Fields / The software landscape

The software landscape

Which tools come closest to binference, across fields? This page compares them. The field pages describe what each tool does in its field. The index gives the date of the survey.

Comparison

The census found no maintained, general Python package that samples the population’s posterior from an arbitrary likelihood matrix, or from posterior samples. The maintained tools take one field’s data products, stop at point estimates, or are written in R or Julia. The tools that sample the full hierarchical model are paper code.

The table is the one home of each tool’s language, licence and status. A dash means the census recorded nothing.

tool	field	language	licence	status	uncertainty on \(\rho\)
RAIL VarInfStackSummarizer	photo-\(z\)	Python	MIT	pz-rail-base 2.0.11, September 2026; smoke tests only	mean-field VB
Photoz-tools	photo-\(z\)	Python notebooks	MIT	last push 2020	Gibbs
BHM-nz	photo-\(z\)	Python	MIT	last push 2023	Gibbs
CHIPPR	photo-\(z\)	Python 2	MIT	README: “stale”, “a prototype”	emcee
frankenz	photo-\(z\)	Python	MIT	last push 2018; PyPI 0.1.5	Gibbs; invalid Metropolis
Starlight	colour–magnitude	Python, Cython	MIT	last commit 2017; never packaged	Gibbs
extreme-deconvolution	densities	C, Python	—	maintained; not on PyPI	none
pygmmis	densities	Python	—	1.2.3, 2022	none
StarFormationHistories.jl	Hess diagrams	Julia	MIT	2.0.0, October 2026	Hessian, HMC, MCMC
exopop	occurrence rates	Python	MIT	2014 paper code	slice sampling
DR25-occurrence-public	occurrence rates	—	—	“not intended to be used by the community”	—
PosteriorStacker	astronomy	Python	GPL/AGPL, inconsistent	0.6.4, September 2026	nested sampling
MRExo	exoplanets	Python	GPL-3	1.1.5, 2023; no tests	bootstrap
dftools	distribution functions	R	GPL-3	last commit 2021	—
gwpopulation	GW	Python	MIT	1.3.1, December 2025	any bilby sampler
BinnedCBCMasses	GW	Stan	MIT	last commit 2019	Stan’s sampler
PixelPop	GW	JAX, NumPyro	MIT	—	NUTS
gppop	GW	PyTorch, Pyro	GPL	not on PyPI; empty test directory	NUTS
GWInferno	GW	—	—	not on PyPI	—
icarogw	GW	—	—	not on PyPI	—
FIGARO	GW	—	—	54-line test suite	—
RooUnfold	unfolding	C++, ROOT	BSD-3	3.1.0; conda-forge, not PyPI	propagated errors, toys
TUnfold	unfolding	C++, ROOT	—	last standalone release 2019	—
PyUnfold	unfolding	Python	MIT	last release 2018	per-bin errors
PyFBU	unfolding	Python, PyMC3, Theano	—	“no longer actively maintained”; needs numpy<1.20	NUTS
mixsqp	empirical Bayes	R, C++	MIT	—	none
REBayes	empirical Bayes	R	GPL (≥ 2)	needs commercial MOSEK	none
ashr	empirical Bayes	R	GPL (≥ 3)	—	mean-field VB
deconvolveR	empirical Bayes	R	GPL (≥ 2)	active 2026; no tests	delta method
ebnm	empirical Bayes	R	GPL (≥ 3)	—	none for \(g\)
Empirikos.jl	empirical Bayes	Julia	MIT	active 2026; MOSEK or Hypatia	frequentist intervals
npeb	empirical Bayes	Python	MIT	alpha; no tests	none
mixsqpx	empirical Bayes	Python, JAX	MIT	alpha; Python 3.12 only	none
QuaPy	quantification	Python	BSD-3	0.2.3, October 2026; 16 test modules	NUTS; bootstrap regions
mlquantify	quantification	Python	BSD-3	—	bootstrap regions
qunfold	quantification, unfolding	Python	—	—	none
rubias	fisheries genetics	R, C++	CC0	archived on CRAN, February 2026	Gibbs
SPAM, ONCOR	fisheries genetics	desktop programs	—	legacy	bootstrap
mixstock	sea turtles	R, C	GPL	archived on CRAN, 2014	Gibbs
SourceTracker2	microbial ecology	Python	modified BSD	last PyPI release 2016	Gibbs
salmon	RNA-seq	Rust	BSD-3	active 2026	bootstrap or Gibbs replicates
kallisto	RNA-seq	C++	BSD-2	active 2026	bootstrap
density-deconvolution	neural	PyTorch	none	one commit, 2020	none
NEB	neural	PyTorch notebooks	MIT	last touched 2022	none
Denoising-Milky-Way-with-NF	neural	PyTorch	MIT	no packaging, no tests	none
neuralG	neural	R, PyTorch	—	no tests	none
diem	neural	JAX	MIT	needs a Slurm cluster	none
flowjax	flows	JAX	MIT	maintained	—
zuko	flows	PyTorch	MIT	maintained	—
nflows	flows	PyTorch	—	no release since 2020	—
flow_matching (Meta)	flows	PyTorch	CC BY-NC 4.0, no commercial use	—	—

The guides describe binference’s own inputs, engines, priors and selection.

Notes on the landscape

• Many tools stop being maintained. CRAN archived rubias because the maintainer’s email bounced. PyFBU stopped working when Theano was discontinued. Several neural repositories have no licence. emcee moved MPI pools and parallel tempering into other packages, because “there aren’t enough resources to maintain” them.

• Tests check that code runs, not that intervals cover. RAIL’s summarizer tests are end-to-end smoke tests. Several tools have no tests at all.

• Uncertainty is where the tools differ most. Most engines converge to the same maximum, but their error bars mean different things. Mean-field VB under-covers. Convex NPMLE solvers return none. Bootstrap regions measure an estimator’s variability. Gibbs and NUTS sample the exact posterior (Priors on the simplex and inference for mixture weights).

• Memory limits the scale. The established remedies merge identical rows (salmon, Empirikos.jl), store sparse matrices (npeb), or compute each likelihood once on fixed bins (Starlight). binference offers the first two (performance guide).

binference’s interface borrows patterns from other scientific Python packages; the design notes list them. The checking guide describes how binference checks a fit.