BUG: Fix object scalar return type of 0-d array indices #4109
Conversation
| return PyArray_Return((PyArrayObject *)ret); | ||
| } | ||
| return ret; | ||
| } |
There was a problem hiding this comment.
Actually the only thing that block does is probably should be to more or less correctly support arr['0', '0'] or broken types, and my new stuff doesn't have that yet... (something like this was probably historically always there)
BUG: Fix object scalar return type of 0-d array indices http://stackoverflow.com/questions/20409334/subclassed-empty-numpy-recarray-losing-its-type-and-added-attributes-in-numpy-1
| return array_subscript_fancy(self, op, fancy); | ||
| } | ||
| PyObject *ret; | ||
| 8000 ret = array_subscript_fancy(self, op, fancy); |
There was a problem hiding this comment.
This line might look better if it were swapped with the following blank line.
|
Looks good modulo style nitpicks. I'm relying on the tests to check that the code does what is claimed ;) |
PyArray_Result was called when it should not have been. This caused object arrays of 0-d arrays have the inner array converted to a scalar: `np.array([np.array(0), None])[0] -> np._float(0)` The same case happened to recarray field access when indexing more than one field. This fixes it.
|
Good, luckily, all cases which should not result in an array go normally through their own special case, so I don't think I can break anything ;). Fixed the style nitpicks, but want confirmation that fixing |
|
Fixed the style nitpicks, but want confirmation that fixing a[['a']] to be identical to a['a'] is good. These distinctions make my head hurt ;) I'm not clear on what the intent of |
BUG: Fix object scalar return type of 0-d array indices
|
I admit, master currently doesn't, I opened a 1.9. blocker about it though. My indexing stuff does mostly (except for the one corner case at this time). |
|
this fix breaks pandas 0.13, you get lots of test failures of these types: the fix is also in the 1.8.x branch so pandas is also broken there whih must be accounted for should a 1.8.1 be made |
|
as pointed out in the pandas issue the maintenance branch is broken by this: surprising we don't have a test for this. @seberg do you have time too look at this or should we just revert it for now? |
|
Oh, I think I will fix it soon. Just need to look through the subtleties and best check what 1.7 did. A mostly fix is simple in any case and quite right may be impossible, but it should match 1.7. if possible... |
|
@juliantaylor, see gh-4223. |
PyArray_Result was called when it should not have been. This caused
object arrays of 0-d arrays have the inner array converted to a
scalar:
np.array([np.array(0), None])[0] -> np._float(0)The same case happened to recarray field access. This fixes it.
(This was an SO question, maybe it will be an issue here, too... It is a regression in 1.8.x). There are some code blocks there I am sure are not used, but I don't want to touch the old mapping.c more then absolutely necessary... My current plan is, that when this is merged I create a blocker issue for 1.9.x. My indexing branch has it fixed anyway, I know that is not the right procedure, so if someone disagrees...