How to train iTransformer for classification with varying number of variates?

[hujoe17]

I am planning to train an iTransformer model for classification with pypots.classification.itransformer.iTransformer, however the number of TS in different observations may be different in my data. The paper claims that it can handle varying number of tokens (i.e., variates), but I am not sure how to handle it in PyPOTS. @WenjieDu

[WenjieDu]

"Varying numbers of tokens" means different sample lengths rather than variates/features.

[hujoe17]

Based on the paper, I believe each token is a time series. Please let me know if think this is different from what has been implemented in PyPOTS.

[WenjieDu]

1. All model implementations in PyPOTS respect their official code as much as possible;
2. Right, in the context of iTransformer, it theoretically accepts samples with different numbers of features. However, to achieve this, you have to ensure samples in a batch have the same number of features, otherwise, you need padding. Even in the official code, they didn't implement these features. If you're interested in this and would like to make contributions, please continue the discussion here and commit your code. I'll help with your questions and code review;