pyhuge Python Package

pyhuge is the native Python package for high-dimensional undirected graph estimation and inference, sharing the same C++ core as the R huge package.

Table of contents

• Background
• Directory structure
• What this package provides
• Requirements
• Installation
• Usage
• Documentation
• Developer workflow
• Citation

Background

pyhuge runs natively in Python with a shared C++ backend and does not require an R runtime.

Directory structure

• pyhuge/: package source code
• pyhuge/data/: packaged datasets (stockdata.npz)
• cpp/: optional pybind11 acceleration kernels
• tests/: unit/e2e/parity tests
• examples/: runnable scripts
• docs/: MkDocs documentation pages
• scripts/: release and docs helper scripts

What this package provides

• Core estimators: huge, huge_mb, huge_glasso, huge_ct, huge_tiger
• Model selection: huge_select (ric, stars, ebic)
• Data transforms and utilities: huge_npn, huge_generator, huge_roc, huge_inference
• Plotting helpers: huge_plot_sparsity, huge_plot_roc, huge_plot_graph_matrix, huge_plot_network, huge_plot
• Dataset helper: huge_stockdata
• Diagnostics: pyhuge.test(), pyhuge-doctor

Requirements

• Python >=3.9
• Runtime packages: numpy, scipy
• Native extension: pyhuge._native_core is required for mb, tiger, glasso
• Native source builds: Linux or macOS; Windows is not currently supported

Optional:

• plotting: matplotlib, networkx
• docs: mkdocs, mkdocs-material

Installation

Published Linux x86_64 wheels bundle OpenBLAS and the OpenMP runtime:

pip install "pyhuge[runtime]"

Linux source builds require a C++17 compiler and OpenBLAS development files (libopenblas-dev on Ubuntu/Debian or openblas-devel on Fedora/RHEL).

From source:

cd python-package
pip install -e ".[runtime]"

Optional extras:

pip install -e ".[viz]"
pip install -e ".[test]"
pip install -e ".[docs]"
pip install -e ".[dev]"

Runtime check:

python -c "import pyhuge; print(pyhuge.test())"
pyhuge-doctor

Usage

import numpy as np
from pyhuge import huge, huge_select

rng = np.random.default_rng(1)
x = rng.normal(size=(120, 30))

fit = huge(x, method="mb", nlambda=8, verbose=False)
sel = huge_select(fit, criterion="ric", verbose=False)

print(fit.method, len(fit.path), sel.opt_lambda, sel.opt_sparsity)

input_type="auto" preserves covariance detection by symmetry. For an observation matrix that is itself square and symmetric, set input_type="data" explicitly; use "covariance" to require covariance or correlation input.

Network visualization:

import matplotlib.pyplot as plt
from pyhuge import huge_plot_network

fig, ax = plt.subplots(figsize=(5, 5))
huge_plot_network(fit, index=-1, ax=ax, layout="spring")
plt.show()

Documentation

• Docs source: python-package/docs
• Function manual pages: python-package/docs/man

Build locally:

cd python-package
mkdocs build --strict

Developer workflow

cd python-package
pytest
bash scripts/build_dist.sh
python scripts/bump_version.py 2.0.1
bash scripts/release.sh 2.0.1

Citation

If you use huge/pyhuge in research, cite:

@article{zhao2012huge,
  title   = {The huge Package for High-dimensional Undirected Graph Estimation in R},
  author  = {Zhao, Tuo and Liu, Han and Roeder, Kathryn and Lafferty, John and Wasserman, Larry},
  journal = {Journal of Machine Learning Research},
  volume  = {13},
  pages   = {1059--1062},
  year    = {2012}
}