python - Transformer for time series forecasting

Question

I have discrete daily features and a target time series and I'm trying to implement a basic Transformer for seq2seq modeling. I construct my supervised data as follows:

 . . .. . .. . . .
       |
 0 0 0 0 0 0 0 0

The next sequence is shifted by one position ahead. Overall I have the input data shape [batch_size, in_sequence_len, num_features] and the target is [batch_size, out_sequence_len, 1]. I understand that encoder input should be of shape [batch_size, in_seq_len, num_features] and decoder takes [batch_size, out_seq_len, num_features]. But how should I transform my batch to be suitable for the transformer input?

Answer 1

等待大神答复