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
633 stars 129 forks source link

Cannot reproduce the SOTA performance on ETTh2 dataset. #58

Open kehuo opened 1 year ago

kehuo commented 1 year ago

Hi,

Thanks for sharing those reproducible code and examples. However, I cannot reproduce the metrics recorded in the paper.

For example, as shown in the below Table 4 in the paper, for ETTh2 dataset, when the horizon is set to 720, MSE = 0.475, MAE = 0.488:

image

However, I ran the below command but cannot get such metrics: [Command I ran]: python run_ETTh.py --data ETTh2 --features M --seq_len 736 --label_len 720 --pred_len 720 --hidden-size 4 --stacks 1 --levels 5 --lr 1e-5 --batch_size 128 --dropout 0.5 --model_name etth2_M_I736_O720_lr1e-5_bs128_dp0.5_h4_s1l5

[Results I got]: Final mean normed mse:1.0782,mae:0.7634

Could you please advice how to repro the Multi-variates forecasting performance on ETTh2 dataset?

Thanks a lot and look forward to your reply.