diff --git a/source/train/Trainer.py b/source/train/Trainer.py
index b7112370ae..0023e4ed12 100644
--- a/source/train/Trainer.py
+++ b/source/train/Trainer.py
@@ -368,7 +368,8 @@ def _init_sess_distrib(self):
         # save_checkpoint_steps = self.save_freq)
 
     def train (self, 
-               data) :
+               data,
+               valid_data) :
         stop_batch = self.stop_batch
         if self.run_opt.is_distrib :
             self._init_sess_distrib()
@@ -417,7 +418,7 @@ def train (self,
             feed_dict_batch[self.place_holders['is_training']] = True
 
             if self.display_in_training and is_first_step :
-                self.test_on_the_fly(fp, data, feed_dict_batch)
+                self.test_on_the_fly(fp, valid_data, feed_dict_batch)
                 is_first_step = False
             if self.timing_in_training : tic = time.time()
             self.sess.run([self.train_op], feed_dict = feed_dict_batch, options=prf_options, run_metadata=prf_run_metadata)
@@ -428,7 +429,7 @@ def train (self,
 
             if self.display_in_training and (cur_batch % self.disp_freq == 0) :
                 tic = time.time()
-                self.test_on_the_fly(fp, data, feed_dict_batch)
+                self.test_on_the_fly(fp, valid_data, feed_dict_batch)
                 toc = time.time()
                 test_time = toc - tic
                 if self.timing_in_training :
@@ -461,12 +462,12 @@ def print_head (self) :
 
     def test_on_the_fly (self,
                          fp,
-                         data,
+                         valid_data,
                          feed_dict_batch) :
         # ! altered by Marián Rynik
         # Do not need to pass numb_test here as data object already knows it.
         # Both DeepmdDataSystem and ClassArg parse the same json file
-        test_data = data.get_test(n_test=data.get_sys_ntest())
+        test_data = valid_data.get_test(n_test=valid_data.get_sys_ntest())
         feed_dict_test = {}
         for kk in test_data.keys():
             if kk == 'find_type' or kk == 'type' :
diff --git a/source/train/train.py b/source/train/train.py
index e7978361b2..55357fa670 100755
--- a/source/train/train.py
+++ b/source/train/train.py
@@ -136,11 +136,30 @@ def _do_work(jdata, run_opt):
                        sys_probs = sys_probs, 
                        auto_prob_style = auto_prob_style)
     data.add_dict(data_requirement)
+
+    ### START modified by ziyao
+
+    # init valid data
+    valid_data = data  # use train data if no validation is specified
+    if 'valid' in jdata.keys():
+        valid_systems = j_must_have(jdata['valid'], 'systems')
+        valid_data = DeepmdDataSystem(valid_systems,
+                                      1,
+                                      test_size,
+                                      rcut,
+                                      set_prefix=set_pfx,
+                                      type_map=ipt_type_map,
+                                      modifier=modifier)
+        valid_data.print_summary(run_opt,
+                                 sys_probs=sys_probs,
+                                 auto_prob_style=auto_prob_style)
+        valid_data.add_dict(data_requirement)
+
     # build the model with stats from the first system
     model.build (data, stop_batch)
     # train the model with the provided systems in a cyclic way
     start_time = time.time()
-    model.train (data)
+    model.train (data, valid_data)
     end_time = time.time()
     run_opt.message("finished training\nwall time: %.3f s" % (end_time-start_time))