Repository navigation
can't index string- and unicode- numpy-scalars with an empty tuple #7267
Description
As discovered in #7259 and #7260, string- and unicode- numpy scalars do not support indexing with an empty tuple, unlike other scalars:
>>> numpy.string_("ABC")[()]
TypeError: string indices must be integers, not tuple
>>> numpy.int32(1)[()]
1
This may be a problem because indexing with an empty tuple is a way of converting possible 0-d arrays to scalars:
>>> a = numpy.int64(1)
>>> b = numpy.array(a)
>>> type(a[()]), type(b[()])
(numpy.int64, numpy.int64)
As of writing this it's not 100% clear to me we actually want this. It could be that instead of just writing
return possiblearr[()]
we should encourage people to write something like
if isinstance(possiblearr, ndarray):
return possiblearr[()]
return possiblearr
But, on the other hand it's probably a good idea for all scalar types to behave similarly. I made a rough outline for a fix in a comment in #7260.
Activity
eric-wieser commentedon Mar 6, 2017
Related:
numpy.string_("ABC")[...]fails in the same waycharris commentedon Mar 6, 2017
OTOH,
eric-wieser commentedon Mar 6, 2017
This manifests itself in
np.ma.array(['ab', 'cd']).take(1)due to the patch in #7586eric-wieser commentedon Jul 3, 2017
Arguably not a bug - we're never going to have this work on object scalars, so having it fail on string scalars too forces the library writer to use a solution that works for object scalars too.
InessaPawson commentedon Sep 18, 2026
Closing this older issue in favor of #31435 for the discussion and implementation to stay in one place.