问不能加载保存策略(TF-代理)
EN

Stack Overflow用户

提问于 2020-12-25 05:03:28

回答 2查看 780关注 0票数 3

我用策略保护程序保存了经过培训的策略如下：

  tf_policy_saver = policy_saver.PolicySaver(agent.policy)
  tf_policy_saver.save(policy_dir)

我想继续用保存的政策训练。因此，我尝试使用保存的策略初始化培训，这会导致一些错误。

agent = dqn_agent.DqnAgent(
tf_env.time_step_spec(),
tf_env.action_spec(),
q_network=q_net,
optimizer=optimizer,
td_errors_loss_fn=common.element_wise_squared_loss,
train_step_counter=train_step_counter)

agent.initialize()

agent.policy=tf.compat.v2.saved_model.load(policy_dir)

错误：

  File "C:/Users/Rohit/PycharmProjects/pythonProject/waypoint.py", line 172, in <module>
agent.policy=tf.compat.v2.saved_model.load('waypoints\\Two_rewards')


File "C:\Users\Rohit\anaconda3\envs\btp36\lib\site-packages\tensorflow\python\training\tracking\tracking.py", line 92, in __setattr__
    super(AutoTrackable, self).__setattr__(name, value)
AttributeError: can't set attribute

我只想从每次再培训中节省时间。如何加载保存的策略并继续培训？

提前感谢

tensorflow

deep-learning

reinforcement-learning

TDSQL-CMySQL免费体验

云原生数据库TDSQL-C MySQL 2核4G 免费体验15天

回答 2

Stack Overflow用户

回答已采纳

发布于 2021-01-01 01:16:44

是的，如前所述，您应该使用Check指针来实现这一点--请看下面的示例代码。

agent = ... # Agent Definition
policy = agent.policy
# Policy --> Y
policy_checkpointer = common.Checkpointer(ckpt_dir='path/to/dir',
                                          policy=policy)

... # Train the agent

# Policy --> X
policy_checkpointer.save(global_step=epoch_counter.numpy())

稍后要重新加载策略时，只需运行相同的初始化代码。

agent = ... # Agent Definition
policy = agent.policy
# Policy --> Y1, possibly Y1==Y depending on agent class you are using, if it's DQN
#               then they are different because of random initialization of network weights
policy_checkpointer = common.Checkpointer(ckpt_dir='path/to/dir',
                                          policy=policy)
# Policy --> X

创建时，policy_checkpointer将自动意识到是否存在任何预先存在的检查点。如果存在，它将在创建时自动更新正在跟踪的变量的值。

有几个笔记要做：

不仅可以保存策略，而且我建议这样做。TF-代理的Check指针对象非常灵活，例如：

train_checkpointer = common.Checkpointer(ckpt_dir=first/dir,
                                         agent=tf_agent,               # tf_agent.TFAgent
                                         train_step=train_step,        # tf.Variable
                                         epoch_counter=epoch_counter,  # tf.Variable
                                         metrics=metric_utils.MetricsGroup(
                                                 train_metrics, 'train_metrics'))

policy_checkpointer = common.Checkpointer(ckpt_dir=second/dir,
                                          policy=agent.policy)

rb_checkpointer = common.Checkpointer(ckpt_dir=third/dir,
                                      max_to_keep=1,
                                      replay_buffer=replay_buffer  # TFUniformReplayBuffer
                                      )

注意到，在DqnAgent的情况下，agent.policy和agent.collect_policy本质上是QNetwork的包装器。下面的代码(查看有关策略变量状态的注释)

显示了这一点的含义。

agent = DqnAgent(...)
policy = agent.policy      # Random initial policy ---> X

dataset = replay_buffer.as_dataset(...)
for data in dataset:
   experience, _ = data
   loss_agent_info = agent.train(experience=experience)

# policy variable stores a trained Policy object ---> Y