Application of the L2HMC algorithm to simulations in lattice QCD.

Overview

l2hmc-qcd CodeFactor

📊 Slides

📒 Example Notebook


Overview

The L2HMC algorithm aims to improve upon HMC by optimizing a carefully chosen loss function which is designed to minimize autocorrelations within the Markov Chain, thereby improving the efficiency of the sampler.

This work is based on the original implementation: brain-research/l2hmc/.

A detailed description of the L2HMC algorithm can be found in the paper:

Generalizing Hamiltonian Monte Carlo with Neural Network

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

Broadly, given an analytically described target distribution, π(x), L2HMC provides a statistically exact sampler that:

  • Quickly converges to the target distribution (fast burn-in).
  • Quickly produces uncorrelated samples (fast mixing).
  • Is able to efficiently mix between energy levels.
  • Is capable of traversing low-density zones to mix between modes (often difficult for generic HMC).

L2HMC for LatticeQCD

Goal: Use L2HMC to efficiently generate gauge configurations for calculating observables in lattice QCD.

A detailed description of the (ongoing) work to apply this algorithm to simulations in lattice QCD (specifically, a 2D U(1) lattice gauge theory model) can be found in doc/main.pdf.

l2hmc-qcd poster

Organization

Dynamics / Network

The base class for the augmented L2HMC leapfrog integrator is implemented in the BaseDynamics (a tf.keras.Model object).

The GaugeDynamics is a subclass of BaseDynamics containing modifications for the 2D U(1) pure gauge theory.

The network is defined in l2hmc-qcd/network/functional_net.py.

Network Architecture

An illustration of the leapfrog layer updating (x, v) --> (x', v') can be seen below.

leapfrog layer

Lattice

Lattice code can be found in lattice.py, specifically the GaugeLattice object that provides the base structure on which our target distribution exists.

Additionally, the GaugeLattice object implements a variety of methods for calculating physical observables such as the average plaquette, ɸₚ, and the topological charge Q,

Training

The training loop is implemented in l2hmc-qcd/utils/training_utils.py .

To train the sampler on a 2D U(1) gauge model using the parameters specified in bin/train_configs.json:

$ python3 /path/to/l2hmc-qcd/l2hmc-qcd/train.py --json_file=/path/to/l2hmc-qcd/bin/train_configs.json

Or via the bin/train.sh script provided in bin/.

Features

  • Distributed training (via horovod): If horovod is installed, the model can be trained across multiple GPUs (or CPUs) by:

    #!/bin/bash
    
    TRAINER=/path/to/l2hmc-qcd/l2hmc-qcd/train.py
    JSON_FILE=/path/to/l2hmc-qcd/bin/train_configs.json
    
    horovodrun -np ${PROCS} python3 ${TRAINER} --json_file=${JSON_FILE}

Contact


Code author: Sam Foreman

Pull requests and issues should be directed to: saforem2

Citation

If you use this code or found this work interesting, please cite our work along with the original paper:

@misc{foreman2021deep,
      title={Deep Learning Hamiltonian Monte Carlo}, 
      author={Sam Foreman and Xiao-Yong Jin and James C. Osborn},
      year={2021},
      eprint={2105.03418},
      archivePrefix={arXiv},
      primaryClass={hep-lat}
}
@article{levy2017generalizing,
  title={Generalizing Hamiltonian Monte Carlo with Neural Networks},
  author={Levy, Daniel and Hoffman, Matthew D. and Sohl-Dickstein, Jascha},
  journal={arXiv preprint arXiv:1711.09268},
  year={2017}
}

Acknowledgement

This research used resources of the Argonne Leadership Computing Facility, which is a DOE Office of Science User Facility supported under contract DE_AC02-06CH11357. This work describes objective technical results and analysis. Any subjective views or opinions that might be expressed in the work do not necessarily represent the views of the U.S. DOE or the United States Government. Declaration of Interests - None.

Hits

Stargazers over time

Comments
  • Remove upper bound on python_requires

    Remove upper bound on python_requires

    (I'm moving between meetings so can iterate on this more later, so excuse the very brief Issue for now).

    At the moment the project has an upper bound on python_requires

    https://github.com/saforem2/l2hmc-qcd/blob/2eb6ee63cc0c53b187e6d716f4c12f418c8b8515/setup.py#L165

    Assuming that you're intending l2hmc to be a library and not an application, then I would highly recommend removing this for the reasons summarized in Henry's detailed blog post on the subject.

    Congrats on getting l2hmc up on PyPI though! :snake: :rocket:

    opened by matthewfeickert 2
  • Alpha

    Alpha

    Pull upstream alpha branch into main

    Major changes

    • new src/ hierarchical module organization
    • Contains skeleton implementation of 4D SU(3) lattice gauge model
    • Framework independent configuration
      • Unified configuration system simplifies logic, same configs used for both tensorflow and pytorch experiments
      • Plan to be able to specify which backend to use through config option
    • Unified (and framework independent) configurations between tensorflow and pytorch implementations

    Note: This is still very much a WIP. Many existing features still need to be re-implemented / updated into new code in src/.

    Todo

    • [ ] Write unit tests
    • [ ] Use simple configs for end-to-end workflow test + integrate into CI
    • [ ] dynamic learning rate scheduling
    • [ ] Test 4D SU(3) numpy code
    • [ ] Write tensorflow and pytorch implementations of LatticeSU3 objects
    • [ ] Improved / simplified ( / trainable?) annealing schedule
    • [ ] Distributed training support
      • [ ] horovod
      • [ ] DDP for pytorch implementation
      • [ ] DeepSpeed from Microsoft??
    • [ ] Testing / inference logic
    • [ ] Automatic checkpointing
    • [ ] Metric logging
      • [ ] Tensorboard?
      • [ ] Sacred?
      • [ ] build custom dashboard? plot.ly?
    • [ ] Setup packaging / distribution through pip
    • [ ] Resolve issue
    opened by saforem2 1
  • Alpha

    Alpha

    opened by saforem2 1
  • Rich

    Rich

    General improvements, rewrote logging methods to use Rich for better formatting.

    • Adds dynamic (trainable) step size eps for each separate x and v updates, seems to generally increase the total energy towards the middle of the trajectory but it remains unclear if this corresponds to an improvement in the tunneling rate
    • Adds methods for calculating autocorrelations of the topological charge, as well as notebooks for generating the plots
    • Updates to the writeup in doc/main.pdf
    • Will likely be last changes to writeup before public release of official draft
    opened by saforem2 1
  • Dev

    Dev

    • Updates to README

    • Ability to load network with new training instance

    • Updates to doc/, removes old sections related to debugging the bias in the plaquette

    opened by saforem2 1
  • Saveable model

    Saveable model

    Complete rewrite of dynamics.xnet and dynamics.vnet models to use tf.keras.functional Models.

    Additional changes include:

    • Non-Compact Projection update for gauge fields
    • Ability to specify convolution structure to be prepended at beginning of gauge network
    opened by saforem2 1
  • Dev

    Dev

    Removes models/gauge_model.py entirely.

    Instead, a base dynamics class is implemented in dynamics/dynamics.py, and an example subclass is provided in dynamics/gauge_dynamics.py.

    opened by saforem2 1
  • Split networks

    Split networks

    Major rewrite of existing codebase.

    This pull request updates everything to be compatible with tensorflow >= 2.2 and removes a bunch of redundant legacy code.

    opened by saforem2 1
  • Dev

    Dev

    • Dynamics object is now compatible with tf >= 2.0
    • Running inference on trained model with tensorflow now creates identical graphs and summary files to numpy inference code
    • Inference with numpy now uses object oriented structure
    • Adds LaTeX + PDF documentation in doc/
    opened by saforem2 1
  • Cooley dev

    Cooley dev

    Adds new GaugeNetwork architecture as the default for training GaugeModel

    Additionally, replaces pickle with joblib for saving data as .z compressed files (as opposed to .pkl files).

    opened by saforem2 1
  • Testing

    Testing

    Implemented nnehmc_loss calculation for an alternative loss function using the approach suggested in https://infoscience.epfl.ch/record/264887/files/robust_parameter_estimation.pdf.

    This modified loss function can be chosen (instead of the standard loss described in the original paper) by passing --use_nnehmc_loss as a command line argument.

    opened by saforem2 1
  • Packaging and PyPI distribution?

    Packaging and PyPI distribution?

    As you've made a library and are using it as such:

    # snippet from toy_distributions.ipynb
    
    # append parent directory to `sys.path`
    # to load from modules in `../l2hmc-qcd/`
    module_path = os.path.join('..')
    if module_path not in sys.path:
        sys.path.append(module_path)
    
    # Local imports
    from utils.attr_dict import AttrDict
    from utils.training_utils import train_dynamics
    from dynamics.config import DynamicsConfig
    from dynamics.base_dynamics import BaseDynamics
    from dynamics.generic_dynamics import GenericDynamics
    from network.config import LearningRateConfig
    from config import (State, NetWeights, MonteCarloStates,
                        BASE_DIR, BIN_DIR, TF_FLOAT)
    
    from utils.distributions import (plot_samples2D, contour_potential,
                                     two_moons_potential, sin_potential,
                                     sin_potential1, sin_potential2)
    

    do you have any plans and/or interest in packaging it as a Python library so it can either be pip installed from GitHub or be distributed on PyPI?

    opened by matthewfeickert 5
Releases(0.12.0)
Owner
Sam Foreman
Computational science Postdoc at Argonne National Laboratory working on applying machine learning to simulations in lattice QCD.
Sam Foreman
NeuroFind - A solution to the to the Task given by the Oberseminar of Messtechnik Institute of TU Dresden in 2021

NeuroFind A solution to the to the Task given by the Oberseminar of Messtechnik

1 Jan 20, 2022
Generic template to bootstrap your PyTorch project with PyTorch Lightning, Hydra, W&B, and DVC.

NN Template Generic template to bootstrap your PyTorch project. Click on Use this Template and avoid writing boilerplate code for: PyTorch Lightning,

Luca Moschella 520 Dec 30, 2022
Learning Domain Invariant Representations in Goal-conditioned Block MDPs

Learning Domain Invariant Representations in Goal-conditioned Block MDPs Beining Han, Chongyi Zheng, Harris Chan, Keiran Paster, Michael R. Zhang, Jim

Chongyi Zheng 3 Apr 12, 2022
[BMVC'21] Official PyTorch Implementation of Grounded Situation Recognition with Transformers

Grounded Situation Recognition with Transformers Paper | Model Checkpoint This is the official PyTorch implementation of Grounded Situation Recognitio

Junhyeong Cho 18 Jul 19, 2022
🍅🍅🍅YOLOv5-Lite: lighter, faster and easier to deploy. Evolved from yolov5 and the size of model is only 1.7M (int8) and 3.3M (fp16). It can reach 10+ FPS on the Raspberry Pi 4B when the input size is 320×320~

YOLOv5-Lite:lighter, faster and easier to deploy Perform a series of ablation experiments on yolov5 to make it lighter (smaller Flops, lower memory, a

pogg 1.5k Jan 05, 2023
Improving Convolutional Networks via Attention Transfer (ICLR 2017)

Attention Transfer PyTorch code for "Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Tran

Sergey Zagoruyko 1.4k Dec 23, 2022
Code, Models and Datasets for OpenViDial Dataset

OpenViDial This repo contains downloading instructions for the OpenViDial dataset in 《OpenViDial: A Large-Scale, Open-Domain Dialogue Dataset with Vis

119 Dec 08, 2022
Code for the IJCAI 2021 paper "Structure Guided Lane Detection"

SGNet Project for the IJCAI 2021 paper "Structure Guided Lane Detection" Abstract Recently, lane detection has made great progress with the rapid deve

Jinming Su 27 Dec 08, 2022
Single object tracking and segmentation.

Single/Multiple Object Tracking and Segmentation Codes and comparison of recent single/multiple object tracking and segmentation. News 💥 AutoMatch is

ZP ZHANG 385 Jan 02, 2023
Coursera - Quiz & Assignment of Coursera

Coursera Assignments This repository is aimed to help Coursera learners who have difficulties in their learning process. The quiz and programming home

浅梦 828 Jan 04, 2023
A flexible submap-based framework towards spatio-temporally consistent volumetric mapping and scene understanding.

Panoptic Mapping This package contains panoptic_mapping, a general framework for semantic volumetric mapping. We provide, among other, a submap-based

ETHZ ASL 194 Dec 20, 2022
Exploring the link between uncertainty estimates obtained via "exact" Bayesian inference and out-of-distribution (OOD) detection.

Uncertainty-based OOD detection Exploring the link between uncertainty estimates obtained by "exact" Bayesian inference and out-of-distribution (OOD)

Christian Henning 1 Nov 05, 2022
Making self-supervised learning work on molecules by using their 3D geometry to pre-train GNNs. Implemented in DGL and Pytorch Geometric.

3D Infomax improves GNNs for Molecular Property Prediction Video | Paper We pre-train GNNs to understand the geometry of molecules given only their 2D

Hannes Stärk 95 Dec 30, 2022
Experimental code for paper: Generative Adversarial Networks as Variational Training of Energy Based Models

Experimental code for paper: Generative Adversarial Networks as Variational Training of Energy Based Models, under review at ICLR 2017 requirements: T

Shuangfei Zhai 18 Mar 05, 2022
KaziText is a tool for modelling common human errors.

KaziText KaziText is a tool for modelling common human errors. It estimates probabilities of individual error types (so called aspects) from grammatic

ÚFAL 3 Nov 24, 2022
A python script to convert images to animated sus among us crewmate twerk jifs as seen on r/196

img_sussifier A python script to convert images to animated sus among us crewmate twerk jifs as seen on r/196 Examples How to use install python pip i

41 Sep 30, 2022
TargetAllDomainObjects - A python wrapper to run a command on against all users/computers/DCs of a Windows Domain

TargetAllDomainObjects A python wrapper to run a command on against all users/co

Podalirius 19 Dec 13, 2022
A library for graph deep learning research

Documentation | Paper [JMLR] | Tutorials | Benchmarks | Examples DIG: Dive into Graphs is a turnkey library for graph deep learning research. Why DIG?

DIVE Lab, Texas A&M University 1.3k Jan 01, 2023
Code for our paper at ECCV 2020: Post-Training Piecewise Linear Quantization for Deep Neural Networks

PWLQ Updates 2020/07/16 - We are working on getting permission from our institution to release our source code. We will release it once we are granted

54 Dec 15, 2022