Skip to content

Setup

Requirements for training the GNN are a machine lots of CPUs, GPUs, and RAM. (Improving the training efficiency of the GNN is WIP). You can either use the prebuilt docker containers, or create your own conda environment.

Start by cloning the repo. If you plan to contribute to the repo, you should work from a fork. 

git clone ssh://git@gitlab.cern.ch:7999/atlas-flavor-tagging-tools/algorithms/GNNJetTagger.git
cd GNNJetTagger

Prebuilt Image#

Prebuilt docker images are the recommended way to work with the GNN. You can run the prebuilt docker images using singulartiy. It is recommended to set SINGULARITY_CACHEDIR in your ~/.bashrc to make sure images are pulled to a directory with enough free space.

export SINGULARITY_CACHEDIR=<some path>/.singularity/

Next, you can pull the image:

singularity pull --docker-login \
    $SINGULARITY_CACHEDIR/gnn_image.simg \
    docker://gitlab-registry.cern.ch/atlas-flavor-tagging-tools/algorithms/gnnjettagger:latest

As the repository is internal at the moment, you'll need to enter your CERN credientials when pulling.

Once you've pulled the docker image locally, you can run it with:

singularity exec -ce --nv --bind $PWD \
    --env-file ./setup/singularity_envs.sh \
    $SINGULARITY_CACHEDIR/gnn_image.simg bash
In order to mount a directory to the image when running singularity exec, use the --bind <path> argument

Within the image, run the setup script. 

python setup.py develop

Running on a system without admistrative access

If you are running on a system without admistrative access, you may encounter 

error: can't create or remove files in install directory
In this case, add the --user flag when running the setup script: 
python setup.py develop --user

Conda#

The conda environment requires packages not able to be installed on MacOS.

If you are running on MacOS, you can use the singularity container.

To get started, clone the repo 

git clone ssh://git@gitlab.cern.ch:7999/atlas-flavor-tagging-tools/algorithms/GNNJetTagger.git
cd GNNJetTagger

Source the conda setup script (which will actually install mamba) 

source ./setup/setup_conda.sh
This script will install mamba, and also create the gnn environment from the setup/environment.yml file.

Make sure you active the conda environent after it is created. 

conda activate gnn

Next, install the gnn_tagger package. 

python setup.py develop