"""Convenience loaders for the checkpoint collection. No data download or training.""" from pathlib import Path import sys def load_model(name, fold=0, split_seed=None, root=None): root = Path(root) if root is not None else Path(__file__).resolve().parent sys.path.insert(0, str(root / 'code')) directory = root / 'checkpoints' partition = 'final' if fold == 'final' else f'fold_{int(fold)}' if split_seed is not None: if partition == 'final': raise ValueError('Split-seed experiments contain CV fold models only') from ampd.split_seed_experiment import SplitSeedPredictor return SplitSeedPredictor(directory / 'three_split_seeds_2026-11-17' / f'seed_{int(split_seed)}' / partition / 'bundle') if name in {'baseline_mlp', 'selected_ae', 'pls'}: if name == 'pls' and partition == 'final': raise ValueError('Original PLS has five CV bundles, no all-patient final bundle') from ampd.retained_models import load_retained return load_retained(name, directory / 'retained' / name / partition / 'bundle') if name in {'pca', 'ablations'}: if partition == 'final': raise ValueError('PCA/ablation packages contain CV fold models only') from ampd.improvement_ablation import AblationPredictor return AblationPredictor(directory / 'improvement_ablation_2026-11-16_v2' / partition / 'bundle') if name == 'top_reproductions': from ampd.top_history_mask import TopCurrentPredictor return TopCurrentPredictor(directory / 'top_nohistory_2026-11-10' / partition / 'inference_bundle') raise ValueError(f'Unknown collection: {name}')