Speed up categorical regressor with numba

docs/release-notes/3353.performance.md

@@ -0,0 +1 @@
+* Speed up for a categorical regressor in {func}`~scanpy.pp.regress_out` {smaller}`S Dicks`

src/scanpy/preprocessing/_simple.py

@@ -637,6 +637,22 @@
 DT = TypeVar("DT")
 
 
+@njit
+def _create_regressor_categorical(
+    X: np.ndarray, number_categories: int, cat_array: np.ndarray
+) -> np.ndarray:
+    # create regressor matrix for categorical variables
+    regressors = np.zeros(X.shape, dtype=X.dtype)
+    # iterate over categories
+    for category in range(number_categories):
+        # iterate over genes and calculate mean expression
+        # for each gene per category
+        mask = category == cat_array
+        for ix in numba.prange(X.T.shape[0]):
+            regressors[mask, ix] = X.T[ix, mask].mean()
+    return regressors
+
+
 @njit
 def get_resid(
     data: np.ndarray,
@@ -732,13 +748,13 @@
             )
             raise ValueError(msg)
         logg.debug("... regressing on per-gene means within categories")
-        regressors = np.zeros(X.shape, dtype="float32")
+
+        # set number of categories to the same dtype as the categories
+        cat_array = adata.obs[keys[0]].cat.codes.to_numpy()
+        number_categories = cat_array.dtype.type(len(adata.obs[keys[0]].cat.categories))
+
         X = _to_dense(X, order="F") if issparse(X) else X
-        # TODO figure out if we should use a numba kernel for this
-        for category in adata.obs[keys[0]].cat.categories:
-            mask = (category == adata.obs[keys[0]]).values
-            for ix, x in enumerate(X.T):
-                regressors[mask, ix] = x[mask].mean()
+        regressors = _create_regressor_categorical(X, number_categories, cat_array)
         variable_is_categorical = True
     # regress on one or several ordinal variables
     else:

tests/test_preprocessing.py

@@ -464,14 +464,21 @@ def test_regress_out_constants():
     assert_equal(adata, adata_copy)
 
 
-def test_regress_out_reproducible():
-    adata = pbmc68k_reduced()
+@pytest.mark.parametrize(
+    ("keys", "test_file"),
+    [
+        (["n_counts", "percent_mito"], "regress_test_small.npy"),
+        (["bulk_labels"], "regress_test_small_cat.npy"),
+    ],
+)
+def test_regress_out_reproducible(keys, test_file):
+    adata = sc.datasets.pbmc68k_reduced()
     adata = adata.raw.to_adata()[:200, :200].copy()
-    sc.pp.regress_out(adata, keys=["n_counts", "percent_mito"])
+    sc.pp.regress_out(adata, keys=keys)
     # This file was generated from the original implementation in version 1.10.3
     # Now we compare new implementation with the old one
-    tester = np.load(DATA_PATH / "regress_test_small.npy")
-    np.testing.assert_allclose(adata.X, tester)
+    tester = np.load(DATA_PATH / test_file)
+    np.testing.assert_array_almost_equal(adata.X, tester)
 
 
 def test_regress_out_constants_equivalent():