MlpPolicy is the only working policy for my custom environment

I tried making a custom Environment which I’m succeeding in training using DQN and A2C using the respective MlpPolicy.

env = FXTradingEnvironment(spot_rates, client_amounts, client_actions)
env = Monitor(env, filename=f'{log_dir}/FXTrading.log', allow_early_resets=True)
# The algorithms require a vectorized environment to run
env = DummyVecEnv([lambda: env])

model_dqn_mlp = DQN(MlpPolicy, env, verbose=0, tensorboard_log="./tmp/dqn_FX_tensorboard/")
model_dqn_cnn = DQN(CnnPolicy, env, verbose=0, tensorboard_log="./tmp/dqn_FX_tensorboard/")

And I get the following error:

model_dqn_cnn = DQN(CnnPolicy, env, verbose=0, tensorboard_log="./tmp/a2c_FX_tensorboard/")
Traceback (most recent call last):

  File "<ipython-input-69-cc2f20ccec35>", line 1, in <module>
    model_dqn_cnn = DQN(CnnPolicy, env, verbose=0, tensorboard_log="./tmp/a2c_FX_tensorboard/")

  File "C:\ProgramData\Anaconda3\lib\site-packages\stable_baselines\deepq\dqn.py", line 105, in __init__
    self.setup_model()

  File "C:\ProgramData\Anaconda3\lib\site-packages\stable_baselines\deepq\dqn.py", line 143, in setup_model
    double_q=self.double_q

  File "C:\ProgramData\Anaconda3\lib\site-packages\stable_baselines\deepq\build_graph.py", line 367, in build_train
    act_f, obs_phs = build_act(q_func, ob_space, ac_space, stochastic_ph, update_eps_ph, sess)

  File "C:\ProgramData\Anaconda3\lib\site-packages\stable_baselines\deepq\build_graph.py", line 141, in build_act
    policy = q_func(sess, ob_space, ac_space, 1, 1, None)

  File "C:\ProgramData\Anaconda3\lib\site-packages\stable_baselines\deepq\policies.py", line 176, in __init__
    layer_norm=False, **_kwargs)

  File "C:\ProgramData\Anaconda3\lib\site-packages\stable_baselines\deepq\policies.py", line 106, in __init__
    extracted_features = cnn_extractor(self.processed_obs, **kwargs)

  File "C:\ProgramData\Anaconda3\lib\site-packages\stable_baselines\common\policies.py", line 24, in nature_cnn
    layer_1 = activ(conv(scaled_images, 'c1', n_filters=32, filter_size=8, stride=4, init_scale=np.sqrt(2), **kwargs))

  File "C:\ProgramData\Anaconda3\lib\site-packages\stable_baselines\a2c\utils.py", line 133, in conv
    n_input = input_tensor.get_shape()[channel_ax].value

  File "C:\ProgramData\Anaconda3\lib\site-packages\tensorflow_core\python\framework\tensor_shape.py", line 870, in __getitem__
    return self._dims[key]

IndexError: list index out of range

Am I doing something wrong? Did I miss something from the docs?

[MentalGear]

Actually, found the issue for "You environment must inherit from gym.Env class cf " error. I used check_env after a (Monitor) env wrapper. Check_env must be used right after init the env and before other env augmentations.

Maybe the error message could be cleared up a bit to help future users quicker identify the issue, by returning: " Your environment must inherit directly from gym.Env class cf Note: Wrappers must be added after checking."

[araffin]

Would you mind telling us how and if you have fixed this bug ?

Your environment must derive from gym.Env class, this is not a bug. Here, the environment is implemented in c++ so this is a corner case.