One way of restoring weights when computation graph structure has changed

restore_using_variable_scope.py

# example code for EC500 K1 / CS591 S2 Deep Learning (Spring 2017)

import tensorflow as tf
import numpy as np


def main_save():
    with tf.Graph().as_default() as g:
        with tf.variable_scope('to_save'):
            a = tf.get_variable('a_name', [100, 100])
            a_dense_original = tf.layers.dense(a, 10)
            a_another_dense_original = tf.layers.dense(a, 20)
        b = tf.get_variable('b', [200, 200])
        to_save_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'to_save')
        for v in to_save_vars:
            print(v.name, tf.shape(v).value)
        saver = tf.train.Saver(to_save_vars)
    
        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())
            saver.save(sess, './saved.ckp')
            print(np.sum(sess.run(a)))
            

def main_load():
    with tf.Graph().as_default() as g:
        with tf.variable_scope('to_save'):
            a = tf.get_variable('a_name', [100, 100])
            a_dense_restored = tf.layers.dense(a, 10)
            a_another_dense_restored = tf.layers.dense(a, 20)
        c = tf.get_variable('b', [300, 300])
        to_save_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'to_save')
        for v in to_save_vars:
            print(v.name, tf.shape(v))
        saver = tf.train.Saver(to_save_vars)
    
        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())
            saver.restore(sess, './saved.ckp')
            print(np.sum(sess.run(a)))


def main():
    main_save()
    main_load()
    
main()