[FIXED] .float() in pytorch changes the value of an int

Issue

import torch
torch.set_printoptions(precision=1, sci_mode=False)

numeric_seq_id = 2021080918959999952

t = torch.tensor(numeric_seq_id)
tt = torch.tensor(numeric_seq_id).float() # !!!

print(t, tt)

output is

tensor(2021080918959999952) tensor(2021080905052848128.)

We could see that tt's value is changed after .float() transform.

Why is here such a difference on the value?

---

ps. pytorch's version = 1.10.1

python's version = 3.8

Solution

This is not pytorch specific, but an artifact of how floats (or doubles) are represented in memory (see this question for more details), which we can also see in numpy:

import numpy as np

np_int = np.int64(2021080918959999952)
np_float = np.float32(2021080918959999952)
np_double = np.float64(2021080918959999952)

print(np_int, int(np_float), int(np_double))

Output:

2021080918959999952 2021080905052848128 2021080918960000000

Answered By - FlyingTeller