E(n)-equivariant Graph Cellular Automata

This repository is a fork of https://github.com/gengala/egnca attempting to improve model expressiveness and remove fixed edges during training

N.B. Sometimes GIFs are not properly loaded in README.md. Please refresh the page.

Angles

python -m trainers.geometric_graph -ds Grid2d -sdg 0.05 -rdg 1.0 -bsc 0 8 2000 16 4000 32 -pats 800 -an --seed 42
python -m trainers.geometric_graph -ds Grid2d -sdg 0.05 -rdg 1.0 -bsc 0 8 2000 16 4000 32 -pats 800     --seed 42

Local loss

Non-persistent	Persistent

python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l local -an --seed 42
python -m trainers.geometric_graph -ds Grid2d -bsc 0 8 2000 16 4000 32 -pats 800 -l local -an --seed 42

Anchors

python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l local -ss -as corners -an --seed 42
python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l local -ss -as corners -asc 1.5 -an --seed 42
python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l local -ss -as simplex -an --seed 42

Penalty

python -m trainers.geometric_graph -ds Grid2d -bsc 0 4 -l local_enp -an --seed 42
python -m trainers.geometric_graph -ds Grid2d -bsc 0 8 2000 16 4000 32 -l local_enp -pd 0.3 -ss -as corners -asc 1.5 -an --seed 42

Different targets

IF the gif does not load, check it out here: https://drive.google.com/uc?export=view&id=1plY13LX7eWIIyKf01JtWft24_8qHUZPa

python -m trainers.geometric_graph -ds Cube -sdg 0.05 -rdg 1.0 -bsc 0 2 -l local_enp -pd 0.2 -ss -as corners -ad 0.15 -asc 0.5 -an --seed 42
python -m trainers.geometric_graph -ds Torus -bsc 0 8 2000 16 -l local_enp -an -at undirected -ss -as corners -ad 0.5
python -m trainers.geometric_graph -ds x_small -bsc 0 8 2000 16 4000 32 -l local_enp -pd 0.2 -ss -as corners --seed 42

OT-based loss

Anchors

python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l ot -ss -as simplex --seed 42

Penalty

python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l ot_p -ss -as simplex --seed 42 -nt nn

Static and dynamic relative edges

Static

python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l ot_p -ss -as simplex -re -en 4 --seed 42 -nt nn

Dynamic, every step

python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l ot_p -ss -as simplex -re -de -en 4 --seed 42

Dynamic, every 5 steps

python -m trainers.geometric_graph -ds Grid2d -bsc 0 8 2000 16 -l ot_p -pd 0.3 -an -ss -as corners -re -de -en 4 -des 5 -ad 0.3 --seed 42
python -m trainers.geometric_graph -ds Grid2d -bsc 0 8 2000 16 -l ot_p -pd 0.3     -ss -as simplex -re -de -en 4 -des 5 -ad 0.3 --seed 42

Dynamic, no penalty

python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l ot -at undirected -an -ffa -ss -as corners -asc 1.5 -uaf -re -de -des 5 -en 5 --seed 123

Beacons

No persistency

python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l ot_p -apd -an -bc -as simplex -re -de -en 4 -des 5 --seed 42'
python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l ot_p -pd 0.3 -bc -as simplex -re -de -en 4 -des 5 --seed 42 -nt nn
python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l ot_p -pd 0.4 -bc -as simplex -re -de -en 4 -des 5 --seed 42 -nt nn
python -m trainers.geometric_graph -ds Grid2d -bsc 0 1 -l ot_p -pd 0.5 -bc -as simplex -re -de -en 4 -des 5 --seed 42 -nt nn

Persistency

IF the gif does not load, check it out here: https://drive.google.com/uc?export=view&id=1v0uKqbdVi8sYXpQh4ufRqNDLKZV7iQM7

python -m trainers.geometric_graph -ds Grid2d -bsc 0 8 2000 16 4000 32 -l ot_p -apd -bc -as simplex -re -de -en 4 -des 5 --seed 42 -nt nn
python -m trainers.geometric_graph -ds Grid2d -bsc 0 8 2000 16 4000 32 -l ot_p -pd 0.3 -bc -as simplex -re -de -en 4 -des 5 --seed 42 -nt nn
python -m trainers.geometric_graph -ds Grid2d -bsc 0 8 2000 16 4000 32 -l ot_p -pd 0.4 -bc -as simplex -re -de -en 4 -des 5 --seed 42 -nt nn
python -m trainers.geometric_graph -ds Grid2d -bsc 0 8 2000 16 4000 32 -l ot_p -pd 0.5 -bc -as simplex -re -de -en 4 -des 5 --seed 42 -nt nn

Robustness

All robustness experiments are performed with the same model:

python -m trainers.geometric_graph -ds Grid2d -bsc 0 8 2000 16 4000 32 -l ot_p -pd 0.3 -bc -as simplex -re -de -en 4 -des 5 --seed 42 -nt nn

Random initial coordinates

Damage

Local damage	Global damage

Increased or decreased node count

IF the gif does not load, check it out here: https://drive.google.com/uc?export=view&id=140wQ-GTYflekb7xVrBeOxdZHsupSkGIJ

Decreased fire rate

IF the gif does not load, check it out here: https://drive.google.com/uc?export=view&id=1qUGOCaR8KIUvhIQ5yuampMZqxfGTuVsr

Rotated structured seeds

E(n) Convergence to 2D-Grid	Adapting to rotated structured seeds

Different targets

X and a

Cube and Torus

For testing, play with notebooks/test_geometric_graph.ipynb and notebooks/visualise_results.ipynb.