search

guopengf / FL-MRCM

Multi-institutional Collaborations for Improving Deep Learning-based Magnetic Resonance Image Reconstruction Using Federated Learning
MIT License
45 stars 10 forks source link

Question about model archeticture. #6

Closed chenslcool closed 3 years ago

chenslcool commented 3 years ago

Thanks for your code, that's a nice job. You use U-Net as the reconstruction network. Howerver, skip-connections are important part of Unet. But it seems that you do not use skip connections. Maybe it will cause performance degradation? I wonder why skip connection is unused for your work. Thanks!

guopengf commented 3 years ago

Thanks for asking. Actually, we use skip-connections. The 'downsample_layer' is from the encoder part. This cat operation can be seen as 'skip-connection'. https://github.com/guopengf/FL-MRCM/blob/ff6c3f56e755b18c81f4a2b13488da9698fa762c/models/unet_model.py#L175

chenslcool commented 3 years ago

Sorry, I should be more careful. Besides, considering the skip-connection, the features of encoder include features output by other stages in encoder, have you ever consider align all these features (not just output of last stage of encoder)?

guopengf commented 3 years ago

Good point. I havn't tried to align all features. While it affects communication efficiency, the ability of dealing with domain shift might be improved.