Blame for grader/grade.py

#!/usr/bin/env -S uv run --script
# /// script
# requires-python = ">=3.9"
# dependencies = ["numpy"]
# ///
"""Standalone grader for the P&W subtask benchmark. Python 3.9+ and numpy only.

    uv run grade.py --gold gold/rollouts/ --pred predictions/seen-policy/ [--pred baselines/terra-zero-shot/ ...]
                    [--tasks connect-dc-jack,...] [--split test|cross-validation] [--cohort rollout|demo]
                    [--common] [--resamples 1000] [--seg-iou 0.8 | --seg-tol 1] [--out results.json]

Files: every *.jsonl under each folder, one episode per line, in P&W's catalog field format:
    {"source": ..., "episode": ..., "task": ..., "subtasks": [{"startSec", "endSec", "label", "rating"}, ...], ...}
Episodes are keyed by (source, episode). Gold lines may also carry "split" and "cohort"; --split / --cohort filter on
the gold's values. A predicted line rates a segment "failed" or not; when "rating" is missing, pSuccess < 0.5 means failed.
An episode is scored for a predictor when it is in the gold (after the filters) and in that predictor's files.
With several --pred folders, deltas against the first are computed on the episodes all of them cover (paired), and
--common scores every predictor on that common set only.

Rules, applied identically to gold and predictions
-------------------------------------------------
1. Labels. P&W's two sort picks ("Pick up a single black screw from the table." and "... silver screw ...") are scored
   as one label, "Pick up a single screw from the table." (screw colour is not scored). --keep-colour turns this off.
2. Clean-up of each episode's list of segments, deterministic, in this order:
   a. segments with endSec <= startSec are dropped;
   b. segments are sorted by (startSec, endSec); ties keep the file's order;
   c. exact duplicates (same startSec, endSec and label) are dropped, keeping the first;
   d. each segment after the first starts where the previous one ends (startSec := previous endSec). This closes gaps
      (the following segment absorbs them) and resolves overlaps (the earlier segment keeps the overlapping time);
      a segment left with no time is dropped.
   Predictions are then clipped to the gold's span [first gold startSec, last gold endSec].
3. Ratings. Gold "failed" counts as failed; "ok", "excellent" (and a missing rating) count as success.

Metrics (pooled over the episodes of a task; the all-tasks row is the unweighted mean of the task values)
-------------------------------------------------------------------------------------------------------
- seg-F1@IoU0.8 (segment F1): adjacent segments with the same label are merged into one run (on both sides). This is
  the standard MS-TCN segmental F1: each predicted run, in time order, claims the unclaimed gold run of the same label
  with the highest IoU; it is a true positive if that IoU >= 0.8, else a false positive; unclaimed gold runs are false
  negatives. F1 = 2TP / (2TP + FP + FN). --seg-iou sets a different IoU threshold (the result key keeps its name).
- seg-F1@T s (--seg-tol T, replaces the IoU rule; the result key keeps its name): on the same merged runs, a predicted
  run and a gold run can match if they have the same label and both |start difference| and |end difference| are <= T
  seconds. Matching is one to one and greedy over all same-label pairs, smallest max(|start diff|, |end diff|) first
  (ties: lower gold index, then lower predicted index). TP = matched pairs, FP = unmatched predicted runs, FN = unmatched
  gold runs; F1 as above. This is the same rule as the internal "seg-F1@1 s" metric when T = 1.
- Boundary error (s): the boundaries are the change points between consecutive merged runs (episode start and end are not
  boundaries). Gold and predicted boundaries are matched one to one, closest first, if within 2 s of each other (labels
  ignored). The metric is the mean absolute time difference of the matched pairs. "boundary recall" is the share of gold
  boundaries matched, reported beside it so that a low error from few matches is visible; "boundary precision" is the
  share of predicted boundaries matched to a gold boundary.
- Failure F1 / precision / recall (attempt level): attempts are the cleaned segments, NOT merged (a retry is its own
  attempt). Predicted and gold attempts are matched one to one if they have the same label and IoU >= 0.5, greedily by
  IoU (ties: lower gold index, then lower predicted index). A predicted failed attempt matched to a gold failed attempt is
  a true positive; every other predicted failed attempt is a false positive; every other gold failed attempt is a false
  negative.
Output
------
JSON on stdout (or --out FILE), with these top-level keys:
- "results": {predictor folder: {task: {metric: {"value", "lo", "hi"}}, "all": {...}}}. Metrics: seg_f1_iou80 (segment
  F1 at the chosen --seg-iou or --seg-tol), boundary_err_s, boundary_recall, boundary_precision, failed_f1, failed_precision,
  failed_recall, plus "n" (episodes scored). F1, precision and recall are fractions in [0, 1]; boundary_err_s is in seconds. "lo"/"hi" are the 95% CI.
- "deltas": {"<pred> - <first pred>": same layout}, paired differences on the common episodes (only with several --pred).
- "results_common": same layout as "results", every predictor scored on the common episodes (with several --pred).
- "coverage": {predictor: {task: {"gold", "scored", "common"}}}, episode counts per task.
- "episodes": {task: n}, the episodes scored per task (the common set when deltas or --common apply).
- "options": the settings used (tasks, split, cohort, resamples, seed, merge_labels, common, seg_iou or seg_tol).

Examples
--------
    uv run grade.py --gold gold/rollouts --pred predictions/seen-policy --split test           # our model, test rollouts
    uv run grade.py --gold gold/rollouts --pred predictions/seen-policy --pred baselines/terra-zero-shot --split test --common
    uv run grade.py --gold gold/rollouts --pred predictions/new-family                        # each family held out
    uv run grade.py --gold gold/rollouts --pred predictions/seen-policy --split test --seg-tol 1 --out tol1.json
    uv run test_grade.py                                                                       # unit tests

Confidence intervals: 95% percentile intervals from an episode bootstrap (--resamples draws, seeded per task), resampling
episodes within each task. Deltas use the same draws for both predictors (paired), on the common episodes.
"""
from __future__ import annotations

import argparse
import json
import sys
import zlib
from pathlib import Path

import numpy as np

SORT_PICKS = ("Pick up a single black screw from the table.", "Pick up a single silver screw from the table.")
SORT_PICK = "Pick up a single screw from the table."
LABEL_GROUPS = {SORT_PICKS[0]: SORT_PICK, SORT_PICKS[1]: SORT_PICK}
SEG_IOU = 0.8
FAIL_IOU = 0.5
BOUNDARY_WINDOW = 2.0
TOL = 1e-6
METRICS = ("seg_f1_iou80", "boundary_err_s", "boundary_recall", "boundary_precision", "failed_f1", "failed_precision", "failed_recall")
COUNTS = ("seg_tp", "seg_fp", "seg_fn", "b_err", "b_matched", "b_gold", "b_pred", "f_tp", "f_fp", "f_fn")


def key(line: dict) -> tuple:
    return (line["source"], line["episode"])


def _failed(seg: dict, is_pred: bool) -> bool:
    rating = seg.get("rating")
    if rating is None and is_pred and seg.get("pSuccess") is not None:
        return float(seg["pSuccess"]) < 0.5
    return rating == "failed"


def normalise(subtasks: list, merge_labels: bool = True, is_pred: bool = False) -> list:
    """[(start, end, label, failed)] after rules 1-2d of the module docstring."""
    rows = []
    for i, s in enumerate(subtasks):
        a, b = float(s["startSec"]), float(s["endSec"])
        if b - a <= TOL:
            continue
        label = LABEL_GROUPS.get(s["label"], s["label"]) if merge_labels else s["label"]
        rows.append((a, b, i, label, _failed(s, is_pred), s["label"]))
    rows.sort(key=lambda r: (r[0], r[1], r[2]))
    seen, out = set(), []
    for a, b, _, label, failed, raw in rows:
        dup = (round(a, 6), round(b, 6), raw)
        if dup in seen:
            continue
        seen.add(dup)
        if out:
            a = out[-1][1]
            if b - a <= TOL:
                continue
        out.append((a, b, label, failed))
    return out


def clip(segs: list, lo: float, hi: float) -> list:
    out = []
    for a, b, label, failed in segs:
        a, b = max(a, lo), min(b, hi)
        if b - a > TOL:
            out.append((a, b, label, failed))
    return out


def runs(segs: list) -> list:
    """Adjacent same-label segments merged: [(start, end, label, any failed)]."""
    out = []
    for a, b, label, failed in segs:
        if out and out[-1][2] == label:
            out[-1] = (out[-1][0], b, label, out[-1][3] or failed)
        else:
            out.append((a, b, label, failed))
    return out


def _iou(x, y) -> float:
    inter = min(x[1], y[1]) - max(x[0], y[0])
    if inter <= 0:
        return 0.0
    return inter / (max(x[1], y[1]) - min(x[0], y[0]))


def seg_counts(gold_runs: list, pred_runs: list, threshold: float = SEG_IOU) -> tuple:
    used = [False] * len(gold_runs)
    tp = fp = 0
    for p in pred_runs:
        best, index = 0.0, -1
        for i, g in enumerate(gold_runs):
            if g[2] == p[2] and not used[i]:
                v = _iou(g, p)
                if v > best:
                    best, index = v, i
        if index >= 0 and best >= threshold - TOL:
            used[index] = True
            tp += 1
        else:
            fp += 1
    return tp, fp, used.count(False)


def seg_counts_tol(gold_runs: list, pred_runs: list, tol: float) -> tuple:
    pairs = sorted((max(abs(g[0] - p[0]), abs(g[1] - p[1])), i, j)
                   for i, g in enumerate(gold_runs) for j, p in enumerate(pred_runs) if g[2] == p[2])
    gi, pj = set(), set()
    for d, i, j in pairs:
        if d > tol + TOL:
            break
        if i not in gi and j not in pj:
            gi.add(i)
            pj.add(j)
    return len(gi), len(pred_runs) - len(gi), len(gold_runs) - len(gi)


def boundary_errors(gold_runs: list, pred_runs: list, window: float = BOUNDARY_WINDOW) -> tuple:
    """(matched absolute errors, number of gold boundaries, number of predicted boundaries)."""
    g = [r[1] for r in gold_runs[:-1]]
    p = [r[1] for r in pred_runs[:-1]]
    pairs = sorted((abs(x - y), i, j) for i, x in enumerate(g) for j, y in enumerate(p) if abs(x - y) <= window + TOL)
    gi, pj, errors = set(), set(), []
    for d, i, j in pairs:
        if i not in gi and j not in pj:
            gi.add(i)
            pj.add(j)
            errors.append(d)
    return errors, len(g), len(p)


def match_attempts(pred: list, gold: list, threshold: float = FAIL_IOU) -> list:
    pairs = sorted((-_iou(p, g), j, i) for i, p in enumerate(pred) for j, g in enumerate(gold) if p[2] == g[2] and _iou(p, g) >= threshold - TOL)
    up, ug, out = set(), set(), []
    for _, j, i in pairs:
        if i not in up and j not in ug:
            up.add(i)
            ug.add(j)
            out.append((i, j))
    return out


def episode_counts(gold_line: dict, pred_line: dict, merge_labels: bool = True, seg_iou: float = SEG_IOU, seg_tol: float | None = None) -> dict:
    gold = normalise(gold_line["subtasks"], merge_labels)
    if not gold:
        return None
    pred = clip(normalise(pred_line["subtasks"], merge_labels, is_pred=True), gold[0][0], gold[-1][1])
    gr, pr = runs(gold), runs(pred)
    tp, fp, fn = seg_counts(gr, pr, seg_iou) if seg_tol is None else seg_counts_tol(gr, pr, seg_tol)
    errors, n_gold_b, n_pred_b = boundary_errors(gr, pr)
    flagged = [i for i, a in enumerate(pred) if a[3]]
    failed = {j for j, a in enumerate(gold) if a[3]}
    hits = sum(1 for i, j in match_attempts(pred, gold) if pred[i][3] and j in failed)
    return {"seg_tp": tp, "seg_fp": fp, "seg_fn": fn, "b_err": float(sum(errors)), "b_matched": len(errors), "b_gold": n_gold_b, "b_pred": n_pred_b,
            "f_tp": hits, "f_fp": len(flagged) - hits, "f_fn": len(failed) - hits}


def _ratio(a, b):
    a, b = np.asarray(a, dtype=float), np.asarray(b, dtype=float)
    with np.errstate(divide="ignore", invalid="ignore"):
        return np.where(b > 0, a / np.where(b > 0, b, 1), np.nan)


def metrics_of(c: np.ndarray) -> dict:
    """c: [..., len(COUNTS)] summed counts -> {metric: array}."""
    k = {name: c[..., i] for i, name in enumerate(COUNTS)}
    return {"seg_f1_iou80": _ratio(2 * k["seg_tp"], 2 * k["seg_tp"] + k["seg_fp"] + k["seg_fn"]),
            "boundary_err_s": _ratio(k["b_err"], k["b_matched"]),
            "boundary_recall": _ratio(k["b_matched"], k["b_gold"]),
            "boundary_precision": _ratio(k["b_matched"], k["b_pred"]),
            "failed_f1": _ratio(2 * k["f_tp"], 2 * k["f_tp"] + k["f_fp"] + k["f_fn"]),
            "failed_precision": _ratio(k["f_tp"], k["f_tp"] + k["f_fp"]),
            "failed_recall": _ratio(k["f_tp"], k["f_tp"] + k["f_fn"])}


def _draws(task: str, n: int, resamples: int, seed: int) -> np.ndarray:
    rng = np.random.default_rng([seed, zlib.crc32(task.encode())])
    return rng.integers(0, n, size=(resamples, n))


def _summary(point: float, draws) -> dict:
    out = {"value": None if point is None or np.isnan(point) else float(point)}
    if draws is not None:
        d = np.asarray(draws, dtype=float)
        d = d[~np.isnan(d)]
        out["lo"] = float(np.percentile(d, 2.5)) if len(d) else None
        out["hi"] = float(np.percentile(d, 97.5)) if len(d) else None
    return out


def grade_lines(gold_lines: list, preds: dict, tasks: list | None = None, split: str | None = None, cohort: str | None = None,
                resamples: int = 1000, seed: int = 0, merge_labels: bool = True, common: bool = False, seg_iou: float = SEG_IOU,
                seg_tol: float | None = None) -> dict:
    """preds: {name: [lines]}, the first is the reference for deltas. Returns {"coverage", "results": {name: {task: {metric: {value, lo, hi}}, "all": ...}},
    "deltas": {name: {task: {metric: ...}}}, "episodes"}."""
    gold = {}
    for x in gold_lines:
        if tasks and x["task"] not in tasks:
            continue
        if split and x.get("split") != split:
            continue
        if cohort and x.get("cohort") != cohort:
            continue
        if key(x) in gold:
            raise ValueError(f"duplicate gold episode {key(x)}")
        gold[key(x)] = x
    by_pred = {}
    for name, lines in preds.items():
        d = {}
        for x in lines:
            if key(x) in d:
                raise ValueError(f"{name}: duplicate predicted episode {key(x)}")
            d[key(x)] = x
        by_pred[name] = d
    names = list(by_pred)
    task_list = sorted({x["task"] for x in gold.values()})
    common_keys = {t: sorted(k for k, x in gold.items() if x["task"] == t and all(k in by_pred[n] for n in names)) for t in task_list}
    out = {"coverage": {}, "results": {}, "deltas": {}, "episodes": {}, "options": {"tasks": tasks, "split": split, "cohort": cohort, "resamples": resamples,
                                                                                  "seed": seed, "merge_labels": merge_labels, "common": common,
                                                                                  "seg_iou": seg_iou}}
    if seg_tol is not None:
        out["options"]["seg_tol"] = seg_tol
    counts_cache = {}

    def counts(name, keys):
        rows = []
        for k in keys:
            ck = (name, k)
            if ck not in counts_cache:
                c = episode_counts(gold[k], by_pred[name][k], merge_labels, seg_iou, seg_tol)
                counts_cache[ck] = None if c is None else [c[n] for n in COUNTS]
            if counts_cache[ck] is not None:
                rows.append(counts_cache[ck])
        return np.asarray(rows, dtype=float).reshape(-1, len(COUNTS))

    def score(name, keys_by_task):
        res, draws_by_task = {}, {}
        for t in task_list:
            keys = keys_by_task[t]
            c = counts(name, keys)
            if not len(c):
                continue
            point = metrics_of(c.sum(0))
            draws = None
            if resamples:
                idx = _draws(t, len(c), resamples, seed)
                draws = metrics_of(c[idx].sum(1))
                draws_by_task[t] = draws
            res[t] = {m: _summary(point[m], None if draws is None else draws[m]) for m in METRICS}
            res[t]["n"] = len(c)
        if res:
            res["all"] = {m: _summary(float(np.nanmean([res[t][m]["value"] if res[t][m]["value"] is not None else np.nan for t in res])),
                                      np.nanmean(np.stack([draws_by_task[t][m] for t in draws_by_task]), axis=0) if resamples and draws_by_task else None) for m in METRICS}
            res["all"]["n"] = sum(res[t]["n"] for t in task_list if t in res)
            res["all"]["tasks"] = sum(1 for t in task_list if t in res)
        return res, draws_by_task

    for name in names:
        own = {t: sorted(k for k, x in gold.items() if x["task"] == t and k in by_pred[name]) for t in task_list}
        out["coverage"][name] = {t: {"gold": sum(1 for x in gold.values() if x["task"] == t), "scored": len(own[t]), "common": len(common_keys[t])} for t in task_list}
        out["results"][name], _ = score(name, common_keys if common else own)
    out["episodes"] = {t: len(common_keys[t]) for t in task_list}
    if len(names) > 1:
        ref_point = {}
        ref_draws = {}
        for name in names:
            res, dr = score(name, common_keys)
            ref_point[name], ref_draws[name] = res, dr
        base = names[0]
        for name in names[1:]:
            delta = {}
            for t in task_list:
                if t not in ref_point[name] or t not in ref_point[base]:
                    continue
                delta[t] = {}
                for m in METRICS:
                    a, b = ref_point[name][t][m]["value"], ref_point[base][t][m]["value"]
                    d = None if a is None or b is None else a - b
                    dd = (ref_draws[name][t][m] - ref_draws[base][t][m]) if resamples else None
                    delta[t][m] = _summary(np.nan if d is None else d, dd)
                delta[t]["n"] = ref_point[name][t]["n"]
            if delta:
                ts = [t for t in task_list if t in delta]
                delta["all"] = {}
                for m in METRICS:
                    vals = [delta[t][m]["value"] for t in ts if delta[t][m]["value"] is not None]
                    dd = np.nanmean(np.stack([ref_draws[name][t][m] - ref_draws[base][t][m] for t in ts]), axis=0) if resamples else None
                    delta["all"][m] = _summary(float(np.mean(vals)) if vals else np.nan, dd)
                delta["all"]["n"] = sum(delta[t]["n"] for t in ts)
            out["deltas"][f"{name} - {base}"] = delta
            out["results_common"] = {n: ref_point[n] for n in names}
    return out


def read_folder(path: str) -> list:
    p = Path(path)
    files = sorted(p.rglob("*.jsonl")) if p.is_dir() else [p]
    lines = []
    for f in files:
        with open(f) as fh:
            lines += [json.loads(x) for x in fh if x.strip()]
    return lines


def _fmt(s: dict, metric: str) -> str:
    if s.get("value") is None:
        return "-"
    scale, digits = (1, 2) if metric == "boundary_err_s" else (100, 1)
    v = f"{s['value'] * scale:.{digits}f}"
    if s.get("lo") is not None:
        v += f" [{s['lo'] * scale:.{digits}f}, {s['hi'] * scale:.{digits}f}]"
    return v


def main(argv=None):
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--gold", required=True)
    ap.add_argument("--pred", required=True, action="append", help="a predictions folder or file; repeat to compare (deltas against the first)")
    ap.add_argument("--tasks", default="", help="comma-separated task names (default: all in the gold)")
    ap.add_argument("--split", default=None, help="test or cross-validation (the gold's split field)")
    ap.add_argument("--cohort", default=None, help="rollout or demo (the gold's cohort field)")
    ap.add_argument("--common", action="store_true", help="score every predictor on the episodes all of them cover")
    ap.add_argument("--resamples", type=int, default=1000)
    ap.add_argument("--seed", type=int, default=0)
    ap.add_argument("--seg-iou", type=float, default=SEG_IOU, help="IoU threshold for segment F1 (default 0.8)")
    ap.add_argument("--seg-tol", type=float, default=None, metavar="SECONDS",
                    help="segment F1 with a boundary tolerance in seconds instead of the IoU rule (see the docstring)")
    ap.add_argument("--keep-colour", action="store_true", help="score the black and silver sort picks as different labels")
    ap.add_argument("--out", default=None, help="write the JSON results to this file instead of stdout")
    a = ap.parse_args(argv)
    preds = {p.rstrip("/"): read_folder(p) for p in a.pred}
    res = grade_lines(read_folder(a.gold), preds, [t for t in a.tasks.split(",") if t] or None, a.split, a.cohort, a.resamples, a.seed, not a.keep_colour, a.common, a.seg_iou, a.seg_tol)
    text = json.dumps(res, indent=1)
    if a.out:
        Path(a.out).write_text(text + "\n")
    else:
        print(text)
    return 0


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