This is a quick tutorial on how to run Graph-MLP on the BwUniCluster2.0 service.
BwUniCluster uses Python 3.6 by default. Pytorch requires at least Python 3.7. Load Python 3.8 by running
module load devel/python/3.8.6_gnu_10.2IMPORTANT: The python version is not persisted! Every time you reconnect you need to run the command above BEFORE activating the virtual environment!
First, create and activate a virtual environment using
python -m venv ./venv
source venv/bin/activateNext, install PyTorch and PyTorch-Geometric (PyG).
Install the correct PyTorch version for your system from the official website. Then install the corresponding PyG version (correct PyTorch version and same CUDA version) from here.
After this is done install the remaining requirements from the requirements.txt by running:
pip install -r requirements.txtOur implementation uses Weights & Biases to easily track your runs.
To change the entity (personal or team account) and the project, change the variables at the top of train.py.
To disable logging in a run, add the --no-wandb flag.
To track runs you have to log in to WandB on your device. To do this activate the virtual environment and run
wandb login
Create a script or use an existing one, and define your SBATCH variables at the very top (see also ../run-scriptsrun.sh). See the documentation for more information. To queue a job, simply run
sbatch run.shIt is required that you define the queue you want to use (this is done in the run.sh script directly).
For all possible queues and their minimum, default and maximum resources, see here.
If you do not comply with the resource limitations, the job will not be accepted!