tft-torch

tft-torch is a Python library that implements "Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting" using pytorch framework. The library provides a complete implementation of a time-series multi-horizon forecasting model with state-of-the-art performance on several benchmark datasets.

This library works for Python 3.7 and higher and PyTorch 1.6.0 and higher.

Installation

This library is distributed on PyPi and can be installed using pip.

$ pip install tft-torch 

The command above will automatically install all the required dependencies. Please visit the installation page for more details.

Getting started

Check out the tutorials for a demonstration how to use the library.

Documentation

For more information, refer to our blogpost and complete documentation.

Reference

This repository suggests an implementation of a model based on the work presented in the following paper:

@misc{lim2020temporal,
      title={Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting}, 
      author={Bryan Lim and Sercan O. Arik and Nicolas Loeff and Tomas Pfister},
      year={2020},
      eprint={1912.09363},
      archivePrefix={arXiv},
      primaryClass={stat.ML}
}

Some parts of the implementation rely on mattsherar's TFT implementation, available as part of the Temporal_Fusion_Transform repository.

Info for developers

The source code of the project is available on GitHub.

$ git clone https://github.com/PlaytikaOSS/tft-torch.git

You can install the library and the dependencies with one of the following commands:

$ pip install .                        # install library + dependencies
$ pip install ".[develop]"             # install library + dependencies + developer-dependencies
$ pip install -r requirements.txt      # install dependencies
$ pip install -r requirements-dev.txt  # install developer-dependencies

For creating the "pip-installable" *.whl file, run the following command (at the root of the repository):

$ python -m build

For creating the HTML documentation of the project, run the following commands:

$ cd docs
$ make clean
$ make html

Run tests

Tests can be executed with pytest running the following commands:

$ cd tests
$ pytest                                      # run all tests
$ pytest test_testmodule.py                   # run all tests within a module
$ pytest test_testmodule.py -k test_testname  # run only 1 test

License

MIT License