the last system act is now updated in environment to be consistent with...

the last system act is now updated in environment to be consistent with pipeline agent and not let the policy change it, which is more errorprone

convlab/dialog_agent/env.py

@@ -28,6 +28,8 @@ class Environment():
         return self.sys_dst.state
 
     def step(self, action):
+        # save last system action
+        self.sys_dst.state['system_action'] = action
         if not self.use_semantic_acts:
             model_response = self.sys_nlg.generate(
                 action) if self.sys_nlg else action
@@ -49,7 +51,6 @@ class Environment():
         self.sys_dst.state['user_action'] = dialog_act
         state = self.sys_dst.update(dialog_act)
         state = deepcopy(state)
-        dialog_act = self.sys_dst.state['user_action']
 
         state['history'].append(["sys", model_response])
         state['history'].append(["usr", observation])

convlab/policy/gdpl/gdpl.py

@@ -76,7 +76,6 @@ class GDPL(Policy):
         # print('True :')
         # print(a)
         action = self.vector.action_devectorize(a.detach().numpy())
-        state['system_action'] = action
         self.info_dict["action_used"] = action
         # for key in state.keys():
         #     print("Key : {} , Value : {}".format(key,state[key]))

convlab/policy/pg/pg.py

@@ -74,7 +74,6 @@ class PG(Policy):
         # print('True :')
         # print(a)
         action = self.vector.action_devectorize(a.detach().numpy())
-        state['system_action'] = action
         self.info_dict["action_used"] = action
         # for key in state.keys():
         #     print("Key : {} , Value : {}".format(key,state[key]))

convlab/policy/ppo/ppo.py

@@ -85,7 +85,6 @@ class PPO(Policy):
         # print('True :')
         # print(a)
         action = self.vector.action_devectorize(a.detach().numpy())
-        state['system_action'] = action
         self.info_dict["action_used"] = action
         # for key in state.keys():
         #     print("Key : {} , Value : {}".format(key,state[key]))