Open zethan1 opened 1 year ago
As with most fine-tuning, you need to just use the pre-trained model as initialization parameters on the downstream task. For this purpose, I recommend using a segmentation network that uses the same backbone, such as Swin-Unet, the same network I used for the downstream task in my paper. If you need more help, please feel free to contact me again.
How do you use the Swin-Unet?
Are you asking how to use Swin-Unet for transfer learning? The method is the same as the original Swin-Unet using the pre-trained model of Swin Transformer. The Swin-MAE trained model is used directly as the pre-trained weights for the encoder and bottleneck of Swin-Unet. And the swin Transformer blocks in the decoder use the same weights of the corresponding layers in the encoder. @DLoboT
How can I fine-tuning for the downstream tasks, such as segmentation? Could you share your code about fine-tuning?