analysis.py

"""Bootstrap confidence intervals and paired model comparisons.

Reads the per-query JSON written by `eval_clap_tf.py --per-query`, so every
number carries an interval derived from the actual queries rather than a single
point estimate.

    python analysis.py results_final            # writes analysis.json + prints
    python analysis.py results_final --paired clotho

Bootstrap: 2000 resamples over queries, fixed seed. The paired comparison uses
the shared fixture fingerprint to guarantee both checkpoints scored the same
clip/caption sequence, then resamples queries *together* (paired bootstrap) and
reports a McNemar exact test on the discordant pairs.
"""
from __future__ import annotations

import argparse
import json
import math
from pathlib import Path

import numpy as np

SEED = 12345
N_BOOT = 2000


def _ci(values: np.ndarray, stat, n_boot: int = N_BOOT, seed: int = SEED):
    """Percentile bootstrap. Returns (low, high, standard error)."""
    rng = np.random.default_rng(seed)
    n = len(values)
    draws = np.empty(n_boot, dtype=np.float64)
    for i in range(n_boot):
        idx = rng.integers(0, n, n)
        draws[i] = stat(values[idx])
    return (float(np.percentile(draws, 2.5)), float(np.percentile(draws, 97.5)),
            float(draws.std(ddof=1)))


def cell_stats(d: dict) -> dict:
    """Recompute headline metrics from per-query arrays and bootstrap them."""
    t2a, a2t = d["metrics"]["t2a"], d["metrics"]["a2t"]
    tr = np.asarray(t2a["first_hit_rank"], dtype=np.int64)
    ar = np.asarray(a2t["first_hit_rank"], dtype=np.int64)
    aap = np.asarray(a2t["ap"], dtype=np.float64)

    out = {"model": d["model"], "dataset": d["dataset"], "tag": d.get("tag"),
           "audio_batch": d["protocol"]["audio_batch"],
           "trim_seconds": d["protocol"]["trim_seconds"],
           "fixture_fingerprint": d.get("fixture_fingerprint"),
           "counts": d["counts"], "metrics": {}}

    specs = [
        ("t2a_r1", tr, lambda v: float((v == 1).mean()), t2a["r@1"]),
        ("t2a_r5", tr, lambda v: float((v <= 5).mean()), t2a["r@5"]),
        ("t2a_r10", tr, lambda v: float((v <= 10).mean()), t2a["r@10"]),
        ("t2a_mrr", tr, lambda v: float((1.0 / v).mean()), t2a["map"]),
        ("a2t_r1", ar, lambda v: float((v == 1).mean()), a2t["r@1"]),
        ("a2t_r5", ar, lambda v: float((v <= 5).mean()), a2t["r@5"]),
        ("a2t_r10", ar, lambda v: float((v <= 10).mean()), a2t["r@10"]),
        ("a2t_map", aap, lambda v: float(v.mean()), a2t["map"]),
    ]
    for name, arr, stat, reported in specs:
        point = stat(arr)
        lo, hi, se = _ci(arr, stat)
        # the bootstrap point must reproduce the harness's own number
        assert abs(point - reported) < 1e-6, (name, point, reported)
        out["metrics"][name] = {"value": point, "ci95": [lo, hi], "se": se}
    out["metrics"]["n_queries"] = {"t2a": int(len(tr)), "a2t": int(len(ar))}
    # kept in memory for the paired tests; stripped before the JSON is written
    out["_pq"] = {"t2a": tr, "a2t": ar, "a2t_ap": aap}
    return out


def paired(a: dict, b: dict, metric: str = "r@1") -> dict:
    """Paired bootstrap on two cells of the same fixture, plus McNemar for hits."""
    assert a["fixture_fingerprint"] == b["fixture_fingerprint"], (
        "fixtures differ; paired comparison is not legitimate")

    if metric.endswith("map"):
        va = a["_pq"]["a2t_ap"].astype(np.float64)
        vb = b["_pq"]["a2t_ap"].astype(np.float64)
        n = len(va)
        diff = float(vb.mean() - va.mean())
        rng = np.random.default_rng(SEED)
        deltas = np.empty(N_BOOT)
        for i in range(N_BOOT):
            idx = rng.integers(0, n, n)
            deltas[i] = vb[idx].mean() - va[idx].mean()
        return {"a": a["model"], "b": b["model"], "metric": metric,
                "delta": diff,
                "ci95": [float(np.percentile(deltas, 2.5)),
                         float(np.percentile(deltas, 97.5))],
                "n": n, "discordant_a_only": None, "discordant_b_only": None,
                "mcnemar_exact_p": None}

    if metric.startswith("t2a"):
        xa, xb = a["_pq"]["t2a"], b["_pq"]["t2a"]
    else:
        xa, xb = a["_pq"]["a2t"], b["_pq"]["a2t"]
    k = int(metric.split("r@")[1])
    ha, hb = xa <= k, xb <= k
    n = len(ha)
    diff = float(hb.mean() - ha.mean())

    rng = np.random.default_rng(SEED)
    deltas = np.empty(N_BOOT)
    for i in range(N_BOOT):
        idx = rng.integers(0, n, n)
        deltas[i] = hb[idx].mean() - ha[idx].mean()
    lo, hi = float(np.percentile(deltas, 2.5)), float(np.percentile(deltas, 97.5))

    n01 = int((~ha & hb).sum())   # a misses, b hits
    n10 = int((ha & ~hb).sum())   # a hits, b misses
    m = n01 + n10
    # exact two-sided binomial under p=0.5
    if m == 0:
        p = 1.0
    else:
        tail = sum(math.comb(m, i) for i in range(0, min(n01, n10) + 1)) / 2 ** m
        p = min(1.0, 2 * tail)

    return {"a": a["model"], "b": b["model"], "metric": metric,
            "delta": diff, "ci95": [lo, hi], "n": n,
            "discordant_a_only": n10, "discordant_b_only": n01,
            "mcnemar_exact_p": p}


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("dir", nargs="?", default="results_final")
    ap.add_argument("--out", default="analysis.json")
    args = ap.parse_args()

    files = sorted(Path(args.dir).glob("*.json"))
    files = [f for f in files if "summary" not in f.name]
    if not files:
        print(f"no result JSONs in {args.dir}")
        return 1
    cells = []
    for f in files:
        d = json.loads(f.read_text(encoding="utf-8"))
        if "first_hit_rank" not in d["metrics"]["t2a"]:
            print(f"skip {f.name}: no per-query arrays (rerun with --per-query)")
            continue
        cells.append(cell_stats(d))

    pairs = []
    by_key = {}
    for c in cells:
        if c["audio_batch"] == 1 and c["trim_seconds"] == 10.0:
            by_key.setdefault(c["dataset"], {})[c["model"]] = c
    for ds, models in by_key.items():
        ids = sorted(models)
        if len(ids) == 2 and "clap-htsat-fused" in ids[0] + ids[1]:
            a, b = models[ids[0]], models[ids[1]]
            for metric in ("t2a r@1", "a2t r@1", "a2t map"):
                pairs.append(paired(a, b, metric))

    print(f"{'model':<28}{'ds':<10}{'metric':<10}{'value':>8}  95% CI")
    for c in cells:
        for k, v in c["metrics"].items():
            if k == "n_queries":
                continue
            print(f"{c['model']:<28}{c['dataset']:<10}{k:<10}{v['value']:>8.3f}  "
                  f"[{v['ci95'][0]:.3f}, {v['ci95'][1]:.3f}]")
    if pairs:
        print("\npaired comparisons (same fixture, resampled together)")
        for p in pairs:
            mcn = ("n/a" if p["mcnemar_exact_p"] is None
                   else f"{p['mcnemar_exact_p']:.3g}")
            disc = ("n/a" if p["discordant_a_only"] is None
                    else f"{p['discordant_a_only']}/{p['discordant_b_only']}")
            print(f"{p['a']} -> {p['b']} | {p['metric']}: delta {p['delta']:+.3f} "
                  f"[{p['ci95'][0]:+.3f}, {p['ci95'][1]:+.3f}], "
                  f"discordant {disc}, McNemar p={mcn}")

    # per-query arrays are large and only needed in memory
    slim = [{k: v for k, v in c.items() if k != "_pq"} for c in cells]
    Path(args.out).write_text(
        json.dumps({"cells": slim, "paired": pairs}, indent=2), encoding="utf-8")
    print(f"\nwrote {args.out}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())