seal-lossnet-v0

SEAL Loss-Net v0

This repository contains a custom PyTorch checkpoint for the Shared Energy SEAL-style loss-net:

E_omega(T, d, c)

The model maps a power-converter topology graph T, continuous duty cycle d, and target specification c to a scalar energy. Lower energy is intended to correspond to better topology-duty-spec combinations.

This is not a Transformers checkpoint. Load it with this project's shared_energy.seal_models.GTNEnergyLossNet class and the checkpoint's model_state_dict.

Files

• seal_lossnet_model.pt: PyTorch checkpoint with model_state_dict, model_config, training config, and best validation metadata.
• manifest.json: Training and evaluation summary.
• training_history.jsonl: Per-epoch training and validation metrics, included when uploaded with --include-eval-artifacts.

Training Objective

The loss-net was trained with supervised energy regression plus intra- and inter-topology ranking losses:

L_E = L_reg + lambda_rank * (rho * L_intra + (1 - rho) * L_inter)

For this run, lambda_rank=0.2 and rho=0.5.

Metrics

• Best epoch: 118
• Epochs: 200

Sample counts:

{
  "dense_test": 11988,
  "dense_train": 55809,
  "dense_val": 11988,
  "inter_test": 5000,
  "inter_train": 30000,
  "inter_val": 5000,
  "intra_test": 4736,
  "intra_train": 22048,
  "intra_val": 4736
}

Final validation metrics:

{
  "dense_count": 11988.0,
  "inter_accuracy": 0.7376,
  "inter_count": 5000.0,
  "inter_loss": 0.011845157063007355,
  "intra_accuracy": 0.7048141891891891,
  "intra_count": 4736.0,
  "intra_loss": 0.00825060364110647,
  "reg_huber": 0.00027953632127988956,
  "reg_mae": 0.012845741039838718,
  "total_proxy": 0.0022891123916912726
}

Final test metrics:

{
  "dense_count": 11988.0,
  "inter_accuracy": 0.7554,
  "inter_count": 5000.0,
  "inter_loss": 0.014154602885246276,
  "intra_accuracy": 0.7185388513513513,
  "intra_count": 4736.0,
  "intra_loss": 0.009359711932169425,
  "reg_huber": 0.00067927523821364,
  "reg_mae": 0.015292200926426533,
  "total_proxy": 0.0030307067199552102
}

Usage

from huggingface_hub import hf_hub_download
import torch

from shared_energy.seal_models import GTNEnergyLossNet

repo_id = "tjwjdgns011119/seal-lossnet-v0"
checkpoint_path = hf_hub_download(repo_id=repo_id, filename="seal_lossnet_model.pt")

payload = torch.load(checkpoint_path, map_location="cpu")
model = GTNEnergyLossNet(**payload["model_config"])
model.load_state_dict(payload["model_state_dict"])
model.eval()

# energy = model(
#     node_features=node_features,
#     neighbor_mask=neighbor_mask,
#     loop_membership=loop_membership,
#     node_mask=node_mask,
#     duty=duty,
#     spec=spec,
# )

The project code must be available in the Python environment, for example with pip install -e . from the repository root.

Notes

The checkpoint is intended for research workflows in this repository. The energy output is a learned proxy, not a direct SPICE simulation result.