create_dataset.py

#!/usr/bin/env python3
"""
Create a HuggingFace dataset from PopSign data.

This script reads PopSign game and non-game subsets, extracts frames from videos
at signing segments, and creates a HuggingFace-compatible dataset with images.
"""

import argparse
import csv
import io
import json
import os
import pickle
import shutil
import tarfile
from functools import partial
from multiprocessing import Pool, cpu_count
from pathlib import Path

import cv2
import pympi
from datasets import Dataset, DatasetDict, Features, Image, Sequence, Value
from PIL import Image as PILImage
from tqdm import tqdm

from pose_utils import get_signing_time_range_from_pose

# Path to the README template
README_TEMPLATE_PATH = Path(__file__).parent / "popsign-images" / "README.md"


def get_video_duration(video_path: str) -> float:
    """Get the duration of a video in seconds."""
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise ValueError(f"Could not open video: {video_path}")

    fps = cap.get(cv2.CAP_PROP_FPS)
    frame_count = cap.get(cv2.CAP_PROP_FRAME_COUNT)
    cap.release()

    if fps <= 0:
        raise ValueError(f"Invalid FPS for video: {video_path}")

    return frame_count / fps


def get_video_fps(video_path: str) -> float:
    """Get the FPS of a video."""
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise ValueError(f"Could not open video: {video_path}")

    fps = cap.get(cv2.CAP_PROP_FPS)
    cap.release()
    return fps


def get_sign_time_range_from_eaf(eaf_path: str) -> tuple[float, float] | None:
    """
    Get the time range of the largest sign segment from an EAF file.

    Returns:
        Tuple of (start_time, end_time) in seconds, or None if no segments found.
    """
    try:
        eaf = pympi.Elan.Eaf(file_path=eaf_path)

        if 'SIGN' not in eaf.get_tier_names():
            return None

        sign_annotations = eaf.get_annotation_data_for_tier('SIGN')

        if not sign_annotations:
            return None

        # Find the largest segment
        largest_segment = max(sign_annotations, key=lambda s: s[1] - s[0])
        start_time = largest_segment[0] / 1000  # Convert ms to seconds
        end_time = largest_segment[1] / 1000

        return start_time, end_time
    except Exception:
        return None


def extract_frames_from_video(
    video_path: str,
    start_time: float,
    end_time: float,
    target_fps: float = 5,
    frame_size: int = 256
) -> list[PILImage.Image]:
    """
    Extract frames from a video between start and end times.

    Args:
        video_path: Path to the video file
        start_time: Start time in seconds
        end_time: End time in seconds
        target_fps: Target frames per second to extract
        frame_size: Size to resize frames to (square)

    Returns:
        List of PIL Images
    """
    cap = cv2.VideoCapture(video_path)
    if not cap.isOpened():
        raise ValueError(f"Could not open video: {video_path}")

    fps = cap.get(cv2.CAP_PROP_FPS)
    if fps <= 0:
        cap.release()
        raise ValueError(f"Invalid FPS for video: {video_path}")

    duration = end_time - start_time
    num_frames = max(2, int(duration * target_fps))

    # Calculate frame indices to extract
    start_frame = int(start_time * fps)
    end_frame = int(end_time * fps)
    duration_frames = end_frame - start_frame

    if duration_frames <= 0:
        cap.release()
        return []

    # Sample frames evenly across the duration
    if num_frames >= duration_frames:
        frame_indices = list(range(start_frame, end_frame + 1))
    else:
        frame_indices = [
            start_frame + int(i * duration_frames / (num_frames - 1))
            for i in range(num_frames - 1)
        ]
        frame_indices.append(end_frame)

    frames = []
    for frame_num in frame_indices:
        cap.set(cv2.CAP_PROP_POS_FRAMES, frame_num)
        ret, frame = cap.read()
        if ret:
            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
            img = PILImage.fromarray(frame_rgb)
            # Resize if needed (videos should already be 256x256)
            if img.size != (frame_size, frame_size):
                img = img.resize((frame_size, frame_size), PILImage.Resampling.LANCZOS)
            frames.append(img)

    cap.release()
    return frames


def process_csv_row(
    row: dict,
    videos_dir: str,
    eaf_dir: str,
    pose_dir: str,
    target_fps: float = 5
) -> dict | None:
    """
    Process a single CSV row and return a dataset entry.

    Uses a cascading approach for segmentation:
    1. First try pose-based segmentation (wrist above elbow heuristic)
    2. If pose covers entire file, fall back to EAF segmentation
    3. If neither works, use full video duration

    Args:
        row: CSV row dictionary
        videos_dir: Directory containing 256x256 videos
        eaf_dir: Directory containing EAF files
        pose_dir: Directory containing pose files
        target_fps: Target FPS for frame extraction

    Returns:
        Dictionary with dataset entry or None if processing failed
    """
    md5 = row['md5']
    video_path = os.path.join(videos_dir, f"{md5}.mp4")
    pose_path = os.path.join(pose_dir, f"{md5}.pose")
    eaf_path = os.path.join(eaf_dir, f"{md5}.eaf")

    # Check if video exists
    if not os.path.exists(video_path):
        return None

    # Cascading segmentation approach:
    # 1. Try pose-based segmentation first
    time_range = None
    if os.path.exists(pose_path):
        time_range = get_signing_time_range_from_pose(pose_path)

    # 2. If pose covers entire file (returns None), try EAF
    if time_range is None and os.path.exists(eaf_path):
        time_range = get_sign_time_range_from_eaf(eaf_path)

    # 3. Fall back to full video duration
    if time_range is not None:
        start_time, end_time = time_range
    else:
        try:
            start_time = 0.0
            end_time = get_video_duration(video_path)
        except Exception:
            return None

    # Validate time range
    if end_time <= start_time:
        return None

    # Extract frames
    try:
        frames = extract_frames_from_video(
            video_path, start_time, end_time, target_fps=target_fps
        )
    except Exception:
        return None

    if not frames:
        return None

    return {
        'file': row['file'],
        'start': round(start_time, 3),
        'end': round(end_time, 3),
        'text': row['text'],
        'images': frames
    }


def load_csv_data(csv_path: str) -> dict[str, list[dict]]:
    """
    Load CSV data and group by split.

    Returns:
        Dictionary mapping split names to lists of row dictionaries
    """
    splits = {'train': [], 'validation': [], 'test': []}

    with open(csv_path, 'r', encoding='utf-8') as f:
        reader = csv.DictReader(f)
        for row in reader:
            split = row['split']
            if split in splits:
                splits[split].append(row)

    return splits


def _process_row_wrapper(args: tuple) -> dict | None:
    """Wrapper for process_csv_row to work with multiprocessing."""
    row, videos_dir, eaf_dir, pose_dir, target_fps = args
    return process_csv_row(row, videos_dir, eaf_dir, pose_dir, target_fps)


def create_subset_dataset(
    csv_path: str,
    videos_dir: str,
    eaf_dir: str,
    pose_dir: str,
    target_fps: float = 5,
    limit: int | None = None,
    num_workers: int | None = None
) -> DatasetDict:
    """
    Create a DatasetDict for a single subset (game or non-game).

    Args:
        csv_path: Path to the index.csv file
        videos_dir: Directory containing 256x256 videos
        eaf_dir: Directory containing EAF files
        pose_dir: Directory containing pose files
        target_fps: Target FPS for frame extraction
        limit: Optional limit on number of samples per split (for testing)
        num_workers: Number of parallel workers (default: CPU count)

    Returns:
        DatasetDict with train, validation, test splits
    """
    splits_data = load_csv_data(csv_path)

    if num_workers is None:
        num_workers = cpu_count()

    features = Features({
        'file': Value('string'),
        'start': Value('float32'),
        'end': Value('float32'),
        'text': Value('string'),
        'images': Sequence(Image())
    })

    dataset_splits = {}

    for split_name, rows in splits_data.items():
        if not rows:
            continue

        if limit is not None:
            rows = rows[:limit]

        # Prepare arguments for parallel processing
        args_list = [(row, videos_dir, eaf_dir, pose_dir, target_fps) for row in rows]

        processed_data = []

        if num_workers > 1:
            with Pool(num_workers) as pool:
                results = list(tqdm(
                    pool.imap(_process_row_wrapper, args_list, chunksize=100),
                    total=len(args_list),
                    desc=f"Processing {split_name}",
                    unit="sample"
                ))
            processed_data = [r for r in results if r is not None]
        else:
            for row in tqdm(rows, desc=f"Processing {split_name}", unit="sample"):
                result = process_csv_row(row, videos_dir, eaf_dir, pose_dir, target_fps)
                if result is not None:
                    processed_data.append(result)

        if processed_data:
            dataset_splits[split_name] = Dataset.from_list(
                processed_data,
                features=features
            )

    return DatasetDict(dataset_splits)


def create_popsign_dataset(
    popsign_dir: str,
    videos_dir: str,
    eaf_dir: str,
    pose_dir: str,
    output_dir: str,
    target_fps: float = 5,
    limit: int | None = None,
    num_workers: int | None = None
):
    """
    Create the complete PopSign HuggingFace dataset with game and non-game subsets.

    Args:
        popsign_dir: Root directory containing game/ and non-game/ subdirectories
        videos_dir: Directory containing 256x256 videos
        eaf_dir: Directory containing EAF files
        pose_dir: Directory containing pose files
        output_dir: Output directory for the dataset
        target_fps: Target FPS for frame extraction
        limit: Optional limit on samples per split (for testing)
        num_workers: Number of parallel workers
    """
    output_path = Path(output_dir)
    output_path.mkdir(parents=True, exist_ok=True)

    subsets = ['game', 'non-game']

    for subset in subsets:
        csv_path = os.path.join(popsign_dir, subset, 'index.csv')

        if not os.path.exists(csv_path):
            print(f"Warning: {csv_path} not found, skipping {subset}")
            continue

        print(f"\nProcessing {subset} subset...")

        dataset_dict = create_subset_dataset(
            csv_path=csv_path,
            videos_dir=videos_dir,
            eaf_dir=eaf_dir,
            pose_dir=pose_dir,
            target_fps=target_fps,
            limit=limit,
            num_workers=num_workers
        )

        # Save the subset
        subset_path = output_path / subset
        dataset_dict.save_to_disk(str(subset_path))
        print(f"Saved {subset} to {subset_path}")

    print(f"\nDataset saved to {output_dir}")


def _save_shard(shard_data: list, shard_idx: int, num_shards: int, split_name: str, subset_data_path: Path, features: Features) -> int:
    """Save a shard to parquet and return the number of samples saved."""
    if not shard_data:
        return 0
    dataset = Dataset.from_list(shard_data, features=features)
    parquet_path = subset_data_path / f"{split_name}-{shard_idx:05d}-of-{num_shards:05d}.parquet"
    dataset.to_parquet(str(parquet_path))
    count = len(shard_data)
    print(f"\n  Saved shard {shard_idx} ({count} samples) -> {parquet_path.name}", flush=True)
    del dataset
    return count


def save_as_parquet(
    popsign_dir: str,
    videos_dir: str,
    eaf_dir: str,
    pose_dir: str,
    output_dir: str,
    target_fps: float = 5,
    limit: int | None = None,
    shard_size: int = 10000,
    num_workers: int | None = None
):
    """
    Create the PopSign dataset and save in Parquet format for HuggingFace Hub upload.

    Saves shards incrementally to minimize RAM usage by processing in small batches.

    Directory structure:
    output_dir/
    ├── README.md
    └── data/
        ├── game/
        │   ├── train-00000-of-NNNNN.parquet
        │   └── ...
        └── non-game/
            └── ...

    Args:
        pose_dir: Directory containing pose files
        shard_size: Number of samples per parquet shard (default: 10000)
        num_workers: Number of parallel workers
    """
    output_path = Path(output_dir)
    data_path = output_path / "data"
    data_path.mkdir(parents=True, exist_ok=True)

    # Copy README.md to output directory (if not already there)
    readme_dest = output_path / "README.md"
    if README_TEMPLATE_PATH.exists() and README_TEMPLATE_PATH.resolve() != readme_dest.resolve():
        shutil.copy(README_TEMPLATE_PATH, readme_dest)
        print(f"Copied README.md to {readme_dest}")

    if num_workers is None:
        num_workers = cpu_count()

    # Process in small batches to limit memory - each batch is processed in parallel,
    # then results are accumulated until we have enough for a shard
    batch_size = min(1000, shard_size)  # Small batches to limit memory
    print(f"Using {num_workers} workers, shard_size={shard_size}, batch_size={batch_size}", flush=True)

    features = Features({
        'file': Value('string'),
        'start': Value('float32'),
        'end': Value('float32'),
        'text': Value('string'),
        'images': Sequence(Image())
    })

    subsets = ['game', 'non-game']

    for subset in subsets:
        csv_path = os.path.join(popsign_dir, subset, 'index.csv')

        if not os.path.exists(csv_path):
            print(f"Warning: {csv_path} not found, skipping {subset}", flush=True)
            continue

        print(f"\nProcessing {subset} subset...", flush=True)

        splits_data = load_csv_data(csv_path)
        subset_data_path = data_path / subset
        subset_data_path.mkdir(parents=True, exist_ok=True)

        for split_name, rows in splits_data.items():
            if not rows:
                continue

            if limit is not None:
                rows = rows[:limit]

            total_rows = len(rows)
            num_shards = max(1, (total_rows + shard_size - 1) // shard_size)
            print(f"Processing {split_name} ({total_rows} rows, ~{num_shards} shards)...", flush=True)

            shard_data = []
            shard_idx = 0
            total_saved = 0
            total_processed = 0

            # Process in small batches to control memory
            pbar = tqdm(total=total_rows, desc=f"  {split_name}", unit="row")

            for batch_start in range(0, total_rows, batch_size):
                batch_end = min(batch_start + batch_size, total_rows)
                batch_rows = rows[batch_start:batch_end]

                # Process this batch
                if num_workers > 1:
                    args_list = [(row, videos_dir, eaf_dir, pose_dir, target_fps) for row in batch_rows]
                    with Pool(num_workers) as pool:
                        results = pool.map(_process_row_wrapper, args_list)
                else:
                    results = [process_csv_row(row, videos_dir, eaf_dir, pose_dir, target_fps) for row in batch_rows]

                # Accumulate successful results
                for result in results:
                    if result is not None:
                        shard_data.append(result)

                        # Save shard when full
                        if len(shard_data) >= shard_size:
                            total_saved += _save_shard(shard_data, shard_idx, num_shards, split_name, subset_data_path, features)
                            shard_data = []
                            shard_idx += 1

                # Free batch memory
                del results
                total_processed += len(batch_rows)
                pbar.update(len(batch_rows))

            pbar.close()

            # Save any remaining data
            if shard_data:
                total_saved += _save_shard(shard_data, shard_idx, num_shards, split_name, subset_data_path, features)
                shard_idx += 1

            print(f"Saved {subset}/{split_name}: {total_saved} samples in {shard_idx} shards", flush=True)

    print(f"\nDataset saved to {output_dir}", flush=True)
    print("\nTo upload to HuggingFace Hub:")
    print(f"  huggingface-cli upload sign/popsign-images {output_dir} .")


def _save_webdataset_shard(
    shard_data: list,
    shard_idx: int,
    num_shards: int,
    split_name: str,
    subset_data_path: Path,
    sample_id_offset: int,
    jpeg_quality: int
) -> int:
    """Save a shard to tar and return the number of samples saved."""
    if not shard_data:
        return 0

    tar_path = subset_data_path / f"{split_name}-{shard_idx:05d}-of-{num_shards:05d}.tar"

    with tarfile.open(tar_path, 'w') as tar:
        for i, sample in enumerate(shard_data):
            sample_id = f"{sample_id_offset + i:06d}"

            # Write metadata JSON
            metadata = {
                'file': sample['file'],
                'start': sample['start'],
                'end': sample['end'],
                'text': sample['text'],
                'num_frames': len(sample['images'])
            }
            json_data = json.dumps(metadata).encode('utf-8')
            json_info = tarfile.TarInfo(name=f"{sample_id}.json")
            json_info.size = len(json_data)
            tar.addfile(json_info, io.BytesIO(json_data))

            # Write all frames as JPEG bytes in a single pickle file
            frames_data = []
            for img in sample['images']:
                jpg_buffer = io.BytesIO()
                img.save(jpg_buffer, format='JPEG', quality=jpeg_quality)
                frames_data.append(jpg_buffer.getvalue())

            pyd_data = pickle.dumps(frames_data)
            pyd_info = tarfile.TarInfo(name=f"{sample_id}.pyd")
            pyd_info.size = len(pyd_data)
            tar.addfile(pyd_info, io.BytesIO(pyd_data))

    count = len(shard_data)
    print(f"\n  Saved shard {shard_idx} ({count} samples) -> {tar_path.name}", flush=True)
    return count


def save_as_webdataset(
    popsign_dir: str,
    videos_dir: str,
    eaf_dir: str,
    pose_dir: str,
    output_dir: str,
    target_fps: float = 5,
    limit: int | None = None,
    shard_size: int = 1000,
    num_workers: int | None = None,
    jpeg_quality: int = 90
):
    """
    Create the PopSign dataset and save in WebDataset format for HuggingFace Hub upload.

    Saves shards incrementally to minimize RAM usage.

    Directory structure:
    output_dir/
    ├── README.md
    └── data/
        ├── game/
        │   ├── train-00000-of-NNNNN.tar
        │   └── ...
        └── non-game/
            └── ...

    Each tar contains:
    - {sample_id:06d}.json  (metadata: file, start, end, text, num_frames)
    - {sample_id:06d}.pyd   (pickled list of JPEG bytes)

    Args:
        pose_dir: Directory containing pose files
        shard_size: Number of samples per tar shard (default: 1000)
        num_workers: Number of parallel workers
        jpeg_quality: JPEG quality for saved images (default: 90)
    """
    output_path = Path(output_dir)
    data_path = output_path / "data"
    data_path.mkdir(parents=True, exist_ok=True)

    # Copy README.md to output directory (if not already there)
    readme_dest = output_path / "README.md"
    if README_TEMPLATE_PATH.exists() and README_TEMPLATE_PATH.resolve() != readme_dest.resolve():
        shutil.copy(README_TEMPLATE_PATH, readme_dest)
        print(f"Copied README.md to {readme_dest}")

    if num_workers is None:
        num_workers = cpu_count()

    # Process in small batches to limit memory
    batch_size = min(1000, shard_size)
    print(f"Using {num_workers} workers, shard_size={shard_size}, batch_size={batch_size}", flush=True)

    subsets = ['game', 'non-game']

    for subset in subsets:
        csv_path = os.path.join(popsign_dir, subset, 'index.csv')

        if not os.path.exists(csv_path):
            print(f"Warning: {csv_path} not found, skipping {subset}", flush=True)
            continue

        print(f"\nProcessing {subset} subset...", flush=True)

        splits_data = load_csv_data(csv_path)
        subset_data_path = data_path / subset
        subset_data_path.mkdir(parents=True, exist_ok=True)

        for split_name, rows in splits_data.items():
            if not rows:
                continue

            if limit is not None:
                rows = rows[:limit]

            total_rows = len(rows)
            num_shards = max(1, (total_rows + shard_size - 1) // shard_size)
            print(f"Processing {split_name} ({total_rows} rows, ~{num_shards} shards)...", flush=True)

            shard_data = []
            shard_idx = 0
            total_saved = 0
            sample_id_offset = 0

            # Process in small batches to control memory
            pbar = tqdm(total=total_rows, desc=f"  {split_name}", unit="row")

            for batch_start in range(0, total_rows, batch_size):
                batch_end = min(batch_start + batch_size, total_rows)
                batch_rows = rows[batch_start:batch_end]

                # Process this batch
                if num_workers > 1:
                    args_list = [(row, videos_dir, eaf_dir, pose_dir, target_fps) for row in batch_rows]
                    with Pool(num_workers) as pool:
                        results = pool.map(_process_row_wrapper, args_list)
                else:
                    results = [process_csv_row(row, videos_dir, eaf_dir, pose_dir, target_fps) for row in batch_rows]

                # Accumulate successful results
                for result in results:
                    if result is not None:
                        shard_data.append(result)

                        # Save shard when full
                        if len(shard_data) >= shard_size:
                            total_saved += _save_webdataset_shard(
                                shard_data, shard_idx, num_shards, split_name,
                                subset_data_path, sample_id_offset, jpeg_quality
                            )
                            sample_id_offset += len(shard_data)
                            shard_data = []
                            shard_idx += 1

                # Free batch memory
                del results
                pbar.update(len(batch_rows))

            pbar.close()

            # Save any remaining data
            if shard_data:
                total_saved += _save_webdataset_shard(
                    shard_data, shard_idx, num_shards, split_name,
                    subset_data_path, sample_id_offset, jpeg_quality
                )
                shard_idx += 1

            print(f"Saved {subset}/{split_name}: {total_saved} samples in {shard_idx} shards", flush=True)

    print(f"\nDataset saved to {output_dir}", flush=True)
    print("\nTo upload to HuggingFace Hub:")
    print(f"  huggingface-cli upload sign/popsign-images {output_dir} .")


def main():
    parser = argparse.ArgumentParser(
        description='Create PopSign HuggingFace dataset with images'
    )
    parser.add_argument(
        '--popsign-dir',
        type=str,
        default='some-path-to/popsign/v1',
        help='Root directory containing game/ and non-game/ subdirectories'
    )
    parser.add_argument(
        '--videos-dir',
        type=str,
        default='some-path-to-videos/256x256',
        help='Directory containing 256x256 videos named by MD5 hash'
    )
    parser.add_argument(
        '--eaf-dir',
        type=str,
        default='some-path-to-segments',
        help='Directory containing EAF segmentation files'
    )
    parser.add_argument(
        '--pose-dir',
        type=str,
        default='some-path-to-poses',
        help='Directory containing pose files for signing boundary detection'
    )
    parser.add_argument(
        '--output-dir',
        type=str,
        default='/shared/popsign-images',
        help='Output directory for the HuggingFace dataset'
    )
    parser.add_argument(
        '--fps',
        type=float,
        default=5,
        help='Target frames per second for frame extraction (default: 5)'
    )
    parser.add_argument(
        '--limit',
        type=int,
        default=None,
        help='Limit number of samples per split (for testing)'
    )
    parser.add_argument(
        '--format',
        type=str,
        choices=['webdataset', 'parquet', 'arrow'],
        default='parquet',
        help='Output format: webdataset (JPEG compressed), parquet, or arrow (for local use)'
    )
    parser.add_argument(
        '--shard-size',
        type=int,
        default=1000,
        help='Number of samples per shard (default: 1000)'
    )
    parser.add_argument(
        '--jpeg-quality',
        type=int,
        default=90,
        help='JPEG quality for WebDataset format (default: 90)'
    )
    parser.add_argument(
        '--workers',
        type=int,
        default=None,
        help='Number of parallel workers (default: CPU count)'
    )

    args = parser.parse_args()

    if args.format == 'webdataset':
        save_as_webdataset(
            popsign_dir=args.popsign_dir,
            videos_dir=args.videos_dir,
            eaf_dir=args.eaf_dir,
            pose_dir=args.pose_dir,
            output_dir=args.output_dir,
            target_fps=args.fps,
            limit=args.limit,
            shard_size=args.shard_size,
            num_workers=args.workers,
            jpeg_quality=args.jpeg_quality
        )
    elif args.format == 'parquet':
        save_as_parquet(
            popsign_dir=args.popsign_dir,
            videos_dir=args.videos_dir,
            eaf_dir=args.eaf_dir,
            pose_dir=args.pose_dir,
            output_dir=args.output_dir,
            target_fps=args.fps,
            limit=args.limit,
            shard_size=args.shard_size,
            num_workers=args.workers
        )
    else:
        create_popsign_dataset(
            popsign_dir=args.popsign_dir,
            videos_dir=args.videos_dir,
            eaf_dir=args.eaf_dir,
            pose_dir=args.pose_dir,
            output_dir=args.output_dir,
            target_fps=args.fps,
            limit=args.limit,
            num_workers=args.workers
        )


if __name__ == '__main__':
    main()