later versions of torch about cmrnet

[cattaneod]

@Sidu28 Thank you for your interest in our work!

The code should be compatible with newer version of Pytorch without problems. However, I can't test it right now, but I can do it next week.

Let me know if you have any problem.

[egirgin]

I get the error below and the only solution I found online is to upgrade the version of the torch. However, I could not run the correlation layer after upgrading. My torch is running without problem with last stable version. Is there anyone check with the last stable version ? (current its 1.10 for torch)

[cattaneod]

@egirgin I suppose that you updated PyTorch after installing the correlation package. in that case, you need to install again both correlation_cuda and visibility packages.

I just tested with PyTorch 1.10 and cuda 10.2 (cuda 11.x is not supported at the moment), and it works.
However I had to update the fille models/CMRNet/correlation_package/setup.py by replacing -std=c++11 with -std=c++14.

Let me know if you are able to solve the problem.

PS: you also need to update the pip dependencies in requirements.txt with the correct version, such as cupy_cuda, torch, etc.

[egirgin]

RESOLVED:EDIT
The solution is to import torch before correlation package

OLD ISSUE
@cattaneod I reinstalled everything from scratch on a different machine with cuda 10.2 and torch 1.10 and the error I get changed. Now I get

ImportError /home/.../models/correlation_package/correlation_cuda.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN2at19UndefinedTensorImpl10_singletonE

Can you specify which dependencies should I replace ?
I changed the followings

cupy_cuda90==6.0.0 to cupy-cuda102
torch==1.0.1.post2 to torch
torchvision==0.2.2 to torchvision

Besides, I installed cudatoolkit=10.2 as it suggested in the pytorch getting started section.

should I change the torch version in the /models/correlation_package/pyproject.toml or nvcc_args in /models/correlation_package/setup.py?? I'm not familiar with the cuda layers tbh.

thanks in advance

[WJ1214]

RESOLVED:EDIT

The solution is to import torch before correlation package
OLD ISSUE @cattaneod I reinstalled everything from scratch on a different machine with cuda 10.2 and torch 1.10 and the error I get changed. Now I get

ImportError /home/.../models/correlation_package/correlation_cuda.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZN2at19UndefinedTensorImpl10_singletonE

Can you specify which dependencies should I replace ? I changed the followings

cupy_cuda90==6.0.0 to cupy-cuda102
torch==1.0.1.post2 to torch
torchvision==0.2.2 to torchvision

Besides, I installed cudatoolkit=10.2 as it suggested in the pytorch getting started section.

should I change the torch version in the /models/correlation_package/pyproject.toml or nvcc_args in /models/correlation_package/setup.py?? I'm not familiar with the cuda layers tbh.

thanks in advance

Hi! I meet the same problem, have you solved it?