Closed inspirit closed 1 year ago
Hi,
As shown in our paper: scenarios 1, 2, and 4 are for univariate TS. However, the TFC model can work on multivariate TS as long as the contrastive encoder able to accept multiple channels (in our scenario 3). The current codes are for single-channel TS, but you can easily revise the code to apply it for multi-dimensional data.
It's not that simple, right? If you change the code to multi-channel, how do you define your constractive loss? How do you combine information from multiple channels into a single representation vector?
I spent almost half a day studying your code, and I found that there was a huge gap with the paper.
Hello,
would you be so kind to explain why do you remove all features except first in all datasets? Does it mean current algorithm only works for univar? thank you
https://github.com/mims-harvard/TFC-pretraining/blob/6d0ce38664d511302249c1fcad4c49590fe6bca0/code/dataloader.py#L29