PolyWeave

Multiplicative (Sigma-Pi) layers and hypernetworks for weight generation.

PolyWeave is a small, modular PyTorch library for building hypernetworks — networks that generate the weights of another network — with an emphasis on multiplicative (Sigma-Pi) computation. Alongside a vanilla additive teacher, it provides Sigma-Pi layers whose multiplicative branch is a genuine geometric product, plus diagnostics that let you measure when multiplicative interactions actually help rather than assuming they do.

The library underpins an ongoing series of papers on multiplicative hypernetworks.

Status

v0.2.0 — alpha. The public API may still shift before it stabilises.

Where to go next

• Getting Started — install the library and run your first Sigma-Pi layer.
• Concepts — what the sigma and pi branches compute, and how the recruitment diagnostics read multiplicative structure.
• API Reference — auto-generated documentation for every public module.
• Blog — plain-language write-ups of the research.

Quick taste

import torch
from polyweave import ConvSigmaPi2d, SigmaPiLinear

# A channels-preserving Sigma-Pi conv block (additive + geometric-product branch).
block = ConvSigmaPi2d(channels=32, kernel_size=3)
block.eval()
y = block(torch.randn(8, 32, 28, 28))
print(block.pi_scale_mean())     # gate amplitude — how strongly pi is recruited
print(block.exponent_abs_mean()) # product shape — how far it departs from a no-op

# A Sigma-Pi fully-connected layer (raw pre-activation, drop-in for nn.Linear).
fc = SigmaPiLinear(in_features=128, out_features=64)
out = fc(torch.randn(8, 128))