Fix bug with distributed training and fixed composition (#308)

src/metatrain/experimental/soap_bpnn/trainer.py

@@ -103,15 +103,18 @@ def train(
                     "user-supplied composition weights"
                 )
                 cur_weight_dict = self.hypers["fixed_composition_weights"][target_name]
-                atomic_types = set()
+                atomic_types = []
                 num_species = len(cur_weight_dict)
                 fixed_weights = torch.zeros(num_species, dtype=dtype, device=device)
 
                 for ii, (key, weight) in enumerate(cur_weight_dict.items()):
-                    atomic_types.add(key)
+                    atomic_types.append(key)
                     fixed_weights[ii] = weight
 
-                if not set(atomic_types) == model.atomic_types:
+                if (
+                    not set(atomic_types)
+                    == (model.module if is_distributed else model).atomic_types
+                ):
                     raise ValueError(
                         "Supplied atomic types are not present in the dataset."
                     )