Installation

MINTverse is two packages, with estiMINT mapping a measured prevalence to transmission intensity and stateMINT emulating the prevalence and case trajectories that follow an intervention. Installing estiMINT with the scenarios extra brings in stateMINT and the run_scenarios pipeline, which is what most work needs.

pip install "estimint[scenarios]"

Or, inside a project managed by uv:

uv add "estimint[scenarios]"

Python 3.12 or newer is required.

Distribution name and import name

stateMINT is published on PyPI as mintstate, so you install mintstate and you import stateMINT:

import stateMINT          # correct
import mintstate          # ModuleNotFoundError

estiMINT has no such split, and its distribution and module are both estimint.

Extras

An extra is an optional set of dependencies, named in square brackets after the package. Plain pip install estimint gives the EIR and HBR models but no emulator and no run_scenarios, so scenarios is the extra most installs want.

Extra	Command	What it adds
`scenarios`	`pip install "estimint[scenarios]"`	stateMINT, so `run_scenarios` and the emulator are available.
`viz`	`pip install "estimint[viz]"`	matplotlib, for the bundled plotting helpers.
`download`	`pip install "estimint[download]"`	The fetchers used to pull training data.
`train`	`pip install "estimint[train]"`	duckdb, scikit-learn and pyarrow, for refitting the XGBoost models.
`all`	`pip install "estimint[all]"`	The four estiMINT extras above.
`gpu`	`pip install "mintstate[gpu]"`	`jax[cuda12]`, so the emulator runs on an NVIDIA GPU with CUDA 12.

The CPU build of JAX is fast enough for inference, where a ten-scenario batch takes about a second, and the gpu extra is only needed if you are training or sweeping thousands of scenarios.

Model weights

estiMINT ships its three XGBoost models (prevalence, HBR and EIR-to-HBR) inside the wheel. load_xgb_model("prevalence") never needs a network.

stateMINT does need one on first use, because its checkpoints live in the Hugging Face repository dide-ic/stateMINT, with a separately trained checkpoint of about 38 MB for each predictor under prevalence/ and cases/, and the first call that needs one fetches it and caches it under ~/.cache/huggingface. Every later call reads the cache. preload_models() fills both caches, loading the XGBoost models and pulling the stateMINT weights.

from estimint import preload_models

preload_models()

Working offline

On a cluster node or an air-gapped machine, warm the cache somewhere with a network first, then set HF_HUB_OFFLINE=1 so that Hugging Face never attempts a request and resolves everything from ~/.cache/huggingface:

export HF_HUB_OFFLINE=1

If the cache directory itself is unavailable, from_pretrained takes a local_dir and will read a directory laid out like the repository instead of calling the Hub.

from stateMINT.model import Mamba2Regressor

artifact = Mamba2Regressor.from_pretrained(
    "dide-ic/stateMINT",
    predictor="prevalence",
    local_dir="/shared/models/stateMINT",
)

Verify the install

importlib.metadata reports the version of an installed distribution rather than of an imported module, so the emulator’s version is recorded under mintstate and not under stateMINT. net_types() asks estiMINT which net formulations it knows about.

import importlib.metadata as md

import estimint
import stateMINT

print("estimint  (import estimint) ", md.version("estimint"))
print("mintstate (import stateMINT)", md.version("mintstate"))
print()
print("net types:", estimint.net_types())

estimint  (import estimint)  1.5.4
mintstate (import stateMINT) 0.3.1

net types: ['pyrethroid_only', 'pyrethroid_pbo', 'pyrethroid_ppf', 'pyrethroid_pyrrole']

Four net formulations and no error confirms both installs.

Extras

Model weights

Working offline

Verify the install

See also