twitch-highlight-detector/detector_explosiones.py

#!/usr/bin/env python3
"""
Detector de EXPLOSIONES de chat:
Busca momentos repentinos de alta actividad en el chat
que suelen indicar momentos épicos/intereses.
"""
import sys
import json
import logging
import torch
import torch.nn.functional as F
import numpy as np
from pathlib import Path

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

def get_device():
    if torch.cuda.is_available():
        return torch.device("cuda")
    return torch.device("cpu")

def detect_chat_explosions(chat_data, device="cuda", window_seconds=10, spike_threshold=3.0):
    """
    Detecta EXPLOSIONES de chat: saltos repentinos en la actividad.
    En lugar de picos sostenidos, busca aumentos bruscos.
    """
    logger.info("=== Detectando EXPLOSIONES de chat ===")

    # Crear timeline de actividad por segundo
    duration = max(
        int(c['content_offset_seconds'])
        for c in chat_data['comments']
    ) + 1

    # Vector de actividad por segundo
    activity = torch.zeros(duration, device=device)
    for comment in chat_data['comments']:
        second = int(comment['content_offset_seconds'])
        if second < duration:
            activity[second] += 1

    # Calcular media móvil para ver tendencia
    window_size = window_seconds  # 10 segundos
    kernel = torch.ones(1, 1, window_size, device=device) / window_size
    activity_reshaped = activity.unsqueeze(0).unsqueeze(0)

    # Padear activity para mantener tamaño después de conv
    padding = window_size // 2
    activity_padded = F.pad(activity_reshaped, (padding, padding))

    activity_smooth = F.conv1d(activity_padded, kernel).squeeze()
    activity_smooth = activity_smooth[:activity.shape[0]]  # Recortar al tamaño original

    # Detectar EXPLOSIONES: saltos bruscos por encima de la tendencia
    # Calcular diferencia con la media móvil
    diff = activity - activity_smooth

    # Buscar spikes donde la actividad real es mucho mayor que la esperada
    mean_diff = torch.mean(diff)
    std_diff = torch.std(diff)

    # Threshold dinámico basado en percentiles
    percentile_90 = torch.quantile(activity[activity > 0], 0.90)
    percentile_95 = torch.quantile(activity[activity > 0], 0.95)
    percentile_99 = torch.quantile(activity[activity > 0], 0.99)

    logger.info(f"Activity stats: p90={percentile_90:.0f}, p95={percentile_95:.0f}, p99={percentile_99:.0f}")

    # Detectar explosiones: actividad > p95 Y diff alto
    explosion_mask = (activity > percentile_95) & (diff > std_diff * spike_threshold)

    # Encontrar regiones contiguas de explosiones
    explosion_indices = torch.where(explosion_mask)[0]

    if len(explosion_indices) == 0:
        logger.warning("No se detectaron explosiones. Bajando threshold...")
        explosion_mask = activity > percentile_90
        explosion_indices = torch.where(explosion_mask)[0]

    # Agrupar en eventos
    events = []
    if len(explosion_indices) > 0:
        start = explosion_indices[0].item()
        prev = explosion_indices[0].item()

        for idx in explosion_indices[1:]:
            second = idx.item()
            if second - prev > 15:  # 15 segundos de gap = nuevo evento
                if prev - start >= 5:  # Mínimo 5 segundos
                    events.append((start, prev))
                start = second
            prev = second

        if prev - start >= 5:
            events.append((start, prev))

    # Calcular "intensidad" de cada evento (pico de actividad)
    events_with_intensity = []
    for start, end in events:
        segment_activity = activity[start:end+1]
        peak = torch.max(segment_activity).item()
        avg = torch.mean(segment_activity).item()
        duration = end - start

        # Score combinado: pico * duración / dispersión
        intensity = (peak * duration) / (1 + (end - start) / 10)
        events_with_intensity.append((start, end, duration, peak, intensity))

    # Ordenar por intensidad (los más "explosivos" primero)
    events_with_intensity.sort(key=lambda x: -x[4])

    logger.info(f"Explosiones detectadas: {len(events)}")

    return events_with_intensity

def main():
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--chat", required=True)
    parser.add_argument("--output", default="explosiones.json")
    parser.add_argument("--top", type=int, default=20, help="Número de eventos a retornar")
    parser.add_argument("--min-duration", type=int, default=8)
    parser.add_argument("--device", default="auto")
    args = parser.parse_args()

    if args.device == "auto":
        device = get_device()
    else:
        device = torch.device(args.device)

    logger.info(f"Usando device: {device}")

    # Cargar chat
    logger.info("Cargando chat...")
    with open(args.chat, 'r') as f:
        chat_data = json.load(f)

    # Detectar explosiones
    events = detect_chat_explosions(chat_data, device)

    # Filtrar por duración mínima y tomar top N
    events_filtered = [(s, e, d, p, i) for s, e, d, p, i in events if d >= args.min_duration]
    events_top = events_filtered[:args.top]

    # Convertir a formato de intervalos
    intervals = [(int(s), int(e)) for s, e, d, p, i in events_top]

    # Guardar
    with open(args.output, 'w') as f:
        json.dump(intervals, f)

    logger.info(f"Guardado en {args.output}")

    # Imprimir resumen
    print(f"\n{'='*70}")
    print(f"EXPLOSIONES DE CHAT DETECTADAS".center(70))
    print(f"{'='*70}")
    print(f"Total: {len(intervals)} eventos (top {args.top} por intensidad)")
    print(f"Duración total: {sum(e-s for s,e in intervals)}s ({sum(e-s for s,e in intervals)/60:.1f} min)")
    print(f"{'-'*70}")

    for i, (start, end) in enumerate(intervals, 1):
        duration = end - start
        h = start // 3600
        m = (start % 3600) // 60
        sec = start % 60
        print(f"{i:2d}. {h:02d}:{m:02d}:{sec:02d} - {duration}s duración")

    print(f"{'='*70}")

if __name__ == "__main__":
    main()