Rename variable scope of saved model in TensorFlow

Question

Is it possible to rename the variable scope of a given model in tensorflow?

For instance, I created a logistic regression model for MNIST digits, based on the tutorial:

with tf.variable_scope('my-first-scope'):
    NUM_IMAGE_PIXELS = 784
    NUM_CLASS_BINS = 10
    x = tf.placeholder(tf.float32, shape=[None, NUM_IMAGE_PIXELS])
    y_ = tf.placeholder(tf.float32, shape=[None, NUM_CLASS_BINS])

    W = tf.Variable(tf.zeros([NUM_IMAGE_PIXELS,NUM_CLASS_BINS]))
    b = tf.Variable(tf.zeros([NUM_CLASS_BINS]))

    y = tf.nn.softmax(tf.matmul(x,W) + b)
    cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y), reduction_indices=[1]))
    saver = tf.train.Saver([W, b])

... # some training happens

saver.save(sess, 'my-model')

Now I want to reload the saved model in the 'my-first-scope' variable scope and then save everything again to a new file and under a new variable scope of 'my-second-scope'.

Answer 1

Based on keveman's answer, I created a python script, which you can execute to rename the variables of any TensorFlow checkpoint:

https://gist.github.com/batzner/7c24802dd9c5e15870b4b56e22135c96

You can replace substrings in the variable names and add a prefix to all names. Call the script with

python tensorflow_rename_variables.py --checkpoint_dir=path/to/dir

with the optional arguments

--replace_from=substr --replace_to=substr --add_prefix=abc --dry_run

Here is the script's core function:

def rename(checkpoint_dir, replace_from, replace_to, add_prefix, dry_run=False):
    checkpoint = tf.train.get_checkpoint_state(checkpoint_dir)
    with tf.Session() as sess:
        for var_name, _ in tf.contrib.framework.list_variables(checkpoint_dir):
            # Load the variable
            var = tf.contrib.framework.load_variable(checkpoint_dir, var_name)

            # Set the new name
            new_name = var_name
            if None not in [replace_from, replace_to]:
                new_name = new_name.replace(replace_from, replace_to)
            if add_prefix:
                new_name = add_prefix + new_name

            if dry_run:
                print('%s would be renamed to %s.' % (var_name, new_name))
            else:
                print('Renaming %s to %s.' % (var_name, new_name))
                # Rename the variable
                var = tf.Variable(var, name=new_name)

        if not dry_run:
            # Save the variables
            saver = tf.train.Saver()
            sess.run(tf.global_variables_initializer())
            saver.save(sess, checkpoint.model_checkpoint_path)

Example:

python tensorflow_rename_variables.py --checkpoint_dir=path/to/dir --replace_from=scope1 --replace_to=scope1/model --add_prefix=abc/

will rename the variable scope1/Variable1 to abc/scope1/model/Variable1.

Answer 2

You can use tf.contrib.framework.list_variables and tf.contrib.framework.load_variable as follows to achieve your goal :

with tf.Graph().as_default(), tf.Session().as_default() as sess:
  with tf.variable_scope('my-first-scope'):
    NUM_IMAGE_PIXELS = 784
    NUM_CLASS_BINS = 10
    x = tf.placeholder(tf.float32, shape=[None, NUM_IMAGE_PIXELS])
    y_ = tf.placeholder(tf.float32, shape=[None, NUM_CLASS_BINS])

    W = tf.Variable(tf.zeros([NUM_IMAGE_PIXELS,NUM_CLASS_BINS]))
    b = tf.Variable(tf.zeros([NUM_CLASS_BINS]))

    y = tf.nn.softmax(tf.matmul(x,W) + b)
    cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y), reduction_indices=[1]))
    saver = tf.train.Saver([W, b])
  sess.run(tf.global_variables_initializer())
  saver.save(sess, 'my-model')

vars = tf.contrib.framework.list_variables('.')
with tf.Graph().as_default(), tf.Session().as_default() as sess:

  new_vars = []
  for name, shape in vars:
    v = tf.contrib.framework.load_variable('.', name)
    new_vars.append(tf.Variable(v, name=name.replace('my-first-scope', 'my-second-scope')))

  saver = tf.train.Saver(new_vars)
  sess.run(tf.global_variables_initializer())
  saver.save(sess, 'my-new-model')