fdstudio-scripts / transform_batch.py
Kavin60606's picture
Upload transform_batch.py with huggingface_hub
c6064aa verified
Raw History Blame Contribute Delete
15.9 kB
"""T3/T4 — batch retarget over selected_clips.json with live, human-readable progress.
Runs under .venv-retarget. For each shortlisted clip: materialize a single-episode parquet
→ run the real M1..M6 engine → emit per-clip friendly-stage progress + metrics into
transform_report.json (written incrementally so the console can poll it live).
"""
import argparse
import contextlib
import io
import json
import os
import ssl
import sys
import time
import urllib.request
from concurrent.futures import ProcessPoolExecutor, as_completed
from pathlib import Path
import numpy as np
import pyarrow.compute as pc
import pyarrow.parquet as pq
# macOS Python's urllib has no CA bundle by default → HF downloads fail with
# CERTIFICATE_VERIFY_FAILED. Use certifi's bundle (fall back to unverified only if absent).
try:
import certifi
_SSL_CTX = ssl.create_default_context(cafile=certifi.where())
except Exception:
_SSL_CTX = ssl._create_unverified_context()
REPO = os.environ.get("FD_EGO_REPO", "Kavin60606/EgoDex-PickPlace-10hr")
PREFIX = os.environ.get("FD_EGO_PREFIX", "") # "" = 10hr root layout; "train/" = griffinlabs
RESOLVE = "https://huggingface.co/datasets/%s/resolve/main/%s"
# customer-facing stage names (replace M1..M6)
STAGES = [
("load", "Reading human demo"),
("retarget", "Retargeting motion to robot"),
("base", "Placing the robot"),
("ik", "Solving arm joints (IK)"),
("collision", "Checking arm collisions"),
("smooth", "Smoothing trajectory"),
("qa", "Quality check"),
("output", "Saving robot trajectory"),
]
STAGE_KEYS = [k for k, _ in STAGES]
# order matters: match "[M3.5]" before "[M3]"
MARKERS = [("[M6]", "output"), ("[M5]", "qa"), ("[M4]", "smooth"),
("[M3.5]", "collision"), ("[M3]", "ik"), ("[M2]", "base"), ("[M1]", "retarget")]
# ── QA verdict: fold FAITHFULNESS (tracking / orientation / arm-collision) into the per-clip
# PASS/WARN/FAIL alongside the motion-quality verdict. Was motion-only, which missed clips that
# move smoothly but don't actually follow the human demo. Thresholds are env-overridable so the
# UI can tune them per run (see config.py for the same defaults). higher value = worse.
def _envf(key: str, default: float) -> float:
try:
return float(os.environ.get(key, default))
except (TypeError, ValueError):
return default
_QA_IK_WARN = _envf("QA_IK_CM_WARN", 3.0)
_QA_IK_FAIL = _envf("QA_IK_CM_FAIL", 6.0)
_QA_ORI_WARN = _envf("QA_ORI_DEG_WARN", 15.0)
_QA_ORI_FAIL = _envf("QA_ORI_DEG_FAIL", 30.0)
_QA_ARMS_STRICT = os.environ.get("QA_ARMS_STRICT", "0") == "1"
_QA_RANK = {"PASS": 0, "WARN": 1, "FAIL": 2, "?": 0}
def _grade_high(v: float, warn: float, fail: float) -> str:
"""Grade a metric where higher is worse against warn/fail bounds."""
if v is None:
return "PASS"
if v >= fail:
return "FAIL"
if v >= warn:
return "WARN"
return "PASS"
def _worst(*verdicts: str) -> str:
return max(verdicts, key=lambda v: _QA_RANK.get(v, 0))
KP_COLS = [f"observation.state.{h}{k}" for h in ("right", "left")
for k in ("ThumbTip", "IndexFingerTip", "MiddleFingerTip", "Hand")]
def _ensure(cache: Path, rel: str, timeout: int = 30, retries: int = 4) -> Path:
"""Download a repo file to cache, ONCE. Bounded per-request timeout + retries + atomic write so a
single stalled host connection can't hang the whole run forever (urlopen has no default timeout —
that infinite block was what froze prewarm at 0% CPU)."""
local = cache / REPO.replace("/", "__") / rel
if local.exists() and local.stat().st_size > 0:
return local
local.parent.mkdir(parents=True, exist_ok=True)
url = RESOLVE % (REPO, rel)
last = None
for attempt in range(retries):
try:
req = urllib.request.Request(url, headers={"User-Agent": "fd-studio"})
with urllib.request.urlopen(req, timeout=timeout, context=_SSL_CTX) as r:
data = r.read()
tmp = local.with_name(local.name + ".part")
tmp.write_bytes(data)
tmp.replace(local) # atomic — a killed/partial download never looks complete
return local
except Exception as e:
last = e
time.sleep(min(5.0, 1.0 * (attempt + 1)))
raise RuntimeError(f"download failed after {retries} tries: {rel}: {last}")
def _episode_range(cache: Path, subset: str, ep: int) -> dict:
base = f"{PREFIX}{subset}"
cols = ["episode_index", "length", "tasks", "data/chunk_index", "data/file_index",
"dataset_from_index", "dataset_to_index"]
for i in range(0, 12): # episodes/chunk-000/file-000.parquet, file-001, ...
rel = f"{base}/meta/episodes/chunk-000/file-{i:03d}.parquet"
try:
local = _ensure(cache, rel)
except Exception:
break
for row in pq.read_table(local, columns=cols).to_pylist():
if row["episode_index"] == ep:
return row
raise KeyError(f"episode {ep} not found in {subset}")
def _materialize(cache: Path, subset: str, ep: int, out_root: Path) -> str:
r = _episode_range(cache, subset, ep)
ci, fi = r["data/chunk_index"], r["data/file_index"]
data_rel = f"{PREFIX}{subset}/data/chunk-{ci:03d}/file-{fi:03d}.parquet"
local = _ensure(cache, data_rel)
# a data file concatenates many episodes; `dataset_from/to_index` are GLOBAL offsets,
# so filter by the file's own episode_index column instead of slicing.
table = pq.read_table(local, columns=KP_COLS + ["episode_index"])
table = table.filter(pc.equal(table["episode_index"], ep)).select(KP_COLS)
mat = out_root / "materialized" / subset / "data" / "chunk-000"
mat.mkdir(parents=True, exist_ok=True)
# Name by SUBSET + episode. The retarget output dir is clip_<this-file-stem>, and episode numbers
# repeat across categories (tools#157, dice_balls#157, …) — naming by episode alone made them
# collide and OVERWRITE, silently dropping ~79% of clips. Subset uses "_" (never "-"), so
# prep_lerobot's `clip.split("-")[1]` still recovers the episode from "<subset>__file-000157".
out = mat / f"{subset}__file-{ep:06d}.parquet"
pq.write_table(table, out)
return str(out)
class Report:
def __init__(self, path: Path, clips: list, teleop: dict):
self.path = path
self.teleop = teleop
self.data = {
"clips_total": len(clips),
"clips": [{
"clip_id": c["clip_id"], "task": c.get("task", ""), "n_frames": c.get("n_frames", 0),
"status": "pending",
"stages": [{"key": k, "label": lbl, "status": "pending"} for k, lbl in STAGES],
"metrics": {}, "error": None, "output": None,
} for c in clips],
"match_report": None, "done": False,
}
self.write()
def write(self):
self.path.write_text(json.dumps(self.data, indent=2))
def advance(self, i: int, key: str):
pos = STAGE_KEYS.index(key)
for j, s in enumerate(self.data["clips"][i]["stages"]):
s["status"] = "done" if j < pos else ("running" if j == pos else "pending")
self.data["clips"][i]["status"] = "running"
def start(self, i: int):
self.data["clips"][i]["status"] = "running"
self.data["clips"][i]["stages"][0]["status"] = "running"
def finish(self, i: int, res: dict, collision: str):
c = self.data["clips"][i]
for s in c["stages"]:
s["status"] = "done"
c["status"] = "done"
c["output"] = res.get("output")
ik_R = round(res.get("ik_R_cm", 0), 2)
ik_L = round(res.get("ik_L_cm", 0), 2)
ori_R = round(res.get("ori_R_deg", 0), 1)
ori_L = round(res.get("ori_L_deg", 0), 1)
motion = res.get("qa_verdict", "?") # m5_qa motion-quality verdict
# grade faithfulness on the worse of the two arms, then combine (worst wins)
track_g = _grade_high(max(ik_R, ik_L), _QA_IK_WARN, _QA_IK_FAIL)
ori_g = _grade_high(max(ori_R, ori_L), _QA_ORI_WARN, _QA_ORI_FAIL)
arms_g = "PASS" if collision == "clean" else ("FAIL" if _QA_ARMS_STRICT else "WARN")
verdict = _worst(motion, track_g, ori_g, arms_g)
c["metrics"] = {
"ik_R_cm": ik_R, "ik_L_cm": ik_L, "ori_R_deg": ori_R, "ori_L_deg": ori_L,
"collision": collision, "qa": verdict, "qa_motion": motion, "dof": 14,
# per-component grades so the UI can show WHAT drove the verdict + re-grade live
"qa_components": {"tracking": track_g, "orientation": ori_g, "arms": arms_g, "motion": motion},
"n_frames": res.get("n_frames_out", c["n_frames"]),
}
def fail(self, i: int, msg: str):
c = self.data["clips"][i]
for s in c["stages"]:
if s["status"] == "running":
s["status"] = "fail"
c["status"] = "failed"
c["error"] = msg
def finalize(self):
done = [c for c in self.data["clips"] if c["status"] == "done"]
failed = sum(1 for c in self.data["clips"] if c["status"] == "failed")
teleop_hz = int(self.teleop.get("fps") or 30)
if done:
iks = [(c["metrics"]["ik_R_cm"] + c["metrics"]["ik_L_cm"]) / 2 for c in done]
ik_mean = round(float(np.mean(iks)), 2)
clean = sum(1 for c in done if c["metrics"]["collision"] == "clean")
fidelity = round(max(0.0, 1 - ik_mean / 10.0), 2)
clean_rate = round(clean / len(done), 2)
else: # nothing succeeded — report honestly, not a fake 100%
ik_mean = fidelity = clean_rate = None
self.data["match_report"] = {
"teleop_hz": teleop_hz, "ego_hz": 30,
"fps_variance": abs(teleop_hz - 30),
"action_hz_match": teleop_hz == 30,
"ik_mean_cm": ik_mean,
"traj_similarity": fidelity,
"collision_clean_rate": clean_rate,
"clips_done": len(done), "clips_failed": failed,
}
self.data["done"] = True
class _Tee(io.TextIOBase):
def __init__(self, orig, on_line):
self.orig, self.on_line, self.buf = orig, on_line, ""
def write(self, s):
self.orig.write(s)
self.buf += s
while "\n" in self.buf:
line, self.buf = self.buf.split("\n", 1)
self.on_line(line)
return len(s)
def flush(self):
self.orig.flush()
def _prewarm(cache: Path, clips: list, workers: int = 24) -> None:
"""Download each clip's episodes-meta + data parquet up front so the retarget workers only read
the cache. Runs in PARALLEL with bounded per-file timeouts — the old sequential + no-timeout
version froze the whole run for good if a single host connection stalled (0% CPU, flat disk).
Existence-check + atomic write make concurrent fetches of a shared file safe/idempotent."""
from concurrent.futures import ThreadPoolExecutor
def _warm(clip):
try:
subset, ep_s = clip["clip_id"].split("#")
r = _episode_range(cache, subset, int(ep_s))
_ensure(cache, f"{PREFIX}{subset}/data/chunk-{r['data/chunk_index']:03d}/file-{r['data/file_index']:03d}.parquet")
except Exception:
pass # a clip that can't prefetch is retried lazily in its worker — never blocks prewarm
with ThreadPoolExecutor(max_workers=min(workers, max(1, len(clips)))) as ex:
list(ex.map(_warm, clips))
def _run_one(payload: dict) -> dict:
"""Worker: materialize + retarget one clip (cache already warm). Returns a picklable result."""
clip = payload["clip"]
cache, out_root = Path(payload["cache"]), Path(payload["out_root"])
from retarget import process_clip
try:
subset, ep_s = clip["clip_id"].split("#")
buf = io.StringIO()
with contextlib.redirect_stdout(buf):
mat = _materialize(cache, subset, int(ep_s), out_root)
res = process_clip(mat, source="lerobot", out_root=str(out_root / "retargeted"))
if res is None:
return {"clip_id": clip["clip_id"], "status": "failed", "error": "clip skipped (QA gate or too few valid frames)"}
txt = buf.getvalue()
collision = "resolved" if any(k in txt for k in ("moved apart", "still colliding", "pushed")) else "clean"
return {"clip_id": clip["clip_id"], "status": "done", "res": res, "collision": collision}
except Exception as e:
return {"clip_id": clip["clip_id"], "status": "failed", "error": str(e)}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--selected", required=True)
ap.add_argument("--report", required=True)
ap.add_argument("--cache", required=True)
ap.add_argument("--out-root", required=True)
ap.add_argument("--limit", type=int, default=3)
ap.add_argument("--workers", type=int, default=0, help="0 = auto (cores-1); 1 = sequential")
args = ap.parse_args()
from retarget import process_clip
selected = json.loads(Path(args.selected).read_text())
all_clips = selected.get("clips", [])
clips = all_clips if args.limit <= 0 else all_clips[: args.limit] # limit<=0 → all
report = Report(Path(args.report), clips, selected.get("teleop", {}))
cache, out_root = Path(args.cache), Path(args.out_root)
workers = args.workers or max(1, min((os.cpu_count() or 2) - 1, len(clips)))
# ---- parallel path: process clips concurrently across cores ----
if workers > 1 and len(clips) > 1:
_prewarm(cache, clips)
idx = {c["clip_id"]: i for i, c in enumerate(clips)}
for i in range(len(clips)):
report.start(i)
report.write()
payloads = [{"clip": c, "cache": str(cache), "out_root": str(out_root)} for c in clips]
with ProcessPoolExecutor(max_workers=workers) as ex:
futs = [ex.submit(_run_one, p) for p in payloads]
for fut in as_completed(futs):
r = fut.result()
i = idx[r["clip_id"]]
if r["status"] == "done":
report.finish(i, r["res"], r["collision"])
else:
report.fail(i, r.get("error", "failed"))
report.write()
report.finalize()
report.write()
print(f"BATCH DONE ({workers} workers)")
return
# ---- sequential path (workers==1): live per-stage streaming ----
for i, clip in enumerate(clips):
report.start(i)
report.write()
try:
subset, ep_s = clip["clip_id"].split("#")
mat = _materialize(cache, subset, int(ep_s), out_root)
state = {"collision": "clean"}
def on_line(line, i=i, state=state):
for mk, key in MARKERS:
if mk in line:
report.advance(i, key)
report.write()
break
if "Collision check: clean" in line:
state["collision"] = "clean"
elif "moved apart" in line or "still colliding" in line or "pushed" in line:
state["collision"] = "resolved"
old = sys.stdout
sys.stdout = _Tee(old, on_line)
try:
res = process_clip(mat, source="lerobot", out_root=str(out_root / "retargeted"))
finally:
sys.stdout = old
if res is None:
raise RuntimeError("clip skipped (QA gate or too few valid frames)")
report.finish(i, res, state["collision"])
except Exception as e: # isolate — one bad clip must not abort the batch
report.fail(i, str(e))
report.write()
report.finalize()
report.write()
print("BATCH DONE")
if __name__ == "__main__":
main()