File size: 2,807 Bytes
c0a9e95 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 | =================================================================
BASE TIER PATCHWORK SOUP
3 experts × 768-d → 256 anchors × 128-d
Device: cuda
=================================================================
=================================================================
LOADING DATA
=================================================================
Train: 118,287 Val: 5,000
clip_l14_openai loaded
dinov2_b14 loaded
siglip_b16_384 loaded
=================================================================
BUILDING MODEL
=================================================================
Parameters:
projectors : 296,064
constellation : 32,768
patchwork : 100,864
classifier : 370,256
total : 799,952
=================================================================
TRAINING
20 epochs, lr=0.001, batch=128
Adam, no weight decay (geometry IS the regularization)
=================================================================
E 1: mAP=0.732 F1=0.609 cv=0.4301 agree=[0.977,0.986,0.974] loss=0.0845 ★
E 2: mAP=0.768 F1=0.679 cv=0.4308 agree=[0.982,0.988,0.980] loss=0.0521 ★
E 3: mAP=0.783 F1=0.726 cv=0.4158 agree=[0.984,0.988,0.981] loss=0.0485 ★
E 4: mAP=0.787 F1=0.714 cv=0.3889 agree=[0.986,0.989,0.982] loss=0.0465 ★
E 5: mAP=0.799 F1=0.742 cv=0.3760 agree=[0.987,0.989,0.983] loss=0.0453 ★
E 6: mAP=0.805 F1=0.738 cv=0.3849 agree=[0.987,0.989,0.983] loss=0.0444 ★
E 7: mAP=0.804 F1=0.737 cv=0.4140 agree=[0.988,0.989,0.984] loss=0.0437
E 8: mAP=0.806 F1=0.748 cv=0.3945 agree=[0.989,0.989,0.984] loss=0.0432 ★
E 9: mAP=0.807 F1=0.735 cv=0.3855 agree=[0.988,0.989,0.984] loss=0.0427 ★
E10: mAP=0.809 F1=0.751 cv=0.3396 agree=[0.989,0.989,0.985] loss=0.0423 ★
E11: mAP=0.815 F1=0.746 cv=0.4014 agree=[0.989,0.989,0.985] loss=0.0417 ★
E12: mAP=0.812 F1=0.729 cv=0.3759 agree=[0.989,0.989,0.985] loss=0.0414
E13: mAP=0.818 F1=0.750 cv=0.3313 agree=[0.990,0.990,0.986] loss=0.0411 ★
E14: mAP=0.817 F1=0.742 cv=0.3890 agree=[0.990,0.989,0.985] loss=0.0408
E15: mAP=0.818 F1=0.749 cv=0.3872 agree=[0.990,0.989,0.985] loss=0.0405 ★
E16: mAP=0.816 F1=0.746 cv=0.3667 agree=[0.990,0.990,0.986] loss=0.0402
E17: mAP=0.824 F1=0.753 cv=0.3517 agree=[0.990,0.989,0.985] loss=0.0399 ★
E18: mAP=0.818 F1=0.747 cv=0.3221 agree=[0.990,0.990,0.986] loss=0.0397
E19: mAP=0.825 F1=0.746 cv=0.3117 agree=[0.990,0.989,0.985] loss=0.0396 ★
E20: mAP=0.816 F1=0.738 cv=0.3320 agree=[0.990,0.990,0.986] loss=0.0392
Best mAP: 0.825
Model: 799,952 params
Anchors: 256 × 128-d
=================================================================
DONE
================================================================= |