tf.variable_scope()和tf.name_scope()

posted @ 2019-07-31 16:12 nxf_rabbit75 阅读(626) 评论(0) 编辑收藏举报

分类: TensorFlow

undefined

1.tf.variable_scope

功能：tf.variable_scope可以让不同命名空间中的变量取相同的名字，无论tf.get_variable或者tf.Variable生成的变量

TensorFlow链接：https://tensorflow.google.cn/api_docs/python/tf/variable_scope?hl=en

举例：

with tf.variable_scope('V1'):
    a1 = tf.get_variable(name='a1', shape=[1], initializer=tf.constant_initializer(1))
    a2 = tf.Variable(tf.random_normal(shape=[2, 3], mean=0, stddev=1), name='a2')
with tf.variable_scope('V2'):
    a3 = tf.get_variable(name='a1', shape=[1], initializer=tf.constant_initializer(1))
    a4 = tf.Variable(tf.random_normal(shape=[2, 3], mean=0, stddev=1), name='a2')
 
with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    print(a1.name)
    print(a2.name)
    print(a3.name)
    print(a4.name)

with tf.variable_scope("foo"):
    v = tf.get_variable("v", [1])
with tf.variable_scope("foo", reuse=True):
    v1 = tf.get_variable("v", [1])
assert v1 == v #不报错

如果想要重用变量，可以设置reuse_variables()

import numpy as np
with tf.variable_scope('V1'):
    a1 = tf.get_variable(name='a1', shape=[1], initializer=tf.constant_initializer(1))
    a2 = tf.Variable(tf.random_normal(shape=[2, 3], mean=0, stddev=1), name='a2')
    tf.get_variable_scope().reuse_variables()
    assert tf.get_variable_scope().reuse == True
    a3 = tf.get_variable(name='a1', shape=[1], initializer=tf.constant_initializer(1))
    a4 = tf.Variable(tf.random_normal(shape=[2, 3], mean=0, stddev=1), name='a2')
 
with tf.variable_scope('V1',reuse=True):
    a5 = tf.get_variable(name='a1', shape=[1], initializer=tf.constant_initializer(1))
 
with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    print(a1.name)
    print(a2.name)
    print(a3.name)
    print(a4.name)
    print(a5.name)

variable重名，虽然name设置的一样，但是实际是不共享同一个变量的；get_variable重name，其实是共享的同一个变量。

2.tf.name_scope

功能：tf.name_scope具有类似的功能，但只限于tf.Variable生成的变量

TensorFlow链接：https://tensorflow.google.cn/api_docs/python/tf/name_scope?hl=en

with tf.name_scope('V1'):
    a1 = tf.get_variable(name='a1', shape=[1], initializer=tf.constant_initializer(1))
    a2 = tf.Variable(tf.random_normal(shape=[2, 3], mean=0, stddev=1), name='a2')
with tf.name_scope('V2'):
    a3 = tf.get_variable(name='a1', shape=[1], initializer=tf.constant_initializer(1))
    a4 = tf.Variable(tf.random_normal(shape=[2, 3], mean=0, stddev=1), name='a2')
 
with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    print(a1.name)
    print(a2.name)
    print(a3.name)
    print(a4.name)

a1，a3会报错：ValueError: Variable a1 already exists, disallowed. Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope? Originally defined at: