TODOs & HOWTOs

[zezhishao]

How-Tos

• How to add new datasets.
• How to config the config file.
• How to inference and save the inference results.
• How to customize loss function.

ToDos

Add more baselines:

• Add DeepAR
• Add LightGBM
• Add NHiTS
• Add GRU
• Add STGODE
• Add SCINet
• Add STSGCN❌, STFGCN❌ (They are built on MXNet)
• Add TimesNet
• Add Vanilla Transformer
• Add GPTs for Time Series
• Add AMP.
• Add early stopping.

Add more datasets:

• Add M4 dataset
• Add PeMSD7(M) and PeMSD7(L).

Others:

• Unify the data preprocssing script.
• Support lr scheduler ReduceLROnPlateau. (see 请问如何修改学习率衰减器 #120 )
• Fix issue clone下来后第一遍正常，运行第二遍generate_training_data.py就报错了，什么都没改 #90 .
• Add AutoML.

[huangst21]

您好，感谢您提供的框架。关于交通预测问题，或许您可以考虑添加上PEMS07(M)和PEMS07(L)这两个数据集？

[zezhishao]

您好，感谢您提供的框架。关于交通预测问题，或许您可以考虑添加上PEMS07(M)和PEMS07(L)这两个数据集？

Of course, could you provide me the source of these two datasets? E.g., links or papers.

[huangst21]

您好，感谢您提供的框架。关于交通预测问题，或许您可以考虑添加上PEMS07(M)和PEMS07(L)这两个数据集？

Of course, could you provide me the source of these two datasets? E.g., links or papers.

STGCN的库里包含了这两个速度数据集： https://github.com/VeritasYin/STGCN_IJCAI-18

[littleheuer]

How to set AMP?

[zezhishao]

How to set AMP?

Currently, BasicTS does not support AMP because these baselines don’t have many parameters.
However, I have also recently tested AMP in BasicTS to facilitate training and inference of large time series models. AMP will be added in the future (probably next version).

If you are want to use AMP, you need to init torch.amp.autocast and torch.cuda.amp.GradScaler in the __init__ function of the runner, and then modify the train_iters, val_iters, test, and backward function.

[WaldenBlues]

It seems that STD-MAE [2024-LJCAI] have mentioned this library. Maybe it time to put that into this repo?