原文 https://github.com/tensorflow/nmt/ The encoder RNN uses zero vectors as its starting states encoder端的rnn使用零向量作为初始状态