tensorflow中命名空间、变量命名的问题

1.简介

对比分析tf.Variable / tf.get_variable | tf.name_scope / tf.variable_scope的异同

2.说明

• tf.Variable创建变量；tf.get_variable创建与获取变量
• tf.Variable自动检测命名冲突并且处理；tf.get_variable在没有设置reuse时会报错
• tf.name_scope没有reuse功能，tf.get_variable在变量冲突时报错；tf.variable_scope有reuse功能，可配合tf.get_variable实现变量共享
• tf.get_variable变量命名不受tf.name_scope的影响；tf.Variable受两者的影响

3.代码示例

3.1 tf.Variable

tf.Variable在命名冲突时自动处理冲突问题

 import tensorflow as tf
 a1 = tf.Variable(tf.constant(1.0, shape=[1]),name="a")
 a2 = tf.Variable(tf.constant(1.0, shape=[1]),name="a")
 print(a1)
 print(a2)
 print(a1==a2)

 ###
10 <tf.Variable 'a:0' shape=(1,) dtype=float32_ref>
11 <tf.Variable 'a_1:0' shape=(1,) dtype=float32_ref>
12 False

3.2 tf.get_variable

tf.get_variable在没有设置命名空间reuse的情况下变量命名冲突时报错

 import tensorflow as tf
 a3 = tf.get_variable("a", shape=[1], initializer=tf.constant_initializer(1.0))
 a4 = tf.get_variable("a", shape=[1], initializer=tf.constant_initializer(1.0))

 ###
7 ValueError: Variable a already exists, disallowed.
8 Did you mean to set reuse=True or reuse=tf.AUTO_REUSE in VarScope?

3.3 tf.name_scope

tf.name_scope没有reuse功能，tf.get_variable命名不受它影响，并且命名冲突时报错；tf.Variable命名受它影响

 import tensorflow as tf
 a = tf.Variable(tf.constant(1.0, shape=[1]),name="a")
 with tf.name_scope('layer2'):
     a1 = tf.Variable(tf.constant(1.0, shape=[1]),name="a")
     a2 = tf.Variable(tf.constant(1.0, shape=[1]),name="a")
     a3 = tf.get_variable("a", shape=[1], initializer=tf.constant_initializer(1.0))
 7 #     a4 = tf.get_variable("a", shape=[1], initializer=tf.constant_initializer(1.0))   该句会报错
     print(a)
     print(a1)
     print(a2)
     print(a3)
     print(a1==a2)

 ###
16 <tf.Variable 'a:0' shape=(1,) dtype=float32_ref>
17 <tf.Variable 'layer2/a:0' shape=(1,) dtype=float32_ref>
18 <tf.Variable 'layer2/a_1:0' shape=(1,) dtype=float32_ref>
19 <tf.Variable 'a_1:0' shape=(1,) dtype=float32_ref>
20 False

3.4 tf.variable_scope

tf.variable_scope可以配tf.get_variable实现变量共享；reuse默认为None，有False/True/tf.AUTO_REUSE可选：

• 设置reuse = None/False时tf.get_variable创建新变量，变量存在则报错
• 设置reuse = True时tf.get_variable只讲获取已存在的变量，变量不存在时报错
• 设置reuse = tf.AUTO_REUSE时tf.get_variable在变量已存在则自动复用，不存在则创建

 import tensorflow as tf
 with tf.variable_scope('layer1',reuse=tf.AUTO_REUSE):
     a1 = tf.Variable(tf.constant(1.0, shape=[1]),name="a")
     a2 = tf.Variable(tf.constant(1.0, shape=[1]),name="a")
     a3 = tf.get_variable("a", shape=[1], initializer=tf.constant_initializer(1.0))
     a4 = tf.get_variable("a", shape=[1], initializer=tf.constant_initializer(1.0))
     print(a1)
     print(a2)
     print(a1==a2)
     print(a3)
     print(a4)
     print(a3==a4)

 ###
16 <tf.Variable 'layer1_1/a:0' shape=(1,) dtype=float32_ref>
17 <tf.Variable 'layer1_1/a_1:0' shape=(1,) dtype=float32_ref>
18 False
19 <tf.Variable 'layer1/a_2:0' shape=(1,) dtype=float32_ref>
20 <tf.Variable 'layer1/a_2:0' shape=(1,) dtype=float32_ref>
21 True

!!!