shm limitation trigers bus error in nnUNetv2

Dear organisers,

I am using a modified version of the nnUNet_baseline, to fit nnUNetv2

I have noticed the docker settings on shared memory size (i.e.: --shm-size="128m") does not seem to allow nnUNetV2's prediction pipeline to run, raising a bus error (SIGBUS) on the preprocessing workers (even if there is only one worker) in my test environment

I am taking a guess that this is also a problem in the validation environment, as I have isolated this limitation as the only problem, as test.sh runs without problem once we set --shm-size="1g".

As a side note, setting the ipc="host" flag also allows the image to run properly

Is this setup locked, or can this limitation be lifted? Or am I simply missing something else?

Regards, Zhack

Wasn't aware of your forum post and ran into the same issue today, see here I found a workaround for this though. No secrets in my process.py so here it is. Feel free to adapt as needed. Since the Docker will always only be applied too one test case at a time (inefficient. meh.) there is no benefit from using nnU-Net's complicated interleaving of preprocessing, prediction and segmentation export.

Google Drive Link

Best and good luck!

Fabian

Thank you so much for your answer!

I'm glad I figured the problem correctly, I can see we went through the same (painful) debugging steps :')

Just to let you know in case this wasn't intended, the drive link requires to send an access request

Best of luck to you, and thanks again!

Zhack

Hey, the link should now work. Apologies! Best, Fabian