Inference stuck in "Generating preview images for step 1" on single niftii file

[szaccount]

When I run inference of the model on a single niftii file (on cpu only) it seems to be stuck indefinitely in step "Generating preview images for step 1".

Is it because its cpu only? I didn't see anything about it in the documentation.

Thanks in advance.

[NathanMolinier]

Hi @szaccount ! Thanks for testing our method ! If I understand correctly, it is running but it is taking an abnormal amount of time when you arrive at this step ?

It looks like it may be caused by the parallel processing. Could you rerun your code with the flag:

--max-workers 1

[szaccount]

@NathanMolinier Yes, but also unlike the previous steps, in the step it is stuck in, it doesn't progress at all (stays on 0%).
I tried your suggestion but it had the same result, of no progress. Could it be that for large files, GPU option has to be activated in order to process the information?

[NathanMolinier]

That's strange indeed, personally I mostly use the GPU but it should also work with CPU. Let me try to reproduce the error. In the meantime, maybe you can try the integration of TotalSpineSeg inside the spinalcordtoolbox.

With the function sct_deepseg -task totalspineseg

[Volkopat]

+1
I have tested this with various studies on GPU mode, however while testing CPU, its always stuck when trying to generate previews.

[NathanMolinier]

I'm not able to reproduce the error unfortunately, what operating system are you using ? I tried on a linux machine and on my personal macos laptop and the inference worked in both cases with only cpu torch installation. I tried also with both python 3.9 and 3.10. Can you provide more information about your experiments ?

[szaccount]

I am using Ubuntu 22.04 and python 3.10.12. Also, tried to run the line specified to install pytorch ("python3 -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118 --upgrade") and now when tried to run inference, get the error:

(in earlier stage then before - Generating preview images for input)

[NathanMolinier]

This issue is actually raised by nnunet due to an error with the last pytorch 2.6 update. I will update the readme to account for this error. Could you reinstall pytorch with this command:

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu118 --upgrade

I updated the command

[szaccount]

Hey @NathanMolinier, indeed this solved the last bug but now it returned to the original problem of it being stuck indefinitely in "Generating preview images for step 1:" stage. Perhaps it is an issue of file size? Do you have an example for the input file you used that I will be able to experiment as well? Thanks.