What is a TensorFlow Session?

I’ve seen a lot of confusion over the rules of tf.Graph and tf.Session in TensorFlow. It’s simple:

Let’s look at an example.

Update 2020-03-04: Sessions are gone in TensorFlow 2. There is one global runtime in the background that executes all computation, whether run eagerly or as a compiled tf.function.

Defining the Graph

We define a graph with a variable and three operations: variable returns the current value of our variable. initialize assigns the initial value of 42 to that variable. assign assigns the new value of 13 to that variable.

graph = tf.Graph()
with graph.as_default():
  variable = tf.Variable(42, name='foo')
  initialize = tf.global_variables_initializer()
  assign = variable.assign(13)

On a side note: TensorFlow creates a default graph for you, so we don’t need the first two lines of the code above. The default graph is also what the sessions in the next section use when not manually specifying a graph.

Running Computations in a Session

To run any of the three defined operations, we need to create a session for that graph. The session will also allocate memory to store the current value of the variable.

with tf.Session(graph=graph) as sess:
# Output: 13

As you can see, the value of our variable is only valid within one session. If we try to query the value afterwards in a second session, TensorFlow will raise an error because the variable is not initialized there.

with tf.Session(graph=graph) as sess:
# Error: Attempting to use uninitialized value foo

Of course, we can use the graph in more than one session, we just have to initialize the variables again. The values in the new session will be completely independent from the first one:

with tf.Session(graph=graph) as sess:
# Output: 42

Hopefully this short workthrough helped you to better understand tf.Session. Feel free to ask questions in the comments.

Updated 2017-07-12: The operation to initialize variables changed in TensorFlow 1.0.

You can use this post under the open CC BY-SA 3.0 license and cite it as:

  author = {Hafner, Danijar},
  title = {What is a TensorFlow Session?},
  year = {2016},
  howpublished = {Blog post},
  url = {https://danijar.com/what-is-a-tensorflow-session/}