Skip to content

[Feature] Improve handling & documentation of encoders and categorical static covariates in .fit_from_dataset() #3006

@tRosenflanz

Description

@tRosenflanz

Describe the issue linked to the documentation
I couldn't find a mention of this in docs but it appears that TorchForecastingModel doesn't use encoders when using fit/predict_from_dataset - it doesn't warn about it either during the call. This seems like a potentially big blind side for the user since it is a silent behavior. Since this method is pretty well documented otherwise, this seems like an unexpected caveat.

I wonder if this should be documented OR actually move encoder logic to be included in the _from_dataset method calls.

Sidenote:
This leads to issues with TFTModel since it just skips training and errors on predict if datasets don't contain future_covariates.

Sidenote2: categorical feature embeddings with TFTModel are also not working as expected (essentially ignored AFAIK) when using the _from_dataset methods since the code uses feature integer index instead of names for lookups but this seems like a completely separate issue.

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    Status

    In progress

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions