search

cure-lab / SCINet

The GitHub repository for the paper: “Time Series is a Special Sequence: Forecasting with Sample Convolution and Interaction“. (NeurIPS 2022)
Apache License 2.0
620 stars 127 forks source link

RevIN is not suitable for pEMS data sets #23

Closed wengwenchao123 closed 2 years ago

wengwenchao123 commented 2 years ago

Hello:
I have tried to use RevIN on perMS data set, the result is far worse than without RevIN, may I ask whether your test result is the same

wengwenchao123 commented 2 years ago

Hello: I have tried to use RevIN on perMS data set, the result is far worse than without RevIN, may I ask whether your test result is the same It's using the hyperparameters that you give it

VEWOXIC commented 2 years ago

We did try the RevIN on the pems dataset and found it did hurt the performance. We presume that it is because the distribution for the training and testing set of PEMS is identical. You can use the tool provided in ./utils/histogram.ipynb to verify the distribution. We recommend enabling RevIN when these two distribution is remarkably different. Thanks for your attention!

nhansendev commented 2 years ago

Is there another name for RevIN, or a definition in the paper? I have not found anything from Google searches. EDIT: Nevermind, found the documentation for it in the docs folder.

For anyone who sees this and is curious, the full name of paper is: Reversible Instance Normalization For Accurate Time-series Forecasting Against Distribution Shift Please add this information somewhere more obvious to avoid confusion (maybe a comment in the code?). :)

liuaoy commented 2 years ago

可能是中继监督