gridfm/genco-pf-transfer

GENCO transfer to IEEE 118 (paper §5.5.3)

Checkpoints for the transfer figure of GENCO: one subgrid-pretrained model, and from-scratch and fine-tuned IEEE 118 models at seven training sizes.

Guide: GENCO §5.5.3.

Files

pretraining/
  best_model_state_dict.pt
  HGNS_PreTrain_subnets_100.yaml
  train.sh
scratch/case118_<N>/                 # N in 100, 1000, 10000, 20000, 50000, 100000, 250000
  best_model_state_dict.pt
  normalizer_stats.pt
  case118_<N>.yaml
  train.sh
finetune/case118_<N>/
  best_model_state_dict.pt
  normalizer_stats.pt
  case118_<N>.yaml
  finetune.sh
data/case118_ieee/raw/               # 9,952 eval scenarios
mlflow/pretrain/1_h_100_32/
mlflow/train/
mlflow/eval/

The pretrained checkpoint has no normalizer_stats.pt. Every scratch and fine-tune folder does. Fine-tuning training data is gridfm/pf_small_case118_ieee. Pretraining grids are gridfm/genco-pf-pretraining.

hf download gridfm/genco-pf-transfer --include "finetune/case118_1000/**" --include "data/**" --local-dir genco-pf-transfer

gridfm_graphkit evaluate
  --config scripts/pretraining_advantage/case118_eval.yaml
  --data_path genco-pf-transfer/data
  --model_path genco-pf-transfer/finetune/case118_1000/best_model_state_dict.pt
  --normalizer_stats genco-pf-transfer/finetune/case118_1000/normalizer_stats.pt
  --compute_dc_ac_metrics