BERT分类提升准确率

在bert最后一层添加一层rnn会使分类效果更好
下面代码基于bert4keras:
例：

model = build_transformer_model(
            self.config_path,
            None if is_predict else self.checkpoint_path,
            with_mlm=False,
            keep_tokens=self.data_deal.keep_tokens
        )
logits_output = keras.layers.Bidirectional(keras.layers.LSTM(256, return_sequences=True), name='Out-RNN')(model.output)
output = keras.layers.Dense(2, name='Out-Dense-Logit', activation='sigmoid')(logits_output)
model = Model(model.input, output)
model.summary()