Closed elephaint closed 4 months ago
Check out this pull request on
See visual diffs & provide feedback on Jupyter Notebooks.
Powered by ReviewNB
@elephaint
n_series=1
work)?futr_exog_list
, hist_exog_list
, and stat_exog_list
for each time series (like in this implementation https://github.com/ditschuk/pytorch-tsmixer?tab=readme-ov-file#tsmixerext)?Thanks
@elephaint
- If we have several thousands of time series (making multivariate forecasting unfeasible), is it possible to use it in an "univariate" way (e.g. would setting
n_series=1
work)?
I wouldn't do it - we use the unique_id
to distinguish individual time series and it would require overwriting this / setting it to the same value for all series in order to function with n_series=1. However, that will introduce a lot of other issues. So, personally I would not do it and then revert to another univariate method.
- Does it correctly support
futr_exog_list
,hist_exog_list
, andstat_exog_list
for each time series (like in this implementation https://github.com/ditschuk/pytorch-tsmixer?tab=readme-ov-file#tsmixerext)?Thanks
No, we will soon be releasing a separate version that supports exogenous inputs ('TSMixerx'). Unfortunately the version supporting exogenous variables is quite different from the original TSMixer. We have a working version of TSMixerx on the AirPassengers Panel data with exogenous inputs, I'm still making sure it works with the rest of our library/examples. Hopefully ready for release next week.
@elephaint Thanks! So basically also TSMixerx will not be suitable for "univariate" + exog variables forecasting (like we do with NHITS, NBEATSx, etc.)?
Correct
This PR adds the TSMixer model into neuralforecast, and demonstrates its applicaiton in a Multivariate_with_TSMixer example notebook.