jejjohnson / knerf

Minimal neural fields with keras-core.

Neural Fields with Keras-Core (In Progress)

This package implements some minimal functionality to train neural fields using keras-core library. It will also feature some extra functionality when working with spatiotemporal datasets. I would eventually like it to have an efficient implementation for introducing modulation within each of the models.

---

Key Algorithms

Fourier Features. The canonical method for Fourier Features.

SIREN. The canonical method that uses sinusoidal activation functions with a special initialization procedure.

FourierNet (TODO)

GaborNet (TODO)

Spherical Harmonics (TODO)

Modulation (TODO)

---

Installation

We can install it directly through pip

pip install git+https://github.com/jejjohnson/knerf

We also use poetry for the development environment.

git clone https://github.com/jejjohnson/knerf.git
cd knerf
conda create -n knerf python=3.11 poetry
poetry install

---

References

Software

• jaxdf - Arbitrary Discretizations
• diffrax (example) - example of diffrax (and equinox) with an finite difference discretization