tf.train.init_from_checkpoint does not initialize variables created with tf.Variable

It seems that tf.train.init_from_checkpoint initalizes variables created via tf.get_variable but not those created via tf.Variable.

For example, let's create two variables and save them:

import tensorflow as tf

tf.Variable(1.0, name='foo')
tf.get_variable('bar',initializer=1.0)
saver = tf.train.Saver()
with tf.Session() as sess:
  tf.global_variables_initializer().run()
  saver.save(sess, './model', global_step=0)

If I load them again via a tf.train.Saver, everything works fine: variables are loaded back to 1 even though they are initialized at zero here:

import tensorflow as tf

foo = tf.Variable(0.0, name='foo')
bar = tf.get_variable('bar', initializer=0.0)
saver = tf.train.Saver()
with tf.Session() as sess:
  saver.restore(sess, './model-0')
  print(f'foo: {foo.eval()}  bar: {bar.eval()}')
  # foo: 1.0  bar: 1.0

However if I use tf.train.init_from_checkpoint I get

import tensorflow as tf

foo = tf.Variable(0.0, name='foo')
bar = tf.get_variable('bar', initializer=0.0)
tf.train.init_from_checkpoint('./model-0', {'/':'/'})
with tf.Session() as sess:
  tf.global_variables_initializer().run()
  print(f'foo: {foo.eval()}  bar: {bar.eval()}')
  # foo: 0.0  bar: 1.0

bar is set back to 1 as expected but foo remains at 0.

Is this the intended behavior? If so, why?

from tf.train.init_from_checkpoint does not initialize variables created with tf.Variable

tf.train.init_from_checkpoint does not initialize variables created with tf.Variable 2019-03-18T08:48:00-07:00 Rating: 4.5 Diposkan Oleh: Admin