File size: 3,829 Bytes
fa4121e | 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 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 | """Paired comparison of two policy evaluations on the same scenes (same --seed in eval_policy.py).
Reads each run's eval_results.json, or eval_episodes.jsonl if the run was stopped, and pairs
episodes by index (only indices both runs finished). Reports success per model, the discordant
pairs with an exact two-sided McNemar test, results per piece, and the grasp measurements.
Run: .venv/bin/python sim/paired_compare.py A=sim/reports/dart_eval/baseline B=sim/reports/dart_eval/dart \
--out sim/reports/dart_eval/paired.md
"""
from __future__ import annotations
import argparse
import json
from math import comb
from pathlib import Path
from statistics import median
def load(path: Path) -> dict:
f = path / "eval_results.json"
if f.exists():
eps = json.loads(f.read_text())["episodes"]
else:
eps = [json.loads(line) for line in (path / "eval_episodes.jsonl").read_text().splitlines() if line.strip()]
return {e["index"]: e for e in eps}
def mcnemar(b: int, c: int) -> float:
n = b + c
if n == 0:
return 1.0
k = min(b, c)
return min(1.0, 2 * sum(comb(n, i) for i in range(k + 1)) / 2 ** n)
def med(values):
values = [v for v in values if v is not None]
return round(median(values), 1) if values else None
def main():
ap = argparse.ArgumentParser(description=__doc__)
ap.add_argument("runs", nargs=2, help="NAME=path to an eval_policy.py output folder")
ap.add_argument("--out")
args = ap.parse_args()
(na, pa), (nb, pb) = (r.split("=", 1) for r in args.runs)
A, B = load(Path(pa)), load(Path(pb))
idx = sorted(set(A) & set(B))
n = len(idx)
sa = sum(A[i]["success"] for i in idx)
sb = sum(B[i]["success"] for i in idx)
only_a = sum(A[i]["success"] and not B[i]["success"] for i in idx)
only_b = sum(B[i]["success"] and not A[i]["success"] for i in idx)
lines = [f"# Paired comparison: {na} vs {nb}", "",
f"{n} scenes finished by both ({na}: {len(A)}, {nb}: {len(B)}).", "",
"| | " + na + " | " + nb + " |", "|---|---|---|",
f"| success | {sa}/{n} ({100 * sa / max(n, 1):.1f}%) | {sb}/{n} ({100 * sb / max(n, 1):.1f}%) |"]
for piece in sorted({A[i]["piece"] for i in idx}):
ids = [i for i in idx if A[i]["piece"] == piece]
lines.append(f"| {piece}s | {sum(A[i]['success'] for i in ids)}/{len(ids)} | {sum(B[i]['success'] for i in ids)}/{len(ids)} |")
for key, label in (("lifted", "piece lifted"),):
lines.append(f"| {label} | {sum(bool(A[i].get(key)) for i in idx)} | {sum(bool(B[i].get(key)) for i in idx)} |")
for key, label in (("close_lateral_mm", "sideways error when closing, median mm"),
("closest_mm", "closest approach to the piece, median mm")):
if any(key in A[i] for i in idx):
lines.append(f"| {label} | {med(A[i].get(key) for i in idx)} | {med(B[i].get(key) for i in idx)} |")
lines += ["", f"Discordant scenes: only {na} succeeded {only_a}, only {nb} succeeded {only_b}; "
f"exact McNemar p = {mcnemar(only_a, only_b):.3g}. Both succeeded "
f"{sum(A[i]['success'] and B[i]['success'] for i in idx)}, neither "
f"{sum(not A[i]['success'] and not B[i]['success'] for i in idx)}."]
moves = sorted({f"{A[i]['source']}-{A[i]['destination']}" for i in idx})
lines += ["", "| move | " + na + " | " + nb + " |", "|---|---|---|"]
for mv in moves:
ids = [i for i in idx if f"{A[i]['source']}-{A[i]['destination']}" == mv]
lines.append(f"| {mv} | {sum(A[i]['success'] for i in ids)}/{len(ids)} | {sum(B[i]['success'] for i in ids)}/{len(ids)} |")
text = "\n".join(lines) + "\n"
if args.out:
Path(args.out).write_text(text)
print(text)
if __name__ == "__main__":
main()
|