| ================================================================= |
| DUAL-STREAM VIT GEOMETRIC DIAGNOSTIC |
| Checkpoint: /content/checkpoints/dual_stream_v3_e100.pt |
| ================================================================= |
|
|
| Epoch: 100 Val acc: 83.5% |
| 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: 83.5% |
| Dual blocks: 2 |
| Patches per image: 64 |
|
|
| ================================================================= |
| SCAN 1: EMBEDDING HEALTH |
| ================================================================= |
| Norms: mean=1.000000 std=0.000000 |
| Self-similarity: mean=0.0197 std=0.1353 |
| ✓ No embedding collapse |
|
|
| Effective dim: 92.1/128 |
| top-3 SVs explain 14.0% |
| top-5 SVs explain 21.8% |
| top-10 SVs explain 36.0% |
| top-20 SVs explain 45.9% |
| top-50 SVs explain 70.6% |
|
|
| Pentachoron CV (GeoLIP structural spectrum): |
| global embedding: CV=0.1149 (✗ outside) [500/500 valid, mean_vol=0.084180] |
| Per-class: |
| airplane : CV=0.2421 (✓ IN BAND) [200/200 valid, mean_vol=0.050191] |
| automobile: CV=0.1793 (✗ outside) [200/200 valid, mean_vol=0.042003] |
| bird : CV=0.2069 (✓ IN BAND) [200/200 valid, mean_vol=0.056354] |
| cat : CV=0.1539 (✗ outside) [200/200 valid, mean_vol=0.063555] |
| deer : CV=0.1997 (✓ IN BAND) [200/200 valid, mean_vol=0.056215] |
| dog : CV=0.1731 (✗ outside) [200/200 valid, mean_vol=0.056146] |
| frog : CV=0.1985 (✓ IN BAND) [200/200 valid, mean_vol=0.049812] |
| horse : CV=0.2266 (✓ IN BAND) [200/200 valid, mean_vol=0.047620] |
| ship : CV=0.2024 (✓ IN BAND) [200/200 valid, mean_vol=0.044833] |
| truck : CV=0.2309 (✓ IN BAND) [200/200 valid, mean_vol=0.046396] |
| Class CV: mean=0.2013 std=0.0274 range=[0.1539, 0.2421] |
| anchor constellation: CV=0.2002 (✓ IN BAND) [200/200 valid, mean_vol=0.077932] |
|
|
| Patch-level CV (from fused patch projections): |
| all patches (flat): CV=0.1593 (✗ outside) [500/500 valid, mean_vol=0.080665] |
|
|
| ================================================================= |
| SCAN 2: ANCHOR HEALTH |
| ================================================================= |
| Anchor pairwise cos: mean=-0.0131 max=0.8867 |
| Anchor effective rank: 30.5/128 |
|
|
| Pooled anchors active: 64/64 |
| Patch anchors active: 64/64 |
| Per-image patch anchors: mean=13.6 min=2 max=31 |
| Entropy: 3.8856/4.1589 (93%) |
| Gini: 0.4081 |
|
|
| ================================================================= |
| SCAN 3: CLASS GEOMETRY |
| ================================================================= |
| Inter-class cos: mean=-0.0060 max=0.4655 min=-0.2154 |
|
|
| Most similar: |
| cat × dog : cos=0.4655 |
| bird × deer : cos=0.2067 |
| bird × cat : cos=0.2001 |
| cat × deer : cos=0.1866 |
| cat × frog : cos=0.1584 |
| Most distant: |
| automobile × deer : cos=-0.2154 |
| automobile × dog : cos=-0.1881 |
| automobile × cat : cos=-0.1869 |
|
|
| Intra-class spread: |
| airplane : spread=0.5070 |
| automobile : spread=0.4499 |
| bird : spread=0.5588 |
| cat : spread=0.6096 |
| deer : spread=0.5566 |
| dog : spread=0.5696 |
| frog : spread=0.4924 |
| horse : spread=0.4935 |
| ship : spread=0.4669 |
| truck : spread=0.4743 |
|
|
| ================================================================= |
| SCAN 4: KSIMPLEX GEOMETRIC FEATURES |
| ================================================================= |
|
|
| ── Dual Block 0 ── |
| Shape: torch.Size([10000, 64, 11]) |
| d² pairs: mean=0.2203 std=0.8032 min=-1.5042 max=2.7124 |
| vol² feat: mean=-2.1158 std=0.2963 |
| Per-feature means: |
| d²_0: mean=0.2962 std=0.8388 |
| d²_1: mean=-0.0692 std=0.1620 |
| d²_2: mean=-0.0306 std=0.3689 |
| d²_3: mean=0.0112 std=0.3136 |
| d²_4: mean=0.0028 std=0.2813 |
| d²_5: mean=0.1604 std=0.2355 |
| d²_6: mean=2.1255 std=0.5473 |
| d²_7: mean=-0.3579 std=0.4393 |
| d²_8: mean=0.0236 std=0.7105 |
| d²_9: mean=0.0407 std=0.2703 |
| vol²: mean=-2.1158 std=0.2963 |
|
|
| Most discriminative features (by class variance): |
| d²_6: var=0.005849 |
| d²_0: var=0.003270 |
| d²_7: var=0.002789 |
| d²_2: var=0.002332 |
| vol²: var=0.002312 |
|
|
| Per-class profiles (11-d mean): |
| airplane : norm=3.0035 strongest=vol²(-2.1586) |
| automobile : norm=3.0534 strongest=vol²(-2.1833) |
| bird : norm=3.0485 strongest=d²_6(2.1739) |
| cat : norm=3.0454 strongest=d²_6(2.1522) |
| deer : norm=3.0802 strongest=d²_6(2.2495) |
| dog : norm=3.0446 strongest=d²_6(2.1520) |
| frog : norm=3.0777 strongest=d²_6(2.2097) |
| horse : norm=3.0267 strongest=d²_6(2.1204) |
| ship : norm=3.0400 strongest=vol²(-2.1387) |
| truck : norm=3.0153 strongest=vol²(-2.1698) |
|
|
| Geo-only class separation (cos on 11-d): |
| mean=0.9977 max=1.0000 min=0.9918 |
| Most similar (geo): |
| cat × dog : cos=1.0000 |
| cat × horse : cos=0.9999 |
| dog × horse : cos=0.9998 |
| Most distinct (geo): |
| deer × truck : cos=0.9918 |
| airplane × deer : cos=0.9938 |
| automobile × deer : cos=0.9943 |
|
|
| ── Dual Block 1 ── |
| Shape: torch.Size([10000, 64, 11]) |
| d² pairs: mean=0.1883 std=0.6192 min=-1.8020 max=2.3287 |
| vol² feat: mean=-2.2199 std=0.2130 |
| Per-feature means: |
| d²_0: mean=-0.2335 std=0.7433 |
| d²_1: mean=0.3329 std=0.7269 |
| d²_2: mean=0.0664 std=0.6665 |
| d²_3: mean=0.0815 std=0.5484 |
| d²_4: mean=0.4962 std=0.5445 |
| d²_5: mean=0.2920 std=0.3366 |
| d²_6: mean=0.1333 std=0.7019 |
| d²_7: mean=0.2054 std=0.4258 |
| d²_8: mean=0.4398 std=0.6412 |
| d²_9: mean=0.0690 std=0.3185 |
| vol²: mean=-2.2199 std=0.2130 |
|
|
| Most discriminative features (by class variance): |
| d²_0: var=0.003839 |
| d²_3: var=0.003223 |
| d²_8: var=0.003008 |
| d²_6: var=0.002997 |
| d²_2: var=0.001378 |
|
|
| Per-class profiles (11-d mean): |
| airplane : norm=2.4099 strongest=vol²(-2.2336) |
| automobile : norm=2.3391 strongest=vol²(-2.1828) |
| bird : norm=2.4203 strongest=vol²(-2.2320) |
| cat : norm=2.3761 strongest=vol²(-2.2202) |
| deer : norm=2.4454 strongest=vol²(-2.2474) |
| dog : norm=2.3706 strongest=vol²(-2.2211) |
| frog : norm=2.4150 strongest=vol²(-2.2294) |
| horse : norm=2.3703 strongest=vol²(-2.2172) |
| ship : norm=2.4024 strongest=vol²(-2.2263) |
| truck : norm=2.3458 strongest=vol²(-2.1891) |
|
|
| Geo-only class separation (cos on 11-d): |
| mean=0.9970 max=0.9996 min=0.9928 |
| Most similar (geo): |
| cat × dog : cos=0.9996 |
| automobile × horse : cos=0.9996 |
| horse × truck : cos=0.9995 |
| Most distinct (geo): |
| airplane × deer : cos=0.9928 |
| airplane × frog : cos=0.9929 |
| deer × dog : cos=0.9940 |
|
|
| ── Geo Evolution Across Blocks ── |
| Block 0: feat_var=0.001972 class_sep=0.0023 |
| Block 1: feat_var=0.001687 class_sep=0.0030 |
|
|
| ================================================================= |
| SCAN 4B: CAYLEY-MENGER VOLUME² PER DUAL BLOCK |
| ================================================================= |
|
|
| ── Dual Block 0 ── |
| Overall valid: 100.0% |
| vol² mean=0.00056016 std=0.00050176 |
| vol² min=0.00002086 max=0.00561523 |
| vol² median=0.00040436 |
| Images with 100% valid patches: 10000/10000 (100.0%) |
| Per-image valid fraction: mean=1.0000 min=1.0000 |
|
|
| Positive vol² distribution (log10): |
| mean=-3.3879 std=0.3458 |
| p 1: -4.1849 (vol²=0.00006533) |
| p10: -3.8331 (vol²=0.00014687) |
| p25: -3.6227 (vol²=0.00023842) |
| p50: -3.3932 (vol²=0.00040436) |
| p75: -3.1467 (vol²=0.00071335) |
| p90: -2.9414 (vol²=0.00114441) |
| p99: -2.5784 (vol²=0.00263977) |
|
|
| Volume CV: 0.8957 (target band: 0.20-0.23) |
|
|
| Per-class vol²: |
| airplane : mean=0.00054835 std=0.00041243 valid=100.0% |
| automobile : mean=0.00067123 std=0.00061501 valid=100.0% |
| bird : mean=0.00049331 std=0.00046131 valid=100.0% |
| cat : mean=0.00057685 std=0.00050456 valid=100.0% |
| deer : mean=0.00045237 std=0.00042224 valid=100.0% |
| dog : mean=0.00054869 std=0.00046722 valid=100.0% |
| frog : mean=0.00048420 std=0.00044561 valid=100.0% |
| horse : mean=0.00053539 std=0.00044204 valid=100.0% |
| ship : mean=0.00062552 std=0.00056308 valid=100.0% |
| truck : mean=0.00066563 std=0.00058618 valid=100.0% |
|
|
| ✓ Zero negative volumes — all simplices valid |
|
|
| ── Dual Block 1 ── |
| Overall valid: 100.0% |
| vol² mean=0.00205160 std=0.00143469 |
| vol² min=0.00007963 max=0.02587891 |
| vol² median=0.00165558 |
| Images with 100% valid patches: 10000/10000 (100.0%) |
| Per-image valid fraction: mean=1.0000 min=1.0000 |
|
|
| Positive vol² distribution (log10): |
| mean=-2.7736 std=0.2705 |
| p 1: -3.3713 (vol²=0.00042534) |
| p10: -3.1175 (vol²=0.00076294) |
| p25: -2.9591 (vol²=0.00109863) |
| p50: -2.7811 (vol²=0.00165558) |
| p75: -2.5912 (vol²=0.00256348) |
| p90: -2.4168 (vol²=0.00382996) |
| p99: -2.1352 (vol²=0.00732422) |
|
|
| Volume CV: 0.6993 (target band: 0.20-0.23) |
|
|
| Per-class vol²: |
| airplane : mean=0.00193671 std=0.00140925 valid=100.0% |
| automobile : mean=0.00205615 std=0.00148104 valid=100.0% |
| bird : mean=0.00198890 std=0.00139486 valid=100.0% |
| cat : mean=0.00213766 std=0.00148557 valid=100.0% |
| deer : mean=0.00193340 std=0.00131106 valid=100.0% |
| dog : mean=0.00216270 std=0.00147388 valid=100.0% |
| frog : mean=0.00191994 std=0.00122417 valid=100.0% |
| horse : mean=0.00210441 std=0.00146341 valid=100.0% |
| ship : mean=0.00213790 std=0.00152762 valid=100.0% |
| truck : mean=0.00213822 std=0.00151685 valid=100.0% |
|
|
| ✓ Zero negative volumes — all simplices valid |
|
|
| ── Volume Evolution Across Blocks ── |
| Block 0: mean=0.00056016 std=0.00050176 valid=100.0% |
| Block 1: mean=0.00205160 std=0.00143469 valid=100.0% |
| Block 1/Block 0 ratio: 3.6626 |
| ⚠ Volumes GROWING through depth — simplices expanding |
|
|
| ================================================================= |
| SCAN 5: PER-CLASS ACCURACY |
| ================================================================= |
| airplane : acc= 85.7% P(correct)=0.853 confuser=bird(0.034) |
| automobile : acc= 92.9% P(correct)=0.928 confuser=truck(0.051) |
| bird : acc= 77.6% P(correct)=0.769 confuser=deer(0.050) |
| cat : acc= 66.1% P(correct)=0.651 confuser=dog(0.145) |
| deer : acc= 77.9% P(correct)=0.767 confuser=bird(0.060) |
| dog : acc= 75.4% P(correct)=0.746 confuser=cat(0.137) |
| frog : acc= 88.9% P(correct)=0.883 confuser=bird(0.032) |
| horse : acc= 88.5% P(correct)=0.876 confuser=deer(0.028) |
| ship : acc= 91.1% P(correct)=0.907 confuser=airplane(0.034) |
| truck : acc= 90.7% P(correct)=0.904 confuser=automobile(0.042) |
|
|
| ================================================================= |
| SCAN 6: EMBEDDING DISCRIMINATION |
| ================================================================= |
| Same-class cos: mean=0.2369 std=0.1500 |
| Diff-class cos: mean=-0.0042 std=0.1097 |
| Gap (same-diff): 0.2411 |
| 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: 6,296,582 |
| 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: 83.5% |
| Eff dim: 92.1/128 |
| Pentachoron CV: 0.1149 (target band: 0.20-0.23) |
| Self-similarity: 0.0197 |
| Pooled anchors: 64/64 |
| Patch anchors: 64/64 |
| Per-img p_anch: 13.6 |
| Entropy: 93% |
| Gini: 0.4081 |
| CM volumes: 200/200 positive |
| Anchor CV: 0.2002 |
| Class CV range: [0.1539, 0.2421] |
| Geo feat var: 0.001687 |
| Block 0 CM valid: 100.0% |
| Block 1 CM valid: 100.0% |
| Same/diff gap: 0.2411 |
| 1-NN accuracy: 77.4% |
|
|
| ✓ No major issues. Geometry is healthy. |
|
|
| ================================================================= |
| DIAGNOSTIC COMPLETE |
| ================================================================= |
