Number of features/channels considered for the experiments in Multivariate data

[atif-hassan]

Hi Authors!!

Thank you so much for this contribution which has really helped the community by leaps and bounds!

I had a very basic doubt.
Since each multivariate dataset has its own number of features/channels, what is the size considered for experimenting with deep learning models wherein the dataloader must be fed with the data of equal size.
Basically, if the dataset is of shape (b,w,n) where b is the batch size, w is the window size and n is the number of channels then what is the value of n?

Once again, thank you so much for this amazing library!

[qhliu26]

Hi! Thank you very much for your interest in our project. We adopt a 'one-for-one' evaluation setup, meaning that we train a separate model for each individual time series. Each time series is paired with its own dedicated dataloader, where each data item retrieved is of the form (w,n), which is then fed into the DNN model.

Let us know of any questions or suggestions!

[atif-hassan]

Hey!

Oh, I see. Yes, a separate dataloader per time series should do the trick!
However, I did not understand what you meant by "we train a separate model for each individual time series". Do you train a model on let's say the Exathlon dataset and then move on to some other dataset with same model?