prep_lerobot.py

#!/usr/bin/env python3
"""prep_lerobot.py — bridge fd-studio's flat transform output to the layout to_lerobot_v21 +
multiview expect, then those scripts run unmodified.

fd-studio's transform writes /workspace/transform/retargeted/clip_file-XXXXXX/retargeted.hdf5
(flat, one dir per episode; the hdf5 carries attrs task=<category> and clip="file-XXXXXX").
to_lerobot_v21.discover() wants /workspace/retargeted/<cat>/clip_file-FFF_epN/retargeted.hdf5 and
finds the ego video at <EGO>/<cat>/videos/observation.images.camera/chunk-CCC/file-FFF.mp4.

This script, per retargeted clip:
  • re-derives the source (chunk, file) via transform_batch._episode_range (the same mapping the
    transform used), so file-FFF matches the ego data+video file index,
  • symlinks the hdf5 into /workspace/retargeted/<cat>/clip_file-FFF_epN/,
  • downloads that clip's ego VIDEO into the cache (transform only fetched keypoints).
EGO for to_lerobot is then just the cache's repo dir (already <cat>/videos + <cat>/meta).

Runs in the retarget venv (h5py). Usage:
  python prep_lerobot.py --flat /workspace/transform/retargeted --cache /workspace/ego_cache \
      [--retgt-out /workspace/retargeted]
Prints a JSON line: {"clips": N, "retgt": ..., "ego": ...}.
"""
import argparse
import glob
import json
import os
import sys
from concurrent.futures import ThreadPoolExecutor
from pathlib import Path

import h5py

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import transform_batch as tb  # noqa: E402  (reuse _episode_range/_ensure/REPO/PREFIX)


def _verdict(m: dict, th: dict) -> str:
    """Re-grade one clip's metrics with the user's thresholds (mirror of the console re-grade)."""
    rank = {"PASS": 0, "WARN": 1, "FAIL": 2, "?": 0}
    ik = max(float(m.get("ik_R_cm", 0) or 0), float(m.get("ik_L_cm", 0) or 0))
    ori = max(float(m.get("ori_R_deg", 0) or 0), float(m.get("ori_L_deg", 0) or 0))
    track = "FAIL" if ik >= th["ik_fail"] else "WARN" if ik >= th["ik_warn"] else "PASS"
    ori_g = "FAIL" if ori >= th["ori_fail"] else "WARN" if ori >= th["ori_warn"] else "PASS"
    arms = "PASS" if m.get("collision") == "clean" else {"allow": "PASS", "warn": "WARN", "fail": "FAIL"}.get(th["arms"], "PASS")
    mraw = m.get("qa_motion") or m.get("qa") or "PASS"
    arm_m = th["motion"]
    motion = "PASS" if arm_m == "off" else ("FAIL" if (arm_m == "strict" and mraw in ("WARN", "FAIL")) else mraw)
    return max([track, ori_g, arms, motion], key=lambda v: rank.get(v, 0))


def _excluded_ids(report_path: str, th: dict) -> set:
    """clip_ids ('<cat>#<ep>') the user's thresholds flag as FAIL — skipped from the dataset."""
    try:
        rep = json.loads(Path(report_path).read_text())
    except Exception:
        return set()
    out = set()
    for c in rep.get("clips", []):
        if c.get("status") == "done" and _verdict(c.get("metrics") or {}, th) == "FAIL":
            out.add(c.get("clip_id"))
    return out


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--flat", required=True, help="/workspace/transform/retargeted")
    ap.add_argument("--cache", required=True, help="/workspace/ego_cache")
    ap.add_argument("--retgt-out", default="/workspace/retargeted")
    # Exclude-enforcement: if --exclude, drop clips the user's thresholds flag as FAIL (read from
    # the retarget report). Off by default — we never auto-drop unless the user opts in.
    ap.add_argument("--report", default="", help="/workspace/report.json (for --exclude)")
    ap.add_argument("--exclude", type=int, default=0, help="1 = drop FAIL-flagged clips")
    ap.add_argument("--ik-warn", type=float, default=3.0)
    ap.add_argument("--ik-fail", type=float, default=6.0)
    ap.add_argument("--ori-warn", type=float, default=15.0)
    ap.add_argument("--ori-fail", type=float, default=30.0)
    ap.add_argument("--arms", default="allow")
    ap.add_argument("--motion", default="off")
    a = ap.parse_args()

    cache = Path(a.cache)
    ego_root = cache / tb.REPO.replace("/", "__")   # to_lerobot EGO = this (<cat>/videos + meta)
    done, vids, skipped, excluded = 0, 0, 0, 0

    th = {"ik_warn": a.ik_warn, "ik_fail": a.ik_fail, "ori_warn": a.ori_warn,
          "ori_fail": a.ori_fail, "arms": a.arms, "motion": a.motion}
    exclude_set = _excluded_ids(a.report, th) if (a.exclude and a.report) else set()
    if exclude_set:
        print(f"[exclude] user opted to drop {len(exclude_set)} FAIL-flagged clips", file=sys.stderr)

    # Pass 1 — resolve each retargeted clip to (cat, ep, chunk, file) and symlink it into the
    # category layout. Cheap (local hdf5 attrs + cached meta parquets), so keep it sequential.
    items = []  # (cat, ep, ci, fi, vid_rel)
    for hp in sorted(glob.glob(f"{a.flat}/*/retargeted.hdf5")):
        try:
            with h5py.File(hp, "r") as h:
                cat = str(h.attrs["task"])
                ep = int(str(h.attrs["clip"]).split("-")[1])
            if f"{cat}#{ep}" in exclude_set:   # user opted to drop this FAIL-flagged clip
                excluded += 1
                continue
            r = tb._episode_range(cache, cat, ep)
            ci, fi = r["data/chunk_index"], r["data/file_index"]
            fyy = f"file-{fi:03d}"

            # Prefer the VALIDATED+CORRECTED trajectory if Data Validation wrote one (clamp+smooth);
            # fall back to the raw retarget. This is what makes the pushed dataset the corrected one.
            corrected = str(Path(hp).with_name("corrected.hdf5"))
            src = corrected if os.path.exists(corrected) else hp

            dst = Path(a.retgt_out) / cat / f"clip_{fyy}_ep{ep}"
            dst.mkdir(parents=True, exist_ok=True)
            link = dst / "retargeted.hdf5"
            if link.is_symlink() or link.exists():
                link.unlink()
            os.symlink(os.path.abspath(src), link)

            vid_rel = f"{tb.PREFIX}{cat}/videos/observation.images.camera/chunk-{ci:03d}/{fyy}.mp4"
            items.append(vid_rel)
            done += 1
        except Exception as e:
            skipped += 1
            print(f"[warn] {hp}: {e}", file=sys.stderr)

    # Pass 2 — download the ego VIDEOS in PARALLEL (transform only fetched keypoints). These are the
    # big files; a sequential loop dominated packaging wall-clock. Dedup + bounded per-file timeout
    # in tb._ensure make concurrent fetches safe. A failed video is a non-fatal warn.
    def _dlvid(vid_rel: str) -> bool:
        try:
            tb._ensure(cache, vid_rel)
            return True
        except Exception as e:
            print(f"[warn] video dl {vid_rel}: {e}", file=sys.stderr)
            return False

    uniq = list(dict.fromkeys(items))  # preserve order, drop dup video files
    with ThreadPoolExecutor(max_workers=min(24, max(1, len(uniq)))) as ex:
        vids = sum(1 for ok in ex.map(_dlvid, uniq) if ok)

    print(json.dumps({"clips": done, "videos": vids, "skipped": skipped, "excluded": excluded,
                      "retgt": a.retgt_out, "ego": str(ego_root)}))


if __name__ == "__main__":
    main()