Tensorflow 最终陷入无限循环

Question

此 tensorflow 代码没有响应，我不明白为什么。请帮忙！

import tensorflow as tf
#reading the file
with tf.name_scope ('File_reading') as scope:
    filename_queue = tf.train.string_input_producer(["forestfires.csv.digested"])
    reader = tf.TextLineReader()
    key, value = reader.read(filename_queue)
    record_defaults = [[1.0], [1.0], [1.0], [1.0], [1.0], [1.0], [1.0], [1.0], [1.0], [1.0], [1.0], [1.0], [0.0]]
    #13 decoded
    col1, col2, col3, col4, col5, col6, col7, col8, col9, col10, col11, col12, col13 = tf.decode_csv(
        value, record_defaults=record_defaults) 


    #12 is feture, and the 13th is the training data
    features = tf.stack([col1, col2, col3, col4, col5, col6, col7, col8, col9, col10, col11, col12],name='data_input')

    with tf.Session() as sess:
        # Start populating the filename queue.
        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(coord=coord)
        for i in range(517):
            # Retrieve a single instance:
            example, label = sess.run([features, col13])

        coord.request_stop()
        coord.join(threads)
with tf.name_scope ('network') as scope:
    W1=tf.Variable(tf.zeros([12, 8]), name='W1')
    b1=tf.Variable(tf.zeros([8]), name="b1")
    h1=tf.add(tf.matmul(tf.expand_dims(features,0), W1),b1, name='hidden_layer')
    W2=tf.Variable(tf.zeros([8, 1]), name='W2')
    b2=tf.Variable(tf.zeros([1]), name="b2")
    output=tf.add(tf.matmul(h1, W2),b2, name='output_layer')
error=tf.add(output,-col13, name='error')
#training
train_step = tf.train.AdamOptimizer(1e-4).minimize(error)
#graphing the output
file_writer = tf.summary.FileWriter('some directory', sess.graph)
with tf.Session() as sess:
    #init
    tf.global_variables_initializer().run()
    print ("\n\n\n\n\n\nTRAINING STARTED\n\n\n\n\n\n")
    print('test1')
    sess.run(error) #this statement causes an infinite loop
    print ('test2')
file_writer.close()

代码运行，并打印 'test1'，但随后什么也不做，甚至不响应 ctrl+c。我已经尝试查找问题，但我的 google 技能不够好，或者它不在 Internet 上。系统：win10 英伟达960M python3.5.2

Answer 1

你构建网络的方式在智力上没有意义。如果您需要从 TextLineReader 中读取 517 个步骤，请使用函数 read_up_to 并提供 517 的值，而不是使用单独的会话。在您构建图表的方式中，输入 reader 和图表的其余部分之间似乎没有明确的联系。

我的建议：

# define graph which includes the input queue
def model(...):
...
  return error, metrics

with tf.Graph.as_default():
  error, metrics = model(...)

  with tf.Session():
    # Start Coordinator
    # Initialise global vars
    # Start queue runners
    # model_error, model_metrics = sess.run([error, metrics])

Answer 2

解决了（针对这个bug），不是无限循环，只是等待输入数据。出于某种原因，如果我将上面的 'with tf.Session() as sess:' 块（没有 with 部分）粘贴到下面的 with 块的顶部，它运行得很好。（可能还有其他一些编码错误，因为从那时起我已经更改了一些东西。）

Tensorflow 最终陷入无限循环

Tensorflow ends up in an infinite loop

tensorflow

tensorflow-gpu