Intervention launch scripts

These scripts are intended to help use the Intervention code on University of Tartu's Slurm cluster, or on a single (multi-)GPU machine.

• ./slurm: Slurm helper scripts
• ./neuron: single-machine helper scripts

Example Slurm usage

This collects episodes of an expert teacher driving. It can be run multiple times in parallel---either across nodes or on the same node. It merges the collected data into a specifiable dataset.

E.g., to run a single job:

$ cd ./slurm
$ sbatch job-collect-teacher-examples.sh

Or to run multiple jobs:

$ cd ./slurm
$ for i in {1..10}; do
>     sbatch job-collect-teacher-examples.sh
> done

Environment variables:

• NUMBER_OF_EPSIODES configures the number of episodes desired
• INTERVENTION_DATASET_DIRECTORY configures the directory the data should be synchronized to

Example single-machine usage

This collects episodes of driving. It can parallelize over multiple GPUs on a single machine.

Configure the collection you want to run:

$ cd ./neuron
$ $EDITOR ./config.py

Then start collection:

$ ./parallel-collect.py

Setup

The scripts assume Conda is installed and available in the current environment. Further, the scripts assume the existence of a conda environment called intervention, providing some dependencies as well as making the intervention learning executable available.

To set this up, run, e.g.:

$ conda env create -f environment.yml
$ conda activate intervention
$ cd path/to/main/intervention/repository
$ python3 -m pip install -r requirements.txt
$ easy_install path/to/carla/PythonAPI/carla/dist/carla-*-py3.7-linux-x86_64.egg

Also see the intervention learning repository for specific installation instructions.

Change exported variables in ./slurm/prepare-carla-env.sh and/or ./neruron/prepare-carla-env.sh to point to a Carla directory.