modify nnunet

[mariaconcettaveraldi]

Hello!
I would like to know how can i modify nnunet to create a single channel architecture with a linear activation function.

Thank you,
Mariaconcetta

[pzaffino]

Hi all,
I have exactly the same problem.
I would have a custom, 3D, single-channel, nnUNet that gives as an output a 3D volume with a linear activation function (I do not want a segmentation).
Any suggestion on how to implement this?

Best,
Paolo

[FabianIsensee]

Hey, nnU-Net is build for segmentation. It cannot do other things. If all you need is the same UNet then please take a look at this function:

nnUNet/nnunetv2/training/nnUNetTrainer/nnUNetTrainer.py

Line 258 in 1635a5e

def build_network_architecture(plans_manager: PlansManager,

Best,
Fabian

[pzaffino]

Thank you @FabianIsensee
I understood.

[ibro45]

Hi @pzaffino @mariaconcettaveraldi have you had any luck with it? I can't seem to successfully train a custom model #1479.

[pzaffino]

Hi,
no, I just switched to another architecture.
I am interested in regression rather than segmentation, and nnunet is specifically built for segmentation.