不同scope的命名方式比较

代码

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]


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)
    var4 = tf.Variable(name='var4', initial_value=[4], dtype=tf.float32)
    var4_reuse = tf.Variable(name='var4', initial_value=[4], dtype=tf.float32)
    scope.reuse_variables()
    var3_reuse = tf.get_variable(name='var3',)

with tf.Session() as sess:
    # tf.initialize_all_variables() no long valid from
    # 2017-03-02 if using tensorflow >= 0.12
    if int((tf.__version__).split('.')[1]) < 12 and int((tf.__version__).split('.')[0]) < 1:
        init = tf.initialize_all_variables()
    else:
        init = tf.global_variables_initializer()
    sess.run(init)
    print(var3.name)            # a_variable_scope/var3:0
    print(sess.run(var3))       # [ 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.]
    print(var3_reuse.name)      # a_variable_scope/var3:0
    print(sess.run(var3_reuse)) # [ 3.]