), dim=-1), None
def encode_images(
self,
full_features: Tensor,
*,
face_features: Tensor | None = None,
face_mask: Tensor | None = None,
anima_features: Tensor | None = None,
conditioning: Tensor | None = None,
) -> Tensor:
"""Encode independent images without padding them into reference sets."""
if full_features.ndim != 3 or full_features.shape[1] != self.external_tokens:
raise ValueError("full_features must have shape [images,external_tokens,width]")
images = full_features.shape[0]
full = self.external_projector(full_features) + self.full_positions
tokens = [full]
masks = [torch.ones(full.shape[:2], dtype=torch.bool, device=full.device)]
if face_features is not None:
if face_features.shape != full_features.shape:
raise ValueError("face_features must match full_features")
if face_mask is None or face_mask.shape != (images,):
raise ValueError("face_mask is required and must contain one value per image")
face = self.external_projector(face_features)
tokens.append(face + self.face_positions)
masks.append(face_mask[:, None].to(torch.bool).expand(-1, self.external_tokens))
if self.anima_projector is not None:
expected = (images, self.anima_tokens)
if anima_features is None or anima_features.shape[:2] != expected:
raise ValueError("anima_features are required by this encoder configuration")
anima = self.anima_projector(anima_features)
tokens.append(anima + self.anima_positions)
masks.append(torch.ones(anima.shape[:2], dtype=torch.bool, device=anima.device))
token_tensor = torch.cat(tokens, dim=1)
token_mask = torch.cat(masks, dim=1)
return self._read_tokens(token_tensor, token_mask, conditioning)[0]
def encode_images_with_factors(
self,
full_features: Tensor,
*,
face_features: Tensor | None = None,
face_mask: Tensor | None = None,
anima_features: Tensor | None = None,
conditioning: Tensor | None = None,
) -> tuple[Tensor, Tensor | None]:
if full_features.ndim != 3 or full_features.shape[1] != self.external_tokens:
raise ValueError("full_features must have shape [images,external_tokens,width]")
images = full_features.shape[0]
full = self.external_projector(full_features) + self.full_positions
tokens = [full]
masks = [torch.ones(full.shape[:2], dtype=torch.bool, device=full.device)]
if face_features is not None:
if face_features.shape != full_features.shape:
raise ValueError("face_features must match full_features")
if face_mask is None or face_mask.shape != (images,):
raise ValueError("face_mask is required and must contain one value per image")
tokens.append(self.external_projector(face_features) + self.face_positions)
masks.append(face_mask[:, None].to(torch.bool).expand(-1, self.external_tokens))
if self.anima_projector is not None:
if anima_features is None or anima_features.shape[:2] != (images, self.anima_tokens):
raise ValueError("anima_features are required by this encoder configuration")
tokens.append(self.anima_projector(anima_features) + self.anima_positions)
masks.append(torch.ones(anima_features.shape[:2], dtype=torch.bool, device=full.device))
return self._read_tokens(torch.cat(tokens, dim=1), torch.cat(masks, dim=1), conditioning)
def encode_reference_sets(
self,
full_features: Tensor,
reference_mask: Tensor,
*,
face_features: Tensor | None = None,
face_mask: Tensor | None = None,
anima_features: Tensor | None = None,
) -> tuple[Tensor, Tensor, Tensor | None]:
batch, references = full_features.shape[:2]
flat_face = None if face_features is None else face_features.flatten(0, 1)
flat_face_mask = None if face_mask is None else face_mask.flatten()
flat_anima = None if anima_features is None else anima_features.flatten(0, 1)
per_image, factors = self.encode_images_with_factors(
full_features.flatten(0, 1),
face_features=flat_face,
face_mask=flat_face_mask,
anima_features=flat_anima,
)
per_image = per_image.reshape(batch, references, -1)
factors = None if factors is None else factors.reshape(batch, references, 4, -1)
base = masked_mean_prototype(per_image, reference_mask)
if not self.reference_grounding:
return base, per_image, factors
consensus, _ = self.consensus_attention(
self.consensus_query.expand(batch, -1, -1),
per_image,
per_image,
key_padding_mask=~reference_mask,
need_weights=False,
)
consensus = F.normalize(consensus[:, 0], dim=-1)
reread, factors = self.encode_images_with_factors(
full_features.flatten(0, 1),
face_features=flat_face,
face_mask=flat_face_mask,
anima_features=flat_anima,
conditioning=consensus[:, None].expand(-1, references, -1).reshape(batch * references, -1),
)
reread = reread.reshape(batch, references, -1)
factors = None if factors is None else factors.reshape(batch, references, 4, -1)
agreement = F.cosine_similarity(reread, consensus[:, None], dim=-1)
centered = agreement - (
(agreement * reference_mask).sum(1, keepdim=True)
/ reference_mask.sum(1, keepdim=True).clamp_min(1)
)
weights = (1.0 + 0.25 * torch.tanh(4.0 * centered)) * reference_mask
grounded = F.normalize((reread * weights[..., None]).sum(1) / weights.sum(1, keepdim=True), dim=-1)
gamma = torch.sigmoid(self.consensus_gamma)
return F.normalize(base + gamma * (grounded - base), dim=-1), reread, factors
def functional_response_loss(self, factor_codes: Tensor, full_features: Tensor) -> Tensor:
"""Reconstruct artist response after removing a shared prompt/seed cell mean."""
if not self.functional_factors:
return full_features.new_zeros(())
groups = ((2, 3, 4, 5), (0, 6, 24), tuple(range(8, 24)), (1, 7, 25))
losses = []
for index, token_indices in enumerate(groups):
teacher = full_features[..., list(token_indices), :].float().mean(dim=-2)
teacher = F.normalize(teacher - teacher.mean(dim=1, keepdim=True), dim=-1)
prediction = F.normalize(self.factor_decoders[index](factor_codes[..., index, :]).float(), dim=-1)
losses.append((1.0 - F.cosine_similarity(prediction, teacher, dim=-1)).mean())
return torch.stack(losses).mean()
def forward(
self,
full_features: Tensor,
reference_mask: Tensor,
*,
face_features: Tensor | None = None,
face_mask: Tensor | None = None,
anima_features: Tensor | None = None,
) -> tuple[Tensor, Tensor]:
if full_features.ndim != 4 or full_features.shape[2] != self.external_tokens:
raise ValueError("full_features must have shape [batch,references,external_tokens,width]")
batch, references = full_features.shape[:2]
if reference_mask.shape != (batch, references):
raise ValueError("reference_mask shape does not match full_features")
prototype, per_image, _ = self.encode_reference_sets(
full_features,
reference_mask,
face_features=face_features,
face_mask=face_mask,
anima_features=anima_features,
)
return prototype, per_image
class AngularPrototypeLoss(nn.Module):
"""Episode prototype softmax with an optional ArcFace-style positive margin."""
def __init__(
self, initial_scale: float = 10.0, initial_bias: float | None = None
) -> None:
super().__init__()
if initial_scale <= 0:
raise ValueError("initial_scale must be positive")
self.raw_scale = nn.Parameter(torch.tensor(math.log(math.expm1(initial_scale))))
# initial_bias remains accepted for feasibility-script compatibility. A
# common softmax bias is a no-op and is not a trainable production parameter.
def forward(
self,
queries: Tensor,
prototypes: Tensor,
targets: Tensor,
*,
margin: float = 0.0,
) -> tuple[Tensor, Tensor]:
if queries.ndim < 2 or prototypes.ndim != queries.ndim:
raise ValueError("queries and prototypes must have matching [episode,...,width] ranks")
if queries.shape[:-2] != prototypes.shape[:-2] or queries.shape[-1] != prototypes.shape[-1]:
raise ValueError("queries and prototypes must have matching episode axes and widths")
if targets.shape != queries.shape[:-1]:
raise ValueError("targets must contain one class index per query")
# Autocast can round a near-perfect BF16 cosine to exactly one. ArcFace's
# sqrt(1-cos²) derivative is then singular, so margin math and CE stay FP32.
cosine = (
F.normalize(queries, dim=-1)
@ F.normalize(prototypes, dim=-1).transpose(-1, -2)
).float()
loss_cosine = cosine
if margin:
if not 0.0 <= margin < math.pi / 2:
raise ValueError("margin must be in [0, pi/2)")
positive = cosine.gather(-1, targets.unsqueeze(-1)).clamp(-1 + 1e-6, 1 - 1e-6)
positive_margin = positive * math.cos(margin) - torch.sqrt(1 - positive.square()) * math.sin(margin)
loss_cosine = cosine.scatter(-1, targets.unsqueeze(-1), positive_margin)
scale = F.softplus(self.raw_scale).clamp_max(100.0)
loss_logits = scale * loss_cosine
metric_logits = scale * cosine
return (
F.cross_entropy(loss_logits.flatten(0, -2), targets.flatten()),
metric_logits,
)
def subset_consistency_loss(first: Tensor, second: Tensor) -> Tensor:
if first.shape != second.shape or first.ndim != 2:
raise ValueError("subset prototypes must have equal [batch,width] shapes")
return (1.0 - F.cosine_similarity(first, second, dim=-1)).mean()
|