Get all tf.Variables related to a specific tensor in TensorFlow

Question

I want to create a utility function where I give it a tensor t and a variable name n , and the function returns (it it exists) the variable whos name contains n and is part of the graph of t .

def get_variable(t, n):
   #code
   return variable

The reason I want to do this is that using Jupyter I often create a graph multiple times while deciding the best structure, but the tensor names keep changing to something like: {name}_{repetition}:0 , so finding them through tf.all_variables() becomes increasingly harder.

It would be easier if I could limit the search to the variables related to a specific tensor because its identifier is of the latest repetition.

Answer 1

(Warning: I am not going to answer the main question about creating an utility function)

The reason the tensors have names like: {name}_{repetition}:0 is not because you add again and again these tensors to the default graph .

A solution is to always specify the graph in which you are working with graph.as_default() :

graph = tf.Graph()
with graph.as_default():
  with tf.variable_scope('foo'):
    var = tf.get_variable('var', [])
    print var.name  # should be 'foo/var:0'
    tensor = tf.constant(2., name='tensor')
    print tensor.name  # should be 'foo/tensor:0'

---

If you run it again, you will see the exact same results because the graph = tf.Graph() line will create a new graph.

graph = tf.Graph()
with graph.as_default():
  with tf.variable_scope('foo'):
    var = tf.get_variable('var', [])
    print var.name  # should be 'foo/var:0'
    tensor = tf.constant(2., name='tensor')
    print tensor.name  # should be 'foo/tensor:0'

---

The drawback is that now you cannot rely on the default graph and should pass the graph to compute for example the list of all variables:

print tf.all_variables()  # returns []

with graph.as_default():
  print tf.all_variables()  # returns [...] with all variables