CODEWITHSUNDEEP
pythontensorflow
R zu
R zu

Reputation: 2064

control_dependencies not as expected

I want tensorflow to do the following in f(...)

  1. get data[index]
  2. cache the value
  3. return data[index]

But tf.control_dependencies doesn't do what I want.

How to fix the control dependency?

Result:

cache_ 0.0
x_ 2.0
AssertionError

Test:

import tensorflow as tf
import numpy as np


def f(a, cache):
    assign_op = tf.assign(cache, a)
    with tf.control_dependencies([assign_op]):
        return a


def main():
    dtype = np.float32
    data = tf.range(5, dtype=dtype)
    cache = tf.Variable(0, dtype=dtype)
    x = f(data[2], cache)
    init_op = tf.global_variables_initializer()
    with tf.Session() as sess:
        sess.run(init_op)
        x_ = sess.run(x)
        cache_ = sess.run(cache)
    print("cache_", cache_)
    print("x_", x_)
    assert np.allclose(cache_, x_)


main()

Upvotes: 0

Views: 58

Answers (1)

McAngus
McAngus

Reputation: 1856

The problem is that return a is Python code. You are not creating any TensorFlow ops in the with block. You can use tf.identity to create an op that will ensure that when a is read from assign_op will be executed first. Here is the updated code:

def f(a, cache):
    assign_op = tf.assign(cache, a)
    with tf.control_dependencies([assign_op]):
        return tf.identity(a)

Upvotes: 2

Related Questions