Correct way of finetuning tochvision deeplab model?

What is the correct way of finetuning a segmentation model??? How can I freeze just the backbone???