Question

我试图通过检查点保存变量，以便为我的程序引入容错。我试图通过使用MonitoredTrainingSession函数来实现这一目标。以下是我的配置： -

import tensorflow as tf

global_step = tf.Variable(10, trainable=False, name='global_step')
x = tf.constant(2)

with tf.device("/job:local/task:0"):
    y1 = tf.Variable(x + 300)

with tf.device("/job:local/task:1"):
    y2 = tf.Variable(x**2)

with tf.device("/job:local/task:2"):
    y3 = tf.Variable(5*x)

with tf.device("/job:local/task:3"):
    y0 = tf.Variable(x - 66)
    y = y0 + y1 + y2 + y3

model = tf.global_variables_initializer()
saver = tf.train.Saver(sharded=True)

chief = tf.train.ChiefSessionCreator(scaffold=None, master='grpc://localhost:2222', config=None, checkpoint_dir='/home/tensorflow/codes/checkpoints')
summary_hook = tf.train.SummarySaverHook(save_steps=None, save_secs=10, output_dir='/home/tensorflow/codes/savepoints', summary_writer=None, scaffold=None, summary_op=tf.summary.tensor_summary(name="y", tensor=y))
saver_hook = tf.train.CheckpointSaverHook(checkpoint_dir='/home/tensorflow/codes/checkpoints', save_secs=None, save_steps=True, saver=saver, checkpoint_basename='model.ckpt', scaffold=None)

# with tf.train.MonitoredSession(session_creator=ChiefSessionCreator,hooks=[saver_hook, summary_hook]) as sess:

with tf.train.MonitoredTrainingSession(master='grpc://localhost:2222', is_chief=True, checkpoint_dir='/home/tensorflow/codes/checkpoints',
    scaffold=None, hooks=[saver_hook,summary_hook], chief_only_hooks=None, save_checkpoint_secs=None, save_summaries_steps=True, config=None) as sess:

    while not sess.should_stop():
        sess.run(tf.global_variables_initializer())

    while not sess.should_stop():
        result = sess.run(y)
        print(result)

我收到以下 RuntimeError ，我无法解决： -

Traceback (most recent call last):
  File "add_1.py", line 39, in <module>
    sess.run(tf.global_variables_initializer())
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variables.py", line 1187, in global_variables_initializer
    return variables_initializer(global_variables())
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variables.py", line 1169, in variables_initializer
    return control_flow_ops.group(*[v.initializer for v in var_list], name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 2773, in group
    deps.append(_GroupControlDeps(dev, ops_on_device[dev]))
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.py", line 2721, in _GroupControlDeps
    return no_op(name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_control_flow_ops.py", line 186, in no_op
    result = _op_def_lib.apply_op("NoOp", name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/op_def_library.py", line 759, in apply_op
    op_def=op_def)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 2199, in create_op
    self._check_not_finalized()
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1925, in _check_not_finalized
    raise RuntimeError("Graph is finalized and cannot be modified.")
RuntimeError: Graph is finalized and cannot be modified.

Answer 1

您的错误的根本原因似乎是MonitoredTrainingSession已完成（冻结）图表，而您的tf.global_variable_initializer()无法再对其进行修改。

话虽如此，有许多事情需要注意：

1）为什么你试着在这里重复初始化所有变量？

while not sess.should_stop():
    sess.run(tf.global_variables_initializer())

2）您的某些代码似乎已包含在MonitoredTrainingSession中，例如ChiefSessionCreator。您能再看一下代码（https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/training/monitored_session.py#L243）或搜索其示例用法并查看应该如何使用MonitoredTrainingSession吗？

Answer 2

这可能不建议用于您的用例，但可以unfinalize a Graph：

sess.graph._unsafe_unfinalize()

Answer 3

如果要在循环上初始化图形，可以使用该函数在循环顶部创建新图形。

import tensorflow as tf

tf.reset_default_graph()
tf.Graph().as_default()

Answer 4

由于您的目标是使用MonitoredTrainingSession来检查点，因此使用方法比您的示例简单得多：

import tensorflow as tf

global_step = tf.contrib.framework.get_or_create_global_step()
x = tf.constant(2)
y1 = x + 300
y2 = x**2
y3 = x * 5
y0 = x - 66
y = y0 + y1 + y2 + y3
step = tf.assign_add(global_step, 1)

with tf.train.MonitoredTrainingSession(checkpoint_dir='/tmp/checkpoints') as sess:
    while not sess.should_stop():
        result, i = sess.run([y, step])
        print(result, i)

保存/恢复检查点的挂钩由MonitoredTrainingSession为您创建。
如果您传入save_checkpoint_secs，则可以从默认的10分钟更改检查点的频率。我发现更高的频率是不值得的：保存检查点不是免费的，因此非常频繁的检查点最终会减慢训练速度。
ChiefSessionCreator和gRPC配置仅用于分布式运行（有关概念的说明，请参阅here。类似于将操作分配给特定设备 - 确保在使用前确实需要执行此操作如果你不小心，它可以减慢速度。
您不需要在tf.Variable()的张量上包含操作结果 - 它们已经是变量。
您可以通过save_summaries_steps传递带有张量板的监控培训，但默认情况下，无论如何都会每100步执行一次。

Tensorflow：图表已完成，无法修改

4 个答案: