diff --git a/models/__init__.py b/models/__init__.py
index fd917c06..4cff8d30 100644
--- a/models/__init__.py
+++ b/models/__init__.py
@@ -53,6 +53,10 @@ def setup(opt, reverse=False):
         # check if all necessary files exist 
         assert os.path.isdir(opt.start_from)," %s must be a a path" % opt.start_from
         assert os.path.isfile(os.path.join(opt.start_from,"infos_"+opt.id+".pkl")),"infos.pkl file does not exist in path %s"%opt.start_from
-        model.load_state_dict(torch.load(os.path.join(opt.start_from, 'model.pth')))
+        forward_dict, backward_dict = torch.load(os.path.join(opt.start_from, 'model.pth'))
+        if reverse:
+            model.load_state_dict(backward_dict)
+        else:
+            model.load_state_dict(forward_dict)
 
     return model
diff --git a/opts.py b/opts.py
index a2addcdc..c3a6409d 100644
--- a/opts.py
+++ b/opts.py
@@ -100,6 +100,7 @@ def parse_opt():
                     help='an id identifying this run/job. used in cross-val and appended when writing progress files')
     parser.add_argument('--train_only', type=int, default=0,
                     help='if true then use 80k, else use 110k')
+    parser.add_argument('--param-l2', type=float, default=0.)
 
     args = parser.parse_args()
 
@@ -117,4 +118,4 @@ def parse_opt():
     assert args.load_best_score == 0 or args.load_best_score == 1, "language_eval should be 0 or 1"
     assert args.train_only == 0 or args.train_only == 1, "language_eval should be 0 or 1"
 
-    return args
\ No newline at end of file
+    return args
diff --git a/train_twinnet_para.py b/train_twinnet_para.py
index b9a25735..ffbb4ef2 100644
--- a/train_twinnet_para.py
+++ b/train_twinnet_para.py
@@ -45,7 +45,7 @@ def train(opt):
         with open(os.path.join(opt.start_from, 'infos_'+opt.id+'.pkl')) as f:
             infos = cPickle.load(f)
             saved_model_opt = infos['opt']
-            need_be_same=["caption_model", "rnn_type", "rnn_size", "num_layers"]
+            need_be_same = ["caption_model", "rnn_type", "rnn_size", "num_layers"]
             for checkme in need_be_same:
                 assert vars(saved_model_opt)[checkme] == vars(opt)[checkme], "Command line argument and saved model disagree on '%s' " % checkme
 
@@ -147,7 +147,9 @@ def train(opt):
         l2_loss = ((affine_states - invert_backstates)** 2).sum(dim=1)[:, 0, :]
         l2_loss = (l2_loss / masks[:, 1:-1].sum(dim=1).expand_as(l2_loss)).mean()
         
-        all_loss = loss + 3.0 * l2_loss + back_loss
+        if opt.param_l2 == 0:
+            back_loss = 0. * back_loss
+        all_loss = loss + opt.param_l2 * l2_loss + back_loss
         
         all_loss.backward()
         #back_loss.backward()
@@ -211,7 +213,7 @@ def train(opt):
                     best_val_score = current_score
                     best_flag = True
                 checkpoint_path = os.path.join(opt.checkpoint_path, 'model.pth')
-                torch.save(model.state_dict(), checkpoint_path)
+                torch.save([model.state_dict(), back_model.state_dict()], checkpoint_path)
                 print("model saved to {}".format(checkpoint_path))
                 optimizer_path = os.path.join(opt.checkpoint_path, 'optimizer.pth')
                 torch.save(optimizer.state_dict(), optimizer_path)