bug: the comment doesn’t align with the code for parameter start_training

[yuzh2001]

In the config comment, it says:

# xuance/configs/matd3/mpe/simple_push_v3.yaml

start_training: 1000  # start training after n episodes

However, in off_policy_marl as an example,

for _ in tqdm(range(n_steps)):
            step_info = {}
            policy_out = self.action(obs_dict=obs_dict, avail_actions_dict=avail_actions, test_mode=False)
            actions_dict = policy_out['actions']
            next_obs_dict, rewards_dict, terminated_dict, truncated, info = self.envs.step(actions_dict)
            next_state = self.envs.buf_state.copy() if self.use_global_state else None
            next_avail_actions = self.envs.buf_avail_actions if self.use_actions_mask else None
            self.store_experience(obs_dict, avail_actions, actions_dict, next_obs_dict, next_avail_actions,
                                  rewards_dict, terminated_dict, info,
                                  **{'state': state, 'next_state': next_state})
+            if self.current_step >= self.start_training and self.current_step % self.training_frequency == 0:
                train_info = self.train_epochs(n_epochs=self.n_epochs)
                self.log_infos(train_info, self.current_step)
                return_info.update(train_info)
            obs_dict = deepcopy(next_obs_dict)
            if self.use_global_state:
                state = deepcopy(next_state)
            if self.use_actions_mask:
                avail_actions = deepcopy(next_avail_actions)

            for i in range(self.n_envs):
                if all(terminated_dict[i].values()) or truncated[i]:
                    obs_dict[i] = info[i]["reset_obs"]
                    self.envs.buf_obs[i] = info[i]["reset_obs"]
                    if self.use_global_state:
                        state = info[i]["reset_state"]
                        self.envs.buf_state[i] = info[i]["reset_state"]
                    if self.use_actions_mask:
                        avail_actions[i] = info[i]["reset_avail_actions"]
                        self.envs.buf_avail_actions[i] = info[i]["reset_avail_actions"]
                    if self.use_wandb:
                        step_info[f"Train-Results/Episode-Steps/rank_{self.rank}/env-%d" % i] = info[i]["episode_step"]
                        step_info[f"Train-Results/Episode-Rewards/rank_{self.rank}/env-%d" % i] = info[i]["episode_score"]
                    else:
                        step_info[f"Train-Results/Episode-Steps/rank_{self.rank}"] = {
                            "env-%d" % i: info[i]["episode_step"]}
                        step_info[f"Train-Results/Episode-Rewards/rank_{self.rank}"] = {
                            "env-%d" % i: np.mean(itemgetter(*self.agent_keys)(info[i]["episode_score"]))}
                    self.log_infos(step_info, self.current_step)
                    return_info.update(step_info)

+            self.current_step += self.n_envs
            self._update_explore_factor()

It is evident that start_training should be configured based on timesteps rather than episodes.

—

When defining environments, there is this parameter self.max_episode_steps. If the start_training parameter should be based on episode counts, maybe it should be calulated as self.envs[1].max_episode_steps * self.start_training.

[wenzhangliu]

Thank you for pointing out the mistake regarding the note for start_training; it should indeed be “start training after n steps”. I’ve corrected it accordingly.