lstm_crf.py 文件源码

def inference(self,X,reuse = None,trainMode=True):
        word_verctors = tf.nn.embedding_lookup(self.words,X)
        length = self.length(word_verctors)
        length_64 = tf.cast(length,tf.int64)
        if trainMode:
            word_verctors = tf.nn.dropout(word_verctors,0.5)

        with tf.variable_scope('rnn_fwbw',reuse =reuse) as scope:
            lstm_fw = rnn.LSTMCell(self.numHidden)
            lsmt_bw = rnn.LSTMCell(self.numHidden)

            inputs = tf.unstack(word_verctors,nlp_segment.flags.max_sentence_len,1)
            output,_,_ = rnn.static_bidirectional_rnn(lstm_fw,lsmt_bw,inputs,sequence_length=length_64,dtype=tf.float32)
            output = tf.reshape(output,[-1,self.numHidden * 2])

        matricized_unary_scores = tf.matmul(output,self.W) + self.b
        unary_scores = tf.reshape(matricized_unary_scores,
                                  [-1,nlp_segment.flags.max_sentence_len,self.distinctTagNum])
        return unary_scores,length