scripts/score_by_difficulty.py

#!/usr/bin/env python
"""Score one predictions.jsonl against multiple difficulty subsets.

The bench script can be run once on the merged 10K test.jsonl (DIFFS=all),
saving 2× model-load time vs running per-difficulty. This script then
re-uses that single predictions.jsonl and scores it against {easy, medium,
hard} qa_roots separately by inner-joining on pair_id, so you still get
per-difficulty metrics with a single bench run.

Usage:
    python scripts/score_by_difficulty.py \\
        --predictions-jsonl <path>/predictions.jsonl \\
        --qa-parent /apdcephfs_cq10/.../all_qa_llm_by_difficulty_v2_filtered_balanced_v4_prompted \\
        --output-dir <path>/scores_by_diff/

Output:
    <output-dir>/easy_score.json
    <output-dir>/medium_score.json
    <output-dir>/hard_score.json
    <output-dir>/all_score.json   (full 10K combined)
"""
from __future__ import annotations

import argparse
import json
import os
import subprocess
import sys
from pathlib import Path


def collect_pair_ids(jsonl_path: Path) -> set[str]:
    ids: set[str] = set()
    with jsonl_path.open() as fh:
        for line in fh:
            line = line.strip()
            if not line:
                continue
            r = json.loads(line)
            pid = r.get("pair_id")
            if pid is not None:
                ids.add(str(pid))
    return ids


def filter_predictions(predictions_jsonl: Path, keep_ids: set[str], out_path: Path) -> int:
    n = 0
    out_path.parent.mkdir(parents=True, exist_ok=True)
    with predictions_jsonl.open() as fin, out_path.open("w") as fout:
        for line in fin:
            line = line.strip()
            if not line:
                continue
            r = json.loads(line)
            pid = r.get("pair_id")
            if pid is None or str(pid) not in keep_ids:
                continue
            fout.write(line + "\n")
            n += 1
    return n


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--predictions-jsonl", required=True)
    ap.add_argument("--qa-parent", required=True,
                    help="Parent dir containing easy/, medium/, hard/, and merged test.jsonl")
    ap.add_argument("--output-dir", required=True)
    ap.add_argument("--diffs", nargs="+", default=["easy", "medium", "hard", "all"])
    ap.add_argument("--score-script", default=None,
                    help="Path to score_test_predictions.py (default: alongside this script)")
    args = ap.parse_args()

    pred_path = Path(args.predictions_jsonl)
    if not pred_path.exists():
        print(f"missing {pred_path}", file=sys.stderr)
        return 1
    out_dir = Path(args.output_dir)
    out_dir.mkdir(parents=True, exist_ok=True)

    score_script = Path(args.score_script) if args.score_script \
        else Path(__file__).parent / "score_test_predictions.py"
    if not score_script.exists():
        print(f"missing scorer: {score_script}", file=sys.stderr)
        return 1

    qa_parent = Path(args.qa_parent)
    for diff in args.diffs:
        qa_dir = qa_parent if diff == "all" else qa_parent / diff
        qa_jsonl = qa_dir / "test.jsonl"
        if not qa_jsonl.exists():
            print(f"[skip] {diff}: missing {qa_jsonl}")
            continue

        # For 'all' we score the full predictions file directly. For per-diff
        # we filter predictions to only pair_ids present in that diff's qa,
        # so the joiner inside score_test_predictions has a clean inner-join.
        if diff == "all":
            sub_pred = pred_path
            n_sub = sum(1 for _ in pred_path.open() if _.strip())
        else:
            ids = collect_pair_ids(qa_jsonl)
            sub_pred = out_dir / f"{diff}_predictions.jsonl"
            n_sub = filter_predictions(pred_path, ids, sub_pred)
            print(f"[{diff}] {n_sub} predictions filtered into {sub_pred.name}")

        score_json = out_dir / f"{diff}_score.json"
        scored_jsonl = out_dir / f"{diff}_scored.jsonl"
        cmd = [
            sys.executable, str(score_script),
            "--predictions-jsonl", str(sub_pred),
            "--qa-root", str(qa_dir),
            "--split", "test",
            "--output-json", str(score_json),
            "--per-record-jsonl", str(scored_jsonl),
        ]
        print(f"[{diff}] scoring n={n_sub} -> {score_json.name}")
        subprocess.run(cmd, check=True)

    # Print summary table
    print()
    print(f"{'DIFF':<8} {'n':>6} {'overall':>8} {'detect_src':>12} {'detect_tm':>10} "
          f"{'azimuth':>8} {'elev':>8} {'identify':>10} {'count':>8}")
    print("-" * 80)
    for diff in args.diffs:
        score_json = out_dir / f"{diff}_score.json"
        if not score_json.exists():
            continue
        d = json.loads(score_json.read_text())
        ov = d.get("overall", {})
        per = d.get("per_task", {})
        def get(t):
            return per.get(t, {}).get("correct_all")
        cells = [
            f"{diff:<8}",
            f"{int(ov.get('n_total', 0)):>6d}",
            f"{ov.get('correct_all', 0):>7.3f} ",
            f"{(get('detect_source') or 0):>11.3f} ",
            f"{(get('detect_time') or 0):>9.3f} ",
            f"{(get('estimate_azimuth') or 0):>7.3f} ",
            f"{(get('estimate_elevation') or 0):>7.3f} ",
            f"{(get('identify_source_by_location') or get('identify_source_by_doa') or 0):>9.3f} ",
            f"{(get('count_sources') or 0):>7.3f} ",
        ]
        print("".join(cells))
    return 0


if __name__ == "__main__":
    sys.exit(main())