@torch.no_grad()
    def decode(self, encoder_output: Tensor, max_lengths: int) -> Tensor:
        r"""
        Decode `encoder_output`.

        Args:
            **encoder_output** (torch.FloatTensor): A output sequence of encoder. `FloatTensor` of size
                ``(seq_length, dimension)``
            **max_lengths** (int): Max decoding time step

        Returns:
            **decode_output** (torch.LongTensor): Result of model predictions
        """
        token_list = list()
        hidden_states = None

        token = torch.LongTensor([[self.decoder.sos_id]])
        if torch.cuda.is_available():
            token = token.cuda()

        for i in range(max_lengths):
            decoder_output, hidden = self.decoder(token, hidden_states=hidden_states)
            output = self.joint(encoder_output[i].view(-1), decoder_output.view(-1))
            prediction_token = output.topk(1)[1]
            token = prediction_token.unsqueeze(1)  # (1, 1)
            prediction_token = int(prediction_token.item())
            token_list.append(prediction_token)

        return torch.LongTensor(token_list)