def call(self, inputs, is_training=True):
bert = self.bert(inputs, is_training)
sequence_output = bert.get_sequence_output() # batch,sequence,768
pre = self.dense(sequence_output)
pre = tf.reshape(pre, [self.batch_size, self.maxlen, -1])
output = tf.math.softmax(pre, axis=-1)
return output
def call(self, inputs, is_training=True):
# 数据切分
input_ids, token_type_ids, input_mask, Y = tf.split(inputs, 4, 0)
input_ids = tf.cast(tf.squeeze(input_ids, axis=0), tf.int64)
token_type_ids = tf.cast(tf.squeeze(token_type_ids, axis=0), tf.int64)
input_mask = tf.cast(tf.squeeze(input_mask, axis=0), tf.int64)
Y = tf.cast(tf.squeeze(Y, axis=0), tf.int64)
# 模型构建
bert = self.bert([input_ids, token_type_ids, input_mask], is_training)
sequence_output = bert.get_sequence_output() # batch,sequence,768
predict = self.dense(sequence_output)
predict = tf.reshape(predict, [self.batch_size, self.maxlen, -1])
# 损失计算
log_likelihood, transition = self.crf(predict, Y, sequence_lengths=tf.reduce_sum(input_mask, 1))
loss = tf.math.reduce_mean(-log_likelihood)
predict, viterbi_score = self.crf.crf_decode(predict, transition,
sequence_length=tf.reduce_sum(input_mask, 1))
return loss, predict