Tunny Icon
TunnyDocs

The next-gen Grasshopper optimization tool.

Violin Plot

Violin Plot uses the same Value, Group by, and Normalize [0,1] controls as Box Plot, with independent selections. Choose one objective or numeric parameter; Group by = None shows one distribution. Categorical parameters and explicitly declared Optuna integer or stepped numeric parameters can split it into groups. Continuous parameters are not grouped or binned. CSV numeric parameters without distribution metadata are Value candidates only; parsed CSV categorical parameters can group the Value.

Normalization uses the finite minimum and maximum of the entire selected Value column before grouping, not each group's range. A constant Value column becomes zero. Missing/non-finite Values and missing/non-finite grouping values are excluded as aligned trial pairs. Actual empty-string categories remain distinct from missing parameters and appear as (empty string). Colliding typed category labels are disambiguated consistently in the chart and CSV. Unavailable Value selections fall back to the first objective, otherwise the first numeric parameter; unavailable Group by selections become None.

Reading the plot

Each filled violin is a Gaussian kernel density estimate of the selected Value in its group. The Y-axis identifies Value; the X-axis labels identify groups. Numeric groups are ordered numerically; categorical groups have deterministic label order. Each violin has the same maximum width and a median marker. Width does not encode the number of trials.

A group with fewer than two finite Values or constant Values cannot be estimated and is omitted. The widget reports the skipped count and reasons, even if every group is skipped. A non-finite or zero KDE bandwidth also prevents estimation. Use Box Plot when singleton or constant groups need to be shown.

CSV export

CSV contains one row per displayed violin/grid point, with columns value, group_by, scale, group, group_identity, grid, and density. The selected Value, grouping parameter, labels, and raw/normalized scale match the chart. Each displayed group has 128 grid points; skipped groups have no rows. No data is exported when all groups are skipped. group is the displayed label; group_identity retains the original category type or numeric group value, keeping identities unambiguous even after CSV safety escaping.