This is a Pytorch implementation of DLinear: "Are Transformers Effective for Time Series Forecasting?".
- Support both Univariate and Multivariate long-term time series forecasting.
- Support visualization of weights.
- Support scripts on different look-back window size.
Beside DLinear, we provide five significant forecasting Transformers to re-implement the results in the paper.
- Transformer (NeuIPS 2017)
- Informer (AAAI 2021 Best paper)
- Autoformer (NeuIPS 2021)
- Pyraformer (ICLR 2022 Oral)
- FEDformer (ICML 2022)
We provide all experiment script files in ./scripts:
| Files | Interpretation |
|---|---|
| EXP-LongForecasting | Long-term Time Series Forecasting Task |
| EXP-LookBackWindow | Study the impact of different look-back window size |
| EXP-Embedding | Study the impact of different embedding strategies |
This code is simply build on the code base of Autoformer. We appreciate the following github repos a lot for their valuable code base or datasets:
The implementation of Autoformer, Informer, Transformer is from https://github.com/thuml/Autoformer
The implementation of FEDformer is from https://github.com/MAZiqing/FEDformer
The implementation of Pyraformer is from https://github.com/alipay/Pyraformer
Although DLinear is simple, it has some compelling characteristics:
- An O(1) maximum signal traversing path length: The shorter the path, the better the dependencies are captured, making DLinear capable of capturing both short-range and long-range temporal relations.
- High-efficiency: As each branch has only one linear layer, it costs much lower memory and fewer parameters and has a faster inference speed than existing Transformers.
- Interpretability: After training, we can visualize weights from the seasonality and trend branches to have some insights on the predicted values.
- Easy-to-use: DLinear can be obtained easily without tuning model hyper-parameters.
In Multivariate long sequence time-series forecasting(left table), DLinear outperforms FEDformer by over 40% on Exchange rate, around 30% on Traffic, Electricity, and Weather, and around 25% on ETTm1.
In Univariate long sequqence time-series forecasting(right table), DLinear outperforms transformer-based methods in most cases.
Comparison of method efficiency on the Electricity dataset with a look-back window size of 96 and forecasting horizon of 720 steps. MACs are the number of multiply-accumulate operations. The inference time is an average result of 5 runs.
First, please make sure you have installed Conda. Then, our environment can be installed by:
conda create -n DLinear python=3.6.9
conda activate DLinear
pip install -r requirements.txt
You can obtain all the nine benchmarks from Google Drive provided in Autoformer. All the datasets are well pre-processed and can be used easily.
mkdir dataset
Please put them in the ./dataset directory
- In
scripts/, we provide the model implementation Dlinear/Autoformer/Informer/Transformer - In
FEDformer/scripts/, we provide the FEDformer implementation - In
Pyraformer/scripts/, we provide the Pyraformer implementation
For example:
To train the DLinear on Exchange-Rate dataset, you can use the scipt scripts/EXP-LongForecasting/DLinear/exchange_rate.sh:
sh scripts/EXP-LongForecasting/DLinear/exchange_rate.sh
It will start to train DLinear, the results will be shown in logs/LongForecasting.
All scripts about using DLinear on long forecasting task is in scripts/EXP-LongForecasting/DLinear/, you can run them in a similar way. The default look-back window in scripts is 96, DLinear generally achieves better results with longer look-back window as dicussed in the paper. For instance, you can simpy change the seq_len (look-back window size) in scripts to 336 to obtain better performance.
Scripts about look-back window size and long forecasting of FEDformer and Pyraformer is in FEDformer/scripts and Pyraformer/scripts, respectively. To run them, you need to first cd FEDformer or cd Pyraformer. Then, you can use sh to run them in a similar way. Logs will store in logs/.
Each experiment in scripts/EXP-LongForecasting/DLinear/ takes 5min-20min. For other Transformer scripts, since we put all related experiments in one script file, directly running them will take 8 hours-1 day. You can keep the experiments you interested in and comment out the others.
As shown in our paper, the weights of DLinear can reveal some charateristic of the data, i.e., the periodicity. We provide the weight visualization of DLinear in weight_plot.py. To run the visualization, you need to input the model path (model_name) of DLinear (the model directory in ./checkpoint by default).
If you find this repository useful for your work, please consider citing it as follows:
@article{Zeng2022AreTE,
title={Are Transformers Effective for Time Series Forecasting?},
author={Ailing Zeng and Muxi Chen and Lei Zhang and Qiang Xu},
journal={arXiv preprint arXiv:2205.13504},
year={2022}
}Please remember to cite all the datasets and compared methods if you use them in your experiments.