CODEWITHSUNDEEP
tensorflow
brown.2179
brown.2179

Reputation: 1880

When should I initialise the local variables?

I am a bit confused by tf.local_variables_initializer. I am not sure when to call it.

I see in other peoples code where they just slather on calls like

init_op = tf.group(tf.global_variables_initializer(),
                   tf.local_variables_initializer())

Should init_op just be called in every TF program? Is this "lazy" TF style of programming?

As an example, consider queues like input_string_producer specified as :

tf.input_string_producer(file_list, num_epochs=None)

does not require tf.local_variables_initializer()

However, when specifying actual number of epochs

tf.input_string_producer(file_list,num_epochs=1)

does in fact require that tf.local_variables_initializer() is executed.

How does the programmer know when (or when not) to initialise such "hidden" local variables? Shouldn't the FIFOQueue init its own variables or have something like queue.initializer?

Upvotes: 3

Views: 665

Answers (1)

Yaroslav Bulatov
Yaroslav Bulatov

Reputation: 57893

For local tensorflow jobs, you need to initialize both local and global. For distributed TensorFlow jobs, the chief worker initializes local+global, whereas remaining workers initialize only local.

Upvotes: 1

Related Questions