siren-pytorch

Why we included this project

SIREN packs a widely studied neural architecture into a small, pip-installable PyTorch library. If you work with coordinate-based models, where one network maps input coordinates to values such as image pixels, audio samples, or spatial fields, this gives you the sine activation, a standalone Siren layer, and a configurable SirenNet where you set omega_0 and the number of layers. The SirenWrapper trains the network on a real image and reconstructs it in a single forward pass, and latent code modulation lets you condition the model for different outputs. The code is short and readable, so it works both as a component to drop into a project and as a base to modify for research. Anyone comparing implicit neural representation approaches saves themselves the work of reimplementing the core from the paper.