Bugfix to adapters: incorrect key caused adapters to not be applied.

Interestingly, random projection seems to get pretty far in the toy
tasks, so this was not immediately obvious. The bug was detected by
observing that the adapter parameters did not change during training.

mammoth/modules/adapters.py

@@ -77,8 +77,9 @@ def is_wrapper(self):
 
     def apply(self, tmp_layer_types, tmp_layer_structs, tmp_layer_dropouts):
         # LoraAdapterLayer wraps the existing feedforward. No norms are added.
-        tmp_layer_structs[0][1] = self.wrap(tmp_layer_structs[0][1])
-        return tmp_layer_types, tmp_layer_structs, tmp_layer_dropouts
+        new_layer_structs = [nn.ModuleList(x) for x in tmp_layer_structs]
+        new_layer_structs[0][1] = self.wrap(new_layer_structs[0][1])
+        return tmp_layer_types, new_layer_structs, tmp_layer_dropouts
 
     def wrap(self, base_layer):
         self.wrapped_base_layer = base_layer
@@ -201,7 +202,7 @@ def _inject_adapters(self):
         ):
             if layer_type == 'f':
                 # Adapters apply to feedforward layers
-                adapter_layers = adapter_layers_by_index[i]
+                adapter_layers = adapter_layers_by_index[f'layer{i}']
                 tmp_layer_types = [layer_type]
                 tmp_layer_structs = [layer_struct]
                 tmp_layer_dropouts = [layer_dropout]