Experimentation

Hyperparameter tuning & experimental setup

Available from academic-skills.org and/or this document (includes advice on writing as well).

Computational Resources

DSLab

DSLab is the internal compute cluster of LIACS. All students have access by default (to the student machines). For this course, we usually provide you additional access to one of the staff machines.

All instructions and an overview of available resources can be found through: https://rel.liacs.nl/labs/dslab
- You need to be inside the REL network to view the website, which can be achieved by:
  - Connecting to the LIACS VPN and setting up a proxy for you browser: https://rsewiki.liacs.nl/liacs_vpn
  - Connecting to Snellius wifi `wlan3' (password can be provided in the lecture)

Connect

To access the DSLab computers you:
- First connect to the REL network: 'ssh [ulcn_name]@ssh.liacs.nl' (or 'ssh [ulcn_name]@silver.liacs.nl')
- Then connect to one of the machines, e.g.: ssh duranium

Available machines

- Apart from the standard student machines, you should all have access to the tritanium staff machine (which has 16 NVIDIA Tesla K80 GPUs).

Memory management

- /home has a quotum, so you need to put your bigger packages (e.g., through conda) under /data
- Watch out: conda may actually still put downloaded packages in cache in your /home folder, causing a "Disk Quota Exceeded Error". The ways to check this, and solutions to this problem, are listed under https://rel.liacs.nl/labs/dslab#storage.

Mujoco installation

- Some papers will require Mujoco installation (https://mujoco.org/), which can be installed at DSLab as well.

ALICE

ALICE is the university-wide compute cluster. You can only access this with permission of a teacher, which we do provide for this course.

You can find general information about ALICE here.
Sign-up information can be found here. Please put the teacher email address in the CC of your email to request access to ALICE, so we can confirm your request.

Third-party

Another option is to use third-party compute resources. Several companies offer a free starting budget, which is usually sufficient for a course.