Different padding used in finetuning and pretraining

[chenzhekl]

Thanks for this great project!

I noticed that no padding was used for pre-training, while padding=2 was set for all fine-tuning tasks. IIUC, the image size used in fine-tuning (1024x768) is perfectly divisible by the patch size (16). Why did you choose to intentionally pad the input to make it not divisible by the patch size?

sapiens/seg/configs/sapiens_normal/normal_general/sapiens_1b_normal_general-1024x768.py

Line 67 in 08dce79

patch_cfg=dict(padding=2),