finalize

src/scib_metrics/nearest_neighbors/_dataclass.py

@@ -28,11 +28,13 @@ def __post_init__(self):
         chex.assert_equal_shape([self.indices, self.distances])
 
     @property
-    def n_samples(self):
+    def n_samples(self) -> np.ndarray:
+        """Number of samples (cells)."""
         return self.indices.shape[0]
 
     @property
-    def n_neighbors(self):
+    def n_neighbors(self) -> np.ndarray:
+        """Number of neighbors."""
         return self.indices.shape[1]
 
     @cached_property
@@ -65,6 +67,7 @@ def knn_graph_connectivities(self) -> coo_matrix:
         return connectivities[0]
 
     def subset_neighbors(self, n: int) -> "NeighborsResults":
+        """Subset down to `n` neighbors."""
         if n > self.n_neighbors:
             raise ValueError("n must be smaller than the number of neighbors")
         return self.__class__(indices=self.indices[:, :n], distances=self.distances[:, :n])

tests/test_metrics.py

@@ -74,7 +74,7 @@ def test_lisi_knn():
         X, pd.DataFrame(labels, columns=["labels"]), label_colnames=["labels"], perplexity=10
     )[:, 0]
     # Slight numerical differences arise due to how self edges are handled. With approximate nearest
-    # neighbors methods, there is no guarantee that the self edge is the first edge. To accomodate this,
+    # neighbors methods, there is no guarantee that the self edge is the first edge. To accommodate this,
     # we mask out self edges internally in lisi computation which causes slight numerical differences.
     np.testing.assert_allclose(lisi_res, harmonypy_lisi_res, rtol=5e-5, atol=5e-5)