search

DAMO-DI-ML / NeurIPS2023-One-Fits-All

The official code for "One Fits All: Power General Time Series Analysis by Pretrained LM (NeurIPS 2023 Spotlight)"
398 stars 57 forks source link

Seq len inconsistency #37

Open GGchen1997 opened 2 months ago

GGchen1997 commented 2 months ago

Hi authors,

Thanks for your interesting work.

I have some issues. Your paper cites the TimesNet paper results as a direct comparison but your setting is inconsistent with the setting of TimesNet. For example, in long-term forecasting, TimesNet uses seq len = 96 while you use seq len = 512, which is unfair, if I understand it correctly.

PSacfc commented 2 months ago

For each baseline, we use the best proposed results from the original paper or reproduced under the same experiments setting.