"""TransNetV2 shot detection wrapper (GPU-accelerated, lazy singleton)."""

from typing import List

_model = None  # lazy-loaded singleton


def _get_model():
    global _model
    if _model is None:
        from transnetv2_pytorch import TransNetV2  # lazy import avoids startup cost
        _model = TransNetV2(device='auto')
        _model.eval()
    return _model


def detect_transnet(video_path: str, threshold: float = 0.5) -> List[int]:
    """Return frame indices of cut points (start of each new shot after the first).

    Uses detect_scenes(), which returns a list of dicts with 'start_frame' and
    'end_frame' keys (integer frame indices, not timestamps).

    If the model detects 1 scene (single-shot video) returns [].

    Args:
        video_path: Path to the video file.
        threshold:  Shot-boundary probability threshold (default 0.5).

    Returns:
        List of integer frame indices where new shots begin (excluding frame 0).
        Empty list if the video contains only one shot.
    """
    model = _get_model()

    # detect_scenes returns List[Dict] with keys:
    #   shot_id, start_frame, end_frame, probability, start_time, end_time
    # This is the cleanest high-level path: no manual numpy conversion needed.
    scenes = model.detect_scenes(video_path, threshold=threshold)

    # Single-shot case: no cuts to report
    if len(scenes) <= 1:
        return []

    # Cut boundaries = start_frame of every scene except the first
    return [int(scene['start_frame']) for scene in scenes[1:]]