为什么类对象中的 tf.Variable 初始化失败？

Question

The way tf.global_variables_initializer() works makes me confused.

This code ends up with uninitialized error:

import tensorflow as tf

class C(object):
    def __init__(self):
        self.a = tf.Variable(tf.ones(()),tf.float32)

op_init = tf.global_variables_initializer()                                                                                                                                       12
with tf.Session() as sess:

    c = C()

    sess.run(op_init)
    print(sess.run(c.a))
    # -> fail. FailedPreconditionError: Attempting to use uninitialized value Variable

But on the other hands, the bellow works.

import tensorflow as tf
class C(object):

    def __init__(self):
        self.a = tf.Variable(tf.ones(()),tf.float32)
        self.op_init = tf.global_variables_initializer()

with tf.Session() as sess:

    c = C()

    sess.run(c.op_init)
    print(sess.run(c.a))
    # -> success. '1.0'

Answer 1

The other answer is incorrect with regards to the "scope" of the initializer. All variables created anywhere are added to the collection of global variables which is a Tensorflow concept and has nothing to do with Python variable scope. Unfortunately I don't have TF 1.X installed anymore so I can't really check, but I suspect that the problem in the first case is that you are creating the initializer before creating the variable. Try reordering like this:

import tensorflow as tf

class C(object):
    def __init__(self):
        self.a = tf.Variable(tf.ones(()),tf.float32)

c = C()
op_init = tf.global_variables_initializer()                                                                                                                                       12
with tf.Session() as sess:
    sess.run(op_init)
    print(sess.run(c.a))

All I did was move the creation of c before the initializer. Please let us know if this works!

Answer 2

The Tensor is not initialized because it is not in the scope of the initializer function. The second way works because it is in the scope. tf.global_variables_initializer() is a convenience for the user but it is not the one and only way you should initialize your Variables.

The correct way in the first option would be to pass an initial value as a fed input:

print(sess.run(c.a, feed_dict={c.a:1}))

This is generally how you want to use a tf.Variable as its contents can vary, meaning that it is probably initialized by the user during every run.