Torch Delaunay - The Delaunay triangulation library for PyTorch

This is a fast library for computing Delaunay triangulation of 2-dimensional points.

The implementation is based on a sweep-algorithm, introduced by David Sinclair¹ and later improved by Volodymyr Agafonkin².

Here is an example of triangulation output produced by Torch Delaunay library for a random set of points:

Installation

The library is distributed as PyPI package, to install the package, execute the following command:

pip install torch_delaunay

You can use the torch_delaunay library for a fast computation of Delaunay triangulation for points defined as PyTorch tensors.

Usage

import torch
from torch_delaunay.functional import shull2d

# Compute Delaunay triangulation for randomly-generated 2-dimensional points.
points = torch.rand((100, 2))
simplices = shull2d(simplices)

Benchmarks

Benchmarks are collected on Apple M1 Pro chip with 32 GiB of memory.

Prob. Distribution	Data Type	Samples	Time	CPU	Iterations
Uniform	Float64	10000	4.18 ms	4.11 ms	171
Uniform	Float64	100000	49.6 ms	48.3 ms	15
Uniform	Float64	1000000	600 ms	592 ms	1
Uniform	Float32	10000	4.37 ms	4.29 ms	161
Uniform	Float32	100000	47.4 ms	46.4 ms	15
Uniform	Float32	1000000	573 ms	568 ms	1
Normal	Float64	10000	4.60 ms	4.49 ms	156
Normal	Float64	100000	48.5 ms	47.4 ms	15
Normal	Float64	1000000	568 ms	563 ms	1
Normal	Float32	10000	4.42 ms	4.33 ms	165
Normal	Float32	100000	46.5 ms	45.4 ms	15
Normal	Float32	1000000	550 ms	547 ms	1

License

The Torch Delaunay is distributed under GPLv3 license. See the LICENSE file for full license text.

Footnotes

1. David Sinclair - S-hull: a fast radial sweep-hull routine for Delaunay triangulation. ↩

2. Volodymyr Agafonkin - Delaunator: An incredibly fast and robust JavaScript library for Delaunay triangulation of 2D points. ↩