为什么Tensorflow初始化一个类中失败的变量？

Question

The simplified code as bellow:

class ExModel(object):
    def __init__(self, graph=None):

        if graph is None:
            self.graph = tf.get_default_graph()
        else:
            self.graph = graph

        self.sess = tf.Session(graph=self.graph)

        self.init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer())
        self.build_model()

        print(self.sess.run(tf.report_uninitialized_variables()))  # [b'global_step']
        print(self.sess.run(self._global_step))  # uninitialized error

    def build_model(self):
        with self.graph.as_default():
            self._global_step = tf.Variable(0, trainable=False, name='global_step')

        self.sess.run(self.init_op)

if __name__ == '__main__':
    model = ExModel()

It shows use uninitialized value error.

It seems no difference with the code bellow, but it can works well.

if __name__ == '__main__':
    # model = ExModel()

    graph = tf.get_default_graph()

    sess = tf.Session(graph=graph)

    with graph.as_default():
        global_step = tf.Variable(0, trainable=False, name='global_step')

    sess.run(tf.global_variables_initializer())
    print(sess.run(global_step))

Please help me, the probelm have tormented me the whole day.

I have read another question . I think my error is samiliar with it, but I don't know how to make the global_variables_initializer be associated with the same graph as the session, and I just use the default graph.

I find when use self.sess.run(tf.global_variables_initializer()) to replace

self.init_op = tf.global_variables_initializer()
self.sess.run(self.init_op)

can work well. But I still don't know the reason.

Answer 1

The probelm is that init_op = tf.global_variables_initializer() should be define at the last of all the variables.

It should be

global_step = tf.Variable(0, trainable=False, name='global_step')
...
init_op = tf.global_variables_initializer()
sess.run(init_op)

but not

init_op = tf.global_variables_initializer()
global_step = tf.Variable(0, trainable=False, name='global_step')
...
sess.run(init_op)