code/train/build_vrbench_navigation_training_data.py

#!/usr/bin/env python3
"""Build SPAgent training data for three VRBench navigation environments.

The resulting prompts and images follow the FrozenLake v3 interaction contract.
Beyond-MMEnv is used only to sample official initial states; rollout transitions
and reward replay use ``spagent.vrbench_navigation_protocol``.
"""

from __future__ import annotations

import argparse
import json
import random
import sys
from concurrent.futures import ProcessPoolExecutor
from pathlib import Path
from typing import Any, Dict, Iterable, List, Sequence


REPO_ROOT = Path(__file__).resolve().parents[1]
THINK_WITH_IMAGES_ROOT = REPO_ROOT.parents[1]
DEFAULT_BEYOND_ROOT = THINK_WITH_IMAGES_ROOT / "Beyond-MMEnv-Scaling"
DEFAULT_OUTPUT_DIR = (
    REPO_ROOT / "outputs" / "grpo_converted" / "vrbench_navigation_v1"
)

if str(REPO_ROOT) not in sys.path:
    sys.path.insert(0, str(REPO_ROOT))

from spagent.vrbench_navigation_protocol import (  # noqa: E402
    MAZE_TASK,
    PATHFINDER_TASK,
    SUPPORTED_TASKS,
    TRAPFIELD_SOLVABLE_SEEDS,
    TRAPFIELD_TASK,
    build_system_prompt,
    build_user_prompt,
    canonical_state,
    encode_initial_state,
    render_state,
    shortest_path,
    state_text,
)

_SAMPLING_ENV: Any = None


def _import_gym_v(beyond_root: Path) -> Any:
    pool_root = beyond_root / "mmenv-pool"
    if not pool_root.is_dir():
        raise FileNotFoundError(f"Beyond-MMEnv mmenv-pool not found: {pool_root}")
    if str(pool_root) not in sys.path:
        sys.path.insert(0, str(pool_root))
    try:
        import gym_v
        import gym_v.envs  # noqa: F401
    except ImportError as exc:
        raise RuntimeError(
            "Could not import gym_v. Run this builder with the mmenv-pool "
            "virtualenv, for example:\n"
            f"  {pool_root}/.venv/bin/python {Path(__file__).resolve()}"
        ) from exc
    return gym_v


def extract_state(env: Any) -> Dict[str, Any]:
    base = env.unwrapped
    return canonical_state(
        grid=base._grid,
        start=(base._pr, base._pc),
        goal=(base._gr, base._gc),
    )


def sample_state(env: Any, seed: int) -> Dict[str, Any]:
    env.reset(seed=seed)
    return extract_state(env)


def task_slug(task_id: str) -> str:
    return {
        PATHFINDER_TASK: "pathfinder",
        MAZE_TASK: "maze",
        TRAPFIELD_TASK: "trapfield",
    }[task_id]


def seed_candidates(
    task_id: str,
    *,
    start_seed: int,
    max_seed_attempts: int,
) -> Iterable[int]:
    stop = start_seed + max_seed_attempts
    if task_id != TRAPFIELD_TASK:
        yield from range(start_seed, stop)
        return
    known = set()
    for seed in TRAPFIELD_SOLVABLE_SEEDS:
        if seed not in known:
            known.add(seed)
            yield seed
    for seed in range(start_seed, stop):
        if seed not in known:
            yield seed


def _init_sampling_worker(beyond_root: str, task_id: str) -> None:
    global _SAMPLING_ENV
    gym_v = _import_gym_v(Path(beyond_root))
    try:
        from loguru import logger

        logger.remove()
    except (ImportError, ValueError):
        pass
    _SAMPLING_ENV = gym_v.make(task_id)


def _sample_seed_batch(
    seed_start: int,
    seed_stop: int,
) -> tuple[int, int, List[tuple[int, Dict[str, Any], List[str]]], int]:
    if _SAMPLING_ENV is None:
        raise RuntimeError("Sampling worker environment was not initialized")
    solvable: List[tuple[int, Dict[str, Any], List[str]]] = []
    errors = 0
    for seed in range(seed_start, seed_stop):
        try:
            state = sample_state(_SAMPLING_ENV, seed)
            solution = shortest_path(state)
        except Exception:
            errors += 1
            continue
        if solution is not None:
            solvable.append((seed, state, solution))
    return seed_start, seed_stop, solvable, errors


def build_record(
    *,
    task_id: str,
    state: Dict[str, Any],
    solution: List[str],
    seed: int,
    index: int,
    output_dir: Path,
) -> Dict[str, Any]:
    slug = task_slug(task_id)
    image_path = output_dir / "images" / slug / f"{slug}_{index:06d}.png"
    render_state(task_id, state, image_path)
    return {
        "id": f"{slug}_{index:06d}",
        "task_id": task_id,
        "source_env": task_id,
        "source_seed": seed,
        "messages": [
            {
                "role": "system",
                "content": build_system_prompt(task_id, state),
            },
            {
                "role": "user",
                "content": build_user_prompt(task_id, state, image_path),
            },
        ],
        "images": [str(image_path.resolve())],
        "initial_state": encode_initial_state(state),
        "state_text": state_text(state, initial=True),
        "solution": " ".join(solution),
        "optimal_length": len(solution),
        "solvable": True,
        "deterministic": True,
        "fully_observed": True,
    }


def _build_record_job(
    task_id: str,
    state: Dict[str, Any],
    solution: List[str],
    seed: int,
    index: int,
    output_dir: str,
) -> Dict[str, Any]:
    return build_record(
        task_id=task_id,
        state=state,
        solution=solution,
        seed=seed,
        index=index,
        output_dir=Path(output_dir),
    )


def collect_records(
    *,
    gym_v: Any,
    task_id: str,
    count: int,
    start_seed: int,
    max_seed_attempts: int,
    output_dir: Path,
    workers: int,
    beyond_root: Path,
) -> tuple[List[Dict[str, Any]], int]:
    selected: List[tuple[int, Dict[str, Any], List[str]]] = []
    seen = set()
    scanned = 0
    errors = 0

    def consume(
        seed: int,
        state: Dict[str, Any],
        solution: List[str],
    ) -> None:
        encoded = encode_initial_state(state)
        if encoded in seen:
            return
        seen.add(encoded)
        selected.append((seed, state, solution))
        if len(selected) % 100 == 0 or len(selected) == count:
            print(
                f"[{task_id}] selected {len(selected)}/{count} "
                f"(scanned {scanned} seeds)",
                flush=True,
            )

    if workers == 1:
        env = gym_v.make(task_id)
        try:
            for seed in seed_candidates(
                task_id,
                start_seed=start_seed,
                max_seed_attempts=max_seed_attempts,
            ):
                if len(selected) >= count or scanned >= max_seed_attempts:
                    break
                scanned += 1
                try:
                    state = sample_state(env, seed)
                    solution = shortest_path(state)
                except Exception as exc:
                    errors += 1
                    print(
                        f"[skip] {task_id} seed={seed}: "
                        f"{type(exc).__name__}: {exc}",
                        flush=True,
                    )
                    continue
                if solution is not None:
                    consume(seed, state, solution)
        finally:
            env.close()
    else:
        next_seed = start_seed
        sample_batch_size = 256
        with ProcessPoolExecutor(
            max_workers=workers,
            initializer=_init_sampling_worker,
            initargs=(str(beyond_root), task_id),
        ) as executor:
            while len(selected) < count and scanned < max_seed_attempts:
                remaining = max_seed_attempts - scanned
                window = min(remaining, workers * sample_batch_size * 8)
                seed_stop = next_seed + window
                ranges = [
                    (
                        batch_start,
                        min(batch_start + sample_batch_size, seed_stop),
                    )
                    for batch_start in range(
                        next_seed,
                        seed_stop,
                        sample_batch_size,
                    )
                ]
                next_seed = seed_stop
                batches = executor.map(
                    _sample_seed_batch,
                    *(zip(*ranges)),
                    chunksize=1,
                )
                done = False
                for batch_start, batch_stop, candidates, batch_errors in batches:
                    scanned += batch_stop - batch_start
                    errors += batch_errors
                    for seed, state, solution in candidates:
                        if len(selected) >= count:
                            done = True
                            break
                        consume(seed, state, solution)
                    if done:
                        break
                    if scanned % 100000 < sample_batch_size:
                        print(
                            f"[{task_id}] scanned {scanned} seeds; "
                            f"selected {len(selected)}/{count}",
                            flush=True,
                        )
    if len(selected) != count:
        raise RuntimeError(
            f"Selected only {len(selected)}/{count} records for {task_id} "
            f"after {scanned} seed attempts ({errors} reset errors)"
        )
    jobs = [
        (
            task_id,
            state,
            solution,
            seed,
            index,
            str(output_dir),
        )
        for index, (seed, state, solution) in enumerate(selected)
    ]
    if workers == 1:
        records = [_build_record_job(*job) for job in jobs]
    else:
        with ProcessPoolExecutor(max_workers=workers) as executor:
            records = list(executor.map(_build_record_job, *zip(*jobs)))
    print(
        f"[{task_id}] rendered {len(records)}/{count} with {workers} workers",
        flush=True,
    )
    return records, scanned


def split_records(
    records: Sequence[Dict[str, Any]],
    *,
    test_ratio: float,
    split_seed: int,
) -> tuple[List[Dict[str, Any]], List[Dict[str, Any]]]:
    shuffled = list(records)
    random.Random(split_seed).shuffle(shuffled)
    if not shuffled:
        return [], []
    test_count = int(len(shuffled) * test_ratio)
    if test_ratio > 0 and len(shuffled) > 1:
        test_count = max(1, test_count)
    test_count = min(test_count, max(0, len(shuffled) - 1))
    return shuffled[test_count:], shuffled[:test_count]


def write_jsonl(path: Path, records: Iterable[Dict[str, Any]]) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    with path.open("w", encoding="utf-8") as handle:
        for record in records:
            handle.write(json.dumps(record, ensure_ascii=False) + "\n")


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument(
        "--beyond-root",
        type=Path,
        default=DEFAULT_BEYOND_ROOT,
    )
    parser.add_argument(
        "--output-dir",
        type=Path,
        default=DEFAULT_OUTPUT_DIR,
    )
    parser.add_argument(
        "--task",
        action="append",
        choices=SUPPORTED_TASKS,
        help="Task to build; repeat for multiple tasks. Defaults to all three.",
    )
    parser.add_argument("--count-per-task", type=int, default=2000)
    parser.add_argument("--test-ratio", type=float, default=0.0)
    parser.add_argument("--start-seed", type=int, default=0)
    parser.add_argument("--split-seed", type=int, default=0)
    parser.add_argument("--max-seed-attempts", type=int, default=1500000)
    parser.add_argument("--workers", type=int, default=16)
    return parser.parse_args()


def main() -> None:
    args = parse_args()
    if args.count_per_task <= 0:
        raise SystemExit("--count-per-task must be positive")
    if not 0 <= args.test_ratio < 1:
        raise SystemExit("--test-ratio must be in [0, 1)")
    if args.workers <= 0:
        raise SystemExit("--workers must be positive")
    tasks = args.task or list(SUPPORTED_TASKS)
    output_dir = args.output_dir.resolve()
    output_dir.mkdir(parents=True, exist_ok=True)
    gym_v = _import_gym_v(args.beyond_root.resolve())

    combined_train: List[Dict[str, Any]] = []
    combined_test: List[Dict[str, Any]] = []
    manifest_tasks: Dict[str, Any] = {}
    for task_offset, task_id in enumerate(tasks):
        records, scanned = collect_records(
            gym_v=gym_v,
            task_id=task_id,
            count=args.count_per_task,
            start_seed=args.start_seed,
            max_seed_attempts=args.max_seed_attempts,
            output_dir=output_dir,
            workers=args.workers,
            beyond_root=args.beyond_root.resolve(),
        )
        unique_initial_states = len(
            {str(record["initial_state"]) for record in records}
        )
        train, test = split_records(
            records,
            test_ratio=args.test_ratio,
            split_seed=args.split_seed + task_offset,
        )
        slug = task_slug(task_id)
        write_jsonl(output_dir / f"{slug}_all_v1.jsonl", records)
        write_jsonl(output_dir / f"{slug}_train_v1.jsonl", train)
        write_jsonl(output_dir / f"{slug}_test_v1.jsonl", test)
        combined_train.extend(train)
        combined_test.extend(test)
        manifest_tasks[task_id] = {
            "records": len(records),
            "train": len(train),
            "test": len(test),
            "unique_initial_states": unique_initial_states,
            "repeated_records": len(records) - unique_initial_states,
            "seed_attempts": scanned,
        }

    random.Random(args.split_seed).shuffle(combined_train)
    random.Random(args.split_seed + 1).shuffle(combined_test)
    write_jsonl(
        output_dir / "vrbench_navigation_train_v1.jsonl",
        combined_train,
    )
    write_jsonl(
        output_dir / "vrbench_navigation_test_v1.jsonl",
        combined_test,
    )
    manifest = {
        "format": "spagent-vrbench-navigation-v1",
        "tasks": manifest_tasks,
        "combined_train": len(combined_train),
        "combined_test": len(combined_test),
        "test_ratio": args.test_ratio,
        "sampling_workers": args.workers,
        "render_workers": args.workers,
        "fully_observed": True,
        "deterministic": True,
        "interaction_contract": "SPAgent/FrozenLake-v3",
    }
    (output_dir / "manifest.json").write_text(
        json.dumps(manifest, indent=2, ensure_ascii=False) + "\n",
        encoding="utf-8",
    )
    print(f"DONE: {output_dir}")
    print(json.dumps(manifest, indent=2, ensure_ascii=False))


if __name__ == "__main__":
    main()