Teaching neural nets to draw with sine waves

What it does

SIREN swaps the usual ReLUs in a multilayer perceptron for sine activations, letting the network represent fine details—an image, a waveform, a 3D shape—as a continuous function of raw coordinates. This repo packages that idea into SirenNet and Siren layers, plus a SirenWrapper that handles the busywork of teaching an MLP to memorize one specific image from (x, y) inputs.

The interesting bit

The first layer’s w0_initial acts like a tuning knob for frequency: too low and the network blurs the details, too high and it hallucinates high-frequency noise. The repo also quietly implements a follow-up paper, letting you steer the whole network with a latent vector that elementwise-modulates every hidden layer.

Key highlights

• SirenNet exposes a full sine-activated MLP for any coordinate-based regression task.
• SirenWrapper abstracts the drudgery of overfitting a single image: instantiate with height and width, train on a tensor, then call it bare to reconstruct.
• Latent conditioning is available via a latent_dim argument, wiring in a modulator feedforward network per the 2021 paper.
• Sine is included as a standalone activation wrapper around torch.sin if you want to roll your own.

Caveats

• The wrapper API is clearly designed to overfit one image at a time; the README is vague on whether it generalizes to batched training across a dataset.
• No benchmarks, pretrained weights, or examples beyond the basic API snippets are shown.

Verdict

Grab it if you are prototyping neural representations, image fitting, or coordinate-based generative tricks. Look elsewhere if you need a general vision backbone or a scalable data-loader-friendly trainer.