TorchVinecopBackend

class TorchVinecopBackend(*, controls=None, structure=None)

PyTorch backend. Wraps pyvinecopulib.torch.TorchVinecop.

Pick this backend for GPU placement (.to("cuda")), autograd through the vine cascade, or composition with other torch.nn.Module code. The default VinecopBackend is generally faster on CPU for the same problem.

Constructing this class imports torch — the explicit opt-in signal that PyTorch is required. The default FitControlsTorchVinecop is materialised lazily at fit time.

Parameters:

controlsFitControlsTorchVinecop, or None, optional: Cascade / placement / precision knobs plus the native (pure-torch) structure-selection knobs (tree_algorithm / seeds / trunc_lvl / tree_criterion / threshold). None resolves to defaults at fit time.
structureRVineStructure, or None, optional: A pre-specified vine structure; when provided, fit_vine skips structure selection.

Methods

cdf
fit_vine
pdf
simulate
with_num_threads

Notes

with_num_threads is a no-op on this backend; for CPU intraop parallelism call torch.set_num_threads(N) globally before evaluating. The sklearn forest’s outer parallelism (via joblib) still applies independently.

Methods

`__init__`
`cdf`
`fit_vine`
`pdf`
`simulate`
`structure_of`
`with_local_random`
`with_num_threads`
`with_random_structure`