Fix pickling GPU models (#632)

GPU models weren't default initializing members like _knn appropriately when
being loaded from pickle appropriately. Fix and update the unittest to test
this out.

implicit/gpu/matrix_factorization_base.py

@@ -221,8 +221,15 @@ def __getstate__(self):
         }
 
     def __setstate__(self, state):
-        self.item_factors = implicit.gpu.Matrix(state["item_factors"])
-        self.user_factors = implicit.gpu.Matrix(state["user_factors"])
+        # default initialize members
+        self.__init__()
+        item_factors = state["item_factors"]
+        if item_factors is not None:
+            self.item_factors = implicit.gpu.Matrix(item_factors)
+
+        user_factors = state["user_factors"]
+        if user_factors is not None:
+            self.user_factors = implicit.gpu.Matrix(user_factors)
 
 
 def check_random_state(random_state):

tests/recommender_base_test.py

@@ -405,12 +405,26 @@ def test_rank_items_batch(self):
             self.assertEqual(set(current_ids), set(selected_items))
 
     def test_pickle(self):
-        item_users = get_checker_board(50)
+        user_items = get_checker_board(50)
         model = self._get_model()
-        model.fit(item_users, show_progress=False)
+        model.fit(user_items, show_progress=False)
 
         pickled = pickle.dumps(model)
-        pickle.loads(pickled)
+        reloaded = pickle.loads(pickled)
+
+        # make sure we can call methods on the reloaded index, and get the same results back
+        # (https://github.com/benfred/implicit/issues/631)
+        ids, _ = model.recommend(0, user_items[0])
+        reloaded_ids, _ = reloaded.recommend(0, user_items[0])
+        assert_array_equal(ids, reloaded_ids)
+
+        ids, _ = model.similar_items(0)
+        reloaded_ids, _ = reloaded.similar_items(0)
+        assert_array_equal(ids, reloaded_ids)
+
+    def test_pickle_unfitted_model(self):
+        model = self._get_model()
+        pickle.loads(pickle.dumps(model))
 
     def test_invalid_user_items(self):