tensorflow摘要需要提供占位符,但我无法理解为什么

时间:2017-04-28 11:27:00

标签: python tensorflow summary

我在更深入地测试摘要之前,我有以下剪切代码

import tensorflow as tf
import numpy as np

def test_placeholders():
    "Simply dump a placeholder to TensorBoard"
    x = tf.placeholder(tf.float32, [])
    sess = tf.Session()

    summary = tf.summary.scalar("x", x)
    train_writer = tf.summary.FileWriter('/tmp/tf/placeholder',
                                         sess.graph, flush_secs=1)

    r = sess.run(tf.global_variables_initializer())
    s = sess.run(summary, feed_dict={x: 1.57})
    train_writer.add_summary(s)

    train_writer.close()


def test_merge():
    "A simple function that make a loop computation and write down into TB"

    x = tf.placeholder(tf.float32)
    k = np.random.random() + 0.1

    # Create a session
    sess = tf.Session()
    sess.run(tf.global_variables_initializer())

    # define a single summary
    summary_x = tf.summary.scalar("x", x)

    train_writer = tf.summary.FileWriter('/tmp/tf/foo',
                                         sess.graph, flush_secs=1)

    # write some summaries
    for i in range(0, 5):
        # WORKS!
        summary = sess.run(summary_x, feed_dict={x: k * i * i})
        train_writer.add_summary(summary, i)

    # write some summaries using merge_all
    # (we have only one define summary)
    merged = tf.summary.merge_all()
    for i in range(5, 10):
        # FAILS: You must feed a value for placeholder ...
        summary = sess.run(merged, feed_dict={x: k * i * i})
        train_writer.add_summary(summary, i)

    train_writer.close()


if __name__ == '__main__':

    test_placeholders()    # if I comment this line ...
    test_merge()           # test_merge() works!?

所以基本上有两个函数可以创建一些循环并为TensorBoard写一些日志。

问题:

每个函数都可以很好地相互隔离,但是,当我按顺序运行时,第二个函数在这里失败

# FAILS: You must feed a value for placeholder ...
summary = sess.run(merged, feed_dict={x: k * i * i})

因为似乎合并包含来自前一个未填充的函数的内容。

tensorflow.python.framework.errors_impl.InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder' with dtype float
[[Node: Placeholder = Placeholder[dtype=DT_FLOAT, shape=[], _device="/job:localhost/replica:0/task:0/cpu:0"]()]]
Caused by op u'Placeholder', defined at:

深入研究代码,我发现TF将便利变量存储到默认容器中,例如:来自之前作品的 _collections ,因此拨打电话

tf.reset_default_graph()

用作重置之前执行的所有内容。

问题:

用于隔离和处理多次TF执行的 tensorflow样式是什么,它们在同一进程中不会产生干扰?

2 个答案:

答案 0 :(得分:1)

您遇到的问题与加载到同一图表的张量有关。 注意到test_merge包含merged = tf.summary.merge_all()这会合并默认图表中收集的所有摘要,并且所有内容都会加载到默认图表中,因此当您尝试评估summary = sess.run(merged, feed_dict={x: k * i * i})时,它需要来自第一个函数的输入同样。如果更改了调用的顺序,您将看到代码执行。如果您需要单独的图表,则可能会出现问题,因此请尝试将所有内容加载到一个图表中 - 但如果您需要,则此答案可能有用Working with multiple graphs in TensorFlow

答案 1 :(得分:0)

使用tf.reset_default_graph()
在代码的开头,将图形恢复为其默认值。有时Tensorflow会记住上一次运行的feed_dict。并且该feed_dict可能与当前会话不同