Sohl-Dickstein / LAHMC

Look Ahead Hamiltonian Monte Carlo

Look Ahead Hamiltonian Monte Carlo

Implements Look Ahead Hamiltonian Monte Carlo (LAHMC) and standard Hamiltonian Monte Carlo (HMC) in both Python and MATLAB.

LAHMC is described in the paper:

Sohl-Dickstein, Jascha and Mudigonda, Mayur and DeWeese, Michael R.
Hamiltonian Monte Carlo Without Detailed Balance.
International Conference on Machine Learning. 2014
http://arxiv.org/abs/1409.5191

Example Python Code

The following code draws samples from an isotropic Gaussian distribution using LAHMC.

from LAHMC import LAHMC
import numpy as np

# Define the energy function and gradient
def E(X, sigma=1.):
    """ Energy function for isotropic Gaussian """
    return np.sum(X**2, axis=0).reshape((1,-1))/2./sigma**2
def dEdX(X, sigma=1.):
    """ Energy function gradient for isotropic Gaussian """
    return X/sigma**2

# Initialize the sample locations -- 2 dimensions, 100 particles
Xinit = np.random.randn(2,100)

# initialize the sampler.
sampler = LAHMC(Xinit, E, dEdX, epsilon=0.1, beta=0.1, kwargs={'sigma':0.1})
# perform 10 sampling steps for all 100 particles
X = sampler.sample(num_steps = 10)
# perform another 10 sampling steps
X = sampler.sample(num_steps = 10)

More detailed documentation, and additional options, can be found in python/LAHMC.py

Example MATLAB Code

The following code draws samples from an isotropic Gaussian distribution using LAHMC.

% opts holds all parameters which will be passed to the sampler
opts = [];
opts.epsilon = 0.1;
opts.beta = 0.1;
% number of sampling steps
opts.T = 10;
% energy function and gradient
opts.E = @E_gauss;
opts.dEdX = @dEdX_gauss;

% state will hold the particle positions and velocities between
% sampler calls, as well as counters for the number of transitions
% and function evaluations
state = []

% Initialize sample locations -- 2 dimensions, 100 particles
opts.Xinit = randn(2,100);
% Gaussian coupling matrix expected by E_gauss and dEdX_gauss
J = eye(2)*100;

% perform 10 sampling steps for all 100 particles
[X, state] = LAHMC(opts, state, J);
% perform another 10 sampling steps
[X, state] = LAHMC(opts, state, J);

More detailed documentation, and additional options, can be found in matlab/LAHMC.m.

Reproduce Figure from the Paper

Code reproducing Figure 2 and Table 1 of the paper, and demonstrating usage of the sampler, can be found in python/generate_figure_2.py and matlab/generate_figure_2.m. The exact plots appearing in the paper were generated using the MATLAB version of the code.