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
617 stars 127 forks source link

Some questions about the results of the PEMS dataset #32

Closed liuaoy closed 2 years ago

liuaoy commented 2 years ago

Hello, your work is very rewarding! But I had some deviations in the MAE, MAPE, RMSE metrics when I performed the replication on the PEMS dataset, and I only used the model part of the code you provided. When using your code completely, MAE, MAPE, RMSE and the results in the paper are about the same, so I would like to ask you if you use any tips in data preprocessing.

liuaoy commented 2 years ago

You mentioned using the same settings as StemGNN, what are the exact operations of these settings? Also, if I use only the code of your model part during training, it may only take 15s. if I use your code completely (including data reading, training code), the training time will be twice as long, so is there some operation in data processing, looking forward to your reply!

ailingzengzzz commented 2 years ago

Hi @liuaoy,

We used the same data preprocessing as previous works shown in our repo [function in the experiments/exp_pems.py#L59]. Data preprocessing will take some time. You can use [time()] to calculate the processing time of each part.