BUG: setting new items with .loc doesn't work for constant sequences

[code-of-kpp]

One can construct DataFrame with frozensets in the index:

 >>> import pandas
 >>> a = frozenset([1])
 >>> b = frozenset([1, 2])
 >>> c = frozenset([1, 2, 3])
 >>> 
 >>> frame = pandas.DataFrame(index=[a, b], columns=[c])
 >>> frame
        (1, 2, 3)
 (1)          NaN
 (1, 2)       NaN

Setting new values for existing items works:

 >>> frame.loc[a] = 0
 >>> frame
        (1, 2, 3)
 (1)            0
 (1, 2)       NaN
 >>> frame[c] = 1
 >>> frame
         (1, 2, 3)
 (1)             1
 (1, 2)          1
 >>> frame[frozenset([4])] = 1
 >>> frame
         (1, 2, 3)  (4)
 (1)             1    1
 (1, 2)          1    1

But new items cannot be created:

 >>> frame.loc[frozenset([3])] = 1
 Traceback (most recent call last):
   File "<stdin>", line 1, in <module>
   File "venv/lib/python2.7/site-packages/pandas/core/indexing.py", line 118, in __setitem__
     indexer = self._convert_to_indexer(key, is_setter=True)
   File "venv/lib/python2.7/site-packages/pandas/core/indexing.py", line 1085, in _convert_to_indexer
     raise KeyError('%s not in index' % objarr[mask])
 KeyError: '[3] not in index'

[jreback]

what are you using this for? a frozenset is really an iterable, yes? (but doesn't have getitem), so not really sure what you would do with this. pandas has a lot of checks for iterable types (and list/tuple.ndarray, etc) in the indexing routines.

[code-of-kpp]

I have a very long sequence and a function that produces a dict with several keys and values (features) for samples from that sequence.

[1, 2, 3 .... 100000000000]

func([1, 2, 3]) == {'a': 0.5, 'b': 0.6}
func([1, 3, 2]) == {'a': 0.5, 'b': 0.6}  # last two are equal

func([1, 3, 5, 9]) == {'a': 0.5, 'b': 0.7}
...

I wanted to make a DataFrame with 100 samples in rows and two columns ('a', 'b').

[jreback]

you can do with with tuples (as a value in an index).

[code-of-kpp]

I have to consider (1, 2, 3), (1, 1, 2, 3) and (1, 3, 2) equal.

I solved this with coding (marshal.dumps and marshal.loads).

I'm just pointing out at inconsistency in API.

[code-of-kpp]

Tuples have a special meaning in .loc

[jreback]

you would need a custom index type for this then and maybe some modifications to the indexing code. If you'd like to submit a pull-request, would be fine.

[jreback]

The API for indexing is not clear that this is supported at all (in fact, IIRC, only: int/label/tuple/list-like/tuple/slice are supported)

[code-of-kpp]

I think a warning should be rised when someone uses a value of another type in the index.

[jreback]

If you can do it w/o sacrificing speed and appropriate test, then we would take it. frozenset as I said is a hybrid list-like, not very friendly to the indexing code.

[jreback]

this is a very thorny problem and basically breaks indexing.

[mroeschke]

I believe we are moving away from making the indexing code more permissive to indexing types such as frozenset. Going to close as a wont fix but happy to reopen if there's more interest from the community