diff --git a/generate_sequence_by_rnn_lm/network_conf.py b/generate_sequence_by_rnn_lm/network_conf.py
index f1aceb0b7d70c6a2aec601cf935d2f34500d20fc..f2e59a94298435f280a561a1e079defa2ea84f62 100644
--- a/generate_sequence_by_rnn_lm/network_conf.py
+++ b/generate_sequence_by_rnn_lm/network_conf.py
@@ -57,4 +57,4 @@ def rnn_lm(vocab_dim,
     else:
         cost = paddle.layer.classification_cost(input=output, label=target)
 
-        return cost, output
+        return cost
diff --git a/generate_sequence_by_rnn_lm/train.py b/generate_sequence_by_rnn_lm/train.py
index d8bffd487a6b8c0b74d28a97b16c689255dd472b..198835829812426f49682b3fdb43f6c35c55a319 100644
--- a/generate_sequence_by_rnn_lm/train.py
+++ b/generate_sequence_by_rnn_lm/train.py
@@ -43,9 +43,14 @@ def train(topology,
 
     # create parameters
     parameters = paddle.parameters.create(topology)
+    # create sum evaluator
+    sum_eval = paddle.evaluator.sum(topology)
     # create trainer
     trainer = paddle.trainer.SGD(
-        cost=topology, parameters=parameters, update_equation=adam_optimizer)
+        cost=topology,
+        parameters=parameters,
+        update_equation=adam_optimizer,
+        extra_layers=sum_eval)
 
     # define the event_handler callback
     def event_handler(event):