GroupBy.apply type upcasting regression

[wesm]

Blocker for 0.11.1. This got broken sometime between 0.11.0 and master; I haven't had time to bisect yet. Essentially, write a mixed type DataFrame and a groupby function that extracts a row. This used to do type inference and convert object back to numeric in the columns of the resulting DataFrame. Didn't have a test so I don't blame anyone for breaking it by accident =)

In [151]: cafdata.dtypes
Out[151]:
id            int64
food         object
fgroup       object
nutrient     object
ngroup       object
units        object
value       float64
dtype: object

In [152]:
def max_value(group):
    return group.ix[group['value'].idxmax()]

max_value(cafdata)

Out[152]:
id                                      14366
food        Tea, instant, unsweetened, powder
fgroup                              Beverages
nutrient                             Caffeine
ngroup                                  Other
units                                      mg
value                                    3680
Name: 336702, dtype: object

In [153]: cafdata.groupby('fgroup').apply(max_value).dtypes
Out[153]:
id          object
food        object
fgroup      object
nutrient    object
ngroup      object
units       object
value       object
dtype: object

[wesm]

Actually not sure this is a regression, here's with 0.11.0:

In [166]:

grouped = data.groupby(['nutrient', 'fgroup'])
results = grouped.apply(max_value)
results.dtypes
Out[166]:
id          object
food        object
fgroup      object
nutrient    object
ngroup      object
units       object
value       object
dtype: object

for some reason making a bar plot of the value column fails in master. i will investigate when i can

[cpcloud]

is the failure a type error? plotting object Series was deprecated by #3572. should do results.value.convert_objects().plot(kind='bar')

[wesm]

Um, really? In this case the resulting plot is totally fine because all of the objects inside are floats. Why not attempt to convert_objects for dtype object and raise if it can't be converted to numeric?

[wesm]

I'm strongly 👎 on that change

[cpcloud]

ok then will change it now

[cpcloud]

@wesm cpcloud/pandas@fe3f01f if u want to check while waiting for travis...

[cpcloud]

@wesm also is that the only issue here? not clear to me if there's still a groupby issue from ur comments above

[cpcloud]

oh i c. still looks like casting back to original types is an issue

[jreback]

seldom used case, but test and fixed now