something wrong when running metadrive

[xc4xpzhf]

File "zoo/metadrive/config/metadrive_sampled_efficientzero_config.py", line 103, in
train_muzero([main_config, create_config], seed=0, max_env_step=max_env_step)
File "/root/autodl-tmp/LightZero/lzero/entry/train_muzero.py", line 154, in train_muzero
stop, reward = evaluator.eval(learner.save_checkpoint, learner.train_iter, collector.envstep)
File "/root/autodl-tmp/LightZero/lzero/worker/muzero_evaluator.py", line 287, in eval
policy_output = self._policy.forward(stack_obs, action_mask, to_play, ready_env_id=ready_env_id)
File "/root/autodl-tmp/LightZero/lzero/policy/sampled_efficientzero.py", line 963, in _forward_eval
network_output = self._eval_model.initial_inference(data)
File "/root/autodl-tmp/LightZero/lzero/model/sampled_efficientzero_model.py", line 270, in initial_inference
latent_state = self._representation(obs)
File "/root/autodl-tmp/LightZero/lzero/model/sampled_efficientzero_model.py", line 330, in _representation
latent_state = self.representation_network(observation)
File "/root/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/root/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
return forward_call(*args, **kwargs)
File "/root/autodl-tmp/LightZero/lzero/model/common.py", line 646, in forward
x = self.downsample_net(x)
File "/root/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/root/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
return forward_call(*args, **kwargs)
File "/root/autodl-tmp/LightZero/lzero/model/common.py", line 336, in forward
x = self.conv1(x)
File "/root/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/root/miniconda3/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1562, in _call_impl
return forward_call(*args, **kwargs)
File "/root/miniconda3/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 458, in forward
return self._conv_forward(input, self.weight, self.bias)
File "/root/miniconda3/lib/python3.8/site-packages/torch/nn/modules/conv.py", line 454, in _conv_forward
return F.conv2d(input, weight, bias, self.stride,
RuntimeError: Given groups=1, weight of size [32, 5, 3, 3], expected input[3, 84, 84, 5] to have 5 channels, but got 84 channels instead

there‘s something wrong when run metadrive_sampled_efficientzero_config.py

[puyuan1996]

Your MetaDrive version probably returns observations in (H, W, C) (NHWC) format, but the network expects (C, H, W) (NCHW). This channel-order mismatch triggers the convolution dimension error. A probable fix solution:

import gym
import numpy as np

class ChannelLast2First(gym.ObservationWrapper):
    def observation(self, obs):
        # (H, W, C) → (C, H, W)
        if obs.ndim == 3:
            obs = np.transpose(obs, (2, 0, 1))
        # (stack, H, W, C) → (stack, C, H, W)
        elif obs.ndim == 4:
            obs = np.transpose(obs, (0, 3, 1, 2))
        return obs.copy()          # avoid negative stride issues

# Wrap the MetaDrive environment immediately after creation
env = ChannelLast2First(env)

Verify that the observation shape is now (B, 5, 84, 84) before passing it to the network; training should then run normally.

[xc4xpzhf]

An error occurred in the later stages of training due to the input changing from the normal shape (n, 1, h, w, c) to (n, 6, h, w, c).

INFO Episode ended! Reason: out_of_road. drive_env.py:214
Observation list shapes: [(6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5), (6, 84, 84, 5)]
First observation type: <class 'numpy.ndarray'>

and sometime There is a mixture of array and list
Observation 185: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 186: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 187: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 188: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 189: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 190: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 191: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 192: shape=(1, 5, 84, 84), type=<class 'list'>
Observation 193: shape=(1, 5, 84, 84), type=<class 'list'>
Observation 194: shape=(1, 5, 84, 84), type=<class 'list'>
Observation 195: shape=(1, 5, 84, 84), type=<class 'list'>
Observation 196: shape=(1, 5, 84, 84), type=<class 'list'>
Observation 197: shape=(1, 5, 84, 84), type=<class 'list'>
Observation 198: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 199: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 200: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 201: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 202: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 203: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 204: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 205: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 206: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 207: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 208: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>
Observation 209: shape=(1, 84, 84, 5), type=<class 'numpy.ndarray'>

[puyuan1996]

Hello,

Thank you for the detailed feedback.

Could you please confirm if you encountered this error right after a fresh installation of the MetaDrive environment and then running the zoo/metadrive/config/metadrive_sampled_efficientzero_config.py script? Have you made any modifications to the related code within the env or lightzero directories?

Regarding the (6, 84, 84, 5) shape, the '6' is likely expected, as it should correspond to unroll_step + 1. This is because EfficientZero needs to sample a sequence of observations during training.

As for the mix of data shapes (numpy.ndarray vs. list), this might be caused by a mismatch between the padding logic in the replay buffer and the observation format provided by the current MetaDrive environment.

To help pinpoint the exact location of the issue, you could check the data type at these key points:

1. After obs, reward, done, info = env.step(action).
2. Before the data is stored in the replay buffer.
3. After data is sampled from the replay buffer.

We will also run some tests locally to investigate this. Thank you for your patience