Getting Error in CMake Libtorch

ritesh_gupta · July 12, 2019, 12:24pm

Hi,
I’m getting CMake error while doing cmake -DCMAKE_PREFIX_PATH=…/…/libtorch … . Yesterday I followed the same step, which is on Pytorch. I’m able to build it successfully on my laptop, but when I’m doing the same I’m getting an error like

I looked all the answer which are relevant to this, but all that didn’t let me to the conclusion. Can someone help me in this regard?
Thank you.

yf225 · July 12, 2019, 2:55pm

Have you checked and made sure that there is a libtorch installation at ../../libtorch? Also another thing to try is to remove the ~/example-app/build folder, recreate it again and then rerun the steps.

ritesh_gupta · July 17, 2019, 7:40am

Hi Will,
I tried the above steps mentioned by you, but it is throwing same error. What to do? I have downloaded the libtorch and extracted it in specified directory.

kevinhaoliu · July 17, 2019, 7:48am

Do you download the right version of libtorch? I download a windows version on my linux machine last time.

ritesh_gupta · July 17, 2019, 9:15am

Hi Kevin,
Yes i did. I also tried without CUDA version also, I’m getting the same error.

yf225 · July 17, 2019, 2:48pm

@ritesh_gupta Could you provide the result of ls ~/libtorch on your machine?

ritesh_gupta · July 17, 2019, 3:15pm

tom · July 17, 2019, 3:59pm

Wouln’t you need to point to the subdir containing the cmake files (share/cmake)?

ritesh_gupta · July 17, 2019, 4:31pm

Sorry, I didn’t get you?

ritesh_gupta · July 17, 2019, 4:46pm

Hi,
I think error could be because of cuda. As same program without CUDA Libtorch runs perfectly well(left terminal), whereas with CUDA Libtorch throws error(right terminal). Any help would be welcome.

Thanks.

yf225 · July 17, 2019, 5:40pm

Thanks for debugging further. I think the answer to the previous problem is that we should use cmake -DCMAKE_PREFIX_PATH=/absolute/path/to/libtorch .. instead of cmake -DCMAKE_PREFIX_PATH=relative/path/to/libtorch ... Putting it here so that other users could see the solution.

tom · July 17, 2019, 5:41pm

Do you have the libculibos.so in the directory it’s looking in? Or one nearby (like cuda-10)?

ritesh_gupta · July 18, 2019, 4:10am

No, I don’t have that file in both places.

ritesh_gupta · July 18, 2019, 4:14am

Thanks Will, But the screenshot I posted has an absolute path, not a relative path. Also any idea on how to correct the library error? I tried giving

by following the command given at Library in ... hidden by files in ... - Jetson TX2 - NVIDIA Developer Forums but to no luck.

yf225 · July 18, 2019, 3:47pm

@ritesh_gupta Did you make sure that CUDA is correctly installed on your machine?

ritesh_gupta · July 19, 2019, 9:14am

@yf225 I’m using DGX server which already has preinstalled CUDA. I can run the pytorch model on GPU. So I think so the installation of CUDA is an issue out here.

beyondpie · July 19, 2019, 9:25am

@ritesh_gupta
Does the link below help you? I think you need to check to make sure that /usr/local/cuda/ or /usr/local/cuda-10.0 you are using.