tensorflow-tf.control_dependencies()作用及用法

Jul 2, 2020 · tensorflow 机器学习 ·

分享到:

tensorflow tf.control_dependencies()作用及用法

在有些机器学习程序中我们想要指定某些操作执行的依赖关系，这时我们可以使用tf.control_dependencies()来实现。 control_dependencies(control_inputs)返回一个控制依赖的上下文管理器，使用with关键字可以让在这个上下文环境中的操作都在control_inputs 执行。

原型分析

1tf.control_dependencies(self, control_inputs)

python

arguments：control_inputs: A list of Operation or Tensor objects which must be executed or computed before running the operations defined in the context. （注意这里control_inputs是list） return： A context manager that specifies control dependencies for all operations constructed within the context.

1with g.control_dependencies([a, b, c]):
2  # `d` and `e` will only run after `a`, `b`, and `c` have executed.
3  d = ...
4  e = ...

python

可以嵌套control_dependencies 使用

1with g.control_dependencies([a, b]):
2  # Ops constructed here run after `a` and `b`.
3  with g.control_dependencies([c, d]):
4    # Ops constructed here run after `a`, `b`, `c`, and `d`.

python

可以传入None 来消除依赖：

1with g.control_dependencies([a, b]):
2  # Ops constructed here run after `a` and `b`.
3  with g.control_dependencies(None):
4    # Ops constructed here run normally, not waiting for either `a` or `b`.
5    with g.control_dependencies([c, d]):
6      # Ops constructed here run after `c` and `d`, also not waiting
7      # for either `a` or `b`.

python

注意：控制依赖只对那些在上下文环境中建立的操作有效，仅仅在context中使用一个操作或张量是没用的

 1# WRONG
 2def my_func(pred, tensor):
 3  t = tf.matmul(tensor, tensor)
 4  with tf.control_dependencies([pred]):
 5    # The matmul op is created outside the context, so no control
 6    # dependency will be added.
 7    return t
 8
 9# RIGHT
10def my_func(pred, tensor):
11  with tf.control_dependencies([pred]):
12    # The matmul op is created in the context, so a control dependency
13    # will be added.
14    return tf.matmul(tensor, tensor)

python

例子：在训练模型时我们每步训练可能要执行两种操作，op a, b 这时我们就可以使用如下代码：

1with tf.control_dependencies([a, b]):
2    c= tf.no_op(name='train')#tf.no_op；什么也不做
3sess.run(c)
4
5# 在这样简单的要求下，可以将上面代码替换为：
6c= tf.group([a, b])
7sess.run(c)

python

其他关于tf.identity()的奇怪操作可见https://blog.csdn.net/u012436149/article/details/72084744

使用tf.no_op()是一个占位符，表示什么都不做，但是会返回一个operation，用以保证tf.control_dependencies()被执行，和tf.group操作类似。