Decluttering Plots¶

Skill version 0.2.0

Dashboard and presentation plots carry Bokeh "chart junk" by default: a toolbar, wheel-zoom that fires when you scroll the page, gridlines on both axes, and repeated axis labels when charts are stacked. This reference covers the .opts() mechanics to strip that — it is the how (mechanical styling). For the what (which series to emphasize, annotation, color as meaning) see Distilling Explanatory Plots. These options apply to any Bokeh-backed HoloViews object, hvPlot output included — see hvPlot vs .opts for the two APIs' division of labour.

Contents¶

Lookup — Bokeh model properties reference
Apply at the top level
Toolbar and tools (disable wheel-zoom)
Gridlines on one axis
Hiding and sharing axes
Legend placement
Nested categorical axes
hvPlot vs .opts

Lookup¶

Bokeh grid/axis/toolbar model properties: https://docs.bokeh.org/en/latest/docs/reference/models/plots.html.

Apply at the top level¶

The most common mistake: setting tool/toolbar options per element on an Overlay or Layout. They often don't stick, because the composite plot builds its own merged toolbar. Apply them to the top-level object you actually render.

# WRONG — per-element on an overlay; the merged toolbar still shows wheel-zoom
overlay = curve.opts(default_tools=[]) * scatter.opts(default_tools=[])

# CORRECT — set on the overlay itself
overlay = (curve * scatter).opts(toolbar=None, default_tools=[], active_tools=[])

A one-line helper keeps every chart consistent — apply it right before you hand the object to pn.pane.HoloViews(...):

def clean(plot):
    return plot.opts(toolbar=None, default_tools=[], active_tools=[])

For a whole family of charts, prefer setting these once for the session with hv.opts.defaults(...) instead of a per-object helper — see Session Defaults, which also covers why hvPlot's sizing options are the one thing that cannot be defaulted this way.

Applying options by element type¶

When different elements in a composite need different styles, chain .opts() calls targeting each element type. This applies the options to every matching element in the composite:

overlay = (curve * scatter).opts(
    "Curve", **curve_opts
).opts(
    "Scatter", **scatter_opts
)

Group-style options (Overlay, NdLayout, NdGrid, …) apply to the container itself, so you can mix container-level and per-element-level options in one chain:

overlay = (curve * scatter).opts(
    "Curve", **curve_opts
).opts(
    "Scatter", **scatter_opts
).opts(
    "Overlay", **overlay_opts
).opts(
    "NdLayout", **ndlayout_opts
)

toolbar=None removes the toolbar entirely (the Bokeh logo + pan/zoom/save icons).
autohide_toolbar=True keeps the toolbar but hides it until the pointer enters the plot. Reach for this instead of toolbar=None when the chart is still meant to be explored and you only want the icons out of the way while it's being read — you keep reset/save/hover without the permanent furniture. It sets Bokeh's toolbar.autohide, and has no effect under toolbar=None, since that leaves no toolbar to reveal.
default_tools=[] strips Bokeh's default tools — pan, wheel_zoom, box_zoom, save, reset, help. This is what actually stops the chart zooming when the user scrolls the page. Hiding the toolbar alone is not enough if wheel-zoom is still an active scroll tool.
active_tools=[] guarantees nothing is active on load (belt-and-suspenders so a scroll does nothing).
Add back only the tools you want with tools=[...], e.g. tools=["hover"]. Don't blanket tools=[] if you want hover — it removes hover too.

# Static/presentation chart — no toolbar at all, nothing zooms on scroll
plot.opts(toolbar=None, default_tools=[], active_tools=[], tools=["hover"])

# Still interactive, but the toolbar only appears on hover
plot.opts(autohide_toolbar=True, default_tools=["reset"], active_tools=[], tools=["hover"])

default_tools=["reset"] keeps just the reset button; default_tools=[] keeps none.

autohide_toolbar is spelled the same in both APIs — it is valid as an hvPlot call kwarg (df.hvplot.line(..., autohide_toolbar=True)) as well as via .opts(), and it takes in hv.opts.defaults(...), so a whole dashboard can adopt hover-only toolbars in one line (see Session Defaults).

Gridlines on one axis¶

show_grid=True draws gridlines on both axes. To keep only horizontal (y) gridlines — usually what aids reading values — set the x-grid's alpha to 0 via gridstyle:

plot.opts(show_grid=True, gridstyle={"xgrid_line_alpha": 0, "ygrid_line_alpha": 0.35})

gridstyle keys are prefixed xgrid_ / ygrid_; zero the *_line_alpha of the axis you want gone. (backend_opts={"xgrid.grid_line_color": None} also works, but gridstyle is the HoloViews-native way.)

xaxis=None / yaxis=None hide an axis entirely (ticks and label). xaxis="bare" keeps the axis line but drops ticks and labels.
When stacking related charts in a column that share one x-axis, show ticks + xlabel only on the bottom chart and set xaxis=None on the rest. This removes the repeated axis that otherwise appears under every panel:

top.opts(xaxis=None, xlabel="")        # and every panel except the last
bottom.opts(xlabel="model / task")     # only the bottom panel labels the shared axis

Rotate long categorical tick labels rather than letting them collide: .opts(xrotation=25) (yrotation= for the other axis).

Legend placement¶

legend_position="top_right" places the legend inside the plot at that corner. Prefer an inside corner over the default, which can render a wide external legend column that squeezes the plot into a sliver.
Positions: top_left, top_right, bottom_left, bottom_right, plus top/bottom/left/right.
show_legend=False drops it entirely. In a stack of related charts, show the legend on the first panel only and set show_legend=False on the rest, so it isn't repeated down the column.
A legend with a single entry distinguishes nothing — suppress it with show_legend=len(groups) > 1 (see Opts System).
Legend options belong on the container (Overlay/NdOverlay), not the elements inside it — see Apply at the top level.

overlay.opts(legend_position="top_right", show_legend=len(groups) > 1)

Nested categorical axes¶

Collapsing two grouping columns into one flat "a · b" label forces long, angled ticks. Bars, Violin, and BoxWhisker accept multiple kdims for a nested categorical x-axis — the inner group sits under the outer group, which reads far cleaner. Order matters: the last kdim is the inner level.

hv.Bars(df, kdims=["model", "task"], vdims=["count"])            # task nested under model
hv.Violin(df, kdims=["model", "task"], vdims=["latency_ms"])

Bars draws the nested axis via multi_level=True (the default). .opts(multi_level=False) flattens it back to one tick per outer group, and .opts(stacked=True) turns the second kdim into stack segments instead of a second axis level. Bars accepts at most 3 kdims.

Caveat: a jittered Scatter/Points overlay can't align to a two-level categorical x. If you need raw points over the violin/box, build a single flat combined key dimension instead:

df["model · task"] = df["model"] + " · " + df["task"]
hv.Violin(df, kdims=["model · task"], vdims=["latency_ms"]) * hv.Scatter(
    df, "model · task", "latency_ms"
).opts(jitter=0.3, alpha=0.3)

Otherwise keep the nested kdims and drop the overlay.

hvPlot vs .opts¶

Everything above is written as .opts() on HoloViews elements. If you arrived here holding hvPlot output, it is a HoloViews object, so every option applies unchanged — but hvPlot spells several of them differently in its own call signature:

This reference (`.opts()`)	hvPlot call kwarg
`xrotation=25`	`rot=25`
`legend_position="top_right"`	`legend="top_right"`
`show_legend=False`	`legend=False`
`show_grid=True` (both axes)	`grid=True`
multiple `kdims=[...]`	`by=[...]`

rot, legend, and grid are hvPlot-only — passing them to .opts() raises. The reverse also holds: gridstyle has no hvPlot kwarg, so one-axis gridlines always go through .opts().

The decluttering options here (toolbar, default_tools, active_tools, tools, show_grid, gridstyle, xaxis) are plot options. Pass them via .opts() — reliable on overlays and layouts — or as hvPlot kwargs for a single element. When in doubt, use .opts() on the final object (see Apply at the top level).
The one thing that does not work via .opts(): responsive=True and height= must be hvPlot call arguments. See Responsive Sizing for why.