Kernel Stein Discrepancy Descent

Sampling by optimization of the Kernel Stein Discrepancy

The paper is available at https://arxiv.org/abs/2105.09994.

The code uses Pytorch, and a numpy backend is available for SVGD.

ksd_picture

The package also contains the code for the celebrated Stein Variational Gradient Descent algorithm and Maximum Mean Discrepancy gradient flow.

Install

The code is available on pip:

$ pip install ksddescent

The github repository is at github.com/pierreablin/ksddescent

Example

The main function is ksdd_lbfgs, which uses the fast L-BFGS algorithm to converge quickly. It takes as input the initial position of the particles, and the score function. For instance, to samples from a Gaussian (where the score is identity), you can use these simple lines of code:

>>> import torch
>>> from ksddescent import ksdd_lbfgs
>>> n, p = 50, 2
>>> x0 = torch.rand(n, p)  # start from uniform distribution
>>> score = lambda x: x  # simple score function
>>> x = ksdd_lbfgs(x0, score)  # run the algorithm

Reference

If you use this code in your project, please cite:

Anna Korba, Pierre-Cyril Aubin-Frankowski, Simon Majewski, Pierre Ablin
Kernel Stein Discrepancy Descent
International Conference on Machine Learning, 2021

Bug reports

Use the github issue tracker to report bugs.