scripts/run_baselines.py

#!/usr/bin/env python3
"""Train the interpretable acoustic baseline or sweep fixed silence policies."""

from __future__ import annotations

import argparse
import hashlib
import json
import sys
from dataclasses import replace
from pathlib import Path

ROOT = Path(__file__).resolve().parents[1]
SRC = ROOT / "src"
if str(SRC) not in sys.path:
    sys.path.insert(0, str(SRC))


def _path(value: str) -> Path:
    candidate = Path(value)
    return candidate if candidate.is_absolute() else ROOT / candidate


def _sha256(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for block in iter(lambda: handle.read(1024 * 1024), b""):
            digest.update(block)
    return digest.hexdigest()


def _portable_path(path: Path) -> str:
    """Prefer a repository-relative path without rejecting external outputs."""

    resolved = path.resolve()
    try:
        return resolved.relative_to(ROOT.resolve()).as_posix()
    except ValueError:
        return str(resolved)


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    subparsers = parser.add_subparsers(dest="command", required=True)

    audio = subparsers.add_parser("audio", help="fit waveform-statistics logistic regression")
    audio.add_argument("--manifest", default="data/processed/partial-iid-splits.jsonl")
    audio.add_argument(
        "--source-root",
        default=".",
        help="base for source_file paths stored in the audited manifest",
    )
    audio.add_argument("--train-split", default="train")
    audio.add_argument("--validation-split", default="validation")
    audio.add_argument(
        "--revision",
        default="e564e2ac567f774d1880aa1db6ce97afb8c519b7",
        help="immutable upstream revision represented by the manifest",
    )
    audio.add_argument("--output", default="artifacts/partial-baseline")
    audio.add_argument(
        "--summary-report",
        help="optional second location for the compact metrics JSON (for tracked reports)",
    )
    audio.add_argument("--max-examples", type=int, help="debug cap per split")
    audio.add_argument("--max-seconds", type=float, default=4.0)
    audio.add_argument("--epochs", type=int, default=800)
    audio.add_argument("--learning-rate", type=float, default=0.05)
    audio.add_argument("--l2", type=float, default=1e-3)
    audio.add_argument("--fpr-budget", type=float, default=0.02)

    timeout = subparsers.add_parser("timeouts", help="evaluate fixed VAD silence timeouts")
    timeout.add_argument("--input", required=True, help="PauseCheckpoint JSONL")
    timeout.add_argument("--output", required=True)
    timeout.add_argument(
        "--timeouts-ms",
        default="200,400,600,800,1200",
        help="comma-separated timeout policies",
    )
    return parser.parse_args()


def _audio(args: argparse.Namespace) -> int:
    from turn_detection.baselines import extract_manifest_features, fit_logistic_baseline
    from turn_detection.data import read_manifest, write_json, write_manifest
    from turn_detection.training.metrics import (
        binary_classification_metrics,
        reliability_bins,
        threshold_at_max_fpr,
    )

    manifest = _path(args.manifest)
    source_root = _path(args.source_root)
    output = _path(args.output)
    output.mkdir(parents=True, exist_ok=True)
    all_rows = list(read_manifest(manifest))
    train_rows = [row for row in all_rows if row.get("split") == args.train_split]
    validation_rows = [row for row in all_rows if row.get("split") == args.validation_split]
    if not train_rows or not validation_rows:
        raise SystemExit("both requested train and validation splits must be non-empty")
    train_groups = {row.get("group_id") for row in train_rows if row.get("group_id")}
    validation_groups = {row.get("group_id") for row in validation_rows if row.get("group_id")}
    overlap = train_groups & validation_groups
    if overlap:
        raise SystemExit(f"refusing leaky manifest: {len(overlap)} group IDs cross splits")
    all_groups = train_groups | validation_groups
    observed_group_rows: dict[str, int] = {}
    for row in all_rows:
        group_id = row.get("group_id")
        if group_id:
            observed_group_rows[str(group_id)] = observed_group_rows.get(str(group_id), 0) + 1
    revisions = {
        str(row["source_revision"]) for row in all_rows if row.get("source_revision") is not None
    }

    train_x, train_y, train_info = extract_manifest_features(
        train_rows,
        source_root=source_root,
        max_examples=args.max_examples,
        max_seconds=args.max_seconds,
    )
    validation_x, validation_y, validation_info = extract_manifest_features(
        validation_rows,
        source_root=source_root,
        max_examples=args.max_examples,
        max_seconds=args.max_seconds,
    )
    model = fit_logistic_baseline(
        train_x,
        train_y,
        epochs=args.epochs,
        learning_rate=args.learning_rate,
        l2=args.l2,
    )
    validation_probabilities = model.predict_proba(validation_x)
    selected = threshold_at_max_fpr(
        validation_y.tolist(),
        validation_probabilities.tolist(),
        max_false_positive_rate=args.fpr_budget,
    )
    model = replace(model, threshold=float(selected["threshold"]))
    train_probabilities = model.predict_proba(train_x)

    model_path = output / "model.json"
    write_json(model_path, model.to_dict())
    report = {
        "status": "development_only",
        "scope": "audited local training shard; official test remains sealed",
        "manifest": args.manifest,
        "manifest_sha256": _sha256(manifest),
        "model": {
            "path": _portable_path(model_path),
            "bytes": model_path.stat().st_size,
            "sha256": _sha256(model_path),
        },
        "data_revision": next(iter(revisions)) if len(revisions) == 1 else args.revision,
        "source_root": args.source_root,
        "train_split": args.train_split,
        "validation_split": args.validation_split,
        "train_examples": int(len(train_y)),
        "validation_examples": int(len(validation_y)),
        "group_overlap_count": 0,
        "group_count": len(all_groups),
        "largest_group_rows": max(observed_group_rows.values(), default=0),
        "grouping_note": (
            "All observed groups are singleton rows; speaker/conversation/voice identity "
            "separation is not established."
            if observed_group_rows and max(observed_group_rows.values()) == 1
            else "Groups contain repeated observed linkage keys."
        ),
        "hyperparameters": {
            "epochs": args.epochs,
            "learning_rate": args.learning_rate,
            "l2": args.l2,
            "max_seconds": args.max_seconds,
            "fpr_budget": args.fpr_budget,
        },
        "threshold_selection": selected,
        "train_metrics_at_validation_threshold": binary_classification_metrics(
            train_y.tolist(), train_probabilities.tolist(), model.threshold
        ),
        "validation_metrics": binary_classification_metrics(
            validation_y.tolist(), validation_probabilities.tolist(), model.threshold
        ),
        "validation_reliability": reliability_bins(
            validation_y.tolist(), validation_probabilities.tolist()
        ),
        "limitations": [
            "One of 83 upstream training shards was locally available.",
            "This is an interpretable sanity baseline, not the submitted neural model.",
            "The threshold was selected on this validation split and must not be tuned on test.",
        ],
    }
    write_json(output / "metrics.json", report)
    if args.summary_report:
        write_json(_path(args.summary_report), report)
    predictions = []
    for info, probability in zip(validation_info, validation_probabilities.tolist(), strict=True):
        predictions.append(
            {
                **info,
                "probability": float(probability),
                "prediction": int(probability >= model.threshold),
                "threshold": model.threshold,
            }
        )
    write_manifest(output / "validation_predictions.jsonl", predictions)
    # Keep an auditable list of the exact training IDs without copying raw audio.
    write_manifest(output / "train_examples.jsonl", train_info)
    print(
        json.dumps({"model": model.to_dict(), "validation": report["validation_metrics"]}, indent=2)
    )
    return 0


def _timeouts(args: argparse.Namespace) -> int:
    from turn_detection.baselines import (
        fixed_timeout_sweep,
        load_checkpoints_jsonl,
    )
    from turn_detection.data import write_json

    try:
        timeouts = [float(value.strip()) for value in args.timeouts_ms.split(",") if value.strip()]
    except ValueError as exc:
        raise SystemExit("--timeouts-ms must contain comma-separated numbers") from exc
    if not timeouts:
        raise SystemExit("provide at least one fixed timeout")
    checkpoints = load_checkpoints_jsonl(_path(args.input))
    results = fixed_timeout_sweep(checkpoints, timeouts)
    output = _path(args.output)
    write_json(
        output,
        {
            "baseline": "fixed_silence_timeout",
            "checkpoint_source": args.input,
            "results": results,
        },
    )
    print(json.dumps(results, indent=2))
    return 0


def main() -> int:
    args = parse_args()
    return _audio(args) if args.command == "audio" else _timeouts(args)


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