fix int for shortcut (#3461) (#3640)

* fix int for shortcut

* add release note

* add order to astype

* (fix): sparse bug + tests

---------



(cherry picked from commit def4f23)

Co-authored-by: Severin Dicks <37635888+Intron7@users.noreply.github.com>
Co-authored-by: Philipp A. <flying-sheep@web.de>

Author: 3 people

docs/release-notes/3461.bugfix.md

@@ -0,0 +1 @@
+Fixes an error where `regress_out` would fail to work with `integer` types {smaller}`S Dicks`

src/scanpy/preprocessing/_simple.py

@@ -755,6 +755,14 @@ def regress_out(
     # if the regressors are not categorical and the matrix is not singular
     # use the shortcut numpy_regress_out
     if not variable_is_categorical and np.linalg.det(regressors.T @ regressors) != 0:
+        # Because we update `X` in `numpy_regress_out`, it needs to be floating point to match
+        # the incoming values.
+        if np.issubdtype(X.dtype, np.integer):
+            target_dtype = np.float32 if X.dtype.itemsize <= 4 else np.float64
+            kwargs = {}
+            if isinstance(X, np.ndarray):
+                kwargs["order"] = "C"
+            X = X.astype(target_dtype, **kwargs)
         X = _to_dense(X, order="C") if isinstance(X, CSBase) else X
         res = numpy_regress_out(X, regressors)
 

tests/test_preprocessing.py

@@ -402,13 +402,22 @@ def test_regress_out_ordinal():
     np.testing.assert_array_equal(single.X, multi.X)
 
 
-def test_regress_out_layer():
+@pytest.mark.parametrize("dtype", [np.int64, np.float64, np.int32])
+def test_regress_out_layer(dtype):
     from scipy.sparse import random
 
-    adata = AnnData(random(1000, 100, density=0.6, format="csr"))
+    adata = AnnData(
+        random(1000, 100, density=0.6, format="csr", dtype=np.uint16).astype(dtype)
+    )
     adata.obs["percent_mito"] = np.random.rand(adata.X.shape[0])
     adata.obs["n_counts"] = adata.X.sum(axis=1)
-    adata.layers["counts"] = adata.X.copy()
+    if dtype == np.float64:
+        dtype_cast = dtype
+    if dtype == np.int64:
+        dtype_cast = np.float64
+    if dtype == np.int32:
+        dtype_cast = np.float32
+    adata.layers["counts"] = adata.X.copy().astype(dtype_cast)
 
     single = sc.pp.regress_out(
         adata, keys=["n_counts", "percent_mito"], n_jobs=1, copy=True
@@ -419,7 +428,7 @@ def test_regress_out_layer():
         adata, layer="counts", keys=["n_counts", "percent_mito"], n_jobs=1, copy=True
     )
 
-    np.testing.assert_array_equal(single.X, layer.layers["counts"])
+    np.testing.assert_allclose(single.X, layer.layers["counts"])
 
 
 def test_regress_out_view():