Indexing TimedeltaIndex with np.timedelta64

[shoyer]

On pandas 0.20.2, TimedeltaIndex.get_loc() worked on np.timedelta64, even though DataFrame.loc fails:

In [2]: df = pd.DataFrame({'x': range(3)}, pd.to_timedelta(range(3), unit='days'))

In [3]: df.loc[df.index[0].to_timedelta64()]
TypeError: cannot do index indexing on <class 'pandas.core.indexes.timedeltas.TimedeltaIndex'> with these indexers [0] of <class 'numpy.timedelta64'>

In [5]: df.index.get_loc(df.index[0].to_timedelta64())
Out[5]: 0

Now, both raise TypeError. (This is similar to the issue in #16896)

I think some xarray users are were likely relying on this (we use get_loc() directly), but I think supporting this could also be useful for pandas users.

[jdeschenes]

Question: Should get_loc only work on 'scalar' values?

if that's the case, then it's pretty easy to fix:

replace is_timedelta64_dtype with is_list_like and it should work.

[jreback]

.get_loc by-definition is only on scalar values. a string and/or a np.timedelta64 are scalars (e.g. not lists/arrays)

.get_indexer works on list-likes

[jreback]

here you can simply pass scalars to Timedelta() and then index