Hydranet - separate prediction heads per class

[GeorgePearse]

Describe the feature

Motivation
I'd love to be able to train an instance segmentation and semantic segmentation model with the same backbone.

Or train a model with the same backbone, but a prediction head per class, so that I could update their weights independently. I guess I could already achieve this by doing something like freezing the backbone, exporting to ONNX, splitting the model appropriately, the same backbone repeatedly, but different heads as needed.

Do any models in MMDetection already make it easy to get started with this?

Any guidance at all would be very helpful.

[GeorgePearse]

Or more precisely, how could I go about adding a classification head, to the SwinTransformer Instance segmentation model?

[GeorgePearse]

Seeing some promising bits for multiple models in extract_model.py of MMDeploy