Questions about deep supervision development

[Strawbili]

I'm adding a customized net to the nnunetv2 toolchain.
Unfortunately, I couldn't use the dynamic architecture in the nnunetv2 since I'm trying to come up with something new. Also, the debug is quite slow since there is so much data. Thus, it's quite hard to implement the deep supervision to the nnunet. Could you please give some hints on the development of the deep supervision with a customized net? What is the order of the output in the UNet?

[Strawbili]

It seems like the code works like this:
the shape of the output list is:

[B, num_classes, H, W] # and the H/W denotes the feature map size belongs to each stage

In the each stage of the decoder, there is a Conv with stride = 1 to make the feature channels to the num_classes