Question about SwinUNETR

Open chaoscls opened this issue 2 years ago • 1 comments

https://github.com/Project-MONAI/MONAI/blob/ab800d8413df5680161ea00fb3d6c1a7aa8dd04b/monai/networks/nets/swin_unetr.py#L304 Why isn't a encoder layer after hidden_states_out[3] like the other hidden outputs?

Mar 01 '23 12:03 chaoscls

Hi @chaoscls , thanks for the interests of the work. The hidden out 3 is a design option, the goal is to keep 4 times of downsampling to avoid the model to be too large and complex. You could freely reformat the hidden to 5 times downsamples and connect all hidden layers to skip connection and outputs. That will increase the # of parameter by a bit. From our observation, the difference is minimal. Thank you.

Mar 13 '23 15:03 tangy5