designspace.Chart

class designspace.Chart(*args, **kwargs)

Bases: Protocol

A monotone map from [0, 1] onto a parameter’s domain.

The single idea behind both sampling and solver integration. Declaring a prior is declaring a chart, so there is no separate “transform” concept: drawing a value means drawing a uniform u and applying from_unit, and a solver proposing in unit coordinates gets type-appropriate behaviour for free: a log-scaled parameter is perturbed multiplicatively, a quantized one snaps to its grid.

Charts are static: they are built once, at resolution, and never depend on a configuration. A parameter with expression bounds gets its chart from the bounds’ envelope, not from the values of a particular draw.

A chart lives on ParamDef.chart, except for a lifted parameter, whose chart is on ListDomain.element_chart.

Examples

>>> s = ds.space(ds.param("lr").real(1e-4, 1e-1).log_scale())
>>> chart = s.params["lr"].chart
>>> round(chart.from_unit(0.0), 6), round(chart.from_unit(1.0), 6)
(0.0001, 0.1)
>>> round(chart.from_unit(0.5), 6)
0.003162

from_unit(u: float) → Any

Map a unit coordinate to a value in the domain.

Parameters:

u (float) – A coordinate in [0, 1].

Returns:

The corresponding value. from_unit(0) is the domain’s lower end and from_unit(1) its upper.

Return type:

Any

to_unit(value: Any) → float

Map a value in the domain back to its unit coordinate.

The inverse of from_unit, where one exists. For a chart that quantizes (an integer or a grid), many values share a coordinate interval, so from_unit(to_unit(v)) == v holds while the reverse round trip does not.

Parameters:

value (Any) – A value in the parameter’s domain.

Returns:

The coordinate in [0, 1].

Return type:

float