vietngth/ph-ensemble-gnn

PH-TSER-Att checkpoints

Trained PH-TSER-Att$_0$ models (graph ensemble over the 38 PH graphs of $\mathcal{G}^{(0)}$) from "Persistent Homology-Induced Graph Ensembles" (code, arXiv:2503.14240), for seed 1 of each network and both training settings of the paper:

folder	network	setting
ci_tuned_seed1	Central Italy	tuned (AdamW, one-cycle, learning rate from a range test)
ci_untuned_seed1	Central Italy	untuned (published TSER-GCN settings)
cw_tuned_seed1	Central-West Italy	tuned
cw_untuned_seed1	Central-West Italy	untuned

Each folder holds best.ckpt (the fold model with the lowest validation loss; weights, hyperparameters and graphs) and results.json (the run's config, the fold and its recorded test metrics). Seed 1 fixes the 80/20 split.

Inference

git clone https://github.com/vietngth/ph-ensemble-gnn && cd ph-ensemble-gnn   # set up as in its README
huggingface-cli download vietngth/ph-ensemble-gnn-data ph-gnn-data.zip --repo-type dataset --local-dir .
python -m zipfile -e ph-gnn-data.zip .
huggingface-cli download vietngth/ph-ensemble-gnn --local-dir checkpoints
python experiments/predict.py --checkpoint checkpoints/ci_tuned_seed1/best.ckpt \
    --results checkpoints/ci_tuned_seed1/results.json --data_root data

predict.py scores the checkpoint on the test set of its seed and checks the MAE against the recorded value. The paper's tables average the five fold models over ten seeds; a single fold model of one seed is not expected to match the table values exactly.