Repository navigation
poor categorical support w/ numpy<1.8 #7988
Description
Numpy versions < 1.8 play poorly with arrays of mixed type, truncating anything not of type string to the length of the longest string. This is inhibiting #7974
So the issue is as follows
>>> import numpy as np
>>> np.__version__
'1.7.1'
>>> np.array(['A', 'A', np.nan, 'B', -np.inf, 3.14], dtype=np.unicode)
array([u'A', u'A', u'n', u'B', u'-', u'3'],
dtype='<U1')two possible solutions:
- Fold back in shims removed in Remove numpy 1.6 specific work-arounds #7484 that existed mainly to solve this problem
- throw a warning if category.py is imported with a numpy<1.8 and skip tests for numpy<1.8
The advantage of 2 is it keeps the code cleaner and this is an edge case of sorts because it's only an issue when the array is of mixed type (strings and numbers). The advantage of 1 is of course edge case support on older systems.
Activity
efiring commentedon Jan 30, 2017
I think option 1 is preferable--less likely to trigger questions from users, and simpler in that it is localized to a single module. Comments could be added to make it very simple to strip out the shims when we no longer support numpy 1.7.
QuLogic commentedon Jan 31, 2017
I think what we can do is add a skip to #7974 so that testing starts working as it should be, and then @story645 can open a PR restoring the shims which will properly be tested once testing works.
tacaswell commentedon Jan 31, 2017
👍 to @QuLogic 's plan
closes matplotlib#7988
closes matplotlib#7988