AutoFormer is a powerful neural network designed for forecasting multiple steps ahead. It consists of two main parts: an encoder and a decoder. These components are specially crafted for handling time series data. The first part uses decomposition layers, a common technique in time series analysis, to understand complex temporal patterns. The second part features an attention mechanism tailored for time series, employing Fourier Transform (FFT) and autocorrelation processes. This unique approach helps identify important parts in a time series, unlike traditional attention mechanisms that focus on individual points. To enhance this process, AutoFormer incorporates time delay aggregations, making it adept at handling intricate time dependencies.
- Speedup in Time Delays Aggregation:
- This project includes the speeding up version for the time delays aggregation process
- during both training and inference phases, as described in the paper.
Here are some example results achieved using AutoFormer for 4 days ahead forecasting:
To run the project using Docker, follow these steps:
-
Make sure you have Docker installed on your machine. If not, you can download it here.
-
Build the Docker image with a specific name:
docker build -t autoformer . -
Run the Docker container:
docker run -v /host/path/for/trained_models:/autoformer/trained_models \ -v host/path/for/results:/autoformer/results \ -v host/path/for/dataset:/autoformer/dataset \ autoformer
If you prefer to install dependencies using pip, you can also use the requirements.txt file:
pip install -r requirements.txt
This project is licensed under the MIT License - see the LICENSE.md file for details.
This project relies on the work described in the following paper: Haixu Wu and Jiehui Xu and Jianmin Wang and Mingsheng Long. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting, Advances in Neural Information Processing Systems, 2021
If you use or reproduce results from this project that are based on the aforementioned paper, please make sure to appropriately cite the original work.