| """Validated backbone handoff between the two Sudoku training stages.""" |
|
|
| from pathlib import Path |
|
|
| import torch |
|
|
| from softasp.sudoku_reporting import file_sha256 |
|
|
|
|
| def extract_backbone(checkpoint: Path, destination: Path) -> dict: |
| """Extract a newly trained perception checkpoint, retaining source identity.""" |
| saved = torch.load(checkpoint, map_location="cpu", weights_only=False) |
| if saved.get("config", {}).get("model", {}).get("name") != "perception_only_sudoku": |
| raise ValueError("Expected a perception_only_sudoku training checkpoint") |
| if saved.get("epoch", -1) < 0: |
| raise ValueError("Backbone checkpoint must follow at least one training epoch") |
| state = { |
| key: value |
| for key, value in saved["model_state_dict"].items() |
| if key.startswith("cell_cnn.") |
| } |
| if not state or any(not torch.isfinite(value).all() for value in state.values()): |
| raise ValueError("Expected finite trained cell_cnn tensors") |
| provenance = { |
| "source_checkpoint": str(checkpoint.resolve()), |
| "source_sha256": file_sha256(checkpoint), |
| "source_epoch": saved["epoch"], |
| "source_metric": saved.get("best_metric"), |
| "dataset_identity": saved["config"].get("resolved_dataset"), |
| } |
| |
| with destination.open("xb") as stream: |
| torch.save({"model_state_dict": state, **provenance}, stream) |
| return {**provenance, "path": str(destination), "sha256": file_sha256(destination)} |
|
|
