Mask decoder finetuning about mobilesam HOT 5 CLOSED

@thangngoc89
were you able to finetune a MobileSAM using this? https://github.com/luca-medeiros/lightning-sam
do you happen to have a code snippet how to do it?

from mobilesam.

This is because of incorrectly normalize image during training and inteference. Closing this

from mobilesam.

This is because of incorrectly normalize image during training and inteference. Closing this

hello, I have the same problem as you, i.e., I frozen the image encoder and prompt encoder, then fine-tuned the mask decoder. However, I found that the total loss did not decrease with the epoch crease, and the accuracy fluctuated. You said the reason is because of incorrectly normalize image during training and inteference, and I use the original image normalize of SAM. Could you give more detailed information. Thank you very much

from mobilesam.

@wm-Githuber actually in my training setting, the accuracy increased but the original issue is that difference normalization scheme when using MobileSAM. You actually need to not apply normalization.

from mobilesam.

@wm-Githuber actually in my training setting, the accuracy increased but the original issue is that difference normalization scheme when using MobileSAM. You actually need to not apply normalization.
Thank you for your reply 👍

from mobilesam.