compatible versions of PyTorch and torchkbnufft.

[RaziehEmami1]

Hi everyone,

I am trying to make a image using torch version 1.10.0 in CUDA and I am hoping to use torchkbnufft as a part of the analysis. I was wondering what is the compatible version of this with torchkbnufft. Otherwise the code crashes and is unable to call fft. I am currently trying torchkbnnufft== 1.3.0 in my remote builder.

Thanks a lot in advance.
RE.

[mmuckley]

I think it should work with PyTorch 1.10. We test using PyTorch 1.10 in our continuous integration. Could you try installing from master and seeing if you still have the fft error?

pip install --upgrade git+https://github.com/mmuckley/torchkbnufft

[RaziehEmami1]

Hi @mmuckley
many thanks for your kind guide. I will absolutely try and let you know what I get. Just to be sure, do not we need to specify the exact version of torchkbufft here?

Thanks,
Razi

[mmuckley]

No - that command will just pull whatever's in the main branch of GitHub. That way I can make sure that the code which is being tested in continuous integration is also not working for you.

[RaziehEmami1]

Fantastic.
Many thanks for your kind helps. I am trying now and will let you know what I get.

Cheers,
Razi

[RaziehEmami1]

Hi
so I tried with the version you said and here is the error that I received

from torchkbnufft import KbNufft, AdjKbNufft
ImportError: cannot import name 'AdjKbNufft' from 'torchkbnufft' (/opt/conda/lib/python3.9/site-packages/torchkbnufft/init.py)

I wonder does is have anythings to do with the python version or is related to the incompatibilities between the version of the pytorch and torchkbnufft?

Let me add that I am using A100 in the cluster. I wonder could that be that this is incompatible with the torch and torchkbnufft that we wish to use?

thanks.

[mmuckley]

Hello @RaziehEmami1, there is no AdjKbNufft in the package. That class is called KbNufftAdjoint (possibly renamed in a past version).

[RaziehEmami1]

Hi, Thanks a lot for your kind helps. Yes, it is working now. The new complain goes to the other lines of the code, as they seem to be renamed in the newer version from torchkbnufft.mri.dcomp_calc import calculate_radial_dcomp_pytorch from torchkbnufft.math import absolute from torch.utils.data import DataLoader, Dataset I was wondering may I ask your opinion about them, please? Here is the error: from torchkbnufft.mri.dcomp_calc import calculate_radial_dcomp_pytorch ModuleNotFoundError: No module named 'torchkbnufft.mri' Thanks a lot, Razi

…

On Apr 7, 2022, at 9:39 AM, Matthew Muckley ***@***.***> wrote: Hello @RaziehEmami1 <https://www.google.com/url?q=https://github.com/RaziehEmami1&source=gmail-imap&ust=1649943569000000&usg=AOvVaw0WS1ORZsZ4xBaF2FD6NWll>, there is no AdjKbNufft in the package. That class is called KbNufftAdjoint (possibly renamed in a past version). — Reply to this email directly, view it on GitHub <https://www.google.com/url?q=https://github.com/mmuckley/torchkbnufft/discussions/59%23discussioncomment-2523820&source=gmail-imap&ust=1649943569000000&usg=AOvVaw00iNjPvR5yC8oAgQSZ_RED>, or unsubscribe <https://www.google.com/url?q=https://github.com/notifications/unsubscribe-auth/AMG6ZJA6BJVAG4CZYPUHNCTVD3QQZANCNFSM5SUGXNZA&source=gmail-imap&ust=1649943569000000&usg=AOvVaw3knS3Ff_SW_3kilhqu0m5a>. You are receiving this because you were mentioned.

[mmuckley]

Please consult the documentation on readthedocs, which should be up-to-date.