Batch processing with variable length sequences

Question

I have a lot of time series with different lengths. I would like to know what are the best practices to fit them to a Bidirectional LSTM model. The problem is a Binary Classification of Sequence to Sequence. So for every time step, I want to predict the binary class.

Currently, I create a tensor for each data frame with the shape of (1, None, #Features). Then I fit every tensor separately to the model.

Is it better to combine every data frame to a tensor of the shape (#Time Series, None, #Features) and fit them all at once? Does this make even a difference?

Or could it be better to go with the sliding window approach and split one time-series into smaller windows?

I can't specify a max-length of the time-series so I think I cannot use pad-sequence from Keras.

Answer 1

Best practices is to have each time series be the same length for Bidirectional LSTM.

You say not specify the maximum length so another approach is to pick a fixed length. If the data is too short, pad. If the data is too long, trim.