How to use ddpg on the TU-Darmstadt Lichtenberg cluster

For general information about the cluster usage see hhlr.

After you have set up password-free login you can access your cluster files via

nautilus sftp://<username>@lcluster2.hrz.tu-darmstadt.de/home/<username> & exit

It might be convenient to create a bash function for this in the `~\.bashrc` on your machine.

```bash function clfiles { nautilus sftp://@lcluster2.hrz.tu-darmstadt.de/home/ & exit } ```

To use python on the cluster we need to load some modules first. Add the following to your ~\.bashrc on the cluster.

module load git gcc intel python/2
export PATH=$PATH:$HOME/.local/bin:$HOME/bin

Tensorflow

Use pip to install the CPU version of TensorFlow. Follow the instructions. Notice that we have to install python packages via pip install --user <package-name> because we do not have root access. TODO: Installing the GPU version of TensorFlow

OpenAI Gym

Install gym via pip install gym --user.

To record video we need ffmpeg. Download the static build from http://johnvansickle.com/ffmpeg/ and unpack to an arbitrary location on the cluster. Lastly, put a symlink to the ffmpeg binary in ~/.local/bin/.

We also need a virtual frame buffer to render the environments on the cluster. You can add the following to the ~\.bashrc.

killall Xvfb
Xvfb :1 -screen 0 1400x900x24 &
trap 'kill $(jobs -p)' EXIT
export DISPLAY=:1

To use the gym mujoco bindings, follow the instructions at https://github.com/openai/mujoco-py.

Usage

Example:

python run.py --outdir ../ddpg-results/experiment1 --env Reacher-v1

Enter python run.py -h to get a complete overview.

Submit a SLURM job via:

python run.py --outdir ../ddpg-results/experiment1 --env Reacher-v1 --job

Dashboard

Example:

python dashboard.py --exdir ../ddpg-results

Enter python dashboard.py -h to get a complete overview.

To use the visualization dashboard remotely you have to set up port forwards. You could add this to the `~\.bashrc` on your machine.

```bash function remote { xdg-open http://localhost:8007/tree/dashboard.ipynb & ssh -c arcfour @lcluster2.hrz.tu-darmstadt.de \ -L 8007:localhost:8007 \ -L 8008:localhost:8008 \ -L 8009:localhost:8009 \ 'python ~/ddpg/dashboard.py --nobrowser --exdir ' } ```