Merge pull request #3125 from jinjiren/enalbe_freezing_unfreezing_wit…

…h_resuming

Fix the problem with resuming training when switching the freezing layers

chainer/optimizer.py

@@ -264,6 +264,12 @@ def serialize(self, serializer):
 
                 for key in self._state:
                     self._state[key] = serializer(key, None)
+                    # leave the update rule state as `None` if the keys are not
+                    # contained in the snapshot, so that these states can be
+                    # automatically initialized with the `_prepare` method
+                    if self._state[key] is None:
+                        self._state = None
+                        break
         else:
             for key in self._state:
                 self._state[key] = serializer(key, self._state[key])