How to initialize a tf.Variable with a tf.constant or a numpy array?

Question

I am trying to initialize a tf.Variable() in a tf.InteractiveSession(). I already have some pre-trained weights which are individual numpy files. How do I effectively initialize the variable with these numpy values ?

I have gone through the following options:

Using tf.assign()
using sess.run() directly during tf.Variable() creation

Seems like the values are not correctly initialized. Following is some code I have tried. Let me know which is the correct one ?

def read_numpy(file):
    return np.fromfile(file,dtype='f')

def build_network():
    with tf.get_default_graph().as_default():
        x = tf.Variable(tf.constant(read_numpy('foo.npy')),name='var1')
        sess = tf.get_default_session()
        with sess.as_default():
            sess.run(tf.global_variables_initializer())

sess = tf.InteractiveSession()
with sess.as_default():
    build_network()

Is this the correct way to do it ? I have printed the session object, and it is the same session used throughout.

edit : Currently it seems like using sess.run(tf.global_variables_initializer()) is calling a random initialize op

Vlad · Accepted Answer · 2019-04-29 19:49:54Z

3

tf.Variable() accepts numpy arrays as initial values:

import tensorflow as tf
import numpy as np

init = np.ones((2, 2))
x = tf.Variable(init) # <-- set initial value to assign to a variable

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer()) # <-- this will assign the init value
    print(x.eval())
# [[1. 1.]
#  [1. 1.]]

So just use the numpy array to initialize, no need to convert it to a tensor first.

Alternatively, you could also use tf.Variable.load() to assign values from numpy array to a variable within a session context:

import tensorflow as tf
import numpy as np

x = tf.Variable(tf.zeros((2, 2)))

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    init = np.ones((2, 2))
    x.load(init)
    print(x.eval())
# [[1. 1.]
#  [1. 1.]]

edited Apr 29, 2019 at 19:49

answered Apr 29, 2019 at 19:44

Vlad

8,6255 gold badges38 silver badges47 bronze badges

Sign up to request clarification or add additional context in comments.

4 Comments

lamo_738 Over a year ago

is it fine to use the tf.get_variable() ? .. I used that and it worked. I had an attribute as initializer=tf.constant(read_numpy('foo.npy'))

Vlad Over a year ago

Yes, you can. The difference between tf.Variable() and tf.get_variable() is that the tf.Variable() creates a new variable, whereas get_variable() may create a new one or may return an existing variable. For more read this answer

lamo_738 Over a year ago

Oh okay, I did get an error on this first time when I used tf.get_variable() since naming was clashing.. but then I included it within a tf.variable_scope() and that was handled. thanks for the response !

Vlad Over a year ago

Glad it helped. I would still use x = tf.Variable(read_numpy('foo.npy')). This way it is shorter.

Collectives™ on Stack Overflow

How to initialize a tf.Variable with a tf.constant or a numpy array?

1 Answer 1

4 Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

4 Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related