CODEWITHSUNDEEP
pythontensorflowpytesttensorflow2.0
joel
joel

Reputation: 7868

How to reset initialization in TensorFlow 2

If I try to change parallelism in TensorFlow 2 after initializing a tf.Variable,

import tensorflow as tf
_ = tf.Variable([1])
tf.config.threading.set_inter_op_parallelism_threads(1)

I get an error

RuntimeError: Inter op parallelism cannot be modified after initialization.

I understand why that could be, but it (and possibly other factors) are causing my tests to interfere with each other. For example

def test_model():  # this test
   v = tf.Variable([1])
   ...

def test_threading():  # is breaking this test
   tf.config.threading.set_inter_op_parallelism_threads(1)
   ...

How do I reset the TensorFlow state so that I can set the threading?

Upvotes: 8

Views: 4704

Answers (2)

Sarah Masud
Sarah Masud

Reputation: 158

Expanding on the solution provided by @thushv89, we can also directly set the context by

from tensorflow.python.eager import context
context.context().intra_op_parallelism_threads = <num_threads>
context.context().inter_op_parallelism_threads = <num_threads>
import tensorflow as tf

Details of the get() and set() thread function available here

Upvotes: 0

thushv89
thushv89

Reputation: 11333

This is achievable in a "hacky" way. But I'd recommend doing this the right way (i.e. by setting up config at the beginning).

import tensorflow as tf
from tensorflow.python.eager import context

_ = tf.Variable([1])

context._context = None
context._create_context()

tf.config.threading.set_inter_op_parallelism_threads(1)

Edit: What is meant by setting up config at the beginning,

import tensorflow as tf
from tensorflow.python.eager import context

tf.config.threading.set_inter_op_parallelism_threads(1)
_ = tf.Variable([1])

But there could be circumstances where you cannot always do this. Merely pointing out the conventional way of setting up config in tf. So if your circumstances don't allow you to fix tf.config at the beginning you have to reset your tf.eager.context as shown in the solution above.

Upvotes: 4

Related Questions