[Errno 2] No such file or directory: 'data/carla_leaderboad2_v8/results/data/sensor_data/data'

[QVTNU7774]

Dear authors of LEAD,

First of all thanks so much for open-source such great work that really improve the community!

We are currently trying out your code. However, we encounter an issue when running postrain:

[Errno 2] No such file or directory: 'data/carla_leaderboad2_v8/results/data/sensor_data/data'

Do we need to run pretrain first before we can run posttrain? Or could we use your pretained chkpt to run posttrain directly?

Thanks so much for your valuable time!

[ln2697]

Hi @QVTNU7774,

Just to clarify, there are two different things that can be confusing here:

1. Pre-trained released checkpoints (HuggingFace models)
   The checkpoints we provide on HuggingFace are already fully trained models.
   They are meant for evaluation/inference only, not for resuming post-training.
   If your goal is only to run closed-loop evaluation, you can directly use those checkpoints and do not need to run any training.

2. Two-stage training pipeline (perception pretrain →planning posttrain)
   If you want to train the model yourself, the pipeline is:

• Stage 1: Pre-training (perception backbone + basic fusion)
• Stage 2: Post-training (adds the planning transformer on top)

You can find the instructions here.

If you follow the instructions in README strictly but still encounter issues, please also provide your training commands. I would be glad to help.

Best,

Long

[EllingtonKirby]

Hello,

I have the same issue when running

python lead/leaderboard_wrapper.py \
  --checkpoint outputs/checkpoints/visiononly_resnet34 \
  --routes data/benchmark_routes/Town13/0.xml \
  --bench2drive

I would add that the B2D routes specified in the README are not included in the project, so I adjusted to python command to point to a route which does exist. Not sure if this causes the problem.

lead/training/config_training.py line 215 seems to point to a different Carla root location than the one specified in the setup:

    # Root directory for CARLA sensor data.
    carla_root = "data/carla_leaderboard2"

Should we be changing this line?

[EllingtonKirby]

I think the issue comes from trying to use the camera only model, which does not seem to have a corresponding target_dataset

I updated it to something like:

@property
def target_dataset(self):
    if "expert_debug" in self.carla_root:
        return TargetDataset.CARLA_LEADERBOARD2_3CAMERAS
    elif "carla_leaderboad2_v3" in self.carla_root:
        return TargetDataset.CARLA_LEADERBOARD2_3CAMERAS
    elif "carla_leaderboard2" in self.carla_root:
        return TargetDataset.CARLA_LEADERBOARD2_3CAMERAS
    elif "carla_leaderboad2_v8" in self.carla_root:
        return TargetDataset.CARLA_LEADERBOARD2_3CAMERAS
    elif self.use_waymo_e2e_data and not self.mixed_data_training:
        return TargetDataset.WAYMO_E2E_2025_3CAMERAS
    elif self.use_navsim_data and not self.use_carla_data:
        return TargetDataset.NAVSIM_4CAMERAS
    elif "carla_leaderboad2_v10" in self.carla_root:
        return TargetDataset.CARLA_LEADERBOARD2_6CAMERAS
    raise ValueError(
        f"Unknown CARLA root path: {self.carla_root}. Please register it in the config."
    )

And it fixed my issue.

@QVTNU7774 were you also trying to use the camera only config and checkpoint?

[ln2697]

I would add that the B2D routes specified in the README are not included in the project, so I adjusted to python command to point to a route which does exist. Not sure if this causes the problem.

No, changing the route should be absolutely fine.

I think the issue comes from trying to use the camera only model, which does not seem to have a corresponding target_dataset
I updated it to something like:

That is a good fix. Thanks, I will update the code.

lead/training/config_training.py line 215 seems to point to a different Carla root location than the one specified in the setup:

For context:

# Root directory for CARLA sensor data.
carla_root = "data/carla_leaderboard2"

This carla_root here is the default value and will be used during training. The training pipeline will stored carla_root in the serialized json config file. At inference time, the json file is loaded again and carla_root is restored with the saved value from training time.

It seems to me like the inference code does not work out of the box for dataset v10. After coming back from vacation I will update the code so every provided checkpoints can be used without issue.

Best,

Long

[QVTNU7774]

Hello,

I have the same issue when running

python lead/leaderboard_wrapper.py \
  --checkpoint outputs/checkpoints/visiononly_resnet34 \
  --routes data/benchmark_routes/Town13/0.xml \
  --bench2drive

I would add that the B2D routes specified in the README are not included in the project, so I adjusted to python command to point to a route which does exist. Not sure if this causes the problem.

lead/training/config_training.py line 215 seems to point to a different Carla root location than the one specified in the setup:

    # Root directory for CARLA sensor data.
    carla_root = "data/carla_leaderboard2"

Should we be changing this line?

I update corresponding path in config.json in the ckpt folder (e.g., tfv6/ tfv6_regnety032), then the problem solved