API: MultiIndex.get_level_values should have label only / level only mode?

[sinhrks]

Currently, MultiIndex.get_level_values handles both names and levels, it return different levels depending on its name (like .ix). Adding an option (or separate method) like .loc and .iloc to avoid any unexpected results?

idx = pd.MultiIndex.from_tuples([(1, 'A'), (2, 'B')])
idx.get_level_values(0)
# Int64Index([1, 2], dtype='int64')

idx = idx.set_names([1, 0])
idx.get_level_values(0)
# Index([u'A', u'B'], dtype='object', name=0)

[shoyer]

Maybe idx.get_level_values and idx.get_ilevel_values? Or have we already locked in the mixed functionality of get_level_values?

[sinhrks]

I think we can't change get_level_values. Options are listed below.

I think users prefer levels to names to avoid ambiguous, cos names can be duplicated or None. Thus option 1 can cover most cases.

1. Add get_ilevel_values which accepts level only. No additional for names.
2. Add get_ilevel_values which accepts level only and get_llevel_values (?) which accepts names only.
3. Add allow_level (?) and allow_names (?) options to get_level_values with default True.

[jreback]

already there is inference to figure out if they mean positional (if no integers are in the names) or label based - eg like ix

but I have never seen this come up as a practical issue

I guess get_ilevel_values would be fine

ironically in .query we allow

ilevel as a search term so this would be consistent at least

[jreback]

http://pandas.pydata.org/pandas-docs/stable/indexing.html#multiindex-query-syntax

[jreback]

just throwing it out there - could allow this:

df.index.ix(0) #current
df.index.loc('foo) #label only
df.index.iloc(0) #positional only

or maybe

df.index.ix[0]

further this might be useful

df.index.ix[0,1] -> grab first 2 levels

[sinhrks]

Right. Adding ix stacks sounds consistent and reasonable. Let me try.

[jorisvandenbossche]

I am -1 on adding ix like above. If we add ix (or loc/iloc to the index), I think the usage should be equivalent to as if you would have a column with the name 'index' and you are indexing frame.index. If it has other typical behaviour, we should just call it differently I think.