BUG: dtypes on empty frame are incorrect

[jreback]

In [1]: df = DataFrame(columns=list('xyz'))

In [2]: df
Out[2]: 
Empty DataFrame
Columns: [x, y, z]
Index: []

In [3]: df.dtypes
Out[3]: 
x   NaN
y   NaN
z   NaN
dtype: float64

In [5]: df['x'].dtype
Out[5]: dtype('O')

expected

In [11]: Series(dict([ (s,np.dtype('O')) for s in list('xyz')]))
Out[11]: 
x    object
y    object
z    object
dtype: object

[hayd]

This is bug in apply with an empty DataFrame/Series, I think I have a fix for it coming up.

I remember it was fixed in #2476 with this commit bb7eaff

[jreback]

I think this got messed up after that.....need an explicit test for it....thanks

[hayd]

I have put together commit which better special cases empty DataFrames and does fix this.

However it breaks test_apply_empty_infer_type whose tests seems a little sketchy to me, for example:

no_cols = DataFrame(index=['a', 'b', 'c'])
no_cols.apply(lambda x: x.mean()) 

Tests that this ought to be a DataFrame... whereas I think this ought to be a Series.

...there really is some ambiguity in when DataFrame should be pushed down to Series in an apply and same with columns names/series name.

[jreback]

This is a series (in master)

In [1]: no_cols = DataFrame(index=['a', 'b', 'c'])

In [2]: no_cols.apply(lambda x: x.mean()) 
Out[2]: Series([], dtype: float64)

[jreback]

@hayd can you revisit this....and see if you can fix the isues here? thxs

[hayd]

@jreback yeah, will have a look at the weekend, will at least write all the cases down (where I think there is ambiguity). IIRC there was an issue where fixing one thing that seemed obviously wrong broke another which had seemed obviously correct.

[jreback]

@hayd actually I am not so sure that this is wrong (nor does it really matter), and default dtype is np.float64, as the dtypes would be changed/inferred if stuff was added.....going to close as not a bug (if you disagree, can pls reopen and move to 0.14)