warn if not csr

scvi/data/_anndata.py

@@ -296,9 +296,9 @@ def _verify_and_correct_data_format(adata, data_registry):
     for k in keys:
         data = get_from_registry(adata, k)
         if isspmatrix(data) and (data.getformat() != "csr"):
-            logger.debug("{} is csc_matrix. Overwriting to csr_matrix.".format(k))
-            data = data.tocsr()
-            _set_data_in_registry(adata, data, k)
+            logger.warning(
+                "Training will be faster when sparse matrix is formatted as CSR. It is safe to cast before model initialization."
+            )
         elif isinstance(data, np.ndarray) and (data.flags["C_CONTIGUOUS"] is False):
             logger.debug(
                 "{} is not C_CONTIGUOUS. Overwriting to C_CONTIGUOUS.".format(k)

tests/dataset/test_anndata.py

@@ -18,7 +18,6 @@
 )
 from scvi import _CONSTANTS
 from scvi.data._anndata import get_from_registry
-from scipy.sparse import csc_matrix
 
 
 def test_transfer_anndata_setup():
@@ -124,28 +123,6 @@ def test_data_format():
         get_from_registry(adata, _CONSTANTS.PROTEIN_EXP_KEY),
     )
 
-    # if data is sparse, check that after setup_anndata, data is csr_matrix
-    adata = synthetic_iid(run_setup_anndata=False)
-    adata.X = csc_matrix(adata.X)
-    adata.obsm["protein_expression"] = csc_matrix(adata.obsm["protein_expression"])
-    old_x = adata.X
-    old_pro = adata.obsm["protein_expression"]
-    setup_anndata(adata, protein_expression_obsm_key="protein_expression")
-
-    assert adata.X.getformat() == "csr"
-    assert adata.obsm["protein_expression"].getformat() == "csr"
-
-    assert np.array_equal(old_x.toarray(), adata.X.toarray())
-    assert np.array_equal(old_pro.toarray(), adata.obsm["protein_expression"].toarray())
-
-    assert np.array_equal(
-        adata.X.toarray(), get_from_registry(adata, _CONSTANTS.X_KEY).toarray()
-    )
-    assert np.array_equal(
-        adata.obsm["protein_expression"].toarray(),
-        get_from_registry(adata, _CONSTANTS.PROTEIN_EXP_KEY).toarray(),
-    )
-
     # if obsm is dataframe, make it C_CONTIGUOUS if it isnt
     adata = synthetic_iid()
     pe = np.asfortranarray(adata.obsm["protein_expression"])