SAM and RobustSAM Parameter initialization

[ljmkgbkq]

Hello, author. This is an excellent piece of work, and I appreciate you making the code open source.

I noticed that the SAM weights are integrated with the new components of RobustSAM (as seen in the robustsam_checkpoint_l.pth file). I have a question regarding whether it is possible to initialize the original SAM parts with the pre-trained SAM weights while training the new parts of RobustSAM from scratch. My downstream task involves a non-visible light scenario, which differs significantly from the conditions of the pre-trained weights. Therefore, I would like to load SAM's pre-trained weights and simultaneously train the new components of RobustSAM from scratch.

Is this feasible with the current version of the code?