diff --git a/deepcrf/main.py b/deepcrf/main.py
index 40b0a57..abeef88 100644
--- a/deepcrf/main.py
+++ b/deepcrf/main.py
@@ -304,27 +304,28 @@ def parse_to_tag_ids(sentences):
                  'n_vocab_add': n_vocab_add,
                  'use_cudnn': args['use_cudnn']}
 
-    deepcrf.util.write_vocab(save_name + '.model_attr', net_attrs)
     model_filename = args['model_filename'] if args['model_filename'] else None
 
-    if args['initial_links_filename']:
-        net_attrs['initial_model'] = BiLSTM_CNN_CRF(
-            **deepcrf.util.load_vocab(args['model_attr_filename']))
-        serializers.load_hdf5(model_filename, net_attrs['initial_model'])
-        fine_tuning_links = set(deepcrf.util.load_vocab(
-            args['initial_links_filename']))
-        model_links = set()
-
-        for link in net_attrs['initial_model'].namedlinks():
-            link = link[0].split('/')[1]
-            if link:
-                model_links.add(link)
-
-        for link in model_links - fine_tuning_links:
-            setattr(net_attrs['initial_model'], link, None)
-
-        deepcrf.util.write_vocab(save_name + '.fine-tuning_links',
-                                 list(model_links & fine_tuning_links))
+    if is_train:
+        deepcrf.util.write_vocab(save_name + '.model_attr', net_attrs)
+        if args['initial_links_filename']:
+            net_attrs['initial_model'] = BiLSTM_CNN_CRF(
+                **deepcrf.util.load_vocab(args['model_attr_filename']))
+            serializers.load_hdf5(model_filename, net_attrs['initial_model'])
+            fine_tuning_links = set(deepcrf.util.load_vocab(
+                args['initial_links_filename']))
+            model_links = set()
+
+            for link in net_attrs['initial_model'].namedlinks():
+                link = link[0].split('/')[1]
+                if link:
+                    model_links.add(link)
+
+            for link in model_links - fine_tuning_links:
+                setattr(net_attrs['initial_model'], link, None)
+
+            deepcrf.util.write_vocab(save_name + '.fine-tuning_links',
+                                     list(model_links & fine_tuning_links))
 
     net = BiLSTM_CNN_CRF(**net_attrs)
     my_cudnn(args['use_cudnn'])
@@ -398,7 +399,7 @@ def eval_loop(x_data, x_char_data, y_data, x_train_additionals=[]):
 
         return predict_lists, sum_loss, predicted_results
 
-    if model_filename is not None and not args['initial_links_filename']:
+    if model_filename is not None and (is_test or (is_train and not args['initial_links_filename'])):
         serializers.load_hdf5(model_filename, net)
 
     if is_test: