| ================================================================= |
| DUAL-STREAM VIT GEOMETRIC DIAGNOSTIC |
| Checkpoint: /content/checkpoints/dual_stream_best.pt |
| ================================================================= |
|
|
| Epoch: 90 Val acc: 85.2% |
| 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: 85.2% |
| Dual blocks: 2 |
| Patches per image: 64 |
|
|
| ================================================================= |
| SCAN 1: EMBEDDING HEALTH |
| ================================================================= |
| Norms: mean=1.000000 std=0.000000 |
| Self-similarity: mean=0.0027 std=0.1443 |
| ✓ No embedding collapse |
|
|
| Effective dim: 58.3/128 |
| top-3 SVs explain 9.0% |
| top-5 SVs explain 14.7% |
| top-10 SVs explain 27.7% |
| top-20 SVs explain 50.9% |
| top-50 SVs explain 98.5% |
|
|
| Pentachoron CV (GeoLIP structural spectrum): |
| global embedding: CV=0.1061 (✗ outside) [500/500 valid, mean_vol=0.086509] |
| Per-class: |
| airplane : CV=0.1405 (✗ outside) [200/200 valid, mean_vol=0.072228] |
| automobile: CV=0.1510 (✗ outside) [200/200 valid, mean_vol=0.067334] |
| bird : CV=0.1404 (✗ outside) [200/200 valid, mean_vol=0.073100] |
| cat : CV=0.1161 (✗ outside) [200/200 valid, mean_vol=0.075927] |
| deer : CV=0.1395 (✗ outside) [200/200 valid, mean_vol=0.071760] |
| dog : CV=0.1401 (✗ outside) [200/200 valid, mean_vol=0.071994] |
| frog : CV=0.1483 (✗ outside) [200/200 valid, mean_vol=0.069985] |
| horse : CV=0.1516 (✗ outside) [200/200 valid, mean_vol=0.070745] |
| ship : CV=0.1484 (✗ outside) [200/200 valid, mean_vol=0.070321] |
| truck : CV=0.1518 (✗ outside) [200/200 valid, mean_vol=0.068549] |
| Class CV: mean=0.1428 std=0.0107 range=[0.1161, 0.1518] |
| anchor constellation: CV=0.1324 (✗ outside) [200/200 valid, mean_vol=0.083785] |
|
|
| Patch-level CV (from fused patch projections): |
| all patches (flat): CV=0.1768 (✗ outside) [500/500 valid, mean_vol=0.082353] |
|
|
| ================================================================= |
| SCAN 2: ANCHOR HEALTH |
| ================================================================= |
| Anchor pairwise cos: mean=-0.0093 max=0.7289 |
| Anchor effective rank: 42.4/128 |
|
|
| Pooled anchors active: 64/64 |
| Patch anchors active: 64/64 |
| Per-image patch anchors: mean=15.2 min=3 max=30 |
| Entropy: 4.0876/4.1589 (98%) |
| Gini: 0.2067 |
|
|
| ================================================================= |
| SCAN 3: CLASS GEOMETRY |
| ================================================================= |
| Inter-class cos: mean=-0.0694 max=0.3587 min=-0.2369 |
|
|
| Most similar: |
| cat × dog : cos=0.3587 |
| bird × deer : cos=0.0938 |
| airplane × bird : cos=0.0613 |
| deer × horse : cos=0.0483 |
| dog × horse : cos=0.0335 |
| Most distant: |
| airplane × cat : cos=-0.2369 |
| bird × truck : cos=-0.2213 |
| deer × truck : cos=-0.1930 |
|
|
| Intra-class spread: |
| airplane : spread=0.7098 |
| automobile : spread=0.6846 |
| bird : spread=0.7192 |
| cat : spread=0.7597 |
| deer : spread=0.7217 |
| dog : spread=0.7265 |
| frog : spread=0.6966 |
| horse : spread=0.6870 |
| ship : spread=0.6929 |
| truck : spread=0.6945 |
|
|
| ================================================================= |
| SCAN 4: KSIMPLEX GEOMETRIC FEATURES |
| ================================================================= |
|
|
| ── Dual Block 0 ── |
| Shape: torch.Size([10000, 64, 11]) |
| d² pairs: mean=0.2629 std=0.7643 min=-1.8340 max=2.8440 |
| vol² feat: mean=-2.3967 std=0.3077 |
| Per-feature means: |
| d²_0: mean=0.0507 std=0.4150 |
| d²_1: mean=0.6021 std=1.3390 |
| d²_2: mean=0.2357 std=0.1966 |
| d²_3: mean=0.1046 std=1.3593 |
| d²_4: mean=0.2869 std=0.4066 |
| d²_5: mean=0.1592 std=0.5049 |
| d²_6: mean=0.1341 std=0.3828 |
| d²_7: mean=0.4269 std=0.4806 |
| d²_8: mean=-0.0711 std=0.2976 |
| d²_9: mean=0.6997 std=0.7489 |
| vol²: mean=-2.3967 std=0.3077 |
|
|
| Most discriminative features (by class variance): |
| d²_3: var=0.005224 |
| d²_9: var=0.003744 |
| d²_8: var=0.001982 |
| d²_1: var=0.001830 |
| d²_4: var=0.001284 |
|
|
| Per-class profiles (11-d mean): |
| airplane : norm=2.6657 strongest=vol²(-2.4019) |
| automobile : norm=2.6744 strongest=vol²(-2.4322) |
| bird : norm=2.6326 strongest=vol²(-2.3826) |
| cat : norm=2.6174 strongest=vol²(-2.3756) |
| deer : norm=2.6352 strongest=vol²(-2.3883) |
| dog : norm=2.6064 strongest=vol²(-2.3653) |
| frog : norm=2.6573 strongest=vol²(-2.4181) |
| horse : norm=2.6208 strongest=vol²(-2.3829) |
| ship : norm=2.6659 strongest=vol²(-2.4003) |
| truck : norm=2.6711 strongest=vol²(-2.4194) |
|
|
| Geo-only class separation (cos on 11-d): |
| mean=0.9974 max=0.9998 min=0.9916 |
| Most similar (geo): |
| cat × dog : cos=0.9998 |
| bird × cat : cos=0.9997 |
| bird × dog : cos=0.9997 |
| Most distinct (geo): |
| airplane × deer : cos=0.9916 |
| airplane × frog : cos=0.9920 |
| deer × ship : cos=0.9933 |
|
|
| ── Dual Block 1 ── |
| Shape: torch.Size([10000, 64, 11]) |
| d² pairs: mean=0.2006 std=0.2409 min=-0.6670 max=1.5775 |
| vol² feat: mean=-2.9146 std=0.0714 |
| Per-feature means: |
| d²_0: mean=0.0490 std=0.1027 |
| d²_1: mean=0.0740 std=0.2235 |
| d²_2: mean=0.2175 std=0.2268 |
| d²_3: mean=0.1087 std=0.1417 |
| d²_4: mean=0.0356 std=0.0973 |
| d²_5: mean=0.4991 std=0.2450 |
| d²_6: mean=0.3985 std=0.2259 |
| d²_7: mean=0.1007 std=0.1207 |
| d²_8: mean=0.3431 std=0.2667 |
| d²_9: mean=0.1805 std=0.0913 |
| vol²: mean=-2.9146 std=0.0714 |
|
|
| Most discriminative features (by class variance): |
| d²_8: var=0.003017 |
| d²_5: var=0.001958 |
| d²_6: var=0.001224 |
| d²_2: var=0.001135 |
| d²_1: var=0.000490 |
|
|
| Per-class profiles (11-d mean): |
| airplane : norm=3.0273 strongest=vol²(-2.9128) |
| automobile : norm=3.0325 strongest=vol²(-2.9309) |
| bird : norm=3.0187 strongest=vol²(-2.9039) |
| cat : norm=3.0163 strongest=vol²(-2.9122) |
| deer : norm=3.0209 strongest=vol²(-2.9074) |
| dog : norm=3.0116 strongest=vol²(-2.9059) |
| frog : norm=3.0203 strongest=vol²(-2.9086) |
| horse : norm=3.0221 strongest=vol²(-2.9157) |
| ship : norm=3.0312 strongest=vol²(-2.9217) |
| truck : norm=3.0315 strongest=vol²(-2.9273) |
|
|
| Geo-only class separation (cos on 11-d): |
| mean=0.9990 max=0.9999 min=0.9977 |
| Most similar (geo): |
| cat × dog : cos=0.9999 |
| bird × deer : cos=0.9999 |
| automobile × truck : cos=0.9999 |
| Most distinct (geo): |
| automobile × deer : cos=0.9977 |
| deer × truck : cos=0.9977 |
| automobile × bird : cos=0.9979 |
|
|
| ── Geo Evolution Across Blocks ── |
| Block 0: feat_var=0.001698 class_sep=0.0026 |
| Block 1: feat_var=0.000838 class_sep=0.0010 |
|
|
| ================================================================= |
| SCAN 4B: CAYLEY-MENGER VOLUME² PER DUAL BLOCK |
| ================================================================= |
|
|
| ── Dual Block 0 ── |
| Overall valid: 100.0% |
| vol² mean=0.00054120 std=0.00061009 |
| vol² min=0.00004506 max=0.00595093 |
| vol² median=0.00030899 |
| Images with 100% valid patches: 10000/10000 (100.0%) |
| Per-image valid fraction: mean=1.0000 min=1.0000 |
|
|
| Positive vol² distribution (log10): |
| mean=-3.4331 std=0.3509 |
| p 1: -4.0316 (vol²=0.00009298) |
| p10: -3.8192 (vol²=0.00015163) |
| p25: -3.6861 (vol²=0.00020599) |
| p50: -3.5101 (vol²=0.00030899) |
| p75: -3.2254 (vol²=0.00059509) |
| p90: -2.9027 (vol²=0.00125122) |
| p99: -2.5047 (vol²=0.00312805) |
|
|
| Volume CV: 1.1273 (target band: 0.20-0.23) |
|
|
| Per-class vol²: |
| airplane : mean=0.00044444 std=0.00050958 valid=100.0% |
| automobile : mean=0.00062647 std=0.00072116 valid=100.0% |
| bird : mean=0.00051964 std=0.00057031 valid=100.0% |
| cat : mean=0.00055682 std=0.00060174 valid=100.0% |
| deer : mean=0.00054031 std=0.00055960 valid=100.0% |
| dog : mean=0.00052790 std=0.00055779 valid=100.0% |
| frog : mean=0.00054843 std=0.00056494 valid=100.0% |
| horse : mean=0.00050269 std=0.00053266 valid=100.0% |
| ship : mean=0.00057502 std=0.00071882 valid=100.0% |
| truck : mean=0.00057030 std=0.00070013 valid=100.0% |
|
|
| ✓ Zero negative volumes — all simplices valid |
|
|
| ── Dual Block 1 ── |
| Overall valid: 100.0% |
| vol² mean=0.00104036 std=0.00015917 |
| vol² min=0.00051117 max=0.00314331 |
| vol² median=0.00102234 |
| Images with 100% valid patches: 10000/10000 (100.0%) |
| Per-image valid fraction: mean=1.0000 min=1.0000 |
|
|
| Positive vol² distribution (log10): |
| mean=-2.9877 std=0.0650 |
| p 1: -3.1375 (vol²=0.00072861) |
| p10: -3.0683 (vol²=0.00085449) |
| p25: -3.0312 (vol²=0.00093079) |
| p50: -2.9904 (vol²=0.00102234) |
| p75: -2.9472 (vol²=0.00112915) |
| p90: -2.9053 (vol²=0.00124359) |
| p99: -2.8253 (vol²=0.00149536) |
|
|
| Volume CV: 0.1530 (target band: 0.20-0.23) |
|
|
| Per-class vol²: |
| airplane : mean=0.00105287 std=0.00015561 valid=100.0% |
| automobile : mean=0.00103683 std=0.00016877 valid=100.0% |
| bird : mean=0.00103337 std=0.00015163 valid=100.0% |
| cat : mean=0.00104821 std=0.00017045 valid=100.0% |
| deer : mean=0.00102021 std=0.00014180 valid=100.0% |
| dog : mean=0.00105887 std=0.00016527 valid=100.0% |
| frog : mean=0.00100447 std=0.00014608 valid=100.0% |
| horse : mean=0.00104700 std=0.00015953 valid=100.0% |
| ship : mean=0.00105534 std=0.00015809 valid=100.0% |
| truck : mean=0.00104644 std=0.00016371 valid=100.0% |
|
|
| ✓ Zero negative volumes — all simplices valid |
|
|
| ── Volume Evolution Across Blocks ── |
| Block 0: mean=0.00054120 std=0.00061009 valid=100.0% |
| Block 1: mean=0.00104036 std=0.00015917 valid=100.0% |
| Block 1/Block 0 ratio: 1.9223 |
| ⚠ Volumes GROWING through depth — simplices expanding |
|
|
| ================================================================= |
| SCAN 5: PER-CLASS ACCURACY |
| ================================================================= |
| airplane : acc= 87.6% P(correct)=0.864 confuser=bird(0.031) |
| automobile : acc= 93.1% P(correct)=0.924 confuser=truck(0.048) |
| bird : acc= 82.5% P(correct)=0.807 confuser=deer(0.045) |
| cat : acc= 68.1% P(correct)=0.668 confuser=dog(0.146) |
| deer : acc= 82.1% P(correct)=0.806 confuser=horse(0.048) |
| dog : acc= 76.1% P(correct)=0.758 confuser=cat(0.124) |
| frog : acc= 89.8% P(correct)=0.889 confuser=cat(0.038) |
| horse : acc= 89.4% P(correct)=0.889 confuser=deer(0.026) |
| ship : acc= 92.1% P(correct)=0.913 confuser=airplane(0.036) |
| truck : acc= 91.7% P(correct)=0.909 confuser=automobile(0.036) |
|
|
| ================================================================= |
| SCAN 6: EMBEDDING DISCRIMINATION |
| ================================================================= |
| Same-class cos: mean=0.0847 std=0.1581 |
| Diff-class cos: mean=-0.0064 std=0.1396 |
| Gap (same-diff): 0.0911 |
| 1-NN accuracy: 57.3% |
|
|
| ================================================================= |
| 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: 6,296,582 |
| Geo stream params: 1,281,340 (20.3%) |
| Std stream params: 1,261,248 (20.0%) |
| Fused block params: 3,159,552 (50.2%) |
| Constellation params:84,480 (1.3%) |
|
|
| ================================================================= |
| DIAGNOSIS SUMMARY |
| ================================================================= |
| Val accuracy: 85.2% |
| Eff dim: 58.3/128 |
| Pentachoron CV: 0.1061 (target band: 0.20-0.23) |
| Self-similarity: 0.0027 |
| Pooled anchors: 64/64 |
| Patch anchors: 64/64 |
| Per-img p_anch: 15.2 |
| Entropy: 98% |
| Gini: 0.2067 |
| CM volumes: 200/200 positive |
| Anchor CV: 0.1324 |
| Class CV range: [0.1161, 0.1518] |
| Geo feat var: 0.000838 |
| Block 0 CM valid: 100.0% |
| Block 1 CM valid: 100.0% |
| Same/diff gap: 0.0911 |
| 1-NN accuracy: 57.3% |
|
|
| ✓ No major issues. Geometry is healthy. |
|
|
| ================================================================= |
| DIAGNOSTIC COMPLETE |
| ================================================================= |
