oineus.diff.cech_delaunay_filtration

oineus.diff.cech_delaunay_filtration(points, eps=0.0, packed=False, print_time=False)[source]

Build a differentiable Cech-Delaunay filtration from a point cloud.

The combinatorics of the alpha complex are computed via diode (CGAL); the filtration values are recomputed differentiably as squared minimum enclosing ball radii of each simplex, so gradients flow back to points.

Parameters:

  • points – (n, d) torch.Tensor with d in {2, 3}. Differentiable.

  • eps (float) – Small value for numerical stability in the MEB computation.

  • packed (bool) – Use the compact bit-packed cell encoding for the Delaunay combinatorics when the vertex ids fit a 64/128-bit word. The values (and gradients) are recomputed here regardless of encoding.

  • print_time (bool) – If True, print per-stage timings.

Returns:

DiffFiltration whose values are squared MEB radii.