"""Evaluate the fusion layer on saved model predictions. Each input is a .npy file holding either positive-class probabilities, shape (N,), or two-class probability vectors, shape (N, 2). Example ------- python scripts/evaluate.py \ --hist data/p_hist_test.npy --risk data/p_risk_test.npy --labels data/y_test.npy \ --calib-hist data/p_hist_val.npy --calib-risk data/p_risk_val.npy --calib-labels data/y_val.npy \ --out results/ """ from __future__ import annotations import argparse import os import sys import numpy as np sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "src")) from otfusion import FusionConfig, format_summary, run_evaluation # noqa: E402 def _alpha(value: str): return value if value == "dynamic" else float(value) def main() -> None: parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter) parser.add_argument("--hist", required=True, help="Histopathology model probabilities (.npy)") parser.add_argument("--risk", required=True, help="Clinical risk model probabilities (.npy)") parser.add_argument("--labels", required=True, help="Reference labels, 0/1 (.npy)") parser.add_argument("--calib-hist", help="Calibration split: histopathology probabilities") parser.add_argument("--calib-risk", help="Calibration split: risk probabilities") parser.add_argument("--calib-labels", help="Calibration split: labels") parser.add_argument("--out", default="results", help="Output directory") defaults = FusionConfig() parser.add_argument("--fp-weight", type=float, default=defaults.fp_weight) parser.add_argument("--fn-weight", type=float, default=defaults.fn_weight) parser.add_argument("--alpha", type=_alpha, default=defaults.alpha, help='float in [0, 1] or "dynamic"') parser.add_argument("--reg", type=float, default=defaults.reg) parser.add_argument("--max-iter", type=int, default=defaults.max_iter) parser.add_argument("--dropout-rates", type=float, nargs="+", default=[0.0, 0.2, 0.5]) parser.add_argument("--dropout-mode", default="either", choices=["either", "hist", "risk", "both"]) parser.add_argument("--threshold", type=float, default=0.5) parser.add_argument("--n-boot", type=int, default=1000) parser.add_argument("--seed", type=int, default=42) parser.add_argument("--no-figures", action="store_true") args = parser.parse_args() calib_args = (args.calib_hist, args.calib_risk, args.calib_labels) if any(calib_args) and not all(calib_args): parser.error("--calib-hist, --calib-risk and --calib-labels must be given together") calibration = tuple(np.load(path) for path in calib_args) if all(calib_args) else None config = FusionConfig( fp_weight=args.fp_weight, fn_weight=args.fn_weight, alpha=args.alpha, reg=args.reg, max_iter=args.max_iter, ) results = run_evaluation( np.load(args.hist), np.load(args.risk), np.load(args.labels), config=config, calibration=calibration, dropout_rates=args.dropout_rates, dropout_mode=args.dropout_mode, threshold=args.threshold, n_boot=args.n_boot, seed=args.seed, out_dir=args.out, make_figures=not args.no_figures, ) print(format_summary(results)) print(f"\nSaved results to {os.path.abspath(args.out)}") if __name__ == "__main__": main()