Question

我在这里使用简单的教程代码https://www.oreilly.com/learning/hello-tensorflow

这是我的版本：

import tensorflow as tf

x = tf.constant(1.0, name='input')
w = tf.Variable(0.8, name='weight')
y = tf.mul(w, x, name='output')
y_ = tf.constant(0.0, name='correct_value')
loss = tf.pow(y - y_, 2, name='loss')
train_step = tf.train.GradientDescentOptimizer(0.025).minimize(loss)
summary_y = tf.scalar_summary('output', loss)

sess = tf.Session()
sess.run(tf.initialize_all_variables())

summary_writer = tf.train.SummaryWriter('outs')
for i in range(1000):
  summary_writer.add_summary(sess.run(summary_y), i)
  sess.run(train_step)

之后，我在张量板中只有914个1000步。这是一次检查：

tensorboard --inspect --logdir=outs

======================================================================
Processing event files... (this can take a few minutes)
======================================================================

Found event files in:
outs

These tags are in outs:
audio -
histograms -
images -
scalars
   output
======================================================================

Event statistics for outs:
audio -
graph -
histograms -
images -
scalars
   first_step           0
   last_step            913
   max_step             913
   min_step             0
   num_steps            914
   outoforder_steps     []
sessionlog:checkpoint -
sessionlog:start -
sessionlog:stop -
======================================================================

如果我打开tensorboard，我会看到正确的914步（从0到913）的图。无论步数多少，都会发生这种情况。例如，如果我采取100步，则总结中只保存93个。

我正在使用Fedora 23（4.6.5-200.fc23.x86_64 GNU / Linux）。 Tensorflow是虚拟环境中的最新安装

pip list | grep tensorflow
tensorflow (0.10.0rc0)

最后一次迭代丢失的想法是什么？

Answer 1

tf.train.SummaryWriter对象在内部缓冲摘要，因此在退出之前调用summary_writer.close()很重要。您还可以调用summary_writer.flush()以确保已将最近编写的摘要写入事件日志。以下对训练循环的简单修改应该有效：

summary_writer = tf.train.SummaryWriter('outs')
for i in range(1000):
  summary_writer.add_summary(sess.run(summary_y), i)
  sess.run(train_step)
summary_writer.close()

Tensorflow SummaryWriter丢失了迭代次数

1 个答案: