Skip to content

Track: Track1; Team name: yeahkitory; Model: Bundle Neural Networks (BuNN)#319

Draft
yeahkitoryisgood wants to merge 22 commits into
geometric-intelligence:mainfrom
yeahkitoryisgood:challenge-2026-bunn
Draft

Track: Track1; Team name: yeahkitory; Model: Bundle Neural Networks (BuNN)#319
yeahkitoryisgood wants to merge 22 commits into
geometric-intelligence:mainfrom
yeahkitoryisgood:challenge-2026-bunn

Conversation

@yeahkitoryisgood
Copy link
Copy Markdown

@yeahkitoryisgood yeahkitoryisgood commented May 16, 2026

Summary

  • Adds a graph BuNN backbone for the challenge track with learned O(2) bundle maps, random-walk heat-kernel Taylor diffusion, residual/norm/dropout channel updates, and a real Hydra config.
  • Review-driven hardening covers tensor shapes, tensor input types, floating node-feature tensors, node-feature widths, node ids, edge-index device placement, bundle shape constraints, scalar/integer/probability hyperparameters, boolean flags, activation config values, finite real-valued non-negative edge weights, non-contiguous edge-weight tensors, boolean edge-weight rejection, unexpected constructor/forward keyword arguments, TopoBench wrapper output-channel introspection, and weighted random-walk Laplacian coverage.
  • The full challenge result artifact remains tracked at 2026_tdl_challenge/outputs/bunn_full/results.json.

Validation

  • .venv\Scripts\python.exe -m ruff check topobench\nn\backbones\graph\bunn.py test\nn\backbones\graph\test_bunn.py -> passed
  • .venv\Scripts\python.exe -m pytest test\nn\backbones\graph\test_bunn.py -q -> 59 passed
  • .venv\Scripts\python.exe -m pytest test\nn\backbones\graph\test_bunn.py test\pipeline\test_pipeline.py -q -> 60 passed
  • Full challenge grid: 72 runs, community accuracy mean 0.1034, triangle normalized MSE 6.7543

The full grid result above was generated before the invalid-input validation commits; the valid-input computation path is unchanged by these hardening patches.

Submission Info

  • Track: Track1
  • Team name: yeahkitory
  • Model: Bundle Neural Networks (BuNN)

Prasanna28Devadiga pushed a commit to Prasanna28Devadiga/TopoBench that referenced this pull request May 17, 2026
Add the artefacts produced by running 2026_tdl_challenge/run_evaluation.ipynb
with MODEL_CONFIG="graph/conn_nsd":

  outputs/conn_nsd_full/
    results.json                                — 72 rows, 3 seeds × 12 cells × 2 tasks
    heatmap_community_detection_accuracy.png    — accuracy mean ± std heatmap
    heatmap_triangle_mse_over_triangles.png     — MSE / Σtriangles heatmap
    OOD/                                        — 6 OOD-delta panels
      OOD_{low,mid,high}_homophily__{community_detection,triangle_counting}.png

Convention mirrors the BuNN submission (PR geometric-intelligence#319), which uses
2026_tdl_challenge/outputs/bunn_full/. Files are force-added past the
.gitignore rule for the broader outputs/ directory.

Headline numbers (10-epoch budget; the spec explicitly does NOT reward
final performance):
  - Community detection: mean test accuracy 0.075 across all 36 runs;
    chance is 1/12 for the 12-community task. Per-seed range [0.04, 0.12].
  - Triangle counting: mean test MSE / Σtriangles = 10.46 across all
    36 runs.

These confirm the pipeline runs end-to-end on every grid cell with no
NaNs in metric fields that should be finite.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants