model-equivariance-auditor

Verify PyTorch neural network equivariance by running systematic end-to-end tests that check if model predictions change predictably under specified symmetries like rotation or translation. This process isolates custom layer violations and confirms global symmetry adherence.

Equivariant models produce inconsistent results under rotation due to subtle bugs in custom layers, configurations, numerical precision, or integration. Layer-wise debugging and gradient symmetry checks isolate the specific layers violating intended symmetry properties.

You can check gradient symmetry to ensure stable training by verifying that model gradients respect intended symmetries. This confirms that parameter updates remain consistent with geometric constraints during backpropagation.

You need PyTorch to test model invariance and equivariance, as the numerical computations and testing frameworks require PyTorch along with numpy and scipy dependencies to execute systematic verification tests.

The best way to debug symmetry violations in custom neural network layers is through layer-wise analysis, which isolates specific layers breaking equivariance properties and tests numerical precision issues within your model architecture.