fdstudio-scripts / validate_batch.py
Kavin60606's picture
Upload validate_batch.py with huggingface_hub
f7d3f89 verified
Raw History Blame Contribute Delete
16.4 kB
#!/usr/bin/env python3
"""Data Validation engine (stage 4) — runs in the retarget venv (numpy<2 + h5py + scipy).
Operates on the transformed output (store/<run>/transform/retargeted/*/retargeted.hdf5) which the
retarget engine already annotates with rich QA (`qa`, `m4`, `ik_errors_*`). This script:
mode=qa → per-clip Trajectory Quality & Hygiene report (feasibility + motion + integrity)
mode=correct → clamp joint limits + smooth vel/acc/jerk → corrected hdf5 (no retime)
mode=similarity → transformed-ego vs teleop action-distribution distance + score
Writes JSON to --out (the API reads it back). Invoked exactly like transform_batch.py.
"""
from __future__ import annotations
import argparse
import glob
import json
import os
import xml.etree.ElementTree as ET
import numpy as np
# ---- tunable thresholds (placeholder defaults; surfaced in the report so the UI can show them) ----
TH = {
"ik_ok_m": 0.02, # per-frame IK error under this (m) = frame "solved"
"ik_success_min": 0.90, # >= this fraction of frames solved → feasibility OK
"limit_viol_warn": 0.001, # fraction of (frame×joint) samples out of joint range
"limit_viol_fail": 0.02,
"selfcol_min_dist_m": 0.02, # min link clearance (m) below which = collision risk
"continuity_max_drad": 0.5, # max per-frame per-joint jump (rad) before "teleport"
"chatter_warn": 2.0, # oscillation rate (from retarget qa) warn level
}
FINGER_NAMES = ("left_finger", "right_finger", "finger")
def joint_limits(xml_path: str) -> dict:
"""Parse <joint name range> from yam_real.xml → {base_name: (lo, hi)} (radians)."""
out = {}
try:
root = ET.parse(xml_path).getroot()
for j in root.iter("joint"):
name, rng = j.get("name"), j.get("range")
if name and rng:
lo, hi = (float(x) for x in rng.split())
out[name] = (lo, hi)
except Exception:
pass
return out
def limits_vector(joint_names: list, lims: dict) -> np.ndarray:
"""Build (14, 2) lo/hi aligned to the clip's joint_names (strip R_/L_ prefix → base)."""
rows = []
for jn in joint_names:
base = jn.split("_", 1)[1] if (jn.startswith("R_") or jn.startswith("L_")) else jn
if base in lims:
rows.append(lims[base])
elif any(f in base for f in FINGER_NAMES):
# fingers: use whichever finger range exists
fr = next((lims[k] for k in lims if "finger" in k), (-1e9, 1e9))
rows.append(fr)
else:
rows.append((-1e9, 1e9))
return np.array(rows, dtype=float)
def _attr_json(h, key):
v = h.attrs.get(key)
if v is None:
return {}
try:
return json.loads(v if isinstance(v, str) else v.decode())
except Exception:
return {}
def qa_one(path: str, lims: dict) -> dict:
import h5py
with h5py.File(path, "r") as h:
jp = np.asarray(h["joint_positions"][:], dtype=float) # (T, 14)
ts = np.asarray(h["timestamps"][:], dtype=float) if "timestamps" in h else None
fps = float(h.attrs.get("fps", 30.0))
jn = json.loads(h.attrs.get("joint_names", "[]"))
ik_L = np.asarray(h["ik_errors_L"][:], dtype=float) if "ik_errors_L" in h else None
ik_R = np.asarray(h["ik_errors_R"][:], dtype=float) if "ik_errors_R" in h else None
qa = _attr_json(h, "qa")
m4 = _attr_json(h, "m4")
clip = h.attrs.get("clip", os.path.basename(os.path.dirname(path)))
task = h.attrs.get("task", "")
T = jp.shape[0]
dt = 1.0 / fps if fps else 1.0 / 30
# --- integrity ---
finite = bool(np.isfinite(jp).all())
dframe = np.abs(np.diff(jp, axis=0)) if T > 1 else np.zeros((0, jp.shape[1]))
max_jump = float(dframe.max()) if dframe.size else 0.0
frame_drops = 0
if ts is not None and ts.size > 2:
gaps = np.diff(ts)
frame_drops = int(np.sum(gaps > 1.8 * np.median(gaps)))
# --- feasibility ---
lo_hi = limits_vector(jn, lims) if jn else np.tile([-1e9, 1e9], (jp.shape[1], 1))
below = jp < lo_hi[:, 0]
above = jp > lo_hi[:, 1]
viol = below | above
limit_viol_frac = float(viol.mean()) if viol.size else 0.0
per_joint_viol = viol.mean(axis=0).tolist() if viol.size else []
def ik_frac(ik):
return float(np.mean(ik <= TH["ik_ok_m"])) if ik is not None and ik.size else None
ik_success = [f for f in (ik_frac(ik_R), ik_frac(ik_L)) if f is not None]
ik_success_min = min(ik_success) if ik_success else None
# self-collision from the retarget qa (min link clearance across arms)
selfcol_dists = []
for arm in ("R", "L"):
d = ((qa.get("arms", {}).get(arm, {}) or {}).get("metrics", {}) or {}).get("selfcol_min_dist")
if d is not None:
selfcol_dists.append(float(d))
selfcol_min = min(selfcol_dists) if selfcol_dists else None
# --- motion (finite-diff on joints; peaks; plus retarget smoothness/chatter/saturation) ---
vel = np.diff(jp, axis=0) / dt if T > 1 else np.zeros((0, jp.shape[1]))
acc = np.diff(vel, axis=0) / dt if vel.shape[0] > 1 else np.zeros((0, jp.shape[1]))
jerk = np.diff(acc, axis=0) / dt if acc.shape[0] > 1 else np.zeros((0, jp.shape[1]))
peak_vel = float(np.abs(vel).max()) if vel.size else 0.0
peak_acc = float(np.abs(acc).max()) if acc.size else 0.0
peak_jerk = float(np.abs(jerk).max()) if jerk.size else 0.0
def arm_metric(key):
vals = []
for arm in ("R", "L"):
v = ((qa.get("arms", {}).get(arm, {}) or {}).get("metrics", {}) or {}).get(key)
if v is not None:
vals.append(float(v))
return vals
ldlj = arm_metric("ldlj") # log dimensionless jerk (smoothness; more negative = jerkier)
chatter = arm_metric("chatter_rate")
saturation = arm_metric("saturation_frac")
# --- verdict: blend our checks with the retarget's own qa_verdict ---
reasons = []
verdict = "PASS"
def demote(level, why):
nonlocal verdict
order = {"PASS": 0, "WARN": 1, "FAIL": 2}
if order[level] > order[verdict]:
verdict = level
reasons.append(why)
if not finite:
demote("FAIL", "non-finite joint values")
if max_jump > TH["continuity_max_drad"]:
demote("WARN", f"discontinuity {max_jump:.2f} rad/frame")
if limit_viol_frac > TH["limit_viol_fail"]:
demote("FAIL", f"joint-limit violation {limit_viol_frac*100:.1f}%")
elif limit_viol_frac > TH["limit_viol_warn"]:
demote("WARN", f"minor limit violation {limit_viol_frac*100:.2f}%")
if ik_success_min is not None and ik_success_min < TH["ik_success_min"]:
demote("FAIL" if ik_success_min < 0.8 else "WARN", f"IK solved {ik_success_min*100:.0f}%")
if selfcol_min is not None and selfcol_min < TH["selfcol_min_dist_m"]:
demote("WARN", f"self-collision clearance {selfcol_min*100:.1f}cm")
if chatter and max(chatter) > TH["chatter_warn"]:
demote("WARN", f"chatter {max(chatter):.1f}")
rv = qa.get("verdict")
if rv == "FAIL":
demote("WARN", "retarget QA flagged FAIL") # respect but don't hard-fail on retarget-only
return {
"clip": clip, "task": task, "n_frames": T, "fps": fps,
"verdict": verdict, "reasons": reasons,
"feasibility": {
"limit_viol_frac": round(limit_viol_frac, 4),
"per_joint_viol": [round(x, 4) for x in per_joint_viol],
"ik_success_min": None if ik_success_min is None else round(ik_success_min, 3),
"selfcol_min_dist_m": None if selfcol_min is None else round(selfcol_min, 4),
},
"motion": {
"peak_vel": round(peak_vel, 3), "peak_acc": round(peak_acc, 2), "peak_jerk": round(peak_jerk, 1),
"ldlj": [round(x, 2) for x in ldlj], "chatter": [round(x, 2) for x in chatter],
"saturation": [round(x, 3) for x in saturation],
},
"integrity": {"finite": finite, "max_jump_rad": round(max_jump, 3), "frame_drops": frame_drops},
"retarget_verdict": rv,
}
def run_qa(retargeted_dir: str, xml_path: str) -> dict:
lims = joint_limits(xml_path)
files = sorted(glob.glob(os.path.join(retargeted_dir, "*", "retargeted.hdf5")))
clips = []
for f in files:
try:
clips.append(qa_one(f, lims))
except Exception as e:
clips.append({"clip": os.path.basename(os.path.dirname(f)), "verdict": "FAIL",
"reasons": [f"read error: {e}"], "error": str(e)})
counts = {v: sum(1 for c in clips if c.get("verdict") == v) for v in ("PASS", "WARN", "FAIL")}
certified = [c["clip"] for c in clips if c.get("verdict") in ("PASS", "WARN")]
return {
"mode": "qa",
"n_clips": len(clips),
"counts": counts,
"certified": certified,
"clips": clips,
"thresholds": TH,
}
def _peak_jerk(jp: np.ndarray, dt: float) -> float:
if jp.shape[0] < 4:
return 0.0
j = np.diff(jp, n=3, axis=0) / (dt ** 3)
return float(np.abs(j).max()) if j.size else 0.0
def correct_one(path: str, lims: dict) -> dict:
"""Clamp to joint limits + smooth (Savitzky-Golay) — no retime. Writes corrected.hdf5 alongside."""
import h5py
from scipy.signal import savgol_filter
with h5py.File(path, "r") as h:
jp = np.asarray(h["joint_positions"][:], dtype=float)
jn = json.loads(h.attrs.get("joint_names", "[]"))
fps = float(h.attrs.get("fps", 30.0))
clip = h.attrs.get("clip", os.path.basename(os.path.dirname(path)))
keep = {k: np.asarray(h[k][:]) for k in h.keys() if k != "joint_positions"}
attrs = dict(h.attrs)
T = jp.shape[0]
dt = 1.0 / fps if fps else 1.0 / 30
lo_hi = limits_vector(jn, lims) if jn else np.tile([-1e9, 1e9], (jp.shape[1], 1))
before = {"limit_viol_frac": round(float(((jp < lo_hi[:, 0]) | (jp > lo_hi[:, 1])).mean()), 4),
"peak_jerk": round(_peak_jerk(jp, dt), 1)}
clamped = np.clip(jp, lo_hi[:, 0], lo_hi[:, 1])
win = min(T if T % 2 else T - 1, 7) # odd window ≤ T, ≤ 7
if win >= 5 and T > win:
smoothed = savgol_filter(clamped, win, 3, axis=0)
smoothed = np.clip(smoothed, lo_hi[:, 0], lo_hi[:, 1]) # re-clamp after smoothing
else:
smoothed = clamped
after = {"limit_viol_frac": round(float(((smoothed < lo_hi[:, 0]) | (smoothed > lo_hi[:, 1])).mean()), 4),
"peak_jerk": round(_peak_jerk(smoothed, dt), 1)}
out_path = os.path.join(os.path.dirname(path), "corrected.hdf5")
with h5py.File(out_path, "w") as o:
o.create_dataset("joint_positions", data=smoothed)
for k, v in keep.items():
o.create_dataset(k, data=v)
for k, v in attrs.items():
o.attrs[k] = v
o.attrs["corrected"] = True
o.attrs["smooth_window"] = int(win)
return {"clip": clip, "before": before, "after": after,
"jerk_reduction": round(before["peak_jerk"] - after["peak_jerk"], 1),
"smooth_window": int(win), "output": out_path}
def run_correct(retargeted_dir: str, xml_path: str) -> dict:
lims = joint_limits(xml_path)
files = sorted(glob.glob(os.path.join(retargeted_dir, "*", "retargeted.hdf5")))
clips, tot_j = [], 0.0
for f in files:
try:
c = correct_one(f, lims)
clips.append(c)
tot_j += max(0.0, c["jerk_reduction"])
except Exception as e:
clips.append({"clip": os.path.basename(os.path.dirname(f)), "error": str(e)})
ok = [c for c in clips if "error" not in c]
return {"mode": "correct", "n_clips": len(clips), "corrected": len(ok),
"mean_jerk_reduction": round(tot_j / max(1, len(ok)), 1), "clips": clips}
def _ego_actions(retargeted_dir: str, max_frames: int = 60000) -> np.ndarray:
import h5py
arrs, n = [], 0
for f in sorted(glob.glob(os.path.join(retargeted_dir, "*", "retargeted.hdf5"))):
with h5py.File(f, "r") as h:
a = np.asarray(h["joint_positions"][:], dtype=float)
arrs.append(a); n += a.shape[0]
if n >= max_frames:
break
return np.concatenate(arrs) if arrs else np.zeros((0, 14))
def _teleop_actions(teleop_dir: str, max_frames: int = 60000) -> np.ndarray:
import pyarrow.parquet as pq
arrs, n = [], 0
for f in sorted(glob.glob(os.path.join(teleop_dir, "**", "*.parquet"), recursive=True)):
try:
col = pq.read_table(f, columns=["action"]).column("action").to_pylist()
except Exception:
continue
a = np.array(col, dtype=float)
if a.ndim != 2:
continue
arrs.append(a); n += a.shape[0]
if n >= max_frames:
break
return np.concatenate(arrs) if arrs else np.zeros((0, 14))
def run_similarity(retargeted_dir: str, teleop_dir: str) -> dict:
"""Task-agnostic execution similarity: do transformed-ego action distributions look like
expert teleop? Per-joint Wasserstein (action + velocity) + range coverage + histograms."""
from scipy.stats import wasserstein_distance
ego = _ego_actions(retargeted_dir)
tel = _teleop_actions(teleop_dir)
if ego.size == 0 or tel.size == 0:
return {"mode": "similarity", "error": "no ego or teleop actions",
"ego_frames": int(ego.shape[0]), "teleop_frames": int(tel.shape[0])}
D = min(ego.shape[1], tel.shape[1])
per_joint = []
for j in range(D):
e, t = ego[:, j], tel[:, j]
s = t.std() or 1.0
wd = wasserstein_distance(e / s, t / s)
ev, tv = np.diff(e), np.diff(t)
sv = tv.std() or 1.0
wdv = wasserstein_distance(ev / sv, tv / sv) if ev.size and tv.size else 0.0
lo, hi = np.percentile(t, [1, 99])
cov = float(np.mean((e >= lo) & (e <= hi)))
# shared-range histograms for the UI overlay
rlo, rhi = float(min(e.min(), t.min())), float(max(e.max(), t.max()))
be, _ = np.histogram(e, bins=20, range=(rlo, rhi), density=True)
bt, edges = np.histogram(t, bins=20, range=(rlo, rhi), density=True)
per_joint.append({
"joint": j, "w_action": round(float(wd), 3), "w_vel": round(float(wdv), 3),
"sim": round(float(np.exp(-wd)), 3), "coverage": round(cov, 3),
"hist": {"edges": [round(x, 3) for x in edges.tolist()],
"ego": [round(x, 3) for x in be.tolist()],
"teleop": [round(x, 3) for x in bt.tolist()]},
})
score = float(np.mean([p["sim"] for p in per_joint]))
coverage = float(np.mean([p["coverage"] for p in per_joint]))
verdict = "PASS" if (score >= 0.6 and coverage >= 0.7) else ("WARN" if score >= 0.4 else "FAIL")
return {"mode": "similarity", "score": round(score, 3), "coverage": round(coverage, 3),
"verdict": verdict, "ego_frames": int(ego.shape[0]), "teleop_frames": int(tel.shape[0]),
"n_joints": D, "per_joint": per_joint}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--mode", default="qa", choices=["qa", "correct", "similarity"])
ap.add_argument("--retargeted", required=True, help="store/<run>/transform/retargeted")
ap.add_argument("--xml", default=os.path.join(os.path.dirname(__file__), "yam_real.xml"))
ap.add_argument("--teleop", default="", help="teleop actions .npy (for similarity)")
ap.add_argument("--thresholds", default="", help="JSON overrides merged into TH")
ap.add_argument("--out", required=True)
a = ap.parse_args()
if a.thresholds:
try:
TH.update({k: float(v) for k, v in json.loads(a.thresholds).items() if k in TH})
except Exception:
pass
if a.mode == "qa":
rep = run_qa(a.retargeted, a.xml)
elif a.mode == "correct":
rep = run_correct(a.retargeted, a.xml)
elif a.mode == "similarity":
rep = run_similarity(a.retargeted, a.teleop)
else:
rep = {"mode": a.mode, "error": "unknown mode"}
with open(a.out, "w") as f:
json.dump(rep, f, indent=2)
print(json.dumps({"mode": a.mode, "n_clips": rep.get("n_clips"), "counts": rep.get("counts"),
"out": a.out}))
if __name__ == "__main__":
main()