seq2seq 做翻译，预测时 用argmax的原因

在这里https://github.com/tensorflow/models/blob/master/tutorials/rnn/translate/translate.py#L281-L282

# This is a greedy decoder - outputs are just argmaxes of output_logits.outputs = [int(np.argmax(logit, axis=1)) for logit in output_logits]

最初很难理解，其实就是在所有词里选概率最大的

因为这段代码的存在https://github.com/tensorflow/models/blob/master/tutorials/rnn/translate/seq2seq_model.py#L168-L173

      if output_projection is not None:        for b in xrange(len(buckets)):          self.outputs[b] = [              tf.matmul(output, output_projection[0]) + output_projection[1]              for output in self.outputs[b]          ]

里面output的维数就是embedding之后的维数，比如1024
output_projection的维数是[1024,vocabulary_size]