elephaint
commented
5 months ago You can use scaler_type='revin' for DLinear.
Unfortunately for some models (like PatchTST, but also TSMixer) we had to separately include revin normalization as a feature to the model, hence the possible confusion. We hope to fix this inconsistency in the future.
Let me know if this works for you.
chaoticAttractoor
Description
The reversible instance normalization is a powerful transformation for the PatchTST model and the literature has shown that, applied to the DLinear model, it can outperforms the classic DLinear model (e.g. here) as well as transformer models with less computational overheads. I request the Revin class be incorporated in the DLinear model.
Use case
Revin - as a feature for DLinear will provide a more competive baseline benchmark for transformers and other SOTA models. Moreover, DLinear + Revin may provide a more accurate and computational efficient forecasting tool than other models in the library currently - improving the quality and functionality of the library.