pawaritl / bayesianlstm Goto Github PK

Was this intentional?

The original paper proposes somewhat different ideas than the current code in this repo, such as:

• Prediction network only uses encoder after the encoder-decoder structure is pre-trained.
• Encoder uses VariationalDropout while Prediction Network uses regular dropout.
• Inherent noise is modeled at inference time using a validation set.
• Encoder reads up to T timestamps and Decoder reconstructs next F timestamps in the window.
• Input is de-trended.

I'm only pointing out these differences because this repo shows up as the first implementation of the original paper on PapersWithCode website (link: https://paperswithcode.com/paper/deep-and-confident-prediction-for-time-series), but the implementation is very different from the idea in the paper.

the code only consider about model uncertaintu, how we can get aleatoric one

Thank you very much for sharing you code!
In your work, the prediction of BayesianLSTM is using past time history to predict one step ahead (which is many to one).
Is it possible to predict multiple step at once based on past time history ( which is many to many)?
Thank you.
Eric