Bugfix: Crash with XGBoost due to old dependency

.github/workflows/tests.yml

@@ -4,7 +4,7 @@ on:
   push:
     branches: [main]
   pull_request:
-    branches: [main, cmd_line]
+    branches: [main, cmd_line, tests_and_CI]
   workflow_dispatch:
 
 concurrency:

flcore/models/xgb/client.py

@@ -144,8 +144,11 @@ def fit(
             }
 
             # If multiclass objective, ensure num_class is correctly set from local data
+            n_classes = len(np.unique(self.local_data['y_train']))
+            if self.xgb_params.get("objective") == "binary:logistic" and n_classes > 2:
+                self.xgb_params["objective"] = "multi:softmax"
+                self.xgb_params["eval_metric"] = "mlogloss"
             if self.xgb_params.get("objective", "").startswith("multi"):
-                n_classes = len(np.unique(self.local_data['y_train']))
                 if n_classes >= 2:
                     self.xgb_params["num_class"] = n_classes
                 print(f"[Client] Multiclass detected: num_class={n_classes}")

flcore/models/xgb/server.py

@@ -11,6 +11,7 @@
 import xgboost as xgb
 
 from flwr.common import (
+    FitIns,
     ArrayRecord, 
     ConfigRecord, 
     Message, 
@@ -174,6 +175,34 @@ def initialize_parameters(self, client_manager):
         empty = np.frombuffer(b"", dtype=np.uint8)
         return ndarrays_to_parameters([empty])
 
+    # ------------------------------------------------------
+    # CONFIGURE FIT
+    # ------------------------------------------------------
+
+    def configure_fit(self, server_round, parameters, client_manager):
+        """Send each client its final learning rate: eta / number of clients."""
+        # Without scaling, the training becomes unstable with more clients
+        instructions = super().configure_fit(server_round, parameters, client_manager)
+        num_clients = len(instructions)
+        if num_clients == 0:
+            return instructions
+
+        base_eta = float(self.xgb_params.get("eta", 0.1))
+        final_eta = base_eta / num_clients
+        print(
+            f"[Round {server_round}] eta: {base_eta} / {num_clients} clients "
+            f"= {final_eta}"
+        )
+
+        # One FitIns per client, each carrying its own final eta
+        return [
+            (
+                client,
+                FitIns(fit_ins.parameters, {**fit_ins.config, "eta": final_eta}),
+            )
+            for client, fit_ins in instructions
+        ]
+
     # ------------------------------------------------------
     # AGGREGATE FIT (CRITICAL FIX)
     # ------------------------------------------------------
@@ -233,8 +262,12 @@ def aggregate_fit(
         total_examples = sum(fit_res.num_examples for _, fit_res in results)
 
         for _, fit_res in results:
-            if "n_out" in fit_res.metrics:
-                self.xgb_params["num_class"] = int(fit_res.metrics["n_out"])
+            n_out = int(fit_res.metrics.get("n_out", 0))
+            if self.xgb_params["objective"] == "binary:logistic" and n_out > 2:
+                self.xgb_params["objective"] = "multi:softmax"
+                self.xgb_params["eval_metric"] = "mlogloss"
+            if self.xgb_params["objective"].startswith("multi"):
+                self.xgb_params["num_class"] = n_out
 
             for key, value in fit_res.metrics.items():
                 if not isinstance(value, (int, float)):
@@ -361,7 +394,7 @@ def get_server_and_strategy(config: dict) -> FedXgbFullyFederated:
         "colsample_bytree": 0.8,
     }
 
-    if task == "binary":
+    if task in ("binary", "classification"):
         xgb_params["objective"] = "binary:logistic"
         xgb_params["eval_metric"] = "auc"
 
@@ -377,7 +410,7 @@ def get_server_and_strategy(config: dict) -> FedXgbFullyFederated:
         xgb_params["eval_metric"] = "rmse"
 
     train_method = xgb_config.get("train_method", "bagging")  # 'bagging' or 'cyclic'
-    num_local_rounds = xgb_config.get("tree_num", 100) // config.get("num_rounds", 10)  # Trees per round
+    num_local_rounds = 1  # Always train one tree per round
 
 
     print("\n" + "=" * 60)

requirements.txt

@@ -11,7 +11,7 @@ scikit-learn==1.3.2
 torch==2.3.1
 torchmetrics==1.4.0
 tqdm==4.65.0
-xgboost==1.7.5
+xgboost==2.1.4
 pdfkit==1.0.0
 scikit-survival==0.22.2
 fpboost==0.1.0

tests/test_federated_training.py

@@ -51,7 +51,6 @@ def case(model, scenario, client_extra=(), id_suffix="", bug=None):
     "lsvc/svm, but SGDClassifier needs one intercept per class (multiclass)"
 )
 SVR_KERNEL_BUG = "SVR with a non-linear kernel has no fit_intercept / coef_"
-XGB_BUG = "xgb server needs iteration_indptr (XGBoost >= 2.0, 1.7.5 is pinned)"
 SURVIVAL_TIME_BUG = "survival client fit() uses `time` without importing it"
 
 CASES = []
@@ -98,7 +97,7 @@ def case(model, scenario, client_extra=(), id_suffix="", bug=None):
     case("weighted_random_forest", "binary", bug=wrf_bug),
     case("weighted_random_forest", "multiclass", bug=wrf_bug),
 ]
-CASES += [case("xgb", s, bug=XGB_BUG) for s in ("binary", "multiclass", "regression")]
+CASES += [case("xgb", s) for s in ("binary", "multiclass", "regression")]
 CASES += [case("nn", s) for s in ("binary", "multiclass", "regression")]
 
 # Survival