Skip to content

Coming from Altair or lattice

The notebooks 05_vega_lite_style and 02_lattice_and_cleveland redo well-known figures of both.

Altair and Vega-Lite

Vega-Lite facets with the row, column and facet channels, and makes scales independent with resolve_scale:

Altair anytrellis
.facet(row="country") Trellis(df, row="country")
.facet(column="island") Trellis(df, col="island")
.facet(facet="year", columns=3) Trellis(df, wrap="year", cols=3)
.resolve_scale(y="independent") share_y=False
mark_point(), mark_line(), mark_bar() .points(), .line(), .bars()
mark_area(), mark_rect(), mark_text(), mark_rule() .area(), .rect(), .text(), .rule()
alt.Color("region") color="region"
alt.Color("cylinders:N"), "cylinders:O" .as_category("cylinders"), then color="cylinders"
alt.Tooltip(["country"]) tooltip=["country"]
alt.Scale(type="log") y_scale="log"
alt.selection_interval() a brush: built in (s, then drag)
alt.selection_point(fields=["region"], bind="legend") a legend click: built in
.repeat(row=[...], column=[...]) of a scatter plot pairs(df, [...])
alt.Chart(three).mark_point().encode(x="fertility", y="life_expect", color="region") \
    .facet(column="year")
import polars as pl

from anytrellis import Trellis
from anytrellis.datasets import load

three = load("gapminder").filter(pl.col("year").is_in([1955, 1980, 2005]))
plot = Trellis(three, col="year", panel_height=240)
plot.points(x="fertility", y="life_expect", color="region", size=6, tooltip=["country"])
plot

What differs: Altair writes a chart's specification for Vega-Lite to draw; anytrellis draws on the GPU in one canvas, so it stays smooth at millions of marks (Altair's default limit is 5,000 rows), and its selection comes back to Python as rows (selected()). Altair has many more kinds of chart and transforms.

Cleveland and lattice

Trellis displays come from Becker, Cleveland and Shyu (1996), built into S-PLUS and later R's lattice. Their rules are anytrellis's defaults: shared scales, axes only on the outer panels, labels in strips over the panels, and panels ordered by what they show.

lattice anytrellis
xyplot(y ~ x \| a) Trellis(df, wrap="a").points(x="x", y="y")
xyplot(y ~ x \| a * b) Trellis(df, row="a", col="b")
xyplot(y ~ x \| a * b * c) (a page per level of c) Trellis(df, row="a", col="b", page="c")
groups = g color="g"
layout = c(2, 3) cols=2
type = c("p", "r") .points() and .smooth(method="linear")
dotplot, bwplot, histogram, densityplot .points(), .box(), .hist(), .density()
splom(~ df[1:4], groups = g) pairs(df, ["a", "b", "c", "d"], color="g")
scales = list(y = "free") share_y=False
scales = list(y = "sliced") share_y="sliced"
index.cond, reorder() .order(...)

Cleveland's barley data, as a dot plot with sites and varieties ordered by their median yield:

dotplot(variety ~ yield | site, data = barley, groups = year, layout = c(2, 3))
barley = load("barley").with_columns(pl.col("year").cast(pl.String))
dots = (
    Trellis(barley, wrap="site", cols=2, panel_height=150, title="Barley yields at six sites",
            x_title="yield (bushels per acre)", y_title="")
    .order("site", by="median", of="yield", descending=True)
    .order("variety", by="median", of="yield")
    .points(x="yield", y="variety", color="year", size=7)
)
dots

Look at Morris: at every other site 1931 yielded more than 1932 for all but at most one variety, and at Morris 1932 yielded more for all ten. Cleveland suspected that the two years had been swapped by mistake. Ordered panels make it stand out.