类实例之间的 Tensorflow 变量共享

Duck 发表于 Dev

鸭

class E():
    def __init__(self):
        self.sess = tf.Session()
        xav_init = tf.contrib.layers.xavier_initializer
        self.b_Wi = tf.get_variable(name='b_Wi', shape=[2,3], dtype=tf.float32,  initializer=xav_init())
e = E()
e1 = E()

执行上述代码时出现以下错误。

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

我知道他们是解决方法，但我更想了解上面背后的逻辑。实例不应该有自己单独的变量。为什么它们与上面的 b/we 和 e1 共享？

张乐纳

from itertools import count
import tensorflow as tf

class E():
    _ids = count(0)
    def __init__(self):
        self.id = next(self._ids)
        self.sess = tf.Session()
        xav_init = tf.contrib.layers.xavier_initializer
        with tf.variable_scope("share") as sp:
            print(self.id)
            if self.id > 0:
                tf.get_variable_scope().reuse_variables()
            self.b_Wi = tf.get_variable(name='b_Wi', shape=[2,3], dtype=tf.float32,  initializer=xav_init())

e1 = E()
e2 = E()

assert(e1.b_Wi == e2.b_Wi) # thus they are exactly the same object in the same graph and hence affect each other.

本文收集自互联网，转载请注明来源。

如有侵权，请联系 [email protected] 删除。