具有加载良好训练模型的多处理.(通过 Python3)

Question

这是我的情况： 我有一个训练有素的语音合成模型。我将通过多处理加快合成速度，在每个CPU中预加载模型，然后继续输入句子以进行文本到语音转换。

这是我的尝试脚本：

####################################################################
#!/usr/bin/python3
####################################################################
from multiprocessing import Process, Pool, cpu_count
import os,time
####################################################################
from tacotron.demo_synthesizer import Synthesizer
from splitting_sent import splitting_para
import tensorflow as tf
from datasets import audio
####################################################################
from pypinyin import pinyin, Style
####################################################################
BASE_DIR = os.path.split(os.path.realpath(__file__))[0]
VOICE = BASE_DIR + "/tmp"
TXT = BASE_DIR + "/txt"
os.makedirs(VOICE, exist_ok=True)
os.makedirs(TXT, exist_ok=True)
####################################################################
def syn(py):
synthesizer = Synthesizer()
synthesizer.load("/path/to/the/model")
wav_name = time.time()
wav_path = VOICE + "/" + str(wav_name)
wav = synthesizer.synthesize(py)
audio.save_wav(wav, wav_path)
if __name__ =='__main__':
with open(os.path.join(TXT, "content.txt"), "r") as f:
lines = f.read().splitlines()
lines = "".join(lines)
sentences = splitting_para(lines)
# splitting paragraph into individual sentences.
py_list = []
for sent in sentences:
py_sent = pinyin(sent, style=Style.TONE3)
py_sent = " ".join([i[0] for i in py_sent if i[0].isalnum()])
py_list.append(py_sent)
# as I am trying the Chinese TTS, it is the inevitable prerequisite step of translating Chinese character into Pinyin.
print('Run the main process (%s).' % (os.getpid()))
mainStart = time.time()
p = Pool(cpu_count())
for py in py_list:
p.apply_async(syn,args=(py,))
print('Waiting for all subprocesses done ...')
p.close()
p.join()
print('All subprocesses done')
mainEnd = time.time()
print('All process ran %0.2f seconds.' % (mainEnd-mainStart))

我被困在这个问题上：我只能将 12 个模型预加载到 12 个进程中以合成随机句子。但是，不可能继续将接下来的 12 个句子输入到预加载模型中。在完成第一组 12 句话后，这些过程终止。我完全迷失在这里。:(

如果有任何建议，我非常感谢。 :)

Answer 1

我从我的朋友那里得到了答案，如下：

####################################################################
import multiprocessing
from multiprocessing import Process, Pool, cpu_count, Queue
import os, time, sys
####################################################################
from tacotron.demo_synthesizer import Synthesizer
from splitting_sent import splitting_para
import tensorflow as tf
from datasets import audio
####################################################################
from pypinyin import pinyin, Style
####################################################################
BASE_DIR = os.path.split(os.path.realpath(__file__))[0]
VOICE = BASE_DIR + "/tmp"
TXT = BASE_DIR + "/txt"
os.makedirs(VOICE, exist_ok=True)
os.makedirs(TXT, exist_ok=True)
####################################################################
num_cpu = cpu_count() - 1
q = []
q_re = []
for tmp in range(num_cpu):
q.append(Queue())
def syn(id):
synthesizer = Synthesizer()
synthesizer.load("/home/chris/Pictures/tts-server/logs-Tacotron/model.ckpt-1255000")
print("LOADED at {} CPU".format(id))
while not q[id].empty():
input_py = q[id].get(True)
print("Starting decode:", id)
wav_name = time.time()
wav_path = VOICE + "/" + str(wav_name)
wav = synthesizer.synthesize(input_py)
audio.save_wav(wav, wav_path)
print("Decoded:", id)
if __name__ =='__main__':
with open(os.path.join(TXT, "content.txt"), "r") as f:
lines = f.read().splitlines()
lines = "".join(lines)
sentences = splitting_para(lines)
py_list = []
for sent in sentences:
py_sent = pinyin(sent, style=Style.TONE3)
py_sent = " ".join([i[0] for i in py_sent if i[0].isalnum()])
py_list.append(py_sent)
for x in range(num_cpu):
p = multiprocessing.Process(target=syn, args=(x,))
p.start()
# decoding
for index, py in enumerate(py_list):
q[index % num_cpu].put(py)