eval_ia_bench.py

"""Evaluate predicted interacted-object boxes against IA-bench ground truth.

Reports the paper's selective-prediction metrics for the predicted *start*
(`initial_object_box`) and *target* (`target_object_box`) boxes:

    acc@IoU    fraction of samples with IoU(pred, gt) >= --iou-threshold
    AUROC      ranking quality of the confidence score (higher better)
    AURC       area under risk-coverage curve (lower better)
    E-AURC     excess AURC vs the optimal selector (lower better)
    cov@90     coverage retained while keeping >=90% accuracy
    cov@95     coverage retained while keeping >=95% accuracy
    R@90/R@95  recall at those accuracy targets

Standalone: needs only numpy (+ pandas for .parquet predictions, and
huggingface_hub if you pass --gt-repo). No dependency on the annotation repo.

Ground truth is read from the IA-bench metadata.jsonl files (boxes only, no
video decode needed):

    # local build / clone:
    python eval_ia_bench.py --predictions preds.jsonl --gt-dir /path/to/IA-bench
    # straight from the Hub:
    python eval_ia_bench.py --predictions preds.jsonl --gt-repo irl-kit/IA-bench

Predictions file (.jsonl or .parquet), one row per (trajectory, subtask):
    {"dataset": "bridge_lerobot", "trajectory_name": "0/41905", "subtask_index": 0,
     "initial_object_box": [x1,y1,x2,y2], "target_object_box": [x1,y1,x2,y2],
     "score": 0.87}
`dataset` may be omitted when --dataset names a single config. A per-box
`initial_score`/`target_score` overrides `score` for that box. GT samples with
no prediction count as incorrect at the lowest score (full-benchmark coverage).
"""
from __future__ import annotations

import argparse
import json
from pathlib import Path

import numpy as np

# ---------------------------------------------------------------------------
# Metrics (vendored verbatim from the paper's scoring code)
# ---------------------------------------------------------------------------


def iou(a, b) -> float:
    a = np.asarray(a, dtype=float)
    b = np.asarray(b, dtype=float)
    ix1, iy1 = max(a[0], b[0]), max(a[1], b[1])
    ix2, iy2 = min(a[2], b[2]), min(a[3], b[3])
    inter = max(0.0, ix2 - ix1) * max(0.0, iy2 - iy1)
    area_a = max(0.0, a[2] - a[0]) * max(0.0, a[3] - a[1])
    area_b = max(0.0, b[2] - b[0]) * max(0.0, b[3] - b[1])
    union = area_a + area_b - inter
    return inter / union if union > 0 else 0.0


def _risk_coverage(pred_scores, correct_flags):
    pred_scores = np.asarray(pred_scores, dtype=float)
    correct_flags = np.asarray(correct_flags, dtype=bool)
    valid = ~np.isnan(pred_scores)
    pred_scores, correct_flags = pred_scores[valid], correct_flags[valid]
    n = len(pred_scores)
    if n == 0:
        return None, None, 0
    order = np.argsort(pred_scores)[::-1]
    cum_correct = np.cumsum(correct_flags[order].astype(float))
    counts = np.arange(1, n + 1, dtype=float)
    risk = 1.0 - cum_correct / counts
    return risk, counts, int(np.sum(correct_flags))


def compute_aurc(pred_scores, correct_flags) -> float:
    risk, _counts, _ = _risk_coverage(pred_scores, correct_flags)
    return float(np.mean(risk)) if risk is not None else np.nan


def compute_eaurc(pred_scores, correct_flags) -> float:
    risk, counts, n_correct = _risk_coverage(pred_scores, correct_flags)
    if risk is None:
        return np.nan
    n = len(counts)
    aurc = float(np.mean(risk))
    opt = np.zeros(n, dtype=float)
    opt[:n_correct] = 1.0
    aurc_opt = float(np.mean(1.0 - np.cumsum(opt) / counts))
    return aurc - aurc_opt


def compute_auroc(pred_scores, correct_flags) -> float:
    pred_scores = np.asarray(pred_scores, dtype=float)
    correct_flags = np.asarray(correct_flags, dtype=bool)
    valid = ~np.isnan(pred_scores)
    pred_scores, correct_flags = pred_scores[valid], correct_flags[valid]
    n_pos = int(np.sum(correct_flags))
    n_neg = len(correct_flags) - n_pos
    if n_pos == 0 or n_neg == 0:
        return np.nan
    order = np.argsort(pred_scores)[::-1]
    sorted_labels = correct_flags[order]
    tpr = np.concatenate([[0.0], np.cumsum(sorted_labels) / n_pos])
    fpr = np.concatenate([[0.0], np.cumsum(~sorted_labels) / n_neg])
    trapz = getattr(np, "trapezoid", None) or np.trapz
    return float(trapz(tpr, fpr))


def compute_recall_at_target_accuracy(pred_scores, correct_flags, target_accuracy=0.9):
    pred_scores = np.asarray(pred_scores, dtype=float)
    correct_flags = np.asarray(correct_flags, dtype=bool)
    valid = ~np.isnan(pred_scores)
    pred_scores, correct_flags = pred_scores[valid], correct_flags[valid]
    total_correct = int(np.sum(correct_flags))
    empty = {"recall": 0.0, "coverage": 0.0, "threshold": np.nan, "kept_accuracy": np.nan}
    if len(pred_scores) == 0 or total_correct == 0:
        return empty
    order = np.argsort(pred_scores)[::-1]
    sorted_scores = pred_scores[order]
    cum_correct = np.cumsum(correct_flags[order].astype(int))
    counts = np.arange(1, len(sorted_scores) + 1)
    acc = cum_correct / counts
    recall = cum_correct / total_correct
    valid_pos = acc >= target_accuracy
    if not np.any(valid_pos):
        return empty
    best = int(np.argmax(np.where(valid_pos, recall, -np.inf)))
    return {
        "recall": float(recall[best]),
        "coverage": float(counts[best] / len(sorted_scores)),
        "threshold": float(sorted_scores[best]),
        "kept_accuracy": float(acc[best]),
    }


# ---------------------------------------------------------------------------
# Data loading
# ---------------------------------------------------------------------------


def _read_jsonl(path: Path) -> list[dict]:
    with Path(path).open() as f:
        return [json.loads(line) for line in f if line.strip()]


def load_predictions(path: Path) -> list[dict]:
    path = Path(path)
    if path.suffix == ".parquet":
        import pandas as pd

        return pd.read_parquet(path).to_dict("records")
    return _read_jsonl(path)


def load_gt(gt_dir: Path | None, gt_repo: str | None, datasets, splits) -> dict:
    """Return {(dataset, trajectory_name, subtask_index): row} from metadata.jsonl."""
    if gt_repo:
        from huggingface_hub import snapshot_download

        gt_dir = Path(snapshot_download(
            repo_id=gt_repo, repo_type="dataset", allow_patterns=["data/*/*/metadata.jsonl"]
        ))
    if gt_dir is None:
        raise SystemExit("provide --gt-dir or --gt-repo")
    gt_dir = Path(gt_dir)

    gt = {}
    for meta in sorted(gt_dir.glob("data/*/*/metadata.jsonl")):
        ds, split = meta.parts[-3], meta.parts[-2]
        if datasets and ds not in datasets:
            continue
        if splits and split not in splits:
            continue
        for row in _read_jsonl(meta):
            gt[(ds, row["trajectory_name"], int(row["subtask_index"]))] = row
    if not gt:
        raise SystemExit(f"no GT metadata found under {gt_dir}/data/*/*/metadata.jsonl")
    return gt


# ---------------------------------------------------------------------------
# Evaluation
# ---------------------------------------------------------------------------

BOX_FIELDS = {"initial": "initial_object_box", "target": "target_object_box"}


def _pred_index(preds, default_dataset):
    idx = {}
    for p in preds:
        ds = p.get("dataset") or default_dataset
        if ds is None:
            raise SystemExit("predictions need a 'dataset' field (or pass --dataset <config>)")
        idx[(ds, p["trajectory_name"], int(p["subtask_index"]))] = p
    return idx


def evaluate(gt, preds, iou_threshold, default_dataset):
    pred_idx = _pred_index(preds, default_dataset)
    # group keys by dataset (+ an 'all' bucket)
    buckets: dict[str, list] = {}
    for key in gt:
        buckets.setdefault(key[0], []).append(key)
        buckets.setdefault("all", []).append(key)

    results = {}
    for box_name, gt_field in BOX_FIELDS.items():
        for bucket, keys in buckets.items():
            scores, correct = [], []
            n_pred = 0
            for key in keys:
                gt_box = gt[key].get(gt_field)
                p = pred_idx.get(key)
                if p is None or p.get(gt_field) is None or gt_box is None:
                    scores.append(-np.inf)   # unpredicted → ranked last, counts as wrong
                    correct.append(False)
                    continue
                n_pred += 1
                score = p.get(f"{box_name}_score", p.get("score", 1.0))
                scores.append(float(score) if score is not None else -np.inf)
                correct.append(iou(p[gt_field], gt_box) >= iou_threshold)
            scores = np.asarray(scores, dtype=float)
            correct = np.asarray(correct, dtype=bool)
            r90 = compute_recall_at_target_accuracy(scores, correct, 0.90)
            r95 = compute_recall_at_target_accuracy(scores, correct, 0.95)
            results[(box_name, bucket)] = {
                "n": len(keys),
                "n_pred": n_pred,
                "acc": float(np.mean(correct)) if len(correct) else np.nan,
                "AUROC": compute_auroc(scores, correct),
                "AURC": compute_aurc(scores, correct),
                "E-AURC": compute_eaurc(scores, correct),
                "cov@90": r90["coverage"],
                "cov@95": r95["coverage"],
                "R@90": r90["recall"],
                "R@95": r95["recall"],
            }
    return results


def _fmt(v):
    if isinstance(v, float):
        return "nan" if np.isnan(v) else f"{v:.3f}"
    return str(v)


def print_table(results, box_name):
    cols = ["n", "n_pred", "acc", "AUROC", "AURC", "E-AURC", "cov@90", "cov@95", "R@90", "R@95"]
    rows = sorted(k[1] for k in results if k[0] == box_name and k[1] != "all")
    rows.append("all")
    print(f"\n=== {box_name} box (IoU correctness) ===")
    print(f"{'dataset':16s} " + " ".join(f"{c:>7s}" for c in cols))
    for bucket in rows:
        m = results[(box_name, bucket)]
        print(f"{bucket:16s} " + " ".join(f"{_fmt(m[c]):>7s}" for c in cols))


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--predictions", required=True, type=Path)
    ap.add_argument("--gt-dir", type=Path, default=None, help="IA-bench root (local build/clone)")
    ap.add_argument("--gt-repo", type=str, default=None, help="HF dataset id, e.g. irl-kit/IA-bench")
    ap.add_argument("--dataset", nargs="*", default=None,
                    help="restrict to these dataset configs (also default 'dataset' for preds)")
    ap.add_argument("--split", nargs="*", default=None, help="restrict to these splits")
    ap.add_argument("--iou-threshold", type=float, default=0.4)
    ap.add_argument("--out", type=Path, default=None, help="optional JSON dump of all metrics")
    args = ap.parse_args()

    gt = load_gt(args.gt_dir, args.gt_repo, set(args.dataset or []), set(args.split or []))
    preds = load_predictions(args.predictions)
    default_dataset = args.dataset[0] if args.dataset and len(args.dataset) == 1 else None
    results = evaluate(gt, preds, args.iou_threshold, default_dataset)

    print(f"GT samples: {len(gt)} | predictions: {len(preds)} | IoU>={args.iou_threshold}")
    print_table(results, "initial")
    print_table(results, "target")

    if args.out:
        dump = {f"{box}/{bucket}": m for (box, bucket), m in results.items()}
        args.out.write_text(json.dumps(dump, indent=2))
        print(f"\nwrote {args.out}")


if __name__ == "__main__":
    main()