Open lucasjinreal opened 2 years ago
Oh... Did you manage to fix it?
No, I suggest using torchvision.ops.boxes.nms to do nms operation, however, which needs modification on detector. But since Alphapose already included that source code, it also worthy to modificate to standard way
Please using my fork version: it works on torch 1.11
I cloned your repo:
git clone https://github.com/jinfagang/AlphaPose.git
And then ran:
python3 setup.py build develop --user
And I get the same error message:
detector/nms/src/nms_kernel.cu:5:10: fatal error: THC/THC.h: No such file or directory
5 | #include <THC/THC.h>
| ^~~~~~~~~~~
compilation terminated.
error: command '/usr/local/cuda/bin/nvcc' failed with exit status 1
Do I need to use a certain branch?
And it turns out that this is a real show-stopper for A40 and A100 cards. They require CUDA 11.1 or higher, thus PyTorch 1.11 or higher, which does not contain the THC component any more. Thus, as of now, I don't see how to run AlphaPose on recent cards.
@jinfagang Still the same error. You didn't modify the code in detector/nms/src/nms_kernel.cu
at branch master.
And it turns out that this is a real show-stopper for A40 and A100 cards. They require CUDA 11.1 or higher, thus PyTorch 1.11 or higher, which does not contain the THC component any more. Thus, as of now, I don't see how to run AlphaPose on recent cards.
@peteruhrig Any good suggestions?
I don't speak C/C++ well enough, but my understanding is that it is possible to replace the functions used with ones that are still supported, as discussed here: https://www.exxactcorp.com/blog/Deep-Learning/pytorch-1-11-0-now-available (scroll down to the CUDA section) My guess is that this is not very difficult for someone who knows what they are doing, and maybe that is what @jinfagang did but forgot to push back to his repo, but I don't understand it enough to be able to do this myself.
For decent pytorch version 1.11 (maybe 1.10), this file removed: