Incorrect timedelta type coercion when doing in-place .loc expansion

[bmcfee]

Code Sample, a copy-pastable example if possible

In [1]: import pandas as pd

In [2]: d = pd.DataFrame(columns=list('abc'))

In [3]: d.loc[0] = dict(a=pd.to_timedelta(5, unit='s'),
                        b=pd.to_timedelta(72, unit='s'),
                        c='23')

In [4]: d
Out[4]: 
         a        b               c
0 00:00:05 00:01:12 00:00:00.000000

In [5]: d.c
Out[5]: 
0   00:00:00.000000
Name: c, dtype: timedelta64[ns]

Expected Output

         a        b  c
0 00:00:05 00:01:12 23

The c column should have type object and the value at loc[0] should be of type string.

Note that this does not happen when only one of the a or b are timedeltas:

In [9]: d2 = pd.DataFrame(columns=list('abc'))

In [10]: d2.loc[0] = dict(a=5, b=pd.to_timedelta(72, unit='s'), c='23')

In [11]: d2
Out[11]: 
     a        b   c
0  5.0 00:01:12  23

In [12]: d3 = pd.DataFrame(columns=list('abc'))

In [13]: d3.loc[0] = dict(a=pd.to_timedelta(5, unit='s'), b=72, c='23')

In [14]: d3
Out[14]: 
         a     b   c
0 00:00:05  72.0  23

In these two cases, the c column retains the correct type.

output of pd.show_versions()

In [6]: pd.show_versions()

INSTALLED VERSIONS
------------------
commit: None
python: 3.5.2.final.0
python-bits: 64
OS: Linux
OS-release: 4.4.0-31-generic
machine: x86_64
processor: x86_64
byteorder: little
LC_ALL: None
LANG: en_US.UTF-8

pandas: 0.18.1
nose: 1.3.7
pip: 8.1.2
setuptools: 20.1.1
Cython: 0.24.1
numpy: 1.11.1
scipy: 0.18.0
statsmodels: None
xarray: None
IPython: 5.0.0
sphinx: 1.4.1
patsy: None
dateutil: 2.5.2
pytz: 2016.4
blosc: None
bottleneck: None
tables: None
numexpr: 2.6.0
matplotlib: 1.5.2
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
lxml: None
bs4: None
html5lib: None
httplib2: None
apiclient: None
sqlalchemy: 1.0.9
pymysql: None
psycopg2: None
jinja2: 2.8
boto: None
pandas_datareader: None

[sinhrks]

Thanks. This looks to be caused by Series type inference is bit aggressive. The conversion is performed here.

• https://github.com/pydata/pandas/blob/master/pandas/core/indexing.py#L380

pd.Series(dict(a=pd.to_timedelta(5, unit='s'),
               b=pd.to_timedelta(72, unit='s'),
               c='23'))
# a          00:00:05
# b          00:01:12
# c   00:00:00.000000
# dtype: timedelta64[ns]

PR is appreciated!

[bmcfee]

Thanks for confirming. I'd be happy to pr, but after looking at the code, really have no idea where to begin. I think this would be better handled by someone already familiar with the internals of pandas.

[jbrockmendel]

This looks right on main. Could use a test (or check if #27303 added a relevant test)