AttributeError:“DummyVecEnv”对象没有属性“shape”

Question

我试图为我的强化学习算法创建一个环境，但是，在调用PPOPolicy时，似乎有点问题。为此，我开发了以下环境envFru：

import gym
import os, sys
import numpy as np
import pandas as pd
from gym import spaces
import random

class envFru(gym.Env):
    metadata ={'render.modes': ['human']}
    
    def __init__(self):
        self.df = df
        
        self.action_space = spaces.Discrete(2)
        self.observation_space = spaces.Box(low=np.array([0,0,0]), high=np.array([1,1,1]), dtype=np.float16)
    
    
    def reset(self):
        pass    
    
    def step(self, action):
        pass
    
    def _next_observation(self):
        pass
    
    def _take_action(self, action):
        pass
        
    def render(self, mode = 'human', close=False):
        pass

from stable_baselines.common.vec_env import DummyVecEnv
from stable_baselines.common.policies import MlpPolicy
from stable_baselines2.ppo.ppo import PPO


envF = DummyVecEnv([lambda : envFru()])

model = PPOPolicy(envF, MlpPolicy, learning_rate= 0.001)
model.learn(total_timesteps=20000)

obs = env.reset()
for i in range(MAX_EPISODES):
    action, _states = model.predict(obs)
    obs, reward,done,info = env.step(action)
    #env.render()

我得到的回溯如下：

AttributeError                            Traceback (most recent call last)
<ipython-input-124-550b8c75c26b> in <module>
     12 envF = DummyVecEnv([lambda : envFruit()])
     13 
---> 14 model = PPOPolicy(envF, MlpPolicy, learning_rate= 0.001)
     15 model.learn(total_timesteps=20000)
     16 

~\Desktop\ImitationLearning\stable_baselines2\ppo\policies.py in __init__(self, observation_space, action_space, learning_rate, net_arch, activation_fn, adam_epsilon, ortho_init, log_std_init)
     29                  ortho_init=True, log_std_init=0.0):
     30         super(PPOPolicy, self).__init__(observation_space, action_space)
---> 31         self.obs_dim = self.observation_space.shape[0]
     32 
     33         # Default network architecture, from stable-baselines

AttributeError: 'DummyVecEnv' object has no attribute 'shape'

Answer 1

你确定这是你的实际代码吗？在上面的代码片段中，甚至没有定义名称PPOPolicy。我们需要查看PPOPolicy的代码。显然，它的构造函数（它的__init__方法）期望它的第一个参数是一个shapeartAttribute，所以我猜，它期望一个pandas数据帧。您的envF没有shape属性，因此会导致错误

从你片段中的名字判断，我想你应该写

model = PPOPolicy(
    envF.observation_space, 
    envF.action_space, 
    MlpPolicy, 
    learning_rate=0.001
)

在相关线路上

我的假设源于错误消息

super(PPOPolicy, self).init(observation_space, action_space)

告诉我们，PPOPolicy的构造函数将两个名为observation_space和action_space的变量传递给它的super()构造函数。由于这些名称会重新出现在您的环境中，我想这就是问题所在。但只要我们看不到正确完整的代码，这只是在迷雾中导航

也许它会帮助您了解如何读取错误消息。这可能会帮助您解决未来的问题。所以，我建议你读一些类似https://www.tutorialsteacher.com/python/error-types-in-python