FitControlsTorchVinecop

class FitControlsTorchVinecop(bicop_controls=<factory>, trunc_lvl=20, tree_criterion='tau', threshold=0.0, tree_algorithm='mst_prim', seeds=<factory>, cache_integrals=True, device=None, dtype=None, batched=False)

Controls for from_data() and the cascade.

Mirrors FitControlsVinecop: bundles all vine-fit knobs into one object. A nested FitControlsTorchBicop controls how each pair-copula is fit; the vine-level fields below carry the structure-selection knobs plus placement / precision / cascade-variant settings.

Attributes:

bicop_controlsFitControlsTorchBicop

Controls applied to every pair-copula fit.

trunc_lvlint, default=20

Maximum number of trees to select when from_data() is called with structure=None.

tree_criterion{“tau”, “rho”, “hoeffd”}, default=”tau”

Dependence measure used to weight candidate edges during structure selection (passed to wdm).

thresholdfloat, default=0.0

Dependence threshold: candidate edges below it are deprioritized during spanning-tree selection.

tree_algorithm{“mst_prim”, “mst_kruskal”, “random_weighted”, “random_unweighted”}, default=”mst_prim”

Spanning-tree algorithm for structure selection: Dissmann’s maximum spanning tree (mst_*) or Wilson’s (weighted / uniform) random spanning tree (random_*).

seedslist of int, default=[]

RNG seeds for the random tree algorithms (ignored by the MST ones).

cache_integralsbool, default=True

If True, precompute the cdf / hfunc / hinv caches on every pair copula’s interpolation grid. Cached lookups are 1–2 orders of magnitude faster than the on-the-fly path with a ~1e-3 IAE cost.

devicetorch.device or None, default=None

Target torch device for the fitted pair copulas. None keeps the input’s device.

dtypetorch.dtype or None, default=None

Target torch dtype. None defaults to torch.float64 (parity with Vinecop).

batchedbool, default=False

If True, fires a single batched bicop call per tree level. Available for pdf / rosenblatt only (inverse_rosenblatt(batched=True) raises).

Notes

Structure selection runs natively on the torch interpolation grids. It is continuous-only and TLL-only, and the criteria for automatic truncation / thresholding (aic / bic / mbicv) are not available here: trunc_lvl is a fixed cap.

Attributes

batched
cache_integrals
device
dtype
threshold
tree_algorithm
tree_criterion
trunc_lvl
bicop_controls
seeds

Methods

__init__