One-line matplotlib plots and themes for notebooks and scripts.
kx is a small, plain-Python package with .mplstyle themes, palettes and colourmaps. It needs nothing beyond
matplotlib, numpy and pandas and works offline, so it also runs in Kaggle competition notebooks.
pip install kx-plot
# adds t-SNE and UMAP to kx.embed (PCA needs neither)
pip install 'kx-plot[embed]'import kx; kx.use('dark')Attach the dataset instead (Add Input → Datasets → vhrabar/kx-plot), then:
import sys; sys.path.append('/kaggle/input/datasets/vhrabar/kx-plot')
import kx; kx.use('dark')kx.use('light', palette='okabe', cmap='cividis') # notebook default: theme, palette, colormap
kx.plot(x, y, z, 'scatter-grid-leg') # spec: kind-theme-flag-...
kx.plot(x, y, 'bar-paper') # a theme in the spec applies to this plot only
with kx.context('paper'): ... # or to every plot in the block
kx.save('fig.png') # to /kaggle/working on Kaggle, else the current folder
for ax, col in zip(kx.grid(3), df.columns): # a figure of 3 panels, a flat array of axes
kx.plot(df, col, spec='hist', ax=ax)
kx.plot(df, 'month', 'y2025', 'bar') # DataFrame columns by name
kx.plot(x, y, 'line', linewidth=3) # other keywords go to matplotlib
kx.plot(df, 'value', spec='kde', by='group') # one density curve per group
kx.plot(df.value, spec='box', by=df.group) # one box per group (also ridge, violin, qq)| Token | Values |
|---|---|
| kind | line step area band scatter bar barh hist kde ridge ecdf box violin |
strip qq count crosstab hexbin heat missing nullmap pie donut radar |
|
parallel pair cluster mosaic treemap sunburst |
|
ML: cm roc pr calib proba thresh gain lift pva resid imp loss |
|
lcurve vcurve elbow embed |
|
Images: imgrid meanimg colorhist imhist segmask saliency miscls |
|
| theme | dark light paper science talk poster thesis |
| flags | grid leg legout tight rot wrap logx logy logxy symlogy zero eq pct |
si facet; per kind: stack norm cum date mk sort ann mean median nocb |
|
sym logc corr all diag fit best raw sig |
stack works with area, bar, barh and hist. norm is a 100% stack for area and a
density (area 1) for hist. cum makes hist cumulative; hist-norm-cum is the CDF.
hist puts every series on the same bins, so the bars line up.
diag adds a dashed y = x line to scatter and hexbin; fit adds a least-squares line to scatter, with
its slope and R² in the legend (one line per group with a categorical z).
sig on box and violin writes the Kruskal-Wallis p and brackets the pairs that differ, with stars (ann: the
p-values), Holm-corrected.
facet draws a plot once per value of a column, kx.plot(df, 'pclass', 'port', 'crosstab-facet', facet='sex'),
with every category in every panel, shared axes, and one colour scale and colorbar for the colour kinds.
Unknown tokens raise an error instead of being guessed, with the closest valid tokens as a hint.
Everything is inspectable: kx.grammar(), kx.grammar('area'), kx.parse('leg-scatter'), kx.show('dark'),
kx.source('scatter').
kx.load('passengers') is a demo table with numbers, categories and NaNs to try these on.
kx.missing(df) # share of NaNs per column, the most missing on top
kx.plot(df, 'nullmap-sort') # where the NaNs are: one pixel row per table row
kx.plot(df, 'heat-corr-ann') # correlations of the numeric columns, -1 to 1
kx.plot(df, 'heat') # any numeric DataFrame or 2-D array as coloured cells
kx.cramers(df, cluster=True) # Cramér's V of the categorical columns, 0 to 1
kx.plot(df, 'cluster-corr') # correlations, similar columns side by side, with trees
kx.plot(df, 'pclass', spec='count', by='survived') # rows per category, split by a second column
kx.plot(df, 'pclass', 'port', 'crosstab-norm-ann') # how often each pair of categories occurs
kx.plot(df, 'port', 'sex', 'survived', spec='mosaic-ann') # tiles sized by count, up to 3 categoricals
kx.plot(df, 'pclass', 'sex', spec='treemap-ann') # a hierarchy of categories; 'sunburst' as rings
kx.plot(df, 'age', 'fare', 'hexbin-logc') # points per hexagon; with a z, its mean per hexagon
kx.plot(df, 'port', 'donut-ann') # shares of a whole, the total in the hole
kx.plot(df[['age', 'fare', 'pclass', 'survived']], 'radar', by='survived') # each group's profile
kx.plot(df[['age', 'fare', 'pclass', 'sex']], 'parallel-mean', by='sex') # every row across columns
kx.plot(df, 'pair', by='sex') # every numeric column against every otherheat, cluster, missing, nullmap, radar, parallel and pair take the whole table and no column names
(radar, parallel and pair take by=, a column of groups); kx.plot(df[cols], 'heat') for some columns.
wrap breaks long tick labels onto several lines.
Image kinds take an (N, H, W) or (N, H, W, C) array (channels first works too), or a list of arrays, PIL images
or file paths, with the labels as y. kx.load('shapes') is a small demo set.
d = kx.load('shapes')
kx.plot(d['images'], d['labels'], 'imgrid') # 6 random images per class, one row each
kx.plot(paths, labels, 'imgrid', idx=[3, 17, 42]) # chosen images, titled with their labels
kx.plot(d['images'], d['labels'], 'meanimg') # the mean image of each class; -std: where pixels vary
kx.plot(d['images'], d['labels'], 'colorhist') # pixel values per channel and class; -hsv, -map
kx.plot(d['images'], 'imhist', idx=[0, 1]) # images beside their own channel histograms
kx.plot(d['images'], d['masks'], pred, 'segmask') # image, mask, outlines; a prediction adds its IoU
kx.plot(d['images'], d['saliency'], 'saliency') # Grad-CAM or any map over its image, smoothed to fit
kx.plot(d['images'], d['labels'], d['proba'], 'miscls') # the mistakes, the most confident firstkx.embed(df, by='label') # PCA of the numeric columns, explained variance on the axes
kx.embed(images, by=labels, method='tsne') # t-SNE or UMAP: pip install 'kx-plot[embed]'
kx.plot(xy, labels, 'embed-ann') # coordinates you already have, names at each groupkx.embed returns the axes, and the coordinates are ax.embedding.
kx.load('binary'), 'multiclass', 'regression', 'history', 'importance' and 'curves' are demo data for
these kinds. Metrics are computed in numpy: no scikit-learn needed.
kx.plot(y_true, score, 'roc') # ROC curve, AUC in the legend
kx.plot(y_true, gbm, logreg, 'pr') # a second model goes in z; AP and a random baseline
kx.plot(y_true, proba_df, 'roc') # a 2-D score: one curve per class, one-vs-rest
kx.plot(y_true, score, 'calib') # reliability diagram, Brier score in the legend
kx.plot(y_true, score, 'proba') # score histogram per true class
kx.plot(y_true, score, 'thresh-best') # precision, recall and F1 against the threshold
kx.plot(y_true, score, 'gain') # also 'lift'
kx.plot(y_true, y_pred, 'cm-norm-ann') # confusion matrix, each row the recall of a class
kx.plot(actual, predicted, 'pva') # predicted vs actual, R², RMSE and MAE
kx.plot(actual, predicted, feature, 'resid') # residuals, coloured by a feature
kx.plot(df, 'feature', 'importance', 'std', 'imp') # feature importance with error bars, top= 20
kx.plot(model.fit(...), 'loss-best') # training curves from a Keras History, dict or DataFrame
kx.plot(sizes, train_scores, val_scores, 'lcurve') # scikit-learn's learning_curve output; also 'vcurve'
kx.plot(ks, inertia, silhouette, 'elbow-best') # choosing k, the knee marked
kx.roc_auc(y_true, score), kx.pr_ap(y_true, score) # the metrics as numbers
kx.evaluate(y_true, gbm, logreg) # a report: the panels picked from the dataThe truth comes first, then the predictions. Binary labels are 0/1 or bool, with 1/True the positive class;
with two other labels the later one in sorted order is positive. For multiclass, pass a 2-D score with one column
per class: the class names are the DataFrame's columns, or else the sorted labels of y_true. roc, pr, calib,
gain and lift draw their legend without leg, because the metric in it is the point of the plot.
kx.evaluate(y_true, *scores) draws roc, pr, calib and proba for a binary task, roc, cm and
proba for a 2-D score, and pva, resid and a residual hist for a numeric y with many values. It returns
the flat array of axes, like kx.grid; spec='dark-grid' styles every panel.
rcParams can be overridden in kx.use with __ for dots: kx.use('dark', figure__figsize=(8, 4)).
Every kind, option, flag and config, with its output, is in the Example Usage notebook.
Palettes for kx.use(theme, palette=...), they swap only the colours; a theme's
other cycled properties (such as paper's linestyles) stay.
| Name | Source |
|---|---|
okabe |
Okabe & Ito (2008) |
tolbright tolvibrant tollight |
Paul Tol |
tolhc tolmc |
Paul Tol |
petroff6 petroff8 petroff10 |
Petroff (2021) |
tab10 tab20 tableaucb |
Tableau |
set1 set2 dark2 paired |
ColorBrewer |
ibm |
IBM Design Colors |
kx.palette('okabe') returns the hex codes. A list of colours works too: palette=['#264653', '#e9c46a'].
Colourmaps for kx.use(theme, cmap=...) colour continuous values, such as a numeric z in scatter.
Without cmap, the theme's own default applies.
Add _r to any name to reverse it, e.g. cmap="thermal_r".
| Name | Type |
|---|---|
viridis magma plasma inferno cividis |
Sequential, perceptually uniform |
turbo |
Sequential, rainbow |
blues greys |
Sequential, ColorBrewer |
coolwarm |
Diverging |
rdbu |
Diverging, ColorBrewer |
berlin managua vanimo |
Diverging, dark centre |
twilight |
Cyclic, for angles and phases |
thermal haline deep |
Sequential, cmocean (vendored) |
balance |
Diverging, cmocean (vendored) |
batlow lipari hawaii oslo davos |
Sequential, Crameri (vendored) |
vik roma bam cork |
Diverging, Crameri (vendored) |
romaO |
Cyclic, Crameri (vendored) |
kx.cmaps() lists the ones that are available in the installed matplotlib as those implemented in matplotlib 3.10 or newer
are not vendored.
Released under the MIT License. See LICENSE.
Copyright © 2026 Vedran Hrabar
Vendored palettes and colourmaps keep their own licences; see kx/styles/palettes/LICENCES.md
and kx/styles/cmaps/LICENCES.md.



