Cannot subclass MultiIndex

[janmedlock]

MultiIndex cannot be subclassed because in several spots, including __new__, the output class is hardcoded as MultiIndex rather than cls or self.__class__ or the like.

This code illustrates the problem:

$ python
Python 2.7.10 (default, Sep 13 2015, 20:30:50) 
[GCC 5.2.1 20150911] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import pandas
>>> class MyMultiIndex(pandas.MultiIndex):
...     pass
... 
>>> multiindex = MyMultiIndex([['a'], ['b']],
...                           [[0], [0]])
>>> print type(multiindex)
<class 'pandas.core.index.MultiIndex'>

The last line should instead read

<class '__main__.MyMultiIndex'>

I have a patch in the works.

>>> pandas.show_versions()

INSTALLED VERSIONS
------------------
commit: None
python: 2.7.10.final.0
python-bits: 64
OS: Linux
OS-release: 4.1.0-2-amd64
machine: x86_64
processor: 
byteorder: little
LC_ALL: None
LANG: en_US.utf8

pandas: 0.16.2.dev
nose: 1.3.6
Cython: 0.23.2
numpy: 1.9.2
scipy: 0.14.1
statsmodels: 0.6.1
IPython: 2.3.0
sphinx: 1.3.1
patsy: 0.4.0
dateutil: 2.2
pytz: 2012c
bottleneck: None
tables: 3.1.1
numexpr: 2.4.3
matplotlib: 1.4.2
openpyxl: None
xlrd: 0.9.4
xlwt: 0.7.5
xlsxwriter: None
lxml: None
bs4: 4.4.0
html5lib: 0.999
httplib2: 0.9.1
apiclient: None
sqlalchemy: None
pymysql: None
psycopg2: None

[janmedlock]

I just submitted pull request #11268 to fix this.

[jreback]

can you provide a use-case for this (not objecting to the change though)

[janmedlock]

Sure. I don't want to need to know the order of the indices in the MultiIndex. E.g.

mi = MultiIndex((('a', 'b'), ('A', 'B')), ((0, 0), (0, 1), (1, 0), (1, 1)), names=('foo', 'bar'))
df = Series(..., index = mi)

I want to be able to write something like

df[mi(foo = 'a', bar = 'b')]

or similar to get the correct Series elements.

Maybe this functionality already exists in MultiIndex, but I didn't see it and decided to subclass.

[jreback]

we've had discussions here

mi = pd.MultiIndex.from_product([list('ab'),list('AB')],names=['foo','bar'])

In [5]: s = Series(range(4),index=mi)

In [6]: s
Out[6]: 
foo  bar
a    A      0
     B      1
b    A      2
     B      3
dtype: int64

to allow:

s.loc[{'foo' : 'a', 'bar' : B'}]

xray calls this .sel, see here

or maybe
s.loc[S(foo='a',bar='B')] where S is simply

class S(object):
    def __call__(self, *kwargs):
         return kwargs

if we allow the dict syntax

you wouldn't normally want to have to do:

s[s.index(foo='a', bar='B')], although not terrible

so, you don't need to sub-class at all, simply patch it in

In [8]: MultiIndex.__call__ = lambda self, **kwargs: tuple([ (kwargs.get(n,slice(None))) for n in self.names ])

In [9]: s.loc[s.index(foo='b',bar='B')]
Out[9]: 3

In [10]: s.loc[s.index(foo='b')]
Out[10]: 
foo  bar
b    A      2
     B      3
dtype: int64

In [11]: s.loc[s.index(bar='B')]
Out[11]: 
foo
a    1
b    3
dtype: int64

In [12]: s.loc[s.index(bar=['B'])]
Out[12]: 
foo  bar
a    B      1
b    B      3
dtype: int64

kind of like that actually :)

cc @shoyer

[janmedlock]

Thanks for the answer!

[jrderuiter]

So this never got changed? Was there any special reason for not persuing this further?

[jreback]

the PR went stale. in principle it is ok if you'd like to pick it up; needs more testing