Specific implementation steps for fine-tuning the SAM HQ model

[Swiftch]

Hello HQ-SAM Team,

Thank you very much for sharing your excellent work!

I am currently working on fine-tuning the SAM HQ model for my specific segmentation task. I would like to ask for some guidance on the concrete implementation steps for fine-tuning. Specifically:

How and where should I integrate LoRA layers into the model? Should these be added in the image encoder, mask decoder, or elsewhere?

Are there any specific training scripts or configurations you recommend for LoRA-based fine-tuning within your codebase?

Do you have any recommended hyperparameters or best practices to ensure effective fine-tuning and good generalization?

During inference, should the flag hq_token_only = True always be set, or only after certain training stages?

I have reviewed your train.py and the model definitions but would appreciate detailed pointers or examples for incorporating LoRA into the training workflow.

Thank you very much for your help!

Best regards,
Swiftch