geolip-cv-experiments / colab_cv_loss_sweep_sequential_outputs.txt
AbstractPhil's picture
Rename colab_loss_sweep_sequential_outputs.txt to colab_cv_loss_sweep_sequential_outputs.txt
01e6910 verified
================================================================================
CV LOSS SWEEP β€” PURE NOISE PREDICTION
Random inputs β†’ MLP β†’ S^(d-1) β†’ constellation β†’ 10 random labels
No data structure. No signal. Pure sphere geometry + optimizer.
200 steps per run, ~2s each.
================================================================================
Running 43 configurations, 200 steps each
Estimated time: ~129s
[ 1/43] no_cv w=0.000 t=0.00 d=128 β†’ CV=0.2319 dim=47 acc=76% (2.9s)
[ 2/43] no_cv_s2 w=0.000 t=0.00 d=128 β†’ CV=0.2288 dim=47 acc=76% (2.7s)
[ 3/43] no_cv_s3 w=0.000 t=0.00 d=128 β†’ CV=0.2532 dim=46 acc=73% (2.8s)
[ 4/43] no_cv_s4 w=0.000 t=0.00 d=128 β†’ CV=0.2468 dim=46 acc=74% (2.7s)
[ 5/43] no_cv_s5 w=0.000 t=0.00 d=128 β†’ CV=0.2174 dim=47 acc=76% (2.9s)
[ 6/43] w0.001_t0.22 w=0.001 t=0.22 d=128 β†’ CV=0.2318 dim=47 acc=76% (12.6s)
[ 7/43] w0.01_t0.22 w=0.010 t=0.22 d=128 β†’ CV=0.2325 dim=47 acc=76% (12.5s)
[ 8/43] w0.1_t0.22 w=0.100 t=0.22 d=128 β†’ CV=0.2357 dim=48 acc=76% (12.6s)
[ 9/43] w0.5_t0.22 w=0.500 t=0.22 d=128 β†’ CV=0.2255 dim=51 acc=72% (12.5s)
[10/43] w1.0_t0.22 w=1.000 t=0.22 d=128 β†’ CV=0.2310 dim=52 acc=66% (12.6s)
[11/43] w5.0_t0.22 w=5.000 t=0.22 d=128 β†’ CV=0.2145 dim=51 acc=27% (12.5s)
[12/43] w10_t0.22 w=10.000 t=0.22 d=128 β†’ CV=0.2388 dim=47 acc=20% (12.5s)
[13/43] w50_t0.22 w=50.000 t=0.22 d=128 β†’ CV=0.2336 dim=45 acc=15% (12.6s)
[14/43] w100_t0.22 w=100.000 t=0.22 d=128 β†’ CV=0.2301 dim=44 acc=13% (12.7s)
[15/43] w0.01_t0.00 w=0.010 t=0.00 d=128 β†’ CV=0.2272 dim=48 acc=76% (12.5s)
[16/43] w0.01_t0.05 w=0.010 t=0.05 d=128 β†’ CV=0.2282 dim=48 acc=76% (12.6s)
[17/43] w0.01_t0.10 w=0.010 t=0.10 d=128 β†’ CV=0.2294 dim=48 acc=76% (12.5s)
[18/43] w0.01_t0.30 w=0.010 t=0.30 d=128 β†’ CV=0.2348 dim=47 acc=76% (12.7s)
[19/43] w0.01_t0.50 w=0.010 t=0.50 d=128 β†’ CV=0.2395 dim=46 acc=76% (12.5s)
[20/43] w0.01_t0.80 w=0.010 t=0.80 d=128 β†’ CV=0.2439 dim=45 acc=74% (12.6s)
[21/43] w0.01_t1.00 w=0.010 t=1.00 d=128 β†’ CV=0.2423 dim=44 acc=73% (12.7s)
[22/43] w0.01_t2.00 w=0.010 t=2.00 d=128 β†’ CV=0.6056 dim=32 acc=55% (13.0s)
[23/43] w1_t0.00 w=1.000 t=0.00 d=128 β†’ CV=0.1358 dim=72 acc=50% (12.8s)
[24/43] w1_t0.05 w=1.000 t=0.05 d=128 β†’ CV=0.1499 dim=68 acc=62% (12.5s)
[25/43] w1_t0.50 w=1.000 t=0.50 d=128 β†’ CV=0.4153 dim=44 acc=27% (12.6s)
[26/43] w1_t0.80 w=1.000 t=0.80 d=128 β†’ CV=0.8266 dim=33 acc=17% (12.7s)
[27/43] w1_t1.00 w=1.000 t=1.00 d=128 β†’ CV=1.0157 dim=20 acc=13% (12.6s)
[28/43] w100_t0.00 w=100.000 t=0.00 d=128 β†’ CV=0.1179 dim=70 acc=15% (12.6s)
[29/43] w100_t0.05 w=100.000 t=0.05 d=128 β†’ CV=0.1176 dim=70 acc=15% (12.5s)
[30/43] w100_t0.10 w=100.000 t=0.10 d=128 β†’ CV=0.1217 dim=69 acc=16% (12.6s)
[31/43] w100_t0.50 w=100.000 t=0.50 d=128 β†’ CV=0.3896 dim=32 acc=12% (12.7s)
[32/43] w100_t0.80 w=100.000 t=0.80 d=128 β†’ CV=0.7022 dim=23 acc=11% (12.5s)
[33/43] w100_t1.00 w=100.000 t=1.00 d=128 β†’ CV=1.0147 dim=18 acc=12% (12.6s)
[34/43] pure_cv_t0.22 w=1.000 t=0.22 d=128 β†’ CV=0.2168 dim=47 acc=11% (11.8s)
[35/43] pure_cv_t0.05 w=1.000 t=0.05 d=128 β†’ CV=0.1050 dim=74 acc=10% (11.8s)
[36/43] pure_cv_t0.50 w=1.000 t=0.50 d=128 β†’ CV=0.4881 dim=38 acc=10% (11.7s)
[37/43] pure_cv_t0.80 w=1.000 t=0.80 d=128 β†’ CV=0.8275 dim=29 acc=10% (11.7s)
[38/43] pure_cv_t1.00 w=1.000 t=1.00 d=128 β†’ CV=1.0491 dim=23 acc=11% (12.0s)
[39/43] dim16 w=0.000 t=0.00 d=16 β†’ CV=0.3753 dim=11 acc=19% (2.7s)
[40/43] dim32 w=0.000 t=0.00 d=32 β†’ CV=0.2839 dim=20 acc=28% (2.7s)
[41/43] dim64 w=0.000 t=0.00 d=64 β†’ CV=0.2491 dim=33 acc=48% (2.7s)
[42/43] dim256 w=0.000 t=0.00 d=256 β†’ CV=0.2905 dim=50 acc=97% (2.9s)
[43/43] dim512 w=0.000 t=0.00 d=512 β†’ CV=0.3334 dim=41 acc=98% (3.2s)
==========================================================================================
LABEL CV_W CV_T DIM FINAL_CV EFF_DIM ACC% CE
──────────────────────────────────────────────────────────────────────────────────────────
no_cv 0.000 0.00 128 0.2319βœ“ 47 76% 0.7452
no_cv_s2 0.000 0.00 128 0.2288βœ“ 47 76% 0.7686
no_cv_s3 0.000 0.00 128 0.2532~ 46 73% 0.8086
no_cv_s4 0.000 0.00 128 0.2468~ 46 74% 0.7998
no_cv_s5 0.000 0.00 128 0.2174βœ“ 47 76% 0.7461
w0.001_t0.22 0.001 0.22 128 0.2318βœ“ 47 76% 0.7451
w0.01_t0.22 0.010 0.22 128 0.2325βœ“ 47 76% 0.7469
w0.1_t0.22 0.100 0.22 128 0.2357βœ“ 48 76% 0.7605
w0.5_t0.22 0.500 0.22 128 0.2255βœ“ 51 72% 0.8654
w1.0_t0.22 1.000 0.22 128 0.2310βœ“ 52 66% 1.0733
w5.0_t0.22 5.000 0.22 128 0.2145βœ“ 51 27% 2.0665
w10_t0.22 10.000 0.22 128 0.2388βœ“ 47 20% 2.1946
w50_t0.22 50.000 0.22 128 0.2336βœ“ 45 15% 2.2746
w100_t0.22 100.000 0.22 128 0.2301βœ“ 44 13% 2.2866
w0.01_t0.00 0.010 0.00 128 0.2272βœ“ 48 76% 0.7389
w0.01_t0.05 0.010 0.05 128 0.2282βœ“ 48 76% 0.7401
w0.01_t0.10 0.010 0.10 128 0.2294βœ“ 48 76% 0.7417
w0.01_t0.30 0.010 0.30 128 0.2348βœ“ 47 76% 0.7505
w0.01_t0.50 0.010 0.50 128 0.2395βœ“ 46 76% 0.7624
w0.01_t0.80 0.010 0.80 128 0.2439~ 45 74% 0.7892
w0.01_t1.00 0.010 1.00 128 0.2423~ 44 73% 0.8181
w0.01_t2.00 0.010 2.00 128 0.6056βœ— 32 55% 1.3255
w1_t0.00 1.000 0.00 128 0.1358βœ— 72 50% 1.4669
w1_t0.05 1.000 0.05 128 0.1499βœ— 68 62% 1.1861
w1_t0.50 1.000 0.50 128 0.4153βœ— 44 27% 2.0607
w1_t0.80 1.000 0.80 128 0.8266βœ— 33 17% 2.2362
w1_t1.00 1.000 1.00 128 1.0157βœ— 20 13% 2.2882
w100_t0.00 100.000 0.00 128 0.1179βœ— 70 15% 2.2663
w100_t0.05 100.000 0.05 128 0.1176βœ— 70 15% 2.2639
w100_t0.10 100.000 0.10 128 0.1217βœ— 69 16% 2.2621
w100_t0.50 100.000 0.50 128 0.3896βœ— 32 12% 2.2983
w100_t0.80 100.000 0.80 128 0.7022βœ— 23 11% 2.2999
w100_t1.00 100.000 1.00 128 1.0147βœ— 18 12% 2.3009
pure_cv_t0.22 1.000 0.22 128 0.2168βœ“ 47 11% 2.3025
pure_cv_t0.05 1.000 0.05 128 0.1050βœ— 74 10% 2.3036
pure_cv_t0.50 1.000 0.50 128 0.4881βœ— 38 10% 2.3036
pure_cv_t0.80 1.000 0.80 128 0.8275βœ— 29 10% 2.3036
pure_cv_t1.00 1.000 1.00 128 1.0491βœ— 23 11% 2.3037
dim16 0.000 0.00 16 0.3753βœ— 11 19% 2.2147
dim32 0.000 0.00 32 0.2839βœ— 20 28% 2.0190
dim64 0.000 0.00 64 0.2491~ 33 48% 1.5276
dim256 0.000 0.00 256 0.2905βœ— 50 97% 0.1247
dim512 0.000 0.00 512 0.3334βœ— 41 98% 0.0645
==========================================================================================
ANALYSIS
==========================================================================================
[1] NO CV LOSS, PURE NOISE (d=128, 5 seeds):
CV: mean=0.2356 min=0.2174 max=0.2532 spread=0.0358
Dim: mean=46.6
Within [0.17, 0.24]: 3/5
[2] WEIGHT SWEEP (target=0.22, d=128):
w= 0.001 β†’ CV=0.2318 acc=76%
w= 0.010 β†’ CV=0.2325 acc=76%
w= 0.100 β†’ CV=0.2357 acc=76%
w= 0.500 β†’ CV=0.2255 acc=72%
w= 1.000 β†’ CV=0.2310 acc=66%
w= 5.000 β†’ CV=0.2145 acc=27%
w= 10.000 β†’ CV=0.2388 acc=20%
w= 50.000 β†’ CV=0.2336 acc=15%
w= 100.000 β†’ CV=0.2301 acc=13%
[3] TARGET SWEEP (w=0.01, d=128):
target=0.00 β†’ CV=0.2272βœ“ acc=76%
target=0.05 β†’ CV=0.2282βœ“ acc=76%
target=0.10 β†’ CV=0.2294βœ“ acc=76%
target=0.22 β†’ CV=0.2325βœ“ acc=76%
target=0.30 β†’ CV=0.2348βœ“ acc=76%
target=0.50 β†’ CV=0.2395βœ“ acc=76%
target=0.80 β†’ CV=0.2439βœ— acc=74%
target=1.00 β†’ CV=0.2423βœ— acc=73%
target=2.00 β†’ CV=0.6056βœ— acc=55%
[3] TARGET SWEEP (w=1.0, d=128):
target=0.00 β†’ CV=0.1358βœ— acc=50%
target=0.05 β†’ CV=0.1499βœ— acc=62%
target=0.22 β†’ CV=0.2310βœ“ acc=66%
target=0.50 β†’ CV=0.4153βœ— acc=27%
target=0.80 β†’ CV=0.8266βœ— acc=17%
target=1.00 β†’ CV=1.0157βœ— acc=13%
[3] TARGET SWEEP (w=100.0, d=128):
target=0.00 β†’ CV=0.1179βœ— acc=15%
target=0.05 β†’ CV=0.1176βœ— acc=15%
target=0.10 β†’ CV=0.1217βœ— acc=16%
target=0.22 β†’ CV=0.2301βœ“ acc=13%
target=0.50 β†’ CV=0.3896βœ— acc=12%
target=0.80 β†’ CV=0.7022βœ— acc=11%
target=1.00 β†’ CV=1.0147βœ— acc=12%
[4] DIMENSION SWEEP (no CV loss):
d= 16 β†’ CV=0.3753 eff_dim=11
d= 32 β†’ CV=0.2839 eff_dim=20
d= 64 β†’ CV=0.2491 eff_dim=33
d= 256 β†’ CV=0.2905 eff_dim=50
d= 512 β†’ CV=0.3334 eff_dim=41
[5] EXTREME FORCE (wβ‰₯100, d=128):
target=0.00 β†’ CV=0.1179 (Ξ” from 0.20: 0.0821) acc=15%
target=0.05 β†’ CV=0.1176 (Ξ” from 0.20: 0.0824) acc=15%
target=0.10 β†’ CV=0.1217 (Ξ” from 0.20: 0.0783) acc=16%
target=0.22 β†’ CV=0.2301 (Ξ” from 0.20: 0.0301) acc=13%
target=0.50 β†’ CV=0.3896 (Ξ” from 0.20: 0.1896) acc=12%
target=0.80 β†’ CV=0.7022 (Ξ” from 0.20: 0.5022) acc=11%
target=1.00 β†’ CV=1.0147 (Ξ” from 0.20: 0.8147) acc=12%
[6] CV TRAJECTORIES (step 0 β†’ step 200):
no_cv : 0.1379 β†’ 0.2519 (Ξ”=+0.1140)
no_cv_s2 : 0.1383 β†’ 0.2360 (Ξ”=+0.0977)
no_cv_s3 : 0.1465 β†’ 0.2392 (Ξ”=+0.0927)
no_cv_s4 : 0.1404 β†’ 0.2460 (Ξ”=+0.1056)
no_cv_s5 : 0.1482 β†’ 0.2344 (Ξ”=+0.0862)
Raw results saved to cv_sweep_results.json
================================================================================
CV SWEEP COMPLETE
================================================================================