Hi wangyihuan123, this looks like a related issue from PyTorch GitHub, please see: https://github.com/pytorch/pytorch/issues/7680
So you may need to leave a comment about it or follow the patch from that issue for the function.
It would be really great if you could patch it for us… it is something very nano related.
I have managed to get yolov3 implementation on pytorch to infer 30fps which is better then anything else I tried, but they also have some SOTA fast models that would be very nice if I could get them to work, one is cornernet the other is fssd instead of ssd. both cause this error.
Building pytorch on the nano is a looooong compilation, and I am not that thrilled about the concept of setting up cross compiling now.
Your best course of action would be to file a bug against PyTorch GitHub (they are responsive in my experience), so they can fix the issue. Then once it’s fixed, we can post the pre-built pip wheel installer built from PyTorch master which picks up the change
there are over 2000 open issues currently on pytorch. I seriously doubt this will be a quick fix…
Oh well, I guess I’ll have to brave the recompilation. Basically only the getdeviceproperties needs to be changed.
If I’ll manage to do that, i’ll post the wheel. Hehehe I am reinventing the wheel
Thanks, dusty_nv, that link is truely helpful.
I have finished recompiling pytorch with num_threads patched to 512. However, I still have other problems. will keep you posted.
Hello wangyihuan123 and dusty_nv,
I have the same problem:
RuntimeError: cuda runtime error (7) : too many resources requested for launch at /media/nvidia/WD_BLUE_2.5_1TB/pytorch/aten/src/THCUNN/generic/SpatialUpSamplingBilinear.cu:67
But I can’t access to /media/nvidia/WD_BLUE_2.5_1TB/, which actually doesn’t exist in my local platform, do I need to correct the path of pytorch? How to do that?
I also download the source from github/pytorch/pytorch, but I can’t find the SpatialUpSamplingBilinear in pytorch/aten/src/THCUNN/generic, do I need to compile it first and then correct SpatialUpSamplingBilinear, and then recomplie pytorch again?
looking forward to your guidance.
Thank you in advance.
I rebuild the source code of pytorch, no errors now
@Xiao,cao
how do you rebulid the source code, i have exactly the same issue than you, and also i try rebuilding the source code from github but doesn’t work.
Hello Alex,
pls refer to “Build Instructions” in
https://devtalk.nvidia.com/default/topic/1049071/jetson-nano/pytorch-for-jetson-nano/