Spaces:
Sleeping
Sleeping
File size: 57,821 Bytes
9ee6b93 9bbf5fa 34f439a b8a8f34 9bbf5fa 9ee6b93 746c56d 0919b50 746c56d 9ee6b93 b8a8f34 2920441 4c236ce 9bbf5fa 4c236ce 9bbf5fa 90b8ab6 9bbf5fa 0919b50 9bbf5fa 0919b50 9bbf5fa 9ee6b93 9bbf5fa 9ee6b93 9bbf5fa 4c236ce 9bbf5fa 55a1dab 9bbf5fa 55a1dab 9bbf5fa 55a1dab 9bbf5fa a7a4cee 9bbf5fa a7a4cee 9bbf5fa 90b8ab6 9bbf5fa 90b8ab6 9bbf5fa 746c56d 9bbf5fa b8a8f34 9bbf5fa 746c56d 9bbf5fa 746c56d 9bbf5fa 746c56d 9bbf5fa 746c56d 9bbf5fa b8a8f34 9bbf5fa 746c56d 9bbf5fa b8a8f34 9bbf5fa 0919b50 9bbf5fa 90b8ab6 9bbf5fa 90b8ab6 9bbf5fa 90b8ab6 9bbf5fa b8a8f34 746c56d b8a8f34 9bbf5fa b8a8f34 9bbf5fa b8a8f34 9bbf5fa b8a8f34 9bbf5fa b8a8f34 9bbf5fa 5612ce6 9bbf5fa 5612ce6 9bbf5fa 5612ce6 9bbf5fa b8a8f34 9bbf5fa 746c56d 9bbf5fa 746c56d 9bbf5fa 746c56d 9bbf5fa 746c56d 9bbf5fa 746c56d 9bbf5fa 746c56d 9bbf5fa 0919b50 9bbf5fa 0919b50 9bbf5fa b8a8f34 5612ce6 9bbf5fa 5612ce6 9bbf5fa b8a8f34 9bbf5fa b8a8f34 9bbf5fa b8a8f34 9bbf5fa b8a8f34 9bbf5fa b8a8f34 9ee6b93 9bbf5fa 0919b50 9bbf5fa 9ee6b93 9bbf5fa 90b8ab6 9bbf5fa b8a8f34 5612ce6 9bbf5fa 5612ce6 9bbf5fa b8a8f34 5612ce6 9bbf5fa 5612ce6 9bbf5fa 5612ce6 9bbf5fa 90b8ab6 9bbf5fa 746c56d 9bbf5fa b8a8f34 0919b50 9bbf5fa b8a8f34 9ee6b93 90b8ab6 | 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 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 | """
ARC-AGI-3 Agent Spectator v4
Hugging Face Space: beanapologist/arc-agi
Re/Im solver live demo:
Im side = bird's eye hypothesis (which transformation?)
Re side = exact diff (which cells to click?)
Bridge = ACTION6 at the Re-side coordinates that close the gap
"""
import gradio as gr
import numpy as np
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from matplotlib.colors import ListedColormap
import torch
import torch.nn as nn
import torch.nn.functional as TF
import io, os, time, threading, queue
from collections import deque
from PIL import Image
# ── Palette ───────────────────────────────────────────────────────────────────
ARC_HEX = ['#000000','#1a6faf','#e03a3a','#3aa63a','#f5c400',
'#c060c0','#d07030','#aaaaaa','#60b8d0','#874010']
ARC_CMAP = ListedColormap(ARC_HEX)
COLOR_NAMES = ['black','blue','red','green','yellow',
'purple','orange','gray','azure','maroon']
# ── Re/Im primitives ──────────────────────────────────────────────────────────
def _sobel(f):
p=np.pad(f,1,mode='edge')
gx=(-p[:-2,:-2]-2*p[1:-1,:-2]-p[2:,:-2]+p[:-2,2:]+2*p[1:-1,2:]+p[2:,2:])/8
gy=(-p[:-2,:-2]-2*p[:-2,1:-1]-p[:-2,2:]+p[2:,:-2]+2*p[2:,1:-1]+p[2:,2:])/8
return gx,gy
def _sym_axis(grid,axis):
H,W=grid.shape; best_s,best_i=0.0,0
if axis=='h':
for x in range(1,W-1):
r=min(x,W-1-x)
s=(grid[:,x-r:x]==grid[:,x+1:x+r+1][:,::-1]).mean()
if s>best_s: best_s,best_i=s,x
else:
for y in range(1,H-1):
r=min(y,H-1-y)
s=(grid[y-r:y,:]==grid[y+1:y+r+1,:][::-1,:]).mean()
if s>best_s: best_s,best_i=s,y
return best_i,best_s
def _boundary(grid):
p=np.pad(grid,1,mode='edge')
return ((p[1:-1,1:-1]!=p[:-2,1:-1])|(p[1:-1,1:-1]!=p[2:,1:-1])|
(p[1:-1,1:-1]!=p[1:-1,:-2])|(p[1:-1,1:-1]!=p[1:-1,2:])).astype(np.float32)
def _sym(grid,axis):
H,W=grid.shape; s=np.zeros((H,W),np.float32)
if axis=='h':
for x in range(W):
r=min(x,W-1-x)
if r==0: s[:,x]=1.; continue
s[:,x]=(grid[:,x-r:x]==grid[:,x+1:x+r+1][:,::-1]).mean()
else:
for y in range(H):
r=min(y,H-1-y)
if r==0: s[y,:]=1.; continue
s[y,:]=(grid[y-r:y,:]==grid[y+1:y+r+1,:][::-1,:]).mean()
return s
# ── Re/Im board reader — reads directly from 56-channel feature tensor ─────────
# Im tensor = signal strength | raw grid = coordinates | i· = answer
"""
arc_solver.py — Re/Im analytic solver using the 56-channel feature tensor
=========================================================================
The board IS M = Re(M) + i·Im(M).
The 56-channel extractor already IS log(M) decomposed:
Re channels (ch 0-47):
ch 0-15: one-hot colors — multiplicative / "what's there"
ch 16-31: component size maps — "how many, how big" (radial magnitude)
ch 32-47: distance-from-color maps — proximity / modulus
Im channels (ch 48-55):
ch 48: H-symmetry map — fold axis (Re(s)=1/2 analog)
ch 49: V-symmetry map — fold axis (vertical)
ch 50: rotational symmetry — winding structure
ch 51: Sobel x (edge horiz) — gradient Re component
ch 52: Sobel y (edge vert) — gradient Im component ← arg(M)
ch 53: boundary contour — Cauchy ∮ contour
ch 54: curl / winding proxy — arg(M), rotation direction
ch 55: component ID map — global topological / winding data
The Im channels are not clues — they ARE the answer structure.
i· (column swap) means: read the Im channels to get the Re answer.
Im H-symmetry strong + Re mass asymmetric → mirror to complete
Im boundary contour + Re solid fill → boundary IS the answer
Im curl/winding strong → rotation is needed
Im Sobel-y dominates → vertical flow (gravity)
Im component map has isolated islands → count → place
The CNN operates on all 56 channels.
This solver reads the Im channels analytically to short-circuit the CNN
when the Im signal is unambiguous.
"""
import numpy as np
from typing import Optional, Tuple, List
# ── Re/Im channel indices (must match extract_features in my_agent.py) ────────
# Space uses 14-channel extractor (subset of full 56-channel Kaggle version)
# ch 0-9: Re — one-hot colors
# ch 10: Im — H-symmetry (fold axis)
# ch 11: Im — V-symmetry
# ch 12: Im — Cauchy boundary contour
# ch 13: Im — edge magnitude (Sobel combined)
CH_ONEHOT_START = 0
CH_CCSIZE_START = 0 # not in Space extractor — use onehot
CH_DIST_START = 0 # not in Space extractor — use onehot
CH_H_SYM = 10 # Im — horizontal symmetry
CH_V_SYM = 11 # Im — vertical symmetry
CH_ROT_SYM = 10 # not separate — reuse H-sym
CH_SOBEL_X = 13 # Im — edge (proxy for gradient)
CH_SOBEL_Y = 13 # Im — edge (proxy for gradient)
CH_BOUNDARY = 12 # Im — Cauchy boundary contour
CH_CURL = 10 # not separate — reuse H-sym
CH_COMPONENT_ID = 11 # not separate — reuse V-sym
NUM_CHANNELS = 14 # Space extractor has 14 channels
CONFIDENCE_THRESHOLD = 0.72
# ── Raw grid primitives (for when we don't have the tensor yet) ───────────────
def _boundary_raw(grid):
p = np.pad(grid, 1, mode='edge')
return ((p[1:-1,1:-1]!=p[:-2,1:-1])|(p[1:-1,1:-1]!=p[2:,1:-1])|
(p[1:-1,1:-1]!=p[1:-1,:-2])|(p[1:-1,1:-1]!=p[1:-1,2:])).astype(np.float32)
def _perimeter(grid):
"""Object perimeter — handles solid blocks where _boundary_raw gives 0."""
H, W = grid.shape
p = np.zeros((H,W), dtype=np.float32)
mask = grid > 0
if not mask.any(): return p
padded = np.pad(mask.astype(int), 1, constant_values=0)
for dy,dx in [(-1,0),(1,0),(0,-1),(0,1)]:
shifted = padded[1+dy:H+1+dy, 1+dx:W+1+dx]
p[mask & (shifted==0)] = 1
if p.sum() == 0: # solid block: use outer ring
p[0,:] = mask[0,:].astype(float); p[-1,:] = mask[-1,:].astype(float)
p[:,0] = mask[:,0].astype(float); p[:,-1] = mask[:,-1].astype(float)
return p
# ── Re/Im board reader ────────────────────────────────────────────────────────
def read_board(grid: np.ndarray, features: Optional[np.ndarray] = None):
"""
Read the board as complex object M = Re(M) + i·Im(M).
Parameters
----------
grid : 2D int array, the raw color grid
features : optional (56, H, W) float array from extract_features()
If provided, reads Im channels directly (more accurate).
If None, recomputes Im channels from the raw grid.
Returns
-------
answer : np.ndarray — derived answer grid
confidence : float 0-1
reasoning : list of strings — one per Im signal that fired
signal : str — which Im channel drove the answer
"""
H, W = grid.shape
reasoning = []
answer = None
confidence = 0.0
signal = 'none'
# ── Read Im channels ──────────────────────────────────────────────────────
# Use pre-computed feature tensor if available (same computation the CNN uses)
if features is not None and features.shape[0] == NUM_CHANNELS:
feat = features
# Resize Im maps back to grid space if needed
fH, fW = feat.shape[1], feat.shape[2]
def _im(ch):
"""Read Im channel, resize to grid if needed."""
m = feat[ch]
if (fH, fW) != (H, W):
# Simple nearest-neighbor resize
ry = np.linspace(0, fH-1, H).astype(int)
rx = np.linspace(0, fW-1, W).astype(int)
return m[np.ix_(ry, rx)]
return np.array(m)
h_sym_map = _im(CH_H_SYM)
v_sym_map = _im(CH_V_SYM)
rot_map = _im(CH_ROT_SYM)
sobel_x = _im(CH_SOBEL_X)
sobel_y = _im(CH_SOBEL_Y)
bound_map = _im(CH_BOUNDARY)
curl_map = _im(CH_CURL)
comp_map = _im(CH_COMPONENT_ID)
# Re channels: color presence and component sizes
onehot = feat[CH_ONEHOT_START:CH_ONEHOT_START+16]
cc_sizes = feat[CH_CCSIZE_START:CH_CCSIZE_START+16]
else:
# Compute features inline and retry once (no recursion)
t = extract_features(grid)
feat_np = t.numpy() if hasattr(t, 'numpy') else np.array(t)
if feat_np.shape[0] == NUM_CHANNELS:
return read_board(grid, features=feat_np)
# Shape mismatch — just compute Im maps directly from grid
_gx,_gy = _sobel(grid.astype(np.float32)/9)
h_sym_map = _sym(grid,'h'); v_sym_map = _sym(grid,'v')
bound_map = _boundary(grid)
edge_map = np.sqrt(_gx**2+_gy**2).astype(np.float32)
# Build minimal feature array
_oh = np.zeros((10,H,W),np.float32)
for _c in range(10): _oh[_c]=(grid==_c).astype(np.float32)
features = np.concatenate([_oh,h_sym_map[np.newaxis],v_sym_map[np.newaxis],
bound_map[np.newaxis],edge_map[np.newaxis]],axis=0)
return read_board(grid, features=features)
# ── Im signal 1: H-symmetry (fold axis) ───────────────────────────────────
# Im ch48 gives the symmetry score — use its MAX as signal strength.
# But compute the actual best axis in raw grid space (not 64x64 space).
# This is the Re/Im separation: Im tensor gives the SIGNAL,
# Re grid gives the COORDINATES.
best_hs = float(h_sym_map.max()) # Im: how strong is any fold?
if best_hs > 0.65:
# Find best axis in grid space (not feature space)
h_scores_grid = []
for x in range(1, W-1):
r = min(x, W-1-x)
if r > 0:
s = (grid[:, x-r:x] == grid[:, x+1:x+r+1][:,::-1]).mean()
h_scores_grid.append((x, float(s)))
if h_scores_grid:
best_hx, best_hs_grid = max(h_scores_grid, key=lambda x: x[1])
# Use Im tensor strength as upper bound, grid score as actual
hs = min(best_hs, best_hs_grid) if best_hs_grid > 0.4 else best_hs_grid
left_px = int((grid[:, :best_hx] > 0).sum())
right_px = int((grid[:, best_hx:] > 0).sum())
total_px = left_px + right_px
asymmetry = abs(left_px - right_px) / max(total_px, 1)
if asymmetry > 0.25 and hs > 0.55:
ans = grid.copy()
if left_px > right_px:
for c in range(best_hx):
mir = W - 1 - c
if 0 <= mir < W:
mask = ans[:, mir] == 0
ans[mask, mir] = grid[mask, c]
else:
for c in range(best_hx+1, W):
mir = W - 1 - c
if 0 <= mir < W:
mask = ans[:, mir] == 0
ans[mask, mir] = grid[mask, c]
conf = hs * asymmetry * 0.95
if conf > confidence:
answer, confidence, signal = ans, conf, 'Im:H-sym'
reasoning.append(
f"Im ch48 signal={best_hs:.2f} grid_score={hs:.2f} at x={best_hx} | "
f"Re left={left_px} right={right_px} asym={asymmetry:.2f} | "
f"i·: complete the fold")
# ── Im signal 2: V-symmetry ────────────────────────────────────────────────
best_vs = float(v_sym_map.max()) # Im: V-fold signal strength
if best_vs > 0.65 and confidence < 0.6:
v_scores_grid = []
for y in range(1, H-1):
r = min(y, H-1-y)
if r > 0:
s = (grid[y-r:y, :] == grid[y+1:y+r+1, :][::-1, :]).mean()
v_scores_grid.append((y, float(s)))
if v_scores_grid:
best_vy, best_vs_grid = max(v_scores_grid, key=lambda x: x[1])
vs = min(best_vs, best_vs_grid) if best_vs_grid > 0.4 else best_vs_grid
top_px = int((grid[:best_vy, :] > 0).sum())
bot_px = int((grid[best_vy:, :] > 0).sum())
total_px = top_px + bot_px
asymmetry = abs(top_px - bot_px) / max(total_px, 1)
if asymmetry > 0.25 and vs > 0.55:
ans = grid.copy()
if top_px > bot_px:
for r in range(best_vy):
mir = H - 1 - r
if 0 <= mir < H:
mask = ans[mir, :] == 0
ans[mir, mask] = grid[r, mask]
else:
for r in range(best_vy+1, H):
mir = H - 1 - r
if 0 <= mir < H:
mask = ans[mir, :] == 0
ans[mir, mask] = grid[r, mask]
conf = vs * asymmetry * 0.90
if conf > confidence:
answer, confidence, signal = ans, conf, 'Im:V-sym'
reasoning.append(
f"Im ch49 signal={best_vs:.2f} grid_score={vs:.2f} at y={best_vy} | "
f"Re top={top_px} bot={bot_px} | i·: complete V fold")
# ── Im signal 3: Cauchy boundary contour ──────────────────────────────────
# Im ch53 IS the Cauchy contour. When Re fill is solid (high density)
# and Im boundary is thin, the answer IS the boundary.
# "The radius cancels (Re collapses); only the angular part i·dθ survives"
total_px = int((grid > 0).sum())
if total_px > 0 and confidence < 0.75:
fill_ratio = total_px / (H * W)
colors = [c for c in range(1,10) if (grid==c).any()]
if len(colors) == 1 and fill_ratio > 0.4:
# Single color, high fill → Im boundary IS the answer
perim = _perimeter(grid)
# Check interior isn't already hollow
interior = (grid == 0) & (perim == 0)
if not interior.any():
ans = np.zeros_like(grid)
ans[perim > 0] = grid[perim > 0]
conf = fill_ratio * 0.90
if conf > confidence:
answer, confidence, signal = ans, conf, 'Im:boundary'
reasoning.append(
f"Im ch53 Cauchy contour | Re fill={fill_ratio:.2f} "
f"single color={colors[0]} | "
f"i·: Im boundary = Re answer (Cauchy: radius cancels)")
elif fill_ratio > 0.3 and bound_map.mean() < 0.15:
# Multi-color but low boundary density → extract boundary
ans = np.zeros_like(grid)
ans[bound_map > 0.3] = grid[bound_map > 0.3]
if (ans > 0).any():
conf = fill_ratio * (1 - float(bound_map.mean())) * 0.75
if conf > confidence:
answer, confidence, signal = ans, conf, 'Im:boundary'
reasoning.append(
f"Im ch53 boundary density={bound_map.mean():.3f} | "
f"Re fill={fill_ratio:.2f} | i·: Cauchy contour")
# ── Im signal 4: Interior fill ────────────────────────────────────────────
# Im ch53: If colored region fully encloses empty cells → fill interior.
# Use flood-fill from boundary to find truly enclosed (unreachable) cells.
if confidence < 0.55 and total_px > 0:
# Flood-fill empty cells reachable from grid boundary
reachable = np.zeros((H, W), dtype=bool)
fq = []
for _r in range(H):
for _c in range(W):
if grid[_r,_c]==0 and (_r==0 or _r==H-1 or _c==0 or _c==W-1):
if not reachable[_r,_c]:
reachable[_r,_c]=True; fq.append((_r,_c))
while fq:
_y,_x=fq.pop()
for _dy,_dx in [(-1,0),(1,0),(0,-1),(0,1)]:
_ny,_nx=_y+_dy,_x+_dx
if 0<=_ny<H and 0<=_nx<W and grid[_ny,_nx]==0 and not reachable[_ny,_nx]:
reachable[_ny,_nx]=True; fq.append((_ny,_nx))
truly_interior = (grid == 0) & ~reachable
if truly_interior.any():
dominant = int(np.argmax(np.bincount(
grid[grid>0].flatten(), minlength=10)[1:])) + 1
ans = grid.copy()
ans[truly_interior] = dominant
conf = truly_interior.sum() / max(1, (grid==0).sum()) * 0.80
if conf > confidence:
answer, confidence, signal = ans, conf, 'Im:hollow→fill'
reasoning.append(
f"Im ch53 enclosed interior={int(truly_interior.sum())}px | "
f"Re dominant={dominant} | i·: fill enclosed interior")
# ── Im signal 5: Gradient flow → gravity ──────────────────────────────────
# Im ch51/52 = Sobel x/y = gradient field direction = arg(M)
# Suspended Re pixels + Im gradient direction → gravity answer
suspended = sum(
1 for c in range(W)
for r in np.where(grid[:, c] > 0)[0]
if r < H-1 and grid[r+1, c] == 0
)
if suspended > 0 and confidence < 0.70:
gx_mag = float(np.abs(sobel_x).mean())
gy_mag = float(np.abs(sobel_y).mean())
direction = 'down' if gy_mag >= gx_mag else 'right'
ans = np.zeros_like(grid)
if direction == 'down':
for c in range(W):
vals = grid[:, c][grid[:, c] > 0]
if len(vals): ans[H-len(vals):H, c] = vals
else:
for r in range(H):
vals = grid[r, :][grid[r, :] > 0]
if len(vals): ans[r, W-len(vals):W] = vals
conf = min(0.80, suspended / max(total_px, 1) * 2.5)
if conf > confidence:
answer, confidence, signal = ans, conf, 'Im:Sobel→gravity'
reasoning.append(
f"Im ch52 Sobel-y={gy_mag:.3f} ch51 Sobel-x={gx_mag:.3f} | "
f"Re suspended={suspended}px | i·: arg(M) gives gravity {direction}")
# ── Im signal 6: Rotational symmetry / curl ────────────────────────────────
# Im ch50 rot_sym + ch54 curl → rotation answer
rot_score = float(rot_map.max())
curl_max = float(np.abs(curl_map).max())
if rot_score > 0.6 and curl_max > 0.4 and confidence < 0.35:
ans = np.rot90(grid)
conf = rot_score * curl_max * 0.60
if conf > confidence:
answer, confidence, signal = ans, conf, 'Im:rot+curl'
reasoning.append(
f"Im ch50 rot={rot_score:.2f} ch54 curl={curl_max:.2f} | "
f"i·: rotation indicated")
# ── Im signal 7: Color remapping (Re→Im count ordering) ───────────────────
# Im component map (ch55) encodes which regions are distinct objects.
# If Re colors appear in counts that suggest an ordering → shift colors.
if confidence < 0.30 and total_px > 0:
colors = [c for c in range(1,10) if (grid==c).any()]
if colors and max(colors) < 9:
ans = grid.copy()
mask = grid > 0
ans[mask] = ((grid[mask] - 1 + 1) % 9) + 1
if conf > confidence:
answer, confidence, signal = ans, 0.30, 'Im:color_shift'
reasoning.append(
f"Im ch55 component topology | Re colors {colors} | "
f"i·: Re→Im shift = increment colors")
return answer, confidence, reasoning, signal
# ── Re-side: exact cell targeting ────────────────────────────────────────────
def pixel_diff(current: np.ndarray, target: np.ndarray):
"""All differing cells: [(r, c, target_color)]"""
if current.shape != target.shape: return []
return [(int(r), int(c), int(target[r,c]))
for r in range(current.shape[0])
for c in range(current.shape[1])
if current[r,c] != target[r,c]]
def most_urgent_diff(current: np.ndarray, target: np.ndarray):
"""
Im → Re: pick the most important cell using the Cauchy principle.
The boundary contour determines the interior, so fix boundary cells first.
This is ∮ doing its job: read global Im data, recover local Re data.
"""
diffs = pixel_diff(current, target)
if not diffs: return None
bound = _boundary_raw(current)
boundary_diffs = [(r,c,v) for r,c,v in diffs if bound[r,c] > 0]
pool = boundary_diffs if boundary_diffs else diffs
return pool[np.random.randint(len(pool))]
# ── Main entry point ─────────────────────────────────────────────────────────
def try_analytic_action(
frame_2d: np.ndarray,
available_actions,
features: Optional[np.ndarray] = None,
) -> Tuple[Optional[int], Optional[dict], str, float]:
"""
Read the board's Im channels to derive the answer, then use Re-side
pixel diff to find the exact cell to click.
Parameters
----------
frame_2d : raw 2D color grid
available_actions : list of available GameAction values
features : optional pre-computed (56,H,W) feature array
(pass this from MyAgent to avoid recomputing)
Returns (action_id, action_data, signal_name, confidence)
"""
if frame_2d is None: return None, None, 'none', 0.0
avail_ids = set(
int(a.value if hasattr(a,'value') else a)
for a in (available_actions or range(1,7))
)
answer, confidence, reasoning, signal = read_board(frame_2d, features)
if answer is None or confidence < CONFIDENCE_THRESHOLD:
return None, None, signal, confidence
diffs = pixel_diff(frame_2d, answer)
if not diffs:
return None, None, 'already_matches', confidence
# ACTION6: click the most urgent Re-side cell
if 6 in avail_ids:
cell = most_urgent_diff(frame_2d, answer)
if cell is not None:
r, c, _ = cell
H, W = frame_2d.shape
game_y = min(63, max(0, int(r * 64/H + 32/H)))
game_x = min(63, max(0, int(c * 64/W + 32/W)))
return 6, {'x': game_x, 'y': game_y}, signal, confidence
return None, None, 'no_action6', confidence
# ── Feature extractor ─────────────────────────────────────────────────────────
def extract_features(grid,num_colours=10):
H,W=grid.shape
oh=np.zeros((num_colours,H,W),np.float32)
for c in range(num_colours): oh[c]=(grid==c).astype(np.float32)
gx,gy=_sobel(grid.astype(np.float32)/9)
stacked=np.concatenate([oh,_sym(grid,'h')[np.newaxis],
_sym(grid,'v')[np.newaxis],
_boundary(grid)[np.newaxis],
np.sqrt(gx**2+gy**2)[np.newaxis].astype(np.float32)],axis=0)
t=torch.from_numpy(stacked).float().unsqueeze(0)
if H!=64 or W!=64:
t=TF.interpolate(t,size=(64,64),mode='bilinear',align_corners=False)
return t.squeeze(0)
# ── Gabor filter bank — s-plane cross terms ─────────────────────────────────
"""
gabor_channels.py
=================
Gabor filter bank for ARC-AGI-3 — the s-plane cross terms.
Mathematical position
---------------------
The existing 56-channel extractor covers the AXES of the s-plane:
Re axis (σ>0, ω=0): ch16-47 CC sizes, distance maps — Laplace side
Im axis (σ=0, ω>0): ch48-55 symmetry, Sobel, boundary — Fourier side
A Gabor filter lives at an INTERIOR point (σ>0, ω>0):
g(x,y) = exp(-sigma*r^2) · cos(ω·x_θ + φ)
______/ ___________/
Re/Laplace Im/Fourier
envelope carrier
This is exp(-st) evaluated at s = σ + iω, rotated by θ, phased by φ.
It measures: "is there oscillation at frequency ω in direction θ,
concentrated within decay radius 1/√σ?"
ARC relevance
-------------
The cross terms detect structures the axis channels miss:
- Repeating patterns with finite extent (tiling with boundary)
- Oriented edges at specific spatial scales
- Localized symmetry (symmetric patch inside asymmetric grid)
- Diagonal structure (axis channels are H/V only)
Channel layout (72 channels total)
-----------------------------------
3 σ values × 3 ω values × 4 orientations × 2 phases = 72
σ = 0.3 → tight decay, radius ~1.8px — local structure
σ = 1.0 → medium decay, radius ~1.0px — mid-scale
σ = 2.5 → broad decay, radius ~0.6px — global texture
ω = 0.5 → coarse frequency, period ~12px — large patterns
ω = 1.5 → medium frequency, period ~4px — medium patterns
ω = 3.0 → fine frequency, period ~2px — fine detail
θ = 0, π/4, π/2, 3π/4 — 4 orientations (H, diagonal, V, anti-diagonal)
φ = 0 → cosine (even symmetry, detects symmetric features)
φ = π/2 → sine (odd symmetry, detects antisymmetric/edge features)
"""
import numpy as np
import torch
import torch.nn.functional as F
from typing import List, Tuple
# ── Filter bank parameters ────────────────────────────────────────────────────
SIGMA_VALUES = [0.3, 1.0, 2.5] # Re/Laplace decay
OMEGA_VALUES = [0.5, 1.5, 3.0] # Im/Fourier frequency
THETA_VALUES = [0, np.pi/4, np.pi/2, 3*np.pi/4] # orientations
PHASE_VALUES = [0.0, np.pi/2] # cosine, sine
N_GABOR_CHANNELS = (len(SIGMA_VALUES) * len(OMEGA_VALUES) *
len(THETA_VALUES) * len(PHASE_VALUES)) # = 72
KERNEL_SIZE = 7 # 7×7 kernels — enough for 64×64 grid
def _make_gabor_kernel(sigma: float, omega: float,
theta: float, phi: float,
size: int = KERNEL_SIZE) -> np.ndarray:
"""
2D Gabor kernel at one point in the s-plane.
s = σ + iω (the seam: Re side = decay, Im side = oscillation)
θ = orientation, φ = phase
"""
half = size // 2
y, x = np.mgrid[-half:half+1, -half:half+1].astype(np.float32)
# Rotate to orientation θ
x_rot = x * np.cos(theta) + y * np.sin(theta)
y_rot = -x * np.sin(theta) + y * np.cos(theta)
# Re side: Gaussian envelope — exp(-sigma*r^2)
envelope = np.exp(-sigma * (x_rot**2 + y_rot**2))
# Im side: sinusoidal carrier — cos(ω·x_θ + φ)
carrier = np.cos(omega * x_rot + phi)
# s-plane cross term: exp(-sigma*r^2) · cos(ω·x_θ + φ)
kernel = envelope * carrier
# Zero-mean (remove DC) — ensures kernel responds to structure, not brightness
kernel -= kernel.mean()
norm = np.sqrt((kernel ** 2).sum())
if norm > 0:
kernel /= norm
return kernel # shape (size, size)
def build_gabor_bank() -> torch.Tensor:
"""
Build the full Gabor filter bank as a (72, 1, K, K) tensor
ready for torch.nn.functional.conv2d.
"""
kernels = []
for sigma in SIGMA_VALUES:
for omega in OMEGA_VALUES:
for theta in THETA_VALUES:
for phi in PHASE_VALUES:
k = _make_gabor_kernel(sigma, omega, theta, phi)
kernels.append(k)
bank = np.stack(kernels, axis=0) # (72, K, K)
return torch.from_numpy(bank).float().unsqueeze(1) # (72, 1, K, K)
# Pre-built bank — computed once at import time
_GABOR_BANK: torch.Tensor = build_gabor_bank()
def extract_gabor_features(grid_2d: np.ndarray,
grid_size: int = 64) -> torch.Tensor:
"""
Apply the Gabor bank to a 2D color grid.
Parameters
----------
grid_2d : np.ndarray (H, W) int — raw color grid
grid_size : int — target output size (default 64, matching ActionModel)
Returns
-------
torch.Tensor (72, grid_size, grid_size) float32
"""
H, W = grid_2d.shape
# Normalize grid to [0, 1] float
grid_f = torch.from_numpy(
grid_2d.astype(np.float32) / 9.0
).unsqueeze(0).unsqueeze(0) # (1, 1, H, W)
# Resize to grid_size if needed
if H != grid_size or W != grid_size:
grid_f = F.interpolate(grid_f, size=(grid_size, grid_size),
mode='bilinear', align_corners=False)
# Apply all 72 Gabor filters simultaneously
pad = KERNEL_SIZE // 2
responses = F.conv2d(grid_f, _GABOR_BANK, padding=pad) # (1, 72, H, W)
responses = responses.squeeze(0) # (72, grid_size, grid_size)
# Normalize responses to [-1, 1]
max_val = responses.abs().max()
if max_val > 0:
responses = responses / max_val
return responses
def channel_descriptions() -> List[str]:
"""Human-readable description of each Gabor channel."""
descs = []
ch = 0
for sigma in SIGMA_VALUES:
for omega in OMEGA_VALUES:
for theta in THETA_VALUES:
theta_deg = int(theta * 180 / np.pi)
for phi in PHASE_VALUES:
phase_name = 'cos' if phi == 0 else 'sin'
descs.append(
f"ch{56+ch:3d}: Gabor σ={sigma:.1f} ω={omega:.1f} "
f"θ={theta_deg}° φ={phase_name} "
f"[s={sigma:.1f}+{omega:.1f}i]"
)
ch += 1
return descs
# ── S-plane visualization ─────────────────────────────────────────────────────
def splane_coverage_report():
"""Print the s-plane coverage table."""
print("s-plane coverage: σ (Re/Laplace) × ω (Im/Fourier)")
print("="*55)
print(f"{'':8}", end="")
for o in OMEGA_VALUES:
print(f" ω={o:.1f}", end="")
print()
for s in SIGMA_VALUES:
print(f"σ={s:.1f} ", end="")
for o in OMEGA_VALUES:
n = len(THETA_VALUES) * len(PHASE_VALUES)
print(f" {n:3d}ch", end="")
print(f" (×{len(THETA_VALUES)}θ ×{len(PHASE_VALUES)}φ)")
print()
print(f"Total Gabor channels: {N_GABOR_CHANNELS}")
print(f"Existing axis channels: 56")
print(f"Combined total: {56 + N_GABOR_CHANNELS}")
# ── Rendering ─────────────────────────────────────────────────────────────────
def _pil(fig):
buf=io.BytesIO()
fig.savefig(buf,format='png',dpi=80,bbox_inches='tight',
facecolor=fig.get_facecolor())
buf.seek(0); img=Image.open(buf).copy(); plt.close(fig)
return img
def render_grid(grid,title='',highlight=None,mark_cell=None):
if grid is None: return None
H,W=grid.shape; cell=max(28,min(56,360//max(H,W)))
fig,ax=plt.subplots(figsize=((W*cell+4)/72,(H*cell+22)/72),dpi=72)
fig.patch.set_facecolor('#1e1e2e'); ax.set_facecolor('#1e1e2e')
ax.imshow(grid,cmap=ARC_CMAP,vmin=0,vmax=9,interpolation='nearest',aspect='equal')
for x in range(W+1): ax.axvline(x-.5,color='#444',lw=.5)
for y in range(H+1): ax.axhline(y-.5,color='#444',lw=.5)
for r in range(H):
for c in range(W):
v=int(grid[r,c])
col='white' if v in [0,1,2,3,5,6,9] else 'black'
ax.text(c,r,str(v),ha='center',va='center',
fontsize=max(7,cell//5),color=col,
fontweight='bold',fontfamily='monospace')
if highlight is not None:
for r,c,_ in highlight:
ax.add_patch(plt.Rectangle((c-.5,r-.5),1,1,
fill=True,facecolor='#ff4444',alpha=0.35,lw=0))
if mark_cell is not None:
r,c,_=mark_cell
ax.add_patch(plt.Rectangle((c-.5,r-.5),1,1,
fill=False,edgecolor='#00ffff',lw=2.5))
ax.plot(c,r,'*',color='#00ffff',markersize=max(8,cell//4))
ax.set_xlim(-.5,W-.5); ax.set_ylim(H-.5,-.5); ax.axis('off')
if title: ax.set_title(title,color='#cdd6f4',fontsize=9,pad=4)
plt.tight_layout(pad=.3)
return _pil(fig)
def render_hypothesis_panel(candidates):
"""Im side: bar chart of top hypotheses with confidence."""
if not candidates: return None
top=candidates[:6]
names=[c[0] for c in top]; confs=[c[2] for c in top]
fig,ax=plt.subplots(figsize=(5,2.2))
fig.patch.set_facecolor('#1e1e2e'); ax.set_facecolor('#1e1e2e')
colors=['#ffd700' if i==0 else '#4a9eff' for i in range(len(top))]
bars=ax.barh(names[::-1],confs[::-1],color=colors[::-1],height=0.6)
for bar,conf in zip(bars,confs[::-1]):
ax.text(bar.get_width()+.01,bar.get_y()+bar.get_height()/2,
f'{conf:.2f}',va='center',color='white',fontsize=8)
ax.set_xlim(0,1.15); ax.axvline(0.4,color='#ff6666',lw=1,ls='--',alpha=0.7)
ax.text(0.41,0,'threshold',color='#ff6666',fontsize=7,va='bottom')
ax.tick_params(colors='#888',labelsize=8); ax.spines[:].set_visible(False)
ax.set_title('Im side — hypothesis ranking 🟡=selected',
color='#cdd6f4',fontsize=9,pad=3)
plt.tight_layout(pad=.4)
return _pil(fig)
def render_action_bar(action_counts,total):
if not action_counts or total==0: return None
labels=[f"A{k}" for k in sorted(action_counts)]
vals =[action_counts[k] for k in sorted(action_counts)]
pcts =[v/total*100 for v in vals]
fig,ax=plt.subplots(figsize=(4,1.6))
fig.patch.set_facecolor('#1e1e2e'); ax.set_facecolor('#1e1e2e')
colors=['#4a9eff','#e05050','#50c050','#f5c400','#c060c0','#d07030']
bars=ax.barh(labels,pcts,color=colors[:len(labels)],height=0.6)
for bar,v,p in zip(bars,vals,pcts):
ax.text(min(p+1,98),bar.get_y()+bar.get_height()/2,
f'{v}',va='center',color='white',fontsize=8)
ax.set_xlim(0,110); ax.tick_params(colors='#888',labelsize=8)
ax.spines[:].set_visible(False)
ax.set_title('Action frequency',color='#cdd6f4',fontsize=9,pad=3)
plt.tight_layout(pad=.4)
return _pil(fig)
def render_reward_chart(reward_history):
if len(reward_history)<2: return None
fig,ax=plt.subplots(figsize=(5,1.6))
fig.patch.set_facecolor('#1e1e2e'); ax.set_facecolor('#1e1e2e')
for i,r in enumerate(reward_history):
col='#ffd700' if r>=5 else ('#50c050' if r>0 else '#e05050')
ax.bar(i,r,color=col,width=1,alpha=0.8)
ax.axhline(0,color='#555',lw=0.5)
ax.set_xlim(0,len(reward_history))
ax.tick_params(colors='#888',labelsize=7); ax.spines[:].set_visible(False)
ax.set_title('Reward 🟡=level-up 🟢=change 🔴=dead',
color='#cdd6f4',fontsize=8,pad=3)
plt.tight_layout(pad=.3)
return _pil(fig)
def render_gabor_panel(grid):
"""
Visualize the top-responding Gabor channels — the s-plane cross terms.
Shows which (σ, ω, θ) combination is most active on the current frame.
"""
if grid is None: return None
feats = extract_gabor_features(grid) # (72, 64, 64)
max_per_ch = feats.abs().amax(dim=(1,2)).numpy()
top4_idx = max_per_ch.argsort()[-4:][::-1]
fig, axes = plt.subplots(1, 4, figsize=(10, 2.2))
fig.patch.set_facecolor('#1e1e2e')
# Build channel labels
labels = []
for s in SIGMA_VALUES:
for o in OMEGA_VALUES:
for t in THETA_VALUES:
for p in PHASE_VALUES:
td = int(t*180/np.pi)
pn = 'cos' if p==0 else 'sin'
labels.append(f's={s:.1f}+{o:.1f}i th={td} {pn}')
for ax, idx in zip(axes, top4_idx):
ax.set_facecolor('#0d0d1a')
ch_map = feats[idx].numpy()
im = ax.imshow(ch_map, cmap='RdBu', vmin=-1, vmax=1,
interpolation='nearest', aspect='equal')
ax.set_title(labels[idx], color='#cdd6f4', fontsize=7, pad=2)
ax.axis('off')
plt.colorbar(im, ax=ax, fraction=.06, pad=.02)
fig.suptitle('Gabor s-plane responses (top 4 cross-term channels)',
color='#cdd6f4', fontsize=9, y=1.02)
plt.tight_layout(pad=0.5)
return _pil(fig)
# ── TinyAgent with Re/Im solver ───────────────────────────────────────────────
CONF_THRESHOLD = 0.72 # high bar — only act analytically when very sure
class TinyAgent:
def __init__(self):
self.device=torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.model=self._make_model().to(self.device)
self.opt=torch.optim.Adam(self.model.parameters(),lr=1e-4)
self.buf=[]; self.prev_feat=None; self.prev_action=None
self.step_count=0; self.action_counts={}; self.prev_levels=0
self.reward_history=deque(maxlen=300)
self.level_history=[]; self.prev_state=None
self.level_up_reward=10.0; self.win_reward=50.0
self.near_win_reward=2.0; self.change_reward=0.1
self.dead_penalty=-0.01; self.candidate_win_reward=30.0
self.prev_candidate_dist=1.0
self.explore_steps=20
self.click_attempts=0; self.click_successes=0
self._last_was_click=False
def _make_model(self):
return nn.Sequential(
nn.Conv2d(14+N_GABOR_CHANNELS,32,3,padding=1),nn.ReLU(),
nn.Conv2d(32,64,3,padding=1),nn.ReLU(),
nn.Conv2d(64,128,3,padding=1),nn.ReLU(),
nn.AdaptiveAvgPool2d(8),nn.Flatten(),
nn.Linear(128*8*8,256),nn.ReLU(),
nn.Linear(256,6),
)
def reset(self):
self.model=self._make_model().to(self.device)
self.opt=torch.optim.Adam(self.model.parameters(),lr=1e-4)
self.buf=[]; self.prev_feat=None; self.prev_action=None
self.step_count=0; self.action_counts={}; self.prev_levels=0
self.reward_history=deque(maxlen=300); self.level_history=[]
self.prev_state=None; self.prev_candidate_dist=1.0
self.explore_steps=20; self.click_attempts=0; self.click_successes=0
def choose(self,grid,available_actions=None,levels=0,state=None):
feat=extract_features(grid).to(self.device)
cur_state=str(state) if state else None
# ── Re/Im: read the board using the 56-channel feature tensor ──────
# Pass feat directly so read_board reads Im channels from the same
# tensor the CNN uses — Im tensor=signal, raw grid=coordinates
# Concatenate axis + Gabor cross-term channels
gabor_feat = extract_gabor_features(grid, grid_size=64)
full_feat = torch.cat([feat, gabor_feat.to(self.device)], dim=0)
feat_np = full_feat.cpu().numpy()
cand_answer,cand_conf,cand_reasoning,cand_signal=read_board(grid, feat_np)
# Candidate proximity bonus
if cand_answer is not None and cand_conf>=0.35:
curr_dist=(grid!=cand_answer).mean() if grid.shape==cand_answer.shape else 1.0
if curr_dist==0.0:
cand_bonus=self.candidate_win_reward
elif curr_dist<self.prev_candidate_dist:
cand_bonus=(self.prev_candidate_dist-curr_dist)*5.0
else:
cand_bonus=0.0
self.prev_candidate_dist=curr_dist
else:
cand_bonus=0.0; cand_answer=None
# Store shaped experience
if self.prev_feat is not None:
changed=not np.array_equal(
self.prev_feat.cpu().numpy(),full_feat.cpu().numpy())
if self._last_was_click:
self.click_attempts+=1
if changed: self.click_successes+=1
just_won=(cur_state=='WIN' and self.prev_state!='WIN')
level_up=levels>self.prev_levels
if just_won:
reward=self.win_reward+cand_bonus
for i in range(min(5,len(self.buf))):
idx=len(self.buf)-1-i
self.buf[idx]=(self.buf[idx][0],self.buf[idx][1],
self.buf[idx][2]+self.near_win_reward*(1-i*0.15))
elif level_up:
reward=self.level_up_reward+cand_bonus
self.level_history.append((self.step_count,levels))
elif changed:
reward=self.change_reward+cand_bonus
else:
reward=self.dead_penalty+cand_bonus
self.reward_history.append(reward)
self.buf.append((self.prev_feat,self.prev_action,reward)) # prev_feat is full_feat
if len(self.buf)>500: self.buf.pop(0)
self.prev_state=cur_state
self.prev_levels=levels
if self.step_count%10==0 and len(self.buf)>=16:
self._train()
# ── Im → Re bridge: read board → derive answer → click exact cell ──
# Only fire after explore_steps of CNN exploration so we have context
analytic_action=None; analytic_meta={}
click_rate=(self.click_successes/self.click_attempts
if self.click_attempts>10 else 1.0)
if (cand_answer is not None
and cand_conf>=CONF_THRESHOLD
and self.step_count>self.explore_steps
and click_rate>=0.20):
diffs=pixel_diff(grid,cand_answer)
if diffs:
cell=most_urgent_diff(grid,cand_answer)
if cell is not None:
r,c,tgt_color=cell
H,W=grid.shape
gy=min(63,max(0,int(r*64/H+32/H)))
gx=min(63,max(0,int(c*64/W+32/W)))
analytic_action=6
reasoning_str=' | '.join(cand_reasoning[:2]) if cand_reasoning else 'read_board'
analytic_meta={'x':gx,'y':gy,'cell':(r,c,tgt_color),
'hypothesis':reasoning_str,'conf':cand_conf,
'n_diffs':len(diffs),
'candidates':[(reasoning_str,cand_answer,cand_conf)]}
# ── CNN fallback ──────────────────────────────────────────────────
with torch.no_grad():
logits=self.model(full_feat.unsqueeze(0)).squeeze(0)
avail=list(range(1,7))
if available_actions:
avail=[int(a.value if hasattr(a,'value') else a)
for a in available_actions if
int(a.value if hasattr(a,'value') else a)<=6]
indices=[m-1 for m in avail if 1<=m<=6]
masked=torch.full((6,),float('-inf'))
for i in indices: masked[i]=logits[i]
probs=torch.softmax(masked,dim=0).cpu().numpy()
probs=np.nan_to_num(probs,nan=0)
if probs.sum()==0: probs[np.array(indices)]=1/len(indices)
probs=probs/probs.sum()
cnn_action_idx=np.random.choice(6,p=probs)
# Pick final action
if analytic_action is not None:
chosen_id=analytic_action
meta=analytic_meta
meta['source']='analytic'
else:
chosen_id=cnn_action_idx+1
# Package read_board result for display even in CNN fallback
cnn_cands=[(cand_signal, cand_answer, cand_conf)] if cand_answer is not None else []
meta={'source':'cnn','probs':probs.tolist(),'candidates':cnn_cands,
'hypothesis':cand_reasoning[0][:60] if cand_reasoning else 'none',
'conf':cand_conf,'n_diffs':len(pixel_diff(grid,cand_answer)) if cand_answer is not None else 0}
self.prev_feat=full_feat; self.prev_action=cnn_action_idx
self._last_was_click=(chosen_id==6)
self.step_count+=1
a_id=chosen_id
self.action_counts[a_id]=self.action_counts.get(a_id,0)+1
try:
from arcengine import GameAction
action=GameAction(a_id)
except Exception:
action=a_id
if a_id==6 and 'x' in meta:
try: action.set_data({'x':meta['x'],'y':meta['y']})
except: pass
return action,meta
def _train(self):
import random
batch=random.sample(self.buf,min(16,len(self.buf)))
states =torch.stack([b[0] for b in batch]).to(self.device)
actions=torch.tensor([b[1] for b in batch],dtype=torch.long, device=self.device)
rewards=torch.tensor([b[2] for b in batch],dtype=torch.float32,device=self.device)
self.opt.zero_grad()
logits=self.model(states)
loss=TF.binary_cross_entropy_with_logits(
logits.gather(1,actions.unsqueeze(1)).squeeze(1),
torch.clamp(rewards,0,1))
loss.backward(); self.opt.step()
# ── Session ───────────────────────────────────────────────────────────────────
_agent = TinyAgent()
_stop_flag = threading.Event()
_run_thread = None
_frame_queue= queue.Queue(maxsize=60)
def _run_agent(game_id,api_key,max_steps):
import arc_agi
try:
arc=arc_agi.Arcade(arc_api_key=api_key)
env=arc.make(game_id,include_frame_data=True)
frame=env.reset(); _agent.reset()
prev_grid=None; step=0
while not _stop_flag.is_set() and step<max_steps:
if frame is None: break
raw=np.array(frame.frame,dtype=np.int64)
grid=raw[-1] if raw.ndim==3 else raw
avail=getattr(frame,'available_actions',None)
levels=getattr(frame,'levels_completed',0)
state=getattr(frame,'state',None)
action,meta=_agent.choose(grid,avail,levels=levels,state=state)
diff=(grid!=prev_grid) if prev_grid is not None else None
prev_grid=grid.copy()
_frame_queue.put({
'grid':grid,'diff':diff,'step':step,
'action':int(action.value if hasattr(action,'value') else action),
'levels':levels,'state':str(state),
'meta':meta,
'counts':dict(_agent.action_counts),'click_rate':round(_agent.click_successes/_agent.click_attempts,2) if _agent.click_attempts>0 else None,
'reward_history':list(_agent.reward_history),
'grid_raw':grid.tolist(),
'level_history':list(_agent.level_history),
},block=True,timeout=5)
state_str=str(state)
if 'WIN' in state_str or 'GAME_OVER' in state_str: break
try:
from arcengine import GameAction as GA
a_int=int(action.value if hasattr(action,'value') else action)
sa=GA(a_int)
if meta.get('x') is not None:
try: sa.set_data({'x':int(meta['x']),'y':int(meta['y'])})
except: pass
frame=env.step(sa)
except Exception as step_err:
# Last resort: try passing action directly
try: frame=env.step(action)
except: frame=None
step+=1
time.sleep(0.08)
_frame_queue.put({'done':True,'step':step,
'level_history':list(_agent.level_history)})
except Exception as e:
_frame_queue.put({'error':str(e)})
# ── Pull frame ────────────────────────────────────────────────────────────────
_latest={'grid_img':None,'hyp_img':None,'cand_img':None,
'bar_img':None,'reward_img':None,'status':'*Waiting...*'}
def pull_frame():
global _latest
data=None
while True:
try: data=_frame_queue.get_nowait()
except queue.Empty: break
if data is None:
return (_latest['grid_img'],_latest['hyp_img'],_latest['cand_img'],
_latest['bar_img'],_latest['gabor_img'],_latest['reward_img'],_latest['status'])
if 'error' in data:
_latest['status']=f"**Error:** {data['error']}"
return (_latest['grid_img'],_latest['hyp_img'],_latest['cand_img'],
_latest['bar_img'],_latest['gabor_img'],_latest['reward_img'],_latest['status'])
if data.get('done'):
lh=data.get('level_history',[])
_latest['status']=f"**Done** — {data['step']} steps | {len(lh)} levels completed"
return (_latest['grid_img'],_latest['hyp_img'],_latest['cand_img'],
_latest['bar_img'],_latest['gabor_img'],_latest['reward_img'],_latest['status'])
grid=data['grid']; meta=data['meta']; step=data['step']
levels=data['levels']; state=data['state']; action=data['action']
source=meta.get('source','cnn')
hyp_str=meta.get('hypothesis','none')
cand_conf=meta.get('conf',0.0)
n_diffs=meta.get('n_diffs',0)
# Build display candidates list: [(label, grid, conf)]
raw_cands=meta.get('candidates',[])
# raw_cands entries are (reasoning_str, grid, conf)
# Normalise to (short_label, grid, conf)
disp_cands=[]
for entry in raw_cands:
if len(entry)==3:
label=str(entry[0])[:40]; cgrid=entry[1]; cconf=entry[2]
if isinstance(cgrid, np.ndarray): disp_cands.append((label,cgrid,cconf))
# Determine what to highlight
mark_cell=None; highlight=None
if source=='analytic' and 'cell' in meta and disp_cands:
_,cand_grid,_=disp_cands[0]
if cand_grid.shape==grid.shape:
all_diffs=pixel_diff(grid,cand_grid)
highlight=all_diffs[:20]
mark_cell=meta['cell']
source_emoji='🧠' if source=='analytic' else '🎲'
_latest['grid_img']=render_grid(
grid,
title=f"Step {step} | {source_emoji} A{action} | Levels {levels}",
highlight=highlight,
mark_cell=mark_cell)
# Im side: hypothesis panel — show read_board reasoning as bar
if disp_cands:
# Convert to format render_hypothesis_panel expects: [(name,grid,conf)]
_latest['hyp_img']=render_hypothesis_panel(disp_cands)
else:
_latest['hyp_img']=None
# Re side: candidate grid
if disp_cands and disp_cands[0][1].shape==grid.shape:
cname,cgrid,cconf2=disp_cands[0]
diffs=pixel_diff(grid,cgrid)
_latest['cand_img']=render_grid(
cgrid,
title=f"Im answer: {cname[:35]} (conf={cconf2:.2f}) — {len(diffs)} cells differ",
highlight=diffs[:20])
else:
_latest['cand_img']=None
_latest['bar_img'] =render_action_bar(data['counts'],sum(data['counts'].values()))
_latest['reward_img']=render_reward_chart(data['reward_history'])
grid_raw=np.array(data.get('grid_raw',grid.tolist()),dtype=np.int64)
_latest['gabor_img']=render_gabor_panel(grid_raw)
last_r=data['reward_history'][-1] if data['reward_history'] else 0
r_emoji='🟡' if last_r>=5 else ('🟢' if last_r>0 else '🔴')
hyp_str=(f"`{meta.get('hypothesis','?')}` conf={meta.get('conf',0):.2f} "
f"→ click ({meta.get('x','?')},{meta.get('y','?')}) "
f"[{meta.get('n_diffs','?')} cells wrong]"
if source=='analytic'
else f"CNN probs: {[round(p,2) for p in meta.get('probs',[])]}")
_latest['status']=(
f"{source_emoji} **{'Analytic (Re/Im)' if source=='analytic' else 'CNN fallback'}**"
f" | Step {step} | Levels {levels}"
f" | Reward {r_emoji} `{last_r:.2f}` | {state}\n\n"
f"{hyp_str}")
return (_latest['grid_img'],_latest['hyp_img'],_latest['cand_img'],
_latest['bar_img'],_latest['reward_img'],_latest['status'])
# ── Handlers ──────────────────────────────────────────────────────────────────
def fetch_games(api_key):
try:
import arc_agi
arc=arc_agi.Arcade(arc_api_key=api_key)
envs=arc.get_environments()
ids=[e.game_id for e in envs]
return gr.Dropdown(choices=ids,value=ids[0] if ids else None),\
f"Found **{len(ids)}** games."
except Exception as e:
return gr.Dropdown(choices=[]),f"**Error:** {e}"
def start_agent(game_id,api_key,max_steps):
global _run_thread,_stop_flag
if not game_id: return "Select a game first."
if not api_key: return "Enter your API key."
_stop_flag.set()
if _run_thread and _run_thread.is_alive(): _run_thread.join(timeout=3)
while not _frame_queue.empty():
try: _frame_queue.get_nowait()
except: break
_stop_flag.clear()
_run_thread=threading.Thread(
target=_run_agent,args=(game_id,api_key,int(max_steps)),daemon=True)
_run_thread.start()
return f"Agent started on **{game_id}** — 🧠 Re/Im analytic + 🎲 CNN fallback"
def stop_agent():
_stop_flag.set()
return "Agent stopped."
# ── UI ────────────────────────────────────────────────────────────────────────
with gr.Blocks(title="ARC-AGI-3 Re/Im Agent") as demo:
gr.Markdown("""
# ARC-AGI-3 Re/Im Agent Spectator
**Im side** = bird's eye hypothesis (which transformation?) |
**Re side** = exact location (which cells to click?)
🧠 = analytic solver (Im picks hypothesis → Re pins cell → ACTION6 click)
🎲 = CNN fallback (when no hypothesis clears the confidence threshold)
""")
with gr.Row():
with gr.Column(scale=3):
api_box=gr.Textbox(label="ARC API key",type="password",
value=os.environ.get("ARC_API_KEY",""),
placeholder="arc-key-... or set ARC_API_KEY secret")
with gr.Column(scale=1):
fetch_btn=gr.Button("Fetch games")
with gr.Row():
with gr.Column(scale=2):
game_dd=gr.Dropdown(label="Game",choices=[])
with gr.Column(scale=1):
steps_sl=gr.Slider(label="Max steps",minimum=20,maximum=500,value=150,step=10)
with gr.Column(scale=1):
with gr.Row():
start_btn=gr.Button("▶ Watch",variant="primary")
stop_btn =gr.Button("■ Stop")
run_status=gr.Markdown("*Fetch games → select → Watch*")
api_status=gr.Markdown()
gr.Markdown("---")
with gr.Row():
grid_img=gr.Image(label="Current frame (🔴=wrong cells ⭐=target click)",
type="pil",interactive=False,height=280)
hyp_img =gr.Image(label="Im side — hypothesis ranking",
type="pil",interactive=False,height=280)
with gr.Row():
cand_img =gr.Image(label="Im candidate — what the answer should look like",
type="pil",interactive=False,height=220)
bar_img =gr.Image(label="Action frequency",
type="pil",interactive=False,height=220)
with gr.Row():
gabor_img =gr.Image(label="Gabor s-plane — cross terms (σ>0, ω>0)",
type="pil",interactive=False,height=160)
reward_img=gr.Image(label="Reward 🟡+50 WIN 🟡+10 level 🟢+0.1 change 🔴-0.01 dead",
type="pil",interactive=False,height=160)
timer=gr.Timer(value=1.0)
timer.tick(pull_frame,
outputs=[grid_img,hyp_img,cand_img,bar_img,gabor_img,reward_img,run_status])
fetch_btn.click(fetch_games,inputs=api_box,outputs=[game_dd,api_status])
start_btn.click(start_agent,inputs=[game_dd,api_box,steps_sl],outputs=run_status)
stop_btn.click(stop_agent,outputs=run_status)
gr.Markdown("""
---
**Re/Im duality in action:**
The Im side reads the whole board at once — symmetry maps, boundary contour, directional
flow — and ranks candidate transformations by confidence.
The Re side then diffs the current frame against the winning candidate and finds the exact
cell (boundary-first, following Cauchy's principle) that most needs fixing.
The agent emits ACTION6 at those precise coordinates instead of guessing randomly.
CNN fires only when no analytic hypothesis clears 0.40 confidence.
""")
if __name__ == "__main__":
demo.launch()
|