File size: 2,726 Bytes
293e6e5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | """Paired comparison of a LeRobot eval run against the official eval_info.json.
lerobot-eval seeds episode i with start_seed + i (see rollout seeding in
lerobot/scripts/lerobot_eval.py), so with --seed=1000 our episode i uses the same
initial state as official episode i (seed 1000 + i).
Usage:
python scripts/compare_with_official.py outputs/eval/official_500ep/eval_info.json
"""
import argparse
import json
import math
from pathlib import Path
OFFICIAL = Path(__file__).resolve().parents[1] / "results/day01/official_eval_info_hub.json"
def wilson(k: int, n: int, z: float = 1.96) -> tuple[float, float]:
p = k / n
denom = 1 + z**2 / n
center = (p + z**2 / (2 * n)) / denom
half = z * math.sqrt(p * (1 - p) / n + z**2 / (4 * n**2)) / denom
return center - half, center + half
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("ours", type=Path, help="eval_info.json written by lerobot-eval")
parser.add_argument("--official", type=Path, default=OFFICIAL)
parser.add_argument("--start-seed", type=int, default=1000)
args = parser.parse_args()
ours = json.loads(args.ours.read_text())["per_task"][0]["metrics"]["successes"]
official = {e["seed"]: e["success"] for e in json.loads(args.official.read_text())["per_episode"]}
seeds = [args.start_seed + i for i in range(len(ours))]
missing = [s for s in seeds if s not in official]
if missing:
raise SystemExit(f"official eval has no episodes for seeds {missing[:5]}...")
theirs = [official[s] for s in seeds]
n = len(ours)
k_ours, k_theirs = sum(ours), sum(theirs)
both = sum(a and b for a, b in zip(ours, theirs))
only_ours = sum(a and not b for a, b in zip(ours, theirs))
only_theirs = sum(b and not a for a, b in zip(ours, theirs))
agree = n - only_ours - only_theirs
lo, hi = wilson(k_ours, n)
lo_t, hi_t = wilson(k_theirs, n)
print(f"seeds {seeds[0]}-{seeds[-1]} (n={n})")
print(f"ours : {k_ours}/{n} = {100 * k_ours / n:.1f}% (95% Wilson CI {100 * lo:.1f}-{100 * hi:.1f}%)")
print(f"official : {k_theirs}/{n} = {100 * k_theirs / n:.1f}% (95% Wilson CI {100 * lo_t:.1f}-{100 * hi_t:.1f}%)")
print(f"same outcome on {agree}/{n} seeds; both succeed {both}, only ours {only_ours}, only official {only_theirs}")
discordant = only_ours + only_theirs
if discordant:
# Exact two-sided McNemar test on the discordant pairs (binomial with p = 0.5).
k = min(only_ours, only_theirs)
p = min(1.0, 2 * sum(math.comb(discordant, i) for i in range(k + 1)) / 2**discordant)
print(f"exact McNemar p = {p:.3f} ({discordant} discordant seeds)")
if __name__ == "__main__":
main()
|