depictr (Python)

Documentation and example gallery: https://pablobernabeu.github.io/depictr-py/

A unified, colourblind-safe toolkit for publication-ready statistical visualisation, built on plotnine.

This is the Python twin of the R package of the same name, which offers the same workflow on top of ggplot2. The two share a design and nearly all of an API, and the Status section below records where this one has yet to catch up.

Gallery

A grouped density (the default palette is the colourblind-safe Okabe-Ito set) and Kaplan-Meier survival curves with a log-rank test and a number-at-risk table, both from a single function call:

There is more in the example gallery. These two use a couple of extra options (a title, legend_inside) shown in the Getting started guide, while the short tour below keeps its calls minimal.

What it is for

Python already has an excellent plot for almost any statistical task, but those plots live in different packages with different defaults. A single figure set for a paper might draw on seaborn, scikit-learn, statsmodels, lifelines and ArviZ, each with its own look, its own API and its own colour scheme, and only one of those defaults to colourblind-safe colours. Making the set consistent and accessible then means repeating the same theming by hand on every plot.

depictr does that work once. It gives the whole workflow one theme, one colourblind-safe palette and one calling convention, and returns plotnine objects you can keep extending with +. Where a specialist package already computes a quantity well, depictr delegates to it and redraws the result under the shared theme, so you keep the trusted computation and gain a coherent, accessible figure.

Accessibility by default

The default palette is the Okabe-Ito set, and the package puts that choice to a test. It ships a simulator of colour-vision deficiency based on the model of Machado, Oliveira and Fernandes (2009), and a CIE-Lab distance test that measures how far apart the palette's colours stay under each deficiency:

import depictr as dp

dp.palette_safety()
# {'min_delta_e': ..., 'by_condition': {'normal': ..., 'protan': ...,
#  'deutan': ..., 'tritan': ...}, 'worst_condition': ..., 'worst_pair': ...,
#  'safe': True, 'threshold': 5.0}

The guarantee belongs to the eight Okabe-Ito colours. A plot with more than eight groups has to interpolate between them, and the interpolated colours no longer clear the distance threshold, so depictr_palette warns when it is asked for a ninth. Facet the groups, or map them to the sequential ramp, when a figure needs more.

A safe palette is not a safe figure, though. Once a plot has been extended with your own scale, shrunk to fit a journal column, or asked to distinguish groups by colour alone, the palette's guarantee no longer describes what a reader will see. check_figure audits the plot you are about to submit and reports what it measured beside the threshold it was measured against, so each verdict can be argued with:

p = dp.explore_distribution(dp.lexical_decision(), "RT", group="condition")
dp.check_figure(p, width_cm=8.9)

It measures colour separability under each dichromacy and in greyscale, the smallest text size at the width the figure will be printed at, the WCAG contrast of text and geometry against their backgrounds, and whether any distinction rests on colour alone. One limitation belongs here, beside the claim it qualifies. The colourblind-safety guarantee covers hue confusion alone, and in greyscale the palette's orange and sky blue differ by only 0.79 in CIE lightness, so they print as the same grey. A figure that may be printed in black and white wants fewer groups, a sequential palette, or a redundant shape or line type.

Installation

depictr is on PyPI:

pip install depictr            # core (plotnine, pandas, numpy, matplotlib, scipy)
pip install depictr[all]       # plus the optional computation back-ends

The classification and survival plots delegate to scikit-learn and lifelines, each an optional dependency installed only when your plots need it (depictr[classification], depictr[survival]). The model and time-series plots delegate to statsmodels, which plotnine itself requires, so it arrives with the core install. The depictr[models] extra remains to pin the tested statsmodels floor.

The development version comes from GitHub:

pip install git+https://github.com/pablobernabeu/depictr-py

A short tour

import depictr as dp

# Exploratory analysis
ld = dp.lexical_decision()
dp.explore_distribution(ld, "RT", group="condition", kind="both")

wb = dp.wellbeing_survey()
dp.correlation_heatmap(wb)
dp.missingness_map(wb)

# Model estimates: a fitted model OR a tidy data frame
import statsmodels.formula.api as smf
# Q() quotes "yield" because it is a Python keyword.
fit = smf.ols('Q("yield") ~ fertiliser + rainfall + soil_ph + treatment',
              dp.crop_yield()).fit()
dp.coefficient_plot(fit)

# Classification: computed by scikit-learn, themed by depictr
ct = dp.clinical_trial()
dp.roc_curve_plot(ct["adverse_event"], ct["biomarker"])

# Survival: Kaplan-Meier with a log-rank test, in one call
dp.survival_plot(ct["time"], ct["event"], group=ct["arm"])

Every function returns a plotnine object, so the usual grammar-of-graphics extensions apply:

from plotnine import labs

dp.roc_curve_plot(ct["adverse_event"], ct["biomarker"]) + labs(title="Adverse event")

The web app

A Streamlit app provides a gallery and a low-friction way to try the package. It loads one of the bundled datasets (or a CSV you upload), draws the chosen plot, shows the exact Python call that produced it, and offers a colourblind-vision toggle that re-renders the figure as each deficiency would be seen. The app lives in the repository rather than the wheel, so run it from a clone:

git clone https://github.com/pablobernabeu/depictr-py
cd depictr-py
pip install -e ".[app]"
streamlit run app/streamlit_app.py

Function families

The exported functions fall into families that follow the stages of an analysis.

Status

This is an early release, but coverage is broad. The colourblind-safe theme, the accessibility check and most of the R package's plotting functions across every family (EDA, estimation, model estimates, diagnostics, classification, multivariate, survival and time series) are in place and tested. Multi-panel composites are built on arrange_plots, which uses plotnine's native plot composition, and they include the four-panel residual_diagnostics_plot, the model_report dashboard, the two-panel Gardner-Altman estimation_plot, the frequentist-over-Bayesian overlay and the survival number-at-risk table.

A few known limitations remain. Compositions in plotnine have no figure-level title, so a grid carries its titles on each panel (the survival and estimation composites place the title on their top panel). survival_plot does not yet draw the confidence band or censor marks the R package draws, though its conf_level argument is accepted for future use. Seven functions from the R package have yet to be ported: optimizer_fixef_plot (there is no direct statsmodels equivalent of lme4::allFit), k_diagnostic, palette_preview, model_fit_table, ts_forecast, the label helper format_terms and the session-wide settings of depictr_options. The monthly_sales dataset is also R-only, so the time-series examples here build their own series.

Relationship to the R package

The two are siblings, with no shared codebase. The R package is built on ggplot2 and this one on plotnine, and the design is common to both. Each gives the whole workflow one accessible theme, takes its input either as a fitted model or as a data frame, and returns plot objects that stay extensible.

Citation

The About page carries the preferred citation with a BibTeX entry and a short note on the developer, and the repository ships CITATION.cff for the Cite this repository button on GitHub.

Licence

MIT. See LICENSE.

Contributing

Issues and pull requests are welcome. The contributing guide describes the development setup and the conventions the package follows, and everyone taking part is asked to honour the Code of Conduct.