I'm trying to do feature extraction in some discretized time series with a variable length, doing that i'm creating an RNN auto encoder. My main problem is to find a way to let the model train with variable length time sequences. I read in TF guides about masking, and use first the padding to make all the sequences have the same length and then the embedding with mask_zero=True, then, i want to ask if it's a proper way to elaborate different length time series array and what i have to do if i have to construct the decoder that is symmetrical to the encoder ( that means that i have to de embed the data coming from the latent layer).
The masking guide that i talk about: https://www.tensorflow.org/guide/keras/masking_and_padding
I'm also open to other path to solve the problem.
