This article is based on the version of tf1.13 tf.variable_scope Class sum tf.name_scope For example, the initialization of corresponding classes is introduced.
- tf.variable_scope Class initialization
def __init__(self,
name_or_scope, #Name or VariableScope, used as the name prefix of op functions in the current scope
default_name=None, #The default name of name_or_scope is globally unique when it is not specified; the variable is invalid when name_or_scope is specified, and [1,2,...] is automatically added after the default name when it is accessed many times.
values=None, #Input parameter list (op input)
initializer=None, #Default initial values for all variables in the current scope (space)
regularizer=None, #Default regularization values for all variables in the current scope
caching_device=None, #Default buffer devices for all variables in the current scope
partitioner=None, #Partitioning of all variables in the current scope
custom_getter=None,
reuse=None, #Whether the variable reuses the flag bit, `True', None, or tf.AUTO_REUSE three values, True represents the value of reuse flag bit inherited from the parent space when the tf.AUTO_REUSE variable does not exist and when it exists, it is rebuilt.
dtype=None, #Variable data type
use_resource=None, #Markers, False for regularized variables, True for customized variables
constraint=None,
auxiliary_name_scope=True): #Flag bit, False means not using aliases, True means creating aliases or spaces.- tf.name_scope Class initialization
def __init__(self,
name, #Class space name, used as the name prefix of op functions in the current scope
default_name=None, #Default name when name=None
values=None): #Input parameter list (op input)These two class functions are often used to restrict tf.Variable and tf.get_variable Class scope of use. The base class of the former is Variable, and the base class of the latter is VariableScope.
| tf.Variable | tf.get_variable | |
| name | Influenced by tf.name_variable and tf.variable_scope | Influenced only by tf.variable_scope |
| Repeated creation |
First created with the name; The number of repetitions ${name} ${number of repetitions, incremental from 1} is added after the name when the repetition is created. |
When the variable name is first created, it is created directly when it does not exist. If reuse=True is set, an error will be reported, and tf.AUTO_REUSE can be used. When creating repeatedly, you need to specify the variable reuse flag as True or tf.AUTO_REUSE |
import tensorflow as tf
in_1=1
in_2=2
name=None
name='foo'
v1 = tf.get_variable('foo/var1', [1]) #If the comment is changed, reuse=tf.AUTO_REUSE of tf.variable_scope will report an error if it is not set or set to True.
for idx in range(2):
with tf.variable_scope(name, 'V0', [in_1, in_2]) as scope:
with tf.name_scope("bar"):
scope.reuse_variables() #Valid only if the variable name already exists, that is, with v1 = tf.get_variable('foo/var1', [1])
#tf.get_variable_scope().reuse_variables() #Valid only if the variable name exists
v1 = tf.get_variable("var1", [1])
v2 = tf.Variable(initial_value=[2.], name='var2', dtype=tf.float32)
v3 = tf.Variable(initial_value=[2.], name='var2', dtype=tf.float32)
with tf.Session() as sess:
print('##ID: {}##'.format(idx))
print('v1 name: {}'.format(v1.name))
print('v2 name: {}'.format(v2.name))
print('v3 name: {}'.format(v3.name))
//The results are as follows:
##ID: 0##
v1 name: foo/var1:0 #The variable name of tf.get_variable only depends on tf.variable_scope as prefix
v2 name: foo/bar/var2:0 #The variable name of tf.Variable depends on tf.variable_scope and tf.name_scope.
v3 name: foo/bar/var2_1:0 #When the current variable space is renamed, the name changes automatically to maintain global uniqueness
##ID: 1##
v1 name: foo/var1:0 #Set scope.reuse_variables(), so get it directly from the existing variables
v2 name: foo_1/bar/var2:0 #When the variable space of tf.Variable is renamed, the name is automatically changed to the name with suffix underlining and the number of repetitions.
v3 name: foo_1/bar/var2_1:0
Reference: