00180d0b1c
- Implementación de detector híbrido (Whisper + Chat + Audio + VLM) - Sistema de detección de gameplay real vs hablando - Scene detection con FFmpeg - Soporte para RTX 3050 y RX 6800 XT - Guía completa en 6800xt.md para próxima IA - Scripts de filtrado visual y análisis de contexto - Pipeline automatizado de generación de videos
447 lines
14 KiB
Python
447 lines
14 KiB
Python
#!/usr/bin/env python3
|
|
"""
|
|
Detector de highlights que REALMENTE usa GPU NVIDIA
|
|
- torchaudio para cargar audio directamente a GPU
|
|
- PyTorch CUDA para todos los cálculos
|
|
- Optimizado para NVIDIA RTX 3050
|
|
"""
|
|
|
|
import sys
|
|
import json
|
|
import logging
|
|
import subprocess
|
|
import torch
|
|
import torch.nn.functional as F
|
|
import torchaudio
|
|
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():
|
|
device = torch.device("cuda")
|
|
logger.info(f"GPU detectada: {torch.cuda.get_device_name(0)}")
|
|
logger.info(
|
|
f"Memoria GPU total: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.1f} GB"
|
|
)
|
|
return device
|
|
return torch.device("cpu")
|
|
|
|
|
|
def load_audio_to_gpu(video_file, device="cuda", target_sr=16000):
|
|
"""
|
|
Carga audio del video a GPU usando ffmpeg + soundfile + PyTorch.
|
|
Extrae audio con ffmpeg a memoria (no disco), luego carga a GPU.
|
|
"""
|
|
import time
|
|
|
|
logger.info(f"Extrayendo audio de {video_file}...")
|
|
t0 = time.time()
|
|
|
|
# Usar ffmpeg para extraer audio a un pipe (memoria, no disco)
|
|
import io
|
|
|
|
cmd = [
|
|
"ffmpeg",
|
|
"-i",
|
|
video_file,
|
|
"-vn",
|
|
"-acodec",
|
|
"pcm_s16le",
|
|
"-ar",
|
|
str(target_sr),
|
|
"-ac",
|
|
"1",
|
|
"-f",
|
|
"wav",
|
|
"pipe:1",
|
|
"-y",
|
|
"-threads",
|
|
"4", # Usar múltiples hilos para acelerar
|
|
]
|
|
|
|
result = subprocess.run(cmd, capture_output=True)
|
|
logger.info(f"FFmpeg audio extraction: {time.time() - t0:.1f}s")
|
|
|
|
# Cargar WAV desde memoria con soundfile
|
|
import soundfile as sf
|
|
|
|
waveform_np, sr = sf.read(io.BytesIO(result.stdout), dtype="float32")
|
|
logger.info(f"Audio decode: {time.time() - t0:.1f}s")
|
|
|
|
# soundfile ya devuelve floats en [-1, 1], no hay que normalizar
|
|
# Convertir a tensor y mover a GPU con pin_memory para transferencia rápida
|
|
t1 = time.time()
|
|
waveform = torch.from_numpy(waveform_np).pin_memory().to(device, non_blocking=True)
|
|
|
|
# Asegurar forma (1, samples) para consistencia
|
|
waveform = (
|
|
waveform.unsqueeze(0)
|
|
if waveform.dim() == 1
|
|
else waveform.mean(dim=0, keepdim=True)
|
|
)
|
|
|
|
logger.info(f"CPU->GPU transfer: {time.time() - t1:.2f}s")
|
|
logger.info(f"Audio cargado: shape={waveform.shape}, SR={sr}")
|
|
logger.info(f"Rango de audio: [{waveform.min():.4f}, {waveform.max():.4f}]")
|
|
return waveform, sr
|
|
|
|
|
|
def detect_audio_peaks_gpu(
|
|
video_file, threshold=1.5, window_seconds=5, device="cuda", skip_intro=600
|
|
):
|
|
"""
|
|
Detecta picos de audio usando GPU completamente.
|
|
Procesa en chunks pequeños para maximizar uso GPU sin OOM en RTX 3050 (4GB).
|
|
"""
|
|
import time
|
|
|
|
# Cargar audio directamente a GPU
|
|
waveform, sr = load_audio_to_gpu(video_file, device=device)
|
|
|
|
# Saltar intro: eliminar primeros N segundos de audio
|
|
skip_samples = skip_intro * sr
|
|
if waveform.shape[-1] > skip_samples:
|
|
waveform = waveform[:, skip_samples:]
|
|
|
|
t0 = time.time()
|
|
# Parámetros
|
|
frame_length = sr * window_seconds
|
|
hop_length = sr # 1 segundo entre ventanas (menos memoria que 0.5s)
|
|
|
|
# Aplanar y mover a CPU para liberar GPU
|
|
waveform = waveform.squeeze(0)
|
|
waveform_cpu = waveform.cpu()
|
|
del waveform
|
|
torch.cuda.empty_cache()
|
|
|
|
# Calcular num_frames para chunking
|
|
total_samples = waveform_cpu.shape[-1]
|
|
num_frames = 1 + (total_samples - frame_length) // hop_length
|
|
|
|
# Chunks más pequeños para RTX 3050 (4GB VRAM)
|
|
chunk_frames = 5000 # frames por chunk (~2GB de memoria temporal)
|
|
num_chunks = (num_frames + chunk_frames - 1) // chunk_frames
|
|
|
|
logger.info(f"Processing {num_frames} frames in {num_chunks} chunks...")
|
|
|
|
all_energies = []
|
|
chunk_times = []
|
|
|
|
for chunk_idx in range(num_chunks):
|
|
chunk_start = chunk_idx * chunk_frames
|
|
chunk_end = min((chunk_idx + 1) * chunk_frames, num_frames)
|
|
actual_frames = chunk_end - chunk_start
|
|
|
|
if actual_frames <= 0:
|
|
break
|
|
|
|
# Calcular índices de muestra para este chunk
|
|
sample_start = chunk_start * hop_length
|
|
sample_end = sample_start + frame_length + (actual_frames - 1) * hop_length
|
|
|
|
if sample_end > total_samples:
|
|
padding_needed = sample_end - total_samples
|
|
chunk_waveform_np = F.pad(waveform_cpu[sample_start:], (0, padding_needed))
|
|
else:
|
|
chunk_waveform_np = waveform_cpu[sample_start:sample_end]
|
|
|
|
# Mover chunk a GPU
|
|
chunk_waveform = chunk_waveform_np.to(device)
|
|
|
|
# unfold para este chunk
|
|
if chunk_waveform.shape[-1] < frame_length:
|
|
del chunk_waveform
|
|
continue
|
|
windows = chunk_waveform.unfold(0, frame_length, hop_length)
|
|
|
|
# Operaciones GPU (visibles en monitoreo)
|
|
ct = time.time()
|
|
|
|
# 1. RMS
|
|
energies = torch.sqrt(torch.mean(windows**2, dim=1))
|
|
|
|
# 2. FFT más pequeño (solo primeras frecuencias)
|
|
window_fft = torch.fft.rfft(windows, n=windows.shape[1] // 4, dim=1)
|
|
spectral_centroid = torch.mean(torch.abs(window_fft), dim=1)
|
|
|
|
# 3. Rolling stats
|
|
kernel = torch.ones(1, 1, 5, device=device) / 5
|
|
energies_reshaped = energies.unsqueeze(0).unsqueeze(0)
|
|
energies_smooth = F.conv1d(energies_reshaped, kernel, padding=2).squeeze()
|
|
|
|
chunk_time = time.time() - ct
|
|
chunk_times.append(chunk_time)
|
|
|
|
# Guardar en CPU y liberar GPU
|
|
all_energies.append(energies.cpu())
|
|
|
|
# Liberar memoria GPU agresivamente
|
|
del (
|
|
chunk_waveform,
|
|
windows,
|
|
energies,
|
|
window_fft,
|
|
spectral_centroid,
|
|
energies_smooth,
|
|
)
|
|
torch.cuda.empty_cache()
|
|
|
|
if chunk_idx < 3:
|
|
logger.info(
|
|
f"Chunk {chunk_idx + 1}/{num_chunks}: {actual_frames} frames, GPU time: {chunk_time:.2f}s, GPU mem: {torch.cuda.memory_allocated(0) / 1024**3:.2f}GB"
|
|
)
|
|
|
|
logger.info(
|
|
f"GPU Processing: {time.time() - t0:.2f}s total, avg chunk: {sum(chunk_times) / len(chunk_times):.2f}s"
|
|
)
|
|
|
|
# Estadísticas finales en GPU
|
|
t1 = time.time()
|
|
all_energies_tensor = torch.cat(all_energies).to(device)
|
|
mean_e = torch.mean(all_energies_tensor)
|
|
std_e = torch.std(all_energies_tensor)
|
|
|
|
logger.info(f"Final stats (GPU): {time.time() - t1:.2f}s")
|
|
logger.info(f"Audio stats: media={mean_e:.4f}, std={std_e:.4f}")
|
|
|
|
# Detectar picos en GPU
|
|
t2 = time.time()
|
|
z_scores = (all_energies_tensor - mean_e) / (std_e + 1e-8)
|
|
peak_mask = z_scores > threshold
|
|
logger.info(f"Peak detection (GPU): {time.time() - t2:.2f}s")
|
|
|
|
# Convertir a diccionario
|
|
audio_scores = {
|
|
i: z_scores[i].item() for i in range(len(z_scores)) if peak_mask[i].item()
|
|
}
|
|
|
|
logger.info(f"Picos de audio detectados: {len(audio_scores)}")
|
|
return audio_scores
|
|
|
|
|
|
def detect_chat_peaks_gpu(chat_data, threshold=1.5, device="cuda", skip_intro=600):
|
|
"""
|
|
Analiza chat usando GPU para estadísticas.
|
|
"""
|
|
# Extraer timestamps del chat (saltar intro)
|
|
chat_times = {}
|
|
for comment in chat_data["comments"]:
|
|
second = int(comment["content_offset_seconds"])
|
|
if second >= skip_intro: # Saltar intro
|
|
chat_times[second] = chat_times.get(second, 0) + 1
|
|
|
|
if not chat_times:
|
|
return {}, {}
|
|
|
|
# Convertir a tensor GPU con pin_memory
|
|
chat_values = list(chat_times.values())
|
|
chat_tensor = torch.tensor(chat_values, dtype=torch.float32, device=device)
|
|
|
|
# Estadísticas en GPU
|
|
mean_c = torch.mean(chat_tensor)
|
|
std_c = torch.std(chat_tensor)
|
|
|
|
logger.info(f"Chat stats: media={mean_c:.1f}, std={std_c:.1f}")
|
|
|
|
# Detectar picos en GPU (vectorizado)
|
|
z_scores = (chat_tensor - mean_c) / (std_c + 1e-8)
|
|
peak_mask = z_scores > threshold
|
|
|
|
chat_scores = {}
|
|
for i, (second, count) in enumerate(chat_times.items()):
|
|
if peak_mask[i].item():
|
|
chat_scores[second] = z_scores[i].item()
|
|
|
|
logger.info(f"Picos de chat: {len(chat_scores)}")
|
|
return chat_scores, chat_times
|
|
|
|
|
|
def detect_video_peaks_fast(video_file, threshold=1.5, window_seconds=5, device="cuda"):
|
|
"""
|
|
Versión optimizada que omite el procesamiento de frames pesado.
|
|
El chat + audio suelen ser suficientes para detectar highlights.
|
|
Si realmente necesitas video, usa OpenCV con CUDA o torchvision.
|
|
"""
|
|
logger.info("Omitiendo análisis de video (lento con ffmpeg CPU)")
|
|
logger.info("Usando solo chat + audio para detección de highlights")
|
|
return {}
|
|
|
|
|
|
def combine_scores_gpu(
|
|
chat_scores,
|
|
audio_scores,
|
|
video_scores,
|
|
duration,
|
|
min_duration,
|
|
device="cuda",
|
|
window=3,
|
|
skip_intro=0,
|
|
):
|
|
"""
|
|
Combina scores usando GPU con ventana de tiempo para permitir coincidencias cercanas.
|
|
"""
|
|
logger.info(
|
|
f"Combinando scores con GPU (ventana={window}s, skip_intro={skip_intro}s)..."
|
|
)
|
|
|
|
# Crear tensores densos para vectorización
|
|
chat_tensor = torch.zeros(duration, device=device)
|
|
for sec, score in chat_scores.items():
|
|
if sec < duration:
|
|
chat_tensor[sec] = score
|
|
|
|
audio_tensor = torch.zeros(duration, device=device)
|
|
for sec, score in audio_scores.items():
|
|
if sec < duration:
|
|
audio_tensor[sec] = score
|
|
|
|
# Aplicar convolución 1D para suavizar con ventana (permite coincidencias cercanas)
|
|
kernel_size = window * 2 + 1
|
|
kernel = torch.ones(1, 1, kernel_size, device=device) / kernel_size
|
|
|
|
# Reshape para conv1d: (batch, channels, length)
|
|
chat_reshaped = chat_tensor.unsqueeze(0).unsqueeze(0)
|
|
audio_reshaped = audio_tensor.unsqueeze(0).unsqueeze(0)
|
|
|
|
# Suavizar con ventana móvil
|
|
chat_smooth = F.conv1d(chat_reshaped, kernel, padding=window).squeeze()
|
|
audio_smooth = F.conv1d(audio_reshaped, kernel, padding=window).squeeze()
|
|
|
|
# Normalizar en GPU
|
|
max_chat = chat_smooth.max()
|
|
max_audio = audio_smooth.max()
|
|
|
|
chat_normalized = chat_smooth / max_chat if max_chat > 0 else chat_smooth
|
|
audio_normalized = audio_smooth / max_audio if max_audio > 0 else audio_smooth
|
|
|
|
# Vectorizado: puntos >= 1 (chat o audio, más permisivo)
|
|
# Antes: puntos >= 2, ahora: puntos >= 1 para encontrar más highlights
|
|
points = (chat_normalized > 0.25).float() + (audio_normalized > 0.25).float()
|
|
highlight_mask = points >= 1
|
|
|
|
# Obtener segundos destacados
|
|
highlight_indices = torch.where(highlight_mask)[0]
|
|
|
|
# Crear intervalos (sumando skip_intro para timestamps reales)
|
|
intervals = []
|
|
if len(highlight_indices) > 0:
|
|
start = highlight_indices[0].item()
|
|
prev = highlight_indices[0].item()
|
|
|
|
for idx in highlight_indices[1:]:
|
|
second = idx.item()
|
|
if second - prev > 1:
|
|
if prev - start >= min_duration:
|
|
intervals.append((int(start + skip_intro), int(prev + skip_intro)))
|
|
start = second
|
|
prev = second
|
|
|
|
if prev - start >= min_duration:
|
|
intervals.append((int(start + skip_intro), int(prev + skip_intro)))
|
|
|
|
return intervals
|
|
|
|
|
|
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")
|
|
parser.add_argument(
|
|
"--skip-intro",
|
|
type=int,
|
|
default=600,
|
|
help="Segundos a saltar del inicio (default: 600s = 10min)",
|
|
)
|
|
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 y analizar chat con GPU
|
|
logger.info("Cargando chat...")
|
|
with open(args.chat, "r") as f:
|
|
chat_data = json.load(f)
|
|
|
|
logger.info(
|
|
f"Saltando intro: primeros {args.skip_intro}s (~{args.skip_intro // 60}min)"
|
|
)
|
|
chat_scores, _ = detect_chat_peaks_gpu(
|
|
chat_data, args.threshold, device=device, skip_intro=args.skip_intro
|
|
)
|
|
|
|
# Analizar audio con GPU (saltando intro)
|
|
audio_scores = detect_audio_peaks_gpu(
|
|
args.video, args.threshold, device=device, skip_intro=args.skip_intro
|
|
)
|
|
|
|
# Analizar video (omitido por rendimiento)
|
|
video_scores = detect_video_peaks_fast(args.video, args.threshold, device=device)
|
|
|
|
# Obtener duración total
|
|
result = subprocess.run(
|
|
[
|
|
"ffprobe",
|
|
"-v",
|
|
"error",
|
|
"-show_entries",
|
|
"format=duration",
|
|
"-of",
|
|
"default=noprint_wrappers=1:nokey=1",
|
|
args.video,
|
|
],
|
|
capture_output=True,
|
|
text=True,
|
|
)
|
|
duration = int(float(result.stdout.strip())) if result.stdout.strip() else 3600
|
|
|
|
# Combinar scores usando GPU (ajustando timestamps por el intro saltado)
|
|
intervals = combine_scores_gpu(
|
|
chat_scores,
|
|
audio_scores,
|
|
video_scores,
|
|
duration,
|
|
args.min_duration,
|
|
device=device,
|
|
skip_intro=args.skip_intro,
|
|
)
|
|
|
|
logger.info(f"Highlights encontrados: {len(intervals)}")
|
|
|
|
# Guardar resultados
|
|
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[:20]):
|
|
print(f" {i + 1}. {s}s - {e}s (duración: {e - s}s)")
|
|
|
|
if len(intervals) > 20:
|
|
print(f" ... y {len(intervals) - 20} más")
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|