Tensorflow name_scope

时间:2021-10-17 08:47:56

在 Tensorflow 当中有两种途径生成变量 variable, 一种是 tf.get_variable(), 另一种是 tf.Variable().

使用tf.get_variable()定义的变量不会被tf.name_scope()当中的名字所影响

 import tensorflow as tf

 with tf.name_scope("a_name_scope"):
initializer = tf.constant_initializer(value=1)
var1 = tf.get_variable(name='var1', shape=[1], dtype=tf.float32, initializer=initializer)
var2 = tf.Variable(name='var2', initial_value=[2], dtype=tf.float32)
var21 = tf.Variable(name='var2', initial_value=[2.1], dtype=tf.float32)
var22 = tf.Variable(name='var2', initial_value=[2.2], dtype=tf.float32) with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
print(var1.name) # var1:0
print(sess.run(var1)) # [ 1.]
print(var2.name) # a_name_scope/var2:0
print(sess.run(var2)) # [ 2.]
print(var21.name) # a_name_scope/var2_1:0
print(sess.run(var21)) # [ 2.0999999]
print(var22.name) # a_name_scope/var2_2:0
print(sess.run(var22)) # [ 2.20000005]

想要达到重复利用变量的效果, 我们就要使用 tf.variable_scope(), 并搭配 tf.get_variable()这种方式产生和提取变量. 不像 tf.Variable() 每次都会产生新的变量, tf.get_variable() 如果遇到了同样名字的变量时, 它会单纯的提取这个同样名字的变量(避免产生新变量). 而在重复使用的时候, 一定要在代码中强调 scope.reuse_variables(), 否则系统将会报错, 以为你只是单纯的不小心重复使用到了一个变量.

 with tf.variable_scope("a_variable_scope") as scope:
initializer = tf.constant_initializer(value=3)
var3 = tf.get_variable(name='var3', shape=[1], dtype=tf.float32, initializer=initializer)
scope.reuse_variables()
var3_reuse = tf.get_variable(name='var3',)
var4 = tf.Variable(name='var4', initial_value=[4], dtype=tf.float32)
var4_reuse = tf.Variable(name='var4', initial_value=[4], dtype=tf.float32) with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print(var3.name) # a_variable_scope/var3:0
print(sess.run(var3)) # [ 3.]
print(var3_reuse.name) # a_variable_scope/var3:0
print(sess.run(var3_reuse)) # [ 3.]
print(var4.name) # a_variable_scope/var4:0
print(sess.run(var4)) # [ 4.]
print(var4_reuse.name) # a_variable_scope/var4_1:0
print(sess.run(var4_reuse)) # [ 4.]

 with tf.variable_scope('foo') as foo_scope:
v = tf.get_variable('v', [1])
with tf.variable_scope('foo', reuse=True):
v1 = tf.get_variable('v')
assert v1 == v

1. 使用tf.Variable()的时候,tf.name_scope()tf.variable_scope() 都会给 Variable 和 op 的 name属性加上前缀。 
2. 使用tf.get_variable()的时候,tf.name_scope()就不会给 tf.get_variable()创建出来的Variable加前缀。