subsampling fix

src/decentralizepy/sharing/SubSampling.py

@@ -101,6 +101,17 @@ class SubSampling(Sharing):
             )
             Path(self.folder_path).mkdir(parents=True, exist_ok=True)
 
+        with torch.no_grad():
+            tensors_to_cat = []
+            for _, v in self.model.state_dict().items():
+                t = v.flatten()
+                tensors_to_cat.append(t)
+            self.init_model = torch.cat(tensors_to_cat, dim=0)
+
+        self.model.shared_parameters_counter = torch.zeros(
+            self.init_model.shape[0], dtype=torch.int32
+        )
+
     def apply_subsampling(self):
         """
         Creates a random binary mask that is used to subsample the parameters that will be shared