MNPE — paper reproduction code

Code and tracked results reproducing the experiments in the MNPE paper (mixed neural posterior estimation for models with both discrete and continuous parameters).

The method itself is not here. MNPE is implemented and maintained in the sbi toolbox; build on that rather than on anything in this repository. What this repository adds is the experiments: the models, the reference posteriors they are scored against, the calibration diagnostics, and a script per paper figure.

Install

git clone https://github.com/mackelab/mnpe-paper.git && cd mnpe-paper
uv sync

No local checkouts are required — uv sync resolves everything, including sbi, from a package index.

Layout

Path	Contents
src/mnpe_paper/	Importable shared machinery: calibration diagnostics, plotting style, path resolution, per-model specifications
experiments/	Runnable scripts, one directory per experiment
figures/	One script per paper figure, plus the published PDFs
results/	Tracked result tables and the compact figure inputs every figure rebuilds from
artifacts/	Untracked. Full pipeline output, written only if you rerun a pipeline
tests/	Behaviour tests, including the architecture-pinning guard

Figures

Every figure below rebuilds from tracked data alone, in a fresh clone, with no pipeline rerun:

uv run python figures/make_coal_figure.py

Figure	Producer	Reads	Experiment
fig1.pdf	figures/make_fig1_panels.py, panels only: they were assembled into fig1.svg by hand in Inkscape and exported to the PDF	nothing	method overview
toy_paper_figure.pdf	figures/make_toy_figure.py	results/toy/figure_inputs.npz	toy
sensitivity_discrete_errors_redesign.pdf	figures/sensitivity_discrete_errors_figure.py	results/toy/sensitivity_discrete_errors.csv	toy
coal_paper_figure.pdf	figures/make_coal_figure.py	results/coal/figure_inputs.npz	coal
queue_paper_figure.pdf	figures/make_queue_figure.py	results/queue/figure_inputs.npz, plus three tracked tables	queue
scaling_figure.pdf	figures/scaling_figure.py	results/scaling/scaling_study_results.csv	scaling
k_sweep_figure.pdf	figures/k_sweep_figure.py	results/scaling/k_sweep_results.csv	scaling
figureHH.pdf	figures/make_hh_figure.py	results/hh/figure_inputs.npz	hh
figureHH_supp_1.pdf	figures/make_hh_supp_1_figure.py	results/hh/figure_inputs.npz	hh
figureHH_supp_2.pdf	figures/make_hh_supp_2_figure.py	results/hh/figure_inputs.npz	hh

Rebuild all figures

uv run python scripts/make_all_figures.py

This runs every script in figures/ and exits non-zero if any fails. Afterwards git status shows no modified files: the tracked PDFs rebuild byte for byte. make_fig1_panels.py regenerates Fig. 1's panels only, not the assembled fig1.pdf.

Tests

uv run pytest
uv run ruff check .

Citation

If you use this code, please cite the paper. The metadata is in CITATION.cff; as BibTeX:

@article{boelts2026mixed,
  title   = {Mixed neural posterior estimation for simulators with discrete and continuous parameters},
  author  = {Boelts, Jan and Schr{\"o}der, Cornelius and Beck, Jonas and Macke, Jakob H. and Deistler, Michael and Gedon, Daniel},
  journal = {Transactions on Machine Learning Research},
  issn    = {2835-8856},
  year    = {2026},
  month   = sep,
  url     = {https://openreview.net/forum?id=fOZZeRfdrC}
}

Reproduction and posture

This repository is a frozen snapshot pinned to sbi==0.27.0. Bit-exact reproduction is not claimed; runs are seeded and CPU-only, but thread counts and BLAS reduction order are not pinned, so a rerun agrees with the tracked numbers within training noise rather than exactly. Rerunning a pipeline rewrites its tracked tables; git checkout results figures restores them. Issues reporting genuine reproduction problems are welcome; improvements to the method itself belong upstream in sbi.