Resample category data with timedelta index

[mapa17]

Hi,

I get a very strange behavior when i try to resample categorical data with and timedelta index, as compared to a datetime index.

>> d1 = pd.DataFrame({'Group_obj': 'A'}, index=pd.date_range('2000-1-1', periods=20, freq='s'))
>> d1['Group'] = d1['Group_obj'].astype('category')
>> d1
                    Group_obj Group
2000-01-01 00:00:00         A     A
2000-01-01 00:00:01         A     A
2000-01-01 00:00:02         A     A
2000-01-01 00:00:03         A     A
2000-01-01 00:00:04         A     A
2000-01-01 00:00:05         A     A
2000-01-01 00:00:06         A     A
2000-01-01 00:00:07         A     A
2000-01-01 00:00:08         A     A
2000-01-01 00:00:09         A     A
2000-01-01 00:00:10         A     A
2000-01-01 00:00:11         A     A
2000-01-01 00:00:12         A     A
2000-01-01 00:00:13         A     A
2000-01-01 00:00:14         A     A
2000-01-01 00:00:15         A     A
2000-01-01 00:00:16         A     A
2000-01-01 00:00:17         A     A
2000-01-01 00:00:18         A     A
2000-01-01 00:00:19         A     A

>> corr = d1.resample('10s', how=lambda x: (x.value_counts().index[0]))
>> corr
                    Group_obj Group
2000-01-01 00:00:00         A     A
2000-01-01 00:00:10         A     A

>> corr.dtypes
Group_obj    object
Group        object
dtype: object

>> d2 = d1.set_index(pd.to_timedelta(list(range(20)), unit='s'))
>> fxx = d2.resample('10s', how=lambda x: (x.value_counts().index[0]))
>> fxx
         Group_obj  Group
00:00:00         A    NaN
00:00:10         A    NaN

>> fxx.dtypes
Group_obj     object
Group        float64
dtype: object

It seems to me the aggregated result in case of using timedelta as an index for the category is always NaN.
Should this be?

Thx

PS: is there a way to specify the dtype for the aggregated columns?

[jreback]

hmm, does appear a little buggy.

you shouldn't need to specify the dtype on aggregations they are inferred. Here I think there is an embedded exception which is caught in stead of actuallly computing correctly.

[jreback]

I look after #11841 as the timedelta resampling is tested a bit more there (but not enough!)

[BranYang]

The root cause of this issue is that, when construct Series from a dict with TimedeltaIndex as key, it will treat the value as float64. See pandas/core/series.py, from line 172 to 185

try:
    if isinstance(index, DatetimeIndex):
        if len(data):
            # coerce back to datetime objects for lookup
            data = _dict_compat(data)
            data = lib.fast_multiget(data, index.astype('O'),
                                     default=np.nan)
        else:
            data = np.nan
    elif isinstance(index, PeriodIndex):
        data = ([data.get(i, nan) for i in index]
                if data else np.nan)
    else:
        data = lib.fast_multiget(data, index.values,
                                 default=np.nan)

I believe just change isinstance(index, PeriodIndex): to isinstance(index, (PeriodIndex, TimedeltaIndex): would solve this issue

Before

In [5]: fxx = d2.resample('10s', how=lambda x: (x.value_counts().index[0]))

In [6]: fxx
Out[6]:
         Group_obj  Group
00:00:00         A    NaN
00:00:10         A    NaN

After

In [5]: fxx = d2.resample('10s', how=lambda x: (x.value_counts().index[0]))

In [6]: fxx
Out[6]:
         Group_obj Group
00:00:00         A     A
00:00:10         A     A