Issue Installing NiftyPET on Linux

[jaimeehh]

Dear NiftyPET,

I am trying to install NiftyPET on a Linux system using a Conda environment, but I am encountering an issue during the final installation step.

Steps I followed:

1. Created a new Conda environment:

   conda create --name niftytest  

2. Activated the environment:

   conda activate niftytest  

3. Installed dependencies:

   conda install -c conda-forge python=3.10 numpy scipy scikit-image matplotlib nibabel pydicom  
   conda install -c conda-forge python=3.10 ipykernel ipywidgets  

4. Set environment variables:

   export PATHTOOLS=$HOME/NiftyPET_tools  
   export HMUDIR=$HOME/mmr_hardwareumaps  

5. Installed required packages:

   pip install --verbose nimpa  
   pip install --verbose nipet  # This is where I encountered the issue  

The error occurs at the last step when trying to install nipet. I have attached the error log file for reference.

System Information:

• OS: Linux (please specify distribution if needed)
• Python version: 3.10 (Conda environment)

I would appreciate any guidance on how to resolve this issue.

Best regards,
Jaime

error_lo3345g.txt

[OliJimbo]

Hi Jaime,

The following got me up and running on docker if that is of any help?

Pull docker image from a previous issue (https://github.com/NiftyPET/NIPET/pkgs/container/nipet):

docker pull ghcr.io/niftypet/nipet:latest

Then you need to set up docker to use cuda (https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/latest/install-guide.html)

curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg \ && curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | \ sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \ sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list

sudo apt-get update

sudo apt-get install -y nvidia-container-toolkit

sudo nvidia-ctk runtime configure --runtime=docker

sudo systemctl restart docker

As per this: https://forums.developer.nvidia.com/t/nvida-container-toolkit-failed-to-initialize-nvml-unknown-error/286219/2
a parameter needs changing in /etc/nvidia-container-runtime/config.toml

no-cgroups = false

Restart Docker
sudo systemctl restart docker

Run docker image with the following parameters:

sudo docker run -it --runtime=nvidia --gpus all ghcr.io/niftypet/nipet

>>> from niftypet import nipet
100%|████████████████████████████████████████| 782k/782k [00:00<00:00, 21.2MB/s]
Extracting: 100%|██████████████████████████| 1.95M/1.95M [00:00<00:00, 69.6MB/s]
>>> from niftypet.nimpa import getnii
>>> from scipy.ndimage.filters import gaussian_filter
<stdin>:1: DeprecationWarning: Please use `gaussian_filter` from the `scipy.ndimage` namespace, the `scipy.ndimage.filters` namespace is deprecated.
>>> from tqdm.auto import trange
>>> import matplotlib.pyplot as plt
>>> import numpy as np
>>> import pydicom

>>> print(nipet.gpuinfo())
[('NVIDIA GeForce GTX 1660', 6144, 7, 5)]

BW

Oli

If Docker is of no use - then these are some of the steps I've tried which may or may not be of use to you. Note that this has been compulsive trial and error rather than planned solutions:

The issue seems to be with the building of the cuvec package from the NiftyPet wheel.

I am able to build cuvec from the source (AMYPAD/cuvec), and am also able to build the individual packages (nipet and nimpa) separately by cloning the repos and removing cuvec as a dependency from the setup.cfg, pyproj.toml, and setup.py files. However only nimpa will run because doing it this way ends up creating circular references - so I think it is essential to install the full NiftyPet module.

I have also tried forking the rep, removing all references to cuvec from each submodules build files, and then installing from pip from my own repo. No joy there.

I have also tried 'tricking' pip into thinking cuvec is installed by moving the package into site-packages and then adding an additional folder called cuvec-6.0.0.dist-info with a METADATA file in it with the following:

Metadata-Version: 2.1
Name: cuvec
Version: 6.0.0

I think this ends up giving a different error message:
/tmp/pip-install-roqcn52a/nipet_cb09cf5bcb594b92a6a91636bf11861e/niftypet/nipet/prj/src/prj_module.cu:14:10: fatal error: pycuvec.cuh: No such file or directory

So I'm now thinking that I need to change the cfg file to require cuvec for setup but not installation. I'm not sure.

[pjmark]

Hi Oli What system are you running it on? Windows or Linux?

[OliJimbo]

This is Linux - Ubuntu 22:04LTW. I've not been able to try it on any other systems since this is the only one I have access to that has an Nvidia gpu

[pjmark]

You shouldn't have any problems. I can help you with this. Contact me via my email.

[jaimeehh]

Hi Oli and Pawel,

Thank you so much for your response and for taking the time to help. I’ve been trying to resolve this issue for many days, but despite all my efforts, I haven’t been able to find a solution. It seems to be related to a subprocess within the pyproject.toml file. I’ve tested different approaches through trial and error, but nothing has worked so far.

I really appreciate the suggestion regarding Docker—I’ll give it a try and see if it helps!

Just in case it’s relevant, here are the details of my system, in case anyone has any insights that might help:

• Operating system: Ubuntu 22.04.5 LTS
• NVIDIA driver version: 470.256.02
• CUDA version: 11.4
• Main error message: ERROR: Could not build wheels for cuvec, which is required to install pyproject.toml-based projects

If you have any suggestions or ideas on what I could try next, I’d really appreciate it.

Thanks again for your help!

Best,
Jaime

[pjmark]

I had this error before - make sure you have cmake installed, relatively recent. Even though this is a Python package, the core programs are written in CUDA C, requiring usual compilation. Happy to connect with you to help with it. I’m also updating the software as other important functions from external packages are deprecated.

[jaimeehh]

Thank you very much for your suggestion! I’ll give it a try over the next few days and see if it helps. I really appreciate your support.

Let’s stay in touch, and I’ll reach out if I need further assistance.

Thanks again!

Best,
Jaime

[OliJimbo]

You shouldn't have any problems. I can help you with this. Contact me via my email.

Thanks Pawel - I think this is probably more of a cuvec issue so I've opened an issue on that repo instead.

[OliJimbo]

Hi @jaimeehh - not sure if you've had any luck yet - but I opened an issue on the cuvec git:
AMYPAD/CuVec#47 - there might be a solution for you over there.

[OliJimbo]

I can now build cuvec thanks to Casper's advice on this issue: AMYPAD/CuVec#47
However niftypet still tries to build cuvec on installation and running pip with no dependencies only installs niftypet (not NIMPA etc).

[jaimeehh]

Hi @oli and @pawel,

Sorry for the delay in getting back to you! I recently managed to install CuVec after applying the update, which allowed me to fully install NiftyPET. Thanks a lot for your help and for opening the issue on the CuVec GitHub—it was really useful.

Regarding your issue, I didn't run into the problem where NiftyPET tries to rebuild CuVec on installation. Maybe you could try reinstalling everything or using the forced CuVec version as suggested in the issue thread. Also, make sure to remove any previous versions of CuVec before reinstalling. You can try the following:

pip uninstall niftypet cuvec -y
pip install --no-deps cuvec
pip install niftypet

For reference, I installed everything using CUDA 11.4, and my system was recognizing it correctly. Maybe forcing the correct version of CuVec and ensuring all dependencies are installed could help.

That said, while running the NiftyPET example tutorial, I encountered an issue where my GPU is not being recognized. Specifically, I get an error at this line:

print(nipet.gpuinfo())

Do you have any suggestions on how to fix this? Could this be related to my system configuration or the CUDA version? I’d really appreciate any insights you might have.

Thanks again for your support! 😊