models.py 文件源码

def _build_sequence_model(self, sequence_input):
        RNN = GRU if self._rnn_type == 'gru' else LSTM

        def rnn():
            rnn = RNN(units=self._rnn_output_size,
                      return_sequences=True,
                      dropout=self._dropout_rate,
                      recurrent_dropout=self._dropout_rate,
                      kernel_regularizer=self._regularizer,
                      kernel_initializer=self._initializer,
                      implementation=2)
            rnn = Bidirectional(rnn) if self._bidirectional_rnn else rnn
            return rnn

        input_ = sequence_input
        for _ in range(self._rnn_layers):
            input_ = BatchNormalization(axis=-1)(input_)
            rnn_out = rnn()(input_)
            input_ = rnn_out
        time_dist_dense = TimeDistributed(Dense(units=self._vocab_size))(rnn_out)

        return time_dist_dense