STITCH_TYPE=aggregator vs dino 두 smoke 결과(job 31199/31200) 비교해 stitch 방식 확정.
--qos=own으로 재제출해 5K step 실 학습 확보.
infer_long.sh의 stale --lora_path 교정 필요.
EgoW/VGGRPO
5·L_cam + 1·L_depth (conf-weighted depth + 4-scale gradient + 9-DoF camera L1). VGGT 원본 대비 충실도 비교표 포함 —
multi-stage pose refinement, filter_by_quantile, point loss는 VGGT-Ω 공개 코드 구조상 생략.
| Job ID | 실행 명령 | 상태 | GPU | 실행 시간 | QOS |
|---|---|---|---|---|---|
| 30920 | bash scripts/find_stitch_layer_vggt_dinov2.sh | COMPLETED | 1 | 00:01:27 | own |
| 30995 | bash scripts/find_stitch_layer_vggt_dinov2.sh | COMPLETED | 1 | 00:00:00 | own |
| 31005 | bash scripts/find_stitch_layer_vggt_dinov2.sh | COMPLETED | 1 | 00:34:44 | own |
| 31193 | MAX_STEPS=5000 STITCH_TYPE=aggregator bash scripts/train_lgm_robocasa_vggt_v0.sh | CANCELLED by 10010 | 4 | 00:00:00 | extra |
| 31194 | MAX_STEPS=5000 STITCH_TYPE=dino bash scripts/train_lgm_robocasa_vggt_v0.sh | CANCELLED by 10010 | 4 | 00:00:00 | extra |
| 31199 | MAX_STEPS=5000 STITCH_TYPE=aggregator bash scripts/train_lgm_robocasa_vggt_v0.sh | COMPLETED | 4 | 00:05:21 | own |
| 31200 | MAX_STEPS=5000 STITCH_TYPE=dino bash scripts/train_lgm_robocasa_vggt_v0.sh | COMPLETED | 4 | 00:09:17 | own |
| Job ID | 프로젝트 | 실행 명령 | 상태 | GPU | 실행 시간 | 노드 |
|---|---|---|---|---|---|---|
| 30920 | VGGRPO | find_stitch_layer_vggt_dinov2.sh | COMPLETED | 1 | 00:01:27 | 1 |
| 30995 | VGGRPO | find_stitch_layer_vggt_dinov2.sh | COMPLETED | 1 | 00:00:00 | 1 |
| 31005 | VGGRPO | find_stitch_layer_vggt_dinov2.sh | COMPLETED | 1 | 00:34:44 | 1 |
| 31193 | VGGRPO | train_lgm_robocasa_vggt_v0.sh (STITCH_TYPE=aggregator) | CANCELLED by 10010 | 4 | 00:00:00 | 1 |
| 31194 | VGGRPO | train_lgm_robocasa_vggt_v0.sh (STITCH_TYPE=dino) | CANCELLED by 10010 | 4 | 00:00:00 | 1 |
| 31198 | FastWAM (외부 프로젝트) |
robocasa_bs1024 — train_zero1.sh 8 task=robocasa_uncond_7d_1e-4_local | RUNNING (15h+) | 8 | 15:03:47+ | 1 |
| 31199 | VGGRPO | train_lgm_robocasa_vggt_v0.sh (STITCH_TYPE=aggregator) | COMPLETED | 4 | 00:05:21 | 1 |
| 31200 | VGGRPO | train_lgm_robocasa_vggt_v0.sh (STITCH_TYPE=dino) | COMPLETED | 4 | 00:09:17 | 1 |
-J robocasa_bs1024로 명명된 FastWAM 로보카사 학습 — 본 리포트 대상 6개 프로젝트 외이므로 카드/우선순위에서 제외.