YOLOv8n Training Report — Phase 3 Ablation
Synthetic LEGO object detection · Scene diversity vs. per-scene augmentation at fixed 100-image budget.
Source: outputs/phase3/results.json ·
Split review
Ablation Results — mAP@50
| Subset | Fold 0 | Fold 1 | Fold 2 | Fold 3 | Fold 4 | Mean | Std |
|---|---|---|---|---|---|---|---|
| 2-view (84 img) | 0.279 | — | — | — | — | 0.279 | 0.000 |
| 5-view (100 img) | 0.343 | — | — | — | — | 0.343 | 0.000 |
| 10-view (100 img) | 0.312 | — | — | — | — | 0.312 | 0.000 |
Best mean highlighted in green. Fold rows use per-val-fold mAP50 from CV runs.
Training Results — Charts
Bar chart not available.
Per-class AP@50
Final runs when available; falls back to best CV fold. Sorted by mean AP.
Cost & Timing — All Runs
| Run | GPU | Compute (s) | Wall (s) | Epochs | mAP50 | mAP50:95 | Cost (USD) | Status |
|---|---|---|---|---|---|---|---|---|
| fold0_10view | T4 | 194 | 198 | 100 | 0.312 | 0.271 | $0.0318 | ✓ ok |
| fold0_2view | T4 | 168 | 176 | 100 | 0.279 | 0.244 | $0.0275 | ✓ ok |
| fold0_5view | T4 | 203 | 211 | 100 | 0.343 | 0.307 | $0.0333 | ✓ ok |
| final_10view | T4 | 195 | 200 | 100 | 0.256 | 0.228 | $0.0320 | ✓ ok |
| final_2view | T4 | 178 | 182 | 100 | 0.314 | 0.280 | $0.0291 | ✓ ok |
| final_5view | T4 | 202 | 206 | 100 | 0.331 | 0.300 | $0.0330 | ✓ ok |
| Total | $0.1867 | |||||||
T4 rate: $0.59/hr = $0.000164/s. Cost = modal_compute_s × rate.
Dataset Split Summary
50
Total scenes
42
Train pool
8
Val hold-out
80
Val images
44
Classes (nc)
5
CV folds
Subset image budgets
| Subset | Views/scene | Scenes | Images |
|---|---|---|---|
| 2view | 2 | 42 | 84 |
| 5view | 5 | 20 | 100 |
| 10view | 10 | 10 | 100 |
Val hold-out: scenes 042–049 (fixed, never in training).