-
Notifications
You must be signed in to change notification settings - Fork 1.3k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
spconv v1.1 unknown device type #70
Comments
Hi MartinHahner88, Was your build spconv 1.1 successful (python setup.py bdist_wheel) and could you list out the versions that you were using for the build especially the version of CUDA and torch you were using for building spconv1.1 Thanks |
Yes, my build of spconv v1.1 was successful.
When I try to build spconv v1.0 (commit 8da6f96) I am stuck with this error message:
|
Hi , !wget -c https://repo.continuum.io/archive/Anaconda3-5.1.0-Linux-x86_64.sh and then I ran the first command you gave ERROR conda.core.link:_execute(481): An error occurred while installing package 'conda-forge::cudatoolkit-dev-10.1.243-h516909a_3' Could you please help? And could you also please share the build log for spconv1.1 that was successful so that we could check if the following line has detected the GPU.. -- Automatic GPU detection failed. Building for common architectures. Thanks |
I think your issue is google colab related and I think this repository also is not really meant to be used by google colab, but I don't have much experience regarding that. Did you make sure that you are connected to a GPU under |
@MartinHahner88 I got exactly the same error when trying to use spconv 1.1 (in a different project that originally used spconv 1.0). I also had trouble installing spconv 1.0, but switching to pytorch 1.1.0 solved that issue for me. |
Unfortunately downgrading to pytorch 1.1.0 did not resolve the issue for me, But instead, I was able to build spconv v1.0 with a fairly up-to-date conda environment, namely:
For more details, see here. But now, after downgrading to spconv v1.0, I can only successfully train PointPillars and SECOND,
My conda environment looks like this:
|
@fregu856: Which version of |
|
Finally SOLVED! The final issue was, that we tried to build spconv on a Tesla K40 GPU, The reason why we always tried to build spconv on a Tesla K40 is that in our lab, it is hard to get a GPU on our GPU cluster interactively. (Usually, you can only submit jobs via qsub/SLURM.) So here is our solution:
Test spconv via:
Done! I hope these instructions help someone else who struggles to build spconv as well. |
Damn, I am stuck in the same place and know this with that old K40m cluster.. |
I successfully make it work by recompiling spconv with GPU (I didn't call GPU in the beginning so it failed). In case some people might have the same situation as me: make sure you should call the GPU when compiling the spconv. Don't forget to clean the cache of previous compiling by 'python setup.py clean' and also 'pip uninstall spconv' before the recompling. |
Did anyone run into this problem and was able to solve it?
It only occurs when training SECOND and PartA^2Net
because they are utilizing spconv in their RPN backbone,
training PointPillars, on the other hand, works fine because it's RPN (PointPillarsScatter)
does not utilize spconv.
Is it a phenomenon only occurring with spconv v1.1 ?
The text was updated successfully, but these errors were encountered: