Awesome

AR-Net

A simple auto-regressive Neural Network for time-series (link to paper).

Install

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 .

Use

View the notebook example_notebooks/arnet.ipynb for an example of how to use the model.

Versions

Current (1.2)

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

Pure PyTorch (0.1)

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.

Now also part of NeuralProphet

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.