Guide for Training DreamerV2 on CarDreaner

This guide assumes you have installed car_dreamer. If not, please follow the instructions in the main README.

First, install the required dependencies for DreamerV2:

cd dreamerv2
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit
pip install -r requirements.txt

Set up CARLA and environment variables:

export CUDNN_PATH=$(dirname $(python -c "import nvidia.cudnn;print(nvidia.cudnn.__file__)"))
export CUSOLVER_PATH=$(dirname $(python -c "import nvidia.cusolver;print(nvidia.cusolver.__file__)"))
export LD_LIBRARY_PATH=$CUDNN_PATH/lib:$CUSOLVER_PATH/lib:$CONDA_PREFIX/lib:$LD_LIBRARY_PATH

Training

Execute the training script with desired configurations:

cd ..
bash train_dm2.sh 2000 0 --task carla_four_lane --dreamerv2.logdir ./logdir/carla_four_lane

2000 is the port number of the CARLA server. The script will automatically start the server so you don't need to start it manually. 0 is the GPU number. --task is the name of the task and --dreamerv2.logdir is the directory to save the training logs. For a complete list of tasks and their configurations, please refer to the documentation.

Visualization

Online data monitoring can be accessed on website on http://localhost:9000/, where the port number should be changed to <carla-port> + 7000 if you don't use the default port number 2000 for CARLA server.

Offline data logging can be accessed through TensorBoard:

tensorboard --logdir ./logdir/carla_four_lane

Go to http://localhost:6006/ in your browser to see the output.