Question

上下文

我是TensorFlow的新手，我试图在this paper中实现一些算法，这些算法需要偶尔从全局共享模型复制到本地线程特定模型。

问题：

完成上述任务的最佳/正确方法是什么？我已经提供了一个虚拟示例，说明了我目前正在执行此操作的方式以及我得到的错误。有人可以解释为什么会发生错误吗？

import tensorflow as tf
import threading

class ExampleModel(object):
  def __init__(self, graph):
    with graph.as_default():
      self.w = tf.Variable(tf.constant(1, shape=[1,2]))

sess = tf.InteractiveSession()
graph = tf.get_default_graph()
global_network = ExampleModel(graph)
sess.run(tf.initialize_all_variables())

def example(i):
  global global_network, graph
  local_network = ExampleModel(graph)
  sess.run(local_network.w.assign(global_network.w))

threads = []
for i in range(5):
  t = threading.Thread(target=example, args=(i,))
  threads.append(t)

for t in threads:
  t.start()

错误：

Exception in thread Thread-3:
Traceback (most recent call last):
  File "/Users/kennyhsu5/anaconda/lib/python2.7/threading.py", line 801, in __bootstrap_inner
    self.run()
  File "/Users/kennyhsu5/anaconda/lib/python2.7/threading.py", line 754, in run
    self.__target(*self.__args, **self.__kwargs)
  File "tmp.py", line 16, in example
    local_network = ExampleModel(graph)
  File "tmp.py", line 7, in __init__
    self.w = tf.Variable(tf.constant(1, shape=[1,2]))
  File "/Users/kennyhsu5/anaconda/lib/python2.7/site-packages/tensorflow/python/ops/variables.py", line 211, in __init__
dtype=dtype)
  File "/Users/kennyhsu5/anaconda/lib/python2.7/site-packages/tensorflow/python/ops/variables.py", line 319, in _init_from_args
    self._snapshot = array_ops.identity(self._variable, name="read")
  File "/Users/kennyhsu5/anaconda/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2976, in __exit__
    self._graph._pop_control_dependencies_controller(self)
  File "/Users/kennyhsu5/anaconda/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2996, in _pop_control_dependencies_controller
    assert self._control_dependencies_stack[-1] is controller
AssertionError

Answer 1

关于Tensorflow中的tf.Graph类：

重要说明：此类对于图构建不是线程安全的。应该从单个线程或外部创建所有操作必须提供同步。除非另有说明，否则全部方法不是线程安全的。

self.w = ...声明和local_network.w.assign(...)操作导致错误。

我知道它基本上会杀死你的多线程实现，但你可以通过将这些声明移动到主线程来修复上面的代码。然后，您可以使用线程实际运行您规定的操作。例如：

import tensorflow as tf
import threading

class ExampleModel(object):
  def __init__(self, graph):
    with graph.as_default():
      self.w = tf.Variable(tf.constant(1, shape=[1,2]))

sess = tf.InteractiveSession()
graph = tf.get_default_graph()
global_network = ExampleModel(graph)
sess.run(tf.global_variables_initializer())

def example(sess, assign_w):
  sess.run(assign_w)

threads = []
for i in range(5):
  local_network = ExampleModel(graph)
  assign_w = local_network.w.assign(global_network.w)
  t = threading.Thread(target=example, args=(sess, assign_w))
  threads.append(t)

for t in threads:
  t.start()

我还建议您通过args参数将变量传递给线程，而不是使用global。

最后，请考虑使用global_variables_initializer而不是弃用initialize_all_variables。

如何将参数从全局模型复制到特定于线程的模型

上下文

问题：

错误：

1 个答案: