talkingAvater_bgk / core /optimization /parallel_processing.py
oKen38461's picture
推論キャッシュと並列処理の機能を追加し、`process_talking_head_optimized`関数をキャッシュと並列処理に対応させました。また、Gradioインターフェースにキャッシュ管理機能を追加しました。
07b71bb
"""
Parallel Processing Module for DittoTalkingHead
Implements concurrent audio and image preprocessing
"""
import asyncio
import concurrent.futures
from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor
import time
from typing import Tuple, Dict, Any, Optional, Callable
import numpy as np
from pathlib import Path
import threading
import queue
import torch
from functools import partial
class ParallelProcessor:
"""
Parallel processing for audio and image preprocessing
"""
def __init__(
self,
num_threads: int = 4,
num_processes: int = 2,
use_cuda_streams: bool = True
):
"""
Initialize parallel processor
Args:
num_threads: Number of threads for I/O operations
num_processes: Number of processes for CPU-intensive tasks
use_cuda_streams: Use CUDA streams for GPU operations
"""
self.num_threads = num_threads
self.num_processes = num_processes
self.use_cuda_streams = use_cuda_streams and torch.cuda.is_available()
# Thread pool for I/O operations
self.thread_executor = ThreadPoolExecutor(max_workers=num_threads)
# Process pool for CPU-intensive operations
self.process_executor = ProcessPoolExecutor(max_workers=num_processes)
# CUDA streams for GPU operations
if self.use_cuda_streams:
self.cuda_streams = [torch.cuda.Stream() for _ in range(2)]
else:
self.cuda_streams = None
print(f"✅ ParallelProcessor initialized: {num_threads} threads, {num_processes} processes")
if self.use_cuda_streams:
print("✅ CUDA streams enabled for GPU parallelism")
def preprocess_audio_parallel(self, audio_path: str) -> Dict[str, Any]:
"""
Preprocess audio file in parallel
Args:
audio_path: Path to audio file
Returns:
Preprocessed audio data
"""
import librosa
# Define subtasks
def load_audio():
return librosa.load(audio_path, sr=16000)
def extract_features(audio, sr):
# Extract various audio features in parallel
features = {}
# MFCC features
features['mfcc'] = librosa.feature.mfcc(y=audio, sr=sr, n_mfcc=13)
# Spectral features
features['spectral_centroid'] = librosa.feature.spectral_centroid(y=audio, sr=sr)
features['spectral_rolloff'] = librosa.feature.spectral_rolloff(y=audio, sr=sr)
return features
# Load audio
audio, sr = load_audio()
# Extract features in parallel (if needed)
features = extract_features(audio, sr)
return {
'audio': audio,
'sample_rate': sr,
'features': features,
'duration': len(audio) / sr
}
def preprocess_image_parallel(self, image_path: str, target_size: int = 320) -> Dict[str, Any]:
"""
Preprocess image file in parallel
Args:
image_path: Path to image file
target_size: Target resolution
Returns:
Preprocessed image data
"""
from PIL import Image
import cv2
# Define subtasks
def load_and_resize():
# Load image
img = Image.open(image_path).convert('RGB')
# Resize
img = img.resize((target_size, target_size), Image.Resampling.LANCZOS)
return np.array(img)
def extract_face_landmarks(img_array):
# Face detection and landmark extraction
# Simplified version - in production, use MediaPipe or similar
return {
'has_face': True,
'landmarks': None # Placeholder
}
# Execute in parallel
future_img = self.thread_executor.submit(load_and_resize)
# Get results
img_array = future_img.result()
# Extract landmarks
landmarks = extract_face_landmarks(img_array)
return {
'image': img_array,
'shape': img_array.shape,
'landmarks': landmarks
}
async def preprocess_parallel_async(
self,
audio_path: str,
image_path: str,
target_size: int = 320
) -> Tuple[Dict[str, Any], Dict[str, Any]]:
"""
Asynchronously preprocess audio and image in parallel
Args:
audio_path: Path to audio file
image_path: Path to image file
target_size: Target image resolution
Returns:
Tuple of (audio_data, image_data)
"""
loop = asyncio.get_event_loop()
# Create tasks for parallel execution
audio_task = loop.run_in_executor(
self.thread_executor,
self.preprocess_audio_parallel,
audio_path
)
image_task = loop.run_in_executor(
self.thread_executor,
partial(self.preprocess_image_parallel, target_size=target_size),
image_path
)
# Wait for both tasks to complete
audio_data, image_data = await asyncio.gather(audio_task, image_task)
return audio_data, image_data
def preprocess_parallel_sync(
self,
audio_path: str,
image_path: str,
target_size: int = 320
) -> Tuple[Dict[str, Any], Dict[str, Any]]:
"""
Synchronously preprocess audio and image in parallel
Args:
audio_path: Path to audio file
image_path: Path to image file
target_size: Target image resolution
Returns:
Tuple of (audio_data, image_data)
"""
# Submit tasks to thread pool
audio_future = self.thread_executor.submit(
self.preprocess_audio_parallel,
audio_path
)
image_future = self.thread_executor.submit(
self.preprocess_image_parallel,
image_path,
target_size
)
# Wait for results
audio_data = audio_future.result()
image_data = image_future.result()
return audio_data, image_data
def process_gpu_parallel(
self,
audio_tensor: torch.Tensor,
image_tensor: torch.Tensor,
model_audio: torch.nn.Module,
model_image: torch.nn.Module
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Process audio and image through models using CUDA streams
Args:
audio_tensor: Audio tensor
image_tensor: Image tensor
model_audio: Audio processing model
model_image: Image processing model
Returns:
Tuple of processed tensors
"""
if not self.use_cuda_streams:
# Fallback to sequential processing
audio_out = model_audio(audio_tensor)
image_out = model_image(image_tensor)
return audio_out, image_out
# Use CUDA streams for parallel GPU processing
with torch.cuda.stream(self.cuda_streams[0]):
audio_out = model_audio(audio_tensor)
with torch.cuda.stream(self.cuda_streams[1]):
image_out = model_image(image_tensor)
# Synchronize streams
torch.cuda.synchronize()
return audio_out, image_out
def shutdown(self):
"""Shutdown executors"""
self.thread_executor.shutdown(wait=True)
self.process_executor.shutdown(wait=True)
print("✅ ParallelProcessor shutdown complete")
class PipelineProcessor:
"""
Pipeline-based processing for continuous operations
"""
def __init__(self, stages: Dict[str, Callable], buffer_size: int = 10):
"""
Initialize pipeline processor
Args:
stages: Dictionary of stage_name -> processing_function
buffer_size: Size of queues between stages
"""
self.stages = stages
self.buffer_size = buffer_size
# Create queues between stages
self.queues = {}
stage_names = list(stages.keys())
for i in range(len(stage_names) - 1):
queue_name = f"{stage_names[i]}_to_{stage_names[i+1]}"
self.queues[queue_name] = queue.Queue(maxsize=buffer_size)
# Input and output queues
self.input_queue = queue.Queue(maxsize=buffer_size)
self.output_queue = queue.Queue(maxsize=buffer_size)
# Worker threads
self.workers = []
self.stop_event = threading.Event()
def _worker(self, stage_name: str, process_func: Callable, input_q: queue.Queue, output_q: queue.Queue):
"""Worker thread for a pipeline stage"""
while not self.stop_event.is_set():
try:
# Get input with timeout
item = input_q.get(timeout=0.1)
if item is None: # Poison pill
output_q.put(None)
break
# Process item
result = process_func(item)
# Put result
output_q.put(result)
except queue.Empty:
continue
except Exception as e:
print(f"Error in stage {stage_name}: {e}")
output_q.put(None)
def start(self):
"""Start pipeline processing"""
stage_names = list(self.stages.keys())
# Create worker threads
for i, (stage_name, process_func) in enumerate(self.stages.items()):
# Determine input and output queues
if i == 0:
input_q = self.input_queue
else:
queue_name = f"{stage_names[i-1]}_to_{stage_names[i]}"
input_q = self.queues[queue_name]
if i == len(stage_names) - 1:
output_q = self.output_queue
else:
queue_name = f"{stage_names[i]}_to_{stage_names[i+1]}"
output_q = self.queues[queue_name]
# Create and start worker
worker = threading.Thread(
target=self._worker,
args=(stage_name, process_func, input_q, output_q)
)
worker.start()
self.workers.append(worker)
print(f"✅ Pipeline started with {len(self.workers)} stages")
def process(self, item: Any) -> Any:
"""Process an item through the pipeline"""
self.input_queue.put(item)
return self.output_queue.get()
def stop(self):
"""Stop pipeline processing"""
self.stop_event.set()
# Send poison pills
self.input_queue.put(None)
# Wait for workers
for worker in self.workers:
worker.join()
print("✅ Pipeline stopped")
def benchmark_parallel_processing():
"""Benchmark parallel vs sequential processing"""
import time
print("\n=== Parallel Processing Benchmark ===")
# Create processor
processor = ParallelProcessor(num_threads=4)
# Test files (using example files)
audio_path = "example/audio.wav"
image_path = "example/image.png"
# Sequential processing
start_seq = time.time()
audio_data_seq = processor.preprocess_audio_parallel(audio_path)
image_data_seq = processor.preprocess_image_parallel(image_path)
time_seq = time.time() - start_seq
# Parallel processing
start_par = time.time()
audio_data_par, image_data_par = processor.preprocess_parallel_sync(audio_path, image_path)
time_par = time.time() - start_par
# Results
print(f"Sequential processing: {time_seq:.3f}s")
print(f"Parallel processing: {time_par:.3f}s")
print(f"Speedup: {time_seq/time_par:.2f}x")
processor.shutdown()
return {
'sequential_time': time_seq,
'parallel_time': time_par,
'speedup': time_seq / time_par
}