saforem2 / l2hmc

Modified TensorFlow implementation for training MCMC samplers on Lattice Gauge Theory models from the paper: Generalizing Hamiltonian Monte Carlo with Neural Network

DEPRECATED

Moved to: l2hmc-qcd

---

L2HMC: Automatic Training of MCMC Samplers

TensorFlow open source implementation for training MCMC samplers from the paper:

Generalizing Hamiltonian Monte Carlo with Neural Networks

by Daniel Levy, Matt D. Hoffman and Jascha Sohl-Dickstein

---

Given an analytically described distributions (implemented as in utils/distributions.py), L2HMC enables training of fast-mixing samplers. We provide an example, in the case of the Strongly-Correlated Gaussian, in the notebook SCGExperiment.ipynb --other details are included in the paper.

---

Forked implementation for Lattice Gauge Theory models.

Forked from original version at brain-research/l2hmc/. The focus of this implementation is on applying the L2HMC algorithm to lattice gauge theory models. Current implementation includes U(1) model.

Additionally, this implementation includes a convolutional neural network architecture that is prepended to the network described in the original paper. The purpose of this additional structure is to better incorporate information about the geometry of the lattice.

Lattice code can be found in l2hmc/lattice/ with the implementation of gauge models in l2hmc/lattice/lattice.py.

Contact

(Original) Code author: Daniel Levy

(Modified) Code author: Sam Foreman

Pull requests and issues for original code: @daniellevy

Pull requests and issues with forked code: @saforem2

Citation

If you use this code, please cite our paper:

@article{levy2017generalizing,
  title={Generalizing Hamiltonian Monte Carlo with Neural Networks},
  author={Levy, Daniel and Hoffman, Matthew D. and Sohl-Dickstein, Jascha},
  journal={International Conference on Learning Representations},
  year={2018}
}

Note

This is not an official Google product.