Closed GeorgeAdamon closed 4 years ago
It seems that the Resample2d package is not installed. Try installing it separately and see if you get any errors in the process. For this you can run
python setup.py install
in the networks/resample2d_package
.
Thanks for the suggestion @zsameem . Running this gives me the following path-related error:
C:\Users\GeorgeAdamon\Documents\GitHub\flownet2-pytorch\networks\resample2d_package>python setup.py install
Traceback (most recent call last):
File "setup.py", line 6, in <module>
from torch.utils.cpp_extension import BuildExtension, CUDAExtension
File "C:\Python37\lib\site-packages\torch\utils\cpp_extension.py", line 61, in <module>
CUDA_HOME = _find_cuda_home()
File "C:\Python37\lib\site-packages\torch\utils\cpp_extension.py", line 30, in _find_cuda_home
if not os.path.exists(cuda_home):
File "C:\Python37\lib\genericpath.py", line 19, in exists
os.stat(path)
TypeError: stat: path should be string, bytes, os.PathLike or integer, not list
My guess is that you don't have the proper setup to compile the CUDA code. I was able to get this running on windows once. You will need the CUDA Toolkit, MS Visual Studio C++ and also the CUDA plugin for Visual Studio. Make sure you have all these things installed and then try running the setup.py again as mentioned in my previous comment.
That's really strange, since I'm actually writing & compiling custom C++ CUDA plugins just fine, and I haven't had problems compiling NVIDIA libraries like Optix.
nvcc -V
C:\Users\GeorgeAdamon\Documents\GitHub\flownet2-pytorch\networks\resample2d_package>nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Wed_Oct_23_19:32:27_Pacific_Daylight_Time_2019
Cuda compilation tools, release 10.2, V10.2.89
From your second post, it seems that the build tool of torch is having trouble finding/setting the CUDA_HOME path. I noticed that your torch version is very old. Did you also try with a newer version of torch. If I remember correctly, it simply worked with torch 1.2/1.3 for me.
Thank you @zsameem !
Upgrading to latest pytorch via
pip install torch===1.5.0 torchvision===0.6.0 -f https://download.pytorch.org/whl/torch_stable.html
fixed the issue!
I also had to go to networks/correlation_package
and networks/channelnorm_package
and "manually" run python setup.py install
.
Finally I need to report that using the latest version of scipy (1.4.0) resulted in scipy.misc.imread
errors, because imread
is deprecated. I had to go back to scipy 1.1.0 for everything to work.
So for future reference, I confirm that the following combination works: | Tool | Version |
---|---|---|
Windows | 10 | |
Visual Studio | 2017 | |
CUDA | 10.2 | |
python | 3.7 | |
pytorch | 1.5.0 | |
scipy | 1.1.0 | |
numpy | 1.17.2 |
I'm trying to run main.py right after running install.sh, and I get the following 2 errors:
The second one is an easy fix, I just removed the relative imports from Resample2d.
But I have no idea how to fix the first one. Apparently resample2d_cuda.py never gets generated when I run install.sh.
Any ideas would be much appreciated.