search

KimMeen / Time-LLM

[ICLR 2024] Official implementation of " 🦙 Time-LLM: Time Series Forecasting by Reprogramming Large Language Models"
https://arxiv.org/abs/2310.01728
Apache License 2.0
1.46k stars 252 forks source link

Question regarding the dataset's StandardScaler() #124

Closed khchul closed 3 months ago

khchul commented 4 months ago

Thank you for your great work! I've been looking through your paper and the implementation, but I have a question regarding the scaling. In the paper, it is said that "Each input channel X(i) is first individually normalized to have zero mean and unit standard deviation". I understood that this is done in the TimeLLM's normalize_layers. However, the data itself has already been scaled through data_loader.py and as far as I know, inverse_transform() doesn't seem to appear elsewhere. Shouldn't the scale variable in the data_loader be set to False?

kwuking commented 3 months ago

Thank you for your great work! I've been looking through your paper and the implementation, but I have a question regarding the scaling. In the paper, it is said that "Each input channel X(i) is first individually normalized to have zero mean and unit standard deviation". I understood that this is done in the TimeLLM's normalize_layers. However, the data itself has already been scaled through data_loader.py and as far as I know, inverse_transform() doesn't seem to appear elsewhere. Shouldn't the scale variable in the data_loader be set to False?

Thank you for your attention. We are currently using the framework template in TSLib. At present, the comparisons for these well-acknowledged benchmarks are all conducted after normalization. You can refer to TSLib for more information.