Uniformise the dataset output format

[madtoinou]

Is your feature request related to a current problem? Please describe.

Dataset batch's format differs depending on the model class and the supported covariates, making it quite counter-intuitive in a lot of cases and difficult to implement new features across models.

Describe proposed solution
Have a single batch format, with clear index for each covariates (even if it's not supported/used) across all Dataset

Describe potential alternatives
Cannot think of one, maybe have a utils function in each class that maps the current format to a unified format?

[alexcolpitts96]

Hey @madtoinou. Just to double check that I understand what you are thinking:

Forward functions are generally x_in: tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]] and you would be looking to convert it to something more explicit like target: torch.Tensor, past_covariates: Optional[torch.Tensor], future_covariates: Optional[torch.Tensor] (and any other potential named fields).

If that is the case, I like the idea.

[madtoinou]

Hi @alexcolpitts96,

This is precisely the idea, also extended to historic past covariates, historic future covariates as well.

Glad that you like the idea.

[dennisbader]

I'll take this one.