fdstudio-scripts / finalize_multiview.py
Kavin60606's picture
Upload finalize_multiview.py with huggingface_hub
d87189c verified
Raw History Blame Contribute Delete
5.25 kB
#!/venv/main/bin/python3
"""
finalize_multiview.py β€” after multiview.py has written all 3 view videos, fold
them into the dataset metadata so /workspace/retargeted_lerobot becomes a clean
LeRobot v2.1 dataset whose image schema MATCHES abc-teleop exactly:
observation.images.top / left_wrist / right_wrist (224x224, h264, yuv420p, 30fps)
It rewrites (backing up first):
meta/info.json β€” features: drop observation.images.camera, add 3 view keys
meta/episodes_stats.jsonl β€” per episode: drop camera stats, add the 3 view stats
(computed exactly during generation), keep state/action/…
The old camera .mp4 files are left on disk (unreferenced) β€” nothing is deleted.
Usage: /venv/main/bin/python3 finalize_multiview.py --ds /workspace/retargeted_lerobot \
[--teleop /workspace/abc-teleop] [--keep-camera]
"""
import os, json, glob, shutil, argparse
KEYS = ["observation.images.top",
"observation.images.left_wrist",
"observation.images.right_wrist"]
VID_H = VID_W = 224
FPS = 30
def video_block():
return {"dtype": "video", "shape": [VID_H, VID_W, 3],
"names": ["height", "width", "channels"],
"info": {"video.height": VID_H, "video.width": VID_W,
"video.codec": "h264", "video.pix_fmt": "yuv420p",
"video.is_depth_map": False, "video.fps": FPS,
"video.channels": 3, "has_audio": False}}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--ds", default="/workspace/retargeted_lerobot")
ap.add_argument("--teleop", default="/workspace/abc-teleop")
ap.add_argument("--keep-camera", action="store_true",
help="also keep observation.images.camera in the schema")
args = ap.parse_args()
ds = args.ds
meta = os.path.join(ds, "meta")
# results -> {episode_index: {key: stats}}
res = {}
with open(os.path.join(ds, "multiview_results.jsonl")) as f:
for line in f:
d = json.loads(line)
if "stats" in d:
res[d["episode_index"]] = d["stats"] # last wins (resume-safe)
print(f"results with stats: {len(res)} episodes")
# sanity: every episode must have all 3 view videos on disk
missing = []
for e in res:
ci = e // 1000
for k in KEYS:
p = f"{ds}/videos/chunk-{ci:03d}/{k}/episode_{e:06d}.mp4"
if not os.path.exists(p):
missing.append(p)
if missing:
print(f"WARNING: {len(missing)} view videos missing on disk, e.g. {missing[:3]}")
# ── info.json ────────────────────────────────────────────────────────────
info_p = os.path.join(meta, "info.json")
info = json.load(open(info_p))
feats = info["features"]
if not args.keep_camera:
feats.pop("observation.images.camera", None)
# insert the 3 view blocks first (ordering cosmetic)
new_feats = {k: video_block() for k in KEYS}
for k, v in feats.items():
if k not in new_feats:
new_feats[k] = v
info["features"] = new_feats
n_ep = info["total_episodes"]
n_keys = len(KEYS) + (1 if args.keep_camera else 0)
# mirror teleop's total_videos convention if available, else episodes*keys
tv = n_ep * n_keys
tinfo_p = os.path.join(args.teleop, "meta", "info.json")
if os.path.exists(tinfo_p):
ti = json.load(open(tinfo_p))
n_tvid_keys = sum(1 for v in ti["features"].values() if v.get("dtype") == "video")
if n_tvid_keys and ti.get("total_episodes"):
per_ep = ti["total_videos"] / ti["total_episodes"] # e.g. 3.0 or 1.0
# teleop counts videos-per-episode == #camera keys -> apply same rule
tv = int(round(per_ep / n_tvid_keys * n_keys * n_ep))
info["total_videos"] = tv
shutil.copy(info_p, info_p + ".camera.bak")
json.dump(info, open(info_p, "w"), indent=4)
print(f"info.json: video keys -> {[k for k,v in info['features'].items() if v.get('dtype')=='video']}"
f" total_videos={tv}")
# ── episodes_stats.jsonl ─────────────────────────────────────────────────
es_p = os.path.join(meta, "episodes_stats.jsonl")
shutil.copy(es_p, es_p + ".camera.bak")
out_lines = []
n_patched = 0
with open(es_p) as f:
for line in f:
d = json.loads(line)
e = d["episode_index"]
st = d.get("stats", {})
if not args.keep_camera:
st.pop("observation.images.camera", None)
if e in res:
for k in KEYS:
st[k] = res[e][k]
n_patched += 1
d["stats"] = st
out_lines.append(json.dumps(d))
with open(es_p, "w") as f:
f.write("\n".join(out_lines) + "\n")
print(f"episodes_stats.jsonl: patched {n_patched}/{len(out_lines)} episodes")
print("done. dataset image schema now matches abc-teleop (top/left_wrist/right_wrist).")
if __name__ == "__main__":
main()