I have a scene graph of an image in the form of the triplet. For example, for image id "213," we have a scene graph triplet as "[man holds cup], [woman drives car]...". I want to create embeddings or features of scene graphs for each image ID using Graph Neural Networks similar to image embedding generation, where I have an image ID and its corresponding image tensors. I have limited experience in GNN compared to vision, so I would like to know what can be the steps for GNN embedding creation and how to consider a number of nodes and features for GNN graph creation. Also, is there a similar thing like pre-trained GNN identical to vision. In vision, we can extract image features from a pre-trained image model.
I tried to create graphs for each scene graph triplets (using networkx) and create a two-layer GNN, but I am not sure how to proceed further. Should I train the GNN model, and how to select a number of nodes and features as a parameter in the GNN model is blocking me?