Mandala-GNN
A 405k-parameter geometric GNN for masked-patch reconstruction on a hand-drawn hexagonal lattice. Its topology β nodes, radial/ring/spoke edges, the pyramid depth β was read off three paint-marker drawings rather than picked from a benchmark.
This repo holds the single checkpoint behind the live demo, plus the cold-state cache and metrics that let a viewer paint the page before torch is loaded.
| params | 405,342 |
| nodes / edges | 203 / 2,092 |
| checkpoint | 1.7 MB (model.pt) |
| training | 400 steps, 88 s, single CPU thread |
| task | predict the 4Γ4 pooled patch of every hidden node |
Headline result
One 45Β° sector of the drawing is held out and never trained on. A position-memorising model cannot fill it in; a model that carries the sketch's angular structure across the graph can.
| wedge r | wedge MSE | |
|---|---|---|
| mean-patch floor | β | 0.0154 |
| plain MLP (no edges, no pyramid) | 0.00 | 0.0286 |
| pyramid pooling, no edges | 0.73 | 0.0177 |
| honeycomb mesh, full model | 0.90 | 0.0030 |
Ablations on the full model: removing the angular harmonics drops wedge r to 0.24; removing the learned edge gate drops it to 0.08. Both the symmetry prior and the multi-scale pyramid earn their keep.
The negative result, which is part of the release: rewiring the edges at random on the same 203 node positions scores within 2 points of the honeycomb. The specific hand-drawn lattice is not measurably doing the work at this task scale β the geometry-aware attention is.
Files
| file | what it is |
|---|---|
model.pt |
{config, state_dict, metrics}, 225 tensors, CPU |
metrics.json |
config + final metrics + training curve |
demo.json |
drawing, mesh, wedge and eval indices β the pre-torch state |
recon.png |
drawing | visible | prediction | composite, 3Γ |
Use it
The checkpoint is only meaningful with the code that built it (same layer names, same mesh). Clone the repo and point the demo server at it:
git clone https://github.com/BoogerCheeseOnRye/mandala-gnn
cd mandala-gnn
python -m venv .venv && .venv/bin/pip install torch numpy pillow
.venv/bin/python -m mandalanet.serve --model runs/main-hex/model.pt --port 8000
metrics.json and demo.json must sit next to model.pt. From Python:
import urllib.request, json
from huggingface_hub import hf_hub_download
d = hf_hub_download("BoogerCheeseOnRye/mandala-gnn", "model.pt")
for f in ("metrics.json", "demo.json"):
hf_hub_download("BoogerCheeseOnRye/mandala-gnn", f, local_dir="runs/main-hex")
from mandalanet.serve import Runner # from the GitHub repo
r = Runner(d, preload=True)
out = r.redraw([]) # one real forward pass
print(out["stats"], out["heads"]) # latency, MSE, per-head attention share
To retrain from the drawing instead (88 s on a phone CPU):
.venv/bin/python -m mandalanet.train --drawing drawings/hex-tessellation.jpg \
--mesh hex --steps 400 --out runs/main-hex
Architecture
local patch per node βββΊ RadialAttention βββΊ ChaosPyramid βββΊ patch head
content + RBF(edge length) (4Γ4 pooled
+ cos/sin n-fold harmonics pixel values)
+ endpoint radii + edge type
Γ learned edge gate
RadialAttentionβ attention logits get a learnable bias from an RBF over edge length, cos/sin harmonics of the edge angle, both endpoint radii and the edge type; messages are scaled by a learned per-head edge gate. 3 depth levels, 2 blocks, 4 heads, 64 dim, 4 harmonics, 8 RBF bins.ChaosPyramidβ spatial cell pooling down Γ3, gated residual skips back up, softmax mixture of per-level mean+max read-outs.
Live demo
The cockpit runs from the repo β python -m mandalanet.serve and open
localhost:8000. It is not hosted: Hugging Face only hosts Docker Spaces for
PRO accounts, so the interactive page runs locally (it needs ~1 s to paint and
~0.25 s per forward pass) and the project page carries a recording.
- Project page and paper: https://boogercheeseonrye.github.io/mandala-gnn/
- Code and full ablations: https://github.com/BoogerCheeseOnRye/mandala-gnn
Scope
A research prototype, not a foundation model. It reconstructs patches of one 128Γ128 drawing of a hex tessellation, and the class head is a small side task. It has no text, no generalisation to arbitrary images, and no interface beyond the demo server. No licence is asserted over the source drawings.