analyze_tune.py

"""Summarise the ST-GFN hyperparameter grid against the TB reference.

The point of this sweep is fairness: a null result for the proposed method is
only meaningful if the method was not simply mis-tuned. We report the best
ST-GFN configuration found anywhere in the grid, per metric, and compare it with
plain TB run under an identical budget.
"""
from __future__ import annotations

import argparse
import glob
import json
import os
from collections import defaultdict

import numpy as np

METRICS = [("modes_found", "Modes", True),
           ("l1_to_target", "L1 to P*", False),
           ("mean_reward", "Mean R", True)]


def load(out_dir):
    by = defaultdict(lambda: defaultdict(list))   # config -> method -> runs
    for path in glob.glob(os.path.join(out_dir, "*", "*.json")):
        cfg = os.path.basename(os.path.dirname(path))
        with open(path) as f:
            r = json.load(f)
        by[cfg][r["method"]].append(r["final"])
    return by


def agg(runs, key):
    vals = [r.get(key) for r in runs if r.get(key) is not None]
    if not vals:
        return None, None
    mu = float(np.mean(vals))
    ci = float(1.96 * np.std(vals, ddof=1) / np.sqrt(len(vals))) if len(vals) > 1 else 0.0
    return mu, ci


if __name__ == "__main__":
    ap = argparse.ArgumentParser()
    ap.add_argument("--out", default="../outputs/tune")
    ap.add_argument("--save", default="../outputs/summary_tune.json")
    args = ap.parse_args()

    by = load(args.out)
    rows, tb_rows = [], []
    for cfg, methods in sorted(by.items()):
        for meth, runs in methods.items():
            entry = {"config": cfg, "method": meth, "n": len(runs)}
            for key, _, _ in METRICS:
                mu, ci = agg(runs, key)
                entry[key] = mu
                entry[key + "_ci"] = ci
            (tb_rows if meth == "tb" else rows).append(entry)

    print(f"{'config':38s} {'n':>2s} " + " ".join(f"{n:>14s}" for _, n, _ in METRICS))
    for r in sorted(rows, key=lambda x: (x["l1_to_target"] is None, x["l1_to_target"])):
        cells = []
        for key, _, _ in METRICS:
            v, ci = r.get(key), r.get(key + "_ci")
            cells.append(f"{v:9.3f}±{ci:4.3f}" if v is not None else f"{'-':>14s}")
        print(f"{r['config']:38s} {r['n']:2d} " + " ".join(cells))

    if tb_rows:
        print("\nTB reference (identical budget):")
        for r in tb_rows:
            cells = []
            for key, _, _ in METRICS:
                v, ci = r.get(key), r.get(key + "_ci")
                cells.append(f"{v:9.3f}±{ci:4.3f}" if v is not None else f"{'-':>14s}")
            print(f"{'tb (baseline)':38s} {r['n']:2d} " + " ".join(cells))

        print("\nBest ST-GFN configuration per metric vs TB:")
        for key, name, hib in METRICS:
            cand = [r for r in rows if r.get(key) is not None]
            if not cand:
                continue
            best = (max if hib else min)(cand, key=lambda r: r[key])
            tb = tb_rows[0].get(key)
            verdict = ("ST-GFN better" if ((best[key] > tb) == hib) else "TB better")
            print(f"  {name:10s} best ST-GFN {best[key]:.3f} ({best['config']})"
                  f"   vs TB {tb:.3f}   -> {verdict}")

    with open(args.save, "w") as f:
        json.dump({"stgfn": rows, "tb": tb_rows}, f, indent=2)
    print(f"\nwrote {args.save}")