Equivariant Mesh Attention Networks

This repository contains the code to reproduce the experiments of Equivariant Mesh Attention Networks published in Transactions on Machine Learning Research (TMLR - 08/2022).

Running experiments

The instructions provided below assume that the python command is triggered from ./eman:

FAUST experiments

python experiments/faust_direct.py --model RelTanEMAN --seed 1 --epochs 1 -equiv_bias

TOSCA experiments

python experiments/tosca_direct.py --model RelTanEMAN --seed 1 --epochs 1 -equiv_bias -null_isolated

Installation instructions

Follow the commands below to create a new conda environment and install all dependencies:

conda create --name eman python=3.7
conda activate eman

# GPU installation
# conda install pytorch=1.11 cudatoolkit=11.3 -c pytorch

# CPU installation
# conda install pytorch=1.11 cpuonly -c pytorch

conda install pyg=2.0.3 -c pyg
pip install wandb pytorch-ignite openmesh opt_einsum trimesh

Project structure

eman
│   README.md
│   LICENSE    
│
└───data
│   │   FAUST/raw/MPI-FAUST.zip  # Download from http://faust.is.tue.mpg.de/
│   │   TOSCA               # Automatically downloaded on first experiment
|
└───eman                    # Implementation of Equivariant Mesh Attention Networks
│   └───nn
│   └───tests
│   └───transform
│   └───utils
|
└───experiments
|   |   faust_direct.py 
|   |   tosca_direct.py 
|   |   paths.json          # Specify dataset locations (default: "./eman/data") 
|   |   ...
|
└───gem_cnn                 # Implementation of Gauge Equivariant CNNs
│   └───nn
│   └───tests
│   └───transform
│   └───utils
│   
└───spiralnet               # Implementation of SpiralNet++
|   |   spiralconv.py
│   └───spiralnet.utils

Citation

Please use the following snippet to cite this work:

@article{basu2022equivariant,
      title={{Equivariant Mesh Attention Networks}}, 
      author={Basu, Sourya and Gallego-Posada, Jose and Vigan\`o, Francesco and Rowbottom, James and Cohen, Taco},
      year={2022},
      month={08},
      journal={Transactions on Machine Learning Research}
}

Acknowledgements

• The code for Gauge Equivariant Mesh CNNs is taken from the official GEM-CNN implementation.
• The code for SpiralNet++ comparison is taken from the official SpiralNet++ implementation.