如何使用张量流准确舍入一半

Question

I'm trying to replicate some C++ based code into Python API of Tensorflow, but I'm having few floating point inaccuracy issues, although I have been able to find one of them.

Generally, Tensorflow seems to round decimals in a round half down manner, meaning that if we have an integer with fractional value exactly equal to 0.5 , the integral part of that decimal will be rounded down to zero:

>>> import tensorflow as tf
>>> tf.Session().run(tf.math.round(2.5))
2.0

whereas, many imperative programming languages that I've encountered do the rounding in a round half up manner. Some of these programming languages are C++ and Python .

In fact, considering that Tensorflow was mostly written in C++, Python and Cuda C++, it seems like a weird convention for them to be using rounding half down method for their functions in API.

---

Question :

Is there any neat way to implement a rounding function that utilizes rounding half up method instead of rounding half down ?

The simplest function I could implement uses tf.floormod method:

>>> def classical_round(x): return tf.cond(tf.math.equal(tf.floormod(x, 1), tf.constant(0.5)), lambda: tf.math.ceil(x), lambda: tf.math.round(x))
...
>>> tf.Session().run(classical_round(4.5))
5.0
>>> tf.Session().run(classical_round(4.49))
4.0
>>> tf.Session().run(classical_round(4.49999999999999))
5.0
>>> tf.Session().run(classical_round(3.2))
3.0

Is this sufficiently accurate method? or can something similar be done by using other tensorflow operations?

Research :

I could only find this Github issue relevant to my problem, I think they ended up adding tf.math.rint for which I couldn't find the place in this problem.

Thank you!

Answer 1

called bankers rounding nothing wrong about that same thing happening in c# too. you can try something like that:

def classical_round(x):
    return tf.math.floor(x+0.5)

sess.run(classical_round(2.5)) #3.0

more information here: https://en.wikipedia.org/wiki/Rounding#Round_half_to_even