Reduce excess matrix manipulations in item-KNN

src/lenskit/knn/item.py

@@ -243,12 +243,14 @@ def __call__(self, query: QueryInput, items: ItemList) -> ItemList:
         assert isinstance(model, csr_array)
         model = model[:, ti_valid_nums]
         assert isinstance(model, csr_array)
-        # convert to CSC so we can count neighbors per target item.
-        model = model.tocsc()
 
-        # count neighborhood sizes
-        sizes = np.diff(model.indptr)
         # which neighborhoods are usable? (at least min neighbors)
+        m_ind = csr_array(
+            (np.ones(model.nnz, np.int32), model.indices, model.indptr), shape=model.shape
+        )
+        sizes = m_ind.sum(0)
+        del m_ind
+        assert isinstance(sizes, np.ndarray) and len(sizes) == model.shape[1]
         scorable = sizes >= self.config.min_nbrs
 
         # fast-path neighborhoods that fit within max neighbors
@@ -268,7 +270,7 @@ def __call__(self, query: QueryInput, items: ItemList) -> ItemList:
         # PyTorch, make a dense matrix (this is usually small enough to be
         # usable), and use the Torch topk function.
         slow_mat = model.T[~fast, :]
-        assert isinstance(slow_mat, csr_array)
+        # assert isinstance(slow_mat, csr_array)
         n_slow, _ = slow_mat.shape
         if n_slow:
             # mask for the slow items.

tests/models/test_knn_item_item.py

@@ -376,7 +376,7 @@ def test_ii_implicit_large(rng, ml_ratings):
 
     for user in users:
         recs = pipe.run("recommender", query=user, n=NRECS)
-        _log.info("user %s recs\n%s", user, recs)
+        _log.info("user %s recs\n%r", user, recs)
         assert isinstance(recs, ItemList)
         assert len(recs) == NRECS
         urates = ml_ratings[ml_ratings["user_id"] == user]