[QUESTION] Models implementation

[william1599]

Hello! I'm using Darts as part of my thesis work (properly citing the repository and giving full credit), and one of my professors has asked me to verify that the implementations of the models here are consistent with the original implementations from their respective sources (particularly the TFT model).

Could you provide any insights on how these implementations were developed? Were they directly adapted? Is there any documentation or validation process that ensures they match the original models? Any information would be greatly appreciated. Thanks in advance!

[dennisbader]

Hi @william1599, we tried to implement the model as close to the original paper as possible.

Our implementation is based on the one form PyTorch Forecasting (see the model description here for more info).

Different to their implementation, we do not support:

• variable input / output window sizes (we use fixed-size input and output windows of size input_chunk_length and output_chunk_length)
• embedding bags for multiple categorical variables

We add additional support for:

• option to use reversible instance normalization
• train a probabilistic model using any of our Darts likelihood models (QuantileRegression (as in source paper), Gaussian Likelihood, Poisson likelihood, ...)

[william1599]

This helps me a lot, thank you!