undefined symbol: __cudaPopCallConfiguration

[xyang35]

I've successfully compiled the cuda dependencies for ROI pooling. Everything looks fine but when I try to call the function RoIPoolFunction(), it comes an error:

ImportError: /home/faster-rcnn.pytorch/lib/model/roi_pooling/_ext/roi_pooling/_roi_pooling.so: undefined symbol: __cudaPopCallConfiguration

Any idea about the problem ? I'm using CUDA 9.2 and pytorch 0.4.

Thanks!

[xyang35]

I changed to CUDA 9.0, and it seems that the problem has gone...

[ghost]

I didn't have any luck moving to cuda 9.0 as well! Anybody else facing the same issue?

[abhishekcvedia]

@xyang35 @codewithsk Any solutions to this? Changing CUDA version is not helping me.

[ghost]

@abhishekcvedia turns out I'd forgotten to set the CUDA PATH in my make file. In my particular case, the gcc version also had to be changed to 4.5 (if I recollect correctly)

[sh8]

I got the same error when using CUDA9.2. After changing the version to CUDA8.0, the issue has gone.

[zhangyuygss]

I had the same problem when using pytorch0.3.1 with cuda9.2, when I changed to torch0.4.1, the issue fixed. I guess the mismatch of pytorch installation(pytorch0.3 with cuda90) and cuda(9.2) caused the problem, since pytorch0.3 doesn't have a cuda92 version, we should either use newer pytorch or older cuda.

[VladVin]

The answer of @zhangyuygss helped me: PyTorch 0.4.1 + CUDA 9.2 worked.

[yuqli]

I use PyTorch 0.4.1 post2 + CUDA 9.2 and it gave the same error.

[wu-yy]

I also meet the problem. SO I uninstall the cuda-9.2 and install the cuda-9.0 . I deal with this problem .

[unclecao]

Also, when I uninstall the cuda-9.2 and instead install the cuda-9.0, then the problem is fixed. I don't know why

[Harathi123]

I got similar issue with CUDA 9.2 + Pytorch 0.4. It got fixed with CUDA 9.0

[yuqli]

I used Pytorch 0.4.0 + CUDA 8.0 and it worked.

[BPiepmatz]

For people with CUDA 9.2 with pytorch 0.4.1: I could solve the issue by installing with conda instead of pip

[muzammil360]

I am also facing the same issue. I am using pytorch '1.0.1.post2' and cuda92. I tried cuda90 but installation gives huge errors. I can't try cuda80 because I can't install it on cluster.

Has anyone been able to run this project with pytorch 1?

[Barbarioli]

Same problem here. Cuda 9.2 Pytorch 1.0.1

[maxfaine]

You need to verify that torch compiled with the correct cuda version. If you have different versions it will not work. You can download the correct precompiled torch pkg that suites cuda from pytorch.org.

Please see the following link for help: https://rusty1s.github.io/pytorch_geometric/build/html/notes/installation.html

[jshi31]

I find the root of the problem is the mismatch of the nvcc version and the torchcuda version. You must make sure they are the same. Check nvcc version: nvcc --version Check torch cuda version, go to the python interface and type import torch torch.version.cuda make sure they are the same and I solved my problem.

[CodHeK]

I am facing the same issue!

nvcc --version -> 9.2

torch.version.cuda -> 9.0

torch.__version__ -> 0.4.1

So I might have to change the nvcc version right? How can install CUDA 9.0 having already installed CUDA 9.2 ?

Thanks!

[muzammil360]

Instead of changing cuda version, Try installing pytorch again.

On Tue, Oct 8, 2019, 7:33 PM Gagan Ganapathy notifications@github.com wrote:

I am facing the same issue!

nvcc --version -> 9.2

torch.version.cuda -> 9.0

torch.version -> 0.4.1

So I might have to change the nvcc version right? How can install CUDA 9.0 having already installed CUDA 9.2 ?

Thanks!

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/jwyang/faster-rcnn.pytorch/issues/190?email_source=notifications&email_token=AE3S5IPOFGQIWU62ZWMMGUTQNSK5JA5CNFSM4FDWDXYKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEAUMKUA#issuecomment-539542864, or mute the thread https://github.com/notifications/unsubscribe-auth/AE3S5IKHJPJX4HUIWDPRDHDQNSK5JANCNFSM4FDWDXYA .

[CodHeK]

What combination of pytorch should I be using instead of 0.4.1? On Wednesday, 9 October, 2019, 08:49:39 am IST, muzammil360 notifications@github.com wrote:

Instead of changing cuda version, Try installing pytorch again.

On Tue, Oct 8, 2019, 7:33 PM Gagan Ganapathy notifications@github.com wrote:

I am facing the same issue!

nvcc --version -> 9.2

torch.version.cuda -> 9.0

torch.version -> 0.4.1

So I might have to change the nvcc version right? How can install CUDA 9.0 having already installed CUDA 9.2 ?

Thanks!

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/jwyang/faster-rcnn.pytorch/issues/190?email_source=notifications&email_token=AE3S5IPOFGQIWU62ZWMMGUTQNSK5JA5CNFSM4FDWDXYKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEAUMKUA#issuecomment-539542864, or mute the thread https://github.com/notifications/unsubscribe-auth/AE3S5IKHJPJX4HUIWDPRDHDQNSK5JANCNFSM4FDWDXYA .

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[Swjtu-only]

just you run conda install pytorch=0.4.1 cuda92 -c pytorch conda install pytorch=0.4.1 cuda(your nvcc-V) -c pytorch

[1194949000]

I find the root of the problem is the mismatch of the nvcc version and the torchcuda version. You must make sure they are the same. Check nvcc version: nvcc --version Check torch cuda version, go to the python interface and type import torch torch.version.cuda make sure they are the same and I solved my problem.

I changed my CUDA version as you suggested and solved my confusion.Thank you very much!

[Vio1etovo]

I change all of it, nvcc -V ->release 9.0, V9.0.176 torch.version.cuda -> 9.0.176 torch.version ->'0.4.1' but this changing is not helping me. I don't know why even I changed gcc5.4 -> 4.8

[ajithvcoder]

@Vio1etovo "Pytorch 0.4.0 is only supported ( it does not support 0.4.1 or higher)" its menitoned in readme. u can try with cuda 8 and torch 0.4.0 .

[henbucuoshanghai]

I use cuda 10.0 cudnn10.0 totch1.1 but it still this error

from torch.utils.data.dataloader import numpy_type_map ImportError: cannot import name 'numpy_type_map'

[ajithvcoder]

@henbucuoshanghai change to cuda 8.0 and gcc 5+ and torch 0.4.0 it works

[henbucuoshanghai]

but my ubuntu is cuda 10,i have to change it ????when i use 10,and i have to change it back to 10.

[ajithvcoder]

you can do it in colab if you share your email id i can send you the notebook. if you want to do it for cuda 10 , yes you have to either change to code to adapt to cuda 10 or else revert to cuda 8 or cuda9

[hanhao2019102935]

you can do it in colab if you share your email id i can send you the notebook. if you want to do it for cuda 10 , yes you have to either change to code to adapt to cuda 10 or else revert to cuda 8 or cuda9 I tried to experiment this project on CoLab, but it failed because of CUDA. Could you please send me your Notebook?My email address is 178337111@qq.com.My VX: H178337111.Looking forward to your reply

[ajithvcoder]

@hanhao2019102935 i dont have notebook it has been a long time but for chaning cuda versions you can refer this .https://forums.developer.nvidia.com/t/install-cuda-8-0-on-google-colab/112482

[hanhao2019102935]

@ajithvallabai First of all, thank you for telling me about the installation of CUDA.But when I implemented this code on coLab.I installed Torch0.4.0 successfully, but when I used print(torch.version) I found that the Torch0.4.0 version had not changed.I installed Cuda9.0 Python 3.6, but the Torch issue never resolved

[hanhao2019102935]

@ajithvallabai I think I know my problem. Now the Python version of coLab is 3.7. Python 3.7 will not run this code.I tried to use Conda to modify the Python version to 3.6, but apparently I only saw superficial changes when I tried to install a torch for 3.6, and it failed.Now I have no way to run it on coLab.

[onkarkris]

torch.version.cuda I have torch.version.cuda '9.0.176'

and nvcc --version Cuda compilation tools, release 9.0, V9.0.176

but still error persist ImportError: lib/model/nms/_ext/nms/_nms.so: undefined symbol: __cudaPopCallConfiguration