"""Portable, metadata-validated checkpoints without pickled model objects.""" from __future__ import annotations import math from pathlib import Path from typing import Any import torch from .model import PlantDiseaseCNN SCHEMA_VERSION = 1 def _validate_metadata(checkpoint: Any) -> None: if ( not isinstance(checkpoint, dict) or type(checkpoint.get("schema_version")) is not int or checkpoint["schema_version"] != SCHEMA_VERSION ): raise ValueError("Unsupported checkpoint schema; expected schema_version=1") model_config = checkpoint.get("model_config") if not isinstance(model_config, dict) or set(model_config) != { "num_classes", "channels", "dropout" }: raise ValueError("Checkpoint has an invalid model_config") names = checkpoint.get("class_names") if ( not isinstance(names, list) or not all(isinstance(name, str) and name for name in names) or len(set(names)) != len(names) or len(names) != model_config["num_classes"] ): raise ValueError("Checkpoint class_names must uniquely match the model outputs") preprocessing = checkpoint.get("preprocessing") if not isinstance(preprocessing, dict): raise ValueError("Checkpoint is missing preprocessing metadata") size = preprocessing.get("image_size") if isinstance(size, bool) or not isinstance(size, int) or size < 16: raise ValueError("Checkpoint image_size must be an integer of at least 16") for name in ("mean", "std"): values = preprocessing.get(name) if ( not isinstance(values, (list, tuple)) or len(values) != 3 or any( isinstance(value, bool) or not isinstance(value, (int, float)) or not math.isfinite(value) or (name == "std" and value <= 0) for value in values ) ): raise ValueError(f"Checkpoint {name} must contain three valid RGB values") state = checkpoint.get("model_state_dict") if not isinstance(state, dict) or not state or any( not isinstance(key, str) or not isinstance(value, torch.Tensor) for key, value in state.items() ): raise ValueError("Checkpoint model_state_dict is missing or invalid") def load_checkpoint( path: str | Path, device: str | torch.device = "cpu" ) -> tuple[PlantDiseaseCNN, dict[str, Any]]: """Load on CPU, validate metadata, then move the reconstructed CNN.""" checkpoint = torch.load(path, map_location="cpu", weights_only=True) _validate_metadata(checkpoint) try: model = PlantDiseaseCNN(**checkpoint["model_config"]) model.load_state_dict(checkpoint["model_state_dict"], strict=True) except (TypeError, ValueError, RuntimeError) as exc: raise ValueError(f"Checkpoint model configuration/weights are incompatible: {exc}") from exc model.to(device) model.eval() return model, checkpoint