Use SAM_Perceptual(train=True, weights=[0,1,1,0]) as training loss instead of IMPACTReg

Add finetune models