diff --git a/train.py b/train.py
index d8f609c..5a0089f 100644
--- a/train.py
+++ b/train.py
@@ -42,8 +42,8 @@
 # Hyper-parameters.
 NUM_EPOCHS = 200
 NUM_HIDDEN = 50
-NUM_LAYERS = 1
-BATCH_SIZE = 1
+NUM_LAYERS = 2
+BATCH_SIZE = 4
 
 # Optimizer parameters.
 INITIAL_LEARNING_RATE = 1e-2
@@ -101,11 +101,12 @@ def main(argv):
             sequence_length_placeholder = tf.placeholder(tf.int32, [None])
 
             # Defining the cell.
-            cell = tf.contrib.rnn.LSTMCell(NUM_HIDDEN, state_is_tuple=True)
+            def lstm_cell():
+                return tf.contrib.rnn.LSTMCell(NUM_HIDDEN, state_is_tuple=True)
 
             # Stacking rnn cells.
-            stack = tf.contrib.rnn.MultiRNNCell([cell] * NUM_LAYERS,
-                                                state_is_tuple=True)
+            stack = tf.contrib.rnn.MultiRNNCell(
+                [lstm_cell() for _ in range(NUM_LAYERS)], state_is_tuple=True)
 
             # Creates a recurrent neural network.
             outputs, _ = tf.nn.dynamic_rnn(stack, inputs_placeholder, sequence_length_placeholder, dtype=tf.float32)