Partial-escnn (WIP)

This repository is an uncleaned and preliminary PyTorch implementation of the method proposed in the paper A Probabilistic Approach to Learning the Degree of Equivariance in Steerable CNNs, accepted at ICML 2024 (see reference below). This paper presents a novel approach to create Steerable CNNs with layer-wise learnable degree of $(\mathbb{R}^n, +) \rtimes H$ equivariance, where $H\leq O(n)$ is learnable.

The implementations in this repository are based on the escnn library, which is a PyTorch library to create steerable CNNs.

Requirements

Basic requirements

Python >= 3.10
torch
torchvision
escnn

Requirements for reproduction of all experiments/plots

wandb
plotly
matplotlib
seaborn
sklearn
pandas

Reference

Paper accepted at ICML 2024

@article{veefkind2024probabilistic,
  title={A Probabilistic Approach to Learning the Degree of Equivariance in Steerable CNNs},
  author={Veefkind, Lars and Cesa, Gabriele},
  journal={arXiv preprint arXiv:2406.03946},
  year={2024}
}