[Bugfix] Fix the case when the input value of wd_l2norm is using scie…

…ntific notation.

python/graphstorm/config/argument.py

@@ -1723,7 +1723,12 @@ def wd_l2norm(self):
         """
         # pylint: disable=no-member
         if hasattr(self, "_wd_l2norm"):
-            return self._wd_l2norm
+            try:
+                wd_l2norm = float(self._wd_l2norm)
+            except:
+                raise ValueError("wd-l2norm must be a floating point " \
+                                 f"but get {self._wd_l2norm}")
+            return wd_l2norm
         return 0
 
     @property

tests/unit-tests/test_config.py

@@ -252,6 +252,7 @@ def create_train_config(tmp_path, file_name):
     yaml_object["gsf"]["hyperparam"] = {
         "topk_model_to_save": 4,
         "save_model_path": os.path.join(tmp_path, "save"),
+        "wd_l2norm": 5e-5,
     }
     with open(os.path.join(tmp_path, file_name+"1.yaml"), "w") as f:
         yaml.dump(yaml_object, f)
@@ -261,6 +262,7 @@ def create_train_config(tmp_path, file_name):
         'save_model_frequency': 2000,
         "topk_model_to_save": 5,
         "save_model_path": os.path.join(tmp_path, "save"),
+        "wd_l2norm": "1e-3",
     }
     with open(os.path.join(tmp_path, file_name+"2.yaml"), "w") as f:
         yaml.dump(yaml_object, f)
@@ -291,6 +293,7 @@ def create_train_config(tmp_path, file_name):
         "use_early_stop": True,
         "early_stop_burnin_rounds": -1,
         "early_stop_rounds": 0,
+        "wd_l2norm": "NA",
     }
 
     with open(os.path.join(tmp_path, file_name+"_fail.yaml"), "w") as f:
@@ -301,6 +304,7 @@ def create_train_config(tmp_path, file_name):
         'save_model_frequency': 2000,
         "topk_model_to_save": 3,
         "save_model_path": os.path.join(tmp_path, "save"),
+        "wd_l2norm": "",
     }
     with open(os.path.join(tmp_path, file_name+"_fail1.yaml"), "w") as f:
         yaml.dump(yaml_object, f)
@@ -350,12 +354,14 @@ def test_train_info():
         args = Namespace(yaml_config_file=os.path.join(Path(tmpdirname), 'train_test1.yaml'), local_rank=0)
         config = GSConfig(args)
         assert config.topk_model_to_save == 4
+        assert config.wd_l2norm == 5e-5
 
         args = Namespace(yaml_config_file=os.path.join(Path(tmpdirname), 'train_test2.yaml'), local_rank=0)
         config = GSConfig(args)
         assert config.eval_frequency == 1000
         assert config.save_model_frequency == 2000
         assert config.topk_model_to_save == 5
+        assert config.wd_l2norm == 1e-3
 
         args = Namespace(yaml_config_file=os.path.join(Path(tmpdirname), 'train_test3.yaml'), local_rank=0)
         config = GSConfig(args)
@@ -380,12 +386,14 @@ def test_train_info():
         check_failure(config, "topk_model_to_save")
         check_failure(config, "early_stop_burnin_rounds")
         check_failure(config, "early_stop_rounds")
+        check_failure(config, "wd_l2norm")
 
         args = Namespace(yaml_config_file=os.path.join(Path(tmpdirname), 'train_test_fail1.yaml'), local_rank=0)
         config = GSConfig(args)
         # in PR # 893 we loose the constraints of model saving frequency and eval frequency
         # so here we do not check failure, but check the topk model argument
         assert config.topk_model_to_save == 3
+        check_failure(config, "wd_l2norm")
 
 def create_rgcn_config(tmp_path, file_name):
     yaml_object = create_dummpy_config_obj()