General-purpose pre-trained adapters

[nawazia]

Hi,

I'm trying to inference your pre-trained Med-SA on the external SAM-Med2D-20M dataset, which covers multiple different datasets and modalities. However, pre-trained weights in the adapter zoo are trained for specific tasks.

Is there any way to implement a more generalisable Med-SA, with a single set of adapters/an optimal way to gate inputs and load the best-suited pre-trained adapters from the model zoo?