diff --git a/dygraph/seq2seq/args.py b/dygraph/seq2seq/args.py
index b754702e21a6fd2d05cd6b7c6191356db21b2269..df787c8c2d947fc2a2716df734d6b4ff9c6559e7 100644
--- a/dygraph/seq2seq/args.py
+++ b/dygraph/seq2seq/args.py
@@ -34,12 +34,6 @@ def parse_args():
     parser.add_argument("--src_lang", type=str, help="source language suffix")
     parser.add_argument("--tar_lang", type=str, help="target language suffix")
 
-    parser.add_argument(
-        "--attention",
-        type=eval,
-        default=False,
-        help="Whether use attention model")
-
     parser.add_argument(
         "--optimizer",
         type=str,
diff --git a/dygraph/seq2seq/attention_model.py b/dygraph/seq2seq/attention_model.py
index bae46d47219bf68288b91b1aeecc138b9f98945d..7fc2e337c911514f49bc03c8a38172fbeb2f7beb 100644
--- a/dygraph/seq2seq/attention_model.py
+++ b/dygraph/seq2seq/attention_model.py
@@ -28,13 +28,12 @@ class AttentionModel(Layer):
                  trg_vocab_size,
                  num_layers=1,
                  init_scale=0.1,
-                 padding_idx=2,
+                 padding_idx=0,
                  dropout=None,
                  beam_size=1,
                  beam_start_token=1,
                  beam_end_token=2,
                  beam_max_step_num=100,
-                 mode='train',
                  dtype="float32"):
         super(AttentionModel, self).__init__()
         self.hidden_size = hidden_size
@@ -47,7 +46,6 @@ class AttentionModel(Layer):
         self.beam_start_token = beam_start_token
         self.beam_end_token = beam_end_token
         self.beam_max_step_num = beam_max_step_num
-        self.mode = mode
         self.kinf = 1e9
 
         self.encoder = Encoder(src_vocab_size, hidden_size, num_layers,
diff --git a/dygraph/seq2seq/reader.py b/dygraph/seq2seq/reader.py
index c76fe7cfb932165db77d755a99e5553f8698e5dc..907cf262262961e7d1e2d8790fa85d7923a5558f 100644
--- a/dygraph/seq2seq/reader.py
+++ b/dygraph/seq2seq/reader.py
@@ -125,14 +125,14 @@ def raw_data(src_lang,
     src_vocab = _build_vocab(src_vocab_file)
     tar_vocab = _build_vocab(tar_vocab_file)
 
-    train_src, train_tar = _para_file_to_ids( src_train_file, tar_train_file, \
-                                              src_vocab, tar_vocab )
+    train_src, train_tar = _para_file_to_ids(src_train_file, tar_train_file, \
+                                              src_vocab, tar_vocab)
     train_src, train_tar = filter_len(
         train_src, train_tar, max_sequence_len=max_sequence_len)
-    eval_src, eval_tar = _para_file_to_ids( src_eval_file, tar_eval_file, \
-                                              src_vocab, tar_vocab )
+    eval_src, eval_tar = _para_file_to_ids(src_eval_file, tar_eval_file, \
+                                              src_vocab, tar_vocab)
 
-    test_src, test_tar = _para_file_to_ids( src_test_file, tar_test_file, \
+    test_src, test_tar = _para_file_to_ids(src_test_file, tar_test_file, \
                                               src_vocab, tar_vocab )
 
     return (train_src, train_tar), (eval_src, eval_tar), (test_src, test_tar),\
@@ -143,8 +143,8 @@ def raw_mono_data(vocab_file, file_path):
 
     src_vocab = _build_vocab(vocab_file)
 
-    test_src, test_tar = _para_file_to_ids( file_path, file_path, \
-                                              src_vocab, src_vocab )
+    test_src, test_tar = _para_file_to_ids(file_path, file_path, \
+                                              src_vocab, src_vocab)
 
     return (test_src, test_tar)
 
@@ -160,7 +160,8 @@ class IWSLTDataset(Dataset):
         src_data, trg_data = raw_data
         data_pair = []
         for src, trg in zip(src_data, trg_data):
-            data_pair.append([src, trg])
+            if len(src) > 0:
+                data_pair.append([src, trg])
         sorted_data_pair = sorted(data_pair, key=lambda k: len(k[0]))
 
         src_data = [data_pair[0] for data_pair in sorted_data_pair]
diff --git a/dygraph/seq2seq/run.sh b/dygraph/seq2seq/run.sh
index 8005ec26af0b415085c02d364c07eecb54af3d72..d8eb0800f717ba6e7191b35552cccdd58ad634e2 100644
--- a/dygraph/seq2seq/run.sh
+++ b/dygraph/seq2seq/run.sh
@@ -3,13 +3,12 @@ export CUDA_VISIBLE_DEVICES=0
 
 python train.py \
         --src_lang en --tar_lang vi \
-        --attention True \
         --num_layers 2 \
         --hidden_size 512 \
         --src_vocab_size 17191 \
         --tar_vocab_size 7709 \
         --batch_size 128 \
-        --dropout 0.0 \
+        --dropout 0.2 \
         --init_scale  0.2 \
         --max_grad_norm 5.0 \
         --train_data_prefix data/en-vi/train \
@@ -20,6 +19,7 @@ python train.py \
         --model_path attention_models \
         --enable_ce \
         --learning_rate 0.002 \
-        --dtype float64 \
+        --dtype float32 \
         --optimizer adam \
-        --max_epoch 1
+        --max_epoch 12 \
+        --padding_idx 2