Why the results of PatchTST in iTransformer paper are worse than those in PatchTST paper?
Closed this issue · 3 comments
I reproduced the results of iTransformer, thank you for the clear and robust code! I also reproduced the results of PatchTST in the original paper(using the official repo), but I found that in iTransformer paper, PatchTST results are worse than those in the original paper, is it because that you reduce the input length of PatchTST from 336/512 to 96?
Take dataset Weather as an example:
PatchTST in iTransformer paper(Table10):
Hello, we follow the long-term forecasting setting of Informer and Autoformer, where the lookback length is fixed and never tuned.
For more details, you may look at the difference in Look-Back-96
and Look-Back-Searching
settings in: https://github.com/thuml/Time-Series-Library
Thanks for your reply! It helps.