"""Data inspection without importing torch.""" from __future__ import annotations from pathlib import Path from typing import Any, Dict from .data.clinical import build_all_data from .data.splits import fold_summary, make_folds from .utils import ensure_dir, save_json def inspect_data(cfg: Dict[str, Any]) -> Dict[str, Any]: bundle = build_all_data(cfg) folds = make_folds(bundle.patients, cfg["cv"]) out = { "n_patients": int(len(bundle.patients)), "n_train_patients": int((bundle.patients["is_supplemental"] == 0).sum()), "n_samples_all": int(len(bundle.samples_all)), "n_samples_eval": int(len(bundle.samples_eval)), "n_features": int(len(bundle.feature_names)), "feature_names": list(bundle.feature_names), "target_cols": list(bundle.target_cols), "masks": {c: int(bundle.samples_eval[c].sum()) for c in bundle.mask_cols}, "cohorts": bundle.patients["cohort"].value_counts().to_dict(), "fold_summary": fold_summary(folds).to_dict(orient="records"), } out_dir = ensure_dir(Path(cfg["experiment"].get("output_dir", "runs")) / cfg["experiment"].get("name", "ampd_exp") / "inspect") save_json(out, out_dir / "data_summary.json") print(f"patients: {out['n_patients']} (train {out['n_train_patients']})") print(f"samples: all {out['n_samples_all']} | eval {out['n_samples_eval']}") print(f"features: {out['n_features']}") print(f"cohorts: {out['cohorts']}") print(f"masks: {out['masks']}") print(f"fold summary:\n{fold_summary(folds).to_string(index=False)}") return out