gp3tools for Python

Documentation: https://stefanosbalaskas.github.io/gp3tools-python/

A Python implementation of the public workflow of gp3tools, the R package for importing, inspecting, analysing, visualising, modelling, and reporting Gazepoint GP3 / Gazepoint Analysis exports.

Status: 0.1.2 stable release. The complete 278-function public name surface frozen from the R v2.3.0 NAMESPACE is present. Core data, QC, pupil, AOI, sequence, event, plotting, reporting, face, simulation, and export workflows are native Python. Statistical functions tied to R-specific backends are Python-native adaptations, not claims of coefficient-for-coefficient equivalence to lme4, glmmTMB, mgcv, or brms.

Installation

Install the validated stable release from PyPI:

python -m pip install "gp3tools==0.1.2"

or allow the latest pre-release explicitly:

python -m pip install --pre gp3tools

PyPI: https://pypi.org/project/gp3tools/

Quick start

import gp3tools as gp3

master = gp3.load_example_master()

sampling = gp3.check_sampling_rate(
    master,
    time_col="TIME",
    group_cols=["subject", "trial_global"],
)

pupil = gp3.preprocess_gazepoint_signals(
    master,
    pupil_col="pupil",
    time_col="TIME",
)

transitions = gp3.compute_gazepoint_aoi_transition_matrix(
    master,
    aoi_col="aoi_current",
)

fig = gp3.plot_gazepoint_heatmap(master)

Public API contract

import gp3tools as gp3

len(gp3.R_EXPORTS)
# 278

gp3.api_status()["status"].value_counts()

The build distinguishes:

• native: direct Python implementation;
• native-adapted: implemented in Python but the R package used a backend or algorithm whose exact numerical identity is not asserted;
• native-adapter: interoperability adapter expressed natively in Python.

See docs/API_PARITY.md and the bundled api_manifest.csv for every exported function.

Main modules

• io: Gazepoint CSV, folder, summary, and face-export import
• qc: master tables, sampling, tracking, missingness, screen bounds, exclusions, readiness
• pupil: artifact flags, blinks, interpolation, baseline correction, smoothing, binocular reconstruction
• aoi: static/dynamic/polygon AOIs, entries, windows, transitions, entropy, scanpaths, clustering
• events: fixation/saccade detection, agreement and benchmark workflows
• stats: model preparation, LMM/GLM/spline adaptations, Bayesian adapters, cluster permutation, bootstrap and sensitivity
• face: external facial-analysis QC and time synchronisation
• interop: eyetrackingR/pupillometryR/gazer/eyetools/HDDM/BIDS/gpbiometrics adapters
• plotting: Matplotlib visualisations
• reporting: CSV/HTML outputs, reporting checklists and end-to-end workflow
• simulation: synthetic Gazepoint data for examples and tests

Scientific compatibility note

A Python port should preserve scientific intent without pretending different statistical engines are identical. Accordingly, the modelling layer documents backend adaptation explicitly. The package is suitable for software migration, workflow prototyping, reproducible preprocessing, data/QC/plotting tasks, and validation development; advanced inferential parity should be checked against the R reference implementation before confirmatory use.

R reference

The public API was frozen against gp3tools R v2.3.0. R_NAMESPACE_REFERENCE.txt is included in the source distribution for auditability.

Citation

Balaskas, S. (2026). gp3tools: An R Package for Reproducible Analysis and Reporting of Gazepoint GP3 Eye-Tracking Exports. Journal of Eye Movement Research, 19(4), 76. DOI: 10.3390/jemr19040076.

License

MIT.

R 2.3.0 parity and validation

The Python package freezes the 278-export R 2.3.0 public API while exposing 285 Python public names in the validated parity branch. Behavioral equality is evidence graded: exact and numerical-tolerance claims are made only where frozen R oracles exist; backend-adapted statistical families are labelled accordingly.

The release gate includes R1–R3 frozen behavioral regressions, R4 canonical structure/compatibility tests plus frozen-oracle integrity checks, whole-surface smoke, plot catalog smoke, runnable examples, Ruff, compilation, strict documentation build, ≥90% test coverage, wheel/sdist construction and isolated-wheel validation. See docs/PARITY_STATUS.md.

October 2026 research article and figure

Native acquisition rate and spatial dispersion documents experimental, source-only validation methods and reproducible synthetic examples. It is available in the website navigation, while independently qualified external empirical claims and stable PyPI APIs remain separate (scientific acceptance criteria).