search

mazurowski-lab / finetune-SAM

This is an official repo for fine-tuning SAM to customized medical images.
https://arxiv.org/abs/2404.09957
Apache License 2.0
114 stars 17 forks source link

How much training data is generally required to fine-tune the model #8

Closed happyday521 closed 3 months ago

happyday521 commented 4 months ago

Hi，

I want to fine-tune the SAM mask decoder to only a single labeled dataset unsing supervised training (i.e., Setup 1 in your paper). Since I have no experience in this field, I would like to ask how much training data is generally required to fine-tune the model in this setting?

Thanks，look forward to your reply！

haoyudong-97 commented 3 months ago

It depends on the dataset, but my educational guess is at least 100 cases if you only want to fine-tune the decoder. In the meantime, I suggest fine-tuning both the encoder and decoder since this strategy is almost always better than only fine-tuning the decoder.

happyday521 commented 3 months ago

Get it. Thanks very much!