Conditional Imitation Learning at CARLA (Unofficial PyTorch Implementation)

This is an unofficial PyTorch implementation of End-to-end Driving via Conditional Imitation Learning. All credit to the original researchers.

Dataset

The dataset can be downloaded here. It is 24GB of HDF5 files. imitation_data.py is a custom Torch dataset class which handles and preprocesses the dataset.

Setup

For Training:

• Clone repo
• put dataset into data-and-checkpoints/imitation_data
• Run docker-compose file to build image and start container.
• From within the container, run train.py.

Training logs will output to host training logs folder for easy Tensorboard access :-)

Note, the Docker container requires Nvidia Docker runtime.

Agent:

• Specify network checkpoint in Agent class init method.
• May need to rewrite certain Carla imports.

.ckpt files for trained network can be downloaded here.

Future Work

• Implement branch-specific backpropagation.