BUG: fix reverse comparison operations for Categorical

[immerrr]

This PR fixes #8658 and also adds pd.lib.unbox_if_zerodim that extracts values from zerodim arrays and adds detection of pd.Period, datetime.date and datetime.time in pd.lib.isscalar.

I tried making unbox_if_zerodim usage more implicit but there's just too much to test when this is introduced (think every method that accepts a scalar must have another test to check it also accepts zerodim array). Definitely too much to add for a single PR, but overall zerodim support could be added later on class-by-class basis.

[jreback]

hmm
what about nan/NaT ?

[immerrr]

what about nan/NaT?

nan is good as long as it's some sort of float.
NaT is good as long as it's an instance of datetime.date:

In [4]: type(pd.NaT).mro()
Out[4]: 
[pandas.tslib.NaTType,
 pandas.tslib._NaT,
 pandas.tslib._Timestamp,
 datetime.datetime,
 datetime.date,
 object]

But OK, I'll add the tests

[jreback]

I don't think u need this PandasScalarTypes
all pandss objects now have a _typ property
so for period u can just check it directly
eg
getsttr(obj,'_typ',None) in ['period]

[immerrr]

In [11]:  getattr(pd.Period('2014-01-01') ,'_typ', 'i dont have a _typ')
Out[11]: 'i dont have a _typ'

PeriodIndex has _typ = 'periodindex', but I'm not sure what does it suppose to mean.

[immerrr]

This can probably be solved by splitting out a PandasScalarObject class from PandasObject. It may add no functionality to PandasObject mixin, but one could inherit from it to underline that it is a scalar and should be treated accordingly.

[jreback]

I realized that I didn't fix Period like I did Categorical, see here:https://github.com/pydata/pandas/blob/master/pandas/core/common.py#L85

if you add _typ to Period then you can create ABCPeriod and then import it in cython and do a direct isinstance check, no need for PandasScalarObject (though you could create that as a congolomerate type as ABCScalar if you want ('categorical','period','timedelta','timestamp')) I think would work.

[immerrr]

This reminds me. I was always wondering why are those implemented in pandas by hand instead of using abc.ABCMeta?

[jreback]

not sure - maybe was a py2/py3 issue? (these were done a while ago)

[jreback]

pls also create an issue to change use of np.isscalar to use lib.isscalar

[immerrr]

I'm also not sure about the name of unbox_if_zerodim, it may create an impression that it's related to timestamp boxing/unboxing, but I couldn't come up with a better yet concise name.

[jreback]

how about _maybe_to_scalar ?

[immerrr]

how about _maybe_to_scalar ?

It kind of loses the zerodim part and, besides, array may contain a non-scalar, e.g. np.array({}).item() == {}. Maybe item_if_zerodim?

[immerrr]

Or, given there's a values_from_object, item_from_zerodim?

[jreback]

item_from_zerodim sounds good (maybe doc better that I did values_from_object :)

[jreback]

merged via 02de853

I changed it very slightly to make the period checking more consistent with internals. (and IIRC need this somewhere else anyhow).

thanks!

these subtle changes are excellent! all towards better code!