Deep4cast: Forecasting for Decision Making under Uncertainty
This package is under active development. Things may change :-).
Deep4Cast is a scalable machine learning package implemented in Python. It has a front-end API designed to be similar to scikit-learn but is based on Pytorch. It was designed for medium to large size time series data sets and allows for modeling of forecast uncertainties. The network architecture is based on DeepMind's WaveNet. Regularization and approximate sampling from posterior predictive distributions of forecasts are achieved via Concrete Dropout.
Package documentation under construction. Please see example notebooks for instructions.
Installation
Source
From the package directory install the requirements and then the package (best in a clean virtual environment)
$ pip install -r requirements.txt
$ python setup.py install
Main Requirements
Examples
Authors:
- Toby Bischoff
- Austin Gross
- Shirley Ren
- Kenneth Tran
References:
- Concrete Dropout is used for approximate posterior Bayesian inference.
- Wavenet is used as encoder network.