VLABench/scripts/run_evaluation_cli.py

#!/usr/bin/env python3
"""Command-line wrapper for tutorials/5.run_evaluation.ipynb."""

from __future__ import annotations

import argparse
import json
import os
import sys
from importlib import import_module
from pathlib import Path


def _resolve_path(path_like: str) -> str:
    return str(Path(path_like).expanduser().resolve())


def _resolve_episode_config_path(
    path_like: str, project_root: Path, vlabench_root: Path
) -> str:
    """Resolve custom episode-config arguments.

    New downstream tasks may ship configs outside of this repo structure.
    To keep the CLI flexible, we attempt a few reasonable fallbacks:
      1) interpret the argument relative to the current working directory;
      2) interpret it relative to the declared project root;
      3) interpret it relative to ``VLABENCH_ROOT``;
      4) if it looks like a track name (e.g. ``chem01``), look under
         ``<VLABENCH_ROOT>/configs/evaluation/tracks`` (with and without
         a ``.json`` suffix).
    """

    project_root = project_root.expanduser().resolve()
    vlabench_root = vlabench_root.expanduser().resolve()
    provided = Path(path_like).expanduser()

    candidates: list[Path] = []

    def _register(path: Path) -> None:
        if path in candidates:
            return
        candidates.append(path)

    if provided.is_absolute():
        _register(provided)
    else:
        _register(Path.cwd() / provided)
        _register(project_root / provided)
        _register(vlabench_root / provided)

    track_dir = vlabench_root / "configs" / "evaluation" / "tracks"
    track_names: list[str] = []
    base_name = provided.name
    if provided.suffix:
        track_names.append(base_name)
    else:
        track_names.extend([base_name, f"{base_name}.json"])
    for name in track_names:
        _register(track_dir / name)

    for candidate in candidates:
        if candidate.exists():
            return str(candidate.resolve())

    searched = "\n  - ".join(str(path) for path in candidates)
    raise FileNotFoundError(
        f"Could not find episode config '{path_like}'. Tried:\n  - {searched}"
    )


def _prepare_environment(project_root: str, vlabench_root: str | None, mujoco_gl: str | None) -> None:
    project_root_path = Path(project_root).expanduser().resolve()
    vlabench_root_path = (
        project_root_path / "VLABench"
        if vlabench_root is None
        else Path(vlabench_root).expanduser().resolve()
    )

    if not project_root_path.exists():
        raise FileNotFoundError(f"project_root does not exist: {project_root_path}")
    if not vlabench_root_path.exists():
        raise FileNotFoundError(f"VLABENCH_ROOT does not exist: {vlabench_root_path}")

    if str(project_root_path) not in sys.path:
        sys.path.append(str(project_root_path))

    src_root = project_root_path / "src"
    if src_root.exists():
        for path in [src_root, *src_root.iterdir()]:
            if path.is_dir():
                path_str = str(path)
                if path_str not in sys.path:
                    sys.path.append(path_str)

    os.environ.setdefault("VLABENCH_ROOT", str(vlabench_root_path))
    if mujoco_gl:
        os.environ["MUJOCO_GL"] = mujoco_gl


def _load_json_arg(value: str | None) -> dict:
    if not value:
        return {}
    candidate = Path(value)
    payload = candidate.read_text() if candidate.exists() else value
    return json.loads(payload)


def _normalize_task_names(task_args: list[str]) -> list[str]:
    normalized = []
    for name in task_args:
        normalized.append(name.split("/")[-1])
    return normalized


def _run_policy_eval(args: argparse.Namespace) -> None:
    from VLABench.evaluation.evaluator import Evaluator
    from VLABench.evaluation.model.policy.base import RandomPolicy

    project_root = Path(args.project_root).expanduser().resolve()
    env_vlabench_root = os.environ.get("VLABENCH_ROOT")
    vlabench_root = (
        Path(env_vlabench_root).expanduser().resolve()
        if env_vlabench_root
        else project_root / "VLABench"
    )

    tasks = _normalize_task_names(args.tasks)
    save_dir = _resolve_path(args.save_dir)
    episode_config = (
        _resolve_episode_config_path(args.episode_config, project_root, vlabench_root)
        if args.episode_config
        else None
    )

    policy_name = args.policy.lower()
    if policy_name == "random":
        policy = RandomPolicy(model=None)
    elif policy_name == "openvla":
        from VLABench.evaluation.model.policy.openvla import OpenVLA

        if not args.model_ckpt:
            raise ValueError("--model-ckpt is required for OpenVLA evaluation")
        if not args.lora_ckpt:
            raise ValueError("--lora-ckpt is required for OpenVLA evaluation")
        norm_config = args.norm_config or os.path.join(
            os.environ["VLABENCH_ROOT"], "configs", "model", "openvla_config.json"
        )
        policy = OpenVLA(
            model_ckpt=_resolve_path(args.model_ckpt),
            lora_ckpt=_resolve_path(args.lora_ckpt),
            norm_config_file=_resolve_path(norm_config),
            device=args.device,
            debug_actions=args.debug_actions,
        )
    elif policy_name == "nora":
        from VLABench.evaluation.model.policy.nora import NoraPolicy

        if not args.model_ckpt:
            raise ValueError("--model-ckpt is required for Nora evaluation")
        policy = NoraPolicy(
            model_ckpt=_resolve_path(args.model_ckpt),
            device=args.device,
            time_horizon=max(1, args.nora_time_horizon),
            debug_tokens=getattr(args, "debug_tokens", False),
            camera_index=args.nora_camera_index,
            action_mode=args.nora_action_mode,
            normalize_gripper=not args.nora_skip_gripper_normalize,
            binarize_gripper=not args.nora_no_gripper_binarize,
            invert_gripper=args.nora_invert_gripper,
            gripper_threshold=args.nora_gripper_threshold,
            lerobot_dataset=args.nora_lerobot_dataset,
        )
    else:
        raise ValueError(f"Unsupported policy: {args.policy}")

    evaluator = Evaluator(
        tasks=tasks,
        n_episodes=args.n_episodes,
        episode_config=episode_config,
        max_substeps=args.max_substeps,
        save_dir=save_dir,
        visulization=args.visualize,
        eval_unseen=args.eval_unseen,
        unnorm_key=args.unnorm_key,
        intention_score_threshold=args.intention_threshold,
    )

    metrics = evaluator.evaluate(policy)
    print(json.dumps(metrics, indent=2, ensure_ascii=False))


def _run_vlm_eval(args: argparse.Namespace) -> None:
    from VLABench.evaluation.evaluator import VLMEvaluator

    vlm_module = import_module("VLABench.evaluation.model.vlm")
    vlm_cls = getattr(vlm_module, args.vlm_name, None)
    if vlm_cls is None:
        raise ValueError(
            f"Unknown VLM '{args.vlm_name}'. Check VLABench.evaluation.model.vlm for valid class names."
        )

    init_kwargs = _load_json_arg(args.vlm_init)
    vlm = vlm_cls(**init_kwargs)

    evaluator = VLMEvaluator(
        tasks=args.vlm_tasks,
        n_episodes=args.n_episodes,
        data_path=_resolve_path(args.data_path),
        save_path=_resolve_path(args.save_path),
        language=args.language,
    )

    evaluator.evaluate(
        vlm,
        task_list=args.vlm_tasks,
        few_shot_num=args.few_shot_num,
        with_CoT=args.with_cot,
        eval_dim=args.eval_dim,
    )
    scores = evaluator.get_final_score_dict(
        vlm_name=vlm.name,
        few_shot_num=args.few_shot_num,
        with_CoT=args.with_cot,
    )
    if scores is not None:
        print(json.dumps(scores, indent=2, ensure_ascii=False))


def build_parser() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(description="Run VLABench policy or VLM evaluations from CLI")
    parser.add_argument(
        "--project-root",
        default=str(Path(__file__).resolve().parents[1]),
        help="Repository root (default: %(default)s)",
    )
    parser.add_argument(
        "--vlabench-root",
        default=None,
        help="Path to importable VLABench package (default: <project-root>/VLABench)",
    )
    parser.add_argument(
        "--mujoco-gl",
        default=None,
        help="Override MUJOCO_GL (set egl for headless)",
    )

    subparsers = parser.add_subparsers(dest="command", required=True)

    policy_parser = subparsers.add_parser("policy", help="Evaluate action policies (VLA)")
    policy_parser.add_argument("--tasks", nargs="+", required=True, help="Task names or task_series/task_name")
    policy_parser.add_argument("--n-episodes", type=int, default=2, help="Episodes per task")
    policy_parser.add_argument("--max-substeps", type=int, default=10, help="Env substeps per action")
    policy_parser.add_argument(
        "--save-dir",
        default=str(Path("./logs/policy_eval")),
        help="Directory for evaluator outputs",
    )
    policy_parser.add_argument(
        "--policy",
        default="random",
        choices=["random", "openvla", "nora"],
        help="Policy backend",
    )
    policy_parser.add_argument("--model-ckpt", default=None, help="OpenVLA base checkpoint path")
    policy_parser.add_argument("--lora-ckpt", default=None, help="OpenVLA LoRA checkpoint path")
    policy_parser.add_argument("--norm-config", default=None, help="Optional OpenVLA normalization config")
    policy_parser.add_argument("--device", default="cuda", help="Device for OpenVLA")
    policy_parser.add_argument(
        "--debug-actions",
        action="store_true",
        help="Enable verbose OpenVLA action debug logs (norm stats + per-step deltas)",
    )
    policy_parser.add_argument("--episode-config", default=None, help="Episode config JSON path")
    policy_parser.add_argument("--visualize", action="store_true", help="Store rollout videos")
    policy_parser.add_argument("--eval-unseen", action="store_true", help="Use unseen-category flag")
    policy_parser.add_argument("--unnorm-key", default="primitive", help="Normalization key for policies")
    policy_parser.add_argument(
        "--intention-threshold",
        type=float,
        default=0.1,
        help="Intention score threshold",
    )
    policy_parser.add_argument(
        "--nora-time-horizon",
        type=int,
        default=1,
        help="Nora action horizon / replan steps",
    )
    policy_parser.add_argument(
        "--nora-camera-index",
        type=int,
        default=2,
        help="Camera index fed to Nora (default front view)",
    )
    policy_parser.add_argument(
        "--nora-action-mode",
        choices=["delta", "absolute"],
        default="delta",
        help="Interpret Nora outputs as delta or absolute poses",
    )
    policy_parser.add_argument(
        "--nora-skip-gripper-normalize",
        action="store_true",
        help="Skip [0,1]->[-1,1] gripper normalization",
    )
    policy_parser.add_argument(
        "--nora-no-gripper-binarize",
        action="store_true",
        help="Keep Nora gripper logits continuous",
    )
    policy_parser.add_argument(
        "--nora-invert-gripper",
        action="store_true",
        help="Flip Nora gripper sign after normalization",
    )
    policy_parser.add_argument(
        "--nora-gripper-threshold",
        type=float,
        default=0.1,
        help="Threshold to decide open/close state",
    )
    policy_parser.add_argument(
        "--nora-lerobot-dataset",
        default=None,
        help="Optional Lerobot dataset name for Unnormalize",
    )
    policy_parser.add_argument(
        "--debug-tokens",
        action="store_true",
        help="(Nora) Print decoded action tokens for debugging",
    )

    repo_root = Path(__file__).resolve().parents[1]
    vlm_parser = subparsers.add_parser("vlm", help="Evaluate vision-language models")
    vlm_parser.add_argument(
        "--vlm-name",
        required=True,
        help="Class exported by VLABench.evaluation.model.vlm (e.g. GPT_4v, Qwen2_VL)",
    )
    vlm_parser.add_argument(
        "--vlm-init",
        default=None,
        help="JSON string or file providing constructor kwargs (API keys, etc.)",
    )
    vlm_parser.add_argument(
        "--vlm-tasks",
        nargs="+",
        required=True,
        help="Tasks in <task_series>/<task_name> format",
    )
    vlm_parser.add_argument("--few-shot-num", type=int, default=0, help="Few-shot example count")
    vlm_parser.add_argument("--with-cot", action="store_true", help="Enable chain-of-thought prompting")
    vlm_parser.add_argument("--n-episodes", type=int, default=2, help="Episodes per task")
    vlm_parser.add_argument(
        "--data-path",
        default=str(repo_root / "dataset" / "vlm"),
        help="Directory containing rendered VLM data",
    )
    vlm_parser.add_argument(
        "--save-path",
        default=str(repo_root / "logs" / "vlm"),
        help="Directory to store VLM outputs",
    )
    vlm_parser.add_argument("--language", choices=["en", "zh"], default="en", help="Prompt language")
    vlm_parser.add_argument("--eval-dim", default="default", help="Evaluation dimension key")

    return parser


def main(argv: list[str] | None = None) -> None:
    parser = build_parser()
    args = parser.parse_args(argv)

    _prepare_environment(args.project_root, args.vlabench_root, args.mujoco_gl)

    if args.command == "policy":
        _run_policy_eval(args)
    elif args.command == "vlm":
        _run_vlm_eval(args)
    else:
        parser.error("Unknown command; use 'policy' or 'vlm'.")


if __name__ == "__main__":
    main()