gridfm/genco-pfdelta

GENCO Base on PFΔ (paper §5.1)

Twenty-four GENCO Base checkpoints used for Figure 4, Figure 18, and Table 4 of GENCO: IEEE 118, eight PFΔ tasks, three seeds (1, 2, 3; YAML seeds 0 / 42 / 1234).

PFNet, CANOS-PF, GNS-S, and Newton–Raphson numbers in the paper are taken from PFΔ; they are not in this repo.

Files

task_{1.1,1.2,1.3,2.3,4.1,4.2,4.3}/seed{1,2,3}/
  last.pt
  HGNS_PF_pfdelta_bs64_seed{1,2,3}.yaml
task_3.1/seed{1,2,3}/
  best_model_state_dict.epoch_299.pt
  HGNS_PF_pfdelta_bs64_seed{1,2,3}.yaml
mlflow/
  {1.1,1.2,1.3,2.3,3.1,4.1,4.2,4.3}/<run_id>/

The training YAML sits in the same folder as the checkpoint. The same three files are on GitHub: scripts/pfdelta/config. One YAML is shared by every task; seed1 / seed2 / seed3 are YAML seeds 0 / 42 / 1234.

last.pt is the last-epoch state_dict. These match the paper CSVs. Task 3.1 was not trained separately: its paper numbers come from evaluating the task 1.3 best-validation snapshot (best_model_state_dict.epoch_299.pt) on task 3.1 data.

The training MLflow runs are under mlflow/ (params, tags, metrics, test CSVs and plots). The paper CSVs are the in-training test dump of those runs.

Install

The paper results were obtained with genco-paper-repro-pfdelta. We only guarantee the same numbers on this branch, because main is under active development. Clone it and install from source:

git clone --branch genco-paper-repro-pfdelta https://github.com/gridfm/gridfm-graphkit.git
cd gridfm-graphkit
python -m venv venv && source venv/bin/activate
pip install -e .
TORCH_CUDA_VERSION=$(python -c "import torch; print(torch.__version__ + ('+cpu' if torch.version.cuda is None else ''))")
pip install torch-scatter -f https://data.pyg.org/whl/torch-${TORCH_CUDA_VERSION}.html

Data

Converted PFΔ parquet (already public). Graphkit loads {data_path} as the case118 folder of a task dataset.

Paper task	Dataset
1.1	gridfm/pfdelta_task1.1
1.2	gridfm/pfdelta_task1.2
1.3 / 2.1	gridfm/pfdelta_task1.3
2.3	gridfm/pfdelta_task2.3
3.1	gridfm/pfdelta_task3.1
4.1	gridfm/pfdelta_task4.1
4.2	gridfm/pfdelta_task4.2
4.3	gridfm/pfdelta_task4.3

Evaluate

From the graphkit repo root. Same YAML for every task; only --data_path and --model_path change.

hf download gridfm/pfdelta_task4.3 --repo-type dataset --local-dir pfdelta_task4.3
hf download gridfm/genco-pfdelta --include "task_4.3/seed1/**" --local-dir genco-pfdelta

gridfm_graphkit evaluate \
  --config scripts/pfdelta/config/HGNS_PF_pfdelta_bs64_seed1.yaml \
  --data_path pfdelta_task4.3/case118 \
  --model_path genco-pfdelta/task_4.3/seed1/last.pt

For task 3.1, download task_3.1/seed1/** and gridfm/pfdelta_task3.1, and pass best_model_state_dict.epoch_299.pt with --data_path pfdelta_task3.1/case118.

Full commands: GENCO §5.1.