From this paperhere, it shows that U_Net initialized by VGG received a better result than the one trained from scratch. Now I want to build a custom u_net which has [32,64,128,256] which is different from vgg.features (I have tried to use U_Net with VGG architecture without pre-trained weights and custom U_Net, they all get the same Iou score .This means that the sample architecture suits for my data. ) How can I get the initialized weight from VGG? As far as I know, knowledge distillation could help me to get a simple net whose weight could be used for initialization for my u-net like ([32,64,128,256]). But knowledge distillation on ImageNet data set would cost lost of time, Is there another way to initialize the ‘smaller’ U-Net([32,64,128,256] )?
You can just copy specific layer weights and set others to random.