seal-tasknet-v0

SEAL Task-Net v0

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

g_psi(T, c) -> d0

The model maps a power-converter topology graph T and target specification c to an initial continuous duty-cycle prediction d0.

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

Files

• seal_tasknet_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 task-net was trained to predict the GTN dense-search best duty while also being guided by a frozen SEAL loss-net energy term:

L_task = lambda_sup * L_sup(d0, d_best) + lambda_energy * E_omega(T, d0, c)

For this run, lambda_sup=1.0 and lambda_energy=0.2.

Metrics

• Best epoch: 9
• Epochs: 200

Sample counts:

{
  "test": 148,
  "train": 689,
  "val": 148
}

Final validation metrics:

{
  "count": 148.0,
  "duty_mae": 0.212996697103655,
  "energy_loss": 0.995880257438969,
  "sup_loss": 0.032279046554420446,
  "total_loss": 0.23145510296563845
}

Final test metrics:

{
  "count": 148.0,
  "duty_mae": 0.23800100829150225,
  "energy_loss": 0.9947394277598407,
  "sup_loss": 0.039557718344636866,
  "total_loss": 0.2385056047020732
}

Usage

from huggingface_hub import hf_hub_download
import torch

from shared_energy.seal_models import GTNTaskNet

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

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

# duty = model(
#     node_features=node_features,
#     neighbor_mask=neighbor_mask,
#     loop_membership=loop_membership,
#     node_mask=node_mask,
#     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 predicted duty is an initialization or proposal, not a direct SPICE-verified optimum.