human/human/audio_inference_handler.py
2024-10-31 08:15:49 +08:00

136 lines
5.0 KiB
Python

#encoding = utf8
import logging
import os
import queue
import time
from queue import Queue
from threading import Event, Thread
import numpy as np
import torch
from human_handler import AudioHandler
from utils import load_model, mirror_index, get_device, SyncQueue
logger = logging.getLogger(__name__)
current_file_path = os.path.dirname(os.path.abspath(__file__))
class AudioInferenceHandler(AudioHandler):
def __init__(self, context, handler):
super().__init__(context, handler)
self._mal_queue = Queue()
self._audio_queue = SyncQueue(context.batch_size * 2)
self._exit_event = Event()
self._run_thread = Thread(target=self.__on_run)
self._exit_event.set()
self._run_thread.start()
logger.info("AudioInferenceHandler init")
def on_handle(self, stream, type_):
if type_ == 1:
self._mal_queue.put(stream)
elif type_ == 0:
self._audio_queue.put(stream)
print('AudioInferenceHandler on_handle', type_, self._audio_queue.size())
def on_message(self, message):
super().on_message(message)
def __on_run(self):
wav2lip_path = os.path.join(current_file_path, '..', 'checkpoints', 'wav2lip.pth')
logger.info(f'AudioInferenceHandler init, path:{wav2lip_path}')
model = load_model(wav2lip_path)
logger.info("Model loaded")
face_list_cycle = self._context.face_list_cycle
length = len(face_list_cycle)
index = 0
count = 0
count_time = 0
logger.info('start inference')
device = get_device()
logger.info(f'use device:{device}')
while True:
if self._exit_event.is_set():
start_time = time.perf_counter()
batch_size = self._context.batch_size
try:
mel_batch = self._mal_queue.get(block=True, timeout=1)
except queue.Empty:
continue
print('origin mel_batch:', len(mel_batch))
is_all_silence = True
audio_frames = []
for _ in range(batch_size * 2):
frame, type_ = self._audio_queue.get()
print('AudioInferenceHandler type_', type_)
audio_frames.append((frame, type_))
if type_ == 0:
is_all_silence = False
if is_all_silence:
for i in range(batch_size):
self.on_next_handle((None, mirror_index(length, index), audio_frames[i * 2:i * 2 + 2]),
0)
index = index + 1
else:
logger.info('infer=======')
t = time.perf_counter()
img_batch = []
for i in range(batch_size):
idx = mirror_index(length, index + i)
face = face_list_cycle[idx]
img_batch.append(face)
print('orign img_batch:', len(img_batch), 'origin mel_batch:', len(mel_batch))
img_batch = np.asarray(img_batch)
mel_batch = np.asarray(mel_batch)
img_masked = img_batch.copy()
img_masked[:, face.shape[0] // 2:] = 0
img_batch = np.concatenate((img_masked, img_batch), axis=3) / 255.
mel_batch = np.reshape(mel_batch,
[len(mel_batch), mel_batch.shape[1], mel_batch.shape[2], 1])
img_batch = torch.FloatTensor(np.transpose(img_batch, (0, 3, 1, 2))).to(device)
mel_batch = torch.FloatTensor(np.transpose(mel_batch, (0, 3, 1, 2))).to(device)
print('img_batch:', img_batch.shape, 'mel_batch:', mel_batch.shape)
with torch.no_grad():
pred = model(mel_batch, img_batch)
pred = pred.cpu().numpy().transpose(0, 2, 3, 1) * 255.
count_time += (time.perf_counter() - t)
count += batch_size
if count >= 100:
logger.info(f"------actual avg infer fps:{count / count_time:.4f}")
count = 0
count_time = 0
for i, res_frame in enumerate(pred):
self.on_next_handle(
(res_frame, mirror_index(length, index), audio_frames[i * 2:i * 2 + 2]),
0)
index = index + 1
logger.info(f'total batch time: {time.perf_counter() - start_time}')
else:
time.sleep(1)
break
logger.info('AudioInferenceHandler inference processor stop')
def stop(self):
self._exit_event.clear()
self._run_thread.join()
def pause_talk(self):
pass
# self._audio_queue.clear()
# self._mal_queue.queue.clear()