search

xinghaochen / TinySAM

Official PyTorch implementation of "TinySAM: Pushing the Envelope for Efficient Segment Anything Model"
Apache License 2.0
403 stars 23 forks source link

Is the mask decoder weight inherited from the teacher models' decoder? #18

Open Vickeyhw opened 9 months ago

Vickeyhw commented 9 months ago

If so, in the full-stage knowledge distillation, the image encoder is randomly initialized, is the mask decoder finetuned at a smaller learning rate than the light weight image encoder? Is this consistent with your implementation?

shuh15 commented 9 months ago

Yes, the weights for mask decoder are inherited from the teacher, and we use a smaller learning rate for mask decoder compared to image encoder.