TensorFlow中Variable和get_variable之间的区别

As far as I know, Variable is the default operation for making a variable, and get_variable is mainly used for weight sharing.

On the one hand, there are some people suggesting using get_variable instead of the primitive Variable operation whenever you need a variable. On the other hand, I merely see any use of get_variable in TensorFlow’s official documents and demos.

Thus I want to know some rules of thumb on how to correctly use these two mechanisms. Are there any “standard” principles?

I’d recommend to always use tf.get_variable(...) — it will make it way easier to refactor your code if you need to share variables at any time, e.g. in a multi-gpu setting (see the multi-gpu CIFAR example). There is no downside to it.

Pure tf.Variable is lower-level; at some point tf.get_variable() did not exist so some code still uses the low-level way.

tf.Variable is a class, and there are several ways to create tf.Variable including tf.Variable.__init__ and tf.get_variable.

tf.Variable.__init__: Creates a new variable with initial_value.

W = tf.Variable(<initial-value>, name=<optional-name>)

tf.get_variable: Gets an existing variable with these parameters or creates a new one. You can also use initializer.

W = tf.get_variable(name, shape=None, dtype=tf.float32, initializer=None,
       regularizer=None, trainable=True, collections=None)

It’s very useful to use initializers such as xavier_initializer:

W = tf.get_variable("W", shape=[784, 256],
       initializer=tf.contrib.layers.xavier_initializer())

More information here.

I can find two main differences between one and the other:

1. First is that tf.Variable will always create a new variable, whereas tf.get_variable gets an existing variable with specified parameters from the graph, and if it doesn’t exist, creates a new one.

2. tf.Variable requires that an initial value be specified.

It is important to clarify that the function tf.get_variable prefixes the name with the current variable scope to perform reuse checks. For example:

with tf.variable_scope("one"):
    a = tf.get_variable("v", [1]) #a.name == "one/v:0"
with tf.variable_scope("one"):
    b = tf.get_variable("v", [1]) #ValueError: Variable one/v already exists
with tf.variable_scope("one", reuse = True):
    c = tf.get_variable("v", [1]) #c.name == "one/v:0"

with tf.variable_scope("two"):
    d = tf.get_variable("v", [1]) #d.name == "two/v:0"
    e = tf.Variable(1, name = "v", expected_shape = [1]) #e.name == "two/v_1:0"

assert(a is c)  #Assertion is true, they refer to the same object.
assert(a is d)  #AssertionError: they are different objects
assert(d is e)  #AssertionError: they are different objects

The last assertion error is interesting: Two variables with the same name under the same scope are supposed to be the same variable. But if you test the names of variables d and e you will realize that Tensorflow changed the name of variable e:

d.name   #d.name == "two/v:0"
e.name   #e.name == "two/v_1:0"

Another difference lies in that one is in ('variable_store',) collection but the other is not.

Please see the source code:

def _get_default_variable_store():
  store = ops.get_collection(_VARSTORE_KEY)
  if store:
    return store[0]
  store = _VariableStore()
  ops.add_to_collection(_VARSTORE_KEY, store)
  return store

Let me illustrate that:

import tensorflow as tf
from tensorflow.python.framework import ops

embedding_1 = tf.Variable(tf.constant(1.0, shape=[30522, 1024]), name="word_embeddings_1", dtype=tf.float32) 
embedding_2 = tf.get_variable("word_embeddings_2", shape=[30522, 1024])

graph = tf.get_default_graph()
collections = graph.collections

for c in collections:
    stores = ops.get_collection(c)
    print('collection %s: ' % str(c))
    for k, store in enumerate(stores):
        try:
            print('\t%d: %s' % (k, str(store._vars)))
        except:
            print('\t%d: %s' % (k, str(store)))
    print('')

The output:

collection ('__variable_store',): 0: {'word_embeddings_2': <tf.Variable 'word_embeddings_2:0' shape=(30522, 1024) dtype=float32_ref>}