fix memory leak in autoregression

bnpm/automatic_regression.py

@@ -293,7 +293,7 @@ def _objective(self, trial: optuna.trial.Trial) -> float:
         if len(loss_train_all) == 1:
             loss_train, loss_test, loss = loss_train_all[0], loss_test_all[0], loss_all[0]        
         else:
-            if isinstance(loss_train_all[0], (np.ndarray, np.generic)):
+            if isinstance(loss_train_all[0], (np.ndarray, np.generic, float, int)):
                 stack = np.stack
             elif isinstance(loss_train_all[0], torch.Tensor):
                 stack = torch.stack
@@ -686,8 +686,8 @@ def __call__(
         elif isinstance(y_train_pred, torch.Tensor):
             from torch.nn.functional import mse_loss
             assert (sample_weight_test is None) and (sample_weight_train is None), 'sample weights not supported for torch tensors.'
-            loss_train = mse_loss(y_train_pred, y_train_true, reduction='mean')
-            loss_test =  mse_loss(y_test_pred,  y_test_true,  reduction='mean')
+            loss_train = mse_loss(y_train_pred, y_train_true, reduction='mean').item()
+            loss_test =  mse_loss(y_test_pred,  y_test_true,  reduction='mean').item()
             loss = self.fn_penalty_testTrainRatio(loss_test, loss_train)
         else:
             raise ValueError(f'Expected y_train_pred to be of type np.ndarray or torch.Tensor, but got type {type(y_train_pred)}.')