Hamiltonian Dynamics with Non-Newtonian Momentum for Rapid Sampling
Code for the paper:
Greg Ver Steeg and Aram Galstyan. "Hamiltonian Dynamics with Non-Newtonian Momentum for Rapid Sampling", NeurIPS 2021. [arxiv] [bibtex]
Non-Newtonian Momentum Animation:
This repo contains code for implementing Energy Sampling Hamiltonian Dynamics, so-called because the Hamiltonian dynamics with this special form of Non-Newtonian momentum ergodically samples from a target un-normalized density specified by an energy function.
Requirements
The core ESH dynamics sampler code (import esh) uses only PyTorch.
python -m pip install git+https://github.com/gregversteeg/esh_dynamics
Use pip install -r requirements.txt
to install requirements for all comparison code.
Usage
Here's a small example where we load a pytorch energy function, then sample Langevin versus ESH trajectories.
import torch as t
import esh # ESH Dynamics integrator
from esh.datasets import ToyDataset # Example energy models
from esh.samplers import hmc_integrate # Sampling comparison methods, like Langevin
# Energy to sample - any pytorch function/module that outputs a scalar per batch item
energy = ToyDataset(toy_type='gmm').energy # Gaussian mixture model
epsilon = 0.01 # Step size should be < 1
n_steps = 100 # Number of steps to take
x0 = t.tensor([[0., 0.5]]) # Initial state, size (batch_size, ...)
xs, vs, rs = esh.leap_integrate_chain(energy, x0, n_steps, epsilon, store=True) # "Store" returns whole trajectory
xs_ula, vs_ula, _ = hmc_integrate(energy, x0, n_steps, epsilon=epsilon, k=1, mh_reject=False) # Unadjusted Langevin Alg
To get just the last state instead of the whole trajectory, set store=False. To do ergodic reservoir sampling, set reservoir=True, store=False.
Generating figures
See the README in the generate_figures
for scripts to generate each figure in the paper, and to see more example usage.
BibTeX
@inproceedings{esh,
title={Hamiltonian Dynamics with Non-Newtonian Momentum for Rapid Sampling},
author={Greg {Ver Steeg} and Aram Galstyan},
Booktitle={Advances in Neural Information Processing Systems},
year={2021}
}