Update src/utils/reproducibility.py

src/utils/reproducibility.py

@@ -1,4 +1,3 @@
-# === 00) Reproducibility: global seeds & deterministic ops ===
 def set_seed(seed: int = 1337):
     import os, random, numpy as np
     os.environ["PYTHONHASHSEED"] = str(seed)
@@ -17,17 +16,12 @@ def set_seed(seed: int = 1337):
     print(f"[seed] set to {seed}")
 set_seed(1337)
 
-# === 01) Grouped CV (by ligand or transporter) ===
 from sklearn.model_selection import StratifiedGroupKFold
 
 def make_grouped_cv(n_splits=5, seed=1337):
     return StratifiedGroupKFold(n_splits=n_splits, shuffle=True, random_state=seed)
 
-# Example placeholders (replace with your arrays/Series):
-# groups = ligand_ids  # or transporter_ids
-# cv = make_grouped_cv()
 
-# === 02) Leakage-safe pipeline (scaler inside Pipeline) + ΔAUPRC evaluation ===
 import numpy as np
 from sklearn.pipeline import Pipeline
 from sklearn.preprocessing import StandardScaler
@@ -49,8 +43,6 @@ def eval_auprc_grouped(X, y, groups, cv=None, estimator=None):
     print(f"[AUPRC] {auprc:.4f} | baseline={base:.4f} | Δ={auprc-base:.4f}")
     return prob, auprc, base
 
-# === 03) CV-safe probability calibration helper ===
-# NOTE: sklearn's CalibratedClassifierCV ignores 'groups' directly; do manual fold loop for true group-aware calibration.
 from sklearn.calibration import calibration_curve
 from sklearn.base import clone
 
@@ -67,7 +59,6 @@ def cv_calibrated_probs(estimator, X, y, groups, cv=None, method="isotonic"):
         proba[test_idx] = cal.predict_proba(X[test_idx])[:,1]
     return proba
 
-# === 04) Active Learning logging utilities ===
 import pandas as pd, os, numpy as np
 from pathlib import Path
 
@@ -84,30 +75,25 @@ def al_log_round(round_id, acquired_idx, scores, y_true, pool_size):
     df.to_csv(out, index=False)
     print(f"[AL] wrote {out} (pool={pool_size}, acquired={len(acquired_idx)})")
 
-# === 05) Domain Generalization baseline placeholders (GroupDRO / CORAL) ===
-# Minimal CORAL regularizer for linear/MLP torch models; for sklearn, emulate via feature alignment preproc.
 def coral_penalty(source, target):
     import torch
     cs = torch.cov(source.T) if source.ndim==2 else torch.cov(source)
     ct = torch.cov(target.T) if target.ndim==2 else torch.cov(target)
     return torch.norm(cs - ct, p='fro')
 
-# GroupDRO would require per-group weighting; keep an interface for later plug-in.
 class GroupDROWrapper:
     def __init__(self, base_estimator):
         self.base = base_estimator
     def fit(self, X, y, groups):
-        # TODO: implement group-wise reweighting loop
         return self.base.fit(X, y)
     def predict_proba(self, X):
         return self.base.predict_proba(X)
 
-# === 07) Runtime leak check helper ===
 def check_scaler_leakage(estimator):
     from sklearn.pipeline import Pipeline
     ok = isinstance(estimator, Pipeline) and any(hasattr(s, "transform") and s.__class__.__name__.endswith("Scaler")
                                                  for _, s in estimator.steps[:-1])
     if not ok:
-        print("⚠️  Potential leakage: estimator is not a Pipeline with scaler → classifier.")
+        print("  Potential leakage: estimator is not a Pipeline with scaler → classifier.")
     else:
-        print("✅ Leakage-safe pipeline detected.")
+        print(" Leakage-safe pipeline detected.")