import _C.nms causes undefined symbol error

[nikisalli]

maybe some useful build flags are missing

[WillLiGitHub]

setup falied.

[ddshan]

Have you solved the issue? Usually, it is related to building the environment. Could you try the steps carefully again? Otherwise, could you share the details about the issue?

[nikisalli]

I followed all the steps carefully again, everything compiles but the demo fails with an undefined symbol error: CUDA_VISIBLE_DEVICES=0 python demo.py --cuda --checkepoch=8 --checkpoint=132028

  File "demo.py", line 34, in <module>
    from model.roi_layers import nms
  File "/home/nik/Desktop/progetti/hand_object_detector/lib/model/roi_layers/__init__.py", line 3, in <module>
    from .nms import nms
  File "/home/nik/Desktop/progetti/hand_object_detector/lib/model/roi_layers/nms.py", line 3, in <module>
    from model import _C
ImportError: /home/nik/Desktop/progetti/hand_object_detector/lib/model/_C.cpython-36m-x86_64-linux-gnu.so: undefined symbol: _ZNSt15__exception_ptr13exception_ptr10_M_releaseEv

[ddshan]

Can you take a look and try this fix (https://github.com/jwyang/faster-rcnn.pytorch/issues/503) from faster-rcnn.pytorch and see if it works?

[nikisalli]

I tried many versions of torch (1.0.0, 1.0.1, 1.3.0) and rebuilt what's under lib many times but I couldn't manage to solve the environment's issues. The best scenario I got was a numpy import error but it wasn't solveable due to many dependency conflicts and I couldn't reproduce it anyways.

[nikisalli]

this are my system's specs if it can be useful in any way Kernel: 5.15.2-arch1-1 x86_64 bits: 64 compiler: gcc v: 11.1.0 Distro: Arch Linux cuda 11.5.0-1 cudnn 8.2.4.15-1

[oliviaylee]

Encountering the same issue, are there any updates on how this was resolved?

[oliviaylee]

Update: this fix works for me for now. However, because it differs from the environment specifications in the ReadMe, I wouldn't be surprised if this caused errors for me further down the line, so I would not consider this resolved unless the code authors can verify that this environment setup also works for the codebase.