Keras-rl2 错误与 Tensorflow 的兼容性

Question

我目前在使用 keras-rl2 和 tensorflow 时遇到问题，我不知道为什么，我只是在互联网上搜索 keras-rl2、tensorflow 和 keras 文档，但没有找到解决方案。

目前，我想将keras-rl2与最新版本的tensorflow（2.16.1和keras 3）一起使用，但在使用时遇到了一些这样的错误

from rl.agents import DKQAgent

ModuleNotFoundError                       Traceback (most recent call last)
Cell In[37], line 1
----> 1 import rl.agents
      3 print("RL agents library version:", rl.agents.__version__)

File D:\Anaconda\Lib\site-packages\rl\agents\__init__.py:2
      1 from .dqn import DQNAgent, NAFAgent, ContinuousDQNAgent
----> 2 from .ddpg import DDPGAgent
      3 from .cem import CEMAgent
      4 from .sarsa import SarsaAgent, SARSAAgent

File D:\Anaconda\Lib\site-packages\rl\agents\dqn.py:8
      5 from tensorflow.keras.layers import Lambda, Input, Layer, Dense
      7 from rl.core import Agent
----> 8 from rl.policy import EpsGreedyQPolicy, GreedyQPolicy
      9 from rl.util import *
     12 def mean_q(y_true, y_pred):

File D:\Anaconda\Lib\site-packages\rl\core.py:8
      4 import numpy as np
      5 from tensorflow.keras.callbacks import History
      7 from rl.callbacks import (
----> 8     CallbackList,
      9     TestLogger,
     10     TrainEpisodeLogger,
     11     TrainIntervalLogger,
     12     Visualizer
     13 )
     16 class Agent:
     17     """Abstract base class for all implemented agents.
     18 
     19     Each agent interacts with the environment (as defined by the `Env` class) by first observing the
   (...)
     37         processor (`Processor` instance): See [Processor](#processor) for details.
     38     """

File D:\Anaconda\Lib\site-packages\rl\callbacks.py:12
      9 from tensorflow.python.keras.callbacks import Callback as KerasCallback, CallbackList as KerasCallbackList
     10 from tensorflow.python.keras.utils.generic_utils import Progbar
---> 12 class Callback(KerasCallback):
     13     def _set_env(self, env):
     14         self.env = env

ModuleNotFoundError: No module named 'keras.utils.generic_utils'

当我认为我只需要将其降级到某个版本（例如 2.13.0 和 keras 2.13.0）时，它仍然会出现这样的错误

---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
Cell In[18], line 1
----> 1 from rl.agents.dqn import DQNAgent

File D:\Anaconda\envs\AI\Lib\site-packages\rl\agents\__init__.py:1
----> 1 from .dqn import DQNAgent, NAFAgent, ContinuousDQNAgent
      2 from .ddpg import DDPGAgent
      3 from .cem import CEMAgent

File D:\Anaconda\envs\AI\Lib\site-packages\rl\agents\dqn.py:7
      4 from tensorflow.keras.models import Model
      5 from tensorflow.keras.layers import Lambda, Input, Layer, Dense
----> 7 from rl.core import Agent
      8 from rl.policy import EpsGreedyQPolicy, GreedyQPolicy
      9 from rl.util import *

File D:\Anaconda\envs\AI\Lib\site-packages\rl\core.py:7
      4 import numpy as np
      5 from tensorflow.keras.callbacks import History
----> 7 from rl.callbacks import (
      8     CallbackList,
      9     TestLogger,
     10     TrainEpisodeLogger,
     11     TrainIntervalLogger,
     12     Visualizer
     13 )
     16 class Agent:
     17     """Abstract base class for all implemented agents.
     18 
     19     Each agent interacts with the environment (as defined by the `Env` class) by first observing the
   (...)
     37         processor (`Processor` instance): See [Processor](#processor) for details.
     38     """

File D:\Anaconda\envs\AI\Lib\site-packages\rl\callbacks.py:8
      6 import numpy as np
      7 import tensorflow as tf
----> 8 from tensorflow.keras import __version__ as KERAS_VERSION
      9 from tensorflow.python.keras.callbacks import Callback as KerasCallback, CallbackList as KerasCallbackList
     10 from tensorflow.python.keras.utils.generic_utils import Progbar

ImportError: cannot import name '__version__' from 'tensorflow.keras' (D:\Anaconda\envs\AI\Lib\site-packages\keras\api\_v2\keras\__init__.py)

任何人都可以给我一个解释或解决方案，为什么它总是错误？

谢谢您的关心

Answer 1

显然我找到了 Tensorflow 的解决方案 <= 2.15.0 to use tke Keras-rl2, which is to add this code on your base code

import tensorflow as tf
tf.keras.__version__ = __version__

这样如果使用keras-rl2就不会出错，而对于现在2.16.0的较新的tensorflow我仍然不知道如何解决它