search

Nixtla / neuralforecast

Scalable and user friendly neural :brain: forecasting algorithms.
https://nixtlaverse.nixtla.io/neuralforecast
Apache License 2.0
3.13k stars 362 forks source link

[ NeuralForecast: TFT|TemporalCovariateEncoder| nn.lstm ] #1185

Closed Otsutsukii closed 2 weeks ago

Otsutsukii commented 1 month ago

Description

Hi could you add a number of layers parameters in the TFT's LSTM so we can do stacking of LSTM?

Like num_layers=1 by defaults?

Use case

I am trying to get a better accuracy for the TFT, by stacking more LSTM layers in the TFT TemporalCovariateEncoder.

marcopeix commented 2 weeks ago

Hi! I experimented with this and it seems the architecture does not allow that unless we rework more parts of the model. To improve accuracy, you can always experiment with different activation functions or number of attention heads.