train_example.py

#!/usr/bin/env python3
"""
Train LeRobot imitation-learning policies on an SO-101 dataset with W&B logging.

Install:
  pip install lerobot huggingface_hub wandb torch pyyaml
  pip install 'lerobot[diffusion]'  # needed for diffusion policy configs

Login once:
  huggingface-cli login   # optional for public dataset, useful generally
  wandb login

Run:
  python train_example.py --config-name act_arch_wide_192
  python train_example.py --config-name act_arch_wide_192 \
    --dataset-repos user/dataset_a user/dataset_b --aggregate-repo user/a_b_merged
  python train_example.py --config-name act_arch_wide_192_hil_finetune

Fine-tuning:
  Set `pretrained_policy_path` to a previous `pretrained_model` directory. This
  starts a new run initialized from that checkpoint. It is different from
  LeRobot `resume`, which continues the exact same interrupted run.
"""

from __future__ import annotations

import argparse
import copy
import hashlib
import os
import shutil
import subprocess
from pathlib import Path
from typing import Any


DEFAULT_DATASET_REPO = "Smencomojica/robotics_class_2"
DEFAULT_CONFIG_FILE = Path(__file__).with_name("train_configs.yaml")


def default_lerobot_home() -> Path:
    hf_home = Path(os.environ.get("HF_HOME", Path.home() / ".cache" / "huggingface"))
    return Path(os.environ.get("HF_LEROBOT_HOME", hf_home / "lerobot"))


def dataset_local_dir(repo_id: str) -> Path:
    return default_lerobot_home() / repo_id


def default_aggregate_repo_id(dataset_repos: list[str]) -> str:
    digest = hashlib.sha1("\n".join(dataset_repos).encode("utf-8")).hexdigest()[:8]
    return f"local/merged_{digest}"


def looks_downloaded(path: Path) -> bool:
    return (
        path.exists()
        and (path / "meta").exists()
        and any(path.rglob("*.parquet"))
    )


def normalize_dataset_repos(value: Any) -> list[str]:
    if value is None:
        return []
    if isinstance(value, str):
        return [value]
    if isinstance(value, list) and all(isinstance(item, str) for item in value):
        return value
    raise ValueError("`dataset_repos` must be a string or a list of strings.")


def normalize_string_list(value: Any, field_name: str) -> list[str]:
    if value is None:
        return []
    if isinstance(value, str):
        return [value]
    if isinstance(value, list) and all(isinstance(item, str) for item in value):
        return value
    raise ValueError(f"`{field_name}` must be a string or a list of strings.")


def pick_device() -> str:
    import torch

    if torch.cuda.is_available():
        return "cuda"
    if getattr(torch.backends, "mps", None) and torch.backends.mps.is_available():
        return "mps"
    return "cpu"


def resolve_config_file(path: str | Path) -> Path:
    config_file = Path(path)
    if config_file.exists():
        return config_file

    script_relative = Path(__file__).parent / config_file
    if script_relative.exists():
        return script_relative

    raise FileNotFoundError(f"Could not find config file: {path}")


def load_sweep_configs(config_file: Path) -> dict[str, Any]:
    try:
        import yaml
    except ModuleNotFoundError as exc:
        raise RuntimeError(
            "PyYAML is required for --config-file support. Install it with `pip install pyyaml`."
        ) from exc

    with config_file.open("r", encoding="utf-8") as f:
        data = yaml.safe_load(f) or {}

    if not isinstance(data, dict):
        raise ValueError(f"{config_file} must contain a YAML mapping.")
    if "configs" not in data or not isinstance(data["configs"], dict):
        raise ValueError(f"{config_file} must define a `configs` mapping.")

    defaults = data.get("defaults", {})
    if defaults is None:
        defaults = {}
    if not isinstance(defaults, dict):
        raise ValueError(f"{config_file} `defaults` must be a mapping.")

    policy_defaults = data.get("policy_defaults", {})
    if policy_defaults is None:
        policy_defaults = {}
    if not isinstance(policy_defaults, dict):
        raise ValueError(f"{config_file} `policy_defaults` must be a mapping.")

    return {
        "defaults": defaults,
        "policy_defaults": policy_defaults,
        "configs": data["configs"],
    }


def deep_merge(base: dict[str, Any], override: dict[str, Any]) -> dict[str, Any]:
    merged = copy.deepcopy(base)
    for key, value in override.items():
        if isinstance(value, dict) and isinstance(merged.get(key), dict):
            merged[key] = deep_merge(merged[key], value)
        else:
            merged[key] = copy.deepcopy(value)
    return merged


def config_by_name(config_file: Path, config_name: str) -> dict[str, Any]:
    data = load_sweep_configs(config_file)
    configs = data["configs"]
    if config_name not in configs:
        available = ", ".join(sorted(configs))
        raise ValueError(
            f"Unknown config name: {config_name}. Available configs: {available}"
        )

    selected = configs[config_name]
    if selected is None:
        selected = {}
    if not isinstance(selected, dict):
        raise ValueError(f"Config `{config_name}` must be a mapping.")

    selected_policy = selected.get("policy", {})
    if selected_policy is None:
        selected_policy = {}
    if not isinstance(selected_policy, dict):
        raise ValueError(f"Config `{config_name}` field `policy` must be a mapping.")

    defaults_policy = data["defaults"].get("policy", {})
    if defaults_policy is None:
        defaults_policy = {}
    if not isinstance(defaults_policy, dict):
        raise ValueError(f"{config_file} `defaults.policy` must be a mapping.")

    policy_type = selected_policy.get("type") or defaults_policy.get("type") or "act"
    policy_defaults = data["policy_defaults"].get(policy_type, {})
    if policy_defaults is None:
        policy_defaults = {}
    if not isinstance(policy_defaults, dict):
        raise ValueError(
            f"{config_file} `policy_defaults.{policy_type}` must be a mapping."
        )

    cfg = deep_merge(data["defaults"], policy_defaults)
    return deep_merge(cfg, selected)


def str_value(value: Any) -> str:
    if isinstance(value, bool):
        return str(value).lower()
    return str(value)


def add_cli_arg(cmd: list[str], key: str, value: Any) -> None:
    if value is None:
        return
    cmd.append(f"--{key}={str_value(value)}")


def looks_like_local_path(value: str) -> bool:
    expanded = Path(value).expanduser()
    if expanded.is_absolute():
        return True
    if value.startswith((".", "~")):
        return True
    # Hugging Face model IDs are commonly "user/repo". Longer slash-separated
    # values are much more likely to be local paths such as outputs/train/...
    return len(expanded.parts) > 2


def validate_pretrained_policy_path(value: Any) -> str | None:
    if value is None:
        return None
    if not isinstance(value, str):
        raise ValueError("`pretrained_policy_path` must be a string.")

    path_value = value.strip()
    if not path_value:
        raise ValueError("`pretrained_policy_path` must not be empty.")

    candidate = Path(path_value).expanduser()
    if candidate.exists() or looks_like_local_path(path_value):
        required_files = ("config.json", "model.safetensors")
        missing = [name for name in required_files if not (candidate / name).is_file()]
        if missing:
            missing_list = ", ".join(missing)
            raise FileNotFoundError(
                f"`pretrained_policy_path` must point to a LeRobot `pretrained_model` "
                f"directory containing {missing_list}: {candidate}"
            )

    return path_value


def apply_cli_overrides(cfg: dict[str, Any], args: argparse.Namespace) -> dict[str, Any]:
    overridden = copy.deepcopy(cfg)
    for key in (
        "dataset_repo",
        "aggregate_repo",
        "output_dir",
        "job_name",
        "steps",
        "batch_size",
        "device",
        "wandb_project",
        "log_freq",
        "save_freq",
        "policy_repo_id",
        "pretrained_policy_path",
    ):
        value = getattr(args, key)
        if value is not None:
            overridden[key] = value
    if args.dataset_repos is not None:
        overridden["dataset_repos"] = args.dataset_repos
    if args.aggregate_drop_features is not None:
        overridden["aggregate_drop_features"] = args.aggregate_drop_features
    return overridden


def download_dataset_if_needed(repo_id: str) -> Path:
    local_dir = dataset_local_dir(repo_id)
    if looks_downloaded(local_dir):
        print(f"Dataset already present: {local_dir}")
        return local_dir

    print(f"Downloading dataset {repo_id} to {local_dir}")
    from huggingface_hub import snapshot_download

    local_dir.parent.mkdir(parents=True, exist_ok=True)
    snapshot_download(
        repo_id=repo_id,
        repo_type="dataset",
        local_dir=str(local_dir),
        local_dir_use_symlinks=False,
    )
    return local_dir


def sanitize_dataset_for_aggregation(
    repo_id: str,
    root: Path,
    drop_features: list[str],
    aggregate_repo: str,
    source_index: int,
) -> tuple[str, Path]:
    if not drop_features:
        return repo_id, root

    from lerobot.datasets.dataset_tools import remove_feature
    from lerobot.datasets.lerobot_dataset import LeRobotDataset

    dataset = LeRobotDataset(repo_id, root=root)
    features_to_drop = [
        feature_name for feature_name in drop_features if feature_name in dataset.meta.features
    ]
    if not features_to_drop:
        return repo_id, root

    sanitized_repo_id = f"{aggregate_repo}_source_{source_index:02d}_sanitized"
    sanitized_root = dataset_local_dir(sanitized_repo_id)

    if looks_downloaded(sanitized_root):
        print(f"Sanitized dataset already present: {sanitized_root}")
        return sanitized_repo_id, sanitized_root

    if sanitized_root.exists():
        raise FileExistsError(
            f"Sanitized dataset path exists but does not look complete: {sanitized_root}. "
            "Remove the incomplete directory or choose a different `aggregate_repo`."
        )

    print(
        f"Creating sanitized copy of {repo_id} without features: "
        + ", ".join(features_to_drop)
    )
    remove_feature(
        dataset=dataset,
        feature_names=features_to_drop,
        output_dir=sanitized_root,
        repo_id=sanitized_repo_id,
    )
    return sanitized_repo_id, sanitized_root


def prepare_training_dataset(cfg: dict[str, Any]) -> tuple[str, Path]:
    dataset_repos = normalize_dataset_repos(cfg.get("dataset_repos"))
    dataset_repo = cfg.get("dataset_repo", DEFAULT_DATASET_REPO)

    if not dataset_repos:
        dataset_repos = [dataset_repo]

    if len(dataset_repos) == 1:
        repo_id = dataset_repos[0]
        return repo_id, download_dataset_if_needed(repo_id)

    source_roots = [download_dataset_if_needed(repo_id) for repo_id in dataset_repos]
    aggregate_repo = cfg.get("aggregate_repo") or default_aggregate_repo_id(dataset_repos)
    aggregate_root = dataset_local_dir(aggregate_repo)
    drop_features = normalize_string_list(
        cfg.get("aggregate_drop_features"), "aggregate_drop_features"
    )

    if looks_downloaded(aggregate_root):
        print(f"Aggregated dataset already present: {aggregate_root}")
        return aggregate_repo, aggregate_root

    if aggregate_root.exists():
        raise FileExistsError(
            f"Aggregate dataset path exists but does not look complete: {aggregate_root}. "
            "Choose a different `aggregate_repo` or remove the incomplete directory."
        )

    aggregate_source_repos = []
    aggregate_source_roots = []
    for source_index, (repo_id, root) in enumerate(zip(dataset_repos, source_roots)):
        sanitized_repo_id, sanitized_root = sanitize_dataset_for_aggregation(
            repo_id=repo_id,
            root=root,
            drop_features=drop_features,
            aggregate_repo=aggregate_repo,
            source_index=source_index,
        )
        aggregate_source_repos.append(sanitized_repo_id)
        aggregate_source_roots.append(sanitized_root)

    print("Aggregating datasets:")
    for repo_id in aggregate_source_repos:
        print(f"  {repo_id}")
    print(f"Aggregate repo id: {aggregate_repo}")
    print(f"Aggregate local dir: {aggregate_root}")

    from lerobot.datasets.aggregate import aggregate_datasets

    aggregate_datasets(
        repo_ids=aggregate_source_repos,
        roots=aggregate_source_roots,
        aggr_repo_id=aggregate_repo,
        aggr_root=aggregate_root,
    )
    return aggregate_repo, aggregate_root


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--config-file", default=str(DEFAULT_CONFIG_FILE))
    parser.add_argument("--config-name")
    parser.add_argument("--list-configs", action="store_true")
    parser.add_argument("--dataset-repo", default=None)
    parser.add_argument("--dataset-repos", nargs="+", default=None)
    parser.add_argument("--aggregate-repo", default=None)
    parser.add_argument("--aggregate-drop-features", nargs="+", default=None)
    parser.add_argument("--output-dir", default=None)
    parser.add_argument("--job-name", default=None)
    parser.add_argument("--steps", type=int, default=None)
    parser.add_argument("--batch-size", type=int, default=None)
    parser.add_argument("--device", default=None)
    parser.add_argument("--wandb-project", default=None)
    parser.add_argument("--log-freq", type=int, default=None)
    parser.add_argument("--save-freq", type=int, default=None)
    parser.add_argument("--policy-repo-id", default=None)
    parser.add_argument("--pretrained-policy-path", default=None)
    args = parser.parse_args()

    config_file = resolve_config_file(args.config_file)
    sweep_data = load_sweep_configs(config_file)

    if args.list_configs:
        print("Available configs:")
        for name in sorted(sweep_data["configs"]):
            print(f"  {name}")
        return

    if not args.config_name:
        available = ", ".join(sorted(sweep_data["configs"]))
        parser.error(f"--config-name is required. Available configs: {available}")

    cfg = apply_cli_overrides(config_by_name(config_file, args.config_name), args)

    output_dir = cfg.get("output_dir", "outputs/train/act_so101_lcc")
    job_name = cfg.get("job_name", args.config_name)
    steps = cfg.get("steps", 20_000)
    batch_size = cfg.get("batch_size", 32)
    device = cfg.get("device") or pick_device()
    wandb_project = cfg.get("wandb_project", "lerobot-so101-act-lcc")
    log_freq = cfg.get("log_freq", 100)
    save_freq = cfg.get("save_freq", 4_000)
    policy_repo_id = cfg.get("policy_repo_id")
    pretrained_policy_path = validate_pretrained_policy_path(
        cfg.get("pretrained_policy_path")
    )
    policy_cfg = cfg.get("policy", {})

    if not isinstance(policy_cfg, dict):
        raise ValueError(f"Config `{args.config_name}` field `policy` must be a mapping.")

    print(f"Selected config: {args.config_name}")
    print("Selected device:", device)

    if shutil.which("lerobot-train") is None:
        raise RuntimeError(
            "Could not find `lerobot-train`. Install LeRobot first: pip install lerobot"
        )

    dataset_repo, local_dir = prepare_training_dataset(cfg)

    os.environ.setdefault("WANDB_PROJECT", wandb_project)

    cmd = [
        "lerobot-train",
        f"--dataset.repo_id={dataset_repo}",
        f"--dataset.root={local_dir}",
        f"--output_dir={output_dir}",
        f"--job_name={job_name}",
        f"--policy.device={device}",
        "--wandb.enable=true",
        "--policy.push_to_hub=false",
        f"--steps={steps}",
        f"--batch_size={batch_size}",
        f"--log_freq={log_freq}",
        f"--save_freq={save_freq}",
    ]

    for key, value in policy_cfg.items():
        add_cli_arg(cmd, f"policy.{key}", value)

    if pretrained_policy_path:
        cmd.append(f"--policy.pretrained_path={pretrained_policy_path}")

    if policy_repo_id:
        cmd.append(f"--policy.repo_id={policy_repo_id}")

    print("\nRunning:\n  " + " \\\n  ".join(cmd) + "\n")
    subprocess.run(cmd, check=True)


if __name__ == "__main__":
    main()