I am learning to use a LSTM model to predict time series data. Specifically, I hope the network should output a sequence (with multiple time steps) only after the input sequence has finished feeding in, as shown in the left figure.
However, most of the LSTM sequence-to-sequence prediction tutorial I have read seems to be the right figure (i.e. each time step of the output sequence is generated after each time step of the input sequence). What's more, as far as I understand, the LSTM implementation in PyTorch (and probably Keras) can only return output sequence corresponding to each time step of the input sequence. It cannot make predictions after the input sequence is over.
I hope to know is there any way to make a sequence-to-sequence LSTM network which starts output only after the input sequence has finished feeding in? And it would be better if someone can show me an example implementation code.