← Run Index
Run summary · Phase 5
lego-iter20000
E.5 iter scaling: lego sh3 20000 iter (収益逓減 plateau 探索)
partial pass
01
設定
- dataset
- /Users/otkrickey/dev/3dgs-workspace/datasets/nerf_synthetic/lego
- iterations
- 20,000
- seed
- 42
- capacity
- 1,000,000 splats
- sh_degree
- 3
- loss
- L1Ssim
- lambda
- 0.200
- ssim
- window=7 sigma=1
- backend.backward
- Simd
- backend.loss_path
- Gpu
02
結果
Val PSNR
23.87dB
Wallclock
14m 8s
Final splats
80,716
Final loss
2.558136e-2
03
描画結果
学習後の eval pipeline から取得した代表視点。
00
01
0204
3D ビューワー
学習結果 PLY を .splat 形式に変換し、antimatter15/splat の WebGL ビューワーで対話表示。
05
推移
Loss curve — 反復ごとの training loss
Splat count — refine の発火と最終構成
06
中間ログ抜粋
| iter | 本 run loss |
|---|---|
| 1 | 3.4534e-1 |
| 500 | 1.1786e-1 |
| 1,000 | 9.8537e-2 |
| 1,500 | 7.9528e-2 |
| 2,000 | 5.5767e-2 |
| 2,500 | 4.5864e-2 |
| 3,000 | 3.9252e-2 |
| 3,500 | 3.8971e-2 |
| 4,000 | 3.4057e-2 |
| 4,500 | 3.4134e-2 |
| 5,000 | 3.1702e-2 |
| 5,500 | 3.2929e-2 |
| 6,000 | 3.0439e-2 |
| 6,500 | 2.9743e-2 |
| 7,000 | 2.8987e-2 |
| 7,500 | 2.9425e-2 |
| 8,000 | 3.0451e-2 |
| 8,500 | 2.8595e-2 |
| 9,000 | 2.7865e-2 |
| 9,500 | 2.7708e-2 |
| 10,000 | 2.7923e-2 |
| 10,500 | 2.7364e-2 |
| 11,000 | 2.6719e-2 |
| 11,500 | 2.7808e-2 |
| 12,000 | 2.7561e-2 |
| 12,500 | 2.6666e-2 |
| 13,000 | 2.7636e-2 |
| 13,500 | 2.6241e-2 |
| 14,000 | 2.6243e-2 |
| 14,500 | 2.8441e-2 |
| 15,000 | 2.6402e-2 |
| 15,500 | 2.6122e-2 |
| 16,000 | 2.7634e-2 |
| 16,500 | 2.5374e-2 |
| 17,000 | 2.7042e-2 |
| 17,500 | 2.6050e-2 |
| 18,000 | 2.6299e-2 |
| 18,500 | 2.6489e-2 |
| 19,000 | 2.5045e-2 |
| 19,500 | 2.6305e-2 |
| 20,000 | 2.5581e-2 |
07
関連 finding
この run を related_runs に含む finding (Positive / Negative どちらも)。