Open svnv-svsv-jm opened 3 months ago
Thank you for pointing on this.
The exemplary LSTMModel in the notebook is not capable of doing multivariate forecasting. Currently, it seems only be able to do univariate time series forecasting.
Thus, I see the necessity to clarify this in the notebook. However, I suppose it is not a bug.
Can you kindly link the Pull Request to which you are referring?
Would tags for the objects be useful here? E.g., capability:multivariate
I have opened an issue to explain: https://github.com/sktime/pytorch-forecasting/issues/1679
Other question: if the model is interfaced in sktime
, it would be able to do multivariate forecasting through broadcasting, although that may be more wasteful than the native capability.
Would tags for the objects be useful here? E.g.,
capability:multivariate
I have opened an issue to explain: https://github.com/sktime/pytorch-forecasting/issues/1679
In this case not sure since this model is implemented in the notebook. Probably to show how to do this with PyTorch-forecasting.
Other question: if the model is interfaced in
sktime
, it would be able to do multivariate forecasting through broadcasting, although that may be more wasteful than the native capability.
Yes it would enable multivariate forecast but the underlying model is still univariate. So enabling multivariate here would lead to a different forecast than broadcasting.
In this case not sure since this model is implemented in the notebook. Probably to show how to do this with PyTorch-forecasting.
That is weird. Why would one implement it in the notebook? I think this should move to the main code base.
I have not seen the notebook, is this some kind of demo? Either way, LSTM are an important class of models, even if not the freshest one, so it should go into the main code base.
Just create this dataset:
And input it to the current
LSTMModel
in the tutorials:And you'll get:
It's just weird that there is still no fix for this, and no LSTM model out-of-the-box. I even made a fix, there is a PR.
Why does no one care about fixing this?
It is totally obscure how
pytorch_forecasting
handles uni-/multi-targets, I've also noticed that if you passtarget=["target"]
toTimeSeriesDataSet
, theTimeSeriesDataSet
behaves very differently w.r.t. if you passedtarget="target"
.Please just review that PR and even merge it, or fix it...