File size: 18,694 Bytes
d9646ca |
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 |
"""
Baseline Vision Transformer with Frozen Pentachora Embeddings
Adapted for L1-normalized pentachora vertices
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from einops import rearrange
import math
from typing import Optional, Tuple, Dict, Any
class PentachoraEmbedding(nn.Module):
"""
A single frozen pentachora embedding (5 vertices in geometric space).
Supports both L1 and L2 normalized vertices.
"""
def __init__(self, vertices: torch.Tensor, norm_type: str = 'l1'):
super().__init__()
self.embed_dim = vertices.shape[-1]
self.norm_type = norm_type
# Store provided vertices as frozen buffer
self.register_buffer('vertices', vertices)
self.vertices.requires_grad = False
# Precompute normalized versions and centroid
with torch.no_grad():
# For L1-normalized data, use L1 norm for consistency
if norm_type == 'l1':
# L1 normalize (sum of abs values = 1)
self.register_buffer('vertices_norm',
vertices / (vertices.abs().sum(dim=-1, keepdim=True) + 1e-8))
else:
# L2 normalize (euclidean norm = 1)
self.register_buffer('vertices_norm', F.normalize(self.vertices, dim=-1))
self.register_buffer('centroid', self.vertices.mean(dim=0))
# Centroid normalization matches vertex normalization
if norm_type == 'l1':
self.register_buffer('centroid_norm',
self.centroid / (self.centroid.abs().sum() + 1e-8))
else:
self.register_buffer('centroid_norm', F.normalize(self.centroid, dim=-1))
def get_vertices(self) -> torch.Tensor:
"""Get all 5 vertices."""
return self.vertices
def get_centroid(self) -> torch.Tensor:
"""Get the centroid of the pentachora."""
return self.centroid
def compute_rose_score(self, features: torch.Tensor) -> torch.Tensor:
"""
Compute Rose similarity score with this pentachora.
Scaled appropriately for L1 norm.
"""
verts = self.vertices.unsqueeze(0) # [1, 5, D]
if features.dim() == 1:
features = features.unsqueeze(0)
B = features.shape[0]
if B > 1:
verts = verts.expand(B, -1, -1)
# For L1 norm, scale the rose score appropriately
score = PentachoronStabilizer.rose_score_magnitude(features, verts)
if self.norm_type == 'l1':
# L1 norm produces smaller values, so amplify the signal
score = score * 10.0
return score
def compute_similarity(self, features: torch.Tensor, mode: str = 'centroid') -> torch.Tensor:
"""
Compute similarity between features and this pentachora.
"""
if mode == 'rose':
return self.compute_rose_score(features)
# Normalize features according to norm type
if self.norm_type == 'l1':
features_norm = features / (features.abs().sum(dim=-1, keepdim=True) + 1e-8)
else:
features_norm = F.normalize(features, dim=-1)
if mode == 'centroid':
# Dot product with centroid
sim = torch.sum(features_norm * self.centroid_norm, dim=-1)
# Scale up L1 similarities to be comparable to L2
if self.norm_type == 'l1':
sim = sim * 10.0
return sim
else: # mode == 'max'
# Max similarity across vertices
sims = torch.matmul(features_norm, self.vertices_norm.T)
if self.norm_type == 'l1':
sims = sims * 10.0
return sims.max(dim=-1)[0]
class TransformerBlock(nn.Module):
"""Standard transformer block with multi-head attention and MLP."""
def __init__(
self,
dim: int,
num_heads: int = 8,
mlp_ratio: float = 4.0,
dropout: float = 0.0,
attn_dropout: float = 0.0
):
super().__init__()
self.norm1 = nn.LayerNorm(dim)
self.attn = nn.MultiheadAttention(
dim,
num_heads,
dropout=attn_dropout,
batch_first=True
)
self.norm2 = nn.LayerNorm(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = nn.Sequential(
nn.Linear(dim, mlp_hidden_dim),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(mlp_hidden_dim, dim),
nn.Dropout(dropout)
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
# Self-attention
x_norm = self.norm1(x)
attn_out, _ = self.attn(x_norm, x_norm, x_norm)
x = x + attn_out
# MLP
x = x + self.mlp(self.norm2(x))
return x
class BaselineViT(nn.Module):
"""
Vision Transformer with frozen pentachora embeddings.
- Preserves L1 law for pentachora geometry.
- Uses L2 angles for RoseFace (ArcFace/CosFace/SphereFace) classification.
"""
def __init__(
self,
pentachora_list: list, # List of torch.Tensor, each [5, vocab_dim]
vocab_dim: int = 256,
img_size: int = 32,
patch_size: int = 4,
embed_dim: int = 512,
depth: int = 12,
num_heads: int = 8,
mlp_ratio: float = 4.0,
dropout: float = 0.0,
attn_dropout: float = 0.0,
similarity_mode: str = 'rose', # legacy similarity (kept for compatibility)
norm_type: str = 'l1', # 'l1' or 'l2' normalization for pentachora law
# --- New RoseFace config ---
head_type: str = 'roseface', # 'roseface' | 'legacy'
prototype_mode: str = 'centroid',# 'centroid' | 'rose5' | 'max_vertex'
margin_type: str = 'cosface', # 'arcface' | 'cosface' | 'sphereface'
margin_m: float = 0.30,
scale_s: float = 30.0,
apply_margin_train_only: bool = False,
):
super().__init__()
# Validate pentachora list
assert isinstance(pentachora_list, list), f"Expected list, got {type(pentachora_list)}"
assert len(pentachora_list) > 0, "Empty pentachora list"
for i, penta in enumerate(pentachora_list):
assert isinstance(penta, torch.Tensor), f"Item {i} is not a tensor"
self.num_classes = len(pentachora_list)
self.embed_dim = embed_dim
self.num_patches = (img_size // patch_size) ** 2
self.similarity_mode = similarity_mode
self.pentachora_dim = vocab_dim
self.norm_type = norm_type
# --- RoseFace config ---
self.head_type = head_type
self.prototype_mode = prototype_mode
self.margin_type = margin_type
self.margin_m = float(margin_m)
self.scale_s = float(scale_s)
self.apply_margin_train_only = apply_margin_train_only
# Create individual pentachora embeddings from list
self.class_pentachora = nn.ModuleList([
PentachoraEmbedding(vertices=penta, norm_type=norm_type)
for penta in pentachora_list
])
# Patch embedding
self.patch_embed = nn.Conv2d(3, embed_dim, kernel_size=patch_size, stride=patch_size)
# CLS token - learnable
self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
# Position embeddings
self.pos_embed = nn.Parameter(torch.zeros(1, 1 + self.num_patches, embed_dim))
self.pos_drop = nn.Dropout(dropout)
# Transformer blocks
self.blocks = nn.ModuleList([
TransformerBlock(
dim=embed_dim,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
dropout=dropout,
attn_dropout=attn_dropout
)
for _ in range(depth)
])
# Final norm
self.norm = nn.LayerNorm(embed_dim)
# Project to pentachora dimension if needed
if self.pentachora_dim != embed_dim:
self.to_pentachora_dim = nn.Linear(embed_dim, self.pentachora_dim)
else:
self.to_pentachora_dim = nn.Identity()
# Legacy temperature (used only if head_type == 'legacy')
if norm_type == 'l1':
self.temperature = nn.Parameter(torch.zeros(1)) # exp(0)=1
else:
self.temperature = nn.Parameter(torch.ones(1) * np.log(1/0.07))
# Precompute all centroids (buffers) for legacy path
self.register_buffer(
'all_centroids',
torch.stack([penta.centroid for penta in self.class_pentachora])
)
if norm_type == 'l1':
centroids_normalized = self.all_centroids / (
self.all_centroids.abs().sum(dim=-1, keepdim=True) + 1e-8)
else:
centroids_normalized = F.normalize(self.all_centroids, dim=-1)
self.register_buffer('all_centroids_norm', centroids_normalized)
# Face weights for rose5 prototypes (10 triads)
face_triplets = torch.tensor([
[0,1,2],[0,1,3],[0,1,4],
[0,2,3],[0,2,4],[0,3,4],
[1,2,3],[1,2,4],[1,3,4],
[2,3,4]
], dtype=torch.long)
face_weights = torch.zeros(10, 5, dtype=torch.float32)
for r, (i,j,k) in enumerate(face_triplets):
face_weights[r, i] = face_weights[r, j] = face_weights[r, k] = 1.0/3.0
self.register_buffer('rose_face_weights', face_weights, persistent=False)
# Initialize weights
self.init_weights()
# Record config for checkpoint saving
self.config = getattr(self, 'config', {})
self.config.update({
'head_type': self.head_type,
'prototype_mode': self.prototype_mode,
'margin_type': self.margin_type,
'margin_m': self.margin_m,
'scale_s': self.scale_s,
'apply_margin_train_only': self.apply_margin_train_only,
'norm_type': self.norm_type,
'similarity_mode': self.similarity_mode,
'pentachora_dim': self.pentachora_dim,
})
def init_weights(self):
nn.init.trunc_normal_(self.cls_token, std=0.02)
nn.init.trunc_normal_(self.pos_embed, std=0.02)
for m in self.modules():
if isinstance(m, nn.Linear):
nn.init.trunc_normal_(m.weight, std=0.02)
if m.bias is not None:
nn.init.zeros_(m.bias)
elif isinstance(m, nn.LayerNorm):
nn.init.ones_(m.weight)
nn.init.zeros_(m.bias)
# ---- Legacy helper (kept) ----
def get_class_centroids(self) -> torch.Tensor:
return self.all_centroids_norm
# ---- Legacy similarity (kept for compatibility & debugging) ----
def compute_pentachora_similarities(self, features: torch.Tensor) -> torch.Tensor:
if self.similarity_mode == 'rose':
all_vertices = torch.stack([penta.vertices for penta in self.class_pentachora])
features_exp = features.unsqueeze(1).expand(-1, self.num_classes, -1)
scores = PentachoronStabilizer.rose_score_magnitude(
features_exp.reshape(-1, self.pentachora_dim),
all_vertices.repeat(features.shape[0], 1, 1)
).reshape(features.shape[0], -1)
if self.norm_type == 'l1':
scores = scores * 10.0
return scores
else:
if self.norm_type == 'l1':
features_norm = features / (features.abs().sum(dim=-1, keepdim=True) + 1e-8)
else:
features_norm = F.normalize(features, dim=-1)
centroids = self.get_class_centroids()
sims = torch.matmul(features_norm, centroids.T)
if self.norm_type == 'l1':
sims = sims * 10.0
return sims
# ---- RoseFace utilities ----
@staticmethod
def _l2_norm(x: torch.Tensor, eps: float = 1e-12) -> torch.Tensor:
return x / (x.norm(p=2, dim=-1, keepdim=True) + eps)
def _get_class_vertices_l2(self) -> torch.Tensor:
"""[C,5,D] L2-normalized vertices for all classes."""
V = torch.stack([p.vertices for p in self.class_pentachora], dim=0)
V = V.to(self.pos_embed.device, dtype=self.pos_embed.dtype)
return self._l2_norm(V)
def _get_prototypes(self, mode: Optional[str] = None) -> Optional[torch.Tensor]:
"""
Prototypes [C,D] for 'centroid'/'rose5'; None for 'max_vertex'.
"""
mode = mode or self.prototype_mode
device = self.pos_embed.device
dtype = self.pos_embed.dtype
if mode == 'centroid':
C = torch.stack([p.centroid for p in self.class_pentachora], dim=0).to(device, dtype)
return self._l2_norm(C)
elif mode == 'rose5':
V_l2 = self._get_class_vertices_l2() # [C,5,D]
W = self.rose_face_weights.to(device=device, dtype=dtype) # [10,5]
faces = torch.einsum('tf,cfd->ctd', W, V_l2) # [C,10,D]
verts_mean = V_l2.mean(dim=1) # [C,D]
faces_mean = faces.mean(dim=1) # [C,D]
alpha, beta = 1.0, 0.5
proto = alpha * verts_mean + beta * faces_mean
return self._l2_norm(proto)
elif mode == 'max_vertex':
return None
else:
raise ValueError(f"Unknown prototype_mode: {mode}")
def _cosine_matrix(self, z_l2: torch.Tensor) -> torch.Tensor:
"""
Pre-margin cosine [B,C] based on prototype_mode.
"""
if self.prototype_mode in ('centroid', 'rose5'):
P = self._get_prototypes(self.prototype_mode) # [C,D]
return torch.matmul(z_l2, P.t()) # [B,C]
elif self.prototype_mode == 'max_vertex':
V_l2 = self._get_class_vertices_l2() # [C,5,D]
cos_cv = torch.einsum('bd,cvd->bcv', z_l2, V_l2) # [B,C,5]
cos_max, _ = cos_cv.max(dim=2) # [B,C]
return cos_max
else:
raise ValueError(f"Unknown prototype_mode: {self.prototype_mode}")
@staticmethod
def _apply_margin(cosine: torch.Tensor, targets: torch.Tensor, m: float, kind: str = 'cosface') -> torch.Tensor:
"""
Apply margin to target class cosines. Returns adjusted cosines [B,C].
"""
eps = 1e-7
B, C = cosine.shape
y = targets.view(-1, 1) # [B,1]
if kind == 'cosface':
cos_m = cosine.clone()
cos_m.scatter_(1, y, (cosine.gather(1, y) - m))
return cos_m
theta = torch.acos(torch.clamp(cosine.gather(1, y), -1.0 + eps, 1.0 - eps)) # [B,1]
if kind == 'arcface':
cos_margin = torch.cos(theta + m)
elif kind == 'sphereface':
cos_margin = torch.cos(m * theta)
else:
raise ValueError(f"Unknown margin type: {kind}")
cos_m = cosine.clone()
cos_m.scatter_(1, y, cos_margin)
return cos_m
def schedule_roseface(
self, epoch: int, warmup_epochs: int = 15, s_start: float = 10.0, s_final: float = 30.0,
m_start: Optional[float] = None, m_final: Optional[float] = None
):
"""
Deterministic cosine ramp for scale s (and optional margin m).
"""
t = max(0.0, min(1.0, epoch / max(1, warmup_epochs)))
# cosine ramp from s_start -> s_final
self.scale_s = float(s_final - 0.5 * (1.0 + np.cos(np.pi * t)) * (s_final - s_start))
if (m_start is not None) and (m_final is not None):
self.margin_m = float(m_final - 0.5 * (1.0 + np.cos(np.pi * t)) * (m_final - m_start))
def forward_features(self, x: torch.Tensor) -> torch.Tensor:
B = x.shape[0]
x = self.patch_embed(x) # [B, embed_dim, H', W']
x = x.flatten(2).transpose(1, 2) # [B, num_patches, embed_dim]
cls_tokens = self.cls_token.expand(B, -1, -1)
x = torch.cat([cls_tokens, x], dim=1)
x = x + self.pos_embed
x = self.pos_drop(x)
for block in self.blocks:
x = block(x)
x = self.norm(x)
return x[:, 0]
def forward(
self,
x: torch.Tensor,
return_features: bool = False,
targets: Optional[torch.Tensor] = None # NEW: required for margin at train time
) -> Dict[str, torch.Tensor]:
features = self.forward_features(x)
output: Dict[str, torch.Tensor] = {}
# Project to pentachora dimension (L1 law applies here)
features_proj = self.to_pentachora_dim(features)
if self.norm_type == 'l1':
features_proj = features_proj / (features_proj.abs().sum(dim=-1, keepdim=True) + 1e-8)
if self.head_type == 'roseface':
# L2 angles for classification head (dual-norm bridge)
z_l2 = features_proj / (features_proj.norm(p=2, dim=-1, keepdim=True) + 1e-12)
# Pre-margin cosines [B,C]
cos_pre = self._cosine_matrix(z_l2)
# Apply margin (train-time if configured)
if (self.apply_margin_train_only and not self.training) or (targets is None):
cos_post = cos_pre
else:
cos_post = self._apply_margin(cos_pre, targets, self.margin_m, self.margin_type)
# Scaled logits
logits = self.scale_s * cos_post
# Emit outputs
output['logits'] = logits # for CE
output['similarities'] = cos_pre # pre-margin (for alignment / diagnostics)
if return_features:
output['features'] = features
output['features_proj'] = features_proj
else:
# Legacy path (kept for compatibility)
similarities = self.compute_pentachora_similarities(features_proj)
logits = similarities * self.temperature.exp()
output['logits'] = logits
output['similarities'] = similarities
if return_features:
output['features'] = features
output['features_proj'] = features_proj
return output
# Test - requires external setup
if __name__ == "__main__":
print("BaselineViT requires:")
print(" 1. PentachoronStabilizer loaded externally")
print(" 2. pentachora_batch tensor [num_classes, 5, vocab_dim]")
print("\nNo random initialization. No fallbacks.") |