seaborn Themes and Palettes

Core Principle: Seaborn theme configurations establish global graphic aesthetics, tick layouts, font scalings, and color palettes across both Seaborn and direct Matplotlib figures.

set_theme

sns.set_theme(
    style="whitegrid",       # darkgrid | whitegrid | dark | white | ticks
    context="paper",         # paper | notebook | talk | poster
    palette="colorblind",
    font="Arial",
    font_scale=1.0,
    rc={"figure.figsize": (5, 3.5), "pdf.fonttype": 42},
)

This writes into matplotlib's rcParams, so it affects everything, not just seaborn calls. See rcParams and Style Sheets.

Styles

style	Look
darkgrid	grey background, white grid (seaborn's default)
whitegrid	white background, grey grid
dark	grey background, no grid
white	white background, no grid
ticks	white, no grid, tick marks: closest to journal convention

For publication, ticks or whitegrid with sns.despine():

sns.set_theme(style="ticks")
sns.despine()                                  # remove top and right spines
sns.despine(trim=True, offset=5)               # trim to data range, detach
sns.despine(fig=fig, left=True)                # whole figure

offset=5 detaching the axes from the plot area is a small touch that reads as deliberate design.

Context

sns.set_context("paper")     # smallest: journal figures
sns.set_context("notebook")  # default
sns.set_context("talk")      # slides
sns.set_context("poster")    # largest
sns.set_context("talk", font_scale=1.2, rc={"lines.linewidth": 2.5})

Context scales all font sizes, line widths and marker sizes together. Converting a paper figure to a talk figure is one changed word, not fifteen changed numbers.

Palettes

sns.color_palette()                        # current default
sns.color_palette("colorblind", 8)
sns.color_palette("deep")                  # seaborn's default
sns.color_palette("muted"); sns.color_palette("pastel")
sns.color_palette("Set2", 6)               # ColorBrewer qualitative
sns.color_palette("husl", 8)               # evenly spaced hues

sns.color_palette("rocket", as_cmap=True)      # sequential
sns.color_palette("mako", as_cmap=True)
sns.cubehelix_palette(8, start=.5, rot=-.75)   # greyscale-safe sequential
sns.diverging_palette(240, 10, as_cmap=True)   # custom diverging
sns.light_palette("seagreen", as_cmap=True)
sns.dark_palette("#69d", reverse=True)

sns.set_palette("colorblind")                   # set globally
sns.color_palette("colorblind")                 # display in a notebook

rocket, mako, flare, crest, vlag, icefire are seaborn's own perceptually uniform maps and are all good defaults. vlag is the diverging one you want for correlation and z-score heatmaps.

Explicit category colours

The habit that makes a multi-panel figure coherent:

CONDITION_COLORS = {
    "control": "#4C72B0",
    "treated": "#DD8452",
    "drug":    "#55A868",
}

sns.scatterplot(data=df, x="a", y="b", hue="condition", palette=CONDITION_COLORS)
sns.boxplot(data=long, x="condition", y="expr", palette=CONDITION_COLORS)
sns.lineplot(data=ts, x="t", y="v", hue="condition", palette=CONDITION_COLORS)

Without an explicit mapping, seaborn assigns colours in the order categories appear in each subset: so "treated" can be blue in panel A and orange in panel B if panel B happens to lack the control group. Define the dict once. See Colormaps and Color.

Category order

sns.boxplot(data=long, x="condition", y="expr",
            order=["control", "low", "high"],
            hue_order=["male", "female"])

Or better, set it once with an ordered categorical, which fixes plot order and sort order and groupby order simultaneously:

long["condition"] = pd.Categorical(
    long["condition"], categories=["control", "low", "high"], ordered=True
)

See Categorical dtype.

hue with continuous data

sns.scatterplot(data=df, x="a", y="b", hue="expression",
                palette="viridis", hue_norm=(0, 10))

seaborn produces a legend of sample values rather than a colorbar. For a proper colorbar, add one manually:

fig, ax = plt.subplots()
sc = ax.scatter(df["a"], df["b"], c=df["expression"], cmap="viridis", s=10)
fig.colorbar(sc, ax=ax, label="log$_2$ CPM")

Raw matplotlib is the better tool for continuous colour encoding.

Temporary changes

with sns.axes_style("white"):
    fig, ax = plt.subplots()
    ...

with sns.plotting_context("talk"):
    ...

with sns.color_palette("Set2"):
    ...

Use these in library code rather than calling set_theme() globally: you should not silently restyle your users' plots.

Inspecting and resetting

sns.axes_style()               # current style dict
sns.plotting_context()         # current context dict
sns.reset_defaults()           # back to matplotlib defaults
sns.reset_orig()               # back to matplotlib's original rcParams

A project header

import matplotlib as mpl, matplotlib.pyplot as plt, seaborn as sns

sns.set_theme(style="ticks", context="paper", palette="colorblind",
              font="Arial", rc={"figure.figsize": (3.5, 2.6)})
mpl.rcParams.update({
    "pdf.fonttype": 42, "ps.fonttype": 42, "svg.fonttype": "none",
    "savefig.bbox": "tight", "savefig.dpi": 300,
    "axes.titlelocation": "left",
    "font.sans-serif": ["Arial", "Helvetica", "DejaVu Sans"],
})

CONDITION_COLORS = {"control": "#4C72B0", "treated": "#DD8452"}
CHROM_ORDER = [f"chr{i}" for i in range(1, 23)] + ["chrX", "chrY", "chrM"]

Put this in a plotstyle.py and import it everywhere in the project. Every figure then matches, and switching the whole paper to a different palette is a one-line change.

Bioinformatics conventions

# expression heatmap: z-scored, diverging, centred
sns.clustermap(mat, z_score=0, cmap="vlag", center=0)

# raw counts heatmap: sequential
sns.heatmap(cpm, cmap="rocket", norm=LogNorm())

# correlation: diverging, symmetric range
sns.heatmap(corr, cmap="vlag", center=0, vmin=-1, vmax=1, square=True)

# categorical sample groups
sns.set_palette("colorblind")

# nucleotides
BASE_COLORS = {"A": "#3DA853", "C": "#4285F4", "G": "#F9AB00", "T": "#EA4335"}

# significance: grey bulk, one accent
sns.scatterplot(data=de, x="log2fc", y="neglog10p", hue="sig",
                palette={True: "crimson", False: "lightgrey"},
                hue_order=[False, True])

hue_order=[False, True] in the last example matters: it draws the grey non-significant points first, so the crimson hits land on top rather than being buried.

Common mistakes

• sns.set(): deprecated alias for set_theme().
• Inconsistent category colours across panels.
• Not setting context="paper" for journal figures, so the fonts are notebook-sized.
• darkgrid in a publication. Grey backgrounds waste ink and reproduce badly.
• Global set_theme() in a library.
• Continuous hue producing a sample legend instead of a colorbar.
• Assuming Arial is installed. Provide fallbacks.
• Forgetting sns.despine() after choosing style="ticks".

See also

Colormaps and Color · rcParams and Style Sheets · Seaborn · Categorical dtype · Saving Figures · Faceting

Exercise

seaborn is not available in this sandbox, so derive the palette values yourself from a matplotlib colormap. Write palette_map(categories, subset_a, subset_b). The first input line is the canonical category order; the second and third are two subsets of it, standing in for two panels of one figure. Sample the viridis colormap evenly over the canonical order: category i of n takes the position i / (n - 1), or 0.0 when n is 1: and convert each colour with matplotlib.colors.to_hex. Build that mapping once from the canonical order and look both subsets up in it; a mapping rebuilt from each subset's own categories is what makes a condition blue in one panel and orange in the next.

Return four lines: palette= followed by category=#hex for every canonical category in order, space separated; A: and B: followed by the same rendering for each subset, or empty when a subset has no categories, with any category absent from the canonical list rendered as unmapped; and shared=<n> consistent=<bool>, counting the categories the two subsets have in common and reporting whether every category either subset names was found in the mapping. Return the single line none when the canonical list is empty.