Make pd.Series(index=values) equivalent to pd.Series(index=pd.Index(values))?

[toobaz]

Code Sample, a copy-pastable example if possible

In [2]: d = {(0, 1) : 2, (3, 4) : 5}

In [3]: pd.Index(d)
Out[3]: Index([(0, 1), (3, 4)], dtype='object')

In [4]: pd.Index(list(d))
Out[4]: 
MultiIndex(levels=[[0, 3], [1, 4]],
           labels=[[0, 1], [0, 1]])

In [5]: pd.Series(d).index
Out[5]: 
MultiIndex(levels=[[0, 3], [1, 4]],
           labels=[[0, 1], [0, 1]])

In [6]: pd.Series(index=list(d)).index
Out[6]: Index([(0, 1), (3, 4)], dtype='object')

Problem description

I guess Out[3]: and Out[6]: are wrong.

Expected Output

Out[4]:

Output of pd.show_versions()

INSTALLED VERSIONS

commit: None
python: 3.5.3.final.0
python-bits: 64
OS: Linux
OS-release: 4.9.0-4-amd64
machine: x86_64
processor:
byteorder: little
LC_ALL: None
LANG: en_GB.UTF-8
LOCALE: en_GB.UTF-8

pandas: 0.22.0.dev0+202.g97bd66ea8
pytest: 3.0.6
pip: 9.0.1
setuptools: 33.1.1
Cython: 0.25.2
numpy: 1.12.1
scipy: 0.18.1
pyarrow: None
xarray: None
IPython: 5.2.2
sphinx: None
patsy: 0.4.1+dev
dateutil: 2.6.0
pytz: 2016.10
blosc: None
bottleneck: 1.2.0
tables: 3.3.0
numexpr: 2.6.1
feather: 0.3.1
matplotlib: 2.0.0
openpyxl: 2.3.0
xlrd: 1.0.0
xlwt: 1.2.0
xlsxwriter: None
lxml: 3.7.1
bs4: 4.5.3
html5lib: 0.999999999
sqlalchemy: 1.0.15
pymysql: None
psycopg2: None
jinja2: 2.8
s3fs: None
fastparquet: None
pandas_gbq: None
pandas_datareader: None

[toobaz]

This suggests the above is partly indended... but there should be a better fix.

[jreback]

In [16]: pd.Index(list(d), tupleize_cols=True)
Out[16]: 
MultiIndex(levels=[[0, 3], [1, 4]],
           labels=[[0, 1], [0, 1]])

In [17]: pd.Index(list(d), tupleize_cols=False)
Out[17]: Index([(0, 1), (3, 4)], dtype='object')

[jreback]

I hate this kw.

[jreback]

@TomAugspurger

[toobaz]

It gets better:

In [2]: pd.Index([[0, 1, 2], [3, 4, 5]])
Out[2]: Index([[0, 1, 2], [3, 4, 5]], dtype='object')

In [3]: pd.Series(index=[[0, 1, 2], [3, 4, 5]])
Out[3]: 
0  3   NaN
1  4   NaN
2  5   NaN
dtype: float64

... which is not "tupleization" (notice the orientation)!

The interpretation of lists as index levels in Series (and DataFrame) is totally unexpected to me... but since it is there, maybe we want to support it in Index too?

[jreback]

[2] is invalid but prob not checked

these are not hashable sub elements

absolutely we do not want to add complexity like this

[jorisvandenbossche]

Yes, and I didn't know this worked like that in Series. I would rather disallow it there as well.

[toobaz]

I'm totally fine with disallowing this feature, but it is precisely the main task of lib.clean_index_list, and it is tested, so it's not accidental.

(and assuming we didn't want to drop it, integrating it into Index would actually simplify the code, which is why I was suggesting it... but again, if you want to drop it, good)

[jorisvandenbossche]

and it is tested, so it's not accidental.

It's not accidental, that's true, but that test is more testing that the Period is preserved and not converted into an int, than testing the "index levels as list" feature I think. Although there are maybe other more explicit tests.

[toobaz]

OK, what about this related behavior? (Just for reference), it is used at ~20 places in testing.

In [2]: s = pd.Series([1, 2, 3])

In [3]: s.index = [['a', 'b', 'b'], ['d', 'd', 'e']]

In [4]: s
Out[4]: 
a  d    1
b  d    2
   e    3
dtype: int64

Another variation:

In [5]: s.index = [pd.Index(['a', 'b', 'b']), pd.Index(['d', 'd', 'e'])]

Do we like them?

[toobaz]

By the way: In [3]: was actually fixed by #18514 (not In [6]). Retitling accordingly.

[jreback]

[3] is pretty clear. We coerce this to a MultiIndex. [5] should work as well (to a MI). I see why you think #18484 (comment) is inconsistent.

are:

a) pd.Index([[....], [....]])

b) infer_to_index([[...], [....]])

the same, where infer_to_index happens upon assignment (e.g. b). I actually could buy that these infer to the same, namely a MultiIndex (an Index with nested elements is completely non-supported however)

[toobaz]

OK. So finally if I understand correctly the plan is to

• make pd.Index([[0, 1, 2], [3, 4, 5]]) return a MultiIndex (without any argument in the vein of tupleize_cols=False, since anyway we don't want lists as labels)
• suppress _ensure_index, which becomes redundant (since it seems to me it isn't catching any other case)

[jreback]

suppress _ensure_index, which becomes redundant (since it seems to me it isn't catching any other case)

the original intent of _ensure_index was I think to infer whether things should be an Index or a MultiIndex. If you can remove code great.