Open Pulse Graph Classifier

Warning

This is a WIP repository, it is not yet stable.

This tool is a graph classifier developped for the EPFL Open Pulse project. The graph is composed of nodes representing Github Organizations, Users and Repositories. The classifier aims to classify if a Repository nodes belongs to EPFL Github community or not.

Important modelling choices

During graph extraction for creating the Pygeometric HeteroData object, the indices of the nodes start at 1 (not zero).

Installation

Docker build (Linux/Windows)

Docker image building :

docker build . --platform linux/amd64 -t openpulse

Docker image run:

docker run -it openpulse

Inside docker:

for training

python open-pulse-graph-classifier/inference/api.py

for inference

python open-pulse-graph-classifier/train/main.py

Local dev

source .venv/bin/activate
uv pip install -r requirements.txt
python src/openpulse_graph_classifier/main.py

RunAI

Training

runai config project <your project>

runai training submit openpulse-training \
    -i ghcr.io/sdsc-ordes/open-pulse-graph-classifier-training:latest \
    --gpu 0.02 \
    --interactive \
    --attach

Inference

runai inference submit openpulse-inference \
  -p openpulse-laure \
  -i ghcr.io/sdsc-ordes/open-pulse-graph-classifier-inference:latest \
  --serving-port=8000 \
  --min-replicas=0 \
  --max-replicas=1 \
  --scale-down-delay-seconds 3600 \
  --image-pull-policy Always \
  -c -- python inference_api.py

(where openpulse is the job name, change according to your needs.)