keras使用多进程

最近在工作中有一个需求：用训练好的模型将数据库中所有数据得出预测结果，并保存到另一张表上。数据库中的数据是一篇篇文章，我训练好的模型是对其中的四个段落分别分类，即我有四个模型，拿到文本后需要提取出这四个段落，并用对应模型分别预测这四个段落的类别，然后存入数据库中。我是用keras训练的模型，backend为tensorflow，因为数据量比较大，自然想到用多进程，但是使用时发现每次都在model.predict上停在那不动了。在Windows上运行一点问题没有，但是在Linux服务器上就有这个问题

模型使用时大致如下：

# -*- coding: utf-8 -*-
import jieba
import numpy as np
import keras
import tensorflow as tf
from keras.preprocessing import sequence
from keras.models import load_model
from config import Config
import json


config_file = ‘data/config.ini‘
model_path = Config(config_file).get_value_str(‘cnn‘, ‘model_path‘)
graph = tf.Graph()
with graph.as_default():
    session = tf.Session()
    with session.as_default():
        model = load_model(model_path)

graph_var = graph
session_var = session


def sentence_process(sentence):
    with open(‘data/words.json‘, encoding=‘utf-8‘) as f:
        words_json = json.load(f)
    words = words_json[‘words‘]
    word_to_id = words_json[‘word_to_id‘]
    max_length = words_json[‘max_length‘]
    segs = jieba.lcut(sentence)
    segs = filter(lambda x: len(x) >= 1, segs)
    segs = [x for x in segs if x]
    vector = []
    for seg in segs:
        if seg in words:
            vector.append(word_to_id[seg])
        else:
            vector.append(4999)
    return vector, max_length


def predict(sentence):
    vector, max_length = sentence_process(sentence)
    vector_np = np.array([vector])
    x_vector = sequence.pad_sequences(vector_np, max_length)
    with graph_var.as_default():
        with session_var.as_default():
            y = model.predict_proba(x_vector)
            if y[0][1] > 0.5:
                predict = 1
            else:
                predict = 0
    return predict

多进程使用大致如下：

from multiprocessing import Pool
from classifaction.classify1 import predict1
from classifaction.classify2 import predict2
from classifaction.classify3 import predict3
from classifaction.classify4 import predict4


def main():
    ‘‘‘
    get texts
    ‘‘‘
    pool = Pool(processes=4, maxtasksperchild=1)
    pool.map(save_to_database, texts)
    pool.close()
    pool.join()


def save_to_database(texts):
    text1, text2, text3, text4 = texts[0], texts[1], texts[2], texts[3]
    label1 = predict1(text1)
    label2 = predict2(text2)
    label3 = predict3(text3)
    label4 = predict4(text4)


if __name__ == ‘__main__‘:
    main()

运行时发现所有进程都停在model.predict上不动了。

Google后发现很多遇到这个问题，也终于找到一个方法。可以看一下链接：

https://github.com/keras-team/keras/issues/9964

有一个方法是

As of TF 1.10, the library seems to be somewhat forkable. So you will have to test what you can do.

Also, something you can try is:
multiprocessing.set_start_method(‘spawn‘, force=True) if you‘re on UNIX and using Python3.

即在使用multiprocessing之前先设置一下。更改后代码如下：

import multiprocessing
from multiprocessing import Pool
from classifaction.classify1 import predict1
from classifaction.classify2 import predict2
from classifaction.classify3 import predict3
from classifaction.classify4 import predict4


def main():
    ‘‘‘
    get texts
    ‘‘‘
    pool = Pool(processes=4, maxtasksperchild=1)
    multiprocessing.set_start_method(‘spawn‘, force=True)
    pool.map(save_to_database, texts)
    pool.close()
    pool.join()


def save_to_database(texts):
    text1, text2, text3, text4 = texts[0], texts[1], texts[2], texts[3]
    label1 = predict1(text1)
    label2 = predict2(text2)
    label3 = predict3(text3)
    label4 = predict4(text4)


if __name__ == ‘__main__‘:
    main()

这样就可以用多进程了