diff --git a/river_dl/evaluate.py b/river_dl/evaluate.py
index 76a858d..06195d4 100755
--- a/river_dl/evaluate.py
+++ b/river_dl/evaluate.py
@@ -201,7 +201,7 @@ def partition_metrics(
         outfile=None,
         val_sites=None,
         test_sites=None,
-
+        train_sites=None,
 ):
     """
     calculate metrics for a certain group (or no group at all) for a given
@@ -222,8 +222,9 @@ def partition_metrics(
     names and dict values are the id values. These are added as columns to the
     metrics information
     :param outfile: [str] file where the metrics should be written
-    :param val_sites: [list] sites to exclude from training metrics
+    :param val_sites: [list] sites to exclude from training and test metrics
     :param test_sites: [list] sites to exclude from validation and training metrics
+    :param train_sites: [list] sites to exclude from test metrics
     :return: [pd dataframe] the condensed metrics
     """
     var_data = fmt_preds_obs(preds, obs_file, spatial_idx_name,
@@ -240,6 +241,10 @@ def partition_metrics(
         # mask out test sites from val partition
         if test_sites and partition=='val':
             data = data[~data[spatial_idx_name].isin(test_sites)]
+        if train_sites and partition=='tst':
+            data = data[~data[spatial_idx_name].isin(train_sites)]
+        if val_sites and partition=='tst':
+            data = data[~data[spatial_idx_name].isin(val_sites)]
 
         if not group:
             metrics = calc_metrics(data)
@@ -286,6 +291,7 @@ def combined_metrics(
     pred_tst=None,
     val_sites=None,
     test_sites=None,
+    train_sites=None,
     spatial_idx_name="seg_id_nat",
     time_idx_name="date",
     group=None,
@@ -349,7 +355,8 @@ def combined_metrics(
                                     id_dict=id_dict,
                                     group=group,
                                     val_sites = val_sites,
-                                    test_sites = test_sites)
+                                    test_sites = test_sites,
+                                    train_sites=train_sites)
         df_all.extend([metrics])
 
     df_all = pd.concat(df_all, axis=0)
diff --git a/river_dl/torch_utils.py b/river_dl/torch_utils.py
index 1751f40..05a7d06 100644
--- a/river_dl/torch_utils.py
+++ b/river_dl/torch_utils.py
@@ -238,19 +238,20 @@ def predict_torch(x_data, model, batch_size):
     @param device: [str] cuda or cpu
     @return: [tensor] predicted values
     """
-    device = next(model.parameters()).device
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    
+    model.to(device)
     data = []
     for i in range(len(x_data)):
         data.append(torch.from_numpy(x_data[i]).float())
 
     dataloader = torch.utils.data.DataLoader(data, batch_size=batch_size, shuffle=False, pin_memory=True)
-    model.to(device)
     model.eval()
     predicted = []
     for iter, x in enumerate(dataloader):
         trainx = x.to(device)
         with torch.no_grad():
-            output = model(trainx.to(device)).cpu()
+            output = model(trainx).detach().cpu()
         predicted.append(output)
     predicted = torch.cat(predicted, dim=0)
     return predicted