Hidden Biases of End-to-End Driving Models
Bernhard Jaeger, Kashyap Chitta, Andreas Geiger
International Conference on Computer Vision (ICCV), 2023This repo contains the code for the paper Hidden Biases of End-to-End Driving Models .
We provide clean, configurable code with documentation as well as pre-trained weights with strong performance.
The repository can serve as a good starting point for end-to-end autonomous driving research on CARLA.
- Setup
- Pre-Trained Models
- Evaluation
- Dataset
- Data generation
- Training
- Additional Documenation
- Citation
Clone the repo, setup CARLA 0.9.10.1, and build the conda environment:
git clone https://github.com/autonomousvision/carla_garage.git
cd carla_garage
chmod +x setup_carla.sh
./setup_carla.sh
conda env create -f environment.yml
conda activate garageBefore running the code, you will need to add the following paths to PYTHONPATH on your system:
export CARLA_ROOT=/path/to/CARLA/root
export WORK_DIR=/path/to/carla_garage
export PYTHONPATH=$PYTHONPATH:${CARLA_ROOT}/PythonAPI
export PYTHONPATH=$PYTHONPATH:${CARLA_ROOT}/PythonAPI/carla
export PYTHONPATH=$PYTHONPATH:$CARLA_ROOT/PythonAPI/carla/dist/carla-0.9.10-py3.7-linux-x86_64.egg
export SCENARIO_RUNNER_ROOT=${WORK_DIR}/scenario_runner
export LEADERBOARD_ROOT=${WORK_DIR}/leaderboard
export PYTHONPATH="${CARLA_ROOT}/PythonAPI/carla/":"${SCENARIO_RUNNER_ROOT}":"${LEADERBOARD_ROOT}":${PYTHONPATH}You can add this in your shell scripts or directly integrate it into your favorite IDE.
E.g. in PyCharm: Settings -> Project -> Python Interpreter -> Show all -> garage (need to add from existing conda environment first) -> Show Interpreter Paths -> add all the absolute paths above (without pythonpath).
We provide a set of pretrained models here.
The models are licensed under CC BY 4.0.
These are the final model weights used in the paper, the folder indicates the benchmark.
For the training and validation towns, we provide 3 models which correspond to 3 different training seeds. The format is approach_trainingsetting_seed. Each folder has an args.txt containing the training settings in text, a config.pickle containing all hyperparameters for the code and a model_0030.pth containing the model weights. Additionally, there are training logs for most models.
To evaluate a model, you need to start a CARLA server:
cd /path/to/CARLA/root
./CarlaUE4.sh -openglAfterward, run leaderboard_evaluator_local.py as the main python file. It is a modified version of the original leaderboard_evaluator.py which has the configurations used in the benchmarks we consider and additionally provides extra logging functionality.
Set the --agent-config option to a folder containing a config.pickle and model_0030.pth.
Set the --agent to sensor_agent.py.
The --routes option should be set to lav.xml or longest6.xml.
The --scenarios option should be set to eval_scenarios.json for both benchmarks.
Set --checkpoint to /path/to/results/result.json
To evaluate on a benchmark, set the respective environment variable: export BENCHMARK=lav or export BENCHMARK=longest6.
Set export SAVE_PATH=/path/to/results to save additional logs or visualizations
Models have inference options that can be set via environment variables.
For the longest6 model you need to set export UNCERTAINTY_THRESHOLD=0.33, for the LAV model export STOP_CONTROL=1 and for the leaderboard model export DIRECT=0.
Other options are correctly set by default.
For an example, you can check out local_evaluation.sh.
After running the evaluation, you need to parse the results file with result_parser.py. It will recompute the metrics (the initial once are incorrect) compute additional statistics and optionally visualize infractions as short video clips.
python ${WORK_DIR}/tools/result_parser.py --xml ${WORK_DIR}/leaderboard/data/lav.xml --results /path/to/results --log_dir /path/to/resultsThe result parser can optionally create short video/gif clips showcasing re-renderings of infractions that happened during evaluation. The code was developed by Luis Winckelmann and Alexander Braun.
To use this feature, you need to prepare some map files once (that are too large to upload to GitHub). For that, start a CARLA sever on your computer and run prepare_map_data.py.
Afterward, you can run the feature by using the --visualize_infractions flag in result_parser.py. The feature requires logs to be available in your results folder, so you need to set export SAVE_PATH=/path/to/results during evaluation.
The instructions above are what you will be using to debug the code. Actually evaluating challenging benchmarks such as longest6, that have over 108 long routes, is very slow in practice. Luckily, CARLA evaluations are embarrassingly parallel. Each of the 108 routes can be evaluated independently in parallel. That means if you have 2 GPUs you can evaluate 2x faster, if you have 108 GPUs you can evaluate 108x faster. While using the same amount of overall compute. To do that, you need access to a scalable cluster system and some scripts to parallelize. We are using SLURM at our institute. To evaluate a model, we are using the script evaluate_routes_slurm.py. It is intended to be run inside a tmux on an interactive node and will spawn evaluation jobs (up till the number set in max_num_jobs.txt). It also monitors the jobs and resubmits jobs where it detected a crash. In the end, the script will run the result parser to aggregate the results. If you are using a different system, you can use this as guidance and write your own script. The CARLA leaderboard benchmarks are the most challenging in the driving scene right now, but if you don't have access to multiple GPUs you might want to use simulators that are less compute intensive for your research. NuPlan is a good option, and our group also provides strong baselines for nuPlan.
We released the dataset we used to train our final models. The dataset is licensed under CC BY 4.0. You can download it using:
cd /path/to/carla_garage/tools
bash download_data.shThe script will download the data to /path/to/carla_garage/data. This is also the path you need to set --root_dir to for training. The script will download and unzip the data with 11 parallel processes. The download is roughly 350 GB large (will be a bit more after unzipping).
Dataset generation is similar to evaluation. You can generate a dataset by changing the --agent option to data_agent.py and the --track option to MAP. In addition, you need to set the following environment flags:
export DATAGEN=1
export BENCHMARK=collection
export CHECKPOINT_ENDPOINT=/path/to/dataset/Routes_{route}_Repetition{repetition}/Dataset_generation_{route}_Repetition{repetition}.json
export SAVE_PATH=/path/to/dataset/Routes_{route}_Repetition{repetition}Again it is too slow to generate our dataset with a single computer, you should be using multiple GPUs. We provide a python script for SLURM clusters, it works in the same fashion as the evaluation script. We will release the dataset we used at a later point.
Agents are trained via the file train.py.
Examples how to use it are provided for shell and SLURM.
You need to activate garage conda environment before running it.
It first sets the relevant environment variables and then launches the training with torchrun.
Torchrun is a pytorch tool that handles multi-gpu training.
If you want to debug on a single gpu simply set --nproc_per_node=1.
The training script has many options to configure your training you can list them with python train.py --help or look through the code.
The most important once are:
--id your_model_000 # Name of your experiment
--batch_size 32 # Batch size per GPU
--setting all # Which towns to withhold during training. Use 'all' for leaderboard, longest6 and '02_05_withheld' for LAV models.
--root_dir /path/to/dataset # Path to the root_dir of your dataset
--logdir /path/to/models # Root dir where the training files will be stored
--use_controller_input_prediction 1 # Whether your model trains with a classification + path prediction head
--use_wp_gru 0 # Whether you model trains with a waypoint head.
--use_discrete_command 1 # Whether to use the navigational command as input to the model
--use_tp 1 # Whether to use the target point as input to your model
--cpu_cores 20 # Total number of cpu cores on your machine
--num_repetitions 3 # How much data to train on (Options are 1,2,3). 1x corresponds to 185k in Table 5, 3x corresponds to 555kAdditionally, to do the two stage training from Table 4 you need the --continue_epoch and --load_file option.
You need to train twice. First train a model with set --use_wp_gru 0 and --use_controller_input_prediction 0, this will only train the perception backbone with auxiliary losses.
Then, train a second model, set e.g. --use_controller_input_prediction 1, --continue_epoch 0 and --load_file /path/to/stage1/model_0030.pth.
The load_file option is usually used to resume a crashed training, but with --continue_epoch 0 the training will start from scratch with the pre-trained weights used for initialization.
You can also run and debug torchrun in PyCharm. To do that you need to set your run/debug configuration as follows:
Set the script path to: /path/to/train.py
Set the interpreter options to:
-m torch.distributed.run --nnodes=1 --nproc_per_node=1 --max_restarts=0 --rdzv_id=123456780 --rdzv_backend=c10dTraining parameters should be set in the Parameters: field and environment variable in Environment Variables:.
Additionally, you need to set up conda environment (variables) as described above.
To submit to the CARLA leaderboard, you need docker installed on your system (as well as the nvidia-container-toolkit to test it). Create the folder team_code/model_ckpt/transfuser. Copy the model.pth files and config.pickle that you want to evaluate to team_code/model_ckpt/transfuser. If you want to evaluate an ensemble, simply copy multiple .pth files into the folder, the code will load all of them and ensemble the predictions.
Edit the environment paths at the top of tools/make_docker.sh and then:
cd tools
./make_docker.shThe script will create a docker image with the name transfuser-agent. Before submitting, you should locally test your image. To do that, start up a CARLA server on your computer (it will be able to communicate with the docker container via ports). Then start your docker container. An example is provided in run_docker.sh. Inside the docker container start your agent using:
cd leaderboard
cd scripts
bash run_evaluation.shYou can stop the evaluation, after confirming that there is no issue, using "ctrl + c + ".
To submit, follow the instructions on the leaderboard to make an account and install alpha.
alpha login
alpha benchmark:submit --split 3 transfuser-agent:latestThe command will upload the docker image to the cloud and evaluate it.
-
Coordinate systems in CARLA repositories are usually a big mess. In this project, we addressed this by changing all data into a unified coordinate frame. Further information about the coordinate system can be found here.
-
The TransFuser model family has grown quite a lot with different variants, which can be confusing for new community members. The history file explains the different versions and which paper you should cite to refer to them.
-
Building a full autonomous driving stack involves quite some engineering. The documentation explains some of the techniques and design philosophies we used in this project.
-
The codebase can run any experiment presented in the paper. It also supports some additional features that we did not end up using. They are documented here.
If you have any questions or suggestions, please feel free to open an issue or contact us at [email protected].
If you find CARLA garage useful, please consider giving us a star 🌟 and citing our paper with the following BibTeX entry.
@InProceedings{Jaeger2023ICCV,
title={Hidden Biases of End-to-End Driving Models},
author={Bernhard Jaeger and Kashyap Chitta and Andreas Geiger},
booktitle={Proc. of the IEEE International Conf. on Computer Vision (ICCV)},
year={2023}
}Open source code like this is build on the shoulders of many other open source repositories. In particularly we would like to thank the following repositories for their contributions:
We also thank the creators of the numerous pip libraries we use. Complex projects like this would not be feasible without your contribution.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent