Hugging Face's logo

AbstractPhil
/
geolip-vit-base-x3

TensorBoard
Safetensors
geolip_vit
geometric-deep-learning
vision
multi-expert
patchwork
hypersphere
from-scratch
custom_code
Model card Files
xet
 Metrics Community
geolip-vit-base-x3
File size: 7,438 Bytes
7c4374c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
=================================================================
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}")