🚀 LTSF: A Baseline Aggregator for Long-Term Time-Series Forecasting

LTSF is a Python 🐍 package that simplifies the process of implementing and testing various state-of-the-art models 🧠 for long-term time-series forecasting 📈. This package supports the majority of current leading baselines, offering a user-friendly interface to perform LTSF tasks effortlessly.

LTSF enables users to employ their desired model with a single-line configuration and provides the tools 🛠️ to create training, validation, and testing datasets effortlessly. By further utilizing LTSF's training interface, users can train and test their models using only a few lines of code.

🔧 Supported Baselines

The table below shows the supported baselines with their corresponding references and links to their original implementation:

Baseline	Reference	Code
PatchTST	ICLR 2023	PatchTST
MICN	ICLR 2023	MICN
FiLM	NIPS 2022	FiLM
TimesNet	ICLR 2023	TimesNet
Crossformer	ICLR 2023	Crossformer
DLinear	AAAI 2023	DLinear
LightTS	arXiv 2022	LightTS
ETSformer	arXiv 2022	ETSformer
Non-stationary Transformer	NeurIPS 2022	Non-stationary Transformer
FEDformer	ICML 2022	FEDformer
Pyraformer	ICLR 2022	Pyraformer
Autoformer	NeurIPS 2021	Autoformer
Informer	AAAI 2021	Informer
Reformer	ICLR 2020	Reformer
Transformer	NeurIPS 2017	Transformer

📚 Usage

Here is an example of how you can use LTSF:

pip install ltsf

import ltsf

# Create a configuration for your desired model and dataset
config = ltsf.Config("Autoformer", "ETTh1") 

# Set your custom configuration
config.set_config({"use_gpu":False})

# Create training, validation, and testing datasets
train_loader, val_loader, test_loader = ltsf.DatasetFactory.create(config, download=True, data_path=".")  

# Create an LTSFTrainer
trainer = ltsf.LTSFTrainer(config) 

# Start the training process
trainer.train(train_loader, val_loader, test_loader)

# Test the trained model
trainer.test(test_loader, res_dir="./result")

⭐ Key Features

• Supports 15 leading baselines for long-term time-series forecasting.
• User-friendly, enabling model training and testing with minimal lines of code.

🎉 Update Notes

• [2023-06-21] LTSF has now been released! 🚀

📈 Future Work

• Planning to support more intelligent parameter settings for improved usability and flexibility.
• Support custom dataset.

🙏 Acknowledgments

We are grateful to the authors of all the papers and their original implementations that made this package possible.

📜 License

Apache License.