What is model equivariance auditing?

It is the process of numerically verifying that a machine learning model respects specific symmetries, ensuring that transforming the input results in a predictable transformation of the output.

Why is gradient equivariance important?

If gradients are not equivariant, the model may 'unlearn' symmetry during training, leading to instability and poor generalization even if the forward pass seems correct.

What frameworks are supported by the auditor?

The auditor provides logic and templates primarily focused on PyTorch, but the methodology can be applied to any tensor-based framework like TensorFlow or JAX.

Can this skill help find which specific layer is broken?

Yes, it includes a layer-wise testing protocol that isolates individual modules to pinpoint exactly where equivariance is lost in the architecture.

Model Equivariance Auditor

Name: Model Equivariance Auditor
Author: lyndonkl

bylyndonkl

•

Data Science & ML

Verifies that machine learning models correctly respect intended symmetries through systematic numerical tests and debugging guidance.

The Model Equivariance Auditor is a specialized capability for developers working with geometric deep learning and symmetry-aware architectures. It provides a structured framework to verify that models claiming equivariance or invariance actually adhere to mathematical constraints. By offering automated testing templates for forward passes, layer-wise isolation, and gradient transformations, the skill helps identify subtle implementation bugs that lead to poor training and inconsistent predictions. It is an essential tool for ensuring the integrity of AI models in fields like computer vision, physics-informed neural networks, and molecule modeling.

Key Features

01Standardized audit reporting with error statistics and diagnostics

02Pre-built test templates for PyTorch and custom model layers

03End-to-end numerical equivariance and invariance verification

04Gradient equivariance checking for training stability analysis

058 GitHub stars

06Layer-wise testing protocols to isolate symmetry-breaking components

Use Cases

01Debugging custom equivariant layers in geometric deep learning projects

02Auditing third-party model implementations for mathematical correctness

03Verifying that data augmentation and model architecture symmetries are aligned

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add lyndonkl/claude model-equivariance-auditor

For use in Claude.ai and ChatGPT

Download Skill