twitch-highlight-detector/gpu_detector.py

#!/usr/bin/env python3
"""
GPU GAMEPLAY DETECTOR
Usa PyTorch + OpenCV en GPU para detectar gameplay en tiempo real
"""

import torch
import cv2
import numpy as np
import json
import subprocess
from pathlib import Path

print(f"🎮 GPU Gameplay Detector")
print(f"Dispositivo: {torch.cuda.get_device_name(0)}")
print(f"VRAM: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.1f} GB")


def extract_frame_batch_gpu(video_path, timestamps):
    """Extrae múltiples frames usando GPU."""
    frames = []

    for ts in timestamps:
        # Extraer frame con ffmpeg
        result = subprocess.run(
            [
                "ffmpeg",
                "-hwaccel",
                "cuda",
                "-i",
                video_path,
                "-ss",
                str(ts),
                "-vframes",
                "1",
                "-f",
                "image2pipe",
                "-vcodec",
                "png",
                "pipe:1",
            ],
            capture_output=True,
        )

        if result.returncode == 0:
            # Decodificar a numpy array
            nparr = np.frombuffer(result.stdout, np.uint8)
            frame = cv2.imdecode(nparr, cv2.IMREAD_COLOR)
            if frame is not None:
                frames.append((ts, frame))

    return frames


def analyze_gameplay_gpu(frames):
    """
    Analiza frames en GPU para detectar gameplay.

    Detecta:
    - Movimiento (optical flow)
    - Bordes (Canny) - UI de LoL tiene bordes característicos
    - Colores - Paleta característica de LoL
    """
    if not frames:
        return []

    results = []

    for ts, frame in frames:
        # Redimensionar para análisis rápido (GPU)
        frame_resized = cv2.resize(frame, (320, 180))

        # Convertir a tensor y mover a GPU
        frame_tensor = torch.from_numpy(frame_resized).float().cuda()

        # Análisis 1: Detectar movimiento (variación entre frames no aplicable aquí)
        # Análisis 2: Detectar colores característicos de LoL
        # LoL tiene muchos verdes (mapa), azules (UI), y colores vivos (campeones)

        mean_color = frame_tensor.mean(dim=(0, 1))
        std_color = frame_tensor.std(dim=(0, 1))

        # Heurísticas de gameplay de LoL:
        # - Alta variación de color (std > umbral)
        # - Presencia de verde (mapa)
        # - No es gris/negro (menu)

        is_colorful = std_color.mean() > 40  # Hay variación de color
        has_green = mean_color[1] > 80  # Canal verde presente (mapa)
        not_dark = frame_tensor.mean() > 30  # No es pantalla negra/menu

        # Score de gameplay (0-1)
        gameplay_score = 0.0
        if is_colorful:
            gameplay_score += 0.4
        if has_green:
            gameplay_score += 0.4
        if not_dark:
            gameplay_score += 0.2

        is_gameplay = gameplay_score > 0.6

        results.append(
            {
                "timestamp": ts,
                "is_gameplay": is_gameplay,
                "score": gameplay_score,
                "color_std": float(std_color.mean()),
                "green_mean": float(mean_color[1]),
            }
        )

        # Liberar memoria GPU
        del frame_tensor

    torch.cuda.empty_cache()
    return results


def scan_video_gpu(video_path, interval=30):
    """Escanea video completo usando GPU."""

    # Obtener duración
    result = subprocess.run(
        [
            "ffprobe",
            "-v",
            "error",
            "-show_entries",
            "format=duration",
            "-of",
            "default=noprint_wrappers=1:nokey=1",
            video_path,
        ],
        capture_output=True,
        text=True,
    )

    duration = float(result.stdout.strip())
    print(f"\n📹 Video: {duration / 60:.1f} minutos")
    print(f"🔍 Analizando cada {interval}s con GPU...")
    print()

    # Generar timestamps
    timestamps = list(range(455, int(duration), interval))

    # Procesar en batches para no saturar VRAM
    batch_size = 10
    all_results = []

    for i in range(0, len(timestamps), batch_size):
        batch_ts = timestamps[i : i + batch_size]
        print(
            f"Procesando batch {i // batch_size + 1}/{(len(timestamps) - 1) // batch_size + 1}..."
        )

        # Extraer frames
        frames = extract_frame_batch_gpu(video_path, batch_ts)

        # Analizar en GPU
        results = analyze_gameplay_gpu(frames)
        all_results.extend(results)

        # Mostrar progreso
        for r in results:
            status = "🎮" if r["is_gameplay"] else "🗣️"
            mins = r["timestamp"] // 60
            secs = r["timestamp"] % 60
            print(f"  {mins:02d}:{secs:02d} {status} Score: {r['score']:.2f}")

    # Convertir a segmentos
    segments = []
    current_start = None

    for r in all_results:
        if r["is_gameplay"]:
            if current_start is None:
                current_start = r["timestamp"]
        else:
            if current_start is not None:
                segments.append(
                    {
                        "start": current_start,
                        "end": r["timestamp"],
                        "duration": r["timestamp"] - current_start,
                    }
                )
                current_start = None

    # Cerrar último
    if current_start is not None:
        segments.append(
            {
                "start": current_start,
                "end": int(duration),
                "duration": int(duration) - current_start,
            }
        )

    return segments


def main():
    video_path = "nuevo_stream_360p.mp4"

    print("=" * 60)
    print("GPU GAMEPLAY DETECTOR")
    print("=" * 60)

    # Escanear
    segments = scan_video_gpu(video_path, interval=30)

    # Guardar
    with open("gameplay_segments_gpu.json", "w") as f:
        json.dump(segments, f, indent=2)

    print(f"\n{'=' * 60}")
    print(f"RESULTADO")
    print(f"{'=' * 60}")
    print(f"Segmentos de gameplay: {len(segments)}")
    total = sum(s["duration"] for s in segments)
    print(f"Tiempo total gameplay: {total // 60}m {total % 60}s")

    for i, seg in enumerate(segments, 1):
        mins_s, secs_s = divmod(seg["start"], 60)
        mins_e, secs_e = divmod(seg["end"], 60)
        print(
            f"{i}. {mins_s:02d}:{secs_s:02d} - {mins_e:02d}:{secs_e:02d} "
            f"({seg['duration'] // 60}m {seg['duration'] % 60}s)"
        )

    print(f"\n💾 Guardado en: gameplay_segments_gpu.json")


if __name__ == "__main__":
    main()