I’m new to DGL and GNN’s and had a question about using BERT embeddings as features in the GNN:
I want to build a graph that uses BERT embeddings as features for the nodes. I understand that I can pre-compute these and assign it as a feature to the node, as long as all the nodes have this feature. As the GNN is trained (for example for a classification task), these initial embeddings will change. Is it possible to link this change back to fine-tune the BERT model through back-prop? If so, how?