experiment: validate beta selection on one paired full condition

scripts/run_baseline.py

@@ -252,6 +252,66 @@ def main(argv=None) -> int:
             "rare_class_recall": rare["rare_class_recall"],
         })
 
+    # ---- global validation evaluation (SELECTION ONLY) -------------------
+    # This block never influences training and is never used to tune any
+    # hyperparameter except the prototype-aggregation beta, which is selected
+    # on validation metrics only. The test split is untouched here.
+    val_embeddings = extract_embeddings(model, val_features)
+    val_pred = nearest_prototype_predict(val_embeddings, global_prototypes)
+    val_metrics = classification_metrics(val_labels, val_pred, bundle.num_classes)
+    val_rare = rare_class_metrics(val_labels, val_pred, bundle.num_classes, union_rare)
+
+    val_eval_assignments, _ = partition_mod.partition_eval_split(
+        val_labels, args.clients, args.alpha, args.dataset, args.seed, "validation"
+    )
+    val_per_client_rows = []
+    for client_id in range(args.clients):
+        idx = val_eval_assignments[client_id]
+        if len(idx) == 0:
+            val_per_client_rows.append({
+                "client_id": client_id, "n_samples": 0, "accuracy": None,
+                "macro_f1": None, "rare_class_macro_f1": None, "rare_class_recall": None,
+            })
+            continue
+        emb = extract_embeddings(model, val_features[idx])
+        pred = nearest_prototype_predict(emb, global_prototypes)
+        metrics = classification_metrics(val_labels[idx], pred, bundle.num_classes)
+        rare_ids = per_client_rare[client_id]
+        if rare_ids:
+            rare = rare_class_metrics(val_labels[idx], pred, bundle.num_classes, rare_ids)
+        else:
+            rare = {"rare_class_macro_f1": None, "rare_class_recall": None}
+        val_per_client_rows.append({
+            "client_id": client_id,
+            "n_samples": int(len(idx)),
+            "accuracy": metrics["accuracy"],
+            "macro_f1": metrics["macro_f1"],
+            "rare_class_macro_f1": rare["rare_class_macro_f1"],
+            "rare_class_recall": rare["rare_class_recall"],
+        })
+
+    def _mean_over_present(rows, key):
+        values = [float(row[key]) for row in rows if row.get(key) is not None]
+        return float(sum(values) / len(values)) if values else None
+
+    val_selection = {
+        "val_rare_class_macro_f1_global": val_rare["rare_class_macro_f1"],
+        "val_rare_class_recall_global": val_rare["rare_class_recall"],
+        "val_macro_f1_global": val_metrics["macro_f1"],
+        "val_weighted_f1_global": val_metrics["weighted_f1"],
+        "val_accuracy_global": val_metrics["accuracy"],
+        "val_rare_class_macro_f1_per_client_mean": _mean_over_present(
+            val_per_client_rows, "rare_class_macro_f1"
+        ),
+        "val_rare_class_recall_per_client_mean": _mean_over_present(
+            val_per_client_rows, "rare_class_recall"
+        ),
+        "val_n_eval_clients": sum(1 for row in val_per_client_rows if row["n_samples"] > 0),
+        "selection_metric": "rare_class_macro_f1_per_client_mean",
+        "selection_split": "validation",
+        "test_used_for_selection": False,
+    }
+
     # ---- structured artifacts --------------------------------------------
     mode_tag = args.aggregation_mode
     if args.aggregation_mode == "support_weighted":
@@ -301,6 +361,19 @@ def main(argv=None) -> int:
         for row in test_metrics["per_class"]
     ])
     write_csv(run_dir / "per_client.csv", per_client_rows)
+    write_json(run_dir / "val_global_metrics.json", {
+        "accuracy": val_metrics["accuracy"],
+        "macro_f1": val_metrics["macro_f1"],
+        "weighted_f1": val_metrics["weighted_f1"],
+        **val_rare,
+    })
+    write_json(run_dir / "val_confusion_matrix.json", val_metrics["confusion_matrix"])
+    write_csv(run_dir / "val_per_class.csv", [
+        {**row, "class_name": bundle.class_names[row["class_id"]]}
+        for row in val_metrics["per_class"]
+    ])
+    write_csv(run_dir / "val_per_client.csv", val_per_client_rows)
+    write_json(run_dir / "val_selection.json", val_selection)
     write_csv(run_dir / "train_support_Nij.csv", [
         {"client_id": i, **{bundle.class_names[c]: int(support[i, c]) for c in range(bundle.num_classes)}}
         for i in range(args.clients)