Synthetic training data
Install the training dependencies with
python -m pip install 'vine-denoising-copula[train]'.
Training examples pair samples with a density evaluated on a uniform midpoint grid. The grid is normalized by its midpoint integral; this discretization is an approximation to the continuous copula, particularly near strong tail peaks. A mixture target is the weighted sum of its component grids. Component sample counts are drawn from the same mixture weights.
The maintained generator supports Gaussian, Student-t, Clayton, Gumbel, Frank, Joe, BB1, BB7 and independence copulas, including rotations and mixtures. Complex grid-defined copulas are also available as individual examples. Joe, BB1 and BB7 sampling and BB-family density evaluation use pyvinecopulib 0.7.5. Tests compare sample margins and joint probabilities with closed-form CDFs, and densities with mixed derivatives of those CDFs.
The conditional_* options in historical configurations
are rejected. They paired a fixed-condition inverse-CDF sample with its
own uniform driver, so the two columns did not form a bivariate copula
and did not match the density target. Use ordinary copula families for a
new training run. Higher-tree copulas also have uniform conditional
margins; fixing one argument of an h-function is not a substitute for
generating such a pair.
These corrections affect new training and generated evaluation data. They do not modify the released checkpoint, its hash, or inference. Historical training streams require their original source pin. Paper experiment instructions retain those pins; running an old configuration with the corrected generator is not an exact replay of the original training run.
The standalone conditional CDF and inverse-CDF functions remain available for conditional diagnostics. Unsupported families and invalid training targets raise errors rather than silently substituting an independence target.