python - Understanding custom policies in stable-baselines3

Question

I was trying to understand the policy networks in stable-baselines3 from this doc page.

1. As explained in this example, to specify custom CNN feature extractor, we extend BaseFeaturesExtractor class and specify it in policy_kwarg.features_extractor_class with first param CnnPolicy:

   model = PPO("CnnPolicy", "BreakoutNoFrameskip-v4", policy_kwargs=policy_kwargs)
   

   Q1. Can we follow same approach for custom MLP feature extractor?

2. As explained in this example, to specify custom MLP feature extractor, we extend ActorCriticPolicy class and override _build_mlp_extractor() and pass it as first param:

   class CustomActorCriticPolicy(ActorCriticPolicy): ...
   model = PPO(CustomActorCriticPolicy, "CartPole-v1", verbose=1)
   

   Q2. Can we follow same approach for custom CNN feature extractor?

3. I feel either we can have CNN extractor or MLP extractor. So it makes no sense to pass MlpPolicy as first param to model and then specify CNN feature extractor in policy_kwarg.features_extractor_class as in this example. This result in following policy (containing both features_extractor and mlp_extractor), which I feel is incorrect:

   ActorCriticPolicy(
     (features_extractor): Net(
       (conv1): Conv2d(1, 32, kernel_size=(3, 3), stride=(1, 1))
       (conv2): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1))
       (fc3): Linear(in_features=384, out_features=512, bias=True)
     )
     (mlp_extractor): MlpExtractor(
       (shared_net): Sequential(
         (0): Linear(in_features=512, out_features=64, bias=True)
         (1): ReLU()
       )
       (policy_net): Sequential(
         (0): Linear(in_features=64, out_features=32, bias=True)
         (1): ReLU()
         (2): Linear(in_features=32, out_features=16, bias=True)
         (3): ReLU()
       )
       (value_net): Sequential(
         (0): Linear(in_features=64, out_features=32, bias=True)
         (1): ReLU()
         (2): Linear(in_features=32, out_features=16, bias=True)
         (3): ReLU()
       )
     )
     (action_net): Linear(in_features=16, out_features=7, bias=True)
     (value_net): Linear(in_features=16, out_features=1, bias=True)
   )  
   

   Q3. Am I correct with this understanding? If yes, then is one of the MLP or CNN feature extractor ignored?

Answer 1

在我浏览完所有库代码后我能说什么。仅在实现的默认类CnnPolicy中有所不同。这仅在您不尝试创建自定义类的情况下才有意义。让我试着解释一下，我们可以看到两种类型的策略：MlpPolicyBaseFeatureExtractionBaseFeatureExtraction

MlpPolicy = ActorCriticPolicy
CnnPolicy = ActorCriticCnnPolicy

我们可以看到这class ActorCriticCnnPolicy(ActorCriticPolicy)只是基于ActorCriticPolicy，我们可以在参数中看到以下内容：

features_extractor_class: Type[BaseFeaturesExtractor] = NatureCNN

哪里NatureCNN是BaseFeatureExtractorCNN 层的简单实现。

现在让我们来回答你的问题！这是一件事，我们总是有以下模型结构： FeatureExtractor -> MlpExtractor -> Policy\Value nets

您可以在 policy_kwargs中指定feature_exctractor_class和。net_arch默认 feature_extractor_class 基于您对CnnPolicyor的选择MlpPolicy，但如果您指定自己的类，则没有区别。所以，你可以只使用MlpPolicy.

net_arch例如，如果您指定，'net_arch':[64, dict(pi=[32, 16], vf=[32, 16])]您将有 1 个用于连接 feature_extractor 输出的 mlp_extractor 密集层。

总而言之，我的建议是遵循您在Q3中找到的示例：

1. feature_extractor使用此网络的任何结构指定您自己的 cnn 或密集或任何您想要的。的输出feature_extractor应该是FC层。
2. net_acrh在以下结构中指定参数：[x1, x2, ..., dict(pi=[px1, px2, ..], vf=[vx1, vx2, ...])]其中第一部分是指定的 FC 层大小，mlp_extractor第二部分（在 dict 内部）对应于策略和价值网络的大小。