search

lss-1138 / SparseTSF

[ICML 2024 Oral] Official repository of the SparseTSF paper: "SparseTSF: Modeling Long-term Time Series Forecasting with 1k Parameters". This work is developed by the Lab of Professor Weiwei Lin (linww@scut.edu.cn), South China University of Technology; Pengcheng Laboratory.
Apache License 2.0
122 stars 7 forks source link

I got SOTA result with hparam search #1

Closed lukaspetersson closed 2 months ago

lukaspetersson commented 4 months ago

I put the Sparse TSF in the PatchTST repo and did a hyper parameter search. I got a MSE of 0.31 for 336 prediction length which is better than SOTA. Am I doing something wrong or are the SOTA results not optimized?

lss-1138 commented 4 months ago

Thank you for your interest. The results reported in the paper were not obtained through hyperparameter search. The hyper parameter we focused on, period_len, is determined by the characteristics of the dataset. As for batch_size and learning_rate, we uniformly chose relatively consistent values. So, it's possible that you achieved better results through hyperparameter search.