Memory issue using Windows (yes I know it is not supported)

[eziolanza]

Hello everyone,
I'm using 32GB ram RTX 3090 Intel I7 CPU. I was able to train all 2d folds on a 180 cases dataset. Was previously able to train 3d_lowres on a small 30 cases dataset; however I am not able to train the 180 dataset using 3d_lowres , I get this error:

"OSserror winerror 8 not enough memory resources are available to process this command "

No matter how I increase the memory swap file this remains. Any workaround for this? Increase RAM? Any way to tweak the system so that It will work with the available hardware?

Thanks for your help

[ancestor-mithril]

You can set the environment variable nnUNet_n_proc_DA, which controls the number of Data Augmentation workers, to a lower value. The default is 8, I think that 4 is ok.
You can also set the environment variable nnUNet_def_n_proc (default: 8) to a lower value such as 4.

[eziolanza]

I tried with 0, checked Get-ChildItem env: and was correctly updated. Still states memory available is insufficient with OSErorr: WinError 8.
I also tried maxing up the swap memory, without luck.
Other ideas?

[ancestor-mithril]

Normally, I would expect nnUNet to work with 32 GB, that's why use python -c "import os; print(os.environ['nnUNet_n_proc_DA'])" to check if it is seen by the python intepreter.

If you use nnUNet_keep_files_open=1, remove it and use nnUNet_keep_files_open=0.
That's all I think of.

[eziolanza]

I'll try that. Curiously enough, the issue arises only when I train 3d_lowres. both 2d and 3d_fullres (unexpectedly) work fine.
Maybe a bug?

[ancestor-mithril]

It could be. If 3d_fullres works, I recommend using it as it's usually better than using 3d_lowres.