Converting from categorical to int ignores NaNs

[hnykda]

Code Sample, a copy-pastable example if possible

In [6]: s = pd.Series([1, 0, None], dtype='category')                                                                                                                                                                                            

In [7]: s                                                                                                                                                                                                                                      
Out[7]: 
0      1
1      0
2    NaN
dtype: category
Categories (2, int64): [0, 1]

In [8]: s.astype(int)                                                                                                                                                                                                                          
Out[8]: 
0                      1
1                      0
2   -9223372036854775808  # <- this is unexpected
dtype: int64

Problem description

When converting categorical series back into Int column, it converts NaN to incorect integer negative value.

Expected Output

I would expect that NaN in category converts to NaN in IntX(nullable integer) or float.

When trying to use d.astype('Int8'), I get an error dtype not understood

Output of pd.show_versions()

In [147]: pd.show_versions()                                                                                                                                                                                                                   

INSTALLED VERSIONS
------------------
commit           : None
python           : 3.7.4.final.0
python-bits      : 64
OS               : Linux
OS-release       : 5.2.13-arch1-1-ARCH
machine          : x86_64
processor        : 
byteorder        : little
LC_ALL           : None
LANG             : en_US.UTF-8
LOCALE           : en_US.UTF-8

pandas           : 0.25.1
numpy            : 1.17.2
pytz             : 2019.2
dateutil         : 2.8.0
pip              : 19.2.3
setuptools       : 41.2.0
Cython           : None
pytest           : 5.1.2
hypothesis       : None
sphinx           : None
blosc            : None
feather          : 0.4.0
xlsxwriter       : None
lxml.etree       : None
html5lib         : None
pymysql          : None
psycopg2         : None
jinja2           : None
IPython          : 7.8.0
pandas_datareader: None
bs4              : None
bottleneck       : None
fastparquet      : None
gcsfs            : None
lxml.etree       : None
matplotlib       : None
numexpr          : 2.7.0
odfpy            : None
openpyxl         : None
pandas_gbq       : None
pyarrow          : 0.14.1
pytables         : None
s3fs             : None
scipy            : None
sqlalchemy       : None
tables           : 3.5.2
xarray           : None
xlrd             : None
xlwt             : None
xlsxwriter       : None

[TomAugspurger]

I would expect s.atype(int) to raise, the same as s.astype(float).astype(int)

To get a nullable integer, s.astype("Int8") should work, but let's leave that as a second issue (we may already have one for it, not sure).

[dsaxton]

I would expect s.atype(int) to raise, the same as s.astype(float).astype(int)

To get a nullable integer, s.astype("Int8") should work, but let's leave that as a second issue (we may already have one for it, not sure).

It seems like the weird value gets introduced here: https://github.com/pandas-dev/pandas/blob/master/pandas/core/arrays/categorical.py#L523 . Would this be a numpy issue?

In [1]: import numpy as np                                                      

In [2]: arr = np.array([1.0, 0.0, np.nan])                                      

In [3]: np.array(arr, dtype=np.int64)                                            
Out[3]: array([                   1,                    0, -9223372036854775808])

[hnykda]

I would expect s.atype(int) to raise, the same as s.astype(float).astype(int)

Well, but that's a bit sad because it means that once you convert to categoricals, you cannot get back :-( .

[TomAugspurger]

It seems like the weird value gets introduced here: https://github.com/pandas-dev/pandas/blob/master/pandas/core/arrays/categorical.py#L523 . Would this be a numpy issue?

Thanks for tracking it down. I think that's just NumPy's defined behavior.

We might want astype_nansafe there (should also test with inf).

Well, but that's a bit sad because it means that once you convert to categoricals, you cannot get back :-( .

What do you mean?

[hnykda]

What do you mean?

Well, I meant something like if you did s = pd.Series([1, None]).astype(int).astype(category) and now wanted to get back to int somehow. The workaround seems to be s.astype('float').astype('Int8') for example.

I understand it's tricky, especially when dealing with nullable integers. It's true that raising is probably better than trying to be somehow clever.

[TomAugspurger]

Sorry, I'm still not understanding. In your example,

s = pd.Series([1, None]).astype(int)

raises. The conversion from float to int is what raises, not to or from categorical.

[hnykda]

Sorry, I wrote it wrongly. I meant if you do:

In [3]: s = pd.Series([1, None], dtype='Int8').astype('category')

and now wanted to get back to numerics, or more specifically e.g. Int8. You have to do

In [6]: s.astype(float).astype('Int8')

while it would be super cool to do just s.astype('Int8')

[TomAugspurger]

s.astype('Int8')

That should work just fine. As I said earlier

To get a nullable integer, s.astype("Int8") should work, but let's leave that as a second issue (we may already have one for it, not sure).

[dsaxton]

Thanks for tracking it down. I think that's just NumPy's defined behavior.

We might want astype_nansafe there (should also test with inf).

Seems that astype_nansafe expects an ndarray but self is Categorical here. Any thoughts on the best workaround?

[TomAugspurger]

Hmm, OK.

It's a bit unfortunate, but I we'll need to include something like

if is_integer_dtype(dtype):
    if self.isna().any():
        raise ValueError(...)