Which libraries does this skill support?

It provides implementation guidance for popular libraries including e3nn (3D/E3), escnn (2D/discrete groups), and pytorch_geometric (permutation/graphs).

What are the benefits of equivariant architectures?

Equivariant architectures require fewer parameters, train faster, and generalize significantly better to unseen orientations or distribution shifts by baking physical symmetries directly into the model structure.

Can this skill help with 3D point clouds?

Yes, it specializes in E(3) and SO(3) symmetries commonly found in 3D data, robotics, and molecular modeling, recommending specific patterns like Tensor Field Networks.

Why do standard nonlinearities like ReLU break equivariance?

Standard nonlinearities often treat individual vector components independently, which can destroy the geometric relationship between features. This skill suggests norm-based or gated nonlinearities that preserve these symmetries.

Equivariant Architecture Designer

Name: Equivariant Architecture Designer
Author: lyndonkl

bylyndonkl

•

Data Science & ML

Designs neural network architectures that preserve symmetry groups to improve model efficiency and robustness.

The Equivariant Architecture Designer skill provides specialized guidance for building neural networks that respect physical or geometric symmetries, such as rotation, translation, and reflection. By leveraging domain-specific libraries like e3nn and escnn, it helps developers select the right architecture families, equivariant layers, and nonlinearities for tasks involving 3D point clouds, images, or molecular structures. This skill streamlines the complex process of building symmetry-aware models that generalize better with fewer parameters than standard deep learning approaches.

Key Features

01Support for 2D/3D domains, graphs, and point clouds

02Guidance on equivariant layers, nonlinearities, and pooling strategies

03Library recommendations for PyTorch and JAX including e3nn and escnn

04Standardized architecture specifications and implementation skeletons

05Symmetry group identification and architecture family mapping

068 GitHub stars

Use Cases

01Building E(3) equivariant networks for molecular property prediction

02Developing rotation-invariant models for medical image analysis

03Designing graph neural networks for permutation-invariant data processing

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add lyndonkl/claude equivariant-architecture-designer

For use in Claude.ai and ChatGPT

Download Skill