PatchTST (ICLR 2023)

This is an offical implementation of PatchTST: A Time Series is Worth 64 Words: Long-term Forecasting with Transformers.

🚩 Our model has been included in GluonTS. Special thanks to the contributor @kashif!

🚩 Our model has been included in NeuralForecast. Special thanks to the contributor @kdgutier and @cchallu!

🚩 Our model has been included in timeseriesAI(tsai). Special thanks to the contributor @oguiza!

We offer a video that provides a concise overview of our paper for individuals seeking a rapid comprehension of its contents: https://www.youtube.com/watch?v=Z3-NrohddJw

Key Designs

🌟 Patching: segmentation of time series into subseries-level patches which are served as input tokens to Transformer.

🌟 Channel-independence: each channel contains a single univariate time series that shares the same embedding and Transformer weights across all the series.

Results

Supervised Learning

Compared with the best results that Transformer-based models can offer, PatchTST/64 achieves an overall 21.0% reduction on MSE and 16.7% reduction on MAE, while PatchTST/42 attains a overall 20.2% reduction on MSE and 16.4% reduction on MAE. It also outperforms other non-Transformer-based models like DLinear.

Self-supervised Learning

We do comparison with other supervised and self-supervised models, and self-supervised PatchTST is able to outperform all the baselines.

We also test the capability of transfering the pre-trained model to downstream tasks.

Efficiency on Long Look-back Windows

Our PatchTST consistently reduces the MSE scores as the look-back window increases, which confirms our model’s capability to learn from longer receptive field.

Getting Started

We seperate our codes for supervised learning and self-supervised learning into 2 folders: PatchTST_supervised and PatchTST_self_supervised. Please choose the one that you want to work with.

Supervised Learning

1. Install requirements. pip install -r requirements.txt

2. Download data. You can download all the datasets from Autoformer. Create a seperate folder ./dataset and put all the csv files in the directory.

3. Training. All the scripts are in the directory ./scripts/PatchTST. The default model is PatchTST/42. For example, if you want to get the multivariate forecasting results for weather dataset, just run the following command, and you can open ./result.txt to see the results once the training is done:

sh ./scripts/PatchTST/weather.sh

You can adjust the hyperparameters based on your needs (e.g. different patch length, different look-back windows and prediction lengths.). We also provide codes for the baseline models.

Self-supervised Learning

1. Follow the first 2 steps above

2. Pre-training: The scirpt patchtst_pretrain.py is to train the PatchTST/64. To run the code with a single GPU on ettm1, just run the following command

python patchtst_pretrain.py --dset ettm1 --mask_ratio 0.4

The model will be saved to the saved_model folder for the downstream tasks. There are several other parameters can be set in the patchtst_pretrain.py script.

3. Fine-tuning: The script patchtst_finetune.py is for fine-tuning step. Either linear_probing or fine-tune the entire network can be applied.

python patchtst_finetune.py --dset ettm1 --pretrained_model <model_name>

Acknowledgement

We appreciate the following github repo very much for the valuable code base and datasets:

https://github.com/cure-lab/LTSF-Linear

https://github.com/zhouhaoyi/Informer2020

https://github.com/thuml/Autoformer

https://github.com/MAZiqing/FEDformer

https://github.com/alipay/Pyraformer

https://github.com/ts-kim/RevIN

https://github.com/timeseriesAI/tsai

Contact

If you have any questions or concerns, please contact us: ynie@princeton.edu or nnguyen@us.ibm.com or submit an issue

Citation

If you find this repo useful in your research, please consider citing our paper as follows:

@inproceedings{Yuqietal-2023-PatchTST,
  title     = {A Time Series is Worth 64 Words: Long-term Forecasting with Transformers},
  author    = {Nie, Yuqi and
               H. Nguyen, Nam and
               Sinthong, Phanwadee and 
               Kalagnanam, Jayant},
  booktitle = {International Conference on Learning Representations},
  year      = {2023}
}