glmax

!!! warning

THIS PROJECT IS STILL IN ALPHA AND SUBJEC TO CHANGE DRAMATICALLY; USE AT YOUR OWN RISK**

glmax provides generalized linear modeling with a grammar-first API built around explicit nouns and top-level verbs.

The canonical workflow is:

1. specify(...) -> GLM
2. fit(model, data, init=None, *, fitter=...) -> FittedGLM
3. predict(model, params, data) for fitted means
4. infer(fitted, inferrer=None, stderr=...) for inferential summaries without refitting
5. check(fitted) for diagnostics without refitting

The canonical nouns are GLMData, Params, FitResult, FittedGLM, InferenceResult, AbstractDiagnostic, GofStats, and InfluenceStats. The package-root also exports inference strategy and stderr types: AbstractTest, WaldTest, ScoreTest, AbstractStdErrEstimator, FisherInfoError, and HuberError.

Params(beta, disp, aux) is the shared parameter carrier across fit, predict, and infer:

• beta stores regression coefficients.
• disp stores GLM dispersion. Gaussian and Gamma use it as EDM dispersion; Poisson, Binomial, and Negative Binomial canonicalize it to 1.0.
• aux stores optional family-specific state. Negative Binomial stores its alpha here while canonical disp remains 1.0.

See docs/index.md for the workflow overview, the specify API for model construction, and the Families & Links guide for the family-specific disp/aux split.

Installation

git clone https://github.com/mancusolab/glmax.git
cd glmax
pip install .

Quickstart

import jax.numpy as jnp
import glmax

from glmax import GLMData
from glmax.family import Gaussian

model = glmax.specify(family=Gaussian())
data = GLMData(
    X=jnp.array([[0.0], [1.0], [2.0], [3.0]]),
    y=jnp.array([0.1, 1.2, 1.9, 3.1]),
)

fitted = glmax.fit(model, data)
params = fitted.params
pred = glmax.predict(model, params, data)
infer_result = glmax.infer(fitted)
# Route through an explicit inferrer when needed.
score_result = glmax.infer(fitted, inferrer=glmax.ScoreTest())
pearson = glmax.check(fitted)

Testing

Use the repository-standard pytest invocation (also wired into project scripts):

pytest -p no:capture tests

Support

• Issues: https://github.com/mancusolab/glmax/issues
• Source: https://github.com/mancusolab/glmax

License

glmax is distributed under the MIT license.