pytorch-forecasting icon indicating copy to clipboard operation
pytorch-forecasting copied to clipboard

[ENH] Is there a way to pass an entire embedding to time_varying_unknown_reals without splitting it into individual components?

Open LinearBasis opened this issue 10 months ago • 2 comments

Hello!

I have a question regarding the time_varying_unknown_reals parameter. Is there any way to pass an entire embedding (e.g., a dense vector representation) directly to time_varying_unknown_reals, rather than splitting it into its individual components?

For example, if I have an embedding of size n, can I pass it as a single feature, or do I need to split it into n separate features?

Thank you in advance for your help!

LinearBasis avatar Mar 06 '25 20:03 LinearBasis

Can you explain what you mean by "splitting into components"?

fkiraly avatar Mar 15 '25 20:03 fkiraly

At the moment, I’m splitting an embedding of size n into n columns and adding them to my dataframe as 'embedding_{i}', and then passing these columns to time_varying_unknown_reals. Is there any way I can let the model know that this is all a single entity by passing everything at once

LinearBasis avatar Mar 16 '25 18:03 LinearBasis