PatchTST: "A Time Series is Worth 64 Words: Long-term Forecasting with Transformers." (official repo: https://github.com/yuqinie98/PatchTST)

Updates:

2023-01: Our paper was accepted to ICLR 2023! The new official repo is: https://github.com/yuqinie98/PatchTST.

Nov 29th: update few instructions and include other baseline models in supervised learning for comparison.

Nov 16th: initial repo created.

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

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

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.

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.

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>

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