I am having trouble to understand what kind of input / output mapping would make most sense if I have a vanilla autoencoder structure (unlike LSTM AE).
I have 6 input sequences (sine signals) that I want to map to two output quasi sine signals. The full dataset has the shape (8, 2000000).
How should one batch look like? (One period is 400 time steps long.)
- Do I take 400 steps of each signal and concatenate them? (this would make one very long vector)
- Do I take 400 steps of each signal and input each signal separately (mapping it to the same output segment)
- Do I take 1 step of each signal and concatenate them?
- Do I take X step of each signal, concatenate them, and slide a moving window on the data?