Restore and continue training Tuner() and AIR

Hi,

I am using ray 2.0.1 to train an RL agent. I am using the new Tuner() module, with AIR library. I am able to successfully checkpoint the model during training. I am also able to restore the model to continue training the model. But the issue I have is that the restore and continue starts from step-0 again and does not continue from the state of the last restore. This is especially a problem for RL training because some parameters such as learning rate, exploration rate and configured based on the number of steps the agent experienced. For example, the restore state looks like this author mentioned:

as you can see the ‘_timesteps_total’: None is saved as none instead of the current timesteps_total

Is there a way to restore and continue training from the last know time_steps?

Does your rl agent report timesteps_total?

I don’t think so. Can you please let me know how my agent can report it? Do I have to create a custom reporting module?

According to this link: A Guide To Callbacks & Metrics in Tune — Ray 2.1.0

timesteps_total is automatically reported.

ah sorry, I meant does your rl agent report timesteps_this_iter?

If I am reading the code correctly

        # self._timesteps_total should only be tracked if increments provided
        if result.get(TIMESTEPS_THIS_ITER) is not None:
            if self._timesteps_total is None:
                self._timesteps_total = 0
            self._timesteps_total += result[TIMESTEPS_THIS_ITER]
            self._timesteps_since_restore += result[TIMESTEPS_THIS_ITER]

^^ in trainable.py
We can only auto generate timesteps_total if timesteps_this_iter is reported in result.

This is the exact problem. I was never getting this ‘timesteps_this_iter’ value. It is always none. How can I get the reporter/train() to get this value?

which rl agent are you using? Are you using rllib’s algo?

If you are implementing a class trainable, in step function, you will return a dict. You will need to make sure that timesteps_this_iter is one key in that dict.
If you are using a function, then make sure that you do report() with one kwargs being timesteps_this_iter.

I’m using default dqn rllib algo along with Tuner(). I am not using any custom trainer. What should I do in that case?

@arturn Does RLlib populate this timesteps_this_iter thing at all? I can’t seem to find it. btw, these are all the fields I see from a random PPO result

Result for PPO_CartPole-v0_364df_00000:
  agent_timesteps_total: 4000
  counters:
    num_agent_steps_sampled: 4000
    num_agent_steps_trained: 4000
    num_env_steps_sampled: 4000
    num_env_steps_trained: 4000
  custom_metrics: {}
  date: 2022-11-09_16-40-56
  done: false
  episode_len_mean: 23.48235294117647
  episode_media: {}
  episode_reward_max: 75.0
  episode_reward_mean: 23.48235294117647
  episode_reward_min: 9.0
  episodes_this_iter: 170
  episodes_total: 170
  experiment_id: 8bffa69bf5e04857aa83536c8ef3736a
  hostname: xw
  info:
    learner:
      default_policy:
        custom_metrics: {}
        learner_stats:
          cur_kl_coeff: 0.20000000298023224
          cur_lr: 4.999999873689376e-05
          entropy: 0.665932297706604
          entropy_coeff: 0.0
          kl: 0.027817677706480026
          policy_loss: -0.039548929780721664
          total_loss: 8.858890533447266
          vf_explained_var: 0.008993614464998245
          vf_loss: 8.892877578735352
        num_agent_steps_trained: 125.0
    num_agent_steps_sampled: 4000
    num_agent_steps_trained: 4000
    num_env_steps_sampled: 4000
    num_env_steps_trained: 4000
  iterations_since_restore: 1
  node_ip: 127.0.0.1
  num_agent_steps_sampled: 4000
  num_agent_steps_trained: 4000
  num_env_steps_sampled: 4000
  num_env_steps_sampled_this_iter: 4000
  num_env_steps_trained: 4000
  num_env_steps_trained_this_iter: 4000
  num_faulty_episodes: 0
  num_healthy_workers: 2
  num_recreated_workers: 0
  num_steps_trained_this_iter: 4000
  perf:
    cpu_util_percent: 14.412244897959184
    ram_util_percent: 55.08775510204081
  pid: 43619
  policy_reward_max: {}
  policy_reward_mean: {}
  policy_reward_min: {}
  sampler_perf:
    mean_action_processing_ms: 0.06086085844030104
    mean_env_render_ms: 0.0
    mean_env_wait_ms: 0.06126414658226595
    mean_inference_ms: 3.9059675275660926
    mean_raw_obs_processing_ms: 0.33546808511207116
  sampler_results:
    custom_metrics: {}
    episode_len_mean: 23.48235294117647
    episode_media: {}
    episode_reward_max: 75.0
    episode_reward_mean: 23.48235294117647
    episode_reward_min: 9.0
    episodes_this_iter: 170
    hist_stats:
      episode_lengths: [17, 15, 20, 15, 19, 15, 18, 15, 18, 15, 25, 23, 14, 20, 75,
        30, 10, 20, 54, 28, 31, 44, 50, 10, 28, 16, 34, 12, 17, 44, 26, 41, 9, 51, 71,
        16, 12, 19, 29, 14, 27, 23, 38, 11, 24, 30, 40, 15, 21, 11, 25, 36, 30, 58,
        19, 18, 37, 29, 24, 32, 17, 13, 26, 13, 22, 25, 14, 25, 21, 15, 13, 15, 9, 24,
        17, 34, 35, 18, 19, 21, 14, 18, 40, 15, 14, 17, 23, 17, 10, 15, 17, 15, 38,
        33, 15, 26, 12, 16, 13, 24, 15, 41, 13, 20, 31, 18, 18, 17, 45, 25, 25, 18,
        15, 12, 18, 33, 30, 12, 22, 50, 13, 23, 25, 18, 56, 13, 11, 24, 22, 67, 23,
        13, 21, 30, 15, 26, 25, 26, 14, 59, 27, 45, 19, 34, 11, 16, 15, 12, 10, 26,
        12, 17, 15, 10, 23, 13, 39, 28, 58, 17, 14, 21, 12, 25, 19, 20, 14, 14, 12,
        16]
      episode_reward: [17.0, 15.0, 20.0, 15.0, 19.0, 15.0, 18.0, 15.0, 18.0, 15.0, 25.0,
        23.0, 14.0, 20.0, 75.0, 30.0, 10.0, 20.0, 54.0, 28.0, 31.0, 44.0, 50.0, 10.0,
        28.0, 16.0, 34.0, 12.0, 17.0, 44.0, 26.0, 41.0, 9.0, 51.0, 71.0, 16.0, 12.0,
        19.0, 29.0, 14.0, 27.0, 23.0, 38.0, 11.0, 24.0, 30.0, 40.0, 15.0, 21.0, 11.0,
        25.0, 36.0, 30.0, 58.0, 19.0, 18.0, 37.0, 29.0, 24.0, 32.0, 17.0, 13.0, 26.0,
        13.0, 22.0, 25.0, 14.0, 25.0, 21.0, 15.0, 13.0, 15.0, 9.0, 24.0, 17.0, 34.0,
        35.0, 18.0, 19.0, 21.0, 14.0, 18.0, 40.0, 15.0, 14.0, 17.0, 23.0, 17.0, 10.0,
        15.0, 17.0, 15.0, 38.0, 33.0, 15.0, 26.0, 12.0, 16.0, 13.0, 24.0, 15.0, 41.0,
        13.0, 20.0, 31.0, 18.0, 18.0, 17.0, 45.0, 25.0, 25.0, 18.0, 15.0, 12.0, 18.0,
        33.0, 30.0, 12.0, 22.0, 50.0, 13.0, 23.0, 25.0, 18.0, 56.0, 13.0, 11.0, 24.0,
        22.0, 67.0, 23.0, 13.0, 21.0, 30.0, 15.0, 26.0, 25.0, 26.0, 14.0, 59.0, 27.0,
        45.0, 19.0, 34.0, 11.0, 16.0, 15.0, 12.0, 10.0, 26.0, 12.0, 17.0, 15.0, 10.0,
        23.0, 13.0, 39.0, 28.0, 58.0, 17.0, 14.0, 21.0, 12.0, 25.0, 19.0, 20.0, 14.0,
        14.0, 12.0, 16.0]
    num_faulty_episodes: 0
    policy_reward_max: {}
    policy_reward_mean: {}
    policy_reward_min: {}
    sampler_perf:
      mean_action_processing_ms: 0.06086085844030104
      mean_env_render_ms: 0.0
      mean_env_wait_ms: 0.06126414658226595
      mean_inference_ms: 3.9059675275660926
      mean_raw_obs_processing_ms: 0.33546808511207116
  time_since_restore: 34.13855314254761
  time_this_iter_s: 34.13855314254761
  time_total_s: 34.13855314254761
  timers:
    learn_throughput: 157.822
    learn_time_ms: 25345.008
    synch_weights_time_ms: 3.063
    training_iteration_time_ms: 34133.688
  timestamp: 1668040856
  timesteps_since_restore: 0
  timesteps_total: 4000
  training_iteration: 1
  trial_id: 364df_00000
  warmup_time: 9.969342947006226

From my experience, I don’t think this is being populated. Here is a sample output after 1 iteration of train:

result: {'custom_metrics': {}, 'episode_media': {}, 'num_recreated_workers': 0, 'info': {'learner': {}, 'num_env_steps_sampled': 100, 'num_env_steps_trained': 0, 'num_agent_steps_sampled': 100, 'num_agent_steps_trained': 0}, 'sampler_results': {'episode_reward_max': nan, 'episode_reward_min': nan, 'episode_reward_mean': nan, 'episode_len_mean': nan, 'episode_media': {}, 'episodes_this_iter': 0, 'policy_reward_min': {}, 'policy_reward_max': {}, 'policy_reward_mean': {}, 'custom_metrics': {}, 'hist_stats': {'episode_reward': [], 'episode_lengths': []}, 'sampler_perf': {}, 'num_faulty_episodes': 0}, 'episode_reward_max': nan, 'episode_reward_min': nan, 'episode_reward_mean': nan, 'episode_len_mean': nan, 'episodes_this_iter': 0, 'policy_reward_min': {}, 'policy_reward_max': {}, 'policy_reward_mean': {}, 'hist_stats': {'episode_reward': [], 'episode_lengths': []}, 'sampler_perf': {}, 'num_faulty_episodes': 0, 'num_healthy_workers': 0, 'num_agent_steps_sampled': 100, 'num_agent_steps_trained': 0, 'num_env_steps_sampled': 100, 'num_env_steps_trained': 0, 'num_env_steps_sampled_this_iter': 100, 'num_env_steps_trained_this_iter': 0, 'timesteps_total': 100, 'num_steps_trained_this_iter': 0, 'agent_timesteps_total': 100, 'timers': {'training_iteration_time_ms': 2307.899}, 'counters': {'num_env_steps_sampled': 100, 'num_env_steps_trained': 0, 'num_agent_steps_sampled': 100, 'num_agent_steps_trained': 0}, 'done': False}

@xwjiang2010 timesteps_this_iter is there for backward compatibility.
These days we use "num_agent_steps_sampled_this_iter" and num_env_steps_sampled_this_iter to discriminate between these sorts of timesteps.

These accumulate to num_env_steps_sampled and num_agent_steps_sampled . Which should always be the same as num_agent_steps_trained and num_env_steps_trained for PPO, since it is a synchronous algorithm.

@varunjammula Please upgrade to Ray 2.1, this has been fixed there.

1 Like

Thanks, it is working now.