BUG: DataFrame quantile with only datetime dtypes should provide better error message

[TomAugspurger]

In [41]: df = DataFrame({"A": [1, 2, 3], "B": [2, 3, 4]})

In [42]: df['C'] = pd.date_range('2014-01-01', periods=3, freq='m')

In [43]: df
Out[43]: 
   A  B          C
0  1  2 2014-01-31
1  2  3 2014-02-28
2  3  4 2014-03-31

In [44]: df[['C']].quantile(.5)
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-44-1e5bdc20c0ca> in <module>()
----> 1 df[['C']].quantile(.5)

/Users/tom/Envs/pandas-dev/lib/python2.7/site-packages/pandas/pandas/core/frame.pyc in quantile(self, q, axis, numeric_only)
   4192         quantiles = [[f(vals, x) for x in per]
   4193                      for (_, vals) in data.iteritems()]
-> 4194         result = DataFrame(quantiles, index=data._info_axis, columns=q).T
   4195         if len(is_dt_col) > 0:
   4196             result[is_dt_col] = result[is_dt_col].applymap(lib.Timestamp)

/Users/tom/Envs/pandas-dev/lib/python2.7/site-packages/pandas/pandas/core/frame.pyc in __init__(self, data, index, columns, dtype, copy)
    253             else:
    254                 mgr = self._init_ndarray(data, index, columns, dtype=dtype,
--> 255                                          copy=copy)
    256         elif isinstance(data, collections.Iterator):
    257             raise TypeError("data argument can't be an iterator")

/Users/tom/Envs/pandas-dev/lib/python2.7/site-packages/pandas/pandas/core/frame.pyc in _init_ndarray(self, values, index, columns, dtype, copy)
    365             columns = _ensure_index(columns)
    366 
--> 367         return create_block_manager_from_blocks([values.T], [columns, index])
    368 
    369     @property

/Users/tom/Envs/pandas-dev/lib/python2.7/site-packages/pandas/pandas/core/internals.pyc in create_block_manager_from_blocks(blocks, axes)
   3183         blocks = [getattr(b, 'values', b) for b in blocks]
   3184         tot_items = sum(b.shape[0] for b in blocks)
-> 3185         construction_error(tot_items, blocks[0].shape[1:], axes, e)
   3186 
   3187 

/Users/tom/Envs/pandas-dev/lib/python2.7/site-packages/pandas/pandas/core/internals.pyc in construction_error(tot_items, block_shape, axes, e)
   3164         raise e
   3165     raise ValueError("Shape of passed values is {0}, indices imply {1}".format(
-> 3166         passed,implied))
   3167 
   3168 

ValueError: Shape of passed values is (0, 0), indices imply (1, 0)

[TomAugspurger]

Actually, it's got nothing to do with just having one column (like my original title suggested).

quantile has a numeric_only kwarg. When you have all non_numerics, maybe we could be better with either

• inferring that numeric_only is what you want, since there are 0 non-numeric cols
• returning a better error message

Maybe I should change the numeric_only kwarg to 'infer'. Where 'infer' means:

1. If any numeric dtypes, exclude all non-numeric
2. If all numeric dtypes, don't exclude non-numeric

and it will still accept True and False as before.

[jreback]

hmm, numeric_only is not in the docstring.

I think this should include datetime/timedelta by default. so numeric_only is really an odd choice here (was prob ok before quantile could handle these). you almost always want to exclude object dtypes (and bool types).

[mroeschke]

Looks to give a reasonable result now. Could use a test

In [32]: In [41]: df = DataFrame({"A": [1, 2, 3], "B": [2, 3, 4]})
    ...:
    ...: In [42]: df['C'] = pd.date_range('2014-01-01', periods=3, freq='m')

In [33]: df[['C']].quantile(.5)
Out[33]: Series([], Name: 0.5, dtype: float64)

In [34]: pd.__version__
Out[34]: '1.1.0.dev0+1108.gcad602e16'

[jbrockmendel]

I think the discussion above suggests this isnt a reasonable result, that we dont want to drop datetime column here

[NumberPiOso]

Nowadays the parameter is defined in the docstring as follows:

numeric_onlybool, default True

    If False, the quantile of datetime and timedelta data will be computed as well.

So the results

df.quantile(0.5)
A    2.0
B    3.0
Name: 0.5, dtype: float64


df.quantile(0.5, numeric_only=False)
A                    2.0
B                    3.0
C    2014-02-28 00:00:00

So this is not a BUG anymore, it looks more like an enchancement about the argument of numeric_only defaulting to False.

[NumberPiOso]

take