How to add control dependency to Tensorflow op

问题:

I want update_op to run before I run summary. Sometimes I just create a tf.summary, and everything works just fine, but sometimes I want to do more fancy stuff, but still have the same control dependency.

Code that doesn't work:

with tf.control_dependencies([update_op]):
    if condition:
        tf.summary.scalar('summary', summary)
    else:
        summary = summary

Bad hack that works

with tf.control_dependencies([update_op]):
    if condition:
        tf.summary.scalar('summary', summary)
    else:
        summary += 0

The problem is that summary=summary doesn't create a new node, so the control dependency is ignored.

---

I am sure that there is a way better way of going about this, any suggestions? :-)

回答1:

I don't think there exists a more elegant solution to this, because this the designed behavior. tf.control_dependencies is a shortcut of tf.Graph.control_dependencies call using a default graph, and here's the quote from its documentation:

N.B. The control dependencies context applies only to ops that are constructed within the context. Merely using an op or tensor in the context does not add a control dependency. The following example illustrates this point:

# WRONG
def my_func(pred, tensor):
  t = tf.matmul(tensor, tensor)
  with tf.control_dependencies([pred]):
    # The matmul op is created outside the context, so no control
    # dependency will be added.
    return t

# RIGHT
def my_func(pred, tensor):
  with tf.control_dependencies([pred]):
    # The matmul op is created in the context, so a control dependency
    # will be added.
    return tf.matmul(tensor, tensor)

So just use tf.identity(summary), as suggested in the comments.