tensorflow - session 恢复后 get_variable() 不起作用

Question

我尝试恢复 session 并调用 get_variable()获取类型的对象
tf.Variable(根据 this answer )。
它无法找到变量。重现该案例的最小示例是
如下。

首先，创建一个变量并保存 session 。

import tensorflow as tf

var = tf.Variable(101)

with tf.Session() as sess:
    with tf.variable_scope(''):
        scoped_var = tf.get_variable('scoped_var', [])

    with tf.variable_scope('', reuse=True):
        new_scoped_var = tf.get_variable('scoped_var', [])

    assert scoped_var is new_scoped_var
    sess.run(tf.global_variables_initializer())
    saver = tf.train.Saver()
    print(sess.run(scoped_var))
    saver.save(sess, 'data/sess')

这里 get_variables在具有 reuse=True 的范围内工作正常。
然后，从文件中恢复 session 并尝试获取变量。

import tensorflow as tf

with tf.Session() as sess:
    saver = tf.train.import_meta_graph('data/sess.meta')
    saver.restore(sess, 'data/sess')

    for v in tf.get_collection('variables'):
        print(v.name)

    print(tf.get_collection(("__variable_store",)))
    # Oops, it's empty!

    with tf.variable_scope('', reuse=True):
        # the next line fails
        new_scoped_var = tf.get_variable('scoped_var', [])

    print("new_scoped_var: ", new_scoped_var)

输出:

Variable:0
scoped_var:0
[]
Traceback (most recent call last):
...
ValueError: Variable scoped_var does not exist, or was not created with tf.get_variable(). Did you mean to set reuse=None in VarScope?

正如我们所见， get_variable()找不到变量。和 ("__variable_store",)集合，由 get_variable() 内部使用,
是空的。

为什么 get_variable失败？

Answer 1

您可以试试这个，而不是处理元图(如果您想修改图以及它的加载方式等，这可能会很有帮助)。

import tensorflow as tf

with tf.Session() as sess:
  with tf.variable_scope(''):
    scoped_var = tf.get_variable('scoped_var', [])

  with tf.variable_scope('', reuse=True):
    new_scoped_var = tf.get_variable('scoped_var', [])

  assert scoped_var is new_scoped_var
  saver = tf.train.Saver()
  path = tf.train.get_checkpoint_state('data/sess')
  if path is not None:
    saver.restore(sess, path.model_checkpoint_path)
  else:
    sess.run(tf.global_variables_initializer())

  print(sess.run(scoped_var))
  saver.save(sess, 'data/sess')

  #now continue to use as you normally would with a restored model

主要区别是您在调用 saver.restore 之前设置了模型