当我增加将数据推送到它的线程数时，我的tensorflow队列的填充速度变慢

Question

我已经写了一些代码将数据推送到tensorflow中的队列中，我的队列处理程序的初始化以及所有线程运行的主要功能如下：

def __init__(self):
    self.X = tf.placeholder(tf.int64)
    self.Y = tf.placeholder(tf.int64)
    self.queue = tf.RandomShuffleQueue(dtypes=[tf.int64, tf.int64],
                                       capacity=100,
                                       min_after_dequeue=20)

    self.enqueue_op = self.queue.enqueue([self.X, self.Y])


def thread_main(self, sess, coord):
    """Cycle through the dataset until the main process says stop."""
    train_fs = open(data_train, 'r')
    while not coord.should_stop():
        X_, Y_ = get_batch(train_fs)
        if not Y: #We're at the end of the file
            train_fs = open(data_train, 'r')
            X, Y = get_batch(train_fs)
        sess.run(self.enqueue_op, feed_dict={self.X:X_, self.Y:Y_}) 

在培训期间，我正在监视队列的大小。由于某些原因，当我增加向其中推送数据的线程数量时，队列的填充速度会变慢。知道为什么吗？是因为我正在同时读取python文件吗？

编辑：

这是我正在使用的代码，在数据和图形旁边它是完全相同的。该代码在此虚拟数据上的行为符合预期。我有两个观察结果：

• 我认为我没有适当地关闭线程，似乎它们在执行后会卡在队列中，而我运行的代码越多，它就会变得越慢。
• 由于多线程在这里发挥作用，我猜失败的仅有两点是图形和读取数据的方式。

首先，生成一个虚拟数据集：

data_train = "./test.txt"

with open(data_train, 'w') as out_stream:
    out_stream.write("""[1,2,3,4,5,6]|1\n[1,2,3,4]|2\n[1,2,3,4,5,6]|0\n[1,2,3,4,5,6]|1\n[1,2,5,6]|1\n[1,2,5,6]|0""")

def get_batch(fs):
    line = fs.readline()
    X, Y = line.split('|')
    X = eval(X)
    Y = eval(Y)
    return X, Y

然后，队列控制器：

import tensorflow as tf
import numpy as np
import threading

tf.reset_default_graph()#Reset the graph essential to use with jupyter else variable conflicts

class QueueCtrl(object):

    def __init__(self):
        self.X = tf.placeholder(tf.int64)
        self.Y = tf.placeholder(tf.int64)
        self.queue = tf.RandomShuffleQueue(dtypes=[tf.int64, tf.int64],
                                           capacity=100,
                                           min_after_dequeue=20)

        self.enqueue_op = self.queue.enqueue([self.X, self.Y])


    def thread_main(self, sess, coord):
        """Cycle through the dataset until the main process says stop."""
        train_fs = open(data_train, 'r')
        while not coord.should_stop():
            X_, Y_ = get_batch(train_fs)
            if not Y_: #We're at the end of the file
                train_fs = open(data_train, 'r')
                X_, Y_ = get_batch(train_fs)
            sess.run(self.enqueue_op, feed_dict={self.X:X_, self.Y:Y_}) 

    def get_batch_from_queue(self):
        """
        Return one batch
        """
        return self.queue.dequeue()

    def start_threads(self, sess, coord, num_threads=2):
        """Start the threads"""
        threads = []
        for _ in range(num_threads):
            t = threading.Thread(target=self.thread_main, args=(sess, coord))
            t.daemon = True
            t.start()
            threads.append(t)
        return threads

然后我们建立一个虚拟图：

queue_ctrl = QueueCtrl()
X_, Y_ = queue_ctrl.get_batch_from_queue()
output = Y_ * tf.reduce_sum(X_)
init = tf.initialize_all_variables()

最后，我们遍历数据：

sess = tf.Session()

sess.run(init)
coord = tf.train.Coordinator()
tf.train.start_queue_runners(sess=sess, coord=coord)
my_thread = queue_ctrl.start_threads(sess, coord, num_threads=6)

for i in range(100):
    out = sess.run(output)
    print("Iter: %d, output: %d, Element in queue: %d" 
              % (i, out, sess.run(queue_ctrl.queue.size())))

coord.request_stop()
for _ in range(len(my_thread)): #if the queue is full at that time then the threads won't see the coord.should_stop
    _ = sess.run([output])

coord.join(my_thread, stop_grace_period_secs=10)
sess.close()

这是带有五个线程的25个第一输出：

Iter: 0, output: 21, Element in queue: 27
Iter: 1, output: 21, Element in queue: 37
Iter: 2, output: 20, Element in queue: 51
Iter: 3, output: 21, Element in queue: 67
Iter: 4, output: 20, Element in queue: 81
Iter: 5, output: 20, Element in queue: 89
Iter: 6, output: 21, Element in queue: 100
Iter: 7, output: 20, Element in queue: 100
Iter: 8, output: 20, Element in queue: 100
Iter: 9, output: 21, Element in queue: 100
Iter: 10, output: 20, Element in queue: 100
Iter: 11, output: 20, Element in queue: 100
Iter: 12, output: 21, Element in queue: 100
Iter: 13, output: 21, Element in queue: 100
Iter: 14, output: 20, Element in queue: 100
Iter: 15, output: 20, Element in queue: 100
Iter: 16, output: 21, Element in queue: 100
Iter: 17, output: 21, Element in queue: 100
Iter: 18, output: 20, Element in queue: 100
Iter: 19, output: 21, Element in queue: 100
Iter: 20, output: 21, Element in queue: 100
Iter: 21, output: 21, Element in queue: 100
Iter: 22, output: 20, Element in queue: 100
Iter: 23, output: 21, Element in queue: 100
Iter: 24, output: 21, Element in queue: 100
Iter: 25, output: 21, Element in queue: 100

有一个线程：

Iter: 0, output: 21, Element in queue: 22
Iter: 1, output: 20, Element in queue: 25
Iter: 2, output: 20, Element in queue: 27
Iter: 3, output: 20, Element in queue: 29
Iter: 4, output: 21, Element in queue: 31
Iter: 5, output: 20, Element in queue: 32
Iter: 6, output: 20, Element in queue: 34
Iter: 7, output: 21, Element in queue: 35
Iter: 8, output: 21, Element in queue: 36
Iter: 9, output: 21, Element in queue: 38
Iter: 10, output: 20, Element in queue: 40
Iter: 11, output: 20, Element in queue: 42
Iter: 12, output: 20, Element in queue: 43
Iter: 13, output: 21, Element in queue: 46
Iter: 14, output: 20, Element in queue: 47
Iter: 15, output: 21, Element in queue: 48
Iter: 16, output: 20, Element in queue: 53
Iter: 17, output: 20, Element in queue: 56
Iter: 18, output: 21, Element in queue: 57
Iter: 19, output: 21, Element in queue: 61
Iter: 20, output: 21, Element in queue: 63
Iter: 21, output: 20, Element in queue: 67
Iter: 22, output: 21, Element in queue: 70
Iter: 23, output: 21, Element in queue: 73
Iter: 24, output: 20, Element in queue: 76
Iter: 25, output: 20, Element in queue: 78

Answer 1

只是想在这里添加一些东西，我实现了一个用于多任务学习的基于多进程的数据馈送管道。它可以实现平均值。GPU利用率> 90％，四核CPU利用率> 95％。不太容易出现内存泄漏，特别适合进行为期一天的训练。并不是说它是完美的，但在我看来至少比当前的TF队列API（1.1）更好。

如果有人感兴趣：https : //hanxiao.github.io/2017/07/07/Get-10x-Speedup-in-Tensorflow-Multi-Task-Learning-using-Python-Multiprocessing/