File size: 40,797 Bytes
1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 b6d45f7 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd b6d45f7 1363636 8d7f7cd b6d45f7 1363636 b6d45f7 1363636 b6d45f7 8d7f7cd 1363636 b6d45f7 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 8d7f7cd 1363636 | 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 | #!/usr/bin/env python3
"""
GeoLIP Tri-Stream ViT v8 β Geometric Arbitration (fixed)
==========================================================
v7βv8 changes:
1. Uniform hypersphere orthogonal init for GAL anchors + constellation
2. Gate init at 1/(2*n_blocks) β geometry enters immediately
3. InfoNCE on emb_b (Stream B survives through contrastive, not BCE)
4. InfoNCE weight on geo_emb raised β geo was starved
5. No residual scaling (per Phil)
6. GAL update interval + lr controlled from trainer
Three processing paths:
Stream A (CE loss): self-attn + FFN, standard cross-entropy
Stream B (BCE+NCE): self-attn + FFN, binary CE + InfoNCE
GAL (geometric): KSimplex features, accumulated over time,
provides cross-attention to shared anchors
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from itertools import combinations
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# UNIFORM HYPERSPHERE INIT
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def uniform_hypersphere_init(n, d):
"""
Generate n points with maximal spread on the d-dimensional unit sphere.
n <= d: orthogonal columns via QR decomposition (perfect spread).
n > d: QR orthogonal basis + iterative repulsion for the rest.
Returns: (n, d) tensor on the unit sphere.
"""
if n <= d:
# Perfect orthogonal set
M = torch.randn(d, n)
Q, _ = torch.linalg.qr(M)
return Q.T.contiguous() # (n, d), each row unit-norm & orthogonal
else:
# Start with d orthogonal vectors, fill remainder
M = torch.randn(d, d)
Q, _ = torch.linalg.qr(M)
basis = Q.T # (d, d)
extra = torch.randn(n - d, d)
extra = F.normalize(extra, dim=-1)
vecs = torch.cat([basis, extra], dim=0) # (n, d)
# Iterative repulsion β push points apart on sphere
for _ in range(200):
sim = vecs @ vecs.T
sim.fill_diagonal_(-2.0) # ignore self
# Find nearest neighbor for each point
nn_idx = sim.argmax(dim=1)
nn_vec = vecs[nn_idx]
# Repel from nearest neighbor
vecs = F.normalize(vecs - 0.05 * nn_vec, dim=-1)
return vecs
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# CAYLEY-MENGER + KSIMPLEX (unchanged)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class CMValidator(nn.Module):
def __init__(self, k):
super().__init__()
self._k = k
self._nv = k + 1
pairs = list(combinations(range(self._nv), 2))
self._npairs = len(pairs)
self.register_buffer('_pi', torch.tensor([p[0] for p in pairs], dtype=torch.long))
self.register_buffer('_pj', torch.tensor([p[1] for p in pairs], dtype=torch.long))
sign = (-1.0) ** (k + 1)
fact = math.factorial(k)
self._prefactor = sign / ((2.0 ** k) * (fact ** 2))
def forward(self, verts):
gram = torch.einsum('...ve,...we->...vw', verts, verts)
norms = torch.diagonal(gram, dim1=-2, dim2=-1)
d2_mat = norms.unsqueeze(-1) + norms.unsqueeze(-2) - 2 * gram
d2_mat = F.relu(d2_mat)
d2_pairs = d2_mat[..., self._pi, self._pj]
shape = d2_mat.shape[:-2]
V = d2_mat.shape[-1]
cm = torch.zeros(*shape, V + 1, V + 1,
device=d2_mat.device, dtype=d2_mat.dtype)
cm[..., 0, 1:] = 1.0; cm[..., 1:, 0] = 1.0
cm[..., 1:, 1:] = d2_mat
vol2 = self._prefactor * torch.linalg.det(cm.float())
vol2 = vol2.to(d2_pairs.dtype)
return d2_pairs, vol2
class KSimplexChannel(nn.Module):
BASE_DEFORM = 0.05
def __init__(self, k, in_dim, edim):
super().__init__()
self._k = k; self._nv = k + 1; self._edim = edim
self._cm = CMValidator(k)
self._out_dim = self._cm._npairs + 1
template = self._make_regular_simplex(k, edim)
self.register_buffer('_template', template)
self._to_deform = nn.Linear(in_dim, self._nv * edim)
self._norm = nn.LayerNorm(self._out_dim)
@staticmethod
def _make_regular_simplex(k, edim):
nv = k + 1
verts = torch.zeros(nv, edim)
for i in range(min(nv, edim)):
verts[i, i] = 1.0
if nv > edim:
for i in range(edim, nv):
v = torch.randn(edim)
verts[i] = v / (v.norm() + 1e-8)
verts = verts - verts.mean(dim=0, keepdim=True)
edge_len = (verts[0] - verts[1]).norm().clamp(min=1e-8)
return verts / edge_len
@property
def out_dim(self):
return self._out_dim
def forward(self, x):
deform = self._to_deform(x).unflatten(-1, (self._nv, self._edim))
verts = self._template + self.BASE_DEFORM * deform
d2, vol2 = self._cm(verts)
geo = torch.cat([d2, vol2.unsqueeze(-1)], dim=-1)
return self._norm(geo), vol2
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# CONSTELLATION + PATCHWORK
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class Constellation(nn.Module):
def __init__(self, n_anchors, dim, anchor_drop=0.0):
super().__init__()
# ββ v8: uniform hypersphere init ββ
init_vecs = uniform_hypersphere_init(n_anchors, dim)
self.anchors = nn.Parameter(init_vecs)
self.anchor_drop = anchor_drop
# Diagnostic
with torch.no_grad():
an = F.normalize(init_vecs, dim=-1)
sim = an @ an.T
mask = ~torch.eye(n_anchors, dtype=torch.bool)
off = sim[mask]
print(f" β Constellation: {n_anchors}Γ{dim} uniform hypersphere")
print(f" pairwise cos: mean={off.mean():.4f} max={off.max():.4f}")
def triangulate(self, emb, training=False):
anchors = F.normalize(self.anchors, dim=-1)
if training and self.anchor_drop > 0:
mask = torch.rand(anchors.shape[0], device=anchors.device) > self.anchor_drop
if mask.sum() < 2: mask[:2] = True
anchors = anchors[mask]
cos = emb @ anchors.T
tri = 1.0 - cos
_, nearest_local = cos.max(dim=-1)
nearest = mask.nonzero(as_tuple=True)[0][nearest_local]
else:
cos = emb @ anchors.T
tri = 1.0 - cos
_, nearest = cos.max(dim=-1)
return tri, nearest
class Patchwork(nn.Module):
def __init__(self, n_anchors, n_comp, d_comp):
super().__init__()
self.n_comp = n_comp; self.d_comp = d_comp
self.register_buffer('asgn', torch.arange(n_anchors) % n_comp)
anchors_per = n_anchors // n_comp
self.comps = nn.ModuleList([nn.Sequential(
nn.Linear(anchors_per, d_comp * 2), nn.GELU(),
nn.Linear(d_comp * 2, d_comp), nn.LayerNorm(d_comp))
for _ in range(n_comp)])
def forward(self, tri):
return torch.cat([self.comps[k](tri[:, self.asgn == k])
for k in range(self.n_comp)], -1)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# EMBEDDING AUTOGRAD (unchanged)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class EmbeddingAutograd(torch.autograd.Function):
@staticmethod
def forward(ctx, x, embedding, anchors, tang, sep):
ctx.save_for_backward(embedding, anchors)
ctx.tang = tang; ctx.sep = sep
return x
@staticmethod
def backward(ctx, grad_output):
embedding, anchors = ctx.saved_tensors
emb_n = F.normalize(embedding.detach().float(), dim=-1)
anchors_n = F.normalize(anchors.detach().float(), dim=-1)
grad_f = grad_output.float()
radial = (grad_f * emb_n).sum(-1, keepdim=True) * emb_n
corrected = (grad_f - radial) + (1.0 - ctx.tang) * radial
if ctx.sep > 0:
cos_to = emb_n @ anchors_n.T
nearest = anchors_n[cos_to.argmax(dim=-1)]
toward = (corrected * nearest).sum(-1, keepdim=True)
corrected = corrected - ctx.sep * (toward > 0).float() * toward * nearest
return corrected.to(grad_output.dtype), None, None, None, None
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# PROCRUSTES ALIGNMENT (unchanged)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def procrustes_align(source, target, whiten=False):
source_c = source.float() - source.float().mean(0, keepdim=True)
target_c = target.float() - target.float().mean(0, keepdim=True)
if whiten:
source_c = source_c / (source_c.std(0, keepdim=True) + 1e-8)
target_c = target_c / (target_c.std(0, keepdim=True) + 1e-8)
M = (source_c.T @ target_c).float()
U, S, Vt = torch.linalg.svd(M)
d = torch.ones(U.shape[0], device=U.device, dtype=U.dtype)
d[-1] = torch.det(U @ Vt).sign()
R = U @ torch.diag(d) @ Vt
return R, S.sum().item()
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# SIMPLEX BUFFER (unchanged)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class SimplexBuffer:
def __init__(self, dim, max_size=50000, device='cuda'):
self.dim = dim; self.max_size = max_size; self.device = device
self._feats = None; self._labels = None
def push(self, feats, labels):
feats = feats.detach().to(self.device)
labels = labels.detach().to(self.device)
if self._feats is None:
self._feats = feats; self._labels = labels
else:
self._feats = torch.cat([self._feats, feats], 0)[-self.max_size:]
self._labels = torch.cat([self._labels, labels], 0)[-self.max_size:]
@property
def size(self):
return 0 if self._feats is None else self._feats.shape[0]
def class_centroids(self, num_classes):
if self._feats is None or self.size < num_classes * 10:
return None
centroids = []
for c in range(num_classes):
mask = self._labels == c
if mask.sum() == 0: return None
centroids.append(self._feats[mask].mean(0))
return torch.stack(centroids)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# GAL β v8: uniform hypersphere anchors
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class GAL(nn.Module):
def __init__(self, stream_dim, n_gal_anchors, n_heads,
ksimplex_k=4, ksimplex_edim=8, dropout=0.1):
super().__init__()
self.stream_dim = stream_dim
self.n_gal_anchors = n_gal_anchors
# ββ v8: uniform hypersphere init for anchors ββ
init_anchors = uniform_hypersphere_init(n_gal_anchors, stream_dim)
self.register_buffer('gal_anchors', init_anchors)
with torch.no_grad():
an = F.normalize(init_anchors, dim=-1)
sim = an @ an.T
mask = ~torch.eye(n_gal_anchors, dtype=torch.bool)
off = sim[mask]
print(f" β GAL anchors: {n_gal_anchors}Γ{stream_dim} "
f"uniform hypersphere")
print(f" pairwise cos: mean={off.mean():.4f} "
f"max={off.max():.4f}")
self.ksimplex = KSimplexChannel(
k=ksimplex_k, in_dim=stream_dim, edim=ksimplex_edim)
self.geo_lift = nn.Sequential(
nn.Linear(self.ksimplex.out_dim, stream_dim), nn.GELU())
self.anchor_proj = nn.Sequential(
nn.Linear(stream_dim, stream_dim), nn.LayerNorm(stream_dim))
@torch.no_grad()
def rotate_anchors(self, rotation_matrix):
self.gal_anchors.copy_(
(self.gal_anchors @ rotation_matrix).contiguous())
def get_anchor_kv(self):
return self.anchor_proj(self.gal_anchors)
class GALBlock(nn.Module):
"""
Per-layer GAL injection with non-zero gate init.
v8: gates start at 1/(2*n_blocks) so geometry enters immediately.
"""
def __init__(self, stream_dim, n_gal_anchors, n_heads,
gate_init=0.055, dropout=0.1):
super().__init__()
self.cross_attn_a = nn.MultiheadAttention(
stream_dim, n_heads, dropout=dropout, batch_first=True)
self.cross_attn_b = nn.MultiheadAttention(
stream_dim, n_heads, dropout=dropout, batch_first=True)
self.norm_ga = nn.LayerNorm(stream_dim)
self.norm_gb = nn.LayerNorm(stream_dim)
self.lift_proj_a = nn.Linear(stream_dim, stream_dim)
self.lift_proj_b = nn.Linear(stream_dim, stream_dim)
# ββ v8: init at small positive value, NOT zero ββ
self.gate_a = nn.Parameter(torch.tensor(gate_init))
self.gate_b = nn.Parameter(torch.tensor(gate_init))
def forward(self, stream_a, stream_b, anchor_kv, geo_lifted):
B = stream_a.shape[0]
kv = anchor_kv.unsqueeze(0).expand(B, -1, -1)
qa = self.norm_ga(stream_a)
ha, _ = self.cross_attn_a(qa, kv, kv, need_weights=False)
qb = self.norm_gb(stream_b)
hb, _ = self.cross_attn_b(qb, kv, kv, need_weights=False)
stream_a = stream_a + self.gate_a * (ha + self.lift_proj_a(geo_lifted))
stream_b = stream_b + self.gate_b * (hb + self.lift_proj_b(geo_lifted))
return stream_a, stream_b
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# TRI-STREAM BLOCK (unchanged structure)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class TriStreamBlock(nn.Module):
def __init__(self, stream_dim, n_gal_anchors, n_heads,
gate_init=0.055, dropout=0.1):
super().__init__()
# Stream A
self.norm_a1 = nn.LayerNorm(stream_dim)
self.attn_a = nn.MultiheadAttention(
stream_dim, n_heads, dropout=dropout, batch_first=True)
self.norm_a2 = nn.LayerNorm(stream_dim)
self.ffn_a = nn.Sequential(
nn.Linear(stream_dim, stream_dim * 4), nn.GELU(),
nn.Dropout(dropout),
nn.Linear(stream_dim * 4, stream_dim), nn.Dropout(dropout))
# Stream B
self.norm_b1 = nn.LayerNorm(stream_dim)
self.attn_b = nn.MultiheadAttention(
stream_dim, n_heads, dropout=dropout, batch_first=True)
self.norm_b2 = nn.LayerNorm(stream_dim)
self.ffn_b = nn.Sequential(
nn.Linear(stream_dim, stream_dim * 4), nn.GELU(),
nn.Dropout(dropout),
nn.Linear(stream_dim * 4, stream_dim), nn.Dropout(dropout))
# GAL block β v8: gate_init passed through
self.gal_block = GALBlock(
stream_dim, n_gal_anchors, n_heads,
gate_init=gate_init, dropout=dropout)
self.geo_combine_norm = nn.LayerNorm(stream_dim)
def forward(self, stream_a, stream_b, gal, anchor_kv):
B, P, D = stream_a.shape
# Stream A
h = self.norm_a1(stream_a)
h, _ = self.attn_a(h, h, h, need_weights=False)
stream_a = stream_a + h
stream_a = stream_a + self.ffn_a(self.norm_a2(stream_a))
# Stream B
h = self.norm_b1(stream_b)
h, _ = self.attn_b(h, h, h, need_weights=False)
stream_b = stream_b + h
stream_b = stream_b + self.ffn_b(self.norm_b2(stream_b))
# GAL
geo_input = self.geo_combine_norm(stream_a + stream_b)
flat = geo_input.reshape(B * P, D)
geo_feats, vol2 = gal.ksimplex(flat)
geo_feats = geo_feats.reshape(B, P, -1)
vol2 = vol2.reshape(B, P)
geo_lifted = gal.geo_lift(geo_feats)
stream_a, stream_b = self.gal_block(
stream_a, stream_b, anchor_kv, geo_lifted)
return stream_a, stream_b, geo_feats, vol2, geo_lifted
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# TRI-STREAM VIT v8
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class TriStreamViT(nn.Module):
def __init__(
self,
num_classes=10,
img_size=32,
patch_size=4,
embed_dim=384,
stream_dim=192,
n_blocks=9,
n_heads=8,
output_dim=256,
n_anchors=128,
n_gal_anchors=64,
n_comp=16,
d_comp=128,
anchor_drop=0.10,
cv_target=0.22,
ksimplex_k=4,
ksimplex_edim=8,
dropout=0.1,
infonce_temp=0.07,
infonce_weight=0.1,
bce_weight=1.0,
cm_weight=0.1,
cv_weight=0.1,
autograd_tang=1.0,
autograd_sep=0.1,
enable_autograd=True,
label_smoothing=0.1,
# ββ v8: stream B + geo InfoNCE weights (separate) ββ
stream_b_nce_weight=0.5,
geo_nce_weight=0.5,
):
super().__init__()
self.num_classes = num_classes
self.num_patches = (img_size // patch_size) ** 2
self.stream_dim = stream_dim
self.output_dim = output_dim
self.cv_target = cv_target
self.infonce_temp = infonce_temp
self.infonce_weight = infonce_weight
self.bce_weight = bce_weight
self.cm_weight = cm_weight
self.cv_weight = cv_weight
self.autograd_tang = autograd_tang
self.autograd_sep = autograd_sep
self.enable_autograd = enable_autograd
self.label_smoothing = label_smoothing
self.stream_b_nce_weight = stream_b_nce_weight
self.geo_nce_weight = geo_nce_weight
self.config = {k: v for k, v in locals().items()
if k != 'self' and not k.startswith('_')}
# ββ v8: gate init from block count ββ
gate_init = 1.0 / (2.0 * n_blocks) # ~0.055 for 9 blocks
print(f" Gate init: {gate_init:.4f} (1/(2Γ{n_blocks}))")
# Shared patch embedding
self.patch_embed = nn.Conv2d(
3, embed_dim, kernel_size=patch_size, stride=patch_size)
self.pos_embed = nn.Parameter(
torch.randn(1, self.num_patches, embed_dim) * 0.02)
# Stream projections
self.proj_a = nn.Sequential(
nn.Linear(embed_dim, stream_dim), nn.LayerNorm(stream_dim))
self.proj_b = nn.Sequential(
nn.Linear(embed_dim, stream_dim), nn.LayerNorm(stream_dim))
# Shared GAL
self.gal = GAL(stream_dim, n_gal_anchors, n_heads,
ksimplex_k, ksimplex_edim, dropout)
# Tri-stream blocks β v8: pass gate_init
self.blocks = nn.ModuleList([
TriStreamBlock(stream_dim, n_gal_anchors, n_heads,
gate_init=gate_init, dropout=dropout)
for _ in range(n_blocks)])
# Output norms
self.norm_a = nn.LayerNorm(stream_dim)
self.norm_b = nn.LayerNorm(stream_dim)
# Sphere projections
self.proj_sphere_a = nn.Sequential(
nn.Linear(stream_dim, output_dim), nn.LayerNorm(output_dim))
self.proj_sphere_b = nn.Sequential(
nn.Linear(stream_dim, output_dim), nn.LayerNorm(output_dim))
self.proj_sphere_geo = nn.Sequential(
nn.Linear(stream_dim, output_dim), nn.LayerNorm(output_dim))
# Constellation + Patchwork (uniform hypersphere via Constellation)
self.constellation = Constellation(n_anchors, output_dim, anchor_drop)
self.patchwork = Patchwork(n_anchors, n_comp, d_comp)
pw_dim = n_comp * d_comp
# Classifiers
self.classifier_a = nn.Sequential(
nn.Linear(pw_dim + output_dim, pw_dim), nn.GELU(),
nn.LayerNorm(pw_dim), nn.Dropout(dropout),
nn.Linear(pw_dim, num_classes))
self.classifier_b = nn.Sequential(
nn.Linear(pw_dim + output_dim, pw_dim), nn.GELU(),
nn.LayerNorm(pw_dim), nn.Dropout(dropout),
nn.Linear(pw_dim, num_classes))
self.geo_classifier = nn.Sequential(
nn.Linear(output_dim, output_dim), nn.GELU(),
nn.Dropout(dropout),
nn.Linear(output_dim, num_classes))
self._init_weights()
def _init_weights(self):
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)
def forward(self, x, apply_autograd=True):
output = {}
B = x.shape[0]
# Patch embedding
tokens = self.patch_embed(x).flatten(2).transpose(1, 2)
tokens = tokens + self.pos_embed
P = tokens.shape[1]
# Split
stream_a = self.proj_a(tokens)
stream_b = self.proj_b(tokens)
# Anchor KV once
anchor_kv = self.gal.get_anchor_kv()
# Process through blocks
all_geo_feats = []
all_vol2 = []
geo_accum = torch.zeros_like(stream_a)
for block in self.blocks:
stream_a, stream_b, geo_feats, vol2, geo_lifted = block(
stream_a, stream_b, self.gal, anchor_kv)
all_geo_feats.append(geo_feats)
all_vol2.append(vol2)
geo_accum = geo_accum + geo_lifted
output['geo_feats'] = all_geo_feats[-1]
output['all_geo_feats'] = torch.stack(all_geo_feats)
output['vol2'] = torch.stack(all_vol2)
# Norms
stream_a = self.norm_a(stream_a)
stream_b = self.norm_b(stream_b)
# Pool
pool_a = stream_a.mean(dim=1)
pool_b = stream_b.mean(dim=1)
pool_geo = geo_accum.mean(dim=1)
# β sphere
emb_a = F.normalize(self.proj_sphere_a(pool_a), dim=-1)
emb_b = F.normalize(self.proj_sphere_b(pool_b), dim=-1)
geo_emb = F.normalize(self.proj_sphere_geo(pool_geo), dim=-1)
# Combined
emb = F.normalize(emb_a + emb_b + geo_emb, dim=-1)
# EmbeddingAutograd
if apply_autograd and self.training and self.enable_autograd:
emb = EmbeddingAutograd.apply(
emb, emb, self.constellation.anchors,
self.autograd_tang, self.autograd_sep)
# ββ v8: autograd on ALL three sub-embeddings ββ
emb_b = EmbeddingAutograd.apply(
emb_b, emb_b, self.constellation.anchors,
self.autograd_tang, self.autograd_sep)
geo_emb = EmbeddingAutograd.apply(
geo_emb, geo_emb, self.constellation.anchors,
self.autograd_tang, self.autograd_sep)
output['embedding'] = emb
output['emb_a'] = emb_a
output['emb_b'] = emb_b
output['geo_emb'] = geo_emb
output['pool_geo'] = pool_geo
# Constellation + Patchwork
tri_full, nearest_full = self.constellation.triangulate(
emb, training=False)
pw = self.patchwork(tri_full)
output['triangulation'] = tri_full
if self.training:
_, nearest = self.constellation.triangulate(emb, training=True)
else:
nearest = nearest_full
output['nearest'] = nearest
# Classifiers
logits_a = self.classifier_a(torch.cat([pw, emb_a], dim=-1))
logits_b = self.classifier_b(torch.cat([pw, emb_b], dim=-1))
geo_logits = self.geo_classifier(geo_emb)
output['logits_a'] = logits_a
output['logits_b'] = logits_b
output['geo_logits'] = geo_logits
# Gate monitoring
gates_a = [b.gal_block.gate_a.item() for b in self.blocks]
gates_b = [b.gal_block.gate_b.item() for b in self.blocks]
output['gates_a'] = gates_a
output['gates_b'] = gates_b
return output
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# PROCRUSTES ANCHOR UPDATE (unchanged)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@torch.no_grad()
def update_gal_anchors(self, simplex_buffer, lr=0.015, whiten=False):
with torch.amp.autocast("cuda", enabled=False):
centroids = simplex_buffer.class_centroids(self.num_classes)
if centroids is None:
return None
anchors = self.gal.gal_anchors.float()
centroid_n = F.normalize(centroids.float(), dim=-1)
anchor_n = F.normalize(anchors, dim=-1)
cos = centroid_n @ anchor_n.T
matched_idx = cos.argmax(dim=1)
matched_anchors = anchors[matched_idx]
R, score = procrustes_align(
matched_anchors, centroids.float(), whiten=whiten)
rotated = anchors @ R
new_anchors = F.normalize(
anchors + lr * (rotated - anchors), dim=-1)
self.gal.gal_anchors.copy_(
new_anchors.to(self.gal.gal_anchors.dtype))
return score
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# LOSS β v8: InfoNCE on emb_b + stronger geo_emb signal
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def compute_loss(self, output, targets, output_aug=None,
mastery_queue=None):
loss_dict = {}
emb = output['embedding']
emb_b = output['emb_b']
geo_emb = output['geo_emb']
B = emb.shape[0]
is_mastery = mastery_queue is not None and mastery_queue.active
# ββ CE on Stream A ββ
l_ce = F.cross_entropy(output['logits_a'], targets)
loss_dict['ce'] = l_ce
acc_a = (output['logits_a'].argmax(-1) == targets).float().mean().item()
loss_dict['acc_a'] = acc_a
# ββ BCE on Stream B ββ
one_hot = F.one_hot(targets, self.num_classes).float()
ls = self.label_smoothing
one_hot_smooth = one_hot * (1.0 - ls) + ls / self.num_classes if ls > 0 else one_hot
l_bce = F.binary_cross_entropy_with_logits(
output['logits_b'], one_hot_smooth)
loss_dict['bce'] = l_bce
acc_b = (output['logits_b'].argmax(-1) == targets).float().mean().item()
loss_dict['acc_b'] = acc_b
# ββ Geo classifier BCE ββ
l_geo_bce = F.binary_cross_entropy_with_logits(
output['geo_logits'], one_hot_smooth)
loss_dict['geo_bce'] = l_geo_bce
geo_acc = (output['geo_logits'].argmax(-1) == targets).float().mean().item()
loss_dict['geo_acc'] = geo_acc
# ββ InfoNCE β v8: on combined, emb_b, AND geo_emb ββ
nce_acc = 0.0
if output_aug is not None:
labels_nce = torch.arange(B, device=emb.device)
# Combined embedding InfoNCE
emb_aug = output_aug['embedding']
sim = emb @ emb_aug.T / self.infonce_temp
l_nce = F.cross_entropy(sim, labels_nce)
nce_acc = (sim.argmax(1) == labels_nce).float().mean().item()
loss_dict['nce'] = l_nce
loss_dict['nce_acc'] = nce_acc
# ββ v8: Stream B InfoNCE (this is what keeps B alive) ββ
emb_b_aug = output_aug.get('emb_b')
if emb_b_aug is not None:
sim_b = emb_b @ emb_b_aug.T / self.infonce_temp
l_nce_b = F.cross_entropy(sim_b, labels_nce)
nce_b_acc = (sim_b.argmax(1) == labels_nce).float().mean().item()
loss_dict['nce_b'] = l_nce_b
loss_dict['nce_b_acc'] = nce_b_acc
# ββ v8: Geo InfoNCE (this is what feeds the geo path) ββ
geo_emb_aug = output_aug.get('geo_emb')
if geo_emb_aug is not None:
sim_g = geo_emb @ geo_emb_aug.T / self.infonce_temp
l_geo_nce = F.cross_entropy(sim_g, labels_nce)
geo_nce_acc = (sim_g.argmax(1) == labels_nce).float().mean().item()
loss_dict['geo_nce'] = l_geo_nce
loss_dict['geo_nce_acc'] = geo_nce_acc
# ββ Mastery (unchanged) ββ
if is_mastery:
q_emb, q_labels = mastery_queue.get()
if q_emb is not None and q_emb.shape[0] >= B:
cross_sim = emb @ q_emb.T
same_mask = targets.unsqueeze(1) == q_labels.unsqueeze(0)
hn_sim = cross_sim.clone(); hn_sim[same_mask] = -1e9
hn_cos = hn_sim.max(dim=1).values
hp_sim = cross_sim.clone(); hp_sim[~same_mask] = 1e9
hp_cos = hp_sim.min(dim=1).values
valid = same_mask.any(1) & (~same_mask).any(1)
if valid.sum() > 0:
margin = mastery_queue.current_margin
l_mastery = F.relu(
hn_cos[valid] - hp_cos[valid] + margin).mean()
loss_dict['mastery'] = l_mastery
loss_dict['hard_neg_cos'] = hn_cos[valid].mean().item()
loss_dict['hard_pos_cos'] = hp_cos[valid].mean().item()
loss_dict['margin'] = margin
mastery_queue.push(emb.detach(), targets.detach())
# ββ CM validity ββ
vol2 = output['vol2']
l_cm = F.relu(-vol2).mean()
loss_dict['cm'] = l_cm
loss_dict['cm_valid'] = (vol2 > 0).float().mean().item()
# ββ CV on combined + geo ββ
l_cv_main = self._cv_loss_fast(emb, target=self.cv_target)
l_cv_geo = self._cv_loss_fast(geo_emb, target=self.cv_target)
l_cv = l_cv_main + l_cv_geo
loss_dict['cv'] = l_cv
loss_dict['cv_main'] = l_cv_main.item() if torch.is_tensor(l_cv_main) else l_cv_main
loss_dict['cv_geo'] = l_cv_geo.item() if torch.is_tensor(l_cv_geo) else l_cv_geo
# ββ Anchor spread ββ
anchors_n = F.normalize(self.constellation.anchors, dim=-1)
anchor_sim = anchors_n @ anchors_n.T
mask_a = ~torch.eye(anchors_n.shape[0], dtype=torch.bool,
device=anchors_n.device)
l_spread = F.relu(anchor_sim[mask_a] - 0.0).mean()
loss_dict['spread'] = l_spread
# ββ Combine β v8: explicit weights for B and geo NCE ββ
loss = (l_ce * self.bce_weight
+ l_bce * self.bce_weight
+ l_geo_bce * self.bce_weight
+ loss_dict.get('nce', 0.0) * self.infonce_weight
+ loss_dict.get('nce_b', 0.0) * self.stream_b_nce_weight
+ loss_dict.get('geo_nce', 0.0) * self.geo_nce_weight
+ loss_dict.get('mastery', 0.0) * self.bce_weight
+ l_cm * self.cm_weight
+ l_cv * self.cv_weight
+ l_spread * 0.001)
loss_dict['total'] = loss
return loss, loss_dict
@staticmethod
def _cv_loss_fast(emb, target=0.22, n_samples=64, n_points=5):
B = emb.shape[0]
if B < n_points:
return torch.tensor(0.0, device=emb.device)
vols = []
for _ in range(n_samples):
idx = torch.randperm(min(B, 512), device=emb.device)[:n_points]
pts = emb[idx].unsqueeze(0)
gram = torch.bmm(pts, pts.transpose(1, 2))
norms = torch.diagonal(gram, dim1=1, dim2=2)
d2 = norms.unsqueeze(2) + norms.unsqueeze(1) - 2 * gram
d2 = F.relu(d2)
N = n_points
cm = torch.zeros(1, N + 1, N + 1,
device=emb.device, dtype=emb.dtype)
cm[:, 0, 1:] = 1; cm[:, 1:, 0] = 1; cm[:, 1:, 1:] = d2
k = N - 1
sign = (-1.0) ** (k + 1)
fact = math.factorial(k)
prefactor = sign / ((2.0 ** k) * (fact ** 2))
vol2 = prefactor * torch.linalg.det(cm.float())
if vol2[0].item() > 1e-20:
vols.append(vol2[0].to(emb.dtype).sqrt())
if len(vols) < 5:
return torch.tensor(0.0, device=emb.device)
vols_t = torch.stack(vols)
cv = vols_t.std() / (vols_t.mean() + 1e-8)
return (cv - target).pow(2)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# MASTERY QUEUE (unchanged)
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class MasteryQueue:
def __init__(self, dim, min_size=1024, max_size=8192, initial_size=4096,
patience=50, device='cuda',
margin_start=0.1, margin_end=0.3, margin_warmup=5000,
resize_step=1024, resize_cooldown=5, overfit_threshold=3.0):
self.dim = dim
self.min_size = min_size; self.max_size = max_size
self._current_max = initial_size
self.patience = patience; self.device = device
self.active = False
self._embs = None; self._labels = None
self._perfect_count = 0; self._total_batches = 0
self._activated_at = None
self._margin_start = margin_start
self._margin_end = margin_end
self._margin_warmup = margin_warmup
self._mastery_steps = 0
self._resize_step = resize_step
self._resize_cooldown = resize_cooldown
self._overfit_threshold = overfit_threshold
self._epochs_since_resize = resize_cooldown
self._gap_history = []; self._gap_window = 5
self._resize_history = []
def check_activation(self, nce_acc):
self._total_batches += 1
if nce_acc >= 0.99:
self._perfect_count += 1
else:
self._perfect_count = 0
if not self.active and self._perfect_count >= self.patience:
self.active = True
self._activated_at = self._total_batches
print(f"\n β
MASTERY ACTIVATED at batch {self._total_batches} "
f"(nce_acc=1.0 for {self.patience} consecutive) "
f"queue={self._current_max}")
if self.active:
self._mastery_steps += 1
def update_size(self, train_acc, val_acc, epoch):
if not self.active: return
self._epochs_since_resize += 1
gap = train_acc - val_acc
self._gap_history.append((epoch, gap))
if self._epochs_since_resize < self._resize_cooldown: return
old_size = self._current_max; reason = None
if gap > self._overfit_threshold * 2:
self._current_max = min(self._current_max + self._resize_step, self.max_size)
reason = f"grow: gap={gap:.1f}%"
elif gap < self._overfit_threshold and gap > 0:
if len(self._gap_history) >= self._gap_window:
recent = [g for _, g in self._gap_history[-self._gap_window:]]
if all(0 < g < self._overfit_threshold for g in recent):
self._current_max = max(self._current_max - self._resize_step, self.min_size)
reason = f"shrink: stable gap={gap:.1f}%"
if reason is None and len(self._gap_history) >= self._gap_window:
drift = gap - self._gap_history[-self._gap_window][1]
if drift > self._overfit_threshold:
self._current_max = min(self._current_max + self._resize_step, self.max_size)
reason = f"drift: {drift:+.1f}%"
elif drift < -self._overfit_threshold and gap > 0:
self._current_max = max(self._current_max - self._resize_step, self.min_size)
reason = f"drift: {drift:+.1f}%"
if self._current_max != old_size:
d = "β" if self._current_max > old_size else "β"
print(f" β Queue {d} {old_size}β{self._current_max} ({reason})")
self._epochs_since_resize = 0
self._resize_history.append((epoch, old_size, self._current_max, gap, reason))
if self._embs is not None and self._embs.shape[0] > self._current_max:
self._embs = self._embs[-self._current_max:]
self._labels = self._labels[-self._current_max:]
@property
def current_margin(self):
if not self.active: return self._margin_start
t = min(self._mastery_steps / max(self._margin_warmup, 1), 1.0)
return self._margin_start + t * (self._margin_end - self._margin_start)
def push(self, emb, labels):
emb = emb.detach().to(self.device)
labels = labels.detach().to(self.device)
if self._embs is None:
self._embs = emb; self._labels = labels
else:
self._embs = torch.cat([self._embs, emb], 0)[-self._current_max:]
self._labels = torch.cat([self._labels, labels], 0)[-self._current_max:]
def get(self):
if self._embs is None: return None, None
return self._embs, self._labels
@property
def size(self):
return 0 if self._embs is None else self._embs.shape[0]
def state_dict(self):
return {
'active': self.active, 'total_batches': self._total_batches,
'activated_at': self._activated_at,
'mastery_steps': self._mastery_steps,
'current_margin': self.current_margin,
'current_max': self._current_max,
'gap_history': self._gap_history[-20:],
'resize_history': self._resize_history,
}
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# FACTORY
# ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def create_tri_stream_vit(**kwargs):
return TriStreamViT(**kwargs) |