BUG: Categoricals shouldn't allow non-strings when object dtype is passed (#13919)

[wcwagner]

• closes ERR: Categoricals should not allow non-strings when an object dtype is passed #13919
• tests added / passed
• passes git diff upstream/master | flake8 --diff
• whatsnew entry

Why this change is needed: Categorical variables are by definition single types, so to allow them to take on different types of values is misleading. Object dtypes should only be allowed when ALL strings or ALL periods are passed (due to the way there are handled internally).

The result of this PR will raise a TypeError when a categorical is created that has an object dtype but doesn't contain all string or all period values.

[jreback]

you already have #14027 you don't need a 2nd
just push to the original
ok for this one

[sinhrks]

can u add tests:

• unicode (for py2, include 2 bytes)
• str and unicode mixed (for py2, it's often the case in asian countries)
• bool (not sure it is used as category)
• period with mixed freq

I think some of them can't be covered because of infer_dtype spec and current impl.

[jreback]

mixed dtypes

[wcwagner]

This change will always fail to build because it conflicts with the way MultiIndex.from_arrays, MultiIndex.from_tuples, and MultiIndex.from_product are implemented.

Regardless of the way I implement it, ensuring that Categoricals aren't mixed dtypes will always raise when a MultiIndex is constructed from some listlike with mixed dtypes :\

Do I disregard this and keep working/pushing to this PR?

[jreback]

you would have 2 show an example

[wcwagner]

@jreback Show an ex. here? the docs? add a test and assert it raises?
If here, then

MultiIndex.from_arrays([[1 ,2], ['foo',  3]])

will raise because this

[jreback]

#13854 refactors those routines

build your fix on top of this PR

[wcwagner]

The refactored routines still call Categorical.from_array (here), so using mixed dtypes with MultiIndex.from_arrays would still raise unfortunately.

[jreback]

can you rebase / update?

[wcwagner]

@jreback I don't think this will work. Even though #13854 refactored MultiIndex.from_arrays, it still creates categoricals, potentially of mixed types. So i will be closing this, as I can't think of a clean way to get around this :\

[jreback]

Categoricals should never be mixed type
not sure why you think that the constructor creates them

if you have an example pls show it

[wcwagner]

@jreback If I call

MultiIndex.from_arrays([[1 ,2], ['foo',  3]])

_factorize_from_iterable will create a Categorical from ['foo', 3] here

I'm definitely misunderstanding something here. In reference to your first comment in #13919, shouldn't that raise on creation?

[jreback]

that should raise its not valid

[wcwagner]

@jreback So back to my original problem. I can't think of a clean way to allow MultiIndex.from_arrays to exist, while simultaneously blocking the creation of Categoricals with mixed dtypes.

My immediate idea is to pass in some parameter to indicate that a MultiIndex is being created, but as you said in another one of my PRs -- you can't pass in an extra parameter to determine state

[jorisvandenbossche]

@jreback from a MultiIndex.from_arrays point of vue, such mixed types is perfectly valid IMO

Personally, I don't think we should check this at Categorical construction, I would rather check for this in the hdf code itself.