PERF: Avoid values in Categorical.set_categories (#17515)

Mater:

```python
In [1]: import pandas as pd; import numpy as np

In [2]: arr = ['s%04d' % i for i in np.random.randint(0, 500000 // 10,
                                                      size=500000)];
s = pd.Series(arr).astype('category')

In [3]: %timeit s.cat.set_categories(s.cat.categories)
68.5 ms ± 846 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
```

HEAD:

```python
In [1]: import pandas as pd; import numpy as np

In [2]: arr = ['s%04d' % i for i in np.random.randint(0, 500000 // 10,
                                                      size=500000)]
s = pd.Series(arr).astype('category')

In [3]: %timeit s.cat.set_categories(s.cat.categories)
7.43 ms ± 110 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
```

Closes #17508

Author: TomAugspurger

asv_bench/benchmarks/categoricals.py

@@ -67,6 +67,9 @@ def time_value_counts_dropna(self):
     def time_rendering(self):
         str(self.sel)
 
+    def time_set_categories(self):
+        self.ts.cat.set_categories(self.ts.cat.categories[::2])
+
 
 class Categoricals3(object):
     goal_time = 0.2

doc/source/whatsnew/v0.21.0.txt

@@ -467,6 +467,7 @@ Performance Improvements
 
 - Improved performance of instantiating :class:`SparseDataFrame` (:issue:`16773`)
 - :attr:`Series.dt` no longer performs frequency inference, yielding a large speedup when accessing the attribute (:issue:`17210`)
+- Improved performance of :meth:`Categorical.set_categories` by not materializing the values (:issue:`17508`)
 - :attr:`Timestamp.microsecond` no longer re-computes on attribute access (:issue:`17331`)
 
 .. _whatsnew_0210.bug_fixes:

pandas/core/categorical.py

@@ -777,8 +777,9 @@ def set_categories(self, new_categories, ordered=None, rename=False,
                 # remove all _codes which are larger and set to -1/NaN
                 self._codes[self._codes >= len(new_categories)] = -1
         else:
-            values = cat.__array__()
-            cat._codes = _get_codes_for_values(values, new_categories)
+            codes = _recode_for_categories(self.codes, self.categories,
+                                           new_categories)
+            cat._codes = codes
         cat._categories = new_categories
 
         if ordered is None:
@@ -2113,6 +2114,38 @@ def _get_codes_for_values(values, categories):
     return coerce_indexer_dtype(t.lookup(vals), cats)
 
 
+def _recode_for_categories(codes, old_categories, new_categories):
+    """
+    Convert a set of codes for to a new set of categories
+
+    Parameters
+    ----------
+    codes : array
+    old_categories, new_categories : Index
+
+    Returns
+    -------
+    new_codes : array
+
+    Examples
+    --------
+    >>> old_cat = pd.Index(['b', 'a', 'c'])
+    >>> new_cat = pd.Index(['a', 'b'])
+    >>> codes = np.array([0, 1, 1, 2])
+    >>> _recode_for_categories(codes, old_cat, new_cat)
+    array([ 1,  0,  0, -1])
+    """
+    from pandas.core.algorithms import take_1d
+
+    if len(old_categories) == 0:
+        # All null anyway, so just retain the nulls
+        return codes
+    indexer = coerce_indexer_dtype(new_categories.get_indexer(old_categories),
+                                   new_categories)
+    new_codes = take_1d(indexer, codes.copy(), fill_value=-1)
+    return new_codes
+
+
 def _convert_to_list_like(list_like):
     if hasattr(list_like, "dtype"):
         return list_like

pandas/core/dtypes/concat.py

@@ -314,6 +314,7 @@ def union_categoricals(to_union, sort_categories=False, ignore_order=False):
     Categories (3, object): [b, c, a]
     """
     from pandas import Index, Categorical, CategoricalIndex, Series
+    from pandas.core.categorical import _recode_for_categories
 
     if len(to_union) == 0:
         raise ValueError('No Categoricals to union')
@@ -359,14 +360,8 @@ def _maybe_unwrap(x):
 
         new_codes = []
         for c in to_union:
-            if len(c.categories) > 0:
-                indexer = categories.get_indexer(c.categories)
-
-                from pandas.core.algorithms import take_1d
-                new_codes.append(take_1d(indexer, c.codes, fill_value=-1))
-            else:
-                # must be all NaN
-                new_codes.append(c.codes)
+            new_codes.append(_recode_for_categories(c.codes, c.categories,
+                                                    categories))
         new_codes = np.concatenate(new_codes)
     else:
         # ordered - to show a proper error message

pandas/tests/test_categorical.py

@@ -26,6 +26,7 @@
                     Interval, IntervalIndex)
 from pandas.compat import range, lrange, u, PY3, PYPY
 from pandas.core.config import option_context
+from pandas.core.categorical import _recode_for_categories
 
 
 class TestCategorical(object):
@@ -963,6 +964,67 @@ def test_rename_categories(self):
         with pytest.raises(ValueError):
             cat.rename_categories([1, 2])
 
+    @pytest.mark.parametrize('codes, old, new, expected', [
+        ([0, 1], ['a', 'b'], ['a', 'b'], [0, 1]),
+        ([0, 1], ['b', 'a'], ['b', 'a'], [0, 1]),
+        ([0, 1], ['a', 'b'], ['b', 'a'], [1, 0]),
+        ([0, 1], ['b', 'a'], ['a', 'b'], [1, 0]),
+        ([0, 1, 0, 1], ['a', 'b'], ['a', 'b', 'c'], [0, 1, 0, 1]),
+        ([0, 1, 2, 2], ['a', 'b', 'c'], ['a', 'b'], [0, 1, -1, -1]),
+        ([0, 1, -1], ['a', 'b', 'c'], ['a', 'b', 'c'], [0, 1, -1]),
+        ([0, 1, -1], ['a', 'b', 'c'], ['b'], [-1, 0, -1]),
+        ([0, 1, -1], ['a', 'b', 'c'], ['d'], [-1, -1, -1]),
+        ([0, 1, -1], ['a', 'b', 'c'], [], [-1, -1, -1]),
+        ([-1, -1], [], ['a', 'b'], [-1, -1]),
+        ([1, 0], ['b', 'a'], ['a', 'b'], [0, 1]),
+    ])
+    def test_recode_to_categories(self, codes, old, new, expected):
+        codes = np.asanyarray(codes, dtype=np.int8)
+        expected = np.asanyarray(expected, dtype=np.int8)
+        old = Index(old)
+        new = Index(new)
+        result = _recode_for_categories(codes, old, new)
+        tm.assert_numpy_array_equal(result, expected)
+
+    def test_recode_to_categories_large(self):
+        N = 1000
+        codes = np.arange(N)
+        old = Index(codes)
+        expected = np.arange(N - 1, -1, -1, dtype=np.int16)
+        new = Index(expected)
+        result = _recode_for_categories(codes, old, new)
+        tm.assert_numpy_array_equal(result, expected)
+
+    @pytest.mark.parametrize('values, categories, new_categories', [
+        # No NaNs, same cats, same order
+        (['a', 'b', 'a'], ['a', 'b'], ['a', 'b'],),
+        # No NaNs, same cats, different order
+        (['a', 'b', 'a'], ['a', 'b'], ['b', 'a'],),
+        # Same, unsorted
+        (['b', 'a', 'a'], ['a', 'b'], ['a', 'b'],),
+        # No NaNs, same cats, different order
+        (['b', 'a', 'a'], ['a', 'b'], ['b', 'a'],),
+        # NaNs
+        (['a', 'b', 'c'], ['a', 'b'], ['a', 'b']),
+        (['a', 'b', 'c'], ['a', 'b'], ['b', 'a']),
+        (['b', 'a', 'c'], ['a', 'b'], ['a', 'b']),
+        (['b', 'a', 'c'], ['a', 'b'], ['a', 'b']),
+        # Introduce NaNs
+        (['a', 'b', 'c'], ['a', 'b'], ['a']),
+        (['a', 'b', 'c'], ['a', 'b'], ['b']),
+        (['b', 'a', 'c'], ['a', 'b'], ['a']),
+        (['b', 'a', 'c'], ['a', 'b'], ['a']),
+        # No overlap
+        (['a', 'b', 'c'], ['a', 'b'], ['d', 'e']),
+    ])
+    @pytest.mark.parametrize('ordered', [True, False])
+    def test_set_categories_many(self, values, categories, new_categories,
+                                 ordered):
+        c = Categorical(values, categories)
+        expected = Categorical(values, new_categories, ordered)
+        result = c.set_categories(new_categories, ordered=ordered)
+        tm.assert_categorical_equal(result, expected)
+
     def test_reorder_categories(self):
         cat = Categorical(["a", "b", "c", "a"], ordered=True)
         old = cat.copy()