Rise of the shadow dolphins

Sinusoidal representation networks (SIREN) can be used to parametrise any scalar or vector field $\mathbb{R}^n \to \mathbb{R}^m$, converge fast during training and are useful as physics-informed neural networks due to continuous differentiability. This repository contains a JAX implementation of SIREN using Equinox. Essentially, this means a multilayer perceptron (MLP) and special initialisation of weights & biases.

Getting started

Make sure you have installed JAX and Equinox. Copy src.py into your own project. Instances of SIREN can be created like

import jax
from src import SIREN

siren = SIREN(
    num_channels_in=2,                 # n (e.g. image grid)
    num_channels_out=3,                # m (e.g. RGB values)
    num_layers=4,
    num_latent_channels=1024,
    omega=30,                          # angular frequency
    rng_key=jax.random.PRNGKey(420)
)

Training SIREN

For an example on how to train the SIREN in Equinox, look at main.py. After installing Optax, scikit-image and tqdm, you can fit an image img.png via

python main.py --path_to_image img.png --num_epochs 1000