为什么round（x）和round（np.float64（x））之间有区别？

Question

据我了解，2.675和numpy.float64（2.675）都是相同的数字。但是，round（2.675，2）得到2.67，而round（np.float64（2.675），2）得到2.68。为什么会这样？

import numpy as np
from decimal import Decimal

x = 2.675
np_x = np.float64(x)
type(x) # float
Decimal(x)    # Decimal('2.67499999999999982236431605997495353221893310546875')
Decimal(np_x) # Decimal('2.67499999999999982236431605997495353221893310546875')
x == np_x # True

# This is the bit that bothers me
round(x, 2) # 2.67
round(np_x, 2) # 2.68

# Using numpy's round gives 2.68 for both the numpy float as well as the Python built-in float...
np.round(x, 2) # 2.68
np.round(np_x, 2) # 2.68

# ... but this is because it might be converting the number to a numpy float before rounding
type(np.round(x, 2)) # numpy.float64

# Versions
# Python 3.6.8 running on 64-bit Windows 10
# Numpy 1.16.2

Answer 1

非常有趣的问题:)看起来它与以 5 结尾的数字有关。Numpy 对它们进行四舍五入，但并不总是......

# list of incoherences between Python Numpy with round(x, 2)
for i in range(1001):
    x = i/1000
    np_x = np.float64(x)
    if round(x, 2) != round(np_x, 2):
        print(x)

# 0.005
# 0.015
# 0.025   <<< some values are missing!
# 0.065
# 0.075
# 0.085
# ...