RoboCasa MG Split 구조 분석

import h5py f = h5py.File('demo_gentex_im128_randcams.hdf5', 'r') list(f["mask"].keys()) # → ['10_demos', '20_demos', '30_demos', '40_demos', '50_demos', '60_demos', # '70_demos', '75_demos', '80_demos', '90_demos', '100_demos', '125_demos', # '150_demos', '200_demos', '250_demos', '300_demos', '400_demos', '500_demos', # '600_demos', '700_demos', '800_demos', '900_demos', '1000_demos', '1500_demos', # '2000_demos', '2500_demos', '3000_demos'] f["mask"]["30_demos"][()] # → array([b'demo_1161', b'demo_1196', ...], dtype='|S9') # 번호순 앞 30개가 아님 — 랜덤 샘플링된 30개!

HDF5 demo_0: 246 frames ↔ LeRobot ep_0: 246 frames ✓ HDF5 demo_1: 244 frames ↔ LeRobot ep_1: 244 frames ✓ HDF5 demo_2: 323 frames ↔ LeRobot ep_2: 323 frames ✓

def get_official_mg_episodes(task_name, mg_size=30): """HDF5 mask에서 공식 MG-N demo ID 추출 → episode 번호 리스트 반환""" hdf5_path = task_to_hdf5[task_name] with h5py.File(hdf5_path, 'r') as f: mask_key = f"{mg_size}_demos" demo_ids = f["mask"][mask_key][()] episode_ids = sorted([int(d.decode().replace("demo_", "")) for d in demo_ids]) return episode_ids # task → HDF5 경로 매핑 base_hdf5 = "/home/nas_main/minhopark/datasets/robot/robocasa/v0.1/single_stage" hdf5_files = sorted(glob.glob(f"{base_hdf5}/**/mg/**/demo_gentex_im128_randcams.hdf5", recursive=True)) task_to_hdf5 = {hp.split("/single_stage/")[1].split("/mg/")[0].split("/")[-1]: hp for hp in hdf5_files}