BUG: Fix packbits to correctly handle empty arrays #8327
Conversation
|
please squash the test and the fix into one commit, it is better to have the testsuite succeed on every commit for easier bisection. |
330c4f9 to
41d10b5
Compare
|
Hi @juliantaylor, thank you for your quick feedback. I squashed the two commits into one commit. |
| @@ -17,6 +17,22 @@ def test_packbits(): | |||
| assert_raises(TypeError, np.packbits, np.array(a, dtype=float)) | |||
|
|
|||
|
|
|||
| def test_packbits_empty(): | |||
| for dtype in [np.bool, np.uint8, np.int]: | |||
There was a problem hiding this comment.
Why just these three types? And why just the last dimension zero? Tests should be exhaustive. Admittedly, most aren't but, strangely enough, when improved often reveal errors.
All integer types are available in np.typecodes['AllInteger'], the boolean type is separate. Or you could just use the string '?bBhHiIlLqQ' for typecodes. If boolean is acceptable I think the docstring need improvement also.
There was a problem hiding this comment.
Also not a good idea to shadow the numpy dtype variable. Use something like dt instead.
There was a problem hiding this comment.
Is it checked anywhere that empty results unpack correctly?
41d10b5 to
fe7448f
Compare
|
Hi @charris, thank you for the detailed and accurate feedback!
According to your comments, I improved the tests in
I updated the statement in
You are correct --- I added tests for unpack against empty arrays, and indeed, I found a similar bug in unpack. So, I also updated function |
| assert_array_equal(b, np.empty((0,))) | ||
|
|
||
|
|
||
| def test_packbits_empty_with_axis(): |
There was a problem hiding this comment.
this function has the wrong (duplicate) name, so only one of them runs
fe7448f to
47c66c9
Compare
|
Sorry for stupid mistake. fixed. |
|
thanks, very nice pull request. |
This small PR addresses several problems of
numpy.packbitspointed out in Issue #8324.As discussed in Issue #8324,
numpy.packbitshave problems with regard to empty input arrays. Specifically:uint8dtype,axisis specified, andnew).The fundamental reason for these problems is common: function
pack_bitsincompiled_base.chandles empty arrays separately, in a rough manner. I could not figure out a good reason why we should handle empty arrays separately. Rather, I found that this part is indeed unnecessary, i.e., without this,pack_bitscan handle empty arrays correctly, as confirmed in my new tests. So, I propose to remove that part to fix these problems.