feat: Initial pipeline for Twitch highlight detection
- New 2-of-3 detection system (chat + audio + color) - GPU support (PyTorch ROCm/CUDA ready) - Draft mode (360p) for fast testing - HD mode (1080p) for final render - Auto download video + chat - CLI pipeline script - Documentation in Spanish
This commit is contained in:
ren
283
detector_gpu.py
Normal file
283
detector_gpu.py
Normal file
@@ -0,0 +1,283 @@
|
||||
import sys
|
||||
import json
|
||||
import logging
|
||||
import subprocess
|
||||
import torch
|
||||
import numpy as np
|
||||
from pathlib import Path
|
||||
|
||||
logging.basicConfig(level=logging.INFO)
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
def get_device():
|
||||
"""Obtiene el dispositivo (GPU o CPU)"""
|
||||
if torch.cuda.is_available():
|
||||
return torch.device("cuda")
|
||||
return torch.device("cpu")
|
||||
|
||||
def extract_audio_gpu(video_file, output_wav="audio.wav"):
|
||||
"""Extrae audio usando ffmpeg"""
|
||||
logger.info(f"Extrayendo audio de {video_file}...")
|
||||
subprocess.run([
|
||||
"ffmpeg", "-i", video_file,
|
||||
"-vn", "-acodec", "pcm_s16le",
|
||||
"-ar", "16000", "-ac", "1", output_wav, "-y"
|
||||
], capture_output=True)
|
||||
return output_wav
|
||||
|
||||
def detect_audio_peaks_gpu(audio_file, threshold=1.5, window_seconds=5, device="cpu"):
|
||||
"""
|
||||
Detecta picos de audio usando PyTorch para procesamiento
|
||||
"""
|
||||
logger.info("Analizando picos de audio con GPU...")
|
||||
|
||||
# Cargar audio con scipy
|
||||
import scipy.io.wavfile as wavfile
|
||||
sr, waveform = wavfile.read(audio_file)
|
||||
|
||||
# Convertir a float
|
||||
waveform = waveform.astype(np.float32) / 32768.0
|
||||
|
||||
# Calcular RMS por ventana usando numpy
|
||||
frame_length = sr * window_seconds
|
||||
hop_length = sr # 1 segundo entre ventanas
|
||||
|
||||
energies = []
|
||||
for i in range(0, len(waveform) - frame_length, hop_length):
|
||||
chunk = waveform[i:i + frame_length]
|
||||
energy = np.sqrt(np.mean(chunk ** 2))
|
||||
energies.append(energy)
|
||||
|
||||
energies = np.array(energies)
|
||||
|
||||
# Detectar picos
|
||||
mean_e = np.mean(energies)
|
||||
std_e = np.std(energies)
|
||||
|
||||
logger.info(f"Audio stats: media={mean_e:.4f}, std={std_e:.4f}")
|
||||
|
||||
audio_scores = {}
|
||||
for i, energy in enumerate(energies):
|
||||
if std_e > 0:
|
||||
z_score = (energy - mean_e) / std_e
|
||||
if z_score > threshold:
|
||||
audio_scores[i] = z_score
|
||||
|
||||
logger.info(f"Picos de audio detectados: {len(audio_scores)}")
|
||||
return audio_scores
|
||||
|
||||
def detect_video_peaks_fast(video_file, threshold=1.5, window_seconds=5):
|
||||
"""
|
||||
Detecta cambios de color/brillo (versión rápida, sin frames)
|
||||
"""
|
||||
logger.info("Analizando picos de color...")
|
||||
|
||||
# Usar ffmpeg para obtener información de brillo por segundo
|
||||
# Esto es mucho más rápido que procesar frames
|
||||
result = subprocess.run([
|
||||
"ffprobe", "-v", "error", "-select_streams", "v:0",
|
||||
"-show_entries", "stream=width,height,r_frame_rate,duration",
|
||||
"-of", "csv=p=0", video_file
|
||||
], capture_output=True, text=True)
|
||||
|
||||
# Extraer frames de referencia en baja resolución
|
||||
video_360 = video_file.replace('.mp4', '_temp_360.mp4')
|
||||
|
||||
# Convertir a 360p para procesamiento rápido
|
||||
logger.info("Convirtiendo a 360p para análisis...")
|
||||
subprocess.run([
|
||||
"ffmpeg", "-i", video_file,
|
||||
"-vf", "scale=-2:360",
|
||||
"-c:v", "libx264", "-preset", "fast",
|
||||
"-crf", "28",
|
||||
"-c:a", "copy",
|
||||
video_360, "-y"
|
||||
], capture_output=True)
|
||||
|
||||
# Extraer un frame cada N segundos
|
||||
frames_dir = Path("frames_temp")
|
||||
frames_dir.mkdir(exist_ok=True)
|
||||
|
||||
subprocess.run([
|
||||
"ffmpeg", "-i", video_360,
|
||||
"-vf", f"fps=1/{window_seconds}",
|
||||
f"{frames_dir}/frame_%04d.png", "-y"
|
||||
], capture_output=True)
|
||||
|
||||
# Procesar frames con PIL
|
||||
from PIL import Image
|
||||
import cv2
|
||||
|
||||
frame_files = sorted(frames_dir.glob("frame_*.png"))
|
||||
|
||||
if not frame_files:
|
||||
logger.warning("No se pudieron extraer frames")
|
||||
return {}
|
||||
|
||||
logger.info(f"Procesando {len(frame_files)} frames...")
|
||||
|
||||
brightness_scores = []
|
||||
for frame_file in frame_files:
|
||||
img = cv2.imread(str(frame_file))
|
||||
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
|
||||
|
||||
# Brillo = Value en HSV
|
||||
brightness = hsv[:,:,2].mean()
|
||||
# Saturación
|
||||
saturation = hsv[:,:,1].mean()
|
||||
|
||||
# Score combinado
|
||||
score = (brightness / 255) + (saturation / 255) * 0.5
|
||||
brightness_scores.append(score)
|
||||
|
||||
brightness_scores = np.array(brightness_scores)
|
||||
|
||||
# Detectar picos
|
||||
mean_b = np.mean(brightness_scores)
|
||||
std_b = np.std(brightness_scores)
|
||||
|
||||
logger.info(f"Brillo stats: media={mean_b:.3f}, std={std_b:.3f}")
|
||||
|
||||
color_scores = {}
|
||||
for i, score in enumerate(brightness_scores):
|
||||
if std_b > 0:
|
||||
z_score = (score - mean_b) / std_b
|
||||
if z_score > threshold:
|
||||
color_scores[i * window_seconds] = z_score
|
||||
|
||||
# Limpiar
|
||||
subprocess.run(["rm", "-rf", str(frames_dir)])
|
||||
subprocess.run(["rm", "-f", video_360], capture_output=True)
|
||||
|
||||
logger.info(f"Picos de color detectados: {len(color_scores)}")
|
||||
return color_scores
|
||||
|
||||
def main():
|
||||
import argparse
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--video", required=True, help="Video file")
|
||||
parser.add_argument("--chat", required=True, help="Chat JSON file")
|
||||
parser.add_argument("--output", default="highlights.json", help="Output JSON")
|
||||
parser.add_argument("--threshold", type=float, default=1.5, help="Threshold for peaks")
|
||||
parser.add_argument("--min-duration", type=int, default=10, help="Min highlight duration")
|
||||
parser.add_argument("--device", default="auto", help="Device: auto, cuda, cpu")
|
||||
args = parser.parse_args()
|
||||
|
||||
# Determinar device
|
||||
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)
|
||||
|
||||
# Extraer timestamps del chat
|
||||
chat_times = {}
|
||||
for comment in chat_data['comments']:
|
||||
second = int(comment['content_offset_seconds'])
|
||||
chat_times[second] = chat_times.get(second, 0) + 1
|
||||
|
||||
# Detectar picos de chat
|
||||
chat_values = list(chat_times.values())
|
||||
mean_c = np.mean(chat_values)
|
||||
std_c = np.std(chat_values)
|
||||
|
||||
logger.info(f"Chat stats: media={mean_c:.1f}, std={std_c:.1f}")
|
||||
|
||||
chat_scores = {}
|
||||
max_chat = max(chat_values) if chat_values else 1
|
||||
for second, count in chat_times.items():
|
||||
if std_c > 0:
|
||||
z_score = (count - mean_c) / std_c
|
||||
if z_score > args.threshold:
|
||||
chat_scores[second] = z_score
|
||||
|
||||
logger.info(f"Picos de chat: {len(chat_scores)}")
|
||||
|
||||
# Extraer y analizar audio
|
||||
audio_file = "temp_audio.wav"
|
||||
extract_audio_gpu(args.video, audio_file)
|
||||
audio_scores = detect_audio_peaks_gpu(audio_file, args.threshold, device=str(device))
|
||||
|
||||
# Limpiar audio temporal
|
||||
Path(audio_file).unlink(missing_ok=True)
|
||||
|
||||
# Analizar video
|
||||
video_scores = detect_video_peaks_fast(args.video, args.threshold)
|
||||
|
||||
# Combinar scores (2 de 3)
|
||||
logger.info("Combinando scores (2 de 3)...")
|
||||
|
||||
# Obtener duración total
|
||||
result = subprocess.run(
|
||||
["ffprobe", "-v", "error", "-show_entries", "format=duration",
|
||||
"-of", "default=noprint_wrokey=1:nokey=1", args.video],
|
||||
capture_output=True, text=True
|
||||
)
|
||||
duration = int(float(result.stdout.strip())) if result.stdout.strip() else 3600
|
||||
|
||||
# Normalizar scores
|
||||
max_audio = max(audio_scores.values()) if audio_scores else 1
|
||||
max_video = max(video_scores.values()) if video_scores else 1
|
||||
max_chat_norm = max(chat_scores.values()) if chat_scores else 1
|
||||
|
||||
# Unir segundos consecutivos
|
||||
highlights = []
|
||||
for second in range(duration):
|
||||
points = 0
|
||||
|
||||
# Chat
|
||||
chat_point = chat_scores.get(second, 0) / max_chat_norm if max_chat_norm > 0 else 0
|
||||
if chat_point > 0.5:
|
||||
points += 1
|
||||
|
||||
# Audio
|
||||
audio_point = audio_scores.get(second, 0) / max_audio if max_audio > 0 else 0
|
||||
if audio_point > 0.5:
|
||||
points += 1
|
||||
|
||||
# Color
|
||||
video_point = video_scores.get(second, 0) / max_video if max_video > 0 else 0
|
||||
if video_point > 0.5:
|
||||
points += 1
|
||||
|
||||
if points >= 2:
|
||||
highlights.append(second)
|
||||
|
||||
# Crear intervalos
|
||||
intervals = []
|
||||
if highlights:
|
||||
start = highlights[0]
|
||||
prev = highlights[0]
|
||||
|
||||
for second in highlights[1:]:
|
||||
if second - prev > 1:
|
||||
if second - start >= args.min_duration:
|
||||
intervals.append((start, prev))
|
||||
start = second
|
||||
prev = second
|
||||
|
||||
if prev - start >= args.min_duration:
|
||||
intervals.append((start, prev))
|
||||
|
||||
logger.info(f"Highlights encontrados: {len(intervals)}")
|
||||
|
||||
# Guardar
|
||||
with open(args.output, 'w') as f:
|
||||
json.dump(intervals, f)
|
||||
|
||||
logger.info(f"Guardado en {args.output}")
|
||||
|
||||
# Imprimir resumen
|
||||
print(f"\nHighlights ({len(intervals)} total):")
|
||||
for i, (s, e) in enumerate(intervals[:10]):
|
||||
print(f" {i+1}. {s}s - {e}s (duración: {e-s}s)")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in New Issue
Block a user