[PYTHON] Save a specific variable in tensorflow.session

Use tensorflow.train.Saver to save Tensorflow Variables to a file. All Variables in Session are saved by the method described in Tutorial. To save / restore only a specific Variable, give the initialization function of tensorflow.train.Saver a list of Variables you want to target in a dictionary type.

This makes it possible to read Variables individually from multiple files.

save.py

import tensorflow as tf

def get_particular_variables(name):
    return {v.name: v for v in tf.all_variables() if v.name.find(name) >= 0}

def define_variables(var0_value, var1_value, var2_value):
    var0 = tf.Variable([var0_value])
    with tf.variable_scope('foo'):
        var1 = tf.Variable([var1_value])
    with tf.variable_scope('bar'):
        var2 = tf.Variable([var2_value])

    return var0, var1, var2


sess = tf.InteractiveSession()

# defines variables
var0, var1, var2 = define_variables(0.0, 0.0, 0.0)

# saving only variables whose name includes foo
saver = tf.train.Saver(get_particular_variables('foo'))

# initializing all of variables
sess.run(tf.initialize_all_variables())

print var0.eval(), var1.eval(), var2.eval()

# saving into file
saver.save(sess, './bar_val')

restore.py

import tensorflow as tf

def get_particular_variables(name):
    return {v.name: v for v in tf.all_variables() if v.name.find(name) >= 0}

def define_variables(var0_value, var1_value, var2_value):
    var0 = tf.Variable([var0_value])
    with tf.variable_scope('foo'):
        var1 = tf.Variable([var1_value])
    with tf.variable_scope('bar'):
        var2 = tf.Variable([var2_value])

    return var0, var1, var2

sess = tf.InteractiveSession()

# defines variables
var0, var1, var2 = define_variables(1.0, 1.0, 1.0)

# restoring only variables whole name includes foo
saver = tf.train.Saver(get_particular_variables('foo'))

# initializing all of variables
sess.run(tf.initialize_all_variables())
print 'before restoring: ', var0.eval(), var1.eval(), var2.eval()

# restoring variable from file
saver.restore(sess, './bar_val')
print 'after restoring only var in foo: ', var0.eval(), var1.eval(), var2.eval()

However, this method requires attention to long names and namespace hierarchies. For example

In such a case, executing get_particular_variables ('foo') above will return var1, var2, and var3. In this way, depending on the search conditions, extra variables are saved, which may cause unexpected bugs when restoring.