.iloc with 1-element integer array behaves badly

[hmeine]

There is an inconsistency in the returned (python) types when indexing .iloc with an integer array. The result should always be a sequence (Series), but if the array contains exactly one index, it returns a scalar value:

>>> import pandas, numpy
>>> column = pandas.Series(numpy.arange(10))
>>> indices, = (column > 4).nonzero()
>>> column.iloc[indices]
5    5
6    6
7    7
8    8
9    9
dtype: int64
>>> indices, = (column == 5).nonzero()
>>> column.iloc[indices]
5
>>> indices, = (column == 99).nonzero()
>>> column.iloc[indices]
Series([], dtype: int64)

I am working around that using numpy.atleast_1d, but it's ugly.

[jreback]

looks like a buglet, not treating the np.ndarray input the same as a list, marking as a bug

In [21]: indices
Out[21]: array([5])

In [22]: type(indices)
Out[22]: numpy.ndarray

In [23]: column.iloc[list(indices)]
Out[23]: 
5    5
dtype: int64

fyi....this would be the preferred usage of what you are doing in any event

In [26]: column.loc[column==5]
Out[26]: 
5    5
dtype: int64

[hmeine]

Thanks, but I have different use cases: I am often using .diff().nonzero() in order to find indices where relevant parameters change (e.g. an error code, depending on some input parameters). Otherwise, I am using boolean selections where possible.

[jreback]

np.....pretty trivial fix, want to do a PR?

[hmeine]

I might have a look later (Monday), but I have to hurry and leave in < 5 minutes. ;-)

[jreback]

gr8.....will leave open for you!

[jreback]

@hmeine doing a PR for this?

[hmeine]

Did not forget about it, but did not yet get around to it, either. Is pandas/tests/test_indexing.py the right place for a regression test? Here's a holiday today, so I have four family days. Can't say yet whether that makes it harder or easier to find time for hacking. ;-)

[jreback]

test_indexing is good for this