메뉴
Use when configuring or replicating LT-DETR v2 (ltdetrv2-s/m/l/x) COCO benchmarks via the lightly-train wrapper — model alias map, global vs per-rank batch size, recipe invariants, ECDet config mapping, cluster pins, and common gotchas.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Use when opening a PR and driving it to green CI — push the branch, write a high-level PR description, watch checks, fix failures, and keep looping until green; also use when an already-open PR gets new commits.
Use when the user starts, resumes, switches, saves, or recalls a named task or ongoing work across sessions (e.g. "remember this", "continue X", "what was I doing on Y", "track this task").
Git commit workflow. Load when finishing any code-change task (after verification) to commit and push, or when explicitly staging/committing.
Use when modifying or adding a transform, loss, model, method, package, task-model, config, docs page, or example in a Lightly AG repo (lightly-ssl or lightly-train) and you need repo-specific conventions.
Spawn a subagent to run a task. Use when you want to delegate work to a separate pi instance.
Use when the user starts, resumes, switches, saves, or recalls a named task or ongoing work across sessions (e.g. "remember this", "continue X", "what was I doing on Y", "track this task").
| name | ltdetr-coco-recipe |
| description | Use when configuring or replicating LT-DETR v2 (ltdetrv2-s/m/l/x) COCO benchmarks via the lightly-train wrapper — model alias map, global vs per-rank batch size, recipe invariants, ECDet config mapping, cluster pins, and common gotchas. |
872e596+)| Alias (canonical) | Underlying model_name | Arch class |
|---|---|---|
ltdetrv2-s (or edgecrafter/ecvitt-ltdetr) | edgecrafter/ecvitt | ViTTiny / DFINE |
ltdetrv2-m (or edgecrafter/ecvittplus-ltdetr) | edgecrafter/ecvittplus | ViTTinyPlus / DFINE |
ltdetrv2-l (or edgecrafter/ecvits-ltdetr) | edgecrafter/ecvits | ViTTinyPlus / DFINE |
ltdetrv2-x (or edgecrafter/ecvitsplus-ltdetr) | edgecrafter/ecvitsplus | ViTTinyPlus / DFINE |
The m/l/x all share the same HybridEncoder / DFINE Transformer / ECViT-backbone-wrapper class — only the backbone weights file differs. The ltdetrv2-{s,m,l,x}-coco aliases (and edgecrafter/ecvitt-ltdetr-coco) auto-download a COCO-pretrained checkpoint; the bare aliases do not.
lightly-train's --batch-size is the GLOBAL (summed over DDP ranks) batch size, NOT per-rank. With --ntasks-per-node=2 each rank gets 16 imgs/step. STEPS_PER_EPOCH = train_imgs / batch_size (118287 / 32 = 3696 for COCO train2017). LR (e.g. 5e-4) is tuned for the global effective bs.
Recipe invariant: if you must reduce per-rank memory, drop augs (--no-copyblend first, then --no-mosaic, then --no-mixup) BEFORE touching --batch-size. The local train.py wrapper does not currently expose --accumulate-grad-batches / gradient_accumulation_steps; MLflow logs show gradient_accumulation_steps=1 for these runs. Halving --batch-size without accumulation silently changes the effective optimizer batch and invalidates the mAP comparison. If the wrapper is extended to expose accumulation, a possible memory-preserving retry is global microbatch 16 + accumulate 2, but verify the resolved config first: effective optimizer batch must remain 32, and step/epoch-derived aug, validation, checkpoint, and scheduler boundaries must still match the intended recipe.
| Setting | ltdetrv2-s (Tiny, ECDet-S) | ltdetrv2-m (Tiny+, ECDet-M) | ltdetrv2-l (Small, ECDet-L) |
|---|---|---|---|
| Epochs | 74 | 62 | 50 |
| Total steps (bs=32, 3696/ep) | 273504 | 229152 | 184800 |
| LR | 5e-4 | 5e-4 | 5e-4 |
| Backbone LR factor | 0.05 | 0.05 | 0.01 |
| Weight decay | 1e-4 | 1e-4 | 1.25e-4 |
| Mixup + mosaic prob | 0.75 | 0.75 | 1.0 |
| Mosaic stop epoch | 36 | 30 | 24 |
| Strong-aug stop epoch | 72 | 60 | 48 |
For ltdetrv2-x (Small+): no ECDet-X reference exists; mirror ltdetrv2-l (same decoder arch class, only backbone weights differ). Revisit if a separate X reference appears.
AUG_START_STEPS = 15000 (≈ epoch 4). Photometric / random zoom-out / random IoU crop, mixup, mosaic, copyblend all turn on at step 15000. Resolved LR schedule (from .err config-dump): warmup 8ep / flat 38ep / no-aug 10ep / cosine; LR target reached at epoch 8. --no-scale-jitter matches the ECDet recipe (best-6x uses scale-jitter instead).
sinfo -p debug: TIMELIMIT=infinite, gpu:4 capacity. Walltime is NOT a job killer.sacct is disabled. Detect job exit via /logs/.../*.err + squeue job absence, NOT sacct./home/lightly/lt_detr_jobs/ is per-job scratch — NOT reliably visible from login context after the job ends. Persist run metadata (mlflow_run_id.txt, log tails) to a stable path before the job exits.ltdetrv2-{m,l,x} jobs (each --gres=gpu:2) fit concurrently on the 4-GPU node. Submit in order; Slurm schedules PENDING→RUNNING back-to-back.TORCH_NCCL_ASYNC_ERROR_HANDLING=1 + TORCH_NCCL_DUMP_ON_TIMEOUT=1 in the job env prevent the 30-min NCCL watchdog hang and zombie GPU-context leaks seen on prior runs.train.py forces mixup / mosaic / copyblend to None unless you pass --mixup --mosaic --copyblend (opt-in). Photometric / random zoom-out / random IoU crop / random flip / scale-jitter are --no-* opt-out. Easy to silently lose all three augs after a config round-trip.oom / error / cancel over .err matches legitimate config-dump keys (random_zoom_out, scheduler_no_aug_steps, <Config CANCELLED at ...>). Use squeue -j $JID returning empty as the crash check; reserve .err-grep for milestone keywords only..err-printed resolved config before reasoning about aug timings.