pysplinecalib

Spline-based probability calibration for binary and multiclass classifiers, in pure Python.

pysplinecalib is a fork of splinecalib by Brian Lucena. It implements the method described in Spline-Based Probability Calibration. The Cython extension has been replaced with NumPy/SciPy, so the package installs with no compiler.

Installation

From PyPI:

pip install pysplinecalib

From Github:

pip install git+https://github.com/Surgical-Recovery-and-Safety-Lab/pysplinecalib.git

For development:

git clone https://github.com/Surgical-Recovery-and-Safety-Lab/pysplinecalib.git
cd pysplinecalib
pip install -e ".[dev]"

Requires Python 3.10 or later.

Usage

Fit the calibrator on model outputs from a held-out calibration set, then use it to calibrate new predictions.

Binary classification

from pysplinecalib import SplineCalib

# p_calib: predicted probability of class 1, shape (n_samples,)
# y_calib: true labels in {0, 1}, shape (n_samples,)
calibrator = SplineCalib()
calibrator.fit(p_calib, y_calib)

p_test_calibrated = calibrator.calibrate(p_test)

A 2-column array of shape (n_samples, 2) is also accepted, and calibrate returns the same shape it was given.

Multiclass classification

# probs_calib: shape (n_samples, n_classes); y_calib: integers in 0..n_classes-1
calibrator = SplineCalib()
calibrator.fit(probs_calib, y_calib)

probs_test_calibrated = calibrator.calibrate(probs_test)  # rows sum to 1

One binary calibrator is fit per class, and the calibrated outputs are renormalised. The individual calibrators are available as calibrator.binary_splinecalibs.

Diagnostics

calibrator.show_calibration_curve()   # calibrated vs. uncalibrated probability
calibrator.show_spline_reg_plot()     # cross-validated loss vs. regularization

For multiclass calibrators, pass class_num= to select a class.

Main parameters

See help(SplineCalib) for the full list.

Differences from splinecalib

• Pure Python: the Cython loss and gradient are now NumPy/SciPy code, and the spline basis is vectorised.
• The final fit in logreg_cv now uses all the data and the requested optimisation method. Upstream refit on the last cross-validation training split, so calibrated values can differ slightly.
• With custom unity_prior_gridpts, the prior weight is no longer applied twice.
• Multiclass calibrators now inherit reg_prec, param_search_mode, force_knot_endpts and logodds_eps from the parent.
• fit no longer modifies the array passed in.

See CHANGELOG.md for the full list.

Development

ruff check src
ruff format --check src
pytest --cov=pysplinecalib

License

MIT. Copyright (c) 2022 numeristical and (c) 2026 Surgical Recovery and Safety Laboratory. See LICENSE.

References

Lucena, B. Spline-Based Probability Calibration. https://arxiv.org/abs/1809.07751