| ================================================================= |
| DUAL-STREAM VIT GEOMETRIC DIAGNOSTIC |
| Checkpoint: /content/checkpoints/dual_stream_v3_e100.pt |
| ================================================================= |
|
|
| Epoch: 100 Val acc: 85.7% |
| Streams: 192-d × 2, 2 dual blocks |
| Fused: 256-d, 4 fused blocks |
| Constellation: 64 anchors × 128-d |
|
|
| Extracting embeddings... |
| Embeddings: torch.Size([10000, 128]) |
| Accuracy: 85.7% |
| Dual blocks: 2 |
| Patches per image: 64 |
|
|
| ================================================================= |
| SCAN 1: EMBEDDING HEALTH |
| ================================================================= |
| Norms: mean=1.000000 std=0.000000 |
| Self-similarity: mean=0.0945 std=0.1840 |
| ✓ No embedding collapse |
|
|
| Effective dim: 69.9/128 |
| top-3 SVs explain 20.3% |
| top-5 SVs explain 27.9% |
| top-10 SVs explain 40.9% |
| top-20 SVs explain 57.7% |
| top-50 SVs explain 91.2% |
|
|
| Pentachoron CV (GeoLIP structural spectrum): |
| global embedding: CV=0.1876 (✓ IN BAND) [500/500 valid, mean_vol=0.069911] |
| Per-class: |
| airplane : CV=0.2340 (✓ IN BAND) [200/200 valid, mean_vol=0.038247] |
| automobile: CV=0.2976 (✗ outside) [200/200 valid, mean_vol=0.041421] |
| bird : CV=0.2056 (✓ IN BAND) [200/200 valid, mean_vol=0.038283] |
| cat : CV=0.1892 (✓ IN BAND) [200/200 valid, mean_vol=0.042475] |
| deer : CV=0.1983 (✓ IN BAND) [200/200 valid, mean_vol=0.039594] |
| dog : CV=0.1847 (✓ IN BAND) [200/200 valid, mean_vol=0.041361] |
| frog : CV=0.2542 (✗ outside) [200/200 valid, mean_vol=0.028879] |
| horse : CV=0.2386 (✓ IN BAND) [200/200 valid, mean_vol=0.033229] |
| ship : CV=0.2906 (✗ outside) [200/200 valid, mean_vol=0.037988] |
| truck : CV=0.2676 (✗ outside) [200/200 valid, mean_vol=0.041951] |
| Class CV: mean=0.2360 std=0.0412 range=[0.1847, 0.2976] |
| anchor constellation: CV=0.0690 (✗ outside) [200/200 valid, mean_vol=0.091371] |
|
|
| Patch-level CV (from fused patch projections): |
| all patches (flat): CV=0.1615 (✗ outside) [500/500 valid, mean_vol=0.075994] |
|
|
| ================================================================= |
| SCAN 2: ANCHOR HEALTH |
| ================================================================= |
| Anchor pairwise cos: mean=-0.0067 max=0.3794 |
| Anchor effective rank: 52.8/128 |
|
|
| Pooled anchors active: 64/64 |
| Patch anchors active: 64/64 |
| Per-image patch anchors: mean=15.4 min=1 max=33 |
| Entropy: 3.5735/4.1589 (86%) |
| Gini: 0.5843 |
|
|
| ================================================================= |
| SCAN 3: CLASS GEOMETRY |
| ================================================================= |
| Inter-class cos: mean=0.2045 max=0.7968 min=-0.3748 |
|
|
| Most similar: |
| cat × dog : cos=0.7968 |
| bird × cat : cos=0.6555 |
| cat × deer : cos=0.6508 |
| cat × frog : cos=0.6479 |
| bird × deer : cos=0.6296 |
| Most distant: |
| automobile × deer : cos=-0.3748 |
| automobile × dog : cos=-0.3692 |
| automobile × cat : cos=-0.3450 |
|
|
| Intra-class spread: |
| airplane : spread=0.4279 |
| automobile : spread=0.4745 |
| bird : spread=0.4194 |
| cat : spread=0.4577 |
| deer : spread=0.4244 |
| dog : spread=0.4450 |
| frog : spread=0.3623 |
| horse : spread=0.3951 |
| ship : spread=0.4391 |
| truck : spread=0.4507 |
|
|
| ================================================================= |
| SCAN 4: KSIMPLEX GEOMETRIC FEATURES |
| ================================================================= |
|
|
| ── Dual Block 0 ── |
| Shape: torch.Size([10000, 64, 11]) |
| d² pairs: mean=0.3292 std=0.2917 min=-0.4463 max=1.7068 |
| vol² feat: mean=-3.2233 std=0.0837 |
| Per-feature means: |
| d²_0: mean=0.6046 std=0.4390 |
| d²_1: mean=0.5711 std=0.3474 |
| d²_2: mean=0.3827 std=0.1808 |
| d²_3: mean=0.3508 std=0.1710 |
| d²_4: mean=0.5717 std=0.1762 |
| d²_5: mean=0.1650 std=0.1221 |
| d²_6: mean=0.1374 std=0.1700 |
| d²_7: mean=0.2348 std=0.1172 |
| d²_8: mean=0.1428 std=0.1575 |
| d²_9: mean=0.1314 std=0.1386 |
| vol²: mean=-3.2233 std=0.0837 |
|
|
| Most discriminative features (by class variance): |
| d²_4: var=0.004094 |
| d²_9: var=0.000866 |
| d²_8: var=0.000674 |
| d²_3: var=0.000505 |
| d²_0: var=0.000451 |
|
|
| Per-class profiles (11-d mean): |
| airplane : norm=3.4447 strongest=vol²(-3.2441) |
| automobile : norm=3.4314 strongest=vol²(-3.2220) |
| bird : norm=3.4395 strongest=vol²(-3.2214) |
| cat : norm=3.4360 strongest=vol²(-3.2123) |
| deer : norm=3.4418 strongest=vol²(-3.2207) |
| dog : norm=3.4386 strongest=vol²(-3.2165) |
| frog : norm=3.4447 strongest=vol²(-3.2209) |
| horse : norm=3.4405 strongest=vol²(-3.2228) |
| ship : norm=3.4379 strongest=vol²(-3.2337) |
| truck : norm=3.4300 strongest=vol²(-3.2184) |
|
|
| Geo-only class separation (cos on 11-d): |
| mean=0.9994 max=1.0000 min=0.9977 |
| Most similar (geo): |
| cat × dog : cos=1.0000 |
| automobile × truck : cos=1.0000 |
| bird × deer : cos=1.0000 |
| Most distinct (geo): |
| airplane × frog : cos=0.9977 |
| airplane × deer : cos=0.9980 |
| frog × ship : cos=0.9982 |
|
|
| ── Dual Block 1 ── |
| Shape: torch.Size([10000, 64, 11]) |
| d² pairs: mean=0.3135 std=0.2532 min=-0.5581 max=1.5650 |
| vol² feat: mean=-3.1620 std=0.0708 |
| Per-feature means: |
| d²_0: mean=0.3060 std=0.1752 |
| d²_1: mean=0.1853 std=0.2129 |
| d²_2: mean=0.2038 std=0.0966 |
| d²_3: mean=0.2170 std=0.2203 |
| d²_4: mean=0.3453 std=0.1957 |
| d²_5: mean=0.7091 std=0.3372 |
| d²_6: mean=0.1514 std=0.1258 |
| d²_7: mean=0.3575 std=0.1338 |
| d²_8: mean=0.2761 std=0.2679 |
| d²_9: mean=0.3837 std=0.1413 |
| vol²: mean=-3.1620 std=0.0708 |
|
|
| Most discriminative features (by class variance): |
| d²_8: var=0.000794 |
| d²_6: var=0.000709 |
| d²_0: var=0.000679 |
| d²_4: var=0.000605 |
| d²_9: var=0.000563 |
|
|
| Per-class profiles (11-d mean): |
| airplane : norm=3.3541 strongest=vol²(-3.1746) |
| automobile : norm=3.3449 strongest=vol²(-3.1601) |
| bird : norm=3.3501 strongest=vol²(-3.1619) |
| cat : norm=3.3458 strongest=vol²(-3.1565) |
| deer : norm=3.3503 strongest=vol²(-3.1589) |
| dog : norm=3.3468 strongest=vol²(-3.1568) |
| frog : norm=3.3497 strongest=vol²(-3.1633) |
| horse : norm=3.3477 strongest=vol²(-3.1589) |
| ship : norm=3.3526 strongest=vol²(-3.1670) |
| truck : norm=3.3471 strongest=vol²(-3.1624) |
|
|
| Geo-only class separation (cos on 11-d): |
| mean=0.9996 max=1.0000 min=0.9990 |
| Most similar (geo): |
| cat × dog : cos=1.0000 |
| automobile × truck : cos=1.0000 |
| cat × horse : cos=1.0000 |
| Most distinct (geo): |
| airplane × deer : cos=0.9990 |
| frog × ship : cos=0.9990 |
| deer × ship : cos=0.9991 |
|
|
| ── Geo Evolution Across Blocks ── |
| Block 0: feat_var=0.000698 class_sep=0.0006 |
| Block 1: feat_var=0.000391 class_sep=0.0004 |
|
|
| ================================================================= |
| SCAN 4B: CAYLEY-MENGER VOLUME² PER DUAL BLOCK |
| ================================================================= |
|
|
| ── Dual Block 0 ── |
| Overall valid: 100.0% |
| vol² mean=0.00059389 std=0.00004124 |
| vol² min=0.00045395 max=0.00080490 |
| vol² median=0.00058746 |
| Images with 100% valid patches: 10000/10000 (100.0%) |
| Per-image valid fraction: mean=1.0000 min=1.0000 |
|
|
| Positive vol² distribution (log10): |
| mean=-3.2273 std=0.0299 |
| p 1: -3.2947 (vol²=0.00050735) |
| p10: -3.2632 (vol²=0.00054550) |
| p25: -3.2483 (vol²=0.00056458) |
| p50: -3.2310 (vol²=0.00058746) |
| p75: -3.2090 (vol²=0.00061798) |
| p90: -3.1855 (vol²=0.00065231) |
| p99: -3.1537 (vol²=0.00070190) |
|
|
| Volume CV: 0.0694 (target band: 0.20-0.23) |
|
|
| Per-class vol²: |
| airplane : mean=0.00060517 std=0.00004063 valid=100.0% |
| automobile : mean=0.00059402 std=0.00004709 valid=100.0% |
| bird : mean=0.00059359 std=0.00003650 valid=100.0% |
| cat : mean=0.00058860 std=0.00003941 valid=100.0% |
| deer : mean=0.00058860 std=0.00003440 valid=100.0% |
| dog : mean=0.00059095 std=0.00003881 valid=100.0% |
| frog : mean=0.00058534 std=0.00003506 valid=100.0% |
| horse : mean=0.00059156 std=0.00004011 valid=100.0% |
| ship : mean=0.00060546 std=0.00004353 valid=100.0% |
| truck : mean=0.00059559 std=0.00004926 valid=100.0% |
|
|
| ✓ Zero negative volumes — all simplices valid |
|
|
| ── Dual Block 1 ── |
| Overall valid: 100.0% |
| vol² mean=0.00060576 std=0.00005404 |
| vol² min=0.00043488 max=0.00091171 |
| vol² median=0.00059891 |
| Images with 100% valid patches: 10000/10000 (100.0%) |
| Per-image valid fraction: mean=1.0000 min=1.0000 |
|
|
| Positive vol² distribution (log10): |
| mean=-3.2194 std=0.0382 |
| p 1: -3.2980 (vol²=0.00050354) |
| p10: -3.2663 (vol²=0.00054169) |
| p25: -3.2454 (vol²=0.00056839) |
| p50: -3.2226 (vol²=0.00059891) |
| p75: -3.1932 (vol²=0.00064087) |
| p90: -3.1681 (vol²=0.00067902) |
| p99: -3.1241 (vol²=0.00075150) |
|
|
| Volume CV: 0.0892 (target band: 0.20-0.23) |
|
|
| Per-class vol²: |
| airplane : mean=0.00059386 std=0.00005703 valid=100.0% |
| automobile : mean=0.00060279 std=0.00005654 valid=100.0% |
| bird : mean=0.00060519 std=0.00005180 valid=100.0% |
| cat : mean=0.00061181 std=0.00005005 valid=100.0% |
| deer : mean=0.00060359 std=0.00004848 valid=100.0% |
| dog : mean=0.00061371 std=0.00005029 valid=100.0% |
| frog : mean=0.00061009 std=0.00004580 valid=100.0% |
| horse : mean=0.00061714 std=0.00005574 valid=100.0% |
| ship : mean=0.00059085 std=0.00005562 valid=100.0% |
| truck : mean=0.00060854 std=0.00006128 valid=100.0% |
|
|
| ✓ Zero negative volumes — all simplices valid |
|
|
| ── Volume Evolution Across Blocks ── |
| Block 0: mean=0.00059389 std=0.00004124 valid=100.0% |
| Block 1: mean=0.00060576 std=0.00005404 valid=100.0% |
| Block 1/Block 0 ratio: 1.0200 |
| ✓ Volume ratio stable |
|
|
| ================================================================= |
| SCAN 5: PER-CLASS ACCURACY |
| ================================================================= |
| airplane : acc= 87.2% P(correct)=0.844 confuser=ship(0.038) |
| automobile : acc= 91.7% P(correct)=0.911 confuser=truck(0.050) |
| bird : acc= 82.5% P(correct)=0.800 confuser=deer(0.050) |
| cat : acc= 69.4% P(correct)=0.645 confuser=dog(0.154) |
| deer : acc= 84.9% P(correct)=0.809 confuser=horse(0.043) |
| dog : acc= 78.1% P(correct)=0.760 confuser=cat(0.130) |
| frog : acc= 90.6% P(correct)=0.890 confuser=cat(0.033) |
| horse : acc= 89.0% P(correct)=0.871 confuser=deer(0.037) |
| ship : acc= 92.5% P(correct)=0.911 confuser=airplane(0.038) |
| truck : acc= 91.2% P(correct)=0.895 confuser=automobile(0.047) |
|
|
| ================================================================= |
| SCAN 6: EMBEDDING DISCRIMINATION |
| ================================================================= |
| Same-class cos: mean=0.3335 std=0.1510 |
| Diff-class cos: mean=0.0742 std=0.1662 |
| Gap (same-diff): 0.2593 |
| 1-NN accuracy: 77.4% |
|
|
| ================================================================= |
| SCAN 7: CAYLEY-MENGER VERIFICATION |
| ================================================================= |
| CM volumes (200 samples): pos=200 neg=0 zero=0 |
| Norms: mean=1.000000 |
|
|
| ================================================================= |
| SCAN 8: ARCHITECTURE SUMMARY |
| ================================================================= |
| Total params: 6,321,542 |
| Geo stream params: 1,306,300 (20.7%) |
| Std stream params: 1,261,248 (20.0%) |
| Fused block params: 3,159,552 (50.0%) |
| Constellation params:84,480 (1.3%) |
|
|
| ================================================================= |
| DIAGNOSIS SUMMARY |
| ================================================================= |
| Val accuracy: 85.7% |
| Eff dim: 69.9/128 |
| Pentachoron CV: 0.1876 (target band: 0.20-0.23) |
| Self-similarity: 0.0945 |
| Pooled anchors: 64/64 |
| Patch anchors: 64/64 |
| Per-img p_anch: 15.4 |
| Entropy: 86% |
| Gini: 0.5843 |
| CM volumes: 200/200 positive |
| Anchor CV: 0.0690 |
| Class CV range: [0.1847, 0.2976] |
| Geo feat var: 0.000391 |
| Block 0 CM valid: 100.0% |
| Block 1 CM valid: 100.0% |
| Same/diff gap: 0.2593 |
| 1-NN accuracy: 77.4% |
|
|
| ✓ No major issues. Geometry is healthy. |
|
|
| ================================================================= |
| DIAGNOSTIC COMPLETE |
| ================================================================= |
