BUG: remove_unused_categories dtype coerces to int64

[sinhrks]

• tests added / passed
• passes git diff upstream/master | flake8 --diff
• whatsnew entry

c = pd.Categorical(['a', 'b'], categories=['a', 'b', 'c'])
c.codes
# array([0, 1], dtype=int8)

# NG, must be int8 dtype
c = c.remove_unused_categories()
c.codes
# array([0, 1])

It is because np.unique uses platform int for unique_inverse

np.unique(np.array([0, 3, 2, 3], dtype=np.int8), return_inverse=True)
(array([0, 2, 3], dtype=int8), array([0, 2, 1, 2]))

[jreback]

can't do this, either use pd.unique() or _coerce_indexer_dtype()

[jreback]

we could add an arg to pd.unique() to return the inverses, which we have (i think directly)

[sinhrks]

Misunderstood that .codes is always np.int8. changed to use_coerce_indexer_dtype().

[jreback]

thank you sir!