Support for np.unique #2902
Support for np.unique #2902
Conversation
Codecov Report
@@ Coverage Diff @@
## master #2902 +/- ##
==========================================
- Coverage 85.84% 85.84% -0.01%
==========================================
Files 326 326
Lines 68309 68329 +20
Branches 7720 7722 +2
==========================================
+ Hits 58640 58654 +14
- Misses 8428 8434 +6
Partials 1241 1241 |
|
Great to see support for more numpy functions in numba! |
|
Thanks for the feedback. I'm pretty sure we should be able to determine the size of the array and branch the implementation based on that, but I don't know how to check for an empty array. |
|
Turns out it's not possible to differentiate on emptiness. Current implementation might replace [x for x in b[:1]] by simply list(b[:1]) (else we add an array to a list), but a cast has to happen somewhere. |
numba/targets/arrayobj.py
Outdated
| @overload(np.unique) | ||
| def np_unique(a): | ||
| def np_unique_impl(a): | ||
| b = np.sort(a.flatten()) |
There was a problem hiding this comment.
Performance nitpick: use .ravel() instead of .flatten() so you can save one copying.
numba/targets/arrayobj.py
Outdated
| def np_unique(a): | ||
| def np_unique_impl(a): | ||
| b = np.sort(a.flatten()) | ||
| return np.array([x for x in b[:1]] + [x for i, x in enumerate(b[1:]) if b[i] != x]) |
There was a problem hiding this comment.
pep8 nitpick: line more than 79col
You can check the size of the array by |
closes #2884
Fairly minimal, just the single array argument is supported,, not any of the return options or the axis argument.