================================================================= BASE TIER SOUP ANALYSIS Device: cuda ================================================================= Loading checkpoint... Loaded: mAP=0.825 cv=0.3117 epoch=19 clip_l14_openai loaded dinov2_b14 loaded siglip_b16_384 loaded Running inference on 5000 val images... Done: fused=torch.Size([5000, 128]) tri=torch.Size([5000, 256]) ================================================================= SCAN 1: ANCHOR GEOMETRY ================================================================= Anchor pairwise cosine: mean=0.0356 std=0.1896 max=0.9542 min=-0.9093 Max neighbor cosine per anchor: mean=0.6949 std=0.2730 max=0.9542 min=0.0639 Anchor norms: mean=1.000000 std=0.000000 Anchor spectral: eff_rank=65.7/128 sv_max=5.0231 sv_10=2.8655 sv_50=0.9697 sv_min=0.125017 Anchor pentachoron CV: 0.2478 mean_vol=0.074751 std_vol=0.018524 ================================================================= SCAN 2: ANCHOR UTILIZATION ================================================================= Active anchors: 1/256 (0.4%) Visit counts: mean=19.5 std=312.5 max=5000 min=0 top 10: [5000, 0, 0, 0, 0, 0, 0, 0, 0, 0] bottom 10: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0] Anchor entropy: -0.0000 / 5.5452 (-0.0%) Per-anchor embedding density: Intra-cluster cosine: mean=0.9693 std=0.0000 ================================================================= SCAN 3: PROJECTOR ANALYSIS ================================================================= clip_l14: norm: mean=1.000000 (should be 1.0) self-sim off-diag: 0.9668 eff_dim: 24.1/128 dinov2_b14: norm: mean=1.000000 (should be 1.0) self-sim off-diag: 0.9678 eff_dim: 25.3/128 siglip_b16: norm: mean=1.000000 (should be 1.0) self-sim off-diag: 0.9501 eff_dim: 23.8/128 Expert agreement (cosine in 128-d): clip_l14 × dinov2_b14 : mean=0.9898 std=0.0066 min=0.8730 clip_l14 × siglip_b16 : mean=0.9893 std=0.0052 min=0.9307 dinov2_b14 × siglip_b16 : mean=0.9855 std=0.0081 min=0.8920 Per-expert nearest anchor agreement: clip_l14 × dinov2_b14 : same_anchor=1.0000 (100.0%) clip_l14 × siglip_b16 : same_anchor=1.0000 (100.0%) dinov2_b14 × siglip_b16 : same_anchor=1.0000 (100.0%) Projector weight comparison: clip_l14 : norm=37.4114 eff_rank=30.5/128 dinov2_b14 : norm=36.3149 eff_rank=23.0/128 siglip_b16 : norm=39.2079 eff_rank=29.0/128 clip_l14 × dinov2_b14 weight_cos=0.0046 clip_l14 × siglip_b16 weight_cos=-0.0049 dinov2_b14 × siglip_b16 weight_cos=-0.0055 ================================================================= SCAN 4: PATCHWORK COMPARTMENTS ================================================================= Comp 0: 32 anchors Comp 1: 32 anchors Comp 2: 32 anchors Comp 3: 32 anchors Comp 4: 32 anchors Comp 5: 32 anchors Comp 6: 32 anchors Comp 7: 32 anchors Patchwork output: torch.Size([5000, 512]) norm: mean=11.6381 std=0.5046 comp 0: norm=2.8604 std_across_dims=0.0010 comp 1: norm=3.7652 std_across_dims=0.1596 comp 2: norm=2.3303 std_across_dims=0.0057 comp 3: norm=3.4802 std_across_dims=0.2053 comp 4: norm=3.3465 std_across_dims=0.1143 comp 5: norm=5.9651 std_across_dims=0.4720 comp 6: norm=6.0775 std_across_dims=0.4946 comp 7: norm=3.2188 std_across_dims=0.0718 ================================================================= SCAN 5: TRIANGULATION PATTERNS ================================================================= Triangulation distances (1-cosine): mean=0.8988 std=0.1301 min=0.0038 max=1.2538 Nearest anchor distance: mean=0.0156 std=0.0042 max=0.0419 min=0.0038 Anchors within cos>0.5 per image: mean=1.0 std=0.0 Top-10 nearest anchor distances: k=0: mean=0.0156 std=0.0042 k=1: mean=0.6646 std=0.0218 k=2: mean=0.6806 std=0.0185 k=3: mean=0.6909 std=0.0173 k=4: mean=0.6977 std=0.0167 k=5: mean=0.7033 std=0.0162 k=6: mean=0.7081 std=0.0158 k=7: mean=0.7126 std=0.0154 k=8: mean=0.7166 std=0.0150 k=9: mean=0.7204 std=0.0147 ================================================================= SCAN 6: PER-CLASS ANCHOR AFFINITY ================================================================= Top-3 anchors per class (first 20 classes): person (n=2693): 65(2693/2693) 1(0/2693) 0(0/2693) bicycle (n= 149): 65(149/149) 1(0/149) 0(0/149) car (n= 535): 65(535/535) 1(0/535) 0(0/535) motorcycle (n= 159): 65(159/159) 1(0/159) 0(0/159) airplane (n= 97): 65(97/97) 1(0/97) 0(0/97) bus (n= 189): 65(189/189) 1(0/189) 0(0/189) train (n= 157): 65(157/157) 1(0/157) 0(0/157) truck (n= 250): 65(250/250) 1(0/250) 0(0/250) boat (n= 121): 65(121/121) 1(0/121) 0(0/121) traffic light (n= 191): 65(191/191) 1(0/191) 0(0/191) fire hydrant (n= 86): 65(86/86) 1(0/86) 0(0/86) stop sign (n= 69): 65(69/69) 1(0/69) 0(0/69) parking meter (n= 37): 65(37/37) 1(0/37) 0(0/37) bench (n= 235): 65(235/235) 1(0/235) 0(0/235) bird (n= 125): 65(125/125) 1(0/125) 0(0/125) cat (n= 184): 65(184/184) 1(0/184) 0(0/184) dog (n= 177): 65(177/177) 1(0/177) 0(0/177) horse (n= 128): 65(128/128) 1(0/128) 0(0/128) sheep (n= 65): 65(65/65) 1(0/65) 0(0/65) cow (n= 87): 65(87/87) 1(0/87) 0(0/87) Anchor specialization: classes per anchor: mean=80.0 std=nan max=80 min=80 ================================================================= SCAN 7: FUSED EMBEDDING GEOMETRY ================================================================= Norms: mean=1.000000 std=0.000000 Self-sim (off-diag): 0.9693 Effective dim: 23.6/128 top-5 SVs explain 44.1% top-10 SVs explain 70.2% top-20 SVs explain 99.2% top-50 SVs explain 100.0% top-100 SVs explain 100.0% Pentachoron CV: 0.3529 ================================================================= SCAN 8: EXPERT CONTRIBUTION ================================================================= clip_l14 : cos_to_fused mean=0.9970 std=0.0016 dinov2_b14 : cos_to_fused mean=0.9957 std=0.0027 siglip_b16 : cos_to_fused mean=0.9955 std=0.0023 Without clip_l14 : cos_to_full=0.9992 (uniqueness=0.0008) Without dinov2_b14 : cos_to_full=0.9989 (uniqueness=0.0011) Without siglip_b16 : cos_to_full=0.9989 (uniqueness=0.0011) Per-image expert disagreement: Agreement: mean=0.9882 std=0.0055 Disagreement: mean=0.0041 std=0.0034 Most agreed image (1449): agreement=0.9978 labels: [22] Most disagreed image (1435): agreement=0.9214 labels: [28] ================================================================= ANALYSIS COMPLETE ================================================================= /tmp/ipykernel_10600/3734699858.py:410: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1857.) f"std={anchor_class_count[anchor_class_count>0].std():.1f}")