Parameter initialization

[guangqian-guo]

Hi, authors,
The parameters of the image encoder and prompt encoder are fixed during training, how are their parameters initialized if I start training from scratch?
According to my understanding, they should be initialized with the parameters of SAM, but I didn't find this part of the code.
Could you help me with this?

[robustsam]

You can use the provided RobustSAM checkpoints for initialization, as those checkpoints retain the original SAM image encoder and prompt encoder parameters.

[ljmkgbkq]

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?