Tensorflow: tf.random_normal get different results with the same initial seed

Question

I want to make a reusable RANDOM TENSOR x and assign the SAME tensor to VARIABLE y. That means they should have the same value during Session.run().

But it turns out not the case. So why does y NOT equal x?

Update: After applying sess.run(x) and sess.run(y) multiple times in line, confirmed that x changes every time while y stays steady. Why?

import tensorflow as tf

x = tf.random_normal([3], seed = 1)
y = tf.Variable(initial_value = x) # expect y get the same random tensor as x

diff = tf.subtract(x, y)
avg = tf.reduce_mean(diff)

sess = tf.InteractiveSession()
sess.run(y.initializer)

print('x0:', sess.run(x))
print('y0:', sess.run(y))
print('x1:', sess.run(x))
print('y1:', sess.run(y))
print('x2:', sess.run(x))
print('y2:', sess.run(y))
print('diff:', sess.run(diff))
print('avg:', sess.run(avg)) # expected as 0.0

sess.close()

Ouputs: TENSOR x changes every sess.run(x)

x0: [ 0.55171245 -0.13107552 -0.04481386]
y0: [-0.8113182   1.4845988   0.06532937]
x1: [-0.67590594  0.28665832  0.3215887 ]
y1: [-0.8113182   1.4845988   0.06532937]
x2: [1.2409041  0.44875884 0.33140722]
y2: [-0.8113182   1.4845988   0.06532937]
diff: [ 1.2404865  -1.4525002   0.05412297]
avg: -0.04116

Please try `sess.run(tf.global_variable_initilizer())` instead of your initialization. Does this help? And maybe set the length of your tensor to 2 so that you can visually compare both. — Lau, Nov 06 '18 at 16:10
@Lau sess.run(tf.global_variable_initilizer()) is the same as sess.run(y.initializer). And apply sess.run(tf.equal(x, y)) you will get a tensor with all False. Have a try. — Kuo, Nov 06 '18 at 17:41
Ok, sry. I tested some code now. `tf.random_normal` draws a normal distribution at every call. This is meant to be, because you want sometime generate noise. If your variables to be static use: `tf.get_variable("test", shape=[3], initializer=tf.random_normal_initializer())`. — Lau, Nov 06 '18 at 22:27

Kuo · Accepted Answer · 2019-11-12T06:17:19.270

The true cause is that: x = tf.random_normal(seed = initial_seed)is evolving every time when applying sess.run() but produces the same tensor series x0-x1-x2 if restart running the script. Here provides some explanation on random seed.

To guarantee the same x after every first run, we need reinitialize it. Not sure there is a decent way for my case. But we can set x as a variable and initialize with a fixed seed. Either tf.get_variable or tf.Variable is OK. I find this answer fit my question.

Here is my final code. It works.

import tensorflow as tf

initializer = tf.random_normal_initializer(seed = 1)
x = tf.get_variable(name = 'x', shape = [3], dtype = tf.float32, initializer = initializer)
y = tf.Variable(initial_value = x)

diff = tf.subtract(x, y)
avg = tf.reduce_mean(diff)

sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())

print('x0:', sess.run(x))
print('y0:', sess.run(y))

print('x1:', sess.run(x))
print('y1:', sess.run(y))

print('x2:', sess.run(x))
print('y2:', sess.run(y))

print('diff:', sess.run(diff))
print('avg:', sess.run(avg))
sess.close()

x0: [-0.8113182   1.4845988   0.06532937]
y0: [-0.8113182   1.4845988   0.06532937]
x1: [-0.8113182   1.4845988   0.06532937]
y1: [-0.8113182   1.4845988   0.06532937]
x2: [-0.8113182   1.4845988   0.06532937]
y2: [-0.8113182   1.4845988   0.06532937]
diff: [0. 0. 0.]
avg: 0.0

Tensorflow: tf.random_normal get different results with the same initial seed

1 Answers1