Merge pull request #7881 from peterzhang2029/rnn_encoder_decoder_fix

Fix the context vector calculation in test_rnn_encoder_decoder.py.

python/paddle/v2/fluid/tests/book/test_machine_translation_encoder_context.py → python/paddle/v2/fluid/tests/book/test_rnn_encoder_decoder.py

@@ -118,12 +118,13 @@ def seq_to_seq_net():
     src_forward, src_backward = bi_lstm_encoder(
         input_seq=src_embedding, hidden_size=encoder_size)
 
-    encoded_vector = fluid.layers.concat(
-        input=[src_forward, src_backward], axis=1)
+    src_forward_last = fluid.layers.sequence_last_step(input=src_forward)
+    src_backward_first = fluid.layers.sequence_first_step(input=src_backward)
 
-    enc_vec_last = fluid.layers.sequence_last_step(input=encoded_vector)
+    encoded_vector = fluid.layers.concat(
+        input=[src_forward_last, src_backward_first], axis=1)
 
-    decoder_boot = fluid.layers.fc(input=enc_vec_last,
+    decoder_boot = fluid.layers.fc(input=encoded_vector,
                                    size=decoder_size,
                                    bias_attr=False,
                                    act='tanh')
@@ -137,7 +138,7 @@ def seq_to_seq_net():
         dtype='float32')
 
     prediction = lstm_decoder_without_attention(trg_embedding, decoder_boot,
-                                                enc_vec_last, decoder_size)
+                                                encoded_vector, decoder_size)
     label = fluid.layers.data(
         name='label_sequence', shape=[1], dtype='int64', lod_level=1)
     cost = fluid.layers.cross_entropy(input=prediction, label=label)