GLUM - High Performance Generalized Linear Models

3.3.0 · active · verified Thu Apr 16

glum is a Python library providing high-performance implementations of Generalized Linear Models (GLMs), including various distributions and link functions. It focuses on speed and feature richness, supporting regularized fitting (L1, L2, ElasticNet) and cross-validation. The current version is 3.3.0, and it maintains an active release cadence with multiple updates throughout the year.

Common errors

```
ValueError: Input contains NaN, infinity or a value too large for dtype('float64').
```
cause Using `InverseGaussianDistribution` with `glum` versions prior to 3.2.3 would cause `log_likelihood` to return NaN, propagating through the model fitting process.

fix Upgrade your `glum` installation to version 3.2.3 or newer: `pip install --upgrade glum`.
```
TypeError: 'numpy.float64' object cannot be interpreted as an integer
```
cause In `glum` versions before 3.1.3, the `theta` setter for `NegativeBinomialDistribution` incorrectly rejected `numpy.number` types, leading to errors when attempting to assign a `theta` value derived from NumPy operations.

fix Upgrade `glum` to version 3.1.3 or later: `pip install --upgrade glum`.
```
Predictions are inconsistent or incorrect when using alpha_search with categorical features.
```
cause A bug in `glum` versions prior to 3.2.1 caused incorrect predictions when specifying a particular `alpha` value after fitting with `alpha_search=True` and having categorical features in the dataset.

fix Ensure you are using `glum` version 3.2.1 or newer. Upgrade using `pip install --upgrade glum`.
```
The `deviance_path_` attribute in `GeneralizedLinearRegressorCV` shows unexpectedly low values.
```
cause Before version 3.1.3, `deviance_path_` was incorrectly scaled down by a factor of `n_folds`, leading to underestimated deviance values.

fix Upgrade your `glum` library to version 3.1.3 or newer to get correctly scaled `deviance_path_`: `pip install --upgrade glum`.

Warnings

breaking In v3.3.0, the `trust-constr` solver's default Hessian calculation changed from `hess="2-point"` (finite-difference) to `SR1()` (quasi-Newton). While improving performance, this might lead to slightly different numerical results for models that previously relied on the finite-difference Hessian.
Fix: If exact reproducibility with older versions is critical, explicitly set `solver_params={'hess': '2-point'}` when initializing `GeneralizedLinearRegressor` or `GeneralizedLinearRegressorCV` if using `solver='trust-constr'`, though `SR1()` is generally superior.
gotcha Prior to v3.2.3, `InverseGaussianDistribution.log_likelihood` contained an incorrect call, causing it to always return NaN. Models using `InverseGaussianDistribution` would fail or produce invalid results.
Fix: Upgrade to `glum>=3.2.3` to ensure correct calculations for the `InverseGaussianDistribution`.
gotcha Versions before 3.2.1 had an error when predicting at a specific `alpha` with categorical features, potentially leading to incorrect or failed predictions.
Fix: Upgrade to `glum>=3.2.1` to resolve prediction issues when working with categorical features and L1/ElasticNet regularization (`alpha_search`).
gotcha In `GeneralizedLinearRegressorCV`, the `deviance_path_` attribute was incorrectly scaled by `n_folds` in versions prior to 3.1.3, leading to misinterpretation of cross-validation deviance values.
Fix: Upgrade to `glum>=3.1.3` for correct `deviance_path_` values in cross-validation.

Install

pip install glum Install stable release

Imports

GeneralizedLinearRegressor

from glum import GeneralizedLinearRegressor

GeneralizedLinearRegressorCV

from glum import GeneralizedLinearRegressorCV

PoissonDistribution
```
from glum import PoissonDistribution
```
GammaDistribution
```
from glum import GammaDistribution
```
LogLink
```
from glum import LogLink
```

Quickstart

This quickstart demonstrates how to fit a Poisson GLM with a log link using `glum.GeneralizedLinearRegressor`. It generates synthetic data, then initializes and fits the model, finally printing the intercept, coefficients, deviance, and predictions for the first few samples.

import numpy as np
import pandas as pd
from glum import GeneralizedLinearRegressor, PoissonDistribution, LogLink

# Generate some synthetic data
np.random.seed(42)
n_samples = 100
X = pd.DataFrame({
    'feature_1': np.random.rand(n_samples) * 10,
    'feature_2': np.random.rand(n_samples) * 5
})

# True coefficients
beta_0 = 1.0
beta_1 = 0.5
beta_2 = 0.2

# Generate Poisson-distributed target variable using a log link
linear_predictor = beta_0 + beta_1 * X['feature_1'] + beta_2 * X['feature_2']
mu = np.exp(linear_predictor)
y = np.random.poisson(mu)

# Create and fit the GLM
glm = GeneralizedLinearRegressor(
    distribution=PoissonDistribution(), 
    link=LogLink(),
    fit_intercept=True
)
glm.fit(X, y)

print(f"Intercept: {glm.intercept_:.4f}")
print(f"Coefficients: {glm.coef_}")
print(f"Deviance: {glm.deviance_:.4f}")
print("Predictions (first 5 samples):\n", glm.predict(X.head()).round(2))

view raw JSON →