iskra ✨ Modern Geometry Processing

Lightweight geometry processing library that is a one-stop shop for all your geometric needs. Iskra is:

• modern and Python-first,
• simple by default, powerful when needed,
• fully differentiable and compatible with machine learning,
• CPU and GPU enabled,
• actievely maintained.

Support the project by starring it:

Computing vertex normals in iskra is as simple as:

import torch

from iskra.geometry import triangle_normals
from iskra.mesh import Mesh
from iskra.topology import face_index, reduce_on_subface

mesh, _ = Mesh.from_path(
    "oded://objects/koala/koala_low_resolution.obj",
    device="cpu"
)
verts = mesh.vertices  # [V, 3]
faces = mesh.faces  # [F, 3]

tris = face_index(verts, faces)  # [F, 3, 3]

tri_normals = triangle_normals(tris)  # [F, 3]

vert_normals = reduce_on_subface(tri_normals, faces, verts.shape[0], "sum")  # [V, 3]
vert_normals = torch.nn.functional.normalize(vert_normals, dim=-1)  # [V, 3]

You will need PyTorch installed for iskra ✨ to work: see PyTorch installation instructions here.

pip install torch --index-url ... # your preferred PyTorch distribution

The code has been tested with torch==2.12, but will likely work with other versions too.

Finally, install iskra ✨ to your active environment using:

pip install -e git+https://github.com/anadodik/iskra/

Lastly, if you plan on contributing, you will need the development dependencies and to compile the C++ extensions in editable mode. This can be done by running the following:

conda env create -f environment.yaml
conda env update -f environment-dev.yaml
conda activate iskra
pip install --no-build-isolation -Ceditable.rebuild=true -ve .

1. Why the name? Iskra means “spark” in Serbo-Croatian: a spark enables using (a) torch. We also expect our system to be the spark that ignites exciting research in geometry. Most importantly, it sounds cool.
2. How much of iskra ✨ is LLM generated? Iskra started back in 2022 at the start of my PhD because I enjoy writing and learning geometry algorithms. It is therefore almost entirely good old fashioned free range human generated slop, except for a select few parts. These are clearly marked in the codebase. [As a side-note, doing it this way had some benefits beyond being fun: being deeply bonded with geometry code in PyTorch lead me to come up with the tensor-based scatter-gather abstraction and I do not think I would have done so had I computer-slopped it together!]