TensorFlow exponential moving average

Question

I can't figure out how to get tf.train.ExponentialMovingAverage to work. Following is a simple code to find w in a simple y_ = x * w equation. m is the moving average. Why is the code returning None for m? How can I get it to return the moving average value?

import numpy as np
import tensorflow as tf

w = tf.Variable(0, dtype=tf.float32)
ema = tf.train.ExponentialMovingAverage(decay=0.9)
m = ema.apply([w])

x = tf.placeholder(tf.float32, [None])
y = tf.placeholder(tf.float32, [None])
y_ = tf.multiply(x, w)

with tf.control_dependencies([m]):
    loss = tf.reduce_sum(tf.square(tf.subtract(y, y_)))
    optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
    train = optimizer.minimize(loss)

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(20):
        _, w_, m_ = sess.run([train, w, m], feed_dict={x: [1], y: [10]})
        print(w_, ',', m_)

The output is:

0.02 , None
0.03996 , None
0.0598801 , None
0.0797603 , None
0.0996008 , None
0.119402 , None
0.139163 , None
0.158884 , None
0.178567 , None
0.19821 , None
0.217813 , None
0.237378 , None
0.256903 , None
0.276389 , None
0.295836 , None
0.315244 , None
0.334614 , None
0.353945 , None
0.373237 , None
0.39249 , None

Answer 1

This is because the m (python) variable doesn't hold the result of the operation but the operation itself. See the doc:

Returns:
  An Operation that updates the moving averages.

To get access to the average value, you need to create a new element in your graph :

av = ema.average(w)

and then:

_, w_, av_ = sess.run([train, w, av], feed_dict={x: [1], y: [10]})
print(w_, ',', av_)

will print

[0.020000001, 0.0]
[0.039960001, 0.0020000006]
[0.059880082, 0.0057960013]
[0.07976032, 0.01120441]
[0.099600799, 0.018060002]

---

Full code

import tensorflow as tf

w = tf.Variable(0, dtype=tf.float32)
ema = tf.train.ExponentialMovingAverage(decay=0.9)
m = ema.apply([w])
av = ema.average(w)

x = tf.placeholder(tf.float32, [None])
y = tf.placeholder(tf.float32, [None])
y_ = tf.multiply(x, w)

with tf.control_dependencies([m]):
    loss = tf.reduce_sum(tf.square(tf.subtract(y, y_)))
    optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001)
    train = optimizer.minimize(loss)

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(20):
        _, w_, av_ = sess.run([train, w, av], feed_dict={x: [1], y: [10]})
        print(w_, ',', av_)