CODEWITHSUNDEEP
tensorflow
Milad P.
Milad P.

Reputation: 5027

Error when adding two large numbers in TensorFlow

In the following code, I am writing a four loop to print the Fibonacci's sequence, to practice using TensorFlow. However, after a few iterations, it makes the numbers negative and then just returns zeros. Why? How can I fix this? It works fine if I use floats by the way. Also, why is this soooo slow compared to a straightforward algorithm? 

import tensorflow as tf

var1 = tf.Variable(1, tf.int8)
var2 = tf.Variable(1, tf.int8)
temp = tf.Variable(0, tf.int8)

var12 = tf.add(var1, var2)
task1 = tf.assign(var1, var2)
task2 = tf.assign(var2, var12)

task3 = tf.assign(var2, tf.add(var1, temp))

init_op = (tf.initialize_all_variables())

with tf.Session() as sess:
    sess.run(init_op)
    for _ in range(50):
        sess.run(var12)
        sess.run(task1)
        sess.run(task2)
        print(var12.eval(), end=',')

Output: 3,6,12,24,48,96,192,384,768,1536,3072,6144,12288,24576,49152,98304,196608,393216,786432,1572864,3145728,6291456,12582912,25165824,50331648,100663296,201326592,402653184,805306368,1610612736,-1073741824,-2147483648,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,

Upvotes: 0

Views: 321

Answers (1)

mdaoust
mdaoust

Reputation: 6367

For your main question: It's an issue with how computers store numbers. What you're experiencing here is called an integer overflow.

The numbers don't fit in the variables, and Tensorflow doesn't try to fix this because that would slow it down (maybe you're used to python handling this for you. 999**999 ftw)

Your variables are of type int32, because that's the default. You're passing tf.int8 to tf.Variable. An int8 would roll over at 128. But it's landing in the Trainable parameter. Switch that to:

tf.Variable(1,dtype=tf.int64)

int64 is larger.

Also, that's not the correct sequence. It's just doubling the previous element. That sess.run(var12) doesn't do anything (var is a bad name for it, it's not a variable). var12 is automatically re-calculated when you run task2.

Why is it slow?

> python fib.py

It's spending almost all of that time importing Tensorflow, it's big.

Good luck, I hope that helps.

Upvotes: 1

Related Questions