如何修复产生不精确结果的numpy浮点运算?

时间:2018-09-05 12:31:57

标签: python pandas numpy comparison floating-accuracy

我需要从已知样本数量的相对频率重建绝对频率。

这应该很容易,但是绝对频率和样本大小为numpy.int64,相对频率为numpy.float64

我知道浮点十进制值通常没有精确的二进制表示形式,并且我们可能会遇到一些精度损失。似乎是这样,浮点运算产生了意外的结果,我不能相信重建的绝对频率。

复制错误的示例代码:

import pandas as pd
import numpy as np

absolutes = np.arange(100000, dtype=np.int64) #numpy.int64
sample_size = absolutes.sum() # numpy.int64
relatives = absolutes / sample_size #float64

# Rebuilding absolutes from relatives

rebuilt_float = relatives * sample_size #float64
rebuilt_int = rebuilt_float.astype(np.int64)

df = pd.DataFrame({'absolutes': absolutes,
                   'relatives': relatives,
                   'rebuilt_float': rebuilt_float,
                   'rebuilt_int': rebuilt_int})

df['check_float'] = df['absolutes'] == df['rebuilt_float']
df['check_int'] = df['absolutes'] == df['rebuilt_int']

print('Failed FLOATS: ', len(df[df['check_float'] == False]))
print('Failed INTS:', len(df[df['check_int'] == False]))
print('Sum of FLOATS:', df['rebuilt_float'].sum())
print('Sum of INTS:', df['rebuilt_int'].sum())

是否可以使用numpy解决问题而不将每个数字都转换为小数?

2 个答案:

答案 0 :(得分:1)

np.isclose(df['absolutes'], df['rebuilt_float'], atol=.99999)

numpy.isclose()是不精确的可感知fp的比较。它有一个额外的参数atolrtol用于相对和绝对公差。

您可以通过更改atol来消除多少舍入错误:

>>> len(np.where( np.isclose(df['absolutes'], df['rebuilt_int'], atol=.99999) == False )[0])
0
>>> len(np.where( np.isclose(df['absolutes'], df['rebuilt_int'], atol=.5) == False )[0])
2767
>>> len(np.where( np.isclose(df['absolutes'], df['rebuilt_int'], atol=1) == False )[0])
0

答案 1 :(得分:0)

如果在将转换后的值转换为整数之前四舍五入,则会得到零个失败的整数。也就是说,使用

rebuilt_int = np.round(rebuilt_float).astype(np.int64)

然后输出

Failed FLOATS:  11062
Failed INTS: 0
Sum of FLOATS: 4999950000.0
Sum of INTS: 4999950000