AbstractPhil commited on
Commit
48e9436
·
verified ·
1 Parent(s): 78b3309

Create hypersphere_convergence.txt

Browse files
Files changed (1) hide show
  1. hypersphere_convergence.txt +397 -0
hypersphere_convergence.txt ADDED
@@ -0,0 +1,397 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ======================================================================
2
+ GEOLIP HYPERSPHERE STRUCTURAL ANALYSIS
3
+ Checkpoint: checkpoints/dual_stream_best.pt
4
+ Device: cuda
5
+ ======================================================================
6
+
7
+ Loading checkpoint...
8
+ mAP=0.838 epoch=17
9
+ D_ANCHOR=256 N_ANCHORS=512 N_EXPERTS=3
10
+ Anchors: torch.Size([512, 256])
11
+ Expert rotations: N/A (dual-stream — perspectives in projectors)
12
+
13
+ Loading expert features...
14
+ Generating embeddings...
15
+ Embeddings: torch.Size([5000, 256])
16
+
17
+ ======================================================================
18
+ SCAN 1: ANCHOR GEOMETRY
19
+ ======================================================================
20
+ Pairwise cosine:
21
+ mean=0.0001 std=0.0442
22
+ max=0.5800 min=-0.5539
23
+ Max neighbor cosine per anchor:
24
+ mean=0.2330 std=0.0823
25
+ max=0.5800 min=0.1204
26
+ pairs with cos > 0.9: 0 (0.00%)
27
+ pairs with cos > 0.8: 0 (0.00%)
28
+ pairs with cos > 0.7: 0 (0.00%)
29
+ pairs with cos > 0.5: 3 (0.00%)
30
+ pairs with cos > 0.3: 47 (0.04%)
31
+ pairs with cos > 0.0: 64430 (49.25%)
32
+
33
+ Anchor spectral:
34
+ effective rank: 256.0/256
35
+ sv_max=1.4151 sv_10=1.4146 sv_50=1.4145 sv_min=1.413725
36
+ top-10 SVs explain 3.9%
37
+ top-25 SVs explain 9.8%
38
+ top-50 SVs explain 19.5%
39
+ top-100 SVs explain 39.1%
40
+ top-128 SVs explain 50.0%
41
+ top-200 SVs explain 78.1%
42
+
43
+ Anchor pentachoron CV: 0.0286
44
+ mean_vol=0.092596 std_vol=0.002649
45
+
46
+ ======================================================================
47
+ SCAN 2: ANCHOR UTILIZATION
48
+ ======================================================================
49
+ Active anchors: 201/512 (39.3%)
50
+ Visit counts: mean=9.8 std=25.6
51
+ max=264 min=1 (among active)
52
+ top 10: [np.float32(264.0), np.float32(194.0), np.float32(161.0), np.float32(140.0), np.float32(129.0), np.float32(114.0), np.float32(111.0), np.float32(103.0), np.float32(96.0), np.float32(90.0)]
53
+ Entropy: 4.5776 / 6.2383 (73.4%)
54
+ Gini coefficient: 0.8568 (0=equal, 1=one anchor gets all)
55
+ anchors with 1-5 visits: 66
56
+ anchors with 5-20 visits: 64
57
+ anchors with 20-50 visits: 39
58
+ anchors with 50-100 visits: 24
59
+ anchors with 100-500 visits: 8
60
+ anchors with 500-5000 visits: 0
61
+
62
+ ======================================================================
63
+ SCAN 3: EMBEDDING MANIFOLD GEOMETRY
64
+ ======================================================================
65
+ Effective dimensionality: 71.6/256
66
+ top-5 SVs explain 33.5%
67
+ top-10 SVs explain 51.1%
68
+ top-20 SVs explain 72.6%
69
+ top-50 SVs explain 94.2%
70
+ top-100 SVs explain 99.4%
71
+ top-128 SVs explain 99.8%
72
+ top-200 SVs explain 100.0%
73
+
74
+ Self-similarity (off-diagonal):
75
+ mean=0.0025 std=0.1920
76
+ max=0.9931 min=-0.5248
77
+
78
+ Norms: mean=1.000000 std=0.000000
79
+
80
+ Global pentachoron CV: 0.2199
81
+ mean_vol=0.080570 std_vol=0.017715
82
+
83
+ ======================================================================
84
+ SCAN 4: EXPERT PERSPECTIVE DIVERGENCE
85
+ ======================================================================
86
+
87
+ Per-image expert agreement:
88
+ clip_l14_openai × dinov2_b14 : mean=1.0000 std=0.0000 min=1.0000
89
+ clip_l14_openai × siglip_b16_384 : mean=1.0000 std=0.0000 min=1.0000
90
+ dinov2_b14 × siglip_b16_384 : mean=1.0000 std=0.0000 min=1.0000
91
+
92
+ Per-anchor expert divergence:
93
+ mean divergence: 0.0000 std: 0.0000
94
+ max divergence: 0.0000 (anchor 0)
95
+ min divergence: 0.0000 (anchor 0)
96
+
97
+ Top 10 most contentious anchors:
98
+ #1 anchor 39: div=0.0000 visits=0
99
+ #2 anchor 38: div=0.0000 visits=0
100
+ #3 anchor 37: div=0.0000 visits=0
101
+ #4 anchor 36: div=0.0000 visits=0
102
+ #5 anchor 35: div=0.0000 visits=0
103
+ #6 anchor 34: div=0.0000 visits=0
104
+ #7 anchor 33: div=0.0000 visits=0
105
+ #8 anchor 32: div=0.0000 visits=2
106
+ #9 anchor 47: div=0.0000 visits=0
107
+ #10 anchor 46: div=0.0000 visits=1
108
+
109
+ Top 10 most unanimous anchors:
110
+ #1 anchor 504: div=0.0000 visits=0
111
+ #2 anchor 505: div=0.0000 visits=2
112
+ #3 anchor 506: div=0.0000 visits=27
113
+ #4 anchor 507: div=0.0000 visits=17
114
+ #5 anchor 508: div=0.0000 visits=4
115
+ #6 anchor 509: div=0.0000 visits=0
116
+ #7 anchor 510: div=0.0000 visits=0
117
+ #8 anchor 511: div=0.0000 visits=0
118
+ #9 anchor 496: div=0.0000 visits=0
119
+ #10 anchor 497: div=0.0000 visits=0
120
+
121
+ Expert rotation eigenspectra:
122
+ clip_l14_openai : ortho_err=0.000000 eval_min=1.0000 eval_max=1.0000
123
+ dinov2_b14 : ortho_err=0.000000 eval_min=1.0000 eval_max=1.0000
124
+ siglip_b16_384 : ortho_err=0.000000 eval_min=1.0000 eval_max=1.0000
125
+
126
+ Expert whitener condition:
127
+ clip_l14_openai : cond=1.00 sv_max=1.0000 sv_min=1.000000
128
+ dinov2_b14 : cond=1.00 sv_max=1.0000 sv_min=1.000000
129
+ siglip_b16_384 : cond=1.00 sv_max=1.0000 sv_min=1.000000
130
+
131
+ ======================================================================
132
+ SCAN 5: NEAREST ANCHOR DISTANCES
133
+ ======================================================================
134
+ k= 0: mean_dist=0.5721 std=0.0978 max=0.8074 min=0.2505
135
+ k= 1: mean_dist=0.6521 std=0.0692 max=0.8282 min=0.4311
136
+ k= 2: mean_dist=0.6931 std=0.0559 max=0.8344 min=0.5052
137
+ k= 4: mean_dist=0.7466 std=0.0436 max=0.8548 min=0.5865
138
+ k= 9: mean_dist=0.8172 std=0.0243 max=0.8985 min=0.7391
139
+ k= 19: mean_dist=0.8814 std=0.0183 max=0.9495 min=0.8327
140
+ k= 49: mean_dist=0.9374 std=0.0132 max=0.9738 min=0.9056
141
+ k= 99: mean_dist=0.9671 std=0.0084 max=0.9896 min=0.9445
142
+ anchors with cos > 0.9: mean=0.0 max=0 min=0
143
+ anchors with cos > 0.8: mean=0.0 max=0 min=0
144
+ anchors with cos > 0.7: mean=0.0 max=1 min=0
145
+ anchors with cos > 0.5: mean=0.3 max=2 min=0
146
+ anchors with cos > 0.3: mean=2.7 max=8 min=0
147
+ anchors with cos > 0.0: mean=227.2 max=284 min=148
148
+
149
+ ======================================================================
150
+ SCAN 6: PER-CLASS ANCHOR AFFINITY
151
+ ======================================================================
152
+ Anchor specialization:
153
+ classes per active anchor: mean=12.8 std=9.4
154
+ max=40 min=1
155
+
156
+ Class spread (anchors per class):
157
+ mean=32.2 std=28.8
158
+ max=170 (person)
159
+ min=2 (bear)
160
+
161
+ Top 10 by anchor spread:
162
+ person : 170 anchors, 2693 images
163
+ chair : 105 anchors, 580 images
164
+ car : 98 anchors, 535 images
165
+ bottle : 85 anchors, 379 images
166
+ dining table : 84 anchors, 501 images
167
+ handbag : 81 anchors, 292 images
168
+ cup : 80 anchors, 390 images
169
+ backpack : 76 anchors, 228 images
170
+ truck : 66 anchors, 250 images
171
+ book : 65 anchors, 230 images
172
+ Bottom 10 by anchor spread:
173
+ sheep : 6 anchors, 65 images
174
+ tennis racket : 5 anchors, 167 images
175
+ giraffe : 5 anchors, 101 images
176
+ hair drier : 4 anchors, 9 images
177
+ toaster : 4 anchors, 8 images
178
+ baseball glove : 4 anchors, 100 images
179
+ elephant : 4 anchors, 89 images
180
+ skis : 4 anchors, 120 images
181
+ zebra : 3 anchors, 85 images
182
+ bear : 2 anchors, 49 images
183
+
184
+ ======================================================================
185
+ SCAN 7: INTER-CLASS GEOMETRIC DISTANCES
186
+ ======================================================================
187
+ Inter-class cosine (78 classes with >10 images):
188
+ mean=0.0531
189
+ max=0.9947 min=-0.4158
190
+
191
+ Most similar class pairs:
192
+ #1: mouse × keyboard cos=0.9947
193
+ #2: baseball bat × baseball glove cos=0.9905
194
+ #3: microwave × refrigerator cos=0.9783
195
+ #4: spoon × bowl cos=0.9757
196
+ #5: skis × snowboard cos=0.9707
197
+ #6: microwave × oven cos=0.9615
198
+ #7: knife × dining table cos=0.9582
199
+ #8: cup × dining table cos=0.9580
200
+ #9: laptop × keyboard cos=0.9577
201
+ #10: fork × knife cos=0.9565
202
+ #11: laptop × mouse cos=0.9531
203
+ #12: oven × refrigerator cos=0.9510
204
+ #13: knife × spoon cos=0.9489
205
+ #14: apple × orange cos=0.9461
206
+ #15: broccoli × carrot cos=0.9433
207
+
208
+ Most distant class pairs:
209
+ #1: person × zebra cos=-0.4158
210
+ #2: person × giraffe cos=-0.3890
211
+ #3: airplane × bottle cos=-0.3845
212
+ #4: snowboard × potted plant cos=-0.3742
213
+ #5: snowboard × chair cos=-0.3692
214
+ #6: snowboard × vase cos=-0.3677
215
+ #7: bench × toothbrush cos=-0.3641
216
+ #8: skis × potted plant cos=-0.3635
217
+ #9: truck × toothbrush cos=-0.3629
218
+ #10: bird × bottle cos=-0.3616
219
+ #11: airplane × cup cos=-0.3575
220
+ #12: car × toothbrush cos=-0.3572
221
+ #13: skis × vase cos=-0.3550
222
+ #14: skis × chair cos=-0.3484
223
+ #15: horse × bottle cos=-0.3445
224
+
225
+ Intra-class spread:
226
+ Tightest 10:
227
+ giraffe : spread=0.0566 (n=101)
228
+ zebra : spread=0.0626 (n=85)
229
+ tennis racket : spread=0.0785 (n=167)
230
+ baseball glove : spread=0.0871 (n=100)
231
+ skis : spread=0.0880 (n=120)
232
+ elephant : spread=0.1058 (n=89)
233
+ snowboard : spread=0.1145 (n=49)
234
+ skateboard : spread=0.1188 (n=127)
235
+ surfboard : spread=0.1377 (n=149)
236
+ baseball bat : spread=0.1512 (n=97)
237
+ Loosest 10:
238
+ potted plant : spread=0.5553 (n=172)
239
+ umbrella : spread=0.5566 (n=174)
240
+ cup : spread=0.5587 (n=390)
241
+ car : spread=0.5765 (n=535)
242
+ bottle : spread=0.6014 (n=379)
243
+ handbag : spread=0.6207 (n=292)
244
+ bench : spread=0.6600 (n=235)
245
+ chair : spread=0.6621 (n=580)
246
+ backpack : spread=0.6764 (n=228)
247
+ person : spread=0.8186 (n=2693)
248
+
249
+ ======================================================================
250
+ SCAN 8: LOCAL PENTACHORON CV
251
+ ======================================================================
252
+ Clusters with 10+ members: 102
253
+ Local CV: mean=0.4302 std=0.1654
254
+ max=1.0834 min=0.1965
255
+ Global CV: 0.2199
256
+ Ratio (local/global): 1.9564
257
+
258
+ Highest local CV (most diverse clusters):
259
+ anchor 86: CV=1.0834 n= 103 mean_vol=0.002920
260
+ anchor 469: CV=0.8698 n= 89 mean_vol=0.000587
261
+ anchor 364: CV=0.8659 n= 194 mean_vol=0.007823
262
+ anchor 175: CV=0.8403 n= 88 mean_vol=0.001932
263
+ anchor 68: CV=0.8222 n= 161 mean_vol=0.000757
264
+ anchor 220: CV=0.7987 n= 114 mean_vol=0.001308
265
+ anchor 437: CV=0.7678 n= 129 mean_vol=0.001834
266
+ anchor 498: CV=0.7410 n= 78 mean_vol=0.001262
267
+ anchor 205: CV=0.6792 n= 48 mean_vol=0.001457
268
+ anchor 339: CV=0.6555 n= 11 mean_vol=0.003612
269
+ Lowest local CV (most uniform clusters):
270
+ anchor 361: CV=0.2631 n= 41 mean_vol=0.012367
271
+ anchor 254: CV=0.2615 n= 32 mean_vol=0.019875
272
+ anchor 155: CV=0.2565 n= 34 mean_vol=0.017756
273
+ anchor 53: CV=0.2527 n= 12 mean_vol=0.025066
274
+ anchor 479: CV=0.2503 n= 43 mean_vol=0.021270
275
+ anchor 157: CV=0.2354 n= 11 mean_vol=0.031356
276
+ anchor 289: CV=0.2137 n= 18 mean_vol=0.018227
277
+ anchor 430: CV=0.2126 n= 15 mean_vol=0.041118
278
+ anchor 486: CV=0.2002 n= 11 mean_vol=0.016117
279
+ anchor 131: CV=0.1965 n= 14 mean_vol=0.030370
280
+
281
+ ======================================================================
282
+ SCAN 9: PROJECTOR ANALYSIS
283
+ ======================================================================
284
+
285
+ clip_l14_openai:
286
+ self-sim: mean=0.0538 std=0.1757
287
+ eff_dim: 76.3/256
288
+ cos→fused: mean=0.9160 std=0.0323
289
+
290
+ dinov2_b14:
291
+ self-sim: mean=0.1254 std=0.1798
292
+ eff_dim: 74.5/256
293
+ cos→fused: mean=0.8665 std=0.0459
294
+
295
+ siglip_b16_384:
296
+ self-sim: mean=0.0332 std=0.1800
297
+ eff_dim: 76.3/256
298
+ cos→fused: mean=0.9303 std=0.0284
299
+
300
+ Cross-expert agreement (projected):
301
+ clip_l14_openai × dinov2_b14 : cos=0.6581 std=0.1092
302
+ clip_l14_openai × siglip_b16_384 : cos=0.8297 std=0.0749
303
+ dinov2_b14 × siglip_b16_384 : cos=0.6966 std=0.1019
304
+
305
+ Expert uniqueness (leave-one-out):
306
+ Without clip_l14_openai : cos_to_full=0.9756 (uniqueness=0.0244)
307
+ Without dinov2_b14 : cos_to_full=0.9652 (uniqueness=0.0348)
308
+ Without siglip_b16_384 : cos_to_full=0.9791 (uniqueness=0.0209)
309
+
310
+ ======================================================================
311
+ SCAN 9.5: DUAL-STREAM ANALYSIS
312
+ ======================================================================
313
+
314
+ Shared × Native cosine per expert:
315
+ clip_l14_openai : mean=0.3337 std=0.0146 min=0.2898 max=0.4650
316
+ dinov2_b14 : mean=0.3576 std=0.0256 min=0.2771 max=0.5046
317
+ siglip_b16_384 : mean=0.3503 std=0.0191 min=0.2433 max=0.4444
318
+
319
+ Displacement (shared - native):
320
+ clip_l14_openai : L2_mean=1.1543 std=0.0127
321
+ dinov2_b14 : L2_mean=1.1332 std=0.0228
322
+ siglip_b16_384 : L2_mean=1.1398 std=0.0168
323
+
324
+ Native effective dimensionality:
325
+ clip_l14_openai : eff_dim=71.3/256
326
+ dinov2_b14 : eff_dim=62.4/256
327
+ siglip_b16_384 : eff_dim=62.8/256
328
+
329
+ Cross-expert native agreement:
330
+ clip_l14_openai × dinov2_b14 : mean=0.2059 std=0.0037
331
+ clip_l14_openai × siglip_b16_384 : mean=0.2078 std=0.0041
332
+ dinov2_b14 × siglip_b16_384 : mean=0.7933 std=0.0048
333
+
334
+ Shared × Other's Native (cross-stream):
335
+ clip_l14_ope_shared × dinov2_b14 _native: mean=0.1277
336
+ clip_l14_ope_shared × siglip_b16_3_native: mean=0.2346
337
+ dinov2_b14_shared × clip_l14_ope_native: mean=0.0136
338
+ dinov2_b14_shared × siglip_b16_3_native: mean=0.2824
339
+ siglip_b16_3_shared × clip_l14_ope_native: mean=0.2211
340
+ siglip_b16_3_shared × dinov2_b14 _native: mean=0.1836
341
+
342
+ Native triangulation divergence from shared:
343
+ clip_l14_openai : tri_cos=0.2093 tri_diff=0.0572
344
+ dinov2_b14 : tri_cos=0.2464 tri_diff=0.0573
345
+ siglip_b16_384 : tri_cos=0.3196 tri_diff=0.0553
346
+
347
+ Native pairwise triangulation correlation:
348
+ clip_l14_openai × dinov2_b14 : mean=0.2059 std=0.0037
349
+ clip_l14_openai × siglip_b16_384 : mean=0.2078 std=0.0041
350
+ dinov2_b14 × siglip_b16_384 : mean=0.7933 std=0.0048
351
+
352
+ Information content:
353
+ Shared tri eff_dim: 71.6
354
+ Combined eff_dim: 126.2
355
+ Info gain from native: +54.5 dims
356
+
357
+ ======================================================================
358
+ SCAN 10: TRIANGULATION STRUCTURE
359
+ ======================================================================
360
+
361
+ clip_l14_openai triangulation:
362
+ mean=0.9993 std=0.0625
363
+ min=0.2505 max=1.3152
364
+ nearest: mean=0.5721 std=0.0978
365
+
366
+ dinov2_b14 triangulation:
367
+ mean=0.9993 std=0.0625
368
+ min=0.2505 max=1.3152
369
+ nearest: mean=0.5721 std=0.0978
370
+
371
+ siglip_b16_384 triangulation:
372
+ mean=0.9993 std=0.0625
373
+ min=0.2505 max=1.3152
374
+ nearest: mean=0.5721 std=0.0978
375
+
376
+ Expert triangulation correlation:
377
+ clip_l14_openai × dinov2_b14 : per_img_cos mean=1.0000 std=0.0000
378
+ clip_l14_openai × siglip_b16_384 : per_img_cos mean=1.0000 std=0.0000
379
+ dinov2_b14 × siglip_b16_384 : per_img_cos mean=1.0000 std=0.0000
380
+
381
+ ======================================================================
382
+ SUMMARY
383
+ ======================================================================
384
+ Checkpoint: checkpoints/dual_stream_best.pt
385
+ mAP: 0.838
386
+ Anchors: 512 × 256-d, 201 active (39%)
387
+ Embedding eff_dim: 71.6/256
388
+ Anchor eff_rank: 256.0/256
389
+ Global CV: 0.2199
390
+ Anchor CV: 0.0286
391
+ Local CV (mean): 0.4302
392
+ Utilization entropy: 73.4%
393
+ Utilization Gini: 0.8568
394
+
395
+ ======================================================================
396
+ ANALYSIS COMPLETE
397
+ ======================================================================