Frequently Asked Questions (FAQ)
If you have your TigerGraph DB setup in a distributed fashion, you will only need to deploy GDPS on the leader’s node.
Yes, ML Workbench supports both CPU and GPUs for training a GNN.
ML Workbench is compatible with some of the most popular Graph ML frameworks such as PyTorch Geometric and DGL. Compatibility with TensorFlow will be available in the upcoming release.
No. Currently, the training will happen outside TigerGraph Database. Our team is actively working on bringing both GNN training and inference within TigerGraph
Yes, whether you have your TigerGraph DB setup with your own on-prem servers, or with a 3rd party Cloud such as AWS or GCP, ML Workbench will be compatible with these setups. One needs to deploy the GDPS component onto the same box as the TigerGraph DB, and ML Workbench will be able to communicate with your setup
This is an upcoming feature of ML Workbench. As for our preview release, ML Workbench is only compatible with TigerGraph on-prem solutions.
The data set used in your example notebooks forms a homogeneous graph? Can we train GNN using heterogeneous graph data?
This is an upcoming feature of ML Workbench, As for our preview release, ML Workbench supports homogeneous graphs only.
Can I leverage the existing algorithms in Graph Data Science Library to generate graph features that would be appropriate for GNN training?
Yes. You can directly leverage our existing graph algorithms, or you can customize your own feature and export them for GNN training. One caveat is that these features will need to be pre-populated prior to loading to your ML dev pipelines.