Setting up a pytorch-forecasting TimeSeriesDataSet for a sequence-to-sequence DataLoader

Hey Folks. I just discovered the pytorch-forecasting package’s TimeSeriesDataSet class, and how it helps with taking data from a pandas dataframe and creating a pytorch DataLoader. They show one example of creating a TimeSeries Dataset, but don’t but don’t have much in the way of a tutorial etc. I was wondering if I can create a sequence-to-sequence dataset using this class, meaning I would need the TimeSeriesDataSet class to cut the time series data into windows, where say the first 10 days are the data and the subsequent 10 days are the labels.

I have code to do this manually right now, and it works. But for the future it would be nice to have a tool like this to setup new datasets faster. So was just wondering if anyone know whether I can setup a Sequence-to-Sequence set of (data, labels) for the DataLoader, and if anyone knows of a good example of this.


The TimeSeriesDataSet does exactly what you are describing. first 10 days are in x attribute encoder_target and the next 10 days in x attribute decorder_target. the next 10 days is also in y attribute target. read about this in to_dataloader method in

I believe the encoder_cont and encoder_real are fed into encoder input. attribute target of y is used as label.