A simple auto-regressive Neural Network for time-series (link to paper).
After downloading the code repository (via git clone), change to the repository directory (cd AR-Net)
and install arnet as python package with pip install .
View the notebook example_notebooks/arnet.ipynb for an example of how to use the model.
The version 1.0 made the model easier to use with your own datasets and requires less hyperparameters for a simpler training procedure. It is built on the fastai library.
Changes (1.1 -> 1.2):
- simplified UI with ARNet as object
- GPU support
- robustified training
- added test cases
- updated example notebooks
Changes (1.0 -> 1.1):
- port beta fastai2 to it's current stable release
- make install as pip package possible
- add black code formatter (and git pre-commit hook)
- add unittests (and git pre-push hook)
- fix issues with new fastai api
- remove old code fragments
Version 0.1 was based on Pytorch and you can still use it if you do not want to use fastai.
See file v0_1/example.py for how to use the v0.1 model.
AR-Net is now part of a more comprehensive package NeuralProphet.
I strongly recommend using it instead of the standalone version, unless you specifically want to use AR-Net, which may make sense if you need to model a highly-autoregressive time-series with sparse long-range dependencies.