File size: 42,975 Bytes
dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 dd3435c d7a1a55 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 |
"""
PentachoraViT: Vision Transformer with Pentachoron Geometric Structure
Enhanced with Geometric Attention for improved head cohesion and generalization
FIXED: All parameters initialized at module creation time (no lazy init)
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from einops import rearrange, repeat
import math
from typing import Optional, Dict, Tuple, List, Any
from dataclasses import dataclass
import warnings
# ============================================
# CONFIGURATION CLASSES
# ============================================
@dataclass
class PentachoraConfig:
"""Configuration for PentachoraViT models."""
img_size: int = 32
patch_size: int = 4
num_classes: int = 100
dim: int = 512
vocab_dim: Optional[int] = None # Vocabulary dimension (can differ from model dim)
depth: int = 12
heads: int = 8
mlp_ratio: float = 4.0
use_mesh_attention: bool = True
preserve_structure_until_layer: int = 6
dropout_rate: float = 0.0
drop_path_rate: float = 0.0
aux_loss_weight: float = 0.0
geo_loss_weight: float = 0.0
vocab: Optional[Any] = None
@property
def num_patches(self) -> int:
return (self.img_size // self.patch_size) ** 2
# ============================================
# GEOMETRIC ATTENTION COMPONENTS (FIXED INIT)
# ============================================
def perfect_4simplex(device):
"""Create perfect 4-simplex (pentachoron) vertices in 4D."""
sqrt5 = math.sqrt(5)
vertices = torch.tensor([
[1, 1, 1, -1/sqrt5],
[1, -1, -1, -1/sqrt5],
[-1, 1, -1, -1/sqrt5],
[-1, -1, 1, -1/sqrt5],
[0, 0, 0, 4/sqrt5]
], device=device, dtype=torch.float32)
return vertices / 2 # Normalize scale
def softmin_over_last(distances, tau):
"""Softmin over last dimension."""
return F.softmax(-distances / tau, dim=-1).sum(dim=-1)
@dataclass
class GeometricConfig:
"""Configuration for geometric attention."""
softmin_tau: float = 0.05
fuse_alpha: float = 0.7
phases: Tuple[float, ...] = (0.0, math.pi/2, math.pi, 3*math.pi/2)
jitter: float = 0.02
shift: float = 0.71
rotate_cycle: int = 11
use_phase_variance: bool = False
geometry_type: str = "pentachoron"
class GeometricNavigator(nn.Module):
"""Maps inputs to geometric regions in 4D space - FIXED with immediate initialization."""
def __init__(self, input_dim: int, num_regions: int, config: GeometricConfig, num_heads: int = 1, device=None):
super().__init__()
self.input_dim = input_dim
self.num_regions = num_regions
self.config = config
self.num_heads = num_heads
# Use CPU by default if device not specified
if device is None:
device = torch.device('cpu')
# Create separate parameters for each head if num_heads > 1
if num_heads > 1:
self.to_nav = nn.Parameter(torch.randn(num_heads, input_dim, 4, device=device) * 0.02)
self.vertex_w = nn.Parameter(torch.zeros(num_heads, num_regions, 5, device=device))
else:
self.to_nav = nn.Linear(input_dim, 4, bias=False)
self.vertex_w = nn.Parameter(torch.zeros(num_regions, 5, device=device))
# Pre-compute phase tensors for vectorization
self.register_buffer('phase_cos', torch.cos(torch.tensor(config.phases, dtype=torch.float32, device=device)))
self.register_buffer('phase_sin', torch.sin(torch.tensor(config.phases, dtype=torch.float32, device=device)))
# Initialize geometry immediately at creation time
self._init_geometry(device)
def _init_geometry(self, device):
"""Initialize geometry at module creation time."""
base = perfect_4simplex(device)
if self.num_heads > 1:
D = torch.zeros(self.num_heads, self.num_regions, 5, 4, device=device)
S = torch.zeros(self.num_heads, self.num_regions, 5, 4, device=device)
for h in range(self.num_heads):
for r in range(self.num_regions):
D[h, r] = base + self.config.jitter * torch.randn_like(base)
theta = torch.tensor(0.27 + 0.05 * (r % self.config.rotate_cycle), device=device)
rot = torch.eye(4, device=device)
c, s_val = torch.cos(theta), torch.sin(theta)
rot[0, 0] = c; rot[0, 1] = -s_val
rot[1, 0] = s_val; rot[1, 1] = c
S[h, r] = (base @ rot) + self.config.shift
S[h, r] += self.config.jitter * torch.randn_like(S[h, r])
else:
D = torch.zeros(self.num_regions, 5, 4, device=device)
S = torch.zeros(self.num_regions, 5, 4, device=device)
for r in range(self.num_regions):
D[r] = base + self.config.jitter * torch.randn_like(base)
theta = torch.tensor(0.27 + 0.05 * (r % self.config.rotate_cycle), device=device)
rot = torch.eye(4, device=device)
c, s_val = torch.cos(theta), torch.sin(theta)
rot[0, 0] = c; rot[0, 1] = -s_val
rot[1, 0] = s_val; rot[1, 1] = c
S[r] = (base @ rot) + self.config.shift
S[r] += self.config.jitter * torch.randn_like(S[r])
self.D = nn.Parameter(D)
self.S = nn.Parameter(S)
def navigate(self, x: torch.Tensor) -> Dict[str, torch.Tensor]:
"""Navigate inputs through geometric space - OPTIMIZED with vectorized phase computation."""
if self.num_heads > 1:
# Batched navigation for multiple heads
BT, H, head_dim = x.shape
# Batched transformation
nav_x = torch.einsum('bhi,hio->bho', x, self.to_nav) # [BT, H, 4]
# Dispatcher scores
nav_x_disp = nav_x.view(BT, H, 1, 1, 4)
D_exp = self.D.unsqueeze(0) # [1, H, regions, 5, 4]
d_disp = torch.norm(nav_x_disp - D_exp, dim=-1)
s_disp = -softmin_over_last(d_disp, self.config.softmin_tau)
# OPTIMIZED: Vectorized phase computation (no loop)
cos_phases = self.phase_cos.view(-1, 1, 1, 1, 1)
sin_phases = self.phase_sin.view(-1, 1, 1, 1, 1)
# Compute all phase variants at once [phases, H, regions, 5, 4]
Vt_all = cos_phases * self.D.unsqueeze(0) + sin_phases * self.S.unsqueeze(0)
# Apply vertex weighting to all phases
w = F.softmax(self.vertex_w, dim=-1)
w_exp = w.unsqueeze(0).unsqueeze(-1) # [1, H, regions, 5, 1]
Vt_mean = Vt_all.mean(dim=3, keepdim=True)
Vt_all = (1.0 - w_exp) * Vt_all + w_exp * Vt_mean
# Compute all ribbon distances at once
nav_x_ribbon = nav_x.view(BT, 1, H, 1, 1, 4)
Vt_exp = Vt_all.unsqueeze(0) # [1, phases, H, regions, 5, 4]
d_ribbon_all = torch.norm(nav_x_ribbon - Vt_exp, dim=-1)
s_ribbon_all = -softmin_over_last(d_ribbon_all, self.config.softmin_tau)
s_ribbon = s_ribbon_all.mean(dim=1) # Average over phases
scores = self.config.fuse_alpha * s_ribbon + (1 - self.config.fuse_alpha) * s_disp
scores = scores.reshape(BT * H, self.num_regions)
else:
# Original single-head navigation
nav_x = self.to_nav(x)
nav_x_exp = nav_x[:, None, None, :]
D_exp = self.D[None, :, :, :]
d_disp = torch.norm(nav_x_exp - D_exp, dim=-1)
s_disp = -softmin_over_last(d_disp, self.config.softmin_tau)
w = F.softmax(self.vertex_w, dim=1)
# OPTIMIZED: Vectorized phase computation for single head
cos_phases = self.phase_cos.view(-1, 1, 1, 1)
sin_phases = self.phase_sin.view(-1, 1, 1, 1)
Vt_all = cos_phases * self.D.unsqueeze(0) + sin_phases * self.S.unsqueeze(0)
w_expanded = w.unsqueeze(0).unsqueeze(-1)
Vt_mean = Vt_all.mean(dim=2, keepdim=True)
Vt_all = (1.0 - w_expanded) * Vt_all + w_expanded * Vt_mean
nav_x_phase = nav_x[:, None, None, None, :]
Vt_exp = Vt_all.unsqueeze(0)
d_ribbon_all = torch.norm(nav_x_phase - Vt_exp, dim=-1)
s_ribbon_all = -softmin_over_last(d_ribbon_all, self.config.softmin_tau)
s_ribbon = s_ribbon_all.mean(dim=1)
scores = self.config.fuse_alpha * s_ribbon + (1 - self.config.fuse_alpha) * s_disp
diagnostics = {
'dispatcher_scores': s_disp.detach() if self.num_heads == 1 else s_disp.reshape(BT * H, -1).detach(),
'ribbon_scores': s_ribbon.detach() if self.num_heads == 1 else s_ribbon.reshape(BT * H, -1).detach()
}
return {'scores': scores, 'diagnostics': diagnostics}
class GeometricAttention(nn.Module):
"""Multi-head geometric attention with Q-K alignment - FIXED with proper device handling."""
def __init__(self, dim: int, num_heads: int = 8, num_regions: Optional[int] = None,
config: Optional[GeometricConfig] = None, dropout: float = 0.0, device=None):
super().__init__()
self.dim = dim
self.num_heads = num_heads
self.head_dim = dim // num_heads
if num_regions is None:
num_regions = min(self.head_dim, 16)
if config is None:
config = GeometricConfig()
self.config = config
self.to_qkv = nn.Linear(dim, dim * 3, bias=False)
# Create batched navigators with device
self.q_navigator = GeometricNavigator(self.head_dim, num_regions, config, num_heads=num_heads, device=device)
self.k_navigator = GeometricNavigator(self.head_dim, num_regions, config, num_heads=num_heads, device=device)
self.out_proj = nn.Linear(dim, dim)
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor, mask: Optional[torch.Tensor] = None,
return_diagnostics: bool = False) -> Tuple[torch.Tensor, Optional[Dict]]:
B, T, D = x.shape
qkv = self.to_qkv(x)
q, k, v = qkv.chunk(3, dim=-1)
q = q.reshape(B, T, self.num_heads, self.head_dim).transpose(1, 2)
k = k.reshape(B, T, self.num_heads, self.head_dim).transpose(1, 2)
v = v.reshape(B, T, self.num_heads, self.head_dim).transpose(1, 2)
# Prepare for batched navigation
q_batched = q.transpose(1, 2).reshape(B * T, self.num_heads, self.head_dim)
k_batched = k.transpose(1, 2).reshape(B * T, self.num_heads, self.head_dim)
# Navigate all heads at once
q_nav = self.q_navigator.navigate(q_batched)
k_nav = self.k_navigator.navigate(k_batched)
# Reshape scores back
q_scores = q_nav['scores'].reshape(B, T, self.num_heads, -1).transpose(1, 2)
k_scores = k_nav['scores'].reshape(B, T, self.num_heads, -1).transpose(1, 2)
# OPTIMIZED: Compute attention for all heads at once using einsum
scale = math.sqrt(q_scores.size(-1))
attn = torch.einsum('bhqr,bhkr->bhqk', q_scores, k_scores) / scale
if mask is not None:
mask_expanded = mask.unsqueeze(1).unsqueeze(2)
attn = attn.masked_fill(mask_expanded == 0, -1e9)
attn = F.softmax(attn, dim=-1)
attn = self.dropout(attn)
# Apply attention to values
out = torch.einsum('bhqk,bhkd->bhqd', attn, v)
out = out.transpose(1, 2).reshape(B, T, D)
output = self.out_proj(out)
output = self.dropout(output)
if return_diagnostics:
return output, {'q_diagnostics': q_nav['diagnostics'], 'k_diagnostics': k_nav['diagnostics']}
return output, None
# ============================================
# DROP PATH (STOCHASTIC DEPTH)
# ============================================
class DropPath(nn.Module):
"""Drop paths (Stochastic Depth) per sample."""
def __init__(self, drop_prob: float = 0.):
super().__init__()
self.drop_prob = drop_prob
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.drop_prob == 0. or not self.training:
return x
keep_prob = 1 - self.drop_prob
shape = (x.shape[0],) + (1,) * (x.ndim - 1)
random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
random_tensor.floor_()
output = x.div(keep_prob) * random_tensor
return output
# ============================================
# HIERARCHICAL CLS WITH PENTACHORA
# ============================================
class HierarchicalPentachoronCLS(nn.Module):
"""
Hierarchical CLS structure with pentachoron geometry.
Uses vocabulary embeddings for CLS tokens.
"""
def __init__(self, dim: int, vocab_dim: int, num_classes: int = 100):
super().__init__()
self.dim = dim
self.vocab_dim = vocab_dim
self.num_classes = num_classes
# Class-specific pentachora from vocabulary
self.register_buffer('class_pentachora', torch.randn(num_classes, 5, vocab_dim) * 0.02)
# Projection from vocabulary dimension to model dimension
if vocab_dim != dim:
self.vocab_to_model = nn.Linear(vocab_dim, dim)
else:
self.vocab_to_model = nn.Identity()
# Learnable aggregation weights
self.vertex_weights = nn.Parameter(torch.ones(5) / 5)
# Optional learnable offset
self.global_offset = nn.Parameter(torch.zeros(1, 1, dim))
# Layer norms
self.vertex_norm = nn.LayerNorm(dim)
self.global_norm = nn.LayerNorm(dim)
def forward(self, batch_size: int, class_indices: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor]:
"""Generate CLS tokens for batch."""
# Get class-specific pentachora
class_pentachora = self.class_pentachora # This is now a computed property
if class_indices is not None and class_indices.shape[0] == batch_size:
vertex_cls_vocab = class_pentachora[class_indices]
else:
vertex_cls_vocab = class_pentachora.mean(dim=0, keepdim=True)
vertex_cls_vocab = vertex_cls_vocab.expand(batch_size, -1, -1)
# Project from vocabulary dimension to model dimension
vertex_cls = self.vocab_to_model(vertex_cls_vocab)
vertex_cls = self.vertex_norm(vertex_cls)
# Create global CLS as weighted combination
weights = F.softmax(self.vertex_weights, dim=0)
global_cls = torch.einsum('bvd,v->bd', vertex_cls, weights).unsqueeze(1)
global_cls = global_cls + self.global_offset
global_cls = self.global_norm(global_cls)
return global_cls, vertex_cls
def get_class_prototypes(self) -> torch.Tensor:
"""Get class prototypes in model dimension."""
class_pentachora = self.class_pentachora # Get computed pentachora
pentachora_model = self.vocab_to_model(class_pentachora)
weights = F.softmax(self.vertex_weights, dim=0)
prototypes = torch.einsum('cvd,v->cd', pentachora_model, weights)
return prototypes
# ============================================
# GEOMETRIC PROJECTION LAYER
# ============================================
class GeometricProjection(nn.Module):
"""Project patches onto pentachoron geometry."""
def __init__(self, dim: int, vocab_dim: int, num_classes: int = 100, dropout: float = 0.1):
super().__init__()
self.dim = dim
self.vocab_dim = vocab_dim
self.num_classes = num_classes
# Projection from model dim to vocab dim
self.to_vocab_space = nn.Linear(dim, vocab_dim)
# Vertex-specific projections
self.vertex_projections = nn.ModuleList([
nn.Linear(vocab_dim, vocab_dim, bias=False) for _ in range(5)
])
# Temperature for alignment scores
self.temperature = nn.Parameter(torch.ones(1))
self.norm = nn.LayerNorm(dim)
self.dropout = nn.Dropout(dropout)
def forward(self, patches: torch.Tensor, pentachora: torch.Tensor) -> torch.Tensor:
"""Compute alignment between patches and class pentachora."""
B, N, D = patches.shape
C = pentachora.shape[0]
# Normalize patches
patches = self.norm(patches)
# Project patches to vocabulary space
patches_vocab = self.to_vocab_space(patches)
patches_vocab = F.normalize(patches_vocab, dim=-1)
# Compute alignment with each vertex
alignments = []
for v in range(5):
patches_v = self.vertex_projections[v](patches_vocab)
patches_v = F.normalize(patches_v, dim=-1)
vertex_v = F.normalize(pentachora[:, v, :], dim=-1)
alignment = torch.matmul(patches_v, vertex_v.T) / self.temperature
alignments.append(alignment)
# Average alignments across vertices
alignments = torch.stack(alignments, dim=-1).mean(dim=-1)
return self.dropout(alignments)
# ============================================
# MLP BLOCK
# ============================================
class MLP(nn.Module):
"""MLP block with GELU activation."""
def __init__(self, in_features: int, hidden_features: Optional[int] = None,
out_features: Optional[int] = None, dropout: float = 0.):
super().__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = nn.GELU()
self.drop1 = nn.Dropout(dropout)
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop2 = nn.Dropout(dropout)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.fc1(x)
x = self.act(x)
x = self.drop1(x)
x = self.fc2(x)
x = self.drop2(x)
return x
# ============================================
# VIT BLOCK WITH GEOMETRIC ATTENTION
# ============================================
class PentachoronViTBlock(nn.Module):
"""ViT block with geometric attention for structured layers."""
def __init__(self, dim: int, heads: int = 8, mlp_ratio: float = 4.0,
use_mesh: bool = True, dropout: float = 0., attn_dropout: float = 0.,
drop_path: float = 0., device=None):
super().__init__()
self.norm1 = nn.LayerNorm(dim)
# Use GeometricAttention for structured layers, standard for others
if use_mesh:
self.attn = GeometricAttention(
dim=dim,
num_heads=heads,
num_regions=min(dim // heads, 16),
config=GeometricConfig(),
dropout=attn_dropout,
device=device
)
else:
# Standard multi-head attention for later layers
self.attn = nn.MultiheadAttention(dim, heads, dropout=attn_dropout, batch_first=True)
self.use_mesh = use_mesh
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = nn.LayerNorm(dim)
mlp_hidden = int(dim * mlp_ratio)
self.mlp = MLP(in_features=dim, hidden_features=mlp_hidden, dropout=dropout)
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def forward(self, x: torch.Tensor, preserve_structure: bool = True) -> torch.Tensor:
if self.use_mesh:
# GeometricAttention
attn_out, _ = self.attn(self.norm1(x))
x = x + self.drop_path1(attn_out)
else:
# Standard attention
normalized = self.norm1(x)
attn_out, _ = self.attn(normalized, normalized, normalized)
x = x + self.drop_path1(attn_out)
x = x + self.drop_path2(self.mlp(self.norm2(x)))
return x
# ============================================
# PATCH EMBEDDING
# ============================================
class PatchEmbed(nn.Module):
"""2D Image to Patch Embedding."""
def __init__(self, img_size: int = 32, patch_size: int = 4,
in_chans: int = 3, embed_dim: int = 512):
super().__init__()
self.img_size = img_size
self.patch_size = patch_size
self.num_patches = (img_size // patch_size) ** 2
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
self.norm = nn.LayerNorm(embed_dim)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.proj(x)
x = rearrange(x, 'b c h w -> b (h w) c')
x = self.norm(x)
return x
# ============================================
# PENTACHORA VISION TRANSFORMER
# ============================================
class PentachoraViT(nn.Module):
"""
Vision Transformer with pentachoron-based hierarchical CLS tokens
and geometric vocabulary integration.
"""
def __init__(self, config: Optional[PentachoraConfig] = None, **kwargs):
super().__init__()
# Use config or kwargs
if config is not None:
cfg = config
else:
cfg = PentachoraConfig(**kwargs)
self.config = cfg
self.num_classes = cfg.num_classes
self.dim = cfg.dim
self.depth = cfg.depth
self.preserve_structure_until_layer = cfg.preserve_structure_until_layer
# Set vocabulary dimension
if cfg.vocab_dim is not None:
self.vocab_dim = cfg.vocab_dim
elif 'vocab_dim' in kwargs:
self.vocab_dim = kwargs['vocab_dim']
else:
self.vocab_dim = cfg.dim
# Patch embedding
self.patch_embed = PatchEmbed(
cfg.img_size, cfg.patch_size, 3, cfg.dim
)
num_patches = self.patch_embed.num_patches
# Positional embedding
self.pos_embed = nn.Parameter(torch.randn(1, num_patches, cfg.dim) * 0.02)
self.pos_drop = nn.Dropout(cfg.dropout_rate)
# CLS tokens with pentachoron structure
self.cls_tokens = HierarchicalPentachoronCLS(cfg.dim, self.vocab_dim, cfg.num_classes)
# Geometric projection layer
self.geometric_proj = GeometricProjection(cfg.dim, self.vocab_dim, cfg.num_classes, cfg.dropout_rate)
# Initialize from vocabulary if provided
if cfg.vocab is not None:
self._init_from_vocab(cfg.vocab)
# Stochastic depth decay rule
dpr = [x.item() for x in torch.linspace(0, cfg.drop_path_rate, cfg.depth)]
# Transformer blocks with geometric attention
self.blocks = nn.ModuleList([
PentachoronViTBlock(
dim=cfg.dim,
heads=cfg.heads,
mlp_ratio=cfg.mlp_ratio,
use_mesh=(cfg.use_mesh_attention and i < cfg.preserve_structure_until_layer),
dropout=cfg.dropout_rate,
attn_dropout=cfg.dropout_rate,
drop_path=dpr[i],
device=torch.device('cpu') # Initialize on CPU, will be moved later
)
for i in range(cfg.depth)
])
# Final norm
self.norm = nn.LayerNorm(cfg.dim)
# Classification heads
self.use_prototype_classifier = True
if self.use_prototype_classifier:
self.head = None
else:
self.head = nn.Linear(cfg.dim, cfg.num_classes)
# Auxiliary head for vertex tokens
self.head_aux = nn.Linear(cfg.dim * 5, cfg.num_classes)
# Initialize weights
self.apply(self._init_weights)
def _init_weights(self, m: nn.Module):
"""Initialize model weights."""
if isinstance(m, nn.Linear):
nn.init.trunc_normal_(m.weight, std=0.02)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
elif isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
if m.bias is not None:
nn.init.constant_(m.bias, 0)
def _init_from_vocab(self, vocab):
"""Initialize class pentachora from geometric vocabulary."""
try:
print("Initializing pentachora from vocabulary...")
if not hasattr(vocab, 'encode_batch'):
print("Vocabulary provided but encode_batch method not found, using random initialization")
return
# Get CIFAR-100 class names
class_names = self._get_cifar100_classes()
# Generate pentachora for all classes
pentachora_list = vocab.encode_batch(class_names[:self.num_classes], generate=True)
pentachora = np.stack(pentachora_list, axis=0)
# Get actual dimensions from the encoded data
actual_vocab_dim = pentachora.shape[-1]
print(f"Encoded pentachora shape: {pentachora.shape}")
print(f"Detected vocabulary dimension: {actual_vocab_dim}")
# Validate basic shape requirements
if pentachora.shape[0] != self.num_classes or pentachora.shape[1] != 5:
print(f"Invalid shape: expected ({self.num_classes}, 5, ?), got {pentachora.shape}")
print("Using random initialization")
return
# Update vocabulary dimension
self.vocab_dim = actual_vocab_dim
self.cls_tokens.vocab_dim = actual_vocab_dim
self.geometric_proj.vocab_dim = actual_vocab_dim
# Replace class_pentachora with the loaded vocabulary
self.cls_tokens.class_pentachora = torch.tensor(pentachora, dtype=torch.float32)
# Update/create projection layer if dimensions differ
if actual_vocab_dim != self.dim:
self.cls_tokens.vocab_to_model = nn.Linear(actual_vocab_dim, self.dim)
else:
self.cls_tokens.vocab_to_model = nn.Identity()
# Rebuild geometric projection components
self.geometric_proj.to_vocab_space = nn.Linear(self.dim, actual_vocab_dim)
self.geometric_proj.vertex_projections = nn.ModuleList([
nn.Linear(actual_vocab_dim, actual_vocab_dim, bias=False) for _ in range(5)
])
# Re-initialize the new layers
nn.init.xavier_uniform_(self.geometric_proj.to_vocab_space.weight)
for proj in self.geometric_proj.vertex_projections:
nn.init.xavier_uniform_(proj.weight)
if actual_vocab_dim != self.dim:
nn.init.xavier_uniform_(self.cls_tokens.vocab_to_model.weight)
print(f"✓ Successfully initialized {self.num_classes} class pentachora from vocabulary")
print(f" Vocabulary dimension: {actual_vocab_dim}")
print(f" Model internal dimension: {self.dim}")
except Exception as e:
print(f"Error initializing from vocabulary: {e}")
print("Using random initialization")
def _get_cifar100_classes(self):
"""Get CIFAR-100 class names."""
return [
'apple', 'aquarium_fish', 'baby', 'bear', 'beaver', 'bed', 'bee', 'beetle',
'bicycle', 'bottle', 'bowl', 'boy', 'bridge', 'bus', 'butterfly', 'camel',
'can', 'castle', 'caterpillar', 'cattle', 'chair', 'chimpanzee', 'clock',
'cloud', 'cockroach', 'couch', 'crab', 'crocodile', 'cup', 'dinosaur',
'dolphin', 'elephant', 'flatfish', 'forest', 'fox', 'girl', 'hamster',
'house', 'kangaroo', 'keyboard', 'lamp', 'lawn_mower', 'leopard', 'lion',
'lizard', 'lobster', 'man', 'maple_tree', 'motorcycle', 'mountain', 'mouse',
'mushroom', 'oak_tree', 'orange', 'orchid', 'otter', 'palm_tree', 'pear',
'pickup_truck', 'pine_tree', 'plain', 'plate', 'poppy', 'porcupine',
'possum', 'rabbit', 'raccoon', 'ray', 'road', 'rocket', 'rose',
'sea', 'seal', 'shark', 'shrew', 'skunk', 'skyscraper', 'snail', 'snake',
'spider', 'squirrel', 'streetcar', 'sunflower', 'sweet_pepper', 'table',
'tank', 'telephone', 'television', 'tiger', 'tractor', 'train', 'trout',
'tulip', 'turtle', 'wardrobe', 'whale', 'willow_tree', 'wolf', 'woman', 'worm'
]
def forward_features(self, x: torch.Tensor, class_indices: Optional[torch.Tensor] = None) -> Dict[str, torch.Tensor]:
"""Extract features from input."""
B = x.shape[0]
# Patch embedding
x = self.patch_embed(x)
x = x + self.pos_embed
x = self.pos_drop(x)
# Get hierarchical CLS tokens
global_cls, vertex_cls = self.cls_tokens(B, class_indices)
# Concatenate CLS tokens with patches
x = torch.cat([global_cls, vertex_cls, x], dim=1)
# Apply transformer blocks
for i, block in enumerate(self.blocks):
preserve = i < self.preserve_structure_until_layer
x = block(x, preserve_structure=preserve)
# Apply final norm
x = self.norm(x)
# Split tokens
global_cls = x[:, 0]
vertex_cls = x[:, 1:6]
patches = x[:, 6:]
return {
'global_cls': global_cls,
'vertex_cls': vertex_cls,
'patches': patches
}
def forward(self, x: torch.Tensor, targets: Optional[torch.Tensor] = None) -> Dict[str, torch.Tensor]:
"""Forward pass through the model."""
# During training, use target labels for class-specific CLS initialization
class_indices = targets if self.training and targets is not None else None
features = self.forward_features(x, class_indices)
# Primary classification using prototype matching
if self.use_prototype_classifier:
prototypes = self.cls_tokens.get_class_prototypes()
prototypes = F.normalize(prototypes, dim=-1)
global_cls_norm = F.normalize(features['global_cls'], dim=-1)
logits = torch.matmul(global_cls_norm, prototypes.T) * 20.0
else:
logits = self.head(features['global_cls'])
# Auxiliary classification using vertex tokens
B = features['vertex_cls'].shape[0]
vertex_flat = features['vertex_cls'].reshape(B, -1)
aux_logits = self.head_aux(vertex_flat)
# Geometric alignment scores
geometric_alignments = self.geometric_proj(
features['patches'],
self.cls_tokens.class_pentachora
)
return {
'logits': logits,
'aux_logits': aux_logits,
'geometric_alignments': geometric_alignments,
'vertex_cls': features['vertex_cls'],
'global_cls': features['global_cls'],
'patches': features['patches']
}
# ============================================
# LOSS FUNCTIONS
# ============================================
class PentachoraLoss(nn.Module):
"""Combined loss for PentachoraViT training."""
def __init__(self, aux_weight: float = 0.3, geo_weight: float = 0.1,
smoothing: float = 0.0):
super().__init__()
self.aux_weight = aux_weight
self.geo_weight = geo_weight
self.criterion = nn.CrossEntropyLoss(label_smoothing=smoothing)
def forward(self, outputs: Dict[str, torch.Tensor], targets: torch.Tensor) -> torch.Tensor:
"""Compute combined loss."""
# Primary classification loss
loss = self.criterion(outputs['logits'], targets)
# Auxiliary loss from vertex tokens
if 'aux_logits' in outputs and self.aux_weight > 0:
aux_loss = self.criterion(outputs['aux_logits'], targets)
loss = loss + self.aux_weight * aux_loss
# Geometric alignment loss
if 'geometric_alignments' in outputs and self.geo_weight > 0:
geo_logits = outputs['geometric_alignments'].mean(dim=1)
geo_loss = self.criterion(geo_logits, targets)
loss = loss + self.geo_weight * geo_loss
return loss
# ============================================
# MODEL REGISTRY AND BUILDERS
# ============================================
MODEL_CONFIGS = {
'pentachora_spark_xs': PentachoraConfig(
dim=100, depth=2, heads=10, mlp_ratio=4.0,
preserve_structure_until_layer=1,
dropout_rate=0.0, drop_path_rate=0.0
),
'pentachora_spark': PentachoraConfig(
dim=100, depth=5, heads=4, mlp_ratio=4.0,
preserve_structure_until_layer=1,
dropout_rate=0.0, drop_path_rate=0.0
),
'pentachora_shock': PentachoraConfig(
dim=100, depth=10, heads=5, mlp_ratio=4.0,
patch_size=5, preserve_structure_until_layer=5,
dropout_rate=0.0, drop_path_rate=0.0
),
'pentachora_shock_xs_32d': PentachoraConfig(
dim=32, depth=2, heads=8, mlp_ratio=4.0,
preserve_structure_until_layer=4,
dropout_rate=0.0, drop_path_rate=0.0
),
'pentachora_shock_xs_64d': PentachoraConfig(
dim=64, depth=2, heads=8, mlp_ratio=1.0,
preserve_structure_until_layer=4,
dropout_rate=0.0, drop_path_rate=0.0
),
'pentachora_shock_xs_128d': PentachoraConfig(
dim=128, depth=2, heads=8, mlp_ratio=2.0,
preserve_structure_until_layer=4,
vocab_dim=256,
dropout_rate=0.0, drop_path_rate=0.0
),
'vit_pixie_256_patch4': PentachoraConfig(
dim=256, depth=10, heads=16, mlp_ratio=1.0,
preserve_structure_until_layer=10,
vocab_dim=256, patch_size=4,
dropout_rate=0.0, drop_path_rate=0.0
),
'vit_pixie_256_patch2': PentachoraConfig(
dim=256, depth=10, heads=16, mlp_ratio=1.0,
preserve_structure_until_layer=10,
vocab_dim=256, patch_size=2,
dropout_rate=0.0, drop_path_rate=0.0
),
'pentachora_shock_xs_256d': PentachoraConfig(
dim=256, depth=2, heads=8, mlp_ratio=4.0,
preserve_structure_until_layer=4,
vocab_dim=128,
dropout_rate=0.0, drop_path_rate=0.0
),
'pentachora_shock_xs_512d': PentachoraConfig(
dim=512, depth=2, heads=8, mlp_ratio=4.0,
preserve_structure_until_layer=4,
dropout_rate=0.0, drop_path_rate=0.0
),
'pentachora_tiny': PentachoraConfig(
dim=384, depth=12, heads=6, mlp_ratio=4.0,
preserve_structure_until_layer=6,
dropout_rate=0.1, drop_path_rate=0.1
),
'pentachora_small': PentachoraConfig(
dim=512, depth=12, heads=8, mlp_ratio=4.0,
preserve_structure_until_layer=6,
dropout_rate=0.1, drop_path_rate=0.1
),
'pentachora_base': PentachoraConfig(
dim=768, depth=12, heads=12, mlp_ratio=4.0,
preserve_structure_until_layer=8,
dropout_rate=0.1, drop_path_rate=0.2
),
'pentachora_large': PentachoraConfig(
dim=1024, depth=24, heads=16, mlp_ratio=4.0,
preserve_structure_until_layer=12,
dropout_rate=0.1, drop_path_rate=0.3
),
}
def create_pentachora_vit(variant: str = 'pentachora_small',
pretrained: bool = False,
**kwargs) -> PentachoraViT:
"""Create PentachoraViT model."""
if variant not in MODEL_CONFIGS:
raise ValueError(f"Unknown variant: {variant}. Choose from {list(MODEL_CONFIGS.keys())}")
config = MODEL_CONFIGS[variant]
# Override config with kwargs
for key, value in kwargs.items():
setattr(config, key, value)
model = PentachoraViT(config)
if pretrained:
warnings.warn("Pretrained weights not available yet")
return model
# Convenience functions for each variant
def pentachora_vit_spark_tiny(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create spark variant (smallest)."""
return create_pentachora_vit('pentachora_spark_xs', pretrained=pretrained, **kwargs)
def pentachora_shock_xs_64d(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create shock xs 64d variant."""
return create_pentachora_vit('pentachora_shock_xs_64d', pretrained=pretrained, **kwargs)
def pentachora_vit_spark(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create spark variant."""
return create_pentachora_vit('pentachora_spark', pretrained=pretrained, **kwargs)
def pentachora_shock_xs_32d(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create shock xs 32d variant."""
return create_pentachora_vit('pentachora_shock_xs_32d', pretrained=pretrained, **kwargs)
def pentachora_shock_xs_256d(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create shock xs 256d variant."""
return create_pentachora_vit('pentachora_shock_xs_256d', pretrained=pretrained, **kwargs)
def pentachora_shock_xs_512d(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create shock xs 512d variant."""
return create_pentachora_vit('pentachora_shock_xs_512d', pretrained=pretrained, **kwargs)
def pentachora_vit_shock(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create shock variant."""
return create_pentachora_vit('pentachora_shock', pretrained=pretrained, **kwargs)
def pentachora_vit_tiny(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create tiny variant."""
return create_pentachora_vit('pentachora_tiny', pretrained=pretrained, **kwargs)
def pentachora_vit_small(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create small variant."""
return create_pentachora_vit('pentachora_small', pretrained=pretrained, **kwargs)
def pentachora_vit_base(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create base variant."""
return create_pentachora_vit('pentachora_base', pretrained=pretrained, **kwargs)
def pentachora_vit_large(pretrained: bool = False, **kwargs) -> PentachoraViT:
"""Create large variant."""
return create_pentachora_vit('pentachora_large', pretrained=pretrained, **kwargs)
# ============================================
# TRAINING UTILITIES
# ============================================
def get_parameter_groups(model: PentachoraViT,
weight_decay: float = 0.05) -> List[Dict[str, Any]]:
"""Get parameter groups for optimizer with weight decay handling."""
no_decay = ['bias', 'norm', 'LayerNorm']
decay_params = []
no_decay_params = []
for name, param in model.named_parameters():
if not param.requires_grad:
continue
if any(nd in name for nd in no_decay):
no_decay_params.append(param)
else:
decay_params.append(param)
return [
{'params': decay_params, 'weight_decay': weight_decay},
{'params': no_decay_params, 'weight_decay': 0.0}
]
def count_parameters(model: nn.Module) -> Dict[str, int]:
"""Count model parameters."""
total = sum(p.numel() for p in model.parameters())
trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
return {
'total': total,
'trainable': trainable,
'non_trainable': total - trainable
}
# ============================================
# INFERENCE UTILITIES
# ============================================
@torch.no_grad()
def extract_features(model: PentachoraViT,
images: torch.Tensor,
feature_type: str = 'global_cls') -> torch.Tensor:
"""Extract features from images using the model."""
model.eval()
features = model.forward_features(images)
return features.get(feature_type, features['global_cls'])
# ============================================
# EXAMPLE USAGE AND TESTING
# ============================================
def test_model():
"""Test model creation and forward pass."""
print("Testing Fixed PentachoraViT Model")
print("=" * 50)
# Test different variants
variants = ['pentachora_spark', 'pentachora_shock_xs_256d', 'pentachora_small']
for variant in variants:
print(f"\nTesting {variant}:")
# Create model with vocab_dim
model = create_pentachora_vit(
variant=variant,
img_size=32,
patch_size=4,
num_classes=100,
vocab_dim=64
)
# Count parameters
params = count_parameters(model)
print(f" Total parameters: {params['total']:,}")
print(f" Trainable parameters: {params['trainable']:,}")
# Test forward pass
x = torch.randn(2, 3, 32, 32)
# Time the forward pass
if torch.cuda.is_available():
model = model.cuda()
x = x.cuda()
torch.cuda.synchronize()
import time
start = time.time()
outputs = model(x)
if torch.cuda.is_available():
torch.cuda.synchronize()
end = time.time()
print(f" Output shapes:")
print(f" Logits: {outputs['logits'].shape}")
print(f" Aux logits: {outputs['aux_logits'].shape}")
print(f" Geometric alignments: {outputs['geometric_alignments'].shape}")
print(f" Forward pass time: {(end - start)*1000:.2f}ms")
# Test loss computation
loss_fn = PentachoraLoss()
targets = torch.randint(0, 100, (2,))
if torch.cuda.is_available():
targets = targets.cuda()
loss = loss_fn(outputs, targets)
print(f" Loss: {loss.item():.4f}")
print("\n" + "=" * 50)
print("All tests passed!")
if __name__ == "__main__":
# Run tests
test_model()
# Example: Create model for training
print("\nExample: Creating model with proper initialization")
model = pentachora_shock_xs_256d(
img_size=32,
num_classes=100,
vocab_dim=100,
dropout_rate=0.0,
drop_path_rate=0.0
)
# All parameters are initialized immediately
print(f"Model has {count_parameters(model)['total']:,} parameters")
print("All geometric parameters initialized at creation time")
# Move model to CUDA if available
if torch.cuda.is_available():
model = model.cuda()
print("Model moved to CUDA")
# Now torch.compile should work without issues
if hasattr(torch, 'compile'):
print("Compiling model with torch.compile...")
try:
model = torch.compile(model)
print("✓ Model compiled successfully")
except Exception as e:
print(f"Compilation warning: {e}")
print("Continuing without compilation")
print("\nModel ready for training with all parameters properly initialized!") |