🦺 safety_rope v20260518
DINOv3-S + RoIAlign HD 1280×720 + cvat #10 force-acceptance (260 task, +17 vs v514)
結果摘要:ep11 early stop @ best ep3, val_AP=0.932 / test_AP=0.902 / F1=0.846
對標 baseline:v20260514 baseline val_AP=0.957 @ ep10/18
📊 結果
| 指標 | v20260514 baseline | v20260518 best | Δ |
| val_AP | 0.957 (ep10/18) | 0.932 (ep3/11) | −2.5pp |
| test_AP | — | 0.902 | — |
| F1 | — | 0.846 | — |
| P / R | — | 0.814 / 0.881 | — |
| TP / FP / FN / TN | — | 1728 / 395 / 233 / 2883 | — |
📦 訓練 stack
- Backbone:DINOv3-S (vit_small_patch16_dinov3) 22.47M params
- Stack:RoIAlign HD 1280×720 + 外擴 1.0/0.2/1.5 + photometric + random_erase + camaug
- Hyperparams:30 ep / batch=8 / lr=1e-4 / class_weights wrong=1.5 correct=1.0
- Dataset:cvat2 #10 + #8: 28,531 rows (vs baseline 16,170, +12k)
- 跳過:11 個 SIEMENS_*_20260427 待標記師 review
📝 觀察
退步推測:新增 17 task 中 8 SIEMENS + 9 JUJIA,distribution shift 比 v20260514 更大。Early stop ep11 表示 model 沒抓到新場域。