Error during preprocessing with latest version (2.6.4), fixed by downgrading

[garance-thv]

Hello Fabian and everyone else involved in this project,

I've been wanting to give nnUNet a try in the context of my PhD for a while, ever since I followed a workshop on it a while back. I'll pass you the details but I followed the instructions rigourously and was stuck on the same error each time during the preprocessing step.

I ran this command :
nnUNetv2_plan_and_preprocess -d 005 --verify_dataset_integrity -np 1

I tested a bunch of different datasets since I thought it originated from my data, this one is the prostate task from the Medical Segmentation Decathlon, which I converted with the built-in method nnUNetv2_convert_MSD_dataset. My python version is 3.12.3 and I got the latest stable version of pytorch for windows (2.10.0) with CUDA 12.6.

And everytime I got the following error :

File "C:\Users\XXX\venv\Lib\site-packages\nnunetv2\preprocessing\preprocessors\default_preprocessor.py", line 412, in run
    _ = [r[i].get() for i in done]
         ^^^^^^^^^^
  File "C:\Users\XXX\Python\Python312\Lib\multiprocessing\pool.py", line 774, in get
    raise self._value
RuntimeError: Error while creating the NDArray

The only thing that fixed it for me was to downgrade my version of nnUNet (I chose the 2.6.0, but it could also work with more recent versions that aren't the 2.6.4, I didn't try).
Am I the only one with this issue on the latest version and is there an explanation for it? I didn't find anything related on here.

Thanks a lot!

[asika-lili]

I also encountered similar problems during training. There are some compatibility issues with the new version. The method of downgrading to solve the problem is very effective.

[FabianIsensee]

Hey thanks for reporting that. Have you been able to narrow this down further? Was this issue on the master or the latest release?