This is an unofficial PyTorch implementation of End-to-end Driving via Conditional Imitation Learning. All credit to the original researchers.
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.
- 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.
- Specify network checkpoint in Agent class init method.
- May need to rewrite certain Carla imports.
.ckpt files for trained network can be downloaded here.
- Implement branch-specific backpropagation.