Seeking Guidance on Joint Training of CNN and GNN for Transductive Node Classification

[amalislam675]

I'm currently working on a transductive node classification task and would really appreciate your thoughts on a design challenge I’m facing involving the joint training of CNN and GNN models.
In my current setup:
I use a GCN trained on one feature source. As expected, I pass the entire adjacency matrix and the full node feature matrix (from this source) to the model, following full-batch training for node classification.
Now, I want to incorporate a second feature source which I intend to process using a CNN. The challenge here is:
The CNN model requires mini-batch training due to GPU memory limitations.
I need to clearly separate training and validation features for the CNN, while the GCN operates over the entire graph, which includes all nodes (training + validation + test).
Ultimately, I want to train both the GCN and CNN end-to-end, fuse their outputs (e.g., by concatenating or summing embeddings), and use the combined representation for final node classification.
My questions are:
Is it feasible to train both models jointly in an end-to-end manner given these constraints?
Would you suggest any specific architecture, training strategy, or resource that might help in aligning the GCN's full-batch paradigm with the CNN's mini-batch requirements?
Any suggestions or pointers would be immensely helpful