symm-learning 0.1.8

Torch modules and utilities of equivariant/invariant learning

Symmetric Learning

Lightweight python package for doing geometric deep learning using ESCNN. This package simply holds:

• Generic equivariant torch models and modules that are not present in ESCNN.
• Linear algebra utilities when working with symmetric vector spaces.
• Statistics utilities for symmetric random variables.

Installation

pip install symm-learning
# or
git clone https://github.com/Danfoa/symmetric_learning
cd symmetric_learning
pip install -e .

Structure:

Linear Algebra

• lstsq: Symmetry-aware computation of the least-squares solution to a linear system of equations with symmetric input-output data.
• invariant_orthogonal_projector: Computes the orthogonal projection to the invariant subspace of a symmetric vector space.

Statistics

• var_mean: Symmetry-aware computation of the variance and mean of a symmetric random variable.
• cov: Symmetry-aware computation of the covariance / cross-covariance of two symmetric random variables.

Models

• iMLP: Invariant MLP for learning invariant functions.
• eMLP: Equivariant MLP for learning equivariant functions.

Torch Modules

• Change2DisentangledBasis: Module for changing the basis of a tensor to a disentangled / isotypic basis.
• IrrepSubspaceNormPooling: Module for extracting invariant features from a geometric tensor, giving one feature per irreducible subspace/representation.