A problem about nms

[bobwan1995]

Hi Rowan, thanks for your codes sharing. However, I find nms.nms_apply(keep, boxes_sorted, nms_thresh) in lib/fpn/nms/functions/nms.py would make pytorch break up. Before I execute that line, the operation of Tensor.cuda()+int() is legal, but after I execute that line this operation will raise error: RuntimeError: cuda runtime error (48) : no kernel image is available for execution on the device at /pytorch/torch/lib/THC/generic/THCTensorMathPairwise.cu:21 when it goes on keep_im + s at line 24. Could you please help me? Thanks!

[rowanz]

what GPU do you have? I think you might have to change the compilation flags; I hardcoded them for my Titan X setup.

[bobwan1995]

Thanks for your kind answering.
My GPU is Tesla P100, so I tried to make the cuda file with /usr/local/cuda/bin/nvcc -c -o file.cu.o file.cu --compiler-options -fPIC -gencode arch=compute_60,code=sm_60, but it still fails with error cudaCheckError() failed : no kernel image is available for execution on the device. I think it might be compilation problem. Except for changing all sm_61 and compute_60 to *_61, do I need to change other places? I've replaced nms and roi_align with other's codes and it works. Thank you a lot!

[rowanz]

Yeah, you'll need to change the LSTM compilation: https://github.com/rowanz/neural-motifs/blob/master/lib/lstm/highway_lstm_cuda/make.sh

I believe that should be everything. Also, have you verified that pytorch usually runs on your machine?

[bobwan1995]

I use pytorch 0.3.1. Now I can run your code, thank you!

[YiwuZhong]

Hi, I've met the same problem using Tesla P100. So did you solve your problem only changing from _61 to _60 in 3 files (nms, roi_align and highway_lstm_cuda)?
And you said "replaced nms and roi_align with other's codes"， could you please tell what "other's codes" did you use to solve the problem? Thanks. :D

[alibabadoufu]

For those who are still solving this issues, here are the steps you should consider to follow:

1. Change SM_# (at places listed below) that suits your current GPU cards (see this for your reference)

• lib/lstm/highway_lstm_cuda/make.sh (line 15)
• lib/fpn/roi_align/src/cuda/Makefile (line 2)
• lib/fpn/nms/src/cuda/Makefile (line 2)

2. Afterwards, you should also delete *.o files (at places listed below) before you try recompiling nms, roi_align and highway_lstm)

• lib/lstm/highway_lstm_cuda/src/highway_lstm_kernel.cu.o
• lib/fpn/roi_align/src/cuda/roi_align.cu.o
• lib/fpn/nms/src/cuda/nms.cu.o

3. Run 'make' in your home directory.

That's all you need to do to change your faulty SM setup to the correct one.