Results
Every row is one run record in results/scripts/report.py. Arms are grouped by backbone, block
configuration and budget, so two different budgets never land in the same average. Read the paired
tables rather than the two means; the protocol and interactive figures are on Experiments ;
cell-by-cell verdicts are on Judgment lines .
Every protocol-matrix best.pt and its full training arguments live on the
checkpoints branchmain; that branch's README states how many): every bucket and routing number recomputes from them.
Fig. 1 Paired deltas across seven generations per-seed values and means of the cells below
Data: results/summary.md
Runs
run
variant
seed
mAP50
mAP50-95
params
wall_s
yolo-master-n-baseline-e1-s3-smokef-fp32-fork-20260912095448
yolo-master-n-baseline@e1f1i800fp32[metaxc500/metax3.3/yolo-master]
3
0.0035
0.0010
2291406
222.6
yolo-master-n-baseline-e120-s0-p800c-20260911165332
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3]
0
0.3597
0.2065
2291406
15410.2
yolo-master-n-baseline-e120-s0-p800c-fork-20260911165202
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3/yolo-master]
0
0.3542
0.2044
2291406
25014.6
yolo-master-n-baseline-e120-s0-p800f-fp32-20260913125527
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3]
0
0.3559
0.2053
2291406
17739.9
yolo-master-n-baseline-e120-s0-p800f-fp32-fork-20260912204616
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
0
0.3532
0.2037
2291406
19895.9
yolo-master-n-baseline-e120-s1-p800d-20260911165331
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3]
1
0.3589
0.2048
2291406
15423.1
yolo-master-n-baseline-e120-s1-p800d-fork-20260911165201
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3/yolo-master]
1
0.3534
0.2027
2291406
25086.3
yolo-master-n-baseline-e120-s1-p800f-fp32-20260913130105
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3]
1
0.3490
0.2020
2291406
17828.0
yolo-master-n-baseline-e120-s1-p800f-fp32-fork-20260912205104
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.3552
0.2047
2291406
19849.1
yolo-master-n-baseline-e120-s2-p800e-20260911204822
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3]
2
0.3598
0.2069
2291406
15448.8
yolo-master-n-baseline-e120-s2-p800e-fork-20260911204651
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3/yolo-master]
2
0.3514
0.2010
2291406
25266.5
yolo-master-n-baseline-e120-s2-p800f-fp32-20260913092752
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3]
2
0.3521
0.2017
2291406
17626.8
yolo-master-n-baseline-e120-s2-p800f-fp32-fork-20260914011826
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
2
0.3598
0.2067
2291406
19861.3
yolo-master-n-esmoe-upstream-w1-e1-s3-smokef-fp32-20260912095449
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e1f1i800fp32[metaxc500/metax3.3]
3
0.0014
0.0003
2814406
263.7
yolo-master-n-esmoe-upstream-w1-e120-s0-p800c-20260911234908
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
0
0.3739
0.2168
2814406
16519.5
yolo-master-n-esmoe-upstream-w1-e120-s0-p800f-fp32-20260913021802
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
0
0.3601
0.2083
2814406
38233.5
yolo-master-n-esmoe-upstream-w1-e120-s1-p800d-20260911235018
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
1
0.3724
0.2163
2814406
17059.5
yolo-master-n-esmoe-upstream-w1-e120-s1-p800f-fp32-20260913022205
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
1
0.3578
0.2067
2814406
38328.1
yolo-master-n-esmoe-upstream-w1-e120-s2-p800e-20260912034809
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
2
0.3698
0.2139
2814406
16814.1
yolo-master-n-esmoe-upstream-w1-e120-s2-p800f-fp32-20260912225003
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
2
0.3677
0.2138
2814406
38258.6
yolo-master-n-upstream-e1-s3-smokef-fp32-fork-20260912095449
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e1f1i800fp32[metaxc500/metax3.3/yolo-master]
3
0.0016
0.0003
2814406
271.3
yolo-master-n-upstream-e120-s0-p800c-fork-20260911060342
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
0
0.3637
0.2098
2814406
38885.7
yolo-master-n-upstream-e120-s0-p800f-fp32-fork-20260912101056
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
0
0.3674
0.2127
2814406
38110.4
yolo-master-n-upstream-e120-s1-p800d-fork-20260911060354
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
1
0.3703
0.2146
2814406
38832.9
yolo-master-n-upstream-e120-s1-p800f-fp32-fork-20260912101226
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.3651
0.2126
2814406
38257.1
yolo-master-n-upstream-e120-s2-p800e-fork-20260911100545
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
2
0.3615
0.2094
2814406
38429.7
yolo-master-n-upstream-e120-s2-p800f-fp32-fork-20260912101356
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
2
0.3669
0.2127
2814406
38126.9
yolo11n-baseline-e120-s0-p800-20260902072608
yolo11n-baseline@e120f1i800[4090/torch2.6]
0
0.3539
0.2027
2591790
5366.7
yolo11n-baseline-e120-s1-p800-20260902121409
yolo11n-baseline@e120f1i800[4090/torch2.6]
1
0.3566
0.2048
2591790
5461.5
yolo11n-baseline-e120-s2-p800-20260902165904
yolo11n-baseline@e120f1i800[4090/torch2.6]
2
0.3545
0.2046
2591790
5451.4
yolo11n-baseline-e20-s0-full-20260830032002
yolo11n-baseline@e20f1i640[4090d/torch2.6]
0
0.1413
0.0716
2591790
666.1
yolo11n-baseline-e20-s1-full-20260830034325
yolo11n-baseline@e20f1i640[4090d/torch2.6]
1
0.1391
0.0701
2591790
657.6
yolo11n-baseline-e20-s2-full-20260830040552
yolo11n-baseline@e20f1i640[4090d/torch2.6]
2
0.1453
0.0743
2591790
644.3
yolo11n-esmoe-e1-s0-compat-20260829113851
yolo11n-e4k2w0.01@e1f0.05i640[4090d/torch2.6]
0
0.0000
0.0000
2906322
30.0
yolo11n-esmoe-e120-s0-p800-20260902085539
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.3567
0.2059
2906322
5934.9
yolo11n-esmoe-e120-s1-p800-20260902134514
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.3593
0.2038
2906322
5813.3
yolo11n-esmoe-e120-s2-p800-20260902182959
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.3530
0.2026
2906322
5862.7
yolo11n-esmoe-e20-s0-full-20260830033112
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
0
0.1458
0.0748
2906322
728.3
yolo11n-esmoe-e20-s1-full-20260830035427
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
1
0.1358
0.0692
2906322
679.8
yolo11n-esmoe-e20-s2-full-20260830041640
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
2
0.1413
0.0715
2906322
702.8
yolo11n-esmoe-rewire-e120-s0-p800-20260902103439
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.3547
0.2024
2906322
5965.0
yolo11n-esmoe-rewire-e120-s1-p800-20260902152212
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.3494
0.2023
2906322
5807.0
yolo11n-esmoe-rewire-e120-s2-p800-20260902200746
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.3622
0.2082
2906322
5751.0
yolo12n-baseline-e120-s0-p800-20260902214347
yolo12n-baseline@e120f1i800[4090/torch2.6]
0
0.3585
0.2072
2569998
7642.4
yolo12n-baseline-e120-s1-p800-20260903020351
yolo12n-baseline@e120f1i800[4090/torch2.6]
1
0.3569
0.2057
2569998
7712.6
yolo12n-baseline-e120-s2-p800-20260903062742
yolo12n-baseline@e120f1i800[4090/torch2.6]
2
0.3571
0.2051
2569998
7550.8
yolo12n-esmoe-e1-s0-compat-20260829113926
yolo12n-e4k2w0.01@e1f0.05i640[4090d/torch2.6]
0
0.0000
0.0000
2884530
35.7
yolo12n-esmoe-e120-s0-p800-20260902235113
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.3563
0.2040
2884530
7953.8
yolo12n-esmoe-e120-s1-p800-20260903041228
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.3543
0.2029
2884530
8109.1
yolo12n-esmoe-e120-s2-p800-20260903083337
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.3565
0.2049
2884530
7963.6
yolo12n-esmoe-rewire-e120-s0-p800-20260903235327
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.3597
0.2084
2884530
7945.0
yolo12n-esmoe-rewire-e120-s1-p800-20260904020557
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.3541
0.2047
2884530
8035.2
yolo12n-esmoe-rewire-e120-s2-p800-20260904041957
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.3590
0.2067
2884530
8116.0
yolo26n-baseline-e120-s0-p800-20260903104625
yolo26n-baseline@e120f1i800[4090/torch2.6]
0
0.3432
0.1939
2507700
7659.5
yolo26n-baseline-e120-s1-p800-20260903150829
yolo26n-baseline@e120f1i800[4090/torch2.6]
1
0.3444
0.1965
2507700
7623.7
yolo26n-baseline-e120-s2-p800-20260903193013
yolo26n-baseline@e120f1i800[4090/torch2.6]
2
0.3416
0.1946
2507700
7682.9
yolo26n-esmoe-e1-s0-compat-20260830113739
yolo26n-e4k2w0.01@e1f0.05i640[4090d/torch2.6]
0
0.0000
0.0000
2822232
27.1
yolo26n-esmoe-e120-s0-p800-20260903125409
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.3374
0.1916
2822232
8055.6
yolo26n-esmoe-e120-s1-p800-20260903171537
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.3410
0.1934
2822232
8072.0
yolo26n-esmoe-e120-s2-p800-20260903213821
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.3406
0.1935
2822232
7988.1
yolo26n-esmoe-rewire-e120-s0-p800-20260904063517
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.3399
0.1945
2822232
7961.4
yolo26n-esmoe-rewire-e120-s1-p800-20260904084803
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.3436
0.1965
2822232
8120.5
yolo26n-esmoe-rewire-e120-s2-p800-20260904110328
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.3442
0.1965
2822232
8154.6
yolov10n-baseline-e120-s0-p800-20260907040937
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
0
0.3506
0.2028
2710940
16687.9
yolov10n-baseline-e120-s1-p800-20260907090359
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
1
0.3540
0.2025
2710940
16804.2
yolov10n-baseline-e120-s2-p800-20260907182750
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
2
0.3454
0.1981
2710940
16893.6
yolov10n-baseline-e120-s3-p800-20260907232405
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
3
0.3532
0.2022
2710940
16914.6
yolov10n-baseline-e120-s4-p800-20260908085304
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
4
0.3523
0.2003
2710940
16870.7
yolov10n-esmoe-norm-e1-s0-newsmoke-20260909122747
yolov10n-e4k2w0.01-gshard@e1f0.02i320[metaxc500/metax3.3]
0
0.0000
0.0000
3025472
49.0
yolov10n-esmoe-dense-e120-s0-p800-20260909232756
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
0
0.3503
0.2005
3025472
17357.6
yolov10n-esmoe-dense-e120-s1-p800-20260910005225
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
1
0.3500
0.2004
3025472
17548.1
yolov10n-esmoe-dense-e120-s2-p800-20260910041726
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
2
0.3500
0.2025
3025472
17069.9
yolov10n-esmoe-e1-s0-smoke-20260907035626
yolov10n-e4k2w0.01@e1f0.02i640[metaxc500/metax3.3]
0
0.0000
0.0000
3025472
81.4
yolov10n-esmoe-e120-s0-p800-20260907040938
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.3513
0.2025
3025472
17650.8
yolov10n-esmoe-e120-s1-p800-20260907134026
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.3511
0.2005
3025472
17232.5
yolov10n-esmoe-e120-s2-p800-20260907183157
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.3476
0.1989
3025472
17516.5
yolov10n-esmoe-e120-s3-p800-20260908040210
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
0.3523
0.2022
3025472
17442.7
yolov10n-esmoe-e120-s4-p800-20260908085441
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
4
0.3525
0.2001
3025472
17716.5
yolov10n-esmoe-norm-e120-s0-p800-20260909184058
yolov10n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
0
0.3506
0.1994
3025984
17207.8
yolov10n-esmoe-rewire-e1-s0-smoke-20260907035758
yolov10n-e4k2w0.01-rewire@e1f0.02i640[metaxc500/metax3.3]
0
0.0000
0.0000
3025472
40.6
yolov10n-esmoe-rewire-e120-s0-p800-20260907084755
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.3505
0.2015
3025472
17538.1
yolov10n-esmoe-rewire-e120-s1-p800-20260907134414
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.3460
0.1971
3025472
17252.3
yolov10n-esmoe-rewire-e120-s2-p800-20260907230936
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.3512
0.2030
3025472
17539.3
yolov10n-esmoe-rewire-e120-s3-p800-20260908040610
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
0.3444
0.1978
3025472
17300.3
yolov10n-esmoe-rewire-e120-s4-p800-20260908133426
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
4
0.3477
0.1984
3025472
11625.8
yolov10n-esmoe-rewire-stages-e1-s0-newsmoke-20260909122638
yolov10n-e4k2w0.01-rewire-gshard-x4@e1f0.02i320[metaxc500/metax3.3]
0
0.0000
0.0000
3153004
57.5
yolov10n-esmoe-rewire-stages-e120-s0-p800-20260909124212
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.3451
0.1969
3153004
21344.8
yolov10n-esmoe-rewire-stages-e120-s1-p800-20260909124342
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.3427
0.1961
3153004
21424.5
yolov10n-esmoe-rewire-stages-e120-s2-p800-20260909183813
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.3408
0.1961
3153004
22441.3
yolov10n-esmoe-rewire-stages-w0.0025-e120-s0-p800-20260910054505
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.3448
0.1967
3153004
21707.9
yolov10n-esmoe-rewire-stages-w0.0025-e120-s1-p800-20260910090207
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.3419
0.1963
3153004
21319.7
yolov10n-esmoe-rewire-stages-w0.0025-e120-s2-p800-20260910114704
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.3294
0.1880
3153004
21938.0
yolov10n-esmoe-w0-e120-s0-p800-20260910145738
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
0
0.3497
0.2005
3025472
16796.3
yolov10n-esmoe-w0-e120-s1-p800-20260910175253
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
1
0.3449
0.1960
3025472
17019.5
yolov10n-esmoe-w0-e120-s2-p800-20260910193750
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
2
0.3474
0.1994
3025472
14753.2
yolov5n-baseline-e120-s0-p800-20260905001430
yolov5n-baseline@e120f1i800[metaxc500/metax3.7]
0
0.3451
0.1970
2510414
10156.3
yolov5n-baseline-e120-s0-p800h2-20260908005500
yolov5n-baseline@e120f1i800[metaxc500/metax3.3]
0
0.3510
0.1989
2510414
10325.1
yolov5n-baseline-e120-s1-p800-20260905060501
yolov5n-baseline@e120f1i800[metaxc500/metax3.7]
1
0.3480
0.1982
2510414
10157.3
yolov5n-baseline-e120-s1-p800h2-20260908065553
yolov5n-baseline@e120f1i800[metaxc500/metax3.3]
1
0.3522
0.2019
2510414
11944.6
yolov5n-baseline-e120-s2-p800-20260905115341
yolov5n-baseline@e120f1i800[metaxc500/metax3.7]
2
0.3514
0.2003
2510414
10007.4
yolov5n-baseline-e120-s2-p800h2-20260908132744
yolov5n-baseline@e120f1i800[metaxc500/metax3.3]
2
0.3494
0.1991
2510414
11702.7
yolov5n-esmoe-dense-e120-s0-p800h2-20260909183321
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
0
0.3530
0.2023
2824946
12451.5
yolov5n-esmoe-dense-e120-s1-p800h2-20260909220105
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
1
0.3547
0.2036
2824946
12143.4
yolov5n-esmoe-dense-e120-s2-p800h2-20260909220236
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
2
0.3542
0.2041
2824946
12168.3
yolov5n-esmoe-e120-s0-p800-20260905030400
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
0
0.3531
0.2015
2824946
10847.5
yolov5n-esmoe-e120-s0-p800h2-20260908034716
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.3562
0.2041
2824946
11303.1
yolov5n-esmoe-e120-s1-p800-20260905085434
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
1
0.3508
0.2001
2824946
10732.2
yolov5n-esmoe-e120-s1-p800h2-20260908100120
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.3526
0.2019
2824946
12372.9
yolov5n-esmoe-e120-s2-p800-20260905144045
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
2
0.3529
0.2019
2824946
10629.4
yolov5n-esmoe-e120-s2-p800h2-20260908134159
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.3604
0.2052
2824946
12839.7
yolov5n-esmoe-gshard-e120-s2-p800h2g-20260910075558
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
2
0.3548
0.2020
2824946
12194.6
yolov5n-esmoe-master-e120-s0-p800h2-20260909113657
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
0
0.3562
0.2019
2824946
12295.2
yolov5n-esmoe-master-e120-s1-p800h2-20260909113657
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
1
0.3506
0.1997
2824946
12359.6
yolov5n-esmoe-master-w0.32-e120-s0-p800h2-20260910144142
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
0
0.3514
0.2016
2824946
12148.7
yolov5n-esmoe-master-w0.32-e120-s1-p800h2-20260910180421
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
1
0.3556
0.2015
2824946
12059.2
yolov5n-esmoe-master-w0.32-e120-s2-p800h2-20260910193230
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
2
0.3519
0.1994
2824946
10385.5
yolov5n-esmoe-norm-e120-s0-p800h2-20260909150352
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
0
0.3516
0.2009
2825458
12468.4
yolov5n-esmoe-norm-e120-s1-p800h2-20260909150522
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
1
0.3569
0.2022
2825458
12452.4
yolov5n-esmoe-norm-e120-s2-p800h2-20260909183150
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
2
0.3556
0.2035
2825458
12546.2
yolov5n-esmoe-rewire-e120-s0-p800-20260905173809
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.7]
0
0.3525
0.2032
2824946
10835.3
yolov5n-esmoe-rewire-e120-s0-p800h2-20260908063042
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.3552
0.2025
2824946
12626.1
yolov5n-esmoe-rewire-e120-s1-p800h2-20260908101509
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.3512
0.2000
2824946
12400.0
yolov5n-esmoe-rewire-e120-s2-p800h2-20260908164258
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.3534
0.2027
2824946
8832.2
yolov5n-esmoe-rewire-master-stages-norm-dense-e1-s0-parity-20260909135148
yolov5n-e4k2w0.01-rewire-master-norm-dense-t0.4-x4@e1f0.02i320[metaxc500/metax3.3]
0
0.0000
0.0000
2953438
55.1
yolov5n-esmoe-rewire-stages-e120-s0-p800h2-20260910031541
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.3432
0.1946
2952478
16805.6
yolov5n-esmoe-rewire-stages-e120-s1-p800h2-20260910031711
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.3498
0.2001
2952478
16819.0
yolov5n-esmoe-rewire-stages-e120-s2-p800h2-20260910143932
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.3448
0.1944
2952478
17567.9
yolov5n-esmoe-w0-e120-s0-p800h2-20260910075740
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
0
0.3479
0.1985
2824946
12098.3
yolov5n-esmoe-w0-e120-s1-p800h2-20260910111929
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
1
0.3531
0.2010
2824946
11991.4
yolov5n-esmoe-w0-e120-s2-p800h2-20260910112059
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
2
0.3532
0.2021
2824946
12032.4
yolov8n-baseline-e100-s0-full-20260830042827
yolov8n-baseline@e100f1i640[4090d/torch2.6]
0
0.2939
0.1649
3012798
3956.6
yolov8n-baseline-e100-s1-full-20260830064202
yolov8n-baseline@e100f1i640[4090d/torch2.6]
1
0.3032
0.1690
3012798
3661.2
yolov8n-baseline-e100-s2-full-20260830084723
yolov8n-baseline@e100f1i640[4090d/torch2.6]
2
0.3042
0.1694
3012798
3694.5
yolov8n-baseline-e120-s0-p800-20260901015653
yolov8n-baseline@e120f1i800[4090/torch2.6]
0
0.3578
0.2058
3012798
4699.0
yolov8n-baseline-e120-s1-p800-20260901044411
yolov8n-baseline@e120f1i800[4090/torch2.6]
1
0.3622
0.2093
3012798
4702.0
yolov8n-baseline-e120-s2-p800-20260901072757
yolov8n-baseline@e120f1i800[4090/torch2.6]
2
0.3615
0.2074
3012798
4809.7
yolov8n-baseline-e20-s0-20260829080043
yolov8n-baseline@e20f0.25i640[4090d/torch2.6]
0
0.0933
0.0419
3012798
362.8
yolov8n-baseline-e20-s0-full-20260829145207
yolov8n-baseline@e20f1i640[4090d/torch2.6]
0
0.1516
0.0776
3012798
789.6
yolov8n-baseline-e20-s1-20260829081257
yolov8n-baseline@e20f0.25i640[4090d/torch2.6]
1
0.0892
0.0393
3012798
365.5
yolov8n-baseline-e20-s1-full-20260829151916
yolov8n-baseline@e20f1i640[4090d/torch2.6]
1
0.1473
0.0744
3012798
805.2
yolov8n-baseline-e20-s2-20260829082514
yolov8n-baseline@e20f0.25i640[4090d/torch2.6]
2
0.0903
0.0402
3012798
389.9
yolov8n-baseline-e20-s2-full-20260829154640
yolov8n-baseline@e20f1i640[4090d/torch2.6]
2
0.1499
0.0756
3012798
790.9
yolov8n-baseline-e50-s0-full-20260829233559
yolov8n-baseline@e50f1i640[4090d/torch2.6]
0
0.2134
0.1128
3012798
1423.2
yolov8n-baseline-e50-s1-full-20260830002429
yolov8n-baseline@e50f1i640[4090d/torch2.6]
1
0.2132
0.1134
3012798
1598.3
yolov8n-baseline-e50-s2-full-20260830011913
yolov8n-baseline@e50f1i640[4090d/torch2.6]
2
0.2126
0.1129
3012798
1536.0
yolov8n-esmoe-e1-s0-compat-20260829113816
yolov8n-e4k2w0.01@e1f0.05i640[4090d/torch2.6]
0
0.0000
0.0000
3327330
31.0
yolov8n-esmoe-e100-s0-full-20260830053429
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
0
0.3043
0.1700
3327330
4049.2
yolov8n-esmoe-e100-s1-full-20260830074307
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
1
0.3021
0.1677
3327330
3851.5
yolov8n-esmoe-e100-s2-full-20260830094902
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
2
0.3086
0.1719
3327330
3875.6
yolov8n-esmoe-e120-s0-p800-20260901031517
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.3608
0.2056
3327330
5329.0
yolov8n-esmoe-e120-s1-p800-20260901060237
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.3602
0.2078
3327330
5115.8
yolov8n-esmoe-e120-s2-p800-20260901084811
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.3681
0.2102
3327330
5149.3
yolov8n-esmoe-e20-s0-20260829080650
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0950
0.0424
3327330
362.7
yolov8n-esmoe-e20-s0-e2k1w0.01-20260829083842
yolov8n-e2k1w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0903
0.0399
3161888
363.0
yolov8n-esmoe-e20-s0-e4k1w0.01-20260829084450
yolov8n-e4k1w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0928
0.0415
3327330
391.4
yolov8n-esmoe-e20-s0-e4k2w0.0-20260829090326
yolov8n-e4k2w0.0@e20f0.25i640[4090d/torch2.6]
0
0.0938
0.0418
3327330
402.0
yolov8n-esmoe-e20-s0-e8k2w0.01-20260829085727
yolov8n-e8k2w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0934
0.0414
3781094
355.0
yolov8n-esmoe-e20-s0-full-20260829150523
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
0
0.1517
0.0779
3327330
828.9
yolov8n-esmoe-e20-s1-20260829081907
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
1
0.0907
0.0399
3327330
362.7
yolov8n-esmoe-e20-s1-full-20260829153247
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
1
0.1505
0.0773
3327330
826.0
yolov8n-esmoe-e20-s2-20260829083149
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
2
0.0932
0.0406
3327330
378.9
yolov8n-esmoe-e20-s2-full-20260829155956
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
2
0.1528
0.0783
3327330
827.9
yolov8n-esmoe-e50-s0-full-20260829235947
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
0
0.2131
0.1135
3327330
1478.3
yolov8n-esmoe-e50-s1-full-20260830005112
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
1
0.2141
0.1137
3327330
1677.1
yolov8n-esmoe-e50-s2-full-20260830014455
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
2
0.2158
0.1142
3327330
1985.8
yolov8n-esmoe-rewire-e120-s0-p800-20260902031114
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.3668
0.2091
3327330
5159.6
yolov8n-esmoe-rewire-e120-s1-p800-20260902043718
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.3634
0.2090
3327330
5066.8
yolov8n-esmoe-rewire-e120-s2-p800-20260902060149
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.3622
0.2079
3327330
5054.6
yolov9t-baseline-e120-s0-p800-20260904155850
yolov9t-baseline@e120f1i800[4090/torch2.6]
0
0.3674
0.2114
2007358
9482.7
yolov9t-baseline-e120-s0-p800-20260907040926
yolov9t-baseline@e120f1i800[metaxc500/metax3.3]
0
0.3661
0.2118
2007358
17909.5
yolov9t-baseline-e120-s1-p800-20260907092351
yolov9t-baseline@e120f1i800[metaxc500/metax3.3]
1
0.3655
0.2118
2007358
17989.0
yolov9t-baseline-e120-s2-p800-20260907194231
yolov9t-baseline@e120f1i800[metaxc500/metax3.3]
2
0.3696
0.2140
2007358
17791.7
yolov9t-esmoe-e1-s0-smoke-20260907035621
yolov9t-e4k2w0.01@e1f0.02i640[metaxc500/metax3.3]
0
0.0000
0.0000
2097042
106.9
yolov9t-esmoe-e120-s0-p800-20260907040927
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.3679
0.2124
2097042
18853.4
yolov9t-esmoe-e120-s1-p800-20260907142354
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.3736
0.2153
2097042
19118.0
yolov9t-esmoe-e120-s2-p800-20260907194245
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.3671
0.2123
2097042
18725.5
yolov9t-esmoe-rewire-e1-s0-smoke-20260907035819
yolov9t-e4k2w0.01-rewire@e1f0.02i640[metaxc500/metax3.3]
0
0.0000
0.0000
2097042
58.0
yolov9t-esmoe-rewire-e120-s0-p800-20260907090808
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.3603
0.2067
2097042
19033.1
yolov9t-esmoe-rewire-e120-s1-p800-20260907142537
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.3633
0.2083
2097042
18986.5
yolov9t-esmoe-rewire-e120-s2-p800-20260908003915
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.3736
0.2163
2097042
21075.9
Across seeds
variant
seeds
mAP50
mAP50-95
yolo-master-n-baseline@e1f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.0035
0.0010
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3]
3
0.3595 ± 0.0005
0.2061 ± 0.0011
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3/yolo-master]
3
0.3530 ± 0.0015
0.2027 ± 0.0017
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3]
3
0.3523 ± 0.0035
0.2030 ± 0.0020
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
3
0.3560 ± 0.0034
0.2050 ± 0.0015
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e1f1i800fp32[metaxc500/metax3.3]
1
0.0014
0.0003
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
3
0.3720 ± 0.0021
0.2157 ± 0.0016
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
3
0.3619 ± 0.0052
0.2096 ± 0.0037
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e1f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.0016
0.0003
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
3
0.3652 ± 0.0046
0.2113 ± 0.0029
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
3
0.3665 ± 0.0012
0.2127 ± 0.0000
yolo11n-baseline@e120f1i800[4090/torch2.6]
3
0.3550 ± 0.0014
0.2041 ± 0.0012
yolo11n-baseline@e20f1i640[4090d/torch2.6]
3
0.1419 ± 0.0032
0.0720 ± 0.0021
yolo11n-e4k2w0.01@e1f0.05i640[4090d/torch2.6]
1
0.0000
0.0000
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
0.3563 ± 0.0032
0.2041 ± 0.0017
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
3
0.1410 ± 0.0050
0.0719 ± 0.0028
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
0.3554 ± 0.0064
0.2043 ± 0.0034
yolo12n-baseline@e120f1i800[4090/torch2.6]
3
0.3575 ± 0.0009
0.2060 ± 0.0011
yolo12n-e4k2w0.01@e1f0.05i640[4090d/torch2.6]
1
0.0000
0.0000
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
0.3557 ± 0.0012
0.2040 ± 0.0010
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
0.3576 ± 0.0031
0.2066 ± 0.0019
yolo26n-baseline@e120f1i800[4090/torch2.6]
3
0.3431 ± 0.0014
0.1950 ± 0.0014
yolo26n-e4k2w0.01@e1f0.05i640[4090d/torch2.6]
1
0.0000
0.0000
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
0.3397 ± 0.0019
0.1929 ± 0.0011
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
0.3426 ± 0.0023
0.1958 ± 0.0012
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
5
0.3511 ± 0.0034
0.2012 ± 0.0020
yolov10n-e4k2w0.01-gshard@e1f0.02i320[metaxc500/metax3.3]
1
0.0000
0.0000
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
3
0.3501 ± 0.0002
0.2011 ± 0.0012
yolov10n-e4k2w0.01@e1f0.02i640[metaxc500/metax3.3]
1
0.0000
0.0000
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
5
0.3509 ± 0.0020
0.2008 ± 0.0015
yolov10n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
1
0.3506
0.1994
yolov10n-e4k2w0.01-rewire@e1f0.02i640[metaxc500/metax3.3]
1
0.0000
0.0000
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
5
0.3480 ± 0.0029
0.1996 ± 0.0026
yolov10n-e4k2w0.01-rewire-gshard-x4@e1f0.02i320[metaxc500/metax3.3]
1
0.0000
0.0000
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
3
0.3429 ± 0.0021
0.1964 ± 0.0005
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
3
0.3387 ± 0.0082
0.1936 ± 0.0049
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
3
0.3473 ± 0.0024
0.1986 ± 0.0023
yolov5n-baseline@e120f1i800[metaxc500/metax3.7]
3
0.3482 ± 0.0031
0.1985 ± 0.0017
yolov5n-baseline@e120f1i800[metaxc500/metax3.3]
3
0.3509 ± 0.0014
0.1999 ± 0.0017
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
3
0.3540 ± 0.0009
0.2033 ± 0.0009
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
3
0.3523 ± 0.0013
0.2011 ± 0.0009
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
0.3564 ± 0.0039
0.2037 ± 0.0016
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
3
0.3539 ± 0.0029
0.2012 ± 0.0013
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
3
0.3530 ± 0.0023
0.2008 ± 0.0013
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
3
0.3547 ± 0.0027
0.2022 ± 0.0013
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.7]
1
0.3525
0.2032
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
0.3532 ± 0.0020
0.2017 ± 0.0015
yolov5n-e4k2w0.01-rewire-master-norm-dense-t0.4-x4@e1f0.02i320[metaxc500/metax3.3]
1
0.0000
0.0000
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
3
0.3459 ± 0.0035
0.1964 ± 0.0032
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
3
0.3514 ± 0.0030
0.2005 ± 0.0018
yolov8n-baseline@e100f1i640[4090d/torch2.6]
3
0.3004 ± 0.0057
0.1678 ± 0.0025
yolov8n-baseline@e120f1i800[4090/torch2.6]
3
0.3605 ± 0.0024
0.2075 ± 0.0018
yolov8n-baseline@e20f0.25i640[4090d/torch2.6]
3
0.0909 ± 0.0022
0.0405 ± 0.0013
yolov8n-baseline@e20f1i640[4090d/torch2.6]
3
0.1496 ± 0.0022
0.0759 ± 0.0016
yolov8n-baseline@e50f1i640[4090d/torch2.6]
3
0.2131 ± 0.0004
0.1130 ± 0.0003
yolov8n-e4k2w0.01@e1f0.05i640[4090d/torch2.6]
1
0.0000
0.0000
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
3
0.3050 ± 0.0033
0.1699 ± 0.0021
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
0.3631 ± 0.0044
0.2079 ± 0.0023
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
3
0.0930 ± 0.0022
0.0410 ± 0.0013
yolov8n-e2k1w0.01@e20f0.25i640[4090d/torch2.6]
1
0.0903
0.0399
yolov8n-e4k1w0.01@e20f0.25i640[4090d/torch2.6]
1
0.0928
0.0415
yolov8n-e4k2w0.0@e20f0.25i640[4090d/torch2.6]
1
0.0938
0.0418
yolov8n-e8k2w0.01@e20f0.25i640[4090d/torch2.6]
1
0.0934
0.0414
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
3
0.1517 ± 0.0011
0.0778 ± 0.0005
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
3
0.2143 ± 0.0014
0.1138 ± 0.0004
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
0.3641 ± 0.0024
0.2087 ± 0.0007
yolov9t-baseline@e120f1i800[4090/torch2.6]
1
0.3674
0.2114
yolov9t-baseline@e120f1i800[metaxc500/metax3.3]
3
0.3671 ± 0.0023
0.2125 ± 0.0013
yolov9t-e4k2w0.01@e1f0.02i640[metaxc500/metax3.3]
1
0.0000
0.0000
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
0.3695 ± 0.0035
0.2134 ± 0.0017
yolov9t-e4k2w0.01-rewire@e1f0.02i640[metaxc500/metax3.3]
1
0.0000
0.0000
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
0.3657 ± 0.0070
0.2104 ± 0.0051
Paired against baseline - metrics/mAP50(B)
variant
seed
baseline
variant
delta
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
0
0.3597
0.3739
+0.0142
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
0
0.3559
0.3601
+0.0042
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
1
0.3589
0.3724
+0.0135
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
1
0.3490
0.3578
+0.0088
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
2
0.3598
0.3698
+0.0100
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
2
0.3521
0.3677
+0.0156
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e1f1i800fp32[metaxc500/metax3.3/yolo-master]
3
0.0035
0.0016
-0.0019
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
0
0.3542
0.3637
+0.0095
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
0
0.3532
0.3674
+0.0142
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
1
0.3534
0.3703
+0.0170
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.3552
0.3651
+0.0099
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
2
0.3514
0.3615
+0.0101
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
2
0.3598
0.3669
+0.0071
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.3539
0.3567
+0.0028
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.3566
0.3593
+0.0027
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.3545
0.3530
-0.0015
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
0
0.1413
0.1458
+0.0045
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
1
0.1391
0.1358
-0.0033
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
2
0.1453
0.1413
-0.0040
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.3539
0.3547
+0.0008
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.3566
0.3494
-0.0071
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.3545
0.3622
+0.0076
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.3585
0.3563
-0.0022
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.3569
0.3543
-0.0026
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.3571
0.3565
-0.0007
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.3585
0.3597
+0.0012
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.3569
0.3541
-0.0028
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.3571
0.3590
+0.0019
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.3432
0.3374
-0.0057
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.3444
0.3410
-0.0035
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.3416
0.3406
-0.0010
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.3432
0.3399
-0.0033
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.3444
0.3436
-0.0008
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.3416
0.3442
+0.0026
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
0
0.3506
0.3503
-0.0003
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
1
0.3540
0.3500
-0.0041
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
2
0.3454
0.3500
+0.0046
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.3506
0.3513
+0.0007
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.3540
0.3511
-0.0029
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.3454
0.3476
+0.0021
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
0.3532
0.3523
-0.0010
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
4
0.3523
0.3525
+0.0002
yolov10n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
0
0.3506
0.3506
+0.0001
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.3506
0.3505
-0.0000
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.3540
0.3460
-0.0081
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.3454
0.3512
+0.0058
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
0.3532
0.3444
-0.0088
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
4
0.3523
0.3477
-0.0046
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.3506
0.3451
-0.0055
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.3540
0.3427
-0.0113
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.3454
0.3408
-0.0046
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.3506
0.3448
-0.0058
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.3540
0.3419
-0.0122
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.3454
0.3294
-0.0160
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
0
0.3506
0.3497
-0.0009
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
1
0.3540
0.3449
-0.0092
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
2
0.3454
0.3474
+0.0019
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
0
0.3510
0.3530
+0.0020
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
1
0.3522
0.3547
+0.0025
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
2
0.3494
0.3542
+0.0048
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
0
0.3451
0.3531
+0.0080
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.3510
0.3562
+0.0052
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
1
0.3480
0.3508
+0.0027
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.3522
0.3526
+0.0004
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
2
0.3514
0.3529
+0.0015
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.3494
0.3604
+0.0109
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
2
0.3494
0.3548
+0.0054
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
0
0.3510
0.3562
+0.0051
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
1
0.3522
0.3506
-0.0016
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
0
0.3510
0.3514
+0.0004
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
1
0.3522
0.3556
+0.0034
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
2
0.3494
0.3519
+0.0024
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
0
0.3510
0.3516
+0.0006
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
1
0.3522
0.3569
+0.0047
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
2
0.3494
0.3556
+0.0061
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.7]
0
0.3451
0.3525
+0.0074
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.3510
0.3552
+0.0042
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.3522
0.3512
-0.0010
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.3494
0.3534
+0.0039
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.3510
0.3432
-0.0079
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.3522
0.3498
-0.0024
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.3494
0.3448
-0.0046
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
0
0.3510
0.3479
-0.0031
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
1
0.3522
0.3531
+0.0009
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
2
0.3494
0.3532
+0.0037
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
0
0.2939
0.3043
+0.0104
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
1
0.3032
0.3021
-0.0011
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
2
0.3042
0.3086
+0.0044
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.3578
0.3608
+0.0030
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.3622
0.3602
-0.0020
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.3615
0.3681
+0.0066
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0933
0.0950
+0.0017
yolov8n-e2k1w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0933
0.0903
-0.0031
yolov8n-e4k1w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0933
0.0928
-0.0006
yolov8n-e4k2w0.0@e20f0.25i640[4090d/torch2.6]
0
0.0933
0.0938
+0.0005
yolov8n-e8k2w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0933
0.0934
+0.0000
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
0
0.1516
0.1517
+0.0001
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
1
0.0892
0.0907
+0.0015
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
1
0.1473
0.1505
+0.0032
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
2
0.0903
0.0932
+0.0029
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
2
0.1499
0.1528
+0.0029
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
0
0.2134
0.2131
-0.0003
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
1
0.2132
0.2141
+0.0009
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
2
0.2126
0.2158
+0.0032
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.3578
0.3668
+0.0090
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.3622
0.3634
+0.0012
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.3615
0.3622
+0.0007
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.3661
0.3679
+0.0019
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.3655
0.3736
+0.0081
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.3696
0.3671
-0.0025
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.3661
0.3603
-0.0058
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.3655
0.3633
-0.0022
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.3696
0.3736
+0.0040
variant
seeds
mean delta
95% CI
wins
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
3
+0.0126
[+0.0070, +0.0181]
3/3
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
3
+0.0095
[-0.0047, +0.0237]
3/3
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e1f1i800fp32[metaxc500/metax3.3/yolo-master]
1
-0.0019
-
0/1
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
3
+0.0122
[+0.0019, +0.0225]
3/3
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
3
+0.0104
[+0.0015, +0.0193]
3/3
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
+0.0013
[-0.0048, +0.0075]
2/3
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
3
-0.0009
[-0.0126, +0.0108]
1/3
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
+0.0004
[-0.0179, +0.0188]
2/3
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
-0.0018
[-0.0044, +0.0008]
0/3
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
+0.0001
[-0.0063, +0.0064]
2/3
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
-0.0034
[-0.0093, +0.0025]
0/3
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
-0.0005
[-0.0079, +0.0068]
1/3
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
3
+0.0001
[-0.0107, +0.0109]
1/3
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
5
-0.0002
[-0.0025, +0.0022]
3/5
yolov10n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
1
+0.0001
-
1/1
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
5
-0.0031
[-0.0107, +0.0044]
1/5
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
3
-0.0071
[-0.0162, +0.0019]
0/3
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
3
-0.0113
[-0.0242, +0.0015]
0/3
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
3
-0.0027
[-0.0170, +0.0116]
1/3
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
3
+0.0031
[-0.0006, +0.0068]
3/3
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
3
+0.0041
[-0.0045, +0.0127]
3/3
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
+0.0055
[-0.0076, +0.0186]
3/3
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
3
+0.0030
[-0.0069, +0.0128]
2/3
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
3
+0.0021
[-0.0018, +0.0059]
3/3
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
3
+0.0038
[-0.0033, +0.0109]
3/3
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.7]
1
+0.0074
-
1/1
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
+0.0024
[-0.0050, +0.0097]
2/3
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
3
-0.0049
[-0.0118, +0.0019]
0/3
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
3
+0.0005
[-0.0080, +0.0090]
2/3
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
3
+0.0046
[-0.0098, +0.0189]
2/3
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
+0.0025
[-0.0082, +0.0133]
2/3
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
3
+0.0021
[+0.0001, +0.0040]
3/3
yolov8n-e2k1w0.01@e20f0.25i640[4090d/torch2.6]
1
-0.0031
-
0/1
yolov8n-e4k1w0.01@e20f0.25i640[4090d/torch2.6]
1
-0.0006
-
0/1
yolov8n-e4k2w0.0@e20f0.25i640[4090d/torch2.6]
1
+0.0005
-
1/1
yolov8n-e8k2w0.01@e20f0.25i640[4090d/torch2.6]
1
+0.0000
-
1/1
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
3
+0.0021
[-0.0022, +0.0063]
3/3
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
3
+0.0013
[-0.0031, +0.0056]
2/3
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
+0.0036
[-0.0080, +0.0152]
3/3
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
+0.0025
[-0.0108, +0.0157]
2/3
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
-0.0013
[-0.0136, +0.0110]
1/3
Paired against baseline - metrics/mAP50-95(B)
variant
seed
baseline
variant
delta
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
0
0.2065
0.2168
+0.0103
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
0
0.2053
0.2083
+0.0030
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
1
0.2048
0.2163
+0.0115
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
1
0.2020
0.2067
+0.0047
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
2
0.2069
0.2139
+0.0070
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
2
0.2017
0.2138
+0.0121
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e1f1i800fp32[metaxc500/metax3.3/yolo-master]
3
0.0010
0.0003
-0.0006
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
0
0.2044
0.2098
+0.0054
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
0
0.2037
0.2127
+0.0090
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
1
0.2027
0.2146
+0.0120
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.2047
0.2126
+0.0079
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
2
0.2010
0.2094
+0.0084
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
2
0.2067
0.2127
+0.0060
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.2027
0.2059
+0.0032
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.2048
0.2038
-0.0010
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.2046
0.2026
-0.0020
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
0
0.0716
0.0748
+0.0032
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
1
0.0701
0.0692
-0.0008
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
2
0.0743
0.0715
-0.0027
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.2027
0.2024
-0.0003
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.2048
0.2023
-0.0025
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.2046
0.2082
+0.0035
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.2072
0.2040
-0.0032
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.2057
0.2029
-0.0028
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.2051
0.2049
-0.0002
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.2072
0.2084
+0.0011
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.2057
0.2047
-0.0010
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.2051
0.2067
+0.0016
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.1939
0.1916
-0.0022
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.1965
0.1934
-0.0031
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.1946
0.1935
-0.0011
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.1939
0.1945
+0.0006
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.1965
0.1965
-0.0001
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.1946
0.1965
+0.0019
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
0
0.2028
0.2005
-0.0023
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
1
0.2025
0.2004
-0.0021
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
2
0.1981
0.2025
+0.0044
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.2028
0.2025
-0.0003
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.2025
0.2005
-0.0020
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.1981
0.1989
+0.0007
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
0.2022
0.2022
+0.0000
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
4
0.2003
0.2001
-0.0003
yolov10n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
0
0.2028
0.1994
-0.0035
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.2028
0.2015
-0.0013
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.2025
0.1971
-0.0054
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.1981
0.2030
+0.0049
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
0.2022
0.1978
-0.0044
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
4
0.2003
0.1984
-0.0020
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.2028
0.1969
-0.0059
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.2025
0.1961
-0.0064
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.1981
0.1961
-0.0021
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.2028
0.1967
-0.0061
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.2025
0.1963
-0.0062
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.1981
0.1880
-0.0102
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
0
0.2028
0.2005
-0.0023
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
1
0.2025
0.1960
-0.0065
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
2
0.1981
0.1994
+0.0012
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
0
0.1989
0.2023
+0.0035
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
1
0.2019
0.2036
+0.0017
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
2
0.1991
0.2041
+0.0050
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
0
0.1970
0.2015
+0.0044
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.1989
0.2041
+0.0052
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
1
0.1982
0.2001
+0.0019
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.2019
0.2019
+0.0001
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
2
0.2003
0.2019
+0.0015
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.1991
0.2052
+0.0061
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
2
0.1991
0.2020
+0.0029
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
0
0.1989
0.2019
+0.0030
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
1
0.2019
0.1997
-0.0022
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
0
0.1989
0.2016
+0.0028
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
1
0.2019
0.2015
-0.0004
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
2
0.1991
0.1994
+0.0003
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
0
0.1989
0.2009
+0.0020
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
1
0.2019
0.2022
+0.0003
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
2
0.1991
0.2035
+0.0044
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.7]
0
0.1970
0.2032
+0.0062
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.1989
0.2025
+0.0037
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.2019
0.2000
-0.0019
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.1991
0.2027
+0.0036
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.1989
0.1946
-0.0042
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.2019
0.2001
-0.0018
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.1991
0.1944
-0.0047
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
0
0.1989
0.1985
-0.0003
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
1
0.2019
0.2010
-0.0009
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
2
0.1991
0.2021
+0.0030
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
0
0.1649
0.1700
+0.0050
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
1
0.1690
0.1677
-0.0013
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
2
0.1694
0.1719
+0.0024
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.2058
0.2056
-0.0002
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.2093
0.2078
-0.0015
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.2074
0.2102
+0.0028
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0419
0.0424
+0.0006
yolov8n-e2k1w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0419
0.0399
-0.0020
yolov8n-e4k1w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0419
0.0415
-0.0004
yolov8n-e4k2w0.0@e20f0.25i640[4090d/torch2.6]
0
0.0419
0.0418
-0.0000
yolov8n-e8k2w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0419
0.0414
-0.0005
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
0
0.0776
0.0779
+0.0003
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
1
0.0393
0.0399
+0.0006
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
1
0.0744
0.0773
+0.0029
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
2
0.0402
0.0406
+0.0005
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
2
0.0756
0.0783
+0.0027
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
0
0.1128
0.1135
+0.0007
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
1
0.1134
0.1137
+0.0003
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
2
0.1129
0.1142
+0.0013
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.2058
0.2091
+0.0033
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.2093
0.2090
-0.0003
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.2074
0.2079
+0.0005
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.2118
0.2124
+0.0007
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.2118
0.2153
+0.0035
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.2140
0.2123
-0.0016
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.2118
0.2067
-0.0051
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.2118
0.2083
-0.0035
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.2140
0.2163
+0.0023
variant
seeds
mean delta
95% CI
wins
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
3
+0.0096
[+0.0038, +0.0154]
3/3
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
3
+0.0066
[-0.0055, +0.0186]
3/3
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e1f1i800fp32[metaxc500/metax3.3/yolo-master]
1
-0.0006
-
0/1
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
3
+0.0086
[+0.0005, +0.0167]
3/3
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
3
+0.0076
[+0.0038, +0.0115]
3/3
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
+0.0001
[-0.0068, +0.0070]
1/3
yolo11n-e4k2w0.01@e20f1i640[4090d/torch2.6]
3
-0.0001
[-0.0076, +0.0074]
1/3
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
+0.0002
[-0.0074, +0.0079]
1/3
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
-0.0021
[-0.0061, +0.0020]
0/3
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
+0.0006
[-0.0029, +0.0040]
2/3
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
-0.0022
[-0.0047, +0.0004]
0/3
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
+0.0008
[-0.0017, +0.0033]
2/3
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
3
-0.0000
[-0.0095, +0.0095]
1/3
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
5
-0.0003
[-0.0016, +0.0009]
2/5
yolov10n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
1
-0.0035
-
0/1
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
5
-0.0016
[-0.0066, +0.0033]
1/5
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
3
-0.0048
[-0.0106, +0.0010]
0/3
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
3
-0.0075
[-0.0132, -0.0018]
0/3
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
3
-0.0025
[-0.0121, +0.0070]
1/3
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
3
+0.0034
[-0.0007, +0.0075]
3/3
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
3
+0.0026
[-0.0013, +0.0066]
3/3
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
+0.0038
[-0.0043, +0.0119]
3/3
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
3
+0.0013
[-0.0061, +0.0086]
2/3
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
3
+0.0009
[-0.0032, +0.0050]
2/3
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
3
+0.0022
[-0.0029, +0.0074]
3/3
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.7]
1
+0.0062
-
1/1
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
+0.0018
[-0.0061, +0.0096]
2/3
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
3
-0.0035
[-0.0074, +0.0003]
0/3
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
3
+0.0006
[-0.0046, +0.0058]
1/3
yolov8n-e4k2w0.01@e100f1i640[4090d/torch2.6]
3
+0.0021
[-0.0058, +0.0099]
2/3
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
3
+0.0004
[-0.0051, +0.0058]
1/3
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
3
+0.0005
[+0.0004, +0.0007]
3/3
yolov8n-e2k1w0.01@e20f0.25i640[4090d/torch2.6]
1
-0.0020
-
0/1
yolov8n-e4k1w0.01@e20f0.25i640[4090d/torch2.6]
1
-0.0004
-
0/1
yolov8n-e4k2w0.0@e20f0.25i640[4090d/torch2.6]
1
-0.0000
-
0/1
yolov8n-e8k2w0.01@e20f0.25i640[4090d/torch2.6]
1
-0.0005
-
0/1
yolov8n-e4k2w0.01@e20f1i640[4090d/torch2.6]
3
+0.0019
[-0.0017, +0.0055]
3/3
yolov8n-e4k2w0.01@e50f1i640[4090d/torch2.6]
3
+0.0008
[-0.0005, +0.0021]
3/3
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
3
+0.0011
[-0.0035, +0.0058]
2/3
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
+0.0009
[-0.0056, +0.0073]
2/3
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
-0.0021
[-0.0117, +0.0076]
1/3
Determinism (repeated runs of an identical config)
variant
seed
mAP50 first
mAP50 repeat
gap
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3]
0
0.3559
0.3560
0.0001
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.3552
0.3519
0.0033
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3]
2
0.3521
0.3625
0.0103
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
0
0.3601
0.3636
0.0035
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.3651
0.3634
0.0018
yolov10n-e4k2w0.01-gshard@e1f0.02i320[metaxc500/metax3.3]
0
0.0000
0.0000
0.0000
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.3513
0.3483
0.0029
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.3511
0.3489
0.0023
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.3562
0.3571
0.0009
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.3526
0.3494
0.0032
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.3604
0.3474
0.0130
yolov8n-e4k2w0.01@e20f0.25i640[4090d/torch2.6]
0
0.0950
0.0950
0.0000
stack
schedule
repeats
mean gap
largest gap
4090d/torch2.6
e20f0.25i640
1
0.0000
0.0000
metaxc500/metax3.3
e120f1i800
5
0.0045
0.0130
metaxc500/metax3.3
e120f1i800fp32
3
0.0046
0.0103
metaxc500/metax3.3
e1f0.02i320
1
0.0000
0.0000
metaxc500/metax3.3/yolo-master
e120f1i800fp32
2
0.0025
0.0033
Same configuration
One model and one protocol, trained on YOLO-Master's fork and on official ultralytics plus this
package, four arms with a seed to a card: A is the fork running its own yolo-master-n (four
ES_MOE blocks), A0 the same fork without them, B official 8.4.101 with this package's four
ESMoE blocks and recipe="upstream", C official without them. A seed's four arms share a card,
so every difference is paired inside one.
The metrics are re-measured by scripts/measure.py: each run's model rebuilt from the config it
trained, loaded with the EMA weights in its last.pt, and evaluated with one set of validation
settings, so neither framework's own validation enters the comparison; what each trainer measured
stays in its record as metrics_by_trainer. Verdicts follow rounds seven (mixed precision) and eight (FP32) of the
judgment lines .
Fig. 2 Four differences with 95% intervals large dot for the mean, small dots for seeds, grey band for the noise floor
Data: results/same_config.md
metrics/mAP50(B)
precision
seed
card
metrics
A
A0
B
C
A - B
A - A0
B - C
(A - A0) - (B - C)
fp32
0
metaxc500/metax3.3
measured
0.3674
0.3532
0.3601
0.3559
+0.0074
+0.0142
+0.0042
+0.0100
fp32
1
metaxc500/metax3.3
measured
0.3651
0.3552
0.3578
0.3490
+0.0074
+0.0099
+0.0088
+0.0011
fp32
2
metaxc500/metax3.3
measured
0.3669
0.3598
0.3677
0.3521
-0.0008
+0.0071
+0.0156
-0.0085
mixed
0
metaxc500/metax3.3
measured
0.3637
0.3542
0.3739
0.3597
-0.0102
+0.0095
+0.0142
-0.0047
mixed
1
metaxc500/metax3.3
measured
0.3703
0.3534
0.3724
0.3589
-0.0021
+0.0170
+0.0135
+0.0035
mixed
2
metaxc500/metax3.3
measured
0.3615
0.3514
0.3698
0.3598
-0.0084
+0.0101
+0.0100
+0.0001
precision
difference
seeds
mean
95% CI
positive
fp32
A - B
3
+0.0046
[-0.0071, +0.0164]
2/3
fp32
A - A0
3
+0.0104
[+0.0015, +0.0193]
3/3
fp32
B - C
3
+0.0095
[-0.0047, +0.0237]
3/3
fp32
(A - A0) - (B - C)
3
+0.0009
[-0.0221, +0.0239]
2/3
mixed
A - B
3
-0.0069
[-0.0175, +0.0037]
0/3
mixed
A - A0
3
+0.0122
[+0.0019, +0.0225]
3/3
mixed
B - C
3
+0.0126
[+0.0070, +0.0181]
3/3
mixed
(A - A0) - (B - C)
3
-0.0004
[-0.0105, +0.0098]
2/3
metrics/mAP50-95(B)
precision
seed
card
metrics
A
A0
B
C
A - B
A - A0
B - C
(A - A0) - (B - C)
fp32
0
metaxc500/metax3.3
measured
0.2127
0.2037
0.2083
0.2053
+0.0044
+0.0090
+0.0030
+0.0061
fp32
1
metaxc500/metax3.3
measured
0.2126
0.2047
0.2067
0.2020
+0.0060
+0.0079
+0.0047
+0.0033
fp32
2
metaxc500/metax3.3
measured
0.2127
0.2067
0.2138
0.2017
-0.0011
+0.0060
+0.0121
-0.0061
mixed
0
metaxc500/metax3.3
measured
0.2098
0.2044
0.2168
0.2065
-0.0070
+0.0054
+0.0103
-0.0049
mixed
1
metaxc500/metax3.3
measured
0.2146
0.2027
0.2163
0.2048
-0.0017
+0.0120
+0.0115
+0.0005
mixed
2
metaxc500/metax3.3
measured
0.2094
0.2010
0.2139
0.2069
-0.0045
+0.0084
+0.0070
+0.0014
precision
difference
seeds
mean
95% CI
positive
fp32
A - B
3
+0.0031
[-0.0062, +0.0123]
2/3
fp32
A - A0
3
+0.0076
[+0.0038, +0.0115]
3/3
fp32
B - C
3
+0.0066
[-0.0055, +0.0186]
3/3
fp32
(A - A0) - (B - C)
3
+0.0011
[-0.0148, +0.0169]
2/3
mixed
A - B
3
-0.0044
[-0.0110, +0.0022]
0/3
mixed
A - A0
3
+0.0086
[+0.0005, +0.0167]
3/3
mixed
B - C
3
+0.0096
[+0.0038, +0.0154]
3/3
mixed
(A - A0) - (B - C)
3
-0.0010
[-0.0094, +0.0074]
2/3
What each run actually trained with
precision
seed
arm
run
amp asked
amp at end
batch at end
epochs replayed
hours
fp32
0
A
yolo-master-n-upstream-e120-s0-p800f-fp32-fork-20260912101056
False
False
32
0
10.59
fp32
0
A0
yolo-master-n-baseline-e120-s0-p800f-fp32-fork-20260912204616
False
False
32
0
5.53
fp32
0
B
yolo-master-n-esmoe-upstream-w1-e120-s0-p800f-fp32-20260913021802
False
False
32
0
10.62
fp32
0
C
yolo-master-n-baseline-e120-s0-p800f-fp32-20260913125527
False
False
32
0
4.93
fp32
1
A
yolo-master-n-upstream-e120-s1-p800f-fp32-fork-20260912101226
False
False
32
0
10.63
fp32
1
A0
yolo-master-n-baseline-e120-s1-p800f-fp32-fork-20260912205104
False
False
32
0
5.51
fp32
1
B
yolo-master-n-esmoe-upstream-w1-e120-s1-p800f-fp32-20260913022205
False
False
32
0
10.65
fp32
1
C
yolo-master-n-baseline-e120-s1-p800f-fp32-20260913130105
False
False
32
0
4.95
fp32
2
A
yolo-master-n-upstream-e120-s2-p800f-fp32-fork-20260912101356
False
False
32
0
10.59
fp32
2
A0
yolo-master-n-baseline-e120-s2-p800f-fp32-fork-20260914011826
False
False
32
0
5.52
fp32
2
B
yolo-master-n-esmoe-upstream-w1-e120-s2-p800f-fp32-20260912225003
False
False
32
0
10.63
fp32
2
C
yolo-master-n-baseline-e120-s2-p800f-fp32-20260913092752
False
False
32
0
4.90
mixed
0
A
yolo-master-n-upstream-e120-s0-p800c-fork-20260911060342
True
False
32
1
10.80
mixed
0
A0
yolo-master-n-baseline-e120-s0-p800c-fork-20260911165202
True
False
32
1
6.95
mixed
0
B
yolo-master-n-esmoe-upstream-w1-e120-s0-p800c-20260911234908
True
True
32
0
4.59
mixed
0
C
yolo-master-n-baseline-e120-s0-p800c-20260911165332
True
True
32
0
4.28
mixed
1
A
yolo-master-n-upstream-e120-s1-p800d-fork-20260911060354
True
False
32
1
10.79
mixed
1
A0
yolo-master-n-baseline-e120-s1-p800d-fork-20260911165201
True
False
32
1
6.97
mixed
1
B
yolo-master-n-esmoe-upstream-w1-e120-s1-p800d-20260911235018
True
True
32
0
4.74
mixed
1
C
yolo-master-n-baseline-e120-s1-p800d-20260911165331
True
True
32
0
4.28
mixed
2
A
yolo-master-n-upstream-e120-s2-p800e-fork-20260911100545
True
False
32
1
10.68
mixed
2
A0
yolo-master-n-baseline-e120-s2-p800e-fork-20260911204651
True
False
32
1
7.02
mixed
2
B
yolo-master-n-esmoe-upstream-w1-e120-s2-p800e-20260912034809
True
True
32
0
4.67
mixed
2
C
yolo-master-n-baseline-e120-s2-p800e-20260911204822
True
True
32
0
4.29
Area buckets
Fig. 3 Seven generations paired deltas by area bucket
Fig. 4 Same configuration A − A0 and B − C
COCO-style buckets (small < 32², medium 32²–96², large ≥ 96²) on ground-truth boxes in the original image, maxDets = 500. This is a definition adopted for this project; VisDrone's official protocol reports no size split.
A row named for a run alone was measured from that run's final EMA weights, as scripts/measure.py rebuilds them; the rest are measured from the run's best.pt.
Ground truth: 26586 small, 11105 medium, 1068 large boxes.
variant
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3/yolo-master]
0
0.2095
0.3640
0.2061
0.1260
0.3122
0.3902
0.3783
0.2826
0.5148
0.5711
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3/yolo-master]
1
0.2080
0.3638
0.2049
0.1262
0.3037
0.3909
0.3777
0.2864
0.5077
0.5903
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3/yolo-master]
2
0.2062
0.3617
0.2024
0.1249
0.3060
0.3751
0.3745
0.2804
0.5078
0.5585
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3]
0
0.2122
0.3708
0.2096
0.1270
0.3153
0.3975
0.3797
0.2848
0.5145
0.5751
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3]
1
0.2100
0.3682
0.2084
0.1248
0.3126
0.3778
0.3780
0.2820
0.5142
0.5395
yolo-master-n-baseline@e120f1i800[metaxc500/metax3.3]
2
0.2122
0.3697
0.2090
0.1293
0.3101
0.4186
0.3796
0.2861
0.5122
0.6126
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
0
0.2090
0.3630
0.2080
0.1230
0.3061
0.3945
0.3751
0.2832
0.5062
0.5640
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.2104
0.3657
0.2096
0.1223
0.3119
0.3881
0.3771
0.2798
0.5130
0.5624
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
2
0.2119
0.3696
0.2092
0.1253
0.3129
0.4172
0.3809
0.2827
0.5179
0.6482
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3]
0
0.2107
0.3654
0.2099
0.1241
0.3118
0.3792
0.3771
0.2799
0.5119
0.5608
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3]
1
0.2077
0.3594
0.2093
0.1225
0.3091
0.3952
0.3737
0.2770
0.5080
0.5704
yolo-master-n-baseline@e120f1i800fp32[metaxc500/metax3.3]
2
0.2072
0.3621
0.2052
0.1229
0.3115
0.3615
0.3765
0.2823
0.5153
0.5337
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
0
0.2150
0.3729
0.2110
0.1303
0.3183
0.4097
0.3861
0.2905
0.5244
0.6029
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
1
0.2196
0.3796
0.2171
0.1321
0.3260
0.3906
0.3843
0.2863
0.5256
0.6093
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3/yolo-master]
2
0.2150
0.3717
0.2139
0.1295
0.3163
0.3960
0.3848
0.2896
0.5220
0.5626
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
0
0.2220
0.3828
0.2214
0.1341
0.3292
0.4253
0.3904
0.2956
0.5264
0.6084
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
1
0.2215
0.3819
0.2223
0.1369
0.3250
0.4129
0.3927
0.2986
0.5268
0.6229
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800[metaxc500/metax3.3]
2
0.2190
0.3789
0.2176
0.1343
0.3221
0.4016
0.3907
0.2953
0.5269
0.6122
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
0
0.2179
0.3775
0.2162
0.1334
0.3216
0.4021
0.3869
0.2904
0.5240
0.5940
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
1
0.2179
0.3747
0.2157
0.1314
0.3221
0.4293
0.3876
0.2915
0.5255
0.5896
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3/yolo-master]
2
0.2180
0.3762
0.2156
0.1317
0.3190
0.4152
0.3903
0.2914
0.5269
0.5912
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
0
0.2143
0.3711
0.2150
0.1291
0.3169
0.4078
0.3890
0.2936
0.5267
0.5794
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
1
0.2122
0.3678
0.2122
0.1262
0.3145
0.3826
0.3846
0.2857
0.5253
0.5961
yolo-master-n-e4k2w1.0-gshard-norm-dense-t0.4-x4@e120f1i800fp32[metaxc500/metax3.3]
2
0.2187
0.3766
0.2202
0.1348
0.3199
0.4068
0.3877
0.2918
0.5259
0.6061
yolo11n-baseline@e120f1i800[4090/torch2.6]
0
0.2087
0.3643
0.2032
0.1244
0.3056
0.3760
0.3783
0.2825
0.5108
0.5443
yolo11n-baseline@e120f1i800[4090/torch2.6]
1
0.2112
0.3670
0.2067
0.1245
0.3092
0.3956
0.3829
0.2883
0.5138
0.6077
yolo11n-baseline@e120f1i800[4090/torch2.6]
2
0.2104
0.3644
0.2106
0.1270
0.3079
0.3870
0.3794
0.2819
0.5169
0.5778
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.2082
0.3655
0.2048
0.1251
0.3050
0.3768
0.3778
0.2835
0.5107
0.5698
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.2091
0.3614
0.2085
0.1240
0.3080
0.3929
0.3764
0.2816
0.5107
0.5701
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.2142
0.3726
0.2098
0.1265
0.3129
0.4108
0.3813
0.2823
0.5185
0.6001
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.2118
0.3675
0.2094
0.1237
0.3130
0.4041
0.3823
0.2866
0.5170
0.5534
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.2101
0.3690
0.2065
0.1275
0.3056
0.4146
0.3772
0.2838
0.5086
0.5925
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.2091
0.3641
0.2043
0.1245
0.3078
0.4127
0.3767
0.2807
0.5124
0.5936
yolo12n-baseline@e120f1i800[4090/torch2.6]
0
0.2131
0.3688
0.2119
0.1260
0.3174
0.3920
0.3807
0.2835
0.5167
0.5843
yolo12n-baseline@e120f1i800[4090/torch2.6]
1
0.2116
0.3676
0.2082
0.1252
0.3119
0.4090
0.3798
0.2831
0.5179
0.5954
yolo12n-baseline@e120f1i800[4090/torch2.6]
2
0.2109
0.3672
0.2098
0.1239
0.3137
0.3875
0.3826
0.2863
0.5184
0.5880
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.2147
0.3712
0.2148
0.1290
0.3177
0.4167
0.3844
0.2887
0.5222
0.6029
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.2109
0.3649
0.2090
0.1276
0.3081
0.3717
0.3796
0.2869
0.5135
0.5782
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.2127
0.3690
0.2079
0.1292
0.3104
0.3909
0.3827
0.2861
0.5199
0.5714
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.2107
0.3662
0.2102
0.1259
0.3110
0.4246
0.3780
0.2800
0.5153
0.6217
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.2098
0.3660
0.2052
0.1265
0.3086
0.4063
0.3811
0.2872
0.5157
0.5852
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.2113
0.3670
0.2100
0.1278
0.3107
0.3779
0.3786
0.2829
0.5151
0.5882
yolo26n-baseline@e120f1i800[4090/torch2.6]
0
0.1997
0.3535
0.1950
0.1242
0.2899
0.3610
0.3725
0.2865
0.4921
0.5253
yolo26n-baseline@e120f1i800[4090/torch2.6]
1
0.2034
0.3556
0.2015
0.1234
0.2907
0.3870
0.3783
0.2901
0.5008
0.5776
yolo26n-baseline@e120f1i800[4090/torch2.6]
2
0.2010
0.3528
0.1961
0.1236
0.2850
0.3673
0.3746
0.2869
0.4941
0.5814
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.2009
0.3520
0.1974
0.1197
0.2904
0.3635
0.3725
0.2813
0.4954
0.5310
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.2029
0.3551
0.1995
0.1252
0.2900
0.3788
0.3782
0.2901
0.4987
0.5562
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.2029
0.3548
0.2003
0.1248
0.2917
0.4097
0.3756
0.2857
0.4983
0.6207
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.1982
0.3489
0.1928
0.1194
0.2845
0.3769
0.3749
0.2824
0.4981
0.5609
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.1994
0.3516
0.1947
0.1178
0.2911
0.3976
0.3714
0.2780
0.4951
0.5549
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.1997
0.3512
0.1974
0.1206
0.2929
0.3662
0.3729
0.2842
0.4975
0.5139
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
0
0.2088
0.3615
0.2071
0.1269
0.3023
0.3578
0.3829
0.2912
0.5108
0.5446
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
1
0.2090
0.3658
0.2053
0.1264
0.3021
0.3814
0.3818
0.2913
0.5062
0.5221
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
2
0.2037
0.3556
0.2001
0.1224
0.2941
0.3914
0.3792
0.2890
0.5019
0.5814
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
3
0.2079
0.3629
0.2037
0.1259
0.3017
0.3822
0.3855
0.2941
0.5138
0.5367
yolov10n-baseline@e120f1i800[metaxc500/metax3.3]
4
0.2062
0.3625
0.2019
0.1266
0.2992
0.3702
0.3771
0.2849
0.5067
0.5674
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.2023
0.3542
0.2019
0.1219
0.2954
0.3986
0.3780
0.2859
0.5060
0.5997
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.2027
0.3534
0.1992
0.1232
0.2957
0.3855
0.3819
0.2892
0.5087
0.6062
yolov10n-e4k2w0.0025-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.1938
0.3401
0.1902
0.1148
0.2850
0.3695
0.3692
0.2759
0.4987
0.5543
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
0
0.2066
0.3613
0.2024
0.1252
0.3008
0.3997
0.3826
0.2903
0.5119
0.5560
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
1
0.2070
0.3607
0.2030
0.1243
0.3030
0.3827
0.3799
0.2873
0.5067
0.5647
yolov10n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
2
0.2087
0.3611
0.2071
0.1210
0.3057
0.3922
0.3817
0.2884
0.5084
0.5555
yolov10n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
0
0.2052
0.3612
0.2000
0.1236
0.3028
0.3926
0.3814
0.2874
0.5108
0.5687
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.2034
0.3564
0.2019
0.1241
0.2948
0.3794
0.3810
0.2916
0.5061
0.5632
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.2020
0.3535
0.1992
0.1205
0.2952
0.3714
0.3785
0.2875
0.5083
0.5429
yolov10n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.2020
0.3518
0.2007
0.1224
0.2941
0.3749
0.3774
0.2858
0.5067
0.5469
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.2076
0.3608
0.2068
0.1263
0.3009
0.3991
0.3819
0.2888
0.5088
0.5637
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.2029
0.3563
0.2013
0.1234
0.2955
0.3587
0.3794
0.2902
0.5079
0.5542
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.2091
0.3618
0.2088
0.1268
0.3003
0.3886
0.3834
0.2914
0.5122
0.6187
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
3
0.2044
0.3561
0.2036
0.1215
0.2973
0.3747
0.3819
0.2927
0.5053
0.5642
yolov10n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
4
0.2042
0.3578
0.2024
0.1244
0.2956
0.3770
0.3816
0.2903
0.5108
0.5474
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.2087
0.3617
0.2082
0.1250
0.3005
0.4045
0.3836
0.2921
0.5101
0.6218
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.2064
0.3616
0.2046
0.1222
0.3024
0.3886
0.3801
0.2898
0.5091
0.5809
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.2052
0.3590
0.2038
0.1238
0.2997
0.3738
0.3819
0.2917
0.5075
0.5496
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
3
0.2084
0.3633
0.2071
0.1250
0.3048
0.3776
0.3819
0.2882
0.5113
0.5610
yolov10n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
4
0.2061
0.3631
0.2023
0.1277
0.2956
0.4011
0.3839
0.2895
0.5125
0.5895
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
0
0.2064
0.3597
0.2043
0.1256
0.2979
0.3876
0.3800
0.2895
0.5081
0.5477
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
1
0.2021
0.3559
0.1970
0.1212
0.2950
0.3633
0.3796
0.2877
0.5083
0.5508
yolov10n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
2
0.2056
0.3581
0.2033
0.1237
0.2974
0.3955
0.3826
0.2917
0.5090
0.5620
yolov5n-baseline@e120f1i800[metaxc500/metax3.3]
0
0.2049
0.3616
0.1997
0.1196
0.3049
0.3842
0.3727
0.2784
0.5046
0.5339
yolov5n-baseline@e120f1i800[metaxc500/metax3.3]
1
0.2084
0.3631
0.2050
0.1258
0.3078
0.3887
0.3769
0.2822
0.5094
0.5848
yolov5n-baseline@e120f1i800[metaxc500/metax3.3]
2
0.2048
0.3596
0.2004
0.1235
0.3008
0.3920
0.3709
0.2767
0.5041
0.5525
yolov5n-baseline@e120f1i800[metaxc500/metax3.7]
0
0.2038
0.3577
0.2000
0.1198
0.3000
0.3922
0.3740
0.2775
0.5117
0.5665
yolov5n-baseline@e120f1i800[metaxc500/metax3.7]
1
0.2041
0.3590
0.1973
0.1198
0.3003
0.4190
0.3709
0.2765
0.4992
0.5406
yolov5n-baseline@e120f1i800[metaxc500/metax3.7]
2
0.2062
0.3616
0.2031
0.1218
0.3028
0.3883
0.3730
0.2790
0.5039
0.5453
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
0
0.2084
0.3639
0.2040
0.1238
0.3079
0.3810
0.3753
0.2775
0.5133
0.5544
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
1
0.2096
0.3658
0.2069
0.1223
0.3114
0.3971
0.3761
0.2757
0.5149
0.5741
yolov5n-e4k2w0.01-dense@e120f1i800[metaxc500/metax3.3]
2
0.2098
0.3657
0.2061
0.1229
0.3098
0.4111
0.3744
0.2754
0.5119
0.6059
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
0
0.2082
0.3660
0.2055
0.1253
0.3051
0.3886
0.3791
0.2808
0.5172
0.5809
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
1
0.2060
0.3612
0.2015
0.1197
0.3057
0.4065
0.3734
0.2778
0.5106
0.5964
yolov5n-e4k2w0.01-gshard@e120f1i800[metaxc500/metax3.3]
2
0.2081
0.3638
0.2033
0.1258
0.3036
0.4000
0.3769
0.2796
0.5148
0.5553
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
0
0.2076
0.3630
0.2042
0.1212
0.3086
0.4144
0.3767
0.2767
0.5184
0.5812
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
1
0.2080
0.3663
0.2033
0.1239
0.3077
0.3788
0.3746
0.2763
0.5156
0.5626
yolov5n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3]
2
0.2092
0.3671
0.2065
0.1240
0.3097
0.4128
0.3733
0.2769
0.5087
0.5931
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
0
0.2011
0.3532
0.1985
0.1211
0.2945
0.3824
0.3695
0.2741
0.5065
0.5549
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
1
0.2068
0.3613
0.2022
0.1202
0.3059
0.4174
0.3725
0.2743
0.5100
0.6029
yolov5n-e4k2w0.01-rewire-x4@e120f1i800[metaxc500/metax3.3]
2
0.2005
0.3552
0.1966
0.1163
0.2990
0.3789
0.3688
0.2730
0.5061
0.5509
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.2087
0.3658
0.2032
0.1263
0.3047
0.3979
0.3753
0.2795
0.5109
0.5758
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.2062
0.3613
0.2011
0.1191
0.3045
0.3973
0.3719
0.2731
0.5080
0.5569
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.2087
0.3633
0.2065
0.1247
0.3108
0.3876
0.3727
0.2784
0.5056
0.5454
yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.7]
0
0.2096
0.3636
0.2067
0.1242
0.3098
0.3887
0.3786
0.2798
0.5180
0.5634
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.2102
0.3654
0.2057
0.1276
0.3068
0.3966
0.3776
0.2823
0.5098
0.5867
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.2076
0.3626
0.2046
0.1226
0.3057
0.4008
0.3763
0.2797
0.5114
0.5899
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.2111
0.3713
0.2084
0.1281
0.3086
0.3819
0.3802
0.2853
0.5133
0.5555
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
0
0.2069
0.3634
0.2038
0.1246
0.3062
0.3970
0.3755
0.2784
0.5146
0.5857
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
1
0.2058
0.3612
0.2027
0.1182
0.3054
0.4338
0.3722
0.2735
0.5105
0.5920
yolov5n-e4k2w0.01@e120f1i800[metaxc500/metax3.7]
2
0.2077
0.3626
0.2048
0.1227
0.3084
0.4001
0.3764
0.2816
0.5105
0.5546
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
0
0.2049
0.3582
0.2017
0.1201
0.3035
0.3963
0.3704
0.2742
0.5052
0.5828
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
1
0.2072
0.3635
0.2051
0.1222
0.3069
0.3914
0.3755
0.2771
0.5166
0.5605
yolov5n-e4k2w0.0@e120f1i800[metaxc500/metax3.3]
2
0.2081
0.3634
0.2030
0.1209
0.3051
0.4102
0.3760
0.2816
0.5096
0.5740
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
0
0.2075
0.3625
0.2029
0.1242
0.3076
0.3939
0.3754
0.2806
0.5112
0.5486
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
1
0.2080
0.3656
0.2045
0.1238
0.3022
0.4016
0.3734
0.2776
0.5077
0.5603
yolov5n-e4k2w0.32-master@e120f1i800[metaxc500/metax3.3]
2
0.2052
0.3618
0.2013
0.1209
0.3065
0.3941
0.3721
0.2748
0.5080
0.5907
yolov8n-baseline@e120f1i800[4090/torch2.6]
0
0.2123
0.3689
0.2120
0.1238
0.3168
0.3820
0.3810
0.2847
0.5144
0.5715
yolov8n-baseline@e120f1i800[4090/torch2.6]
1
0.2152
0.3723
0.2122
0.1285
0.3158
0.3931
0.3853
0.2877
0.5222
0.5715
yolov8n-baseline@e120f1i800[4090/torch2.6]
2
0.2135
0.3722
0.2134
0.1261
0.3171
0.4049
0.3793
0.2836
0.5127
0.5660
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
0
0.2156
0.3772
0.2121
0.1278
0.3205
0.3981
0.3827
0.2834
0.5221
0.5675
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
1
0.2147
0.3729
0.2111
0.1288
0.3136
0.3962
0.3790
0.2815
0.5138
0.5763
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6]
2
0.2136
0.3725
0.2114
0.1223
0.3183
0.4045
0.3856
0.2857
0.5237
0.5745
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
0
0.2114
0.3706
0.2055
0.1237
0.3120
0.3789
0.3824
0.2855
0.5167
0.5708
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
1
0.2140
0.3711
0.2113
0.1280
0.3122
0.3837
0.3835
0.2899
0.5134
0.5526
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6]
2
0.2164
0.3778
0.2124
0.1299
0.3200
0.3861
0.3837
0.2884
0.5206
0.5468
yolov9t-baseline@e120f1i800[metaxc500/metax3.3]
0
0.2181
0.3768
0.2132
0.1264
0.3213
0.4363
0.3833
0.2833
0.5227
0.6136
yolov9t-baseline@e120f1i800[metaxc500/metax3.3]
1
0.2177
0.3764
0.2154
0.1284
0.3207
0.4517
0.3835
0.2859
0.5204
0.6422
yolov9t-baseline@e120f1i800[metaxc500/metax3.3]
2
0.2201
0.3806
0.2185
0.1290
0.3262
0.4140
0.3823
0.2815
0.5227
0.6369
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
0
0.2132
0.3719
0.2118
0.1289
0.3163
0.4334
0.3794
0.2796
0.5202
0.6262
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
1
0.2146
0.3741
0.2104
0.1272
0.3177
0.4214
0.3798
0.2806
0.5189
0.6523
yolov9t-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.3]
2
0.2224
0.3829
0.2227
0.1285
0.3320
0.4209
0.3846
0.2844
0.5266
0.5916
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
0
0.2187
0.3784
0.2171
0.1311
0.3225
0.4232
0.3823
0.2854
0.5192
0.6008
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
1
0.2213
0.3836
0.2166
0.1335
0.3261
0.4196
0.3847
0.2878
0.5215
0.6212
yolov9t-e4k2w0.01@e120f1i800[metaxc500/metax3.3]
2
0.2184
0.3769
0.2174
0.1269
0.3226
0.4327
0.3831
0.2826
0.5241
0.6102
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0054
+0.0089
+0.0049
+0.0043
+0.0061
+0.0195
+0.0078
+0.0080
+0.0096
+0.0318
1
+0.0116
+0.0158
+0.0122
+0.0059
+0.0224
-0.0004
+0.0066
-0.0001
+0.0179
+0.0190
2
+0.0088
+0.0099
+0.0115
+0.0046
+0.0103
+0.0210
+0.0103
+0.0092
+0.0142
+0.0041
mean
+0.0086
+0.0116
+0.0095
+0.0049
+0.0129
+0.0133
+0.0082
+0.0057
+0.0139
+0.0183
wins
3/3
3/3
3/3
3/3
3/3
2/3
3/3
2/3
3/3
3/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0097
+0.0120
+0.0118
+0.0070
+0.0139
+0.0279
+0.0107
+0.0108
+0.0119
+0.0333
1
+0.0115
+0.0137
+0.0139
+0.0121
+0.0124
+0.0352
+0.0147
+0.0166
+0.0126
+0.0834
2
+0.0068
+0.0092
+0.0086
+0.0050
+0.0120
-0.0169
+0.0111
+0.0092
+0.0146
-0.0004
mean
+0.0093
+0.0116
+0.0114
+0.0080
+0.0128
+0.0154
+0.0122
+0.0122
+0.0130
+0.0388
wins
3/3
3/3
3/3
3/3
3/3
2/3
3/3
3/3
3/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0089
+0.0145
+0.0081
+0.0104
+0.0155
+0.0076
+0.0118
+0.0072
+0.0178
+0.0300
1
+0.0074
+0.0090
+0.0061
+0.0091
+0.0102
+0.0412
+0.0105
+0.0117
+0.0126
+0.0272
2
+0.0060
+0.0066
+0.0063
+0.0064
+0.0061
-0.0020
+0.0094
+0.0087
+0.0090
-0.0569
mean
+0.0075
+0.0100
+0.0069
+0.0086
+0.0106
+0.0156
+0.0106
+0.0092
+0.0131
+0.0001
wins
3/3
3/3
3/3
3/3
3/3
2/3
3/3
3/3
3/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0036
+0.0057
+0.0051
+0.0051
+0.0051
+0.0286
+0.0119
+0.0136
+0.0148
+0.0186
1
+0.0045
+0.0084
+0.0029
+0.0036
+0.0053
-0.0126
+0.0109
+0.0087
+0.0173
+0.0257
2
+0.0115
+0.0145
+0.0150
+0.0119
+0.0084
+0.0453
+0.0112
+0.0095
+0.0107
+0.0723
mean
+0.0065
+0.0095
+0.0077
+0.0069
+0.0063
+0.0204
+0.0114
+0.0106
+0.0143
+0.0389
wins
3/3
3/3
3/3
3/3
3/3
2/3
3/3
3/3
3/3
3/3
yolo11n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6] minus baseline, per seed
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0005
+0.0011
+0.0017
+0.0007
-0.0005
+0.0008
-0.0005
+0.0010
-0.0001
+0.0255
1
-0.0021
-0.0056
+0.0018
-0.0006
-0.0012
-0.0028
-0.0065
-0.0067
-0.0031
-0.0376
2
+0.0038
+0.0082
-0.0008
-0.0005
+0.0050
+0.0238
+0.0019
+0.0004
+0.0016
+0.0223
mean
+0.0004
+0.0012
+0.0009
-0.0001
+0.0011
+0.0073
-0.0017
-0.0018
-0.0005
+0.0034
wins
1/3
2/3
2/3
1/3
1/3
2/3
1/3
2/3
1/3
2/3
yolo11n-e4k2w0.01@e120f1i800[4090/torch2.6] minus baseline, per seed
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0031
+0.0032
+0.0062
-0.0007
+0.0075
+0.0282
+0.0040
+0.0041
+0.0062
+0.0091
1
-0.0010
+0.0020
-0.0002
+0.0029
-0.0036
+0.0189
-0.0056
-0.0045
-0.0052
-0.0152
2
-0.0013
-0.0003
-0.0063
-0.0025
-0.0001
+0.0257
-0.0027
-0.0011
-0.0045
+0.0158
mean
+0.0003
+0.0016
-0.0001
-0.0001
+0.0013
+0.0243
-0.0014
-0.0005
-0.0011
+0.0032
wins
1/3
2/3
1/3
1/3
1/3
3/3
1/3
1/3
1/3
2/3
yolo12n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6] minus baseline, per seed
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0016
+0.0023
+0.0029
+0.0030
+0.0003
+0.0246
+0.0037
+0.0052
+0.0055
+0.0186
1
-0.0007
-0.0027
+0.0008
+0.0024
-0.0039
-0.0373
-0.0002
+0.0038
-0.0044
-0.0171
2
+0.0019
+0.0018
-0.0020
+0.0053
-0.0033
+0.0035
+0.0001
-0.0002
+0.0015
-0.0167
mean
+0.0009
+0.0005
+0.0006
+0.0035
-0.0023
-0.0031
+0.0012
+0.0029
+0.0009
-0.0051
wins
2/3
2/3
2/3
3/3
1/3
2/3
2/3
2/3
2/3
1/3
yolo12n-e4k2w0.01@e120f1i800[4090/torch2.6] minus baseline, per seed
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0024
-0.0026
-0.0017
-0.0001
-0.0064
+0.0326
-0.0027
-0.0036
-0.0014
+0.0374
1
-0.0019
-0.0016
-0.0030
+0.0013
-0.0033
-0.0026
+0.0013
+0.0041
-0.0022
-0.0101
2
+0.0005
-0.0002
+0.0002
+0.0039
-0.0030
-0.0096
-0.0040
-0.0034
-0.0033
+0.0001
mean
-0.0013
-0.0015
-0.0015
+0.0017
-0.0043
+0.0068
-0.0018
-0.0010
-0.0023
+0.0091
wins
1/3
0/3
1/3
2/3
0/3
1/3
1/3
1/3
0/3
2/3
yolo26n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6] minus baseline, per seed
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0012
-0.0016
+0.0024
-0.0045
+0.0005
+0.0025
+0.0001
-0.0052
+0.0033
+0.0057
1
-0.0005
-0.0006
-0.0021
+0.0018
-0.0007
-0.0082
-0.0002
+0.0001
-0.0022
-0.0214
2
+0.0019
+0.0020
+0.0042
+0.0012
+0.0067
+0.0424
+0.0010
-0.0012
+0.0042
+0.0393
mean
+0.0009
-0.0000
+0.0015
-0.0005
+0.0022
+0.0122
+0.0003
-0.0021
+0.0018
+0.0079
wins
2/3
1/3
2/3
2/3
2/3
2/3
2/3
1/3
2/3
2/3
yolo26n-e4k2w0.01@e120f1i800[4090/torch2.6] minus baseline, per seed
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0015
-0.0046
-0.0022
-0.0048
-0.0054
+0.0159
+0.0024
-0.0041
+0.0060
+0.0356
1
-0.0040
-0.0040
-0.0069
-0.0056
+0.0004
+0.0106
-0.0070
-0.0120
-0.0057
-0.0227
2
-0.0013
-0.0016
+0.0013
-0.0030
+0.0079
-0.0011
-0.0017
-0.0027
+0.0034
-0.0676
mean
-0.0022
-0.0034
-0.0026
-0.0045
+0.0010
+0.0085
-0.0021
-0.0063
+0.0012
-0.0182
wins
0/3
0/3
1/3
0/3
2/3
2/3
1/3
0/3
2/3
1/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0066
-0.0072
-0.0051
-0.0050
-0.0068
+0.0409
-0.0049
-0.0053
-0.0049
+0.0552
1
-0.0063
-0.0123
-0.0061
-0.0032
-0.0064
+0.0041
+0.0000
-0.0021
+0.0025
+0.0841
2
-0.0099
-0.0155
-0.0099
-0.0076
-0.0090
-0.0219
-0.0099
-0.0131
-0.0031
-0.0270
mean
-0.0076
-0.0117
-0.0071
-0.0053
-0.0074
+0.0077
-0.0049
-0.0068
-0.0018
+0.0374
wins
0/3
0/3
0/3
0/3
0/3
2/3
1/3
0/3
1/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0023
-0.0002
-0.0047
-0.0017
-0.0015
+0.0419
-0.0003
-0.0009
+0.0010
+0.0115
1
-0.0020
-0.0050
-0.0022
-0.0022
+0.0009
+0.0013
-0.0020
-0.0040
+0.0004
+0.0426
2
+0.0050
+0.0056
+0.0070
-0.0014
+0.0116
+0.0008
+0.0026
-0.0006
+0.0065
-0.0259
mean
+0.0002
+0.0001
+0.0000
-0.0018
+0.0037
+0.0147
+0.0001
-0.0018
+0.0027
+0.0094
wins
1/3
1/3
1/3
0/3
2/3
3/3
1/3
0/3
3/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0036
-0.0003
-0.0070
-0.0032
+0.0005
+0.0349
-0.0015
-0.0038
-0.0001
+0.0241
mean
-0.0036
-0.0003
-0.0070
-0.0032
+0.0005
+0.0349
-0.0015
-0.0038
-0.0001
+0.0241
wins
0/1
0/1
0/1
0/1
1/1
1/1
0/1
0/1
0/1
1/1
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0054
-0.0051
-0.0052
-0.0027
-0.0075
+0.0217
-0.0019
+0.0004
-0.0048
+0.0186
1
-0.0071
-0.0123
-0.0061
-0.0059
-0.0069
-0.0100
-0.0033
-0.0038
+0.0021
+0.0208
2
-0.0017
-0.0038
+0.0007
-0.0000
-0.0000
-0.0165
-0.0017
-0.0032
+0.0048
-0.0345
mean
-0.0047
-0.0070
-0.0035
-0.0029
-0.0048
-0.0016
-0.0023
-0.0022
+0.0007
+0.0016
wins
0/3
0/3
1/3
0/3
0/3
1/3
0/3
1/3
2/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0013
-0.0007
-0.0003
-0.0006
-0.0014
+0.0413
-0.0010
-0.0024
-0.0020
+0.0192
1
-0.0061
-0.0095
-0.0040
-0.0031
-0.0066
-0.0227
-0.0024
-0.0011
+0.0016
+0.0321
2
+0.0054
+0.0063
+0.0087
+0.0044
+0.0062
-0.0028
+0.0042
+0.0024
+0.0104
+0.0374
3
-0.0035
-0.0067
-0.0001
-0.0044
-0.0044
-0.0076
-0.0036
-0.0014
-0.0085
+0.0275
4
-0.0019
-0.0047
+0.0005
-0.0022
-0.0036
+0.0069
+0.0044
+0.0054
+0.0041
-0.0200
mean
-0.0015
-0.0031
+0.0010
-0.0012
-0.0019
+0.0030
+0.0003
+0.0006
+0.0011
+0.0192
wins
1/5
1/5
2/5
1/5
1/5
2/5
2/5
2/5
3/5
4/5
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0001
+0.0002
+0.0011
-0.0019
-0.0018
+0.0468
+0.0007
+0.0008
-0.0007
+0.0773
1
-0.0026
-0.0041
-0.0007
-0.0042
+0.0004
+0.0072
-0.0017
-0.0015
+0.0029
+0.0588
2
+0.0015
+0.0034
+0.0038
+0.0013
+0.0056
-0.0175
+0.0027
+0.0027
+0.0057
-0.0318
3
+0.0005
+0.0004
+0.0034
-0.0009
+0.0031
-0.0046
-0.0036
-0.0059
-0.0025
+0.0243
4
-0.0000
+0.0005
+0.0003
+0.0011
-0.0036
+0.0309
+0.0068
+0.0046
+0.0058
+0.0221
mean
-0.0002
+0.0001
+0.0016
-0.0009
+0.0007
+0.0126
+0.0010
+0.0001
+0.0022
+0.0301
wins
2/5
4/5
4/5
2/5
3/5
3/5
3/5
3/5
3/5
4/5
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0024
-0.0018
-0.0028
-0.0013
-0.0043
+0.0298
-0.0029
-0.0017
-0.0027
+0.0031
1
-0.0070
-0.0098
-0.0083
-0.0052
-0.0071
-0.0181
-0.0022
-0.0036
+0.0021
+0.0287
2
+0.0019
+0.0025
+0.0032
+0.0013
+0.0034
+0.0041
+0.0034
+0.0027
+0.0072
-0.0194
mean
-0.0025
-0.0030
-0.0026
-0.0017
-0.0027
+0.0053
-0.0006
-0.0009
+0.0022
+0.0042
wins
1/3
1/3
1/3
1/3
1/3
2/3
1/3
1/3
2/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0034
+0.0023
+0.0043
+0.0042
+0.0030
-0.0032
+0.0026
-0.0010
+0.0087
+0.0204
1
+0.0012
+0.0026
+0.0019
-0.0035
+0.0036
+0.0084
-0.0008
-0.0065
+0.0055
-0.0107
2
+0.0051
+0.0061
+0.0057
-0.0007
+0.0090
+0.0190
+0.0036
-0.0013
+0.0078
+0.0534
mean
+0.0032
+0.0037
+0.0039
+0.0000
+0.0052
+0.0081
+0.0018
-0.0029
+0.0074
+0.0210
wins
3/3
3/3
3/3
1/3
3/3
2/3
2/3
0/3
3/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0033
+0.0043
+0.0058
+0.0057
+0.0001
+0.0044
+0.0064
+0.0024
+0.0125
+0.0470
1
-0.0024
-0.0019
-0.0035
-0.0061
-0.0021
+0.0178
-0.0035
-0.0044
+0.0013
+0.0116
2
+0.0034
+0.0042
+0.0029
+0.0023
+0.0028
+0.0080
+0.0060
+0.0029
+0.0108
+0.0028
mean
+0.0014
+0.0022
+0.0017
+0.0006
+0.0003
+0.0101
+0.0029
+0.0003
+0.0082
+0.0204
wins
2/3
2/3
2/3
2/3
2/3
3/3
2/3
2/3
3/3
3/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0026
+0.0013
+0.0045
+0.0017
+0.0037
+0.0303
+0.0039
-0.0017
+0.0138
+0.0472
1
-0.0005
+0.0032
-0.0017
-0.0020
-0.0000
-0.0098
-0.0023
-0.0059
+0.0063
-0.0222
2
+0.0045
+0.0075
+0.0061
+0.0005
+0.0089
+0.0207
+0.0024
+0.0002
+0.0046
+0.0407
mean
+0.0022
+0.0040
+0.0030
+0.0001
+0.0042
+0.0137
+0.0014
-0.0025
+0.0082
+0.0219
wins
2/3
3/3
2/3
2/3
2/3
2/3
2/3
1/3
3/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0038
-0.0085
-0.0012
+0.0015
-0.0105
-0.0017
-0.0033
-0.0043
+0.0019
+0.0210
1
-0.0016
-0.0018
-0.0028
-0.0057
-0.0019
+0.0288
-0.0044
-0.0078
+0.0006
+0.0181
2
-0.0043
-0.0044
-0.0037
-0.0073
-0.0018
-0.0132
-0.0021
-0.0037
+0.0020
-0.0015
mean
-0.0033
-0.0049
-0.0026
-0.0038
-0.0047
+0.0046
-0.0033
-0.0053
+0.0015
+0.0125
wins
0/3
0/3
0/3
1/3
0/3
1/3
0/3
0/3
3/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0038
+0.0042
+0.0035
+0.0067
-0.0002
+0.0137
+0.0026
+0.0011
+0.0062
+0.0418
1
-0.0023
-0.0018
-0.0038
-0.0067
-0.0033
+0.0086
-0.0050
-0.0091
-0.0014
-0.0279
2
+0.0039
+0.0038
+0.0062
+0.0012
+0.0099
-0.0044
+0.0019
+0.0017
+0.0016
-0.0071
mean
+0.0018
+0.0021
+0.0020
+0.0004
+0.0021
+0.0060
-0.0002
-0.0021
+0.0021
+0.0023
wins
2/3
2/3
2/3
2/3
1/3
2/3
2/3
2/3
2/3
1/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0058
+0.0059
+0.0067
+0.0045
+0.0098
-0.0035
+0.0046
+0.0023
+0.0063
-0.0031
mean
+0.0058
+0.0059
+0.0067
+0.0045
+0.0098
-0.0035
+0.0046
+0.0023
+0.0063
-0.0031
wins
1/1
1/1
1/1
1/1
1/1
0/1
1/1
1/1
1/1
0/1
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0053
+0.0038
+0.0060
+0.0080
+0.0018
+0.0125
+0.0049
+0.0039
+0.0052
+0.0528
1
-0.0009
-0.0005
-0.0004
-0.0033
-0.0021
+0.0122
-0.0006
-0.0024
+0.0020
+0.0051
2
+0.0064
+0.0117
+0.0080
+0.0045
+0.0078
-0.0101
+0.0093
+0.0086
+0.0092
+0.0030
mean
+0.0036
+0.0050
+0.0045
+0.0031
+0.0025
+0.0048
+0.0045
+0.0034
+0.0055
+0.0203
wins
2/3
2/3
2/3
2/3
2/3
2/3
2/3
2/3
3/3
3/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0031
+0.0057
+0.0038
+0.0048
+0.0062
+0.0048
+0.0016
+0.0009
+0.0029
+0.0192
1
+0.0016
+0.0022
+0.0055
-0.0016
+0.0050
+0.0148
+0.0013
-0.0030
+0.0114
+0.0515
2
+0.0015
+0.0010
+0.0017
+0.0009
+0.0056
+0.0118
+0.0034
+0.0025
+0.0066
+0.0094
mean
+0.0021
+0.0030
+0.0037
+0.0014
+0.0056
+0.0105
+0.0021
+0.0002
+0.0069
+0.0267
wins
3/3
3/3
3/3
2/3
3/3
3/3
3/3
2/3
3/3
3/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0000
-0.0035
+0.0019
+0.0005
-0.0015
+0.0121
-0.0023
-0.0042
+0.0006
+0.0488
1
-0.0012
+0.0004
+0.0001
-0.0036
-0.0009
+0.0028
-0.0014
-0.0051
+0.0072
-0.0243
2
+0.0033
+0.0038
+0.0027
-0.0027
+0.0043
+0.0182
+0.0052
+0.0049
+0.0056
+0.0216
mean
+0.0007
+0.0003
+0.0016
-0.0019
+0.0006
+0.0110
+0.0005
-0.0015
+0.0045
+0.0154
wins
1/3
2/3
3/3
1/3
1/3
3/3
1/3
1/3
3/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0026
+0.0009
+0.0032
+0.0046
+0.0027
+0.0097
+0.0026
+0.0022
+0.0066
+0.0147
1
-0.0005
+0.0025
-0.0005
-0.0020
-0.0055
+0.0129
-0.0035
-0.0046
-0.0016
-0.0245
2
+0.0004
+0.0022
+0.0009
-0.0027
+0.0057
+0.0020
+0.0012
-0.0019
+0.0039
+0.0382
mean
+0.0009
+0.0019
+0.0012
-0.0000
+0.0009
+0.0082
+0.0001
-0.0014
+0.0030
+0.0095
wins
2/3
3/3
2/3
1/3
2/3
3/3
2/3
1/3
2/3
2/3
yolov8n-e4k2w0.01-rewire@e120f1i800[4090/torch2.6] minus baseline, per seed
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0033
+0.0083
+0.0001
+0.0040
+0.0037
+0.0162
+0.0017
-0.0013
+0.0077
-0.0040
1
-0.0005
+0.0007
-0.0011
+0.0003
-0.0022
+0.0031
-0.0063
-0.0061
-0.0084
+0.0048
2
+0.0002
+0.0002
-0.0019
-0.0037
+0.0012
-0.0003
+0.0063
+0.0020
+0.0109
+0.0085
mean
+0.0010
+0.0031
-0.0010
+0.0002
+0.0009
+0.0063
+0.0006
-0.0018
+0.0034
+0.0031
wins
2/3
3/3
1/3
2/3
2/3
2/3
2/3
1/3
2/3
2/3
yolov8n-e4k2w0.01@e120f1i800[4090/torch2.6] minus baseline, per seed
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0009
+0.0017
-0.0065
-0.0001
-0.0047
-0.0030
+0.0014
+0.0008
+0.0022
-0.0007
1
-0.0012
-0.0012
-0.0010
-0.0005
-0.0036
-0.0094
-0.0018
+0.0022
-0.0088
-0.0189
2
+0.0029
+0.0056
-0.0010
+0.0038
+0.0029
-0.0187
+0.0045
+0.0048
+0.0078
-0.0192
mean
+0.0003
+0.0020
-0.0028
+0.0011
-0.0018
-0.0104
+0.0014
+0.0026
+0.0004
-0.0129
wins
1/3
2/3
0/3
1/3
1/3
0/3
2/3
3/3
2/3
0/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
-0.0049
-0.0049
-0.0014
+0.0025
-0.0050
-0.0029
-0.0039
-0.0037
-0.0026
+0.0126
1
-0.0030
-0.0023
-0.0050
-0.0012
-0.0030
-0.0303
-0.0037
-0.0052
-0.0015
+0.0101
2
+0.0023
+0.0022
+0.0041
-0.0005
+0.0058
+0.0069
+0.0023
+0.0029
+0.0040
-0.0453
mean
-0.0019
-0.0016
-0.0008
+0.0003
-0.0008
-0.0088
-0.0018
-0.0020
-0.0000
-0.0075
wins
1/3
1/3
1/3
1/3
1/3
1/3
1/3
1/3
1/3
2/3
seed
AP
AP50
AP75
APs
APm
APl
AR
ARs
ARm
ARl
0
+0.0006
+0.0015
+0.0039
+0.0047
+0.0012
-0.0131
-0.0010
+0.0021
-0.0035
-0.0129
1
+0.0036
+0.0073
+0.0011
+0.0051
+0.0054
-0.0321
+0.0012
+0.0019
+0.0012
-0.0210
2
-0.0017
-0.0038
-0.0011
-0.0021
-0.0037
+0.0187
+0.0009
+0.0010
+0.0014
-0.0266
mean
+0.0008
+0.0017
+0.0013
+0.0025
+0.0010
-0.0088
+0.0004
+0.0017
-0.0003
-0.0202
wins
2/3
2/3
2/3
2/3
2/3
1/3
2/3
3/3
2/3
0/3
Router behaviour
Fig. 5 Concentration against paired delta large dots for four-block layouts
Fig. 6 Dead experts by balance term counted per checkpoint
Per checkpoint: the share of images on which each expert is the top-1 choice, the share on which it is in the top-2, the mean routing probability, and the correlation of that probability with the mean object size and the object count of the image.
Does concentration cost accuracy?
Over the 81 runs that have both a paired delta and a routing analysis, the leading expert's top-1 share runs 0.47 to 1.00 and the paired mAP50 delta -0.0160 to +0.0109. Their correlation is r = +0.044 : on this evidence a concentrated dispatch does not cost accuracy.
yolo-master-n-esmoe-upstream-w1-e120-s0-p800c
548 images through block 0 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: [0, 2], mean entropy 1.3658 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.000
0.222
-0.22
+0.18
1
5
0.002
1.000
0.280
+0.20
-0.35
2
7
0.000
0.000
0.184
+0.28
-0.24
3
9
0.998
1.000
0.314
-0.20
+0.32
548 images through block 1 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [2, 3], mean entropy 1.382 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.622
1.000
0.270
-0.06
+0.34
1
5
0.378
1.000
0.269
+0.23
-0.17
2
7
0.000
0.000
0.214
+0.12
-0.11
3
9
0.000
0.000
0.247
-0.28
-0.06
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1, 3], mean entropy 1.3814 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.854
1.000
0.277
-0.30
+0.35
1
5
0.000
0.000
0.218
+0.20
-0.16
2
7
0.146
1.000
0.270
+0.18
-0.28
3
9
0.000
0.000
0.235
+0.25
-0.31
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3703 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.084
0.285
0.229
+0.25
-0.22
1
5
0.055
0.540
0.238
-0.27
+0.33
2
7
0.790
0.872
0.301
+0.11
-0.02
3
9
0.071
0.303
0.231
-0.33
+0.07
yolo-master-n-esmoe-upstream-w1-e120-s0-p800f-fp32
548 images through block 0 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: [0, 2], mean entropy 1.3678 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.000
0.222
-0.32
+0.17
1
5
0.011
1.000
0.279
+0.21
-0.16
2
7
0.000
0.000
0.189
+0.29
-0.23
3
9
0.989
1.000
0.310
-0.15
+0.19
548 images through block 1 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [2, 3], mean entropy 1.377 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.007
1.000
0.278
-0.17
+0.18
1
5
0.993
1.000
0.290
-0.13
+0.20
2
7
0.000
0.000
0.212
+0.33
-0.45
3
9
0.000
0.000
0.221
-0.15
+0.21
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1, 3], mean entropy 1.3821 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
1.000
0.263
-0.21
+0.08
1
5
0.000
0.000
0.221
+0.28
-0.02
2
7
1.000
1.000
0.279
-0.25
-0.10
3
9
0.000
0.000
0.236
+0.18
+0.06
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3633 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.015
0.257
0.219
+0.53
-0.49
1
5
0.009
0.257
0.218
-0.09
+0.25
2
7
0.918
0.951
0.325
-0.12
+0.06
3
9
0.058
0.535
0.238
-0.17
+0.10
yolo-master-n-esmoe-upstream-w1-e120-s1-p800d
548 images through block 0 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1, 3], mean entropy 1.372 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.007
1.000
0.274
+0.28
-0.29
1
5
0.000
0.000
0.223
-0.30
+0.27
2
7
0.993
1.000
0.305
-0.17
+0.26
3
9
0.000
0.000
0.198
+0.01
-0.10
548 images through block 1 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1, 2], mean entropy 1.3681 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.113
1.000
0.293
+0.09
+0.05
1
5
0.000
0.000
0.195
+0.05
-0.18
2
7
0.000
0.000
0.211
-0.27
+0.13
3
9
0.887
1.000
0.301
+0.01
-0.03
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1], mean entropy 1.3616 of 1.3863, distinct top-2 pairs seen: 3 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.387
0.504
0.275
-0.26
+0.51
1
5
0.000
0.000
0.179
+0.23
-0.45
2
7
0.124
0.533
0.262
+0.34
-0.55
3
9
0.489
0.964
0.284
+0.11
-0.39
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3684 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.064
0.153
0.219
+0.12
-0.17
1
5
0.015
0.268
0.232
-0.02
+0.17
2
7
0.166
0.681
0.251
-0.23
+0.24
3
9
0.755
0.898
0.298
+0.03
-0.06
yolo-master-n-esmoe-upstream-w1-e120-s1-p800f-fp32
548 images through block 0 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1, 3], mean entropy 1.3674 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.009
1.000
0.275
+0.21
-0.28
1
5
0.000
0.000
0.201
-0.31
+0.15
2
7
0.991
1.000
0.317
-0.17
+0.29
3
9
0.000
0.000
0.207
+0.20
-0.12
548 images through block 1 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [2, 3], mean entropy 1.3723 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
1.000
0.281
+0.12
+0.20
1
5
1.000
1.000
0.301
+0.02
+0.01
2
7
0.000
0.000
0.213
-0.25
+0.19
3
9
0.000
0.000
0.205
+0.09
-0.45
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1], mean entropy 1.3679 of 1.3863, distinct top-2 pairs seen: 3 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.049
0.150
0.227
-0.18
+0.15
1
5
0.000
0.000
0.196
+0.14
-0.47
2
7
0.613
0.883
0.294
+0.32
-0.29
3
9
0.338
0.967
0.283
-0.38
+0.46
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3608 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.073
0.489
0.236
+0.45
-0.38
1
5
0.009
0.268
0.216
+0.02
-0.04
2
7
0.874
0.936
0.325
-0.28
+0.31
3
9
0.044
0.307
0.223
-0.04
-0.03
yolo-master-n-esmoe-upstream-w1-e120-s2-p800e
548 images through block 0 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1, 3], mean entropy 1.373 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.036
1.000
0.273
+0.12
-0.12
1
5
0.000
0.000
0.212
+0.24
-0.28
2
7
0.964
1.000
0.305
-0.24
+0.30
3
9
0.000
0.000
0.210
+0.17
-0.27
548 images through block 1 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1, 2], mean entropy 1.373 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.139
1.000
0.287
-0.27
+0.23
1
5
0.000
0.000
0.220
+0.23
-0.24
2
7
0.000
0.000
0.201
+0.28
-0.28
3
9
0.861
1.000
0.293
-0.08
+0.12
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [0, 2], mean entropy 1.3797 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.000
0.227
+0.00
-0.36
1
5
0.779
1.000
0.281
-0.22
+0.52
2
7
0.000
0.000
0.216
+0.19
-0.47
3
9
0.221
1.000
0.275
+0.23
-0.06
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.373 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.237
0.228
-0.08
+0.25
1
5
0.688
0.792
0.288
+0.31
-0.16
2
7
0.159
0.540
0.244
-0.32
+0.28
3
9
0.151
0.431
0.239
-0.19
-0.07
yolo-master-n-esmoe-upstream-w1-e120-s2-p800f-fp32
548 images through block 0 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1, 3], mean entropy 1.367 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.004
1.000
0.277
+0.25
-0.18
1
5
0.000
0.000
0.198
-0.24
+0.07
2
7
0.996
1.000
0.316
-0.15
+0.26
3
9
0.000
0.000
0.208
+0.05
-0.20
548 images through block 1 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [1, 2], mean entropy 1.3696 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
1.000
0.281
-0.28
+0.44
1
5
0.000
0.000
0.216
+0.23
-0.49
2
7
0.000
0.000
0.196
+0.28
-0.31
3
9
1.000
1.000
0.307
-0.15
+0.22
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: [0, 2], mean entropy 1.3793 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.000
0.231
-0.14
+0.25
1
5
0.872
1.000
0.282
-0.15
+0.42
2
7
0.000
0.000
0.212
+0.08
-0.39
3
9
0.128
1.000
0.275
+0.24
-0.38
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3671 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.007
0.166
0.215
-0.24
+0.14
1
5
0.071
0.383
0.236
-0.24
+0.14
2
7
0.746
0.869
0.304
+0.12
-0.07
3
9
0.175
0.582
0.245
+0.13
-0.08
yolo11n-esmoe-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3335 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.168
0.233
+0.23
-0.46
1
5
0.653
0.911
0.309
+0.26
-0.31
2
7
0.257
0.465
0.210
-0.31
+0.51
3
9
0.089
0.456
0.247
+0.17
-0.40
yolo11n-esmoe-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3486 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.027
0.135
0.230
-0.02
-0.04
1
5
0.190
0.527
0.229
-0.30
+0.47
2
7
0.133
0.500
0.238
+0.31
-0.18
3
9
0.650
0.838
0.303
+0.05
-0.29
yolo11n-esmoe-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2719 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.009
0.620
0.264
+0.26
-0.46
1
5
0.084
0.354
0.246
-0.06
+0.11
2
7
0.786
0.849
0.352
+0.18
-0.22
3
9
0.120
0.177
0.138
-0.21
+0.29
yolo11n-esmoe-rewire-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3233 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.210
0.226
+0.27
-0.62
1
5
0.788
0.943
0.338
+0.18
-0.18
2
7
0.153
0.392
0.194
-0.29
+0.48
3
9
0.058
0.454
0.241
+0.20
-0.41
yolo11n-esmoe-rewire-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3112 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.179
0.207
+0.14
-0.37
1
5
0.031
0.316
0.211
+0.01
-0.28
2
7
0.319
0.628
0.241
-0.27
+0.47
3
9
0.650
0.878
0.341
+0.24
-0.28
yolo11n-esmoe-rewire-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2885 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.009
0.509
0.251
+0.47
-0.45
1
5
0.797
0.901
0.361
+0.20
-0.21
2
7
0.119
0.363
0.227
-0.12
+0.29
3
9
0.075
0.226
0.162
-0.32
+0.21
yolo12n-esmoe-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2869 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.703
0.914
0.348
+0.12
-0.08
1
5
0.100
0.297
0.221
-0.23
+0.29
2
7
0.089
0.133
0.149
-0.24
+0.24
3
9
0.108
0.655
0.282
+0.39
-0.48
yolo12n-esmoe-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.334 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.013
0.108
0.215
+0.07
-0.11
1
5
0.564
0.814
0.298
-0.49
+0.60
2
7
0.042
0.161
0.184
+0.38
-0.50
3
9
0.381
0.918
0.302
+0.26
-0.29
yolo12n-esmoe-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2781 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.011
0.427
0.224
+0.07
-0.06
1
5
0.907
0.953
0.405
-0.15
+0.15
2
7
0.038
0.529
0.227
+0.59
-0.42
3
9
0.044
0.091
0.143
-0.15
+0.07
yolo12n-esmoe-rewire-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3312 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.571
0.236
+0.08
+0.04
1
5
0.064
0.177
0.210
+0.33
-0.29
2
7
0.119
0.359
0.210
-0.28
+0.33
3
9
0.816
0.892
0.344
-0.05
-0.04
yolo12n-esmoe-rewire-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3062 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.009
0.329
0.226
+0.24
-0.14
1
5
0.124
0.256
0.178
-0.30
+0.32
2
7
0.863
0.920
0.366
+0.18
-0.17
3
9
0.004
0.496
0.230
+0.21
-0.45
yolo12n-esmoe-rewire-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2758 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.790
0.920
0.364
+0.14
-0.10
1
5
0.055
0.204
0.206
-0.17
+0.30
2
7
0.099
0.126
0.145
-0.20
+0.15
3
9
0.057
0.750
0.285
+0.33
-0.43
yolo26n-esmoe-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3339 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.009
0.336
0.230
+0.17
-0.08
1
5
0.031
0.213
0.206
+0.57
-0.39
2
7
0.237
0.553
0.228
-0.29
+0.37
3
9
0.723
0.898
0.336
-0.01
-0.16
yolo26n-esmoe-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3069 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.084
0.655
0.278
+0.32
-0.42
1
5
0.374
0.520
0.272
-0.18
+0.34
2
7
0.051
0.104
0.145
-0.26
+0.27
3
9
0.491
0.721
0.305
+0.24
-0.31
yolo26n-esmoe-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3395 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.188
0.222
+0.11
-0.06
1
5
0.166
0.613
0.234
-0.34
+0.44
2
7
0.015
0.285
0.206
+0.46
-0.50
3
9
0.819
0.914
0.339
+0.04
-0.13
yolo26n-esmoe-rewire-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3129 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.035
0.491
0.239
+0.15
-0.22
1
5
0.064
0.126
0.171
-0.20
+0.15
2
7
0.027
0.443
0.225
+0.51
-0.56
3
9
0.874
0.940
0.366
-0.10
+0.21
yolo26n-esmoe-rewire-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3065 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.380
0.232
+0.12
-0.19
1
5
0.086
0.184
0.170
-0.17
+0.20
2
7
0.896
0.929
0.364
+0.03
+0.02
3
9
0.018
0.507
0.234
+0.28
-0.44
yolo26n-esmoe-rewire-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3394 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.193
0.224
+0.13
-0.27
1
5
0.217
0.560
0.236
-0.33
+0.46
2
7
0.111
0.376
0.216
+0.49
-0.47
3
9
0.671
0.870
0.324
-0.08
-0.02
yolov10n-esmoe-dense-e120-s0-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3585 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.148
0.229
-0.05
+0.10
1
5
0.467
0.737
0.286
+0.14
-0.12
2
7
0.239
0.473
0.234
+0.14
-0.34
3
9
0.292
0.642
0.251
-0.30
+0.46
yolov10n-esmoe-dense-e120-s1-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3379 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.007
0.409
0.240
-0.01
+0.08
1
5
0.166
0.352
0.214
-0.30
+0.44
2
7
0.188
0.438
0.230
+0.32
-0.40
3
9
0.639
0.801
0.315
+0.04
-0.13
yolov10n-esmoe-dense-e120-s2-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3207 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.007
0.181
0.211
+0.17
-0.18
1
5
0.015
0.283
0.205
+0.25
-0.34
2
7
0.234
0.628
0.232
-0.27
+0.49
3
9
0.745
0.909
0.352
+0.12
-0.28
yolov10n-esmoe-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3466 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.265
0.226
-0.03
-0.09
1
5
0.159
0.418
0.240
-0.17
+0.06
2
7
0.316
0.584
0.243
-0.26
+0.36
3
9
0.525
0.734
0.291
+0.35
-0.34
yolov10n-esmoe-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3308 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.018
0.442
0.257
+0.46
-0.44
1
5
0.248
0.544
0.256
-0.08
+0.25
2
7
0.117
0.256
0.184
-0.28
+0.36
3
9
0.617
0.759
0.303
+0.15
-0.33
yolov10n-esmoe-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.293 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.350
0.210
-0.03
-0.07
1
5
0.060
0.292
0.187
-0.15
+0.23
2
7
0.892
0.942
0.399
+0.10
-0.02
3
9
0.047
0.416
0.205
+0.02
-0.20
yolov10n-esmoe-e120-s3-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3373 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.144
0.225
+0.09
-0.23
1
5
0.097
0.367
0.205
-0.29
+0.39
2
7
0.832
0.916
0.336
+0.12
-0.07
3
9
0.069
0.573
0.234
+0.17
-0.32
yolov10n-esmoe-e120-s4-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2983 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.235
0.225
+0.30
-0.46
1
5
0.847
0.925
0.369
+0.04
+0.10
2
7
0.110
0.270
0.169
-0.26
+0.27
3
9
0.044
0.569
0.237
+0.28
-0.47
yolov10n-esmoe-norm-e120-s0-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3316 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.024
0.339
0.246
+0.16
-0.09
1
5
0.829
0.914
0.326
+0.09
+0.07
2
7
0.080
0.571
0.250
+0.29
-0.40
3
9
0.068
0.175
0.178
-0.30
+0.19
yolov10n-esmoe-rewire-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3054 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.429
0.218
-0.09
+0.12
1
5
0.088
0.394
0.209
-0.23
+0.37
2
7
0.066
0.250
0.190
+0.10
-0.24
3
9
0.845
0.927
0.383
+0.12
-0.13
yolov10n-esmoe-rewire-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3155 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.020
0.571
0.230
-0.12
+0.09
1
5
0.079
0.217
0.191
-0.18
+0.27
2
7
0.816
0.894
0.369
+0.17
-0.06
3
9
0.086
0.318
0.210
-0.00
-0.28
yolov10n-esmoe-rewire-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3121 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.025
0.153
0.215
-0.01
-0.12
1
5
0.089
0.531
0.214
-0.23
+0.31
2
7
0.841
0.916
0.368
-0.09
+0.19
3
9
0.044
0.400
0.202
+0.47
-0.62
yolov10n-esmoe-rewire-e120-s3-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3006 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.438
0.222
+0.23
-0.25
1
5
0.093
0.292
0.181
-0.24
+0.29
2
7
0.890
0.923
0.380
+0.04
-0.03
3
9
0.015
0.347
0.217
+0.30
-0.43
yolov10n-esmoe-rewire-e120-s4-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3337 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.213
0.230
+0.09
-0.33
1
5
0.755
0.909
0.332
+0.11
-0.08
2
7
0.203
0.460
0.214
-0.25
+0.39
3
9
0.042
0.418
0.224
+0.24
-0.45
yolov10n-esmoe-rewire-stages-e120-s0-p800-best
548 images through block 0 (32 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.352 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.708
0.810
0.310
-0.21
+0.32
1
5
0.122
0.256
0.204
+0.08
-0.12
2
7
0.157
0.370
0.242
+0.22
-0.37
3
9
0.013
0.564
0.243
-0.03
+0.16
548 images through block 1 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2674 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.416
0.487
0.244
+0.18
-0.16
1
5
0.566
0.611
0.260
-0.28
+0.28
2
7
0.002
0.259
0.248
+0.36
-0.41
3
9
0.016
0.642
0.248
+0.36
-0.41
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3498 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.332
0.662
0.257
+0.27
-0.21
1
5
0.606
0.781
0.303
-0.32
+0.51
2
7
0.053
0.265
0.222
+0.18
-0.59
3
9
0.009
0.292
0.218
-0.03
-0.22
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.317 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.276
0.211
+0.05
-0.20
1
5
0.057
0.330
0.196
-0.21
+0.47
2
7
0.161
0.489
0.231
+0.28
-0.25
3
9
0.783
0.905
0.362
-0.10
-0.05
yolov10n-esmoe-rewire-stages-e120-s1-p800-best
548 images through block 0 (32 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3319 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.161
0.391
0.220
+0.14
-0.15
1
5
0.752
0.814
0.340
-0.23
+0.35
2
7
0.084
0.219
0.212
+0.24
-0.39
3
9
0.004
0.577
0.229
+0.19
-0.42
548 images through block 1 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3559 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.661
0.734
0.290
-0.10
+0.38
1
5
0.279
0.504
0.234
-0.03
-0.08
2
7
0.057
0.454
0.238
+0.25
-0.58
3
9
0.004
0.308
0.238
+0.25
-0.57
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3577 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.305
0.586
0.261
+0.34
-0.37
1
5
0.068
0.279
0.226
+0.15
-0.55
2
7
0.004
0.332
0.218
-0.28
+0.19
3
9
0.624
0.803
0.296
-0.28
+0.59
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2896 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.164
0.195
+0.06
-0.23
1
5
0.073
0.531
0.218
+0.09
-0.16
2
7
0.905
0.954
0.405
+0.01
+0.10
3
9
0.022
0.350
0.181
-0.21
+0.17
yolov10n-esmoe-rewire-stages-e120-s2-p800-best
548 images through block 0 (32 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3745 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.093
0.770
0.261
-0.12
+0.29
1
5
0.655
0.792
0.281
-0.26
+0.10
2
7
0.184
0.268
0.232
+0.15
-0.08
3
9
0.068
0.170
0.226
+0.20
-0.23
548 images through block 1 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3824 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.699
0.863
0.271
-0.25
+0.54
1
5
0.031
0.620
0.246
-0.21
+0.21
2
7
0.013
0.122
0.239
-0.03
-0.08
3
9
0.257
0.394
0.243
+0.31
-0.52
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3351 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.179
0.693
0.267
-0.16
+0.30
1
5
0.588
0.783
0.296
-0.17
+0.43
2
7
0.069
0.157
0.199
-0.03
-0.44
3
9
0.164
0.367
0.238
+0.38
-0.43
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2818 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.411
0.201
+0.25
-0.30
1
5
0.053
0.237
0.203
+0.07
-0.16
2
7
0.046
0.400
0.182
-0.23
+0.40
3
9
0.901
0.953
0.414
+0.04
-0.07
yolov10n-esmoe-rewire-stages-w0.0025-e120-s0-p800-best
548 images through block 0 (32 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3848 of 1.3863, distinct top-2 pairs seen: 3 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.630
0.244
-0.29
+0.36
1
5
0.998
1.000
0.272
+0.33
-0.31
2
7
0.000
0.192
0.242
-0.24
-0.03
3
9
0.000
0.179
0.242
-0.17
-0.17
548 images through block 1 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: [2, 3], mean entropy 1.3842 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
1.000
1.000
0.276
+0.31
-0.45
1
5
0.000
1.000
0.246
-0.31
+0.47
2
7
0.000
0.000
0.239
-0.31
+0.45
3
9
0.000
0.000
0.239
-0.30
+0.42
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3828 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.708
0.945
0.272
-0.28
+0.40
1
5
0.292
0.624
0.250
+0.23
-0.22
2
7
0.000
0.033
0.239
+0.05
-0.40
3
9
0.000
0.398
0.239
+0.04
-0.39
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2901 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.148
0.233
+0.23
-0.54
1
5
0.657
0.894
0.342
+0.02
+0.09
2
7
0.146
0.256
0.159
-0.23
+0.36
3
9
0.197
0.703
0.266
+0.26
-0.48
yolov10n-esmoe-rewire-stages-w0.0025-e120-s1-p800-best
548 images through block 0 (32 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3829 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.097
0.195
0.234
+0.22
-0.19
1
5
0.901
0.943
0.279
-0.24
+0.21
2
7
0.000
0.584
0.244
+0.21
-0.14
3
9
0.002
0.277
0.244
+0.21
-0.20
548 images through block 1 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3845 of 1.3863, distinct top-2 pairs seen: 3 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.447
0.242
+0.13
-0.38
1
5
0.000
0.425
0.241
+0.01
-0.03
2
7
1.000
1.000
0.275
-0.08
+0.22
3
9
0.000
0.128
0.242
+0.20
-0.54
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3729 of 1.3863, distinct top-2 pairs seen: 4 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.235
0.330
0.229
+0.34
-0.48
1
5
0.000
0.599
0.239
-0.13
-0.32
2
7
0.000
0.204
0.239
-0.12
-0.34
3
9
0.765
0.867
0.292
-0.32
+0.57
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3119 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.721
0.243
+0.47
-0.51
1
5
0.878
0.911
0.369
+0.10
-0.21
2
7
0.077
0.184
0.195
-0.22
+0.31
3
9
0.044
0.184
0.193
-0.12
+0.24
yolov10n-esmoe-rewire-stages-w0.0025-e120-s2-p800-best
548 images through block 0 (32 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3824 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.053
0.367
0.248
+0.28
-0.23
1
5
0.659
0.688
0.267
-0.33
+0.32
2
7
0.097
0.124
0.234
+0.26
-0.30
3
9
0.192
0.821
0.251
+0.31
-0.26
548 images through block 1 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.378 of 1.3863, distinct top-2 pairs seen: 3 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.321
0.232
+0.08
-0.52
1
5
0.000
0.301
0.232
+0.09
-0.51
2
7
0.000
0.378
0.231
+0.19
-0.39
3
9
1.000
1.000
0.306
-0.16
+0.47
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3587 of 1.3863, distinct top-2 pairs seen: 4 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.617
0.234
-0.05
+0.04
1
5
0.208
0.312
0.223
+0.28
-0.50
2
7
0.792
0.872
0.308
-0.30
+0.56
3
9
0.000
0.199
0.234
-0.04
+0.04
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2756 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.055
0.211
+0.31
-0.60
1
5
0.058
0.257
0.155
-0.20
+0.32
2
7
0.892
0.945
0.397
-0.11
+0.19
3
9
0.049
0.743
0.236
+0.35
-0.53
yolov10n-esmoe-w0-e120-s0-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [0, 3], mean entropy 1.3779 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.000
0.206
-0.36
+0.17
1
5
0.036
1.000
0.269
+0.20
+0.02
2
7
0.964
1.000
0.290
+0.18
-0.22
3
9
0.000
0.000
0.235
-0.28
+0.20
yolov10n-esmoe-w0-e120-s1-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [1, 3], mean entropy 1.3705 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
1.000
0.261
+0.16
+0.17
1
5
0.000
0.000
0.211
-0.20
+0.11
2
7
1.000
1.000
0.318
+0.09
-0.35
3
9
0.000
0.000
0.209
-0.05
+0.39
yolov10n-esmoe-w0-e120-s2-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [1, 2], mean entropy 1.3771 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
1.000
0.268
-0.07
+0.37
1
5
0.000
0.000
0.220
+0.24
-0.35
2
7
0.000
0.000
0.216
-0.15
+0.07
3
9
1.000
1.000
0.297
-0.02
-0.01
yolov5n-esmoe-dense-e120-s0-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3124 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.547
0.230
+0.42
-0.46
1
5
0.000
0.060
0.218
+0.37
-0.41
2
7
0.547
0.626
0.269
-0.38
+0.48
3
9
0.451
0.766
0.283
+0.31
-0.43
yolov5n-esmoe-dense-e120-s1-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2917 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.161
0.203
+0.01
-0.03
1
5
0.027
0.436
0.212
+0.10
-0.14
2
7
0.847
0.916
0.388
+0.18
-0.32
3
9
0.126
0.487
0.197
-0.31
+0.52
yolov5n-esmoe-dense-e120-s2-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3096 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.004
0.666
0.244
+0.14
-0.30
1
5
0.075
0.182
0.191
-0.10
+0.07
2
7
0.080
0.263
0.208
-0.12
+0.38
3
9
0.841
0.889
0.356
+0.13
-0.22
yolov5n-esmoe-e120-s0-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3236 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.024
0.263
0.215
-0.15
+0.08
1
5
0.016
0.148
0.199
-0.09
+0.22
2
7
0.141
0.642
0.233
+0.14
+0.01
3
9
0.819
0.947
0.353
+0.02
-0.15
yolov5n-esmoe-e120-s0-p800h2-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2538 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.525
0.229
+0.37
-0.44
1
5
0.007
0.219
0.218
+0.18
-0.19
2
7
0.825
0.905
0.387
+0.20
-0.21
3
9
0.166
0.350
0.167
-0.29
+0.32
yolov5n-esmoe-e120-s1-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3227 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.319
0.222
+0.04
+0.02
1
5
0.031
0.407
0.217
+0.16
-0.24
2
7
0.120
0.376
0.204
-0.24
+0.41
3
9
0.847
0.898
0.356
+0.10
-0.22
yolov5n-esmoe-e120-s1-p800h2-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2969 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.221
0.201
-0.03
+0.04
1
5
0.027
0.170
0.188
+0.05
-0.21
2
7
0.752
0.885
0.376
+0.22
-0.29
3
9
0.221
0.725
0.235
-0.33
+0.53
yolov5n-esmoe-e120-s2-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3268 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.004
0.555
0.227
-0.04
-0.04
1
5
0.051
0.327
0.224
+0.17
-0.22
2
7
0.058
0.201
0.187
-0.23
+0.35
3
9
0.887
0.918
0.363
+0.07
-0.11
yolov5n-esmoe-e120-s2-p800h2-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2445 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.734
0.243
+0.22
-0.31
1
5
0.005
0.126
0.224
+0.14
-0.14
2
7
0.137
0.228
0.146
-0.20
+0.27
3
9
0.858
0.912
0.387
+0.16
-0.22
yolov5n-esmoe-gshard-e120-s2-p800h2g-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [1, 2], mean entropy 1.3767 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.931
1.000
0.295
-0.09
-0.15
1
5
0.000
0.000
0.215
+0.24
-0.14
2
7
0.000
0.000
0.220
+0.20
+0.19
3
9
0.069
1.000
0.271
-0.13
+0.20
yolov5n-esmoe-master-e120-s0-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [0], mean entropy 1.344 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.009
0.203
-0.16
-0.21
1
5
0.000
0.381
0.213
-0.19
-0.10
2
7
0.272
0.715
0.265
+0.29
-0.35
3
9
0.728
0.894
0.319
-0.25
+0.46
yolov5n-esmoe-master-e120-s1-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3597 of 1.3863, distinct top-2 pairs seen: 4 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.058
0.215
+0.03
-0.44
1
5
0.000
0.044
0.195
-0.20
-0.01
2
7
0.122
0.905
0.264
+0.32
-0.22
3
9
0.878
0.993
0.327
-0.30
+0.39
yolov5n-esmoe-master-w0.32-e120-s0-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [1], mean entropy 1.3657 of 1.3863, distinct top-2 pairs seen: 3 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.212
0.845
0.269
+0.11
+0.20
1
5
0.000
0.000
0.209
-0.01
+0.15
2
7
0.179
0.296
0.229
-0.22
+0.24
3
9
0.610
0.860
0.293
+0.19
-0.42
yolov5n-esmoe-master-w0.32-e120-s1-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [0, 3], mean entropy 1.3612 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.000
0.190
+0.10
-0.20
1
5
0.113
1.000
0.286
-0.12
+0.21
2
7
0.887
1.000
0.320
+0.20
-0.18
3
9
0.000
0.000
0.204
-0.46
+0.25
yolov5n-esmoe-master-w0.32-e120-s2-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [2], mean entropy 1.362 of 1.3863, distinct top-2 pairs seen: 3 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.593
0.794
0.296
+0.31
-0.46
1
5
0.164
0.308
0.228
-0.18
+0.26
2
7
0.000
0.000
0.200
-0.20
+0.11
3
9
0.243
0.898
0.276
-0.17
+0.31
yolov5n-esmoe-norm-e120-s0-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3489 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.035
0.175
0.228
-0.06
+0.05
1
5
0.009
0.173
0.207
+0.08
+0.07
2
7
0.892
0.960
0.332
-0.12
+0.19
3
9
0.064
0.692
0.234
+0.21
-0.41
yolov5n-esmoe-norm-e120-s1-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3193 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.376
0.221
+0.04
-0.03
1
5
0.036
0.451
0.226
-0.01
+0.01
2
7
0.057
0.235
0.186
-0.25
+0.28
3
9
0.907
0.938
0.367
+0.17
-0.20
yolov5n-esmoe-norm-e120-s2-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3152 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.414
0.212
-0.10
+0.05
1
5
0.029
0.244
0.203
+0.10
-0.15
2
7
0.069
0.405
0.205
-0.17
+0.38
3
9
0.901
0.936
0.380
+0.09
-0.20
yolov5n-esmoe-rewire-e120-s0-p800-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3094 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.057
0.376
0.221
-0.07
+0.11
1
5
0.016
0.244
0.197
-0.02
+0.22
2
7
0.027
0.447
0.200
+0.46
-0.48
3
9
0.900
0.932
0.383
-0.17
+0.05
yolov5n-esmoe-rewire-e120-s0-p800h2-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3447 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.029
0.086
0.216
-0.08
+0.08
1
5
0.016
0.184
0.204
-0.06
+0.10
2
7
0.805
0.943
0.329
-0.05
+0.29
3
9
0.150
0.786
0.251
+0.15
-0.39
yolov5n-esmoe-rewire-e120-s1-p800h2-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2799 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.208
0.197
-0.07
+0.05
1
5
0.029
0.464
0.203
+0.02
-0.09
2
7
0.921
0.945
0.417
+0.15
-0.19
3
9
0.049
0.383
0.183
-0.24
+0.37
yolov5n-esmoe-rewire-e120-s2-p800h2-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3355 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.022
0.549
0.237
-0.23
+0.25
1
5
0.047
0.303
0.214
+0.37
-0.32
2
7
0.179
0.323
0.217
-0.24
+0.42
3
9
0.752
0.825
0.332
+0.09
-0.26
yolov5n-esmoe-rewire-stages-e120-s0-p800h2-best
548 images through block 0 (32 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.367 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.058
0.808
0.254
-0.12
+0.28
1
5
0.810
0.883
0.304
-0.19
+0.14
2
7
0.104
0.150
0.207
+0.16
-0.21
3
9
0.027
0.159
0.234
+0.19
-0.08
548 images through block 1 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3554 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.310
0.591
0.243
-0.15
+0.47
1
5
0.688
0.818
0.308
+0.12
-0.36
2
7
0.000
0.197
0.225
-0.04
+0.07
3
9
0.002
0.394
0.225
-0.03
+0.07
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3431 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.363
0.642
0.258
-0.04
+0.38
1
5
0.611
0.734
0.288
-0.03
-0.09
2
7
0.016
0.358
0.227
+0.10
-0.38
3
9
0.009
0.266
0.227
+0.11
-0.39
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2542 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.356
0.186
-0.04
-0.12
1
5
0.058
0.350
0.193
-0.02
-0.19
2
7
0.921
0.947
0.445
+0.07
+0.10
3
9
0.020
0.347
0.175
-0.11
+0.16
yolov5n-esmoe-rewire-stages-e120-s1-p800h2-best
548 images through block 0 (32 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3514 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.115
0.779
0.267
-0.15
+0.32
1
5
0.093
0.173
0.195
+0.20
-0.25
2
7
0.775
0.914
0.319
-0.16
+0.13
3
9
0.016
0.133
0.219
+0.10
-0.19
548 images through block 1 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3615 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.874
0.892
0.327
-0.23
+0.44
1
5
0.126
0.182
0.215
+0.23
-0.41
2
7
0.000
0.699
0.229
+0.23
-0.46
3
9
0.000
0.226
0.229
+0.23
-0.46
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3572 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.595
0.234
+0.25
-0.38
1
5
0.168
0.307
0.217
+0.33
-0.47
2
7
0.807
0.841
0.315
-0.33
+0.47
3
9
0.025
0.257
0.234
+0.26
-0.40
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2836 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.035
0.367
0.218
-0.01
-0.05
1
5
0.025
0.613
0.221
+0.23
-0.29
2
7
0.004
0.068
0.147
-0.07
+0.22
3
9
0.936
0.953
0.413
-0.07
+0.07
yolov5n-esmoe-rewire-stages-e120-s2-p800h2-best
548 images through block 0 (32 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3585 of 1.3863, distinct top-2 pairs seen: 4 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.235
0.332
0.214
+0.25
-0.20
1
5
0.000
0.489
0.236
-0.14
+0.19
2
7
0.765
0.916
0.314
-0.26
+0.18
3
9
0.000
0.263
0.236
-0.14
+0.18
548 images through block 1 (64 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3425 of 1.3863, distinct top-2 pairs seen: 4 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.511
0.222
+0.17
-0.35
1
5
0.748
0.847
0.330
-0.26
+0.43
2
7
0.252
0.359
0.225
+0.23
-0.36
3
9
0.000
0.283
0.222
+0.17
-0.36
548 images through block 2 (128 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3485 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.522
0.697
0.278
-0.32
+0.52
1
5
0.474
0.599
0.258
+0.33
-0.40
2
7
0.002
0.432
0.232
-0.12
-0.28
3
9
0.002
0.272
0.232
-0.11
-0.29
548 images through block 3 (256 channels), kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2901 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.135
0.198
+0.03
-0.15
1
5
0.120
0.666
0.232
-0.03
-0.17
2
7
0.860
0.929
0.397
+0.10
+0.02
3
9
0.020
0.270
0.174
-0.15
+0.31
yolov5n-esmoe-w0-e120-s0-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [1, 3], mean entropy 1.3516 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
1.000
0.253
-0.17
+0.38
1
5
0.000
0.000
0.199
-0.29
+0.15
2
7
0.998
1.000
0.356
+0.24
-0.25
3
9
0.000
0.000
0.192
-0.16
+0.06
yolov5n-esmoe-w0-e120-s1-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [0, 3], mean entropy 1.3456 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.000
0.189
-0.24
+0.15
1
5
0.004
1.000
0.253
-0.03
+0.21
2
7
0.996
1.000
0.364
+0.21
-0.25
3
9
0.000
0.000
0.194
-0.30
+0.23
yolov5n-esmoe-w0-e120-s2-p800h2-best
548 images, kernels [3, 5, 7, 9], top-2, dead experts: [1, 2], mean entropy 1.3762 of 1.3863, distinct top-2 pairs seen: 1 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.608
1.000
0.288
+0.22
-0.22
1
5
0.000
0.000
0.216
-0.17
+0.12
2
7
0.000
0.000
0.215
-0.30
+0.20
3
9
0.392
1.000
0.280
+0.01
+0.05
yolov8n-esmoe-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3023 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.055
0.202
+0.39
-0.45
1
5
0.011
0.551
0.220
+0.47
-0.42
2
7
0.246
0.474
0.218
-0.30
+0.48
3
9
0.743
0.920
0.360
+0.10
-0.31
yolov8n-esmoe-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2586 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.213
0.187
+0.11
-0.23
1
5
0.011
0.381
0.191
+0.12
-0.14
2
7
0.920
0.943
0.437
+0.11
-0.14
3
9
0.069
0.462
0.185
-0.27
+0.38
yolov8n-esmoe-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3032 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.170
0.195
+0.32
-0.20
1
5
0.004
0.053
0.175
+0.26
-0.35
2
7
0.381
0.876
0.293
-0.45
+0.49
3
9
0.615
0.901
0.337
+0.26
-0.30
yolov8n-esmoe-rewire-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2849 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.239
0.197
-0.04
-0.01
1
5
0.038
0.402
0.204
+0.02
-0.08
2
7
0.046
0.409
0.192
-0.16
+0.30
3
9
0.916
0.951
0.408
+0.10
-0.13
yolov8n-esmoe-rewire-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3135 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.015
0.173
0.214
-0.06
+0.20
1
5
0.057
0.571
0.221
-0.17
+0.42
2
7
0.885
0.931
0.376
-0.04
-0.01
3
9
0.044
0.325
0.190
+0.21
-0.41
yolov8n-esmoe-rewire-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.292 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.002
0.411
0.206
+0.14
-0.12
1
5
0.036
0.179
0.197
+0.02
-0.09
2
7
0.100
0.473
0.206
-0.19
+0.26
3
9
0.861
0.938
0.391
+0.09
-0.08
yolov9t-esmoe-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2878 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.005
0.195
0.206
+0.12
-0.22
1
5
0.170
0.476
0.213
-0.28
+0.44
2
7
0.016
0.420
0.200
+0.37
-0.46
3
9
0.808
0.909
0.381
+0.07
-0.15
yolov9t-esmoe-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.2605 of 1.3863, distinct top-2 pairs seen: 5 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.036
0.164
+0.07
-0.24
1
5
0.179
0.790
0.265
-0.29
+0.57
2
7
0.027
0.230
0.171
+0.16
-0.32
3
9
0.794
0.943
0.400
+0.12
-0.19
yolov9t-esmoe-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.331 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.004
0.110
0.210
-0.01
-0.19
1
5
0.182
0.555
0.230
-0.33
+0.34
2
7
0.779
0.901
0.344
+0.04
+0.03
3
9
0.035
0.434
0.216
+0.50
-0.53
yolov9t-esmoe-rewire-e120-s0-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3339 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.000
0.084
0.204
+0.04
-0.32
1
5
0.244
0.527
0.242
-0.35
+0.59
2
7
0.212
0.533
0.237
+0.19
-0.22
3
9
0.544
0.856
0.316
+0.18
-0.33
yolov9t-esmoe-rewire-e120-s1-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3183 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.016
0.234
0.212
-0.12
+0.04
1
5
0.825
0.942
0.360
-0.22
+0.22
2
7
0.025
0.073
0.168
-0.16
+0.07
3
9
0.133
0.752
0.259
+0.47
-0.36
yolov9t-esmoe-rewire-e120-s2-p800-best.pt
548 images, kernels [3, 5, 7, 9], top-2, dead experts: none, mean entropy 1.3481 of 1.3863, distinct top-2 pairs seen: 6 of 6.
expert
kernel
top-1 share
top-2 share
mean prob
corr. size
corr. count
0
3
0.139
0.639
0.257
+0.11
-0.36
1
5
0.491
0.611
0.266
-0.41
+0.58
2
7
0.343
0.626
0.276
+0.34
-0.33
3
9
0.027
0.124
0.202
+0.06
-0.22
Balancing pressure
The gradient each objective puts on the router's logits at one weight, measured at the mean routing probabilities each checkpoint converged to. Generated by scripts/pressure.py.
Fig. 7 Gradient of the balance terms at one weight, median over checkpoints
Data: results/pressure.md
Mean |d term / d logit| at the mean routing probabilities each checkpoint converged to,
over 548 rows scattered by the spread column. Generated by scripts/pressure.py.
checkpoint
spread
switch
gshard_probs
gshard(gate)
master
gate/switch
yolo-master-n-esmoe-upstream-w1-e120-s0-p800c
0.15
7.152e-04
1.635e-04
2.635e-04
1.647e-05
0.37
yolo-master-n-esmoe-upstream-w1-e120-s0-p800c
0.5
2.917e-04
1.526e-04
2.228e-04
1.393e-05
0.76
yolo-master-n-esmoe-upstream-w1-e120-s0-p800c
1.5
8.990e-05
9.196e-05
8.646e-05
5.404e-06
0.96
yolo-master-n-esmoe-upstream-w1-e120-s0-p800f-fp32
0.15
7.086e-04
1.576e-04
2.545e-04
1.591e-05
0.36
yolo-master-n-esmoe-upstream-w1-e120-s0-p800f-fp32
0.5
2.840e-04
1.472e-04
2.144e-04
1.340e-05
0.75
yolo-master-n-esmoe-upstream-w1-e120-s0-p800f-fp32
1.5
8.791e-05
8.879e-05
8.305e-05
5.190e-06
0.94
yolo-master-n-esmoe-upstream-w1-e120-s1-p800d
0.15
6.632e-04
1.369e-04
3.194e-04
1.997e-05
0.48
yolo-master-n-esmoe-upstream-w1-e120-s1-p800d
0.5
1.937e-04
1.171e-04
1.908e-04
1.192e-05
0.98
yolo-master-n-esmoe-upstream-w1-e120-s1-p800d
1.5
3.885e-05
5.665e-05
5.155e-05
3.222e-06
1.33
yolo-master-n-esmoe-upstream-w1-e120-s1-p800f-fp32
0.15
7.446e-04
1.616e-04
3.020e-04
1.887e-05
0.41
yolo-master-n-esmoe-upstream-w1-e120-s1-p800f-fp32
0.5
2.338e-04
1.411e-04
2.357e-04
1.473e-05
1.01
yolo-master-n-esmoe-upstream-w1-e120-s1-p800f-fp32
1.5
3.867e-05
6.539e-05
5.590e-05
3.493e-06
1.45
yolo-master-n-esmoe-upstream-w1-e120-s2-p800e
0.15
6.809e-04
1.357e-04
3.203e-04
2.002e-05
0.47
yolo-master-n-esmoe-upstream-w1-e120-s2-p800e
0.5
1.891e-04
1.164e-04
1.872e-04
1.170e-05
0.99
yolo-master-n-esmoe-upstream-w1-e120-s2-p800e
1.5
3.176e-05
5.293e-05
4.610e-05
2.881e-06
1.45
yolo-master-n-esmoe-upstream-w1-e120-s2-p800f-fp32
0.15
7.526e-04
1.631e-04
2.802e-04
1.752e-05
0.37
yolo-master-n-esmoe-upstream-w1-e120-s2-p800f-fp32
0.5
2.336e-04
1.420e-04
2.304e-04
1.440e-05
0.99
yolo-master-n-esmoe-upstream-w1-e120-s2-p800f-fp32
1.5
3.889e-05
6.541e-05
5.519e-05
3.449e-06
1.42
yolo11n-esmoe-e120-s0-p800-best
0.15
4.804e-04
1.255e-04
5.640e-04
3.525e-05
1.17
yolo11n-esmoe-e120-s0-p800-best
0.5
2.072e-04
1.219e-04
2.248e-04
1.405e-05
1.08
yolo11n-esmoe-e120-s0-p800-best
1.5
6.772e-05
7.801e-05
7.898e-05
4.936e-06
1.17
yolo11n-esmoe-e120-s1-p800-best
0.15
4.159e-04
1.058e-04
5.260e-04
3.288e-05
1.26
yolo11n-esmoe-e120-s1-p800-best
0.5
1.662e-04
9.108e-05
1.741e-04
1.088e-05
1.05
yolo11n-esmoe-e120-s1-p800-best
1.5
4.836e-05
5.319e-05
5.466e-05
3.416e-06
1.13
yolo11n-esmoe-e120-s2-p800-best
0.15
4.844e-04
2.006e-04
5.648e-04
3.530e-05
1.17
yolo11n-esmoe-e120-s2-p800-best
0.5
2.683e-04
1.831e-04
2.694e-04
1.684e-05
1.00
yolo11n-esmoe-e120-s2-p800-best
1.5
9.321e-05
1.126e-04
1.101e-04
6.879e-06
1.18
yolo11n-esmoe-rewire-e120-s0-p800-best
0.15
5.479e-04
1.928e-04
6.418e-04
4.011e-05
1.17
yolo11n-esmoe-rewire-e120-s0-p800-best
0.5
2.704e-04
1.803e-04
3.125e-04
1.953e-05
1.16
yolo11n-esmoe-rewire-e120-s0-p800-best
1.5
8.413e-05
1.048e-04
1.045e-04
6.533e-06
1.24
yolo11n-esmoe-rewire-e120-s1-p800-best
0.15
5.567e-04
1.923e-04
6.532e-04
4.082e-05
1.17
yolo11n-esmoe-rewire-e120-s1-p800-best
0.5
2.559e-04
1.659e-04
3.007e-04
1.879e-05
1.17
yolo11n-esmoe-rewire-e120-s1-p800-best
1.5
7.797e-05
8.691e-05
8.990e-05
5.619e-06
1.15
yolo11n-esmoe-rewire-e120-s2-p800-best
0.15
5.808e-04
2.411e-04
6.123e-04
3.827e-05
1.05
yolo11n-esmoe-rewire-e120-s2-p800-best
0.5
3.055e-04
2.243e-04
3.657e-04
2.286e-05
1.20
yolo11n-esmoe-rewire-e120-s2-p800-best
1.5
1.044e-04
1.296e-04
1.271e-04
7.942e-06
1.22
yolo12n-esmoe-e120-s0-p800-best
0.15
7.369e-04
2.091e-04
3.311e-04
2.070e-05
0.45
yolo12n-esmoe-e120-s0-p800-best
0.5
2.955e-04
1.820e-04
2.513e-04
1.570e-05
0.85
yolo12n-esmoe-e120-s0-p800-best
1.5
8.160e-05
8.630e-05
8.221e-05
5.138e-06
1.01
yolo12n-esmoe-e120-s1-p800-best
0.15
7.687e-04
1.745e-04
9.176e-05
5.735e-06
0.12
yolo12n-esmoe-e120-s1-p800-best
0.5
3.035e-04
1.629e-04
1.930e-04
1.206e-05
0.64
yolo12n-esmoe-e120-s1-p800-best
1.5
9.114e-05
9.792e-05
9.031e-05
5.644e-06
0.99
yolo12n-esmoe-e120-s2-p800-best
0.15
5.426e-04
3.510e-04
7.401e-04
4.625e-05
1.36
yolo12n-esmoe-e120-s2-p800-best
0.5
4.017e-04
3.175e-04
5.102e-04
3.189e-05
1.27
yolo12n-esmoe-e120-s2-p800-best
1.5
1.235e-04
1.727e-04
1.634e-04
1.021e-05
1.32
yolo12n-esmoe-rewire-e120-s0-p800-best
0.15
5.293e-04
2.000e-04
7.148e-04
4.468e-05
1.35
yolo12n-esmoe-rewire-e120-s0-p800-best
0.5
2.701e-04
1.727e-04
3.152e-04
1.970e-05
1.17
yolo12n-esmoe-rewire-e120-s0-p800-best
1.5
7.773e-05
8.421e-05
8.409e-05
5.256e-06
1.08
yolo12n-esmoe-rewire-e120-s1-p800-best
0.15
5.382e-04
2.554e-04
7.047e-04
4.404e-05
1.31
yolo12n-esmoe-rewire-e120-s1-p800-best
0.5
3.355e-04
2.241e-04
4.072e-04
2.545e-05
1.21
yolo12n-esmoe-rewire-e120-s1-p800-best
1.5
6.439e-05
1.064e-04
8.813e-05
5.508e-06
1.37
yolo12n-esmoe-rewire-e120-s2-p800-best
0.15
7.587e-04
2.420e-04
3.165e-04
1.978e-05
0.42
yolo12n-esmoe-rewire-e120-s2-p800-best
0.5
3.513e-04
2.141e-04
3.076e-04
1.922e-05
0.88
yolo12n-esmoe-rewire-e120-s2-p800-best
1.5
9.520e-05
1.014e-04
9.744e-05
6.090e-06
1.02
yolo26n-esmoe-e120-s0-p800-best
0.15
5.211e-04
1.815e-04
7.114e-04
4.446e-05
1.37
yolo26n-esmoe-e120-s0-p800-best
0.5
2.493e-04
1.566e-04
2.839e-04
1.775e-05
1.14
yolo26n-esmoe-e120-s0-p800-best
1.5
7.196e-05
7.791e-05
7.811e-05
4.882e-06
1.09
yolo26n-esmoe-e120-s1-p800-best
0.15
3.092e-04
1.262e-04
2.090e-04
1.306e-05
0.68
yolo26n-esmoe-e120-s1-p800-best
0.5
2.066e-04
1.203e-04
1.391e-04
8.695e-06
0.67
yolo26n-esmoe-e120-s1-p800-best
1.5
7.954e-05
8.469e-05
7.666e-05
4.791e-06
0.96
yolo26n-esmoe-e120-s2-p800-best
0.15
5.184e-04
1.872e-04
6.747e-04
4.217e-05
1.30
yolo26n-esmoe-e120-s2-p800-best
0.5
2.577e-04
1.611e-04
2.983e-04
1.864e-05
1.16
yolo26n-esmoe-e120-s2-p800-best
1.5
8.129e-05
8.651e-05
9.012e-05
5.632e-06
1.11
yolo26n-esmoe-rewire-e120-s0-p800-best
0.15
5.397e-04
2.506e-04
6.689e-04
4.180e-05
1.24
yolo26n-esmoe-rewire-e120-s0-p800-best
0.5
3.312e-04
2.178e-04
4.034e-04
2.521e-05
1.22
yolo26n-esmoe-rewire-e120-s0-p800-best
1.5
9.346e-05
1.085e-04
1.115e-04
6.968e-06
1.19
yolo26n-esmoe-rewire-e120-s1-p800-best
0.15
5.305e-04
2.490e-04
6.806e-04
4.254e-05
1.28
yolo26n-esmoe-rewire-e120-s1-p800-best
0.5
3.302e-04
2.194e-04
3.981e-04
2.488e-05
1.21
yolo26n-esmoe-rewire-e120-s1-p800-best
1.5
6.693e-05
1.067e-04
9.158e-05
5.724e-06
1.37
yolo26n-esmoe-rewire-e120-s2-p800-best
0.15
4.804e-04
1.520e-04
6.237e-04
3.898e-05
1.30
yolo26n-esmoe-rewire-e120-s2-p800-best
0.5
2.233e-04
1.312e-04
2.502e-04
1.564e-05
1.12
yolo26n-esmoe-rewire-e120-s2-p800-best
1.5
6.949e-05
7.376e-05
7.730e-05
4.831e-06
1.11
yolov10n-esmoe-dense-e120-s0-p800-best
0.15
3.887e-04
7.573e-05
3.982e-04
2.489e-05
1.02
yolov10n-esmoe-dense-e120-s0-p800-best
0.5
1.366e-04
7.834e-05
1.391e-04
8.691e-06
1.02
yolov10n-esmoe-dense-e120-s0-p800-best
1.5
4.752e-05
5.878e-05
6.056e-05
3.785e-06
1.27
yolov10n-esmoe-dense-e120-s1-p800-best
0.15
4.565e-04
1.326e-04
5.887e-04
3.679e-05
1.29
yolov10n-esmoe-dense-e120-s1-p800-best
0.5
1.944e-04
1.149e-04
2.088e-04
1.305e-05
1.07
yolov10n-esmoe-dense-e120-s1-p800-best
1.5
5.777e-05
5.910e-05
5.852e-05
3.657e-06
1.01
yolov10n-esmoe-dense-e120-s2-p800-best
0.15
5.437e-04
2.200e-04
7.293e-04
4.558e-05
1.34
yolov10n-esmoe-dense-e120-s2-p800-best
0.5
2.758e-04
1.890e-04
3.283e-04
2.052e-05
1.19
yolov10n-esmoe-dense-e120-s2-p800-best
1.5
7.781e-05
9.447e-05
9.332e-05
5.833e-06
1.20
yolov10n-esmoe-e120-s0-p800-best
0.15
3.649e-04
7.976e-05
4.319e-04
2.699e-05
1.18
yolov10n-esmoe-e120-s0-p800-best
0.5
1.437e-04
6.937e-05
1.513e-04
9.457e-06
1.05
yolov10n-esmoe-e120-s0-p800-best
1.5
3.885e-05
4.833e-05
4.808e-05
3.005e-06
1.24
yolov10n-esmoe-e120-s1-p800-best
0.15
3.909e-04
1.025e-04
3.577e-04
2.236e-05
0.92
yolov10n-esmoe-e120-s1-p800-best
0.5
1.887e-04
1.019e-04
1.802e-04
1.127e-05
0.96
yolov10n-esmoe-e120-s1-p800-best
1.5
6.243e-05
6.839e-05
6.693e-05
4.183e-06
1.07
yolov10n-esmoe-e120-s2-p800-best
0.15
5.731e-04
3.413e-04
8.405e-04
5.253e-05
1.47
yolov10n-esmoe-e120-s2-p800-best
0.5
4.077e-04
2.964e-04
5.202e-04
3.251e-05
1.28
yolov10n-esmoe-e120-s2-p800-best
1.5
9.280e-05
1.337e-04
1.193e-04
7.459e-06
1.29
yolov10n-esmoe-e120-s3-p800-best
0.15
5.219e-04
1.832e-04
6.983e-04
4.365e-05
1.34
yolov10n-esmoe-e120-s3-p800-best
0.5
2.479e-04
1.614e-04
2.817e-04
1.761e-05
1.14
yolov10n-esmoe-e120-s3-p800-best
1.5
4.564e-05
7.806e-05
6.902e-05
4.314e-06
1.51
yolov10n-esmoe-e120-s4-p800-best
0.15
5.487e-04
2.665e-04
6.767e-04
4.229e-05
1.23
yolov10n-esmoe-e120-s4-p800-best
0.5
3.456e-04
2.445e-04
4.196e-04
2.623e-05
1.21
yolov10n-esmoe-e120-s4-p800-best
1.5
1.128e-04
1.356e-04
1.376e-04
8.598e-06
1.22
yolov10n-esmoe-norm-e120-s0-p800-best
0.15
4.646e-04
1.581e-04
5.258e-04
3.286e-05
1.13
yolov10n-esmoe-norm-e120-s0-p800-best
0.5
2.481e-04
1.538e-04
2.782e-04
1.739e-05
1.12
yolov10n-esmoe-norm-e120-s0-p800-best
1.5
7.846e-05
1.004e-04
9.712e-05
6.070e-06
1.24
yolov10n-esmoe-rewire-e120-s0-p800-best
0.15
5.640e-04
2.976e-04
8.054e-04
5.034e-05
1.43
yolov10n-esmoe-rewire-e120-s0-p800-best
0.5
3.517e-04
2.552e-04
4.426e-04
2.766e-05
1.26
yolov10n-esmoe-rewire-e120-s0-p800-best
1.5
9.922e-05
1.213e-04
1.165e-04
7.284e-06
1.17
yolov10n-esmoe-rewire-e120-s1-p800-best
0.15
5.493e-04
2.643e-04
7.489e-04
4.681e-05
1.36
yolov10n-esmoe-rewire-e120-s1-p800-best
0.5
3.313e-04
2.313e-04
4.011e-04
2.507e-05
1.21
yolov10n-esmoe-rewire-e120-s1-p800-best
1.5
6.898e-05
1.059e-04
9.044e-05
5.653e-06
1.31
yolov10n-esmoe-rewire-e120-s2-p800-best
0.15
5.579e-04
2.645e-04
8.033e-04
5.021e-05
1.44
yolov10n-esmoe-rewire-e120-s2-p800-best
0.5
3.353e-04
2.304e-04
4.084e-04
2.553e-05
1.22
yolov10n-esmoe-rewire-e120-s2-p800-best
1.5
7.354e-05
1.096e-04
1.005e-04
6.279e-06
1.37
yolov10n-esmoe-rewire-e120-s3-p800-best
0.15
5.536e-04
2.923e-04
7.592e-04
4.745e-05
1.37
yolov10n-esmoe-rewire-e120-s3-p800-best
0.5
3.648e-04
2.553e-04
4.518e-04
2.824e-05
1.24
yolov10n-esmoe-rewire-e120-s3-p800-best
1.5
7.478e-05
1.172e-04
9.790e-05
6.119e-06
1.31
yolov10n-esmoe-rewire-e120-s4-p800-best
0.15
5.282e-04
1.808e-04
7.291e-04
4.557e-05
1.38
yolov10n-esmoe-rewire-e120-s4-p800-best
0.5
2.455e-04
1.692e-04
2.807e-04
1.754e-05
1.14
yolov10n-esmoe-rewire-e120-s4-p800-best
1.5
7.275e-05
9.505e-05
8.996e-05
5.622e-06
1.24
yolov10n-esmoe-rewire-stages-e120-s0-p800-best
0.15
4.512e-04
1.176e-04
5.311e-04
3.319e-05
1.18
yolov10n-esmoe-rewire-stages-e120-s0-p800-best
0.5
1.499e-04
9.355e-05
1.658e-04
1.036e-05
1.11
yolov10n-esmoe-rewire-stages-e120-s0-p800-best
1.5
3.147e-05
2.984e-05
3.240e-05
2.025e-06
1.03
yolov10n-esmoe-rewire-stages-e120-s1-p800-best
0.15
5.392e-04
1.992e-04
7.450e-04
4.656e-05
1.38
yolov10n-esmoe-rewire-stages-e120-s1-p800-best
0.5
2.719e-04
1.849e-04
3.167e-04
1.979e-05
1.16
yolov10n-esmoe-rewire-stages-e120-s1-p800-best
1.5
7.684e-05
1.033e-04
1.001e-04
6.254e-06
1.30
yolov10n-esmoe-rewire-stages-e120-s2-p800-best
0.15
4.035e-04
7.658e-05
3.494e-04
2.184e-05
0.87
yolov10n-esmoe-rewire-stages-e120-s2-p800-best
0.5
1.199e-04
7.217e-05
1.229e-04
7.682e-06
1.02
yolov10n-esmoe-rewire-stages-e120-s2-p800-best
1.5
4.148e-05
4.749e-05
4.459e-05
2.787e-06
1.07
yolov10n-esmoe-rewire-stages-w0.0025-e120-s0-p800-best
0.15
2.280e-04
4.559e-05
2.365e-04
1.478e-05
1.04
yolov10n-esmoe-rewire-stages-w0.0025-e120-s0-p800-best
0.5
7.811e-05
4.956e-05
7.583e-05
4.740e-06
0.97
yolov10n-esmoe-rewire-stages-w0.0025-e120-s0-p800-best
1.5
2.798e-05
4.222e-05
3.992e-05
2.495e-06
1.43
yolov10n-esmoe-rewire-stages-w0.0025-e120-s1-p800-best
0.15
2.983e-04
5.976e-05
3.356e-04
2.098e-05
1.13
yolov10n-esmoe-rewire-stages-w0.0025-e120-s1-p800-best
0.5
1.020e-04
6.247e-05
1.091e-04
6.820e-06
1.07
yolov10n-esmoe-rewire-stages-w0.0025-e120-s1-p800-best
1.5
3.473e-05
5.044e-05
4.895e-05
3.060e-06
1.41
yolov10n-esmoe-rewire-stages-w0.0025-e120-s2-p800-best
0.15
2.055e-04
3.570e-05
1.928e-04
1.205e-05
0.94
yolov10n-esmoe-rewire-stages-w0.0025-e120-s2-p800-best
0.5
7.987e-05
4.163e-05
8.207e-05
5.129e-06
1.03
yolov10n-esmoe-rewire-stages-w0.0025-e120-s2-p800-best
1.5
2.856e-05
3.875e-05
3.793e-05
2.371e-06
1.33
yolov10n-esmoe-w0-e120-s0-p800-best
0.15
5.338e-04
1.068e-04
3.084e-04
1.928e-05
0.58
yolov10n-esmoe-w0-e120-s0-p800-best
0.5
1.824e-04
1.043e-04
1.583e-04
9.895e-06
0.87
yolov10n-esmoe-w0-e120-s0-p800-best
1.5
4.925e-05
6.875e-05
6.464e-05
4.040e-06
1.31
yolov10n-esmoe-w0-e120-s1-p800-best
0.15
6.437e-04
1.445e-04
4.791e-04
2.995e-05
0.74
yolov10n-esmoe-w0-e120-s1-p800-best
0.5
2.088e-04
1.307e-04
2.297e-04
1.435e-05
1.10
yolov10n-esmoe-w0-e120-s1-p800-best
1.5
3.661e-05
6.270e-05
5.350e-05
3.344e-06
1.46
yolov10n-esmoe-w0-e120-s2-p800-best
0.15
5.729e-04
1.120e-04
3.643e-04
2.277e-05
0.64
yolov10n-esmoe-w0-e120-s2-p800-best
0.5
1.675e-04
9.287e-05
1.738e-04
1.087e-05
1.04
yolov10n-esmoe-w0-e120-s2-p800-best
1.5
4.829e-05
4.420e-05
4.606e-05
2.879e-06
0.95
yolov5n-esmoe-dense-e120-s0-p800h2-best
0.15
5.080e-04
9.264e-05
2.734e-04
1.709e-05
0.54
yolov5n-esmoe-dense-e120-s0-p800h2-best
0.5
1.402e-04
8.282e-05
1.359e-04
8.495e-06
0.97
yolov5n-esmoe-dense-e120-s0-p800h2-best
1.5
2.777e-05
4.762e-05
4.132e-05
2.583e-06
1.49
yolov5n-esmoe-dense-e120-s1-p800h2-best
0.15
5.676e-04
3.146e-04
8.169e-04
5.106e-05
1.44
yolov5n-esmoe-dense-e120-s1-p800h2-best
0.5
3.740e-04
2.728e-04
4.732e-04
2.958e-05
1.27
yolov5n-esmoe-dense-e120-s1-p800h2-best
1.5
9.347e-05
1.285e-04
1.221e-04
7.631e-06
1.31
yolov5n-esmoe-dense-e120-s2-p800h2-best
0.15
6.327e-04
2.286e-04
6.128e-04
3.830e-05
0.97
yolov5n-esmoe-dense-e120-s2-p800h2-best
0.5
3.029e-04
1.988e-04
3.641e-04
2.276e-05
1.20
yolov5n-esmoe-dense-e120-s2-p800h2-best
1.5
8.655e-05
9.582e-05
9.362e-05
5.851e-06
1.08
yolov5n-esmoe-e120-s0-p800-best
0.15
5.417e-04
2.235e-04
7.239e-04
4.524e-05
1.34
yolov5n-esmoe-e120-s0-p800-best
0.5
2.801e-04
1.923e-04
3.333e-04
2.083e-05
1.19
yolov5n-esmoe-e120-s0-p800-best
1.5
7.859e-05
9.568e-05
9.454e-05
5.908e-06
1.20
yolov5n-esmoe-e120-s0-p800h2-best
0.15
5.481e-04
3.061e-04
7.347e-04
4.592e-05
1.34
yolov5n-esmoe-e120-s0-p800h2-best
0.5
3.643e-04
2.668e-04
4.545e-04
2.841e-05
1.25
yolov5n-esmoe-e120-s0-p800h2-best
1.5
1.022e-04
1.308e-04
1.283e-04
8.022e-06
1.26
yolov5n-esmoe-e120-s1-p800-best
0.15
5.468e-04
2.310e-04
7.753e-04
4.846e-05
1.42
yolov5n-esmoe-e120-s1-p800-best
0.5
2.945e-04
1.984e-04
3.521e-04
2.201e-05
1.20
yolov5n-esmoe-e120-s1-p800-best
1.5
8.494e-05
9.752e-05
9.746e-05
6.091e-06
1.15
yolov5n-esmoe-e120-s1-p800h2-best
0.15
6.344e-04
2.817e-04
6.703e-04
4.190e-05
1.06
yolov5n-esmoe-e120-s1-p800h2-best
0.5
3.531e-04
2.457e-04
4.333e-04
2.708e-05
1.23
yolov5n-esmoe-e120-s1-p800h2-best
1.5
7.842e-05
1.168e-04
1.059e-04
6.619e-06
1.35
yolov5n-esmoe-e120-s2-p800-best
0.15
5.418e-04
2.454e-04
7.317e-04
4.573e-05
1.35
yolov5n-esmoe-e120-s2-p800-best
0.5
3.074e-04
2.110e-04
3.740e-04
2.337e-05
1.22
yolov5n-esmoe-e120-s2-p800-best
1.5
9.203e-05
1.061e-04
1.068e-04
6.676e-06
1.16
yolov5n-esmoe-e120-s2-p800h2-best
0.15
5.324e-04
2.937e-04
6.810e-04
4.256e-05
1.28
yolov5n-esmoe-e120-s2-p800h2-best
0.5
3.423e-04
2.544e-04
4.140e-04
2.587e-05
1.21
yolov5n-esmoe-e120-s2-p800h2-best
1.5
1.171e-04
1.321e-04
1.353e-04
8.456e-06
1.16
yolov5n-esmoe-gshard-e120-s2-p800h2g-best
0.15
6.050e-04
1.134e-04
3.053e-04
1.908e-05
0.50
yolov5n-esmoe-gshard-e120-s2-p800h2g-best
0.5
1.694e-04
9.263e-05
1.395e-04
8.716e-06
0.82
yolov5n-esmoe-gshard-e120-s2-p800h2g-best
1.5
3.643e-05
3.321e-05
3.391e-05
2.120e-06
0.93
yolov5n-esmoe-master-e120-s0-p800h2-best
0.15
6.679e-04
1.500e-04
4.199e-04
2.624e-05
0.63
yolov5n-esmoe-master-e120-s0-p800h2-best
0.5
2.327e-04
1.349e-04
2.421e-04
1.513e-05
1.04
yolov5n-esmoe-master-e120-s0-p800h2-best
1.5
6.842e-05
7.664e-05
8.011e-05
5.007e-06
1.17
yolov5n-esmoe-master-e120-s1-p800h2-best
0.15
7.016e-04
1.637e-04
4.139e-04
2.587e-05
0.59
yolov5n-esmoe-master-e120-s1-p800h2-best
0.5
2.500e-04
1.497e-04
2.711e-04
1.694e-05
1.08
yolov5n-esmoe-master-e120-s1-p800h2-best
1.5
7.354e-05
8.312e-05
8.478e-05
5.299e-06
1.15
yolov5n-esmoe-master-w0.32-e120-s0-p800h2-best
0.15
5.559e-04
1.064e-04
3.318e-04
2.074e-05
0.60
yolov5n-esmoe-master-w0.32-e120-s0-p800h2-best
0.5
1.583e-04
8.783e-05
1.578e-04
9.860e-06
1.00
yolov5n-esmoe-master-w0.32-e120-s0-p800h2-best
1.5
4.414e-05
4.037e-05
4.129e-05
2.581e-06
0.94
yolov5n-esmoe-master-w0.32-e120-s1-p800h2-best
0.15
8.001e-04
1.873e-04
1.875e-04
1.172e-05
0.23
yolov5n-esmoe-master-w0.32-e120-s1-p800h2-best
0.5
3.214e-04
1.766e-04
2.569e-04
1.606e-05
0.80
yolov5n-esmoe-master-w0.32-e120-s1-p800h2-best
1.5
8.171e-05
1.053e-04
9.187e-05
5.742e-06
1.12
yolov5n-esmoe-master-w0.32-e120-s2-p800h2-best
0.15
6.125e-04
1.220e-04
2.781e-04
1.738e-05
0.45
yolov5n-esmoe-master-w0.32-e120-s2-p800h2-best
0.5
1.851e-04
1.002e-04
1.459e-04
9.120e-06
0.79
yolov5n-esmoe-master-w0.32-e120-s2-p800h2-best
1.5
4.674e-05
4.085e-05
4.276e-05
2.673e-06
0.91
yolov5n-esmoe-norm-e120-s0-p800h2-best
0.15
5.158e-04
1.737e-04
6.929e-04
4.330e-05
1.34
yolov5n-esmoe-norm-e120-s0-p800h2-best
0.5
2.339e-04
1.533e-04
2.635e-04
1.647e-05
1.13
yolov5n-esmoe-norm-e120-s0-p800h2-best
1.5
4.509e-05
7.430e-05
6.671e-05
4.169e-06
1.48
yolov5n-esmoe-norm-e120-s1-p800h2-best
0.15
5.458e-04
2.565e-04
7.273e-04
4.545e-05
1.33
yolov5n-esmoe-norm-e120-s1-p800h2-best
0.5
3.129e-04
2.202e-04
3.848e-04
2.405e-05
1.23
yolov5n-esmoe-norm-e120-s1-p800h2-best
1.5
9.415e-05
1.110e-04
1.103e-04
6.896e-06
1.17
yolov5n-esmoe-norm-e120-s2-p800h2-best
0.15
5.674e-04
2.905e-04
8.313e-04
5.196e-05
1.47
yolov5n-esmoe-norm-e120-s2-p800h2-best
0.5
3.544e-04
2.490e-04
4.388e-04
2.743e-05
1.24
yolov5n-esmoe-norm-e120-s2-p800h2-best
1.5
1.002e-04
1.172e-04
1.162e-04
7.263e-06
1.16
yolov5n-esmoe-rewire-e120-s0-p800-best
0.15
5.650e-04
2.989e-04
8.055e-04
5.034e-05
1.43
yolov5n-esmoe-rewire-e120-s0-p800-best
0.5
3.590e-04
2.567e-04
4.505e-04
2.815e-05
1.25
yolov5n-esmoe-rewire-e120-s0-p800-best
1.5
9.920e-05
1.198e-04
1.131e-04
7.069e-06
1.14
yolov5n-esmoe-rewire-e120-s0-p800h2-best
0.15
6.090e-04
1.642e-04
5.533e-04
3.458e-05
0.91
yolov5n-esmoe-rewire-e120-s0-p800h2-best
0.5
2.373e-04
1.477e-04
2.683e-04
1.677e-05
1.13
yolov5n-esmoe-rewire-e120-s0-p800h2-best
1.5
4.766e-05
7.744e-05
7.075e-05
4.422e-06
1.48
yolov5n-esmoe-rewire-e120-s1-p800h2-best
0.15
5.825e-04
3.899e-04
8.552e-04
5.345e-05
1.47
yolov5n-esmoe-rewire-e120-s1-p800h2-best
0.5
4.298e-04
3.372e-04
5.687e-04
3.554e-05
1.32
yolov5n-esmoe-rewire-e120-s1-p800h2-best
1.5
1.182e-04
1.553e-04
1.505e-04
9.406e-06
1.27
yolov5n-esmoe-rewire-e120-s2-p800h2-best
0.15
5.060e-04
1.727e-04
6.772e-04
4.232e-05
1.34
yolov5n-esmoe-rewire-e120-s2-p800h2-best
0.5
2.463e-04
1.492e-04
2.773e-04
1.733e-05
1.13
yolov5n-esmoe-rewire-e120-s2-p800h2-best
1.5
7.188e-05
7.413e-05
7.505e-05
4.690e-06
1.04
yolov5n-esmoe-rewire-stages-e120-s0-p800h2-best
0.15
4.750e-04
1.142e-04
5.103e-04
3.190e-05
1.07
yolov5n-esmoe-rewire-stages-e120-s0-p800h2-best
0.5
1.895e-04
1.127e-04
2.095e-04
1.309e-05
1.11
yolov5n-esmoe-rewire-stages-e120-s0-p800h2-best
1.5
6.293e-05
7.036e-05
7.219e-05
4.512e-06
1.15
yolov5n-esmoe-rewire-stages-e120-s1-p800h2-best
0.15
6.742e-04
1.518e-04
4.191e-04
2.619e-05
0.62
yolov5n-esmoe-rewire-stages-e120-s1-p800h2-best
0.5
2.176e-04
1.363e-04
2.310e-04
1.444e-05
1.06
yolov5n-esmoe-rewire-stages-e120-s1-p800h2-best
1.5
3.647e-05
6.426e-05
5.233e-05
3.271e-06
1.43
yolov5n-esmoe-rewire-stages-e120-s2-p800h2-best
0.15
4.789e-04
1.315e-04
6.030e-04
3.769e-05
1.26
yolov5n-esmoe-rewire-stages-e120-s2-p800h2-best
0.5
2.059e-04
1.165e-04
2.300e-04
1.438e-05
1.12
yolov5n-esmoe-rewire-stages-e120-s2-p800h2-best
1.5
4.171e-05
6.745e-05
6.573e-05
4.108e-06
1.58
yolov5n-esmoe-w0-e120-s0-p800h2-best
0.15
7.114e-04
2.289e-04
5.158e-04
3.224e-05
0.73
yolov5n-esmoe-w0-e120-s0-p800h2-best
0.5
3.067e-04
2.036e-04
3.643e-04
2.277e-05
1.19
yolov5n-esmoe-w0-e120-s0-p800h2-best
1.5
6.966e-05
9.684e-05
8.716e-05
5.448e-06
1.25
yolov5n-esmoe-w0-e120-s1-p800h2-best
0.15
7.351e-04
2.464e-04
4.816e-04
3.010e-05
0.66
yolov5n-esmoe-w0-e120-s1-p800h2-best
0.5
3.506e-04
2.207e-04
3.803e-04
2.377e-05
1.08
yolov5n-esmoe-w0-e120-s1-p800h2-best
1.5
9.661e-05
1.224e-04
1.176e-04
7.352e-06
1.22
yolov5n-esmoe-w0-e120-s2-p800h2-best
0.15
6.187e-04
1.171e-04
2.198e-04
1.374e-05
0.36
yolov5n-esmoe-w0-e120-s2-p800h2-best
0.5
1.779e-04
9.623e-05
1.361e-04
8.507e-06
0.76
yolov5n-esmoe-w0-e120-s2-p800h2-best
1.5
4.250e-05
3.733e-05
4.019e-05
2.512e-06
0.95
yolov8n-esmoe-e120-s0-p800-best
0.15
5.494e-04
2.403e-04
7.688e-04
4.805e-05
1.40
yolov8n-esmoe-e120-s0-p800-best
0.5
3.041e-04
2.053e-04
3.679e-04
2.299e-05
1.21
yolov8n-esmoe-e120-s0-p800-best
1.5
9.203e-05
1.027e-04
1.083e-04
6.770e-06
1.18
yolov8n-esmoe-e120-s1-p800-best
0.15
5.928e-04
4.409e-04
8.912e-04
5.570e-05
1.50
yolov8n-esmoe-e120-s1-p800-best
0.5
4.601e-04
3.809e-04
6.253e-04
3.908e-05
1.36
yolov8n-esmoe-e120-s1-p800-best
1.5
1.379e-04
1.721e-04
1.772e-04
1.108e-05
1.29
yolov8n-esmoe-e120-s2-p800-best
0.15
8.122e-04
2.198e-04
1.749e-04
1.093e-05
0.22
yolov8n-esmoe-e120-s2-p800-best
0.5
3.514e-04
1.979e-04
2.835e-04
1.772e-05
0.81
yolov8n-esmoe-e120-s2-p800-best
1.5
8.544e-05
1.086e-04
9.942e-05
6.214e-06
1.16
yolov8n-esmoe-rewire-e120-s0-p800-best
0.15
5.802e-04
3.622e-04
8.566e-04
5.354e-05
1.48
yolov8n-esmoe-rewire-e120-s0-p800-best
0.5
4.012e-04
3.094e-04
5.235e-04
3.272e-05
1.31
yolov8n-esmoe-rewire-e120-s0-p800-best
1.5
1.231e-04
1.439e-04
1.510e-04
9.435e-06
1.23
yolov8n-esmoe-rewire-e120-s1-p800-best
0.15
5.547e-04
2.812e-04
7.610e-04
4.756e-05
1.37
yolov8n-esmoe-rewire-e120-s1-p800-best
0.5
3.481e-04
2.446e-04
4.310e-04
2.694e-05
1.24
yolov8n-esmoe-rewire-e120-s1-p800-best
1.5
8.323e-05
1.193e-04
1.113e-04
6.958e-06
1.34
yolov8n-esmoe-rewire-e120-s2-p800-best
0.15
5.736e-04
3.196e-04
8.468e-04
5.293e-05
1.48
yolov8n-esmoe-rewire-e120-s2-p800-best
0.5
3.736e-04
2.736e-04
4.717e-04
2.948e-05
1.26
yolov8n-esmoe-rewire-e120-s2-p800-best
1.5
1.094e-04
1.275e-04
1.308e-04
8.176e-06
1.20
yolov9t-esmoe-e120-s0-p800-best
0.15
5.663e-04
2.937e-04
8.183e-04
5.114e-05
1.44
yolov9t-esmoe-e120-s0-p800-best
0.5
3.480e-04
2.511e-04
4.363e-04
2.727e-05
1.25
yolov9t-esmoe-e120-s0-p800-best
1.5
1.055e-04
1.203e-04
1.256e-04
7.852e-06
1.19
yolov9t-esmoe-e120-s1-p800-best
0.15
7.803e-04
3.243e-04
3.890e-04
2.431e-05
0.50
yolov9t-esmoe-e120-s1-p800-best
0.5
4.218e-04
2.863e-04
4.281e-04
2.676e-05
1.01
yolov9t-esmoe-e120-s1-p800-best
1.5
1.365e-04
1.556e-04
1.594e-04
9.964e-06
1.17
yolov9t-esmoe-e120-s2-p800-best
0.15
5.381e-04
2.030e-04
7.268e-04
4.543e-05
1.35
yolov9t-esmoe-e120-s2-p800-best
0.5
2.775e-04
1.776e-04
3.239e-04
2.024e-05
1.17
yolov9t-esmoe-e120-s2-p800-best
1.5
5.981e-05
9.188e-05
8.762e-05
5.476e-06
1.46
yolov9t-esmoe-rewire-e120-s0-p800-best
0.15
4.629e-04
1.333e-04
5.424e-04
3.390e-05
1.17
yolov9t-esmoe-rewire-e120-s0-p800-best
0.5
2.051e-04
1.164e-04
2.279e-04
1.425e-05
1.11
yolov9t-esmoe-rewire-e120-s0-p800-best
1.5
7.021e-05
7.181e-05
7.872e-05
4.920e-06
1.12
yolov9t-esmoe-rewire-e120-s1-p800-best
0.15
6.970e-04
2.386e-04
4.853e-04
3.033e-05
0.70
yolov9t-esmoe-rewire-e120-s1-p800-best
0.5
3.539e-04
2.253e-04
3.759e-04
2.349e-05
1.06
yolov9t-esmoe-rewire-e120-s1-p800-best
1.5
1.119e-04
1.327e-04
1.339e-04
8.370e-06
1.20
yolov9t-esmoe-rewire-e120-s2-p800-best
0.15
3.443e-04
7.608e-05
1.606e-04
1.003e-05
0.47
yolov9t-esmoe-rewire-e120-s2-p800-best
0.5
1.426e-04
7.789e-05
1.129e-04
7.059e-06
0.79
yolov9t-esmoe-rewire-e120-s2-p800-best
1.5
4.007e-05
5.607e-05
5.130e-05
3.206e-06
1.28
backbone
checkpoints
gate/switch
yolo-master-n
18
0.36x - 1.45x
yolo11n
18
1.00x - 1.26x
yolo12n
18
0.12x - 1.37x
yolo26n
18
0.67x - 1.37x
yolov10n
69
0.58x - 1.51x
yolov5n
84
0.23x - 1.58x
yolov8n
18
0.22x - 1.50x
yolov9t
18
0.47x - 1.46x
Reading the gate costs 0.12x to 1.58x what the Switch term costs at the same weight, so the two are pushed comparably; the paper's form is 10x to 134x weaker, which is the 1/E^2 factor. The low ratios all sit at the tightest spread, where a router whose mean is already skewed gives the Switch term more to pull on.
Release audit
Each row below is derived from the records, the git history and the published artifacts by scripts/closure.py, and can be re-checked by running it.
Generated by scripts/closure.py; every row is derived from this repository.
criterion
met
evidence
provenance
yes
142 protocol records carry revision, environment and config hash
controls
yes
100 arm runs paired against a same-seed baseline on the same card
statistics
yes
31 compared cells, 29 at three seeds or more; below three: yolov10n-e4k2w0.01-norm@e120f1i800[metaxc500/metax3.3], yolov5n-e4k2w0.01-rewire@e120f1i800[metaxc500/metax3.7]; declared on the experiments page
negative results
yes
11 registrations and 3 published corrections over 32 commits
upstream fix
yes
#241 merged, sections: Problem, Fix, Tests
secrets
yes
no credential pattern in tracked files
aux loss reaches the optimiser
yes
94 runs optimise it; 10 checks in results/verify.json
deliverable
present
evidence
plugin package
yes
esmoe on PyPI, importable, no fork of ultralytics
compatibility
yes
8 backbones measured: yolo-master-n, yolo11n, yolo12n, yolo26n, yolov10n, yolov5n, yolov8n, yolov9t; CI workflows: canary.yml, ci.yml, docs.yml, release.yml
multi-seed ablation
yes
142 protocol runs in results/summary.md
README / Colab
yes
README.md and notebooks/quickstart.ipynb
release notes
yes
docs/RELEASE.md, both languages
0 unmet .