search

time-series-foundation-models / lag-llama

Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting
Apache License 2.0
1.08k stars 121 forks source link

Adding our own covariates #58

Open FabioGadegbeku opened 1 month ago

FabioGadegbeku commented 1 month ago

First of all thank you for this very interesting model and paper! However I was a bit confused does the model solely used the lags predefined as covariates or we can add our own covariates ie : seasons/day of the week/ holiday or not etc.. ? If not, is there a simple way to implement this possibility on my own ? Thanks in advance.

ashok-arjun commented 1 month ago

Hi, thanks for the kind words!

You could add your own covariates, but you would have to pre-train the model from scratch with the covariates.

In order to adapt the codebase to use covariates, you'd have to implement having covariates in several functions such as this one. This requires a bit of looking into GluonTS documentation. Are you looking to do that? in that case, I can try to point you to some other GluonTS model implementations where covariates are implemented.

Sigurd-git commented 1 month ago

Hi, thanks for the kind words!

You could add your own covariates, but you would have to pre-train the model from scratch with the covariates.

In order to adapt the codebase to use covariates, you'd have to implement having covariates in several functions such as this one. This requires a bit of looking into GluonTS documentation. Are you looking to do that? in that case, I can try to point you to some other GluonTS model implementations where covariates are implemented.

Hello Arjun, I have a similar requirement. My variable is measured in days, while my additional covariates are measured in months. Could you provide a brief demonstration on how to finetune the model accordingly?

Sigurd-git commented 2 weeks ago

Hi, thanks for the kind words! You could add your own covariates, but you would have to pre-train the model from scratch with the covariates. In order to adapt the codebase to use covariates, you'd have to implement having covariates in several functions such as this one. This requires a bit of looking into GluonTS documentation. Are you looking to do that? in that case, I can try to point you to some other GluonTS model implementations where covariates are implemented.

Hello Arjun, I have a similar requirement. My variable is measured in days, while my additional covariates are measured in months. Could you provide a brief demonstration on how to finetune the model accordingly?

@ashok-arjun Just following up to ensure you have seen this.