Stable Baselines3运行RL自动驾驶代码报numpy负步长错误求助

问题基本信息 #

• 开发场景：编写基于强化学习的自动驾驶代码，技术栈为Stable Baselines3框架 + OpenAI Gym环境，运行载体为Jupyter Notebook
• 运行环境：Windows 10系统、Anaconda管理的Python 3.8.13环境、CPU版本PyTorch

触发报错的测试代码 #

# Testing our model
episodes = 5 # test the environment 5 times
for episodes in range(1,episodes+1): # looping through each episodes
    bs = env.reset() # observation space
    # Taking the obs and passing it through our model
    # tells that which kind of the action is best for our work
    done = False 
    score = 0
    while not done:
        env.render()
        action, _ = model.predict(obs) # now using model here # returns model action and next state
        # take that action to get the best reward
        # for observation space we get the box environment
        # rather than getting random action we are using model.predict(obs) on our obs for an curr env to gen the action inorder to get best possible reward
        obs, reward, done, info = env.step(action)  # gies state, reward whose value is 1
        # reward is 1 for every step including the termination step
        score += reward
    print('Episode:{},Score:{}'.format(episodes,score))
env.close()

报错信息 #

ValueError: At least one stride in the given numpy array is negative, and tensors with negative strides are not currently supported

问题根因 #

• 代码存在笔误：env.reset()返回的初始观测值被赋值给了bs变量，但首次调用model.predict(obs)时obs未被正确初始化，会导致传入模型的数组内存布局异常
• Gym环境返回的观测numpy数组存在负步长，PyTorch CPU版本目前不支持直接将带负步长的numpy数组转换为张量，Stable Baselines3内部做数据格式转换时就会触发该报错
• 额外逻辑问题：循环计数变量和总测试轮次变量重名，均命名为episodes，循环执行时会覆盖总轮次参数，容易引发逻辑异常

修复方案 #

• 修正变量命名错误：将env.reset()的返回值直接赋值给obs，同时将单轮循环变量重命名为ep，避免变量覆盖
• 对传入模型的观测数组做内存连续化处理，使用np.ascontiguousarray()消除负步长，适配PyTorch张量转换要求
• 修复后的可运行代码如下：

import numpy as np

# 模型测试逻辑
episodes = 5
for ep in range(1, episodes + 1):
    obs = env.reset()  # 修正：初始观测直接赋值给obs
    done = False
    score = 0
    while not done:
        env.render()
        # 核心修复：将观测转为内存连续数组，解决负步长问题
        action, _ = model.predict(np.ascontiguousarray(obs))
        obs, reward, done, info = env.step(action)
        score += reward
    print(f'Episode:{ep}, Score:{score}')
env.close()

补充说明：如果上述修改后仍触发同类报错，可以在自定义Gym环境的reset()、step()方法返回观测值前，统一对观测数组执行np.ascontiguousarray(obs)处理，从数据源层面保证数组内存格式符合要求。

内容的提问来源于stack exchange，提问作者Shaina Mehta