从熊猫的数据框中删除无限值？

what is the quickest/simplest way to drop nan and inf/-inf values from a pandas DataFrame without resetting mode.use_inf_as_null? I’d like to be able to use the subset and how arguments of dropna, except with inf values considered missing, like:

df.dropna(subset=["col1", "col2"], how="all", with_inf=True)

is this possible? Is there a way to tell dropna to include inf in its definition of missing values?

The simplest way would be to first replace infs to NaN:

df.replace([np.inf, -np.inf], np.nan)

and then use the dropna:

df.replace([np.inf, -np.inf], np.nan).dropna(subset=["col1", "col2"], how="all")

For example:

In [11]: df = pd.DataFrame([1, 2, np.inf, -np.inf])

In [12]: df.replace([np.inf, -np.inf], np.nan)
Out[12]:
    0
0   1
1   2
2 NaN
3 NaN

The same method would work for a Series.

With option context, this is possible without permanently setting use_inf_as_na. For example:

with pd.option_context('mode.use_inf_as_na', True):
    df = df.dropna(subset=['col1', 'col2'], how='all')

Of course it can be set to treat inf as NaN permanently with

pd.set_option('use_inf_as_na', True)

For older versions, replace use_inf_as_na with use_inf_as_null.

Here is another method using .loc to replace inf with nan on a Series:

s.loc[(~np.isfinite(s)) & s.notnull()] = np.nan

So, in response to the original question:

df = pd.DataFrame(np.ones((3, 3)), columns=list('ABC'))

for i in range(3): 
    df.iat[i, i] = np.inf

df
          A         B         C
0       inf  1.000000  1.000000
1  1.000000       inf  1.000000
2  1.000000  1.000000       inf

df.sum()
A    inf
B    inf
C    inf
dtype: float64

df.apply(lambda s: s[np.isfinite(s)].dropna()).sum()
A    2
B    2
C    2
dtype: float64

Use (fast and simple):

df = df[np.isfinite(df).all(1)]

This answer is based on DougR’s answer in an other question. Here an example code:

import pandas as pd
import numpy as np
df=pd.DataFrame([1,2,3,np.nan,4,np.inf,5,-np.inf,6])
print('Input:\n',df,sep='')
df = df[np.isfinite(df).all(1)]
print('\nDropped:\n',df,sep='')

Result:

Input:
    0
0  1.0000
1  2.0000
2  3.0000
3     NaN
4  4.0000
5     inf
6  5.0000
7    -inf
8  6.0000

Dropped:
     0
0  1.0
1  2.0
2  3.0
4  4.0
6  5.0
8  6.0

Yet another solution would be to use the isin method. Use it to determine whether each value is infinite or missing and then chain the all method to determine if all the values in the rows are infinite or missing.

Finally, use the negation of that result to select the rows that don’t have all infinite or missing values via boolean indexing.

all_inf_or_nan = df.isin([np.inf, -np.inf, np.nan]).all(axis='columns')
df[~all_inf_or_nan]

The above solution will modify the infs that are not in the target columns. To remedy that,

lst = [np.inf, -np.inf]
to_replace = {v: lst for v in ['col1', 'col2']}
df.replace(to_replace, np.nan)

You can use with . You should ensure first your dataframe series are all of type float. Then use dropna with your existing logic.

print(df)

       col1      col2
0 -0.441406       inf
1 -0.321105      -inf
2 -0.412857  2.223047
3 -0.356610  2.513048

df = df.mask(np.isinf(df))

print(df)

       col1      col2
0 -0.441406       NaN
1 -0.321105       NaN
2 -0.412857  2.223047
3 -0.356610  2.513048