/VIND

Variational Inference with Numerical Derivatives: variance reduction through coupling

Primary LanguagePython

VIND

Variational Inference using Numerical Differentiation for Non-Reparameterisable Parameters

Dependencies

Python 3.6 with following packages

numpy, scipy, scikit-learn, pytorch, matplotlib, sacred, pandas, tqdm

For reproducing experiments, please run the following files using python {filename} and subsequently plot the used figures with python create_plots.py.

  • linear_regression.py
  • mse_grad_gamma.py
  • wishart_student_normal.py
  • wishart_normal_normal.py

For the stationarity test on the included data, use R to run stationarity.R.

Data

Used standard ML benchmark data sets as well as openly accessible historical data from yahoo finance. For more information on the individual files used, see here.