CODEWITHSUNDEEP
pythontensorflow
njk
njk

Reputation: 655

Order of addends in a sum affects the result?

In the following code, op1 and op2 should be identical. The only difference is in the order the addends appear in the sum. However, np.array_equal(res1, res2) returns False. Why does this happen? (I'm using TensorFlow version 0.12.1).

import numpy as np
import tensorflow as tf

V = tf.constant([[-3.47429895, -5.99409866],
                 [-4.84888363, 5.3685813],
                 [8.32318401, 0.62552071]], dtype=tf.float32)

rs = tf.reduce_sum(V, axis=1, keep_dims=True)
m = -tf.matmul(V, V, transpose_b=True)
op1 = m + rs + tf.transpose(rs)
op2 = rs + tf.transpose(rs) + m

with tf.Session() as session:
    res1 = session.run(op1)
    res2 = session.run(op2)
    assert np.array_equal(res1, res2) # False, why?

Upvotes: 1

Views: 77

Answers (1)

Priyatham
Priyatham

Reputation: 2897

It's happening because floating point operations are not accurate and changing the order can change the results albeit, the discrepancy is not significant.

The problem here is with np.array_equal, the error between res1 and res2 is more than it can tolerate. Try printing res1 - res2 to get a better idea. You might want to use numpy.allclose with properly chosen tolerance values.

Upvotes: 2

Related Questions