ImportError: Cannot import 'detrex._C', therefore 'MultiScaleDeformableAttention' is not available.

[Sjyhne]

ImportError: Cannot import 'detrex._C', therefore 'MultiScaleDeformableAttention' is not available.

[rentainhe]

@Sjyhne Hello! Please make sure you have CUDA runtime, you can check it by print CUDA_HOME to see if it's None, If None, you can set it by export CUDA_HOME=/path/to/cuda-11.3/ (just an example) and rebuild detrex again

[rentainhe]

Hello!

I'm following the installation guide, but facing issues when trying to run models with MultiScaleDeformableAttention. As the title states, I'm receiving the following error:

ImportError: Cannot import 'detrex._C', therefore 'MultiScaleDeformableAttention' is not available.

Relevant packages from conda list:

detectron2 0.6 dev_0 detrex 0.1.0 dev_0 pytorch 1.10.1 py3.7_cuda10.2_cudnn7.6.5_0 pytorch pytorch-mutex 1.0 cuda pytorch torchaudio 0.10.1 py37_cu102 pytorch torchvision 0.11.2 py37_cu102 pytorch

1. Clean conda environment with python version 3.7
2. Install pytorch according to the PyTorch installation guides.
3. Clone detrex
4. Init and update submodules
5. python -m pip install -e detectron2
6. pip install -e .

However when running:

python tools/train_net.py --config-file projects/dino/configs/dino_r50_4scale_12ep.py

I get the error above.

Is there any known solutions for this issue?

You can also follow this issue to see if it can help you to fix the installation problem. #85 (comment)

[Sjyhne]

Yeah, thanks! This might be my issue. However, I am not able to access the CUDA_HOME, as cuda is not installed locally on the server I'm training on. Are there any other possible solutions to the issue? Circumventing the CUDA_HOME?

[rentainhe]

Yeah, thanks! This might be my issue. However, I am not able to access the CUDA_HOME, as cuda is not installed locally on the server I'm training on. Are there any other possible solutions to the issue? Circumventing the CUDA_HOME?

I'm not sure if you're using slrum for training~, maybe you can install CUDA Runtime locally and successfully build detrex in your local environment and using this environment on the server, or build detrex directly on the server

[rentainhe]

Yeah, thanks! This might be my issue. However, I am not able to access the CUDA_HOME, as cuda is not installed locally on the server I'm training on. Are there any other possible solutions to the issue? Circumventing the CUDA_HOME?

Maybe you can try to build detrex on that server~, find the right cuda path and then export CUDA_HOME=/path/to/cuda

[rentainhe]

As there is no more activity, I am closing the issue~ Feel free to reopen it if necessary. Or you can leave a new issue if you meet some other problems.