This notebook contains the samplers from the papers "Stochastic Gradient MCMC with Repulsive Forces" and "Accelerating Stochastic Gradient Markov Chain Monte Carlo with Momentum and Repulsive Forces" (to appear soon).

All the samplers are implemented in jax. You can open the notebooks and then run it in Colab.

The implemented samplers/optimizers are located in Samplers_jax.ipynb. The implemented ones are:

• SVGD (Stein Variational Gradient Descend)

• SGLD+R (Stochastic Gradient Langevin Dynamics plus Repulsion)

• SGD (Stochastic Gradient Descent)

• SGLD (Stochastic Gradient Langevin Dynamics)

• SGDm (SGD plus Momentum)

• SGDm+R (SGD plus Momentum and Repulsion)

All the samplers are vectorized and use jit for increased performance, with an emphasis on simplicity. The notebook constains a standard Gaussian as the target distribution.

See the notebook Gaussian_example_jax.ipynb for a comparison between SVGD and SGLD+R.