prlq's picture
lora_kit: per-folder download (repo file list capped at 100k)
5f94a5e verified
|
Raw History Blame Contribute Delete
3.43 kB

MobileManipBench per-cell π0.5 LoRA kit (536 / nova, 2026-11-16)

One script prepares everything on a fresh GPU machine and trains a 10K-step LoRA for one task × robot, continued from pi05_v3/80000:

curl -sfLO https://huggingface.co/datasets/prlq/mobilemanipbench-data/resolve/main/debug/536/lora_kit/train_lora_cell.sh
bash train_lora_cell.sh N025 sawyer_umi /data/mmb_lora          # TASK robot workdir

Steps (each skipped if already done): (1) openpi @215abfb + pi05_v3/code/openpi_mmb.patch, uv sync; (2) pi05_v3/80000 (12.4 GB); (3) this cell's data; (4) canonical_action_v2/convert_merge.py --v2 → $WORK/lerobot/mmb/<cell>; (5) add_lora_config.py → TrainConfig pi05_mmb_v3_lora_<cell> and scripts/train.py. Cell names: <task lower>_<go2|saw|str>, e.g. n025_saw.

Recipe (same as the runs on 536)

LoRA on PaliGemma (gemma_2b_lora) + action expert (gemma_300m_lora), backbone frozen; init = all parameters of pi05_v3/80000; data definition and per-robot normalization identical to pi05_mmb_v3 (stats asset mmb/v2t of the 80K checkpoint; state 14-D canonical, action chunk H=10 [base cmd | EEF target xyz+rot6d | gripper]); batch 32, cosine LR 500 warm-up → 2.5e-5 → 2.5e-6 over 10K, no EMA, checkpoint every 2K. Measured on one L40S 46 GB: 3.0 s/step → 10K ≈ 8.3 h. Multi-GPU: all visible GPUs are used (data parallel, global batch stays 32). Env: STEPS, BATCH, SAVE, RESUME=1, SKIP_TRAIN=1, CUDA_VISIBLE_DEVICES.

Data (HF prlq/mobilemanipbench-data)

Training set of a cell = <TASK>/<robot>/dr_off/ (original, visual DR off, 50) + <TASK>/<robot>/dr_on_perturb/ (NEW: visual DR on + DART-style perturbation: the robot executes the expert command plus temporary offsets, the record keeps the clean expert command; ~50). The original dr_on/ is NOT used here (owner decision). Download: the script fetches each folder with hf_fetch.py (list_repo_tree + 32 parallel hf_hub_download). Do NOT use snapshot_download(allow_patterns=...) on this repo: the Hub's file list of the repo stops at 100,000 files (it ends inside N021), so later folders (N025+, canonical_action_v2, pi05_v3) come back as "Fetching 0 files" (fixed 2026-11-16 04:45 CST). Status on HF at 2026-11-16 03:35 CST (perturbed episodes; collection running, 50 per cell is the target):

task go2_arm sawyer_umi stretch3
N025 52 ✅ 51 ✅ 47
N033 20 18 11
N051 20 19 15
N058 18 11 12
N004 4 12 15

Evaluate a checkpoint

Serve with serve_mmb_policy_lora.py (= pi05_v3/code/serve_mmb_policy.py + the stats lookup under asset id mmb/v2t, needed for LoRA configs):

cd $WORK/openpi && OPENPI_SKIP_IMAGE_KEYS=right_wrist_0_rgb uv run --no-sync python ../serve_mmb_policy_lora.py \
    --config pi05_mmb_v3_lora_n025_saw --checkpoint $WORK/openpi/checkpoints/pi05_mmb_v3_lora_n025_saw/n025_saw_10k/2000 --port 8000

Eval (MagicSim mobile-ready 7fc2f64009 + two eval fixes: ep.cs0 set for every server in Eval/controller.py, and --walltime-s raised in scripts/eval/eval_cell.sh / eval_family.sh): scripts/eval/run_eval.sh N025 sawyer_umi --server ws://host:8000 --episodes 5 --seed-start 90001 --dr on --visual-dr off --policy pi05. Baseline 80K on these 7 cells (N025 str/saw, N011 saw, N058 go2, N004 go2, N051 saw, N033 go2), seeds 90001-90005, visual DR off: 0 / 41.