如何在批量学习中为 tf-agents 定义正确的形状

Question

我正在尝试使用 tf_agents 库来训练具有批量学习的 DDPG 代理。 但是，我需要定义一个观察规范和动作规范，其中 state 代理将接收的张量的形状。 我已经设法创建了可以提供数据的轨迹，但是这些轨迹和代理本身的形状不匹配

我尝试使用代理定义更改观察和操作规范。 这是我的代理定义：

observation_spec = TensorSpec(shape = (1,),dtype =  tf.float32)
time_step_spec = time_step.time_step_spec(observation_spec)
action_spec = BoundedTensorSpec([1],tf.float32,minimum = -100, maximum = 100)
actor_network = ActorNetwork(
        input_tensor_spec=observation_spec,
        output_tensor_spec=action_spec,
        fc_layer_params=(100,200,100),
        name="ddpg_ActorNetwork"
    )
critic_net_input_specs = (observation_spec, action_spec)
critic_network = CriticNetwork(
    input_tensor_spec=critic_net_input_specs,
    observation_fc_layer_params=(200,100),
    joint_fc_layer_params=(100,200),
    action_fc_layer_params=None,
    name="ddpg_CriticNetwork"
)



agent = ddpg_agent.DdpgAgent(
    time_step_spec=time_step_spec,
    action_spec=action_spec,
    actor_network=actor_network,
    critic_network=critic_network,
    actor_optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
    critic_optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)
)

这就是轨迹的样子

Trajectory(step_type=<tf.Variable 'Variable:0' shape=(1, 2) dtype=int32, numpy=array([[0, 1]], dtype=int32)>, observation=<tf.Variable 'Variable:0' shape=(1, 2) dtype=int32, numpy=array([[280, 280]], dtype=int32)>, action=<tf.Variable 'Variable:0' shape=(1, 2) dtype=float64, numpy=array([[nan,  0.]])>, policy_info=(), next_step_type=<tf.Variable 'Variable:0' shape=(1, 2) dtype=int32, numpy=array([[1, 1]], dtype=int32)>, reward=<tf.Variable 'Variable:0' shape=(1, 2) dtype=float64, numpy=array([[ -6.93147181, -12.14113521]])>, discount=<tf.Variable 'Variable:0' shape=(1, 2) dtype=float32, numpy=array([[0.9, 0.9]], dtype=float32)>)

我应该能够调用 agent.train(trajectory) 并且它会工作但是我收到以下错误：

ValueError                                Traceback (most recent call last)
<ipython-input-325-bf162a5dc8d7> in <module>
----> 1 agent.train(trajs[0])

~/.local/lib/python3.7/site-packages/tf_agents/agents/tf_agent.py in train(self, experience, weights)
    213           "experience must be type Trajectory, saw type: %s" % type(experience))
    214 
--> 215     self._check_trajectory_dimensions(experience)
    216 
    217     if self._enable_functions:

~/.local/lib/python3.7/site-packages/tf_agents/agents/tf_agent.py in _check_trajectory_dimensions(self, experience)
    137     if not nest_utils.is_batched_nested_tensors(
    138         experience, self.collect_data_spec,
--> 139         num_outer_dims=self._num_outer_dims):
    140       debug_str_1 = tf.nest.map_structure(lambda tp: tp.shape, experience)
    141       debug_str_2 = tf.nest.map_structure(lambda spec: spec.shape,

~/.local/lib/python3.7/site-packages/tf_agents/utils/nest_utils.py in is_batched_nested_tensors(tensors, specs, num_outer_dims)
    142       'And spec_shapes:\n   %s' %
    143       (num_outer_dims, tf.nest.pack_sequence_as(tensors, tensor_shapes),
--> 144        tf.nest.pack_sequence_as(specs, spec_shapes)))
    145 
    146 

ValueError: Received a mix of batched and unbatched Tensors, or Tensors are not compatible with Specs.  num_outer_dims: 2.
Saw tensor_shapes:
   Trajectory(step_type=TensorShape([1, 2]), observation=TensorShape([1, 2]), action=TensorShape([1, 2]), policy_info=(), next_step_type=TensorShape([1, 2]), reward=TensorShape([1, 2]), discount=TensorShape([1, 2]))
And spec_shapes:
   Trajectory(step_type=TensorShape([]), observation=TensorShape([1]), action=TensorShape([1]), policy_info=(), next_step_type=TensorShape([]), reward=TensorShape([]), discount=TensorShape([]))

Answer 1

这可以通过使用环境轻松解决。 在 TF-Agents 中，环境需要遵循PyEnvironment class（然后用TFPyEnvironment包装它以并行执行多个 env）。 如果你已经定义了你的环境来匹配这个类的规范，那么你的环境应该已经为你提供了两个方法env.time_step_spec()和env.action_spec() 。 只需将这两个提供给您的代理，您就应该完成。

如果您想从您的环境中获得多个输出，而不是所有 go 进入您的代理，它会变得有点复杂。 在这种情况下，您需要定义一个observation_and_action_constraint_splitter函数以传递给您的代理。 有关如何正确设置 TensorSpecs/ArraySpecs 的更多详细信息，以及可行的示例，请参见我的回答here