ENH: add .dropinf() and/or .dropna(drop_inf=True) methods

[adrivsh]

There are known ways of droping inf and nan (.replace(np.inf, np.nan).dropna()) but still, the API could be expanded:

.dropinf(inplace=False) 

.dropna(drop_inf = False)

Are both self explanatory wishes and it is no stretch to see how they are usefull.
Note for instance that today's API does not provide a one liner to drop inf in place while keeping nans.

Best

PS: not sure how to submit wishes.

[jreback]

In [1]: s = Series(np.random.randn(5))

In [2]: s[1] = np.nan

In [3]: s[2] = np.inf

In [4]: s[3] = -np.inf

In [5]: s
Out[5]: 
0    0.760816
1         NaN
2         inf
3        -inf
4    0.441029
dtype: float64

In [6]: s.dropna()
Out[6]: 
0    0.760816
2         inf
3        -inf
4    0.441029
dtype: float64

You can already do this

In [12]: with pd.option_context('use_inf_as_null',True):
   ....:     print(s.dropna())
   ....:     
0    0.760816
4    0.441029
dtype: float64

[TomAugspurger]

Using the context manager is OK, but I wouldn't want to encourage setting this option globally.

I'll typically use s[np.isfinite(s)] when I have infinities (which is often). This also doesn't make it clear that you're excluding NaNs (which apparently are defined to not be finite).

+0.5 on adding dropinf a method.

[jreback]

yep, just showing that as an example. I agree s[np.isifinite(s)] is pretty clear IMHO. maybe add to the docs? (could also add as a related example in .dropna doc-string)

[TomAugspurger]

The nice thing about dropna is you get the how=any/all and axis=0/1 behavior too. If we go the doc route we should include examples on a DataFrame with those. And yeah, np.isfinite's docstring states quite clearly that NaN is not finite.

[adrivsh]

I agree s[np.isifinite(s)] is clear and works. So is s[np.isnull(s)].
We nonetheless have dropna(). And we use dropna. It is convenient and pandaic.
Why not also have dropinf() ? With all associated options: inplace, axis, how, etc.