Hello,
I’m working on LSTM-VAE to make anomaly dectection on timeseries (with no constant duration). I made the code using the dropout parameter included in pytorch LSTM module.
When I activate the dropout during the train phase, I have a pretty good reconstruction of the timeserie for the train phase but not for the eval phase : I only have a horizontal line for the reconstructed timeserie.
When I fix dropout=0.0 (LSTM module parameter), I have horizontal line too during the train phase.
Do you know where come from this behavior ? Is-it possible that the train phase is totally catched by the dropout ? If yes, what ist the solution ?
Thanks by advance.