scripts/combine_datasets.py

"""Combine multiple LeRobot v2 datasets into a single dataset.

Combines three g1_procedural_room_navigation datasets by:
- Re-indexing episodes across all source datasets
- Copying parquet files with updated episode_id columns
- Symlinking video files to originals
- Merging episodes.jsonl with new indices
- Creating origin.yaml to track provenance

Properly handles chunk boundaries: when the new episode index crosses
chunks_size, files are placed into chunk-001/, chunk-002/, etc.
"""

import json
import os
import shutil
from pathlib import Path

import pyarrow.parquet as pq
import pyarrow as pa
import yaml


DEMO_DATA_DIR = Path("demo_data")

SOURCE_DATASETS = [
    "g1_procedural_room_navigation_20260206_062009",
    "g1_procedural_room_navigation_20260206_080307",
    "g1_procedural_room_navigation_20260206_095145",
]

OUTPUT_DATASET = "g1_procedural_room_navigation_combined"


def count_episodes(dataset_dir: Path) -> int:
    with open(dataset_dir / "meta" / "episodes.jsonl") as f:
        return sum(1 for line in f if line.strip())


def get_video_keys(info: dict) -> list[str]:
    """Extract video feature keys from info.json features."""
    return [
        key for key, feat in info["features"].items()
        if feat.get("dtype") == "video"
    ]


def main():
    output_dir = DEMO_DATA_DIR / OUTPUT_DATASET
    if output_dir.exists():
        print(f"Output directory {output_dir} already exists. Aborting.")
        return

    source_dirs = [DEMO_DATA_DIR / name for name in SOURCE_DATASETS]
    for src in source_dirs:
        if not src.exists():
            print(f"Source dataset {src} not found. Aborting.")
            return

    # Load info.json to get chunks_size and video keys
    with open(source_dirs[0] / "meta" / "info.json") as f:
        info = json.load(f)
    chunks_size = info["chunks_size"]
    data_path_template = info["data_path"]
    video_path_template = info["video_path"]
    video_keys = get_video_keys(info)

    print(f"chunks_size: {chunks_size}")
    print(f"Video keys: {video_keys}")

    # Create output directory structure
    (output_dir / "meta").mkdir(parents=True)

    # 1. Copy info.json, modality.json, tasks.jsonl from first source (all identical)
    first_src = source_dirs[0]
    for filename in ["info.json", "modality.json", "tasks.jsonl"]:
        shutil.copy2(first_src / "meta" / filename, output_dir / "meta" / filename)
    print("Copied info.json, modality.json, tasks.jsonl")

    # 2. Merge episodes.jsonl with re-indexed episode_index
    all_episodes = []
    episode_offset = 0
    for src in source_dirs:
        with open(src / "meta" / "episodes.jsonl") as f:
            for line in f:
                line = line.strip()
                if not line:
                    continue
                ep = json.loads(line)
                ep["episode_index"] = ep["episode_index"] + episode_offset
                all_episodes.append(ep)
        episode_offset += count_episodes(src)

    with open(output_dir / "meta" / "episodes.jsonl", "w") as f:
        for ep in all_episodes:
            f.write(json.dumps(ep) + "\n")
    print(f"Merged episodes.jsonl: {len(all_episodes)} total episodes")

    # 3. Copy parquet files with updated episode_id, respecting chunk boundaries
    episode_offset = 0
    for src in source_dirs:
        num_episodes = count_episodes(src)
        src_chunks_size = chunks_size  # assume same chunks_size across sources

        for local_idx in range(num_episodes):
            new_idx = local_idx + episode_offset
            src_chunk = local_idx // src_chunks_size
            dst_chunk = new_idx // chunks_size

            src_parquet = src / data_path_template.format(
                episode_chunk=src_chunk, episode_index=local_idx
            )
            dst_parquet = output_dir / data_path_template.format(
                episode_chunk=dst_chunk, episode_index=new_idx
            )
            dst_parquet.parent.mkdir(parents=True, exist_ok=True)

            table = pq.read_table(src_parquet)
            new_episode_id = pa.array([new_idx] * len(table), type=pa.int64())
            col_idx = table.schema.get_field_index("episode_id")
            table = table.set_column(col_idx, "episode_id", new_episode_id)
            pq.write_table(table, dst_parquet)

        print(f"Copied {num_episodes} parquet files from {src.name} (episodes {episode_offset}-{episode_offset + num_episodes - 1})")
        episode_offset += num_episodes

    # 4. Symlink video files, respecting chunk boundaries
    episode_offset = 0
    for src in source_dirs:
        num_episodes = count_episodes(src)
        src_chunks_size = chunks_size

        for local_idx in range(num_episodes):
            new_idx = local_idx + episode_offset
            src_chunk = local_idx // src_chunks_size
            dst_chunk = new_idx // chunks_size

            for video_key in video_keys:
                src_video = src / video_path_template.format(
                    episode_chunk=src_chunk, video_key=video_key, episode_index=local_idx
                )
                src_video = src_video.resolve()
                dst_video = output_dir / video_path_template.format(
                    episode_chunk=dst_chunk, video_key=video_key, episode_index=new_idx
                )
                dst_video.parent.mkdir(parents=True, exist_ok=True)
                os.symlink(src_video, dst_video)

        print(f"Symlinked {num_episodes} videos from {src.name}")
        episode_offset += num_episodes

    # 5. Create origin.yaml
    origin = {
        "description": "Combined dataset from multiple collection sessions",
        "sources": [],
    }
    episode_offset = 0
    for src in source_dirs:
        num_episodes = count_episodes(src)
        origin["sources"].append({
            "name": src.name,
            "path": str(src.resolve()),
            "original_episodes": num_episodes,
            "mapped_range": [episode_offset, episode_offset + num_episodes - 1],
        })
        episode_offset += num_episodes
    origin["total_episodes"] = episode_offset

    with open(output_dir / "meta" / "origin.yaml", "w") as f:
        yaml.dump(origin, f, default_flow_style=False, sort_keys=False)
    print("Created origin.yaml")

    print(f"\nDone! Combined dataset at: {output_dir}")
    print(f"Total episodes: {episode_offset}")


if __name__ == "__main__":
    main()