Import Error mmdet/ops/dcn/deform_conv_cuda/ undefined symbol: _ZNSt15__exception_ptr13exception_ptr9_M_addrefEv

[hiepbk]

Hi,
Thanks for great work.
I tried to run cityscape_demo.py in tools and got this error.
Traceback (most recent call last):
File "cityscapes_demo.py", line 15, in
from mmdet.apis import (
File "/home/deltax/EfficientPS/mmdet/apis/init.py", line 1, in
from .inference import (async_inference_detector, inference_detector,
File "/home/deltax/EfficientPS/mmdet/apis/inference.py", line 11, in
from mmdet.core import get_classes
File "/home/deltax/EfficientPS/mmdet/core/init.py", line 7, in
from .post_processing import * # noqa: F401, F403
File "/home/deltax/EfficientPS/mmdet/core/post_processing/init.py", line 1, in
from .bbox_nms import multiclass_nms
File "/home/deltax/EfficientPS/mmdet/core/post_processing/bbox_nms.py", line 3, in
from mmdet.ops.nms import nms_wrapper
File "/home/deltax/EfficientPS/mmdet/ops/init.py", line 2, in
from .conv import build_conv_layer
File "/home/deltax/EfficientPS/mmdet/ops/conv.py", line 4, in
from .dcn import DeformConvPack, ModulatedDeformConvPack
File "/home/deltax/EfficientPS/mmdet/ops/dcn/init.py", line 1, in
from .deform_conv import (DeformConv, DeformConvPack, ModulatedDeformConv,
File "/home/deltax/EfficientPS/mmdet/ops/dcn/deform_conv.py", line 11, in
from . import deform_conv_cuda
ImportError: /home/deltax/EfficientPS/mmdet/ops/dcn/deform_conv_cuda.cpython-37m-x86_64-linux-gnu.so: undefined symbol: _ZNSt15__exception_ptr13exception_ptr9_M_addrefEv

My conda env was built with Python 3.7, Pytorch 1.7.1+cud110.
I'm using RTX 3060 with cuda 11.6, ubuntu 22.04.
It was successfully built efficientNet and mmdet source when I run python setup.py develop.
I could see some .so files were generated in the mmdet/ops.
I searched some solution, this issue might come from new GPU like RTX 3060 could not work with old version of mmdet.
How can I solve this issues ?

[hiepbk]

I solved this problem:

• Firstly, the Cuda version in host ubuntu and conda env should be same (in my case is cuda 11.0). I have some problems when I install cuda 11.0 in Ubuntu 22.04 because the default GCC version of Ubuntu 22.04 is GCC-11, and cuda 11.0 just supports the maximum version of GCC 9. So I need to change GCC 11 to GCC 9. Refer Using multiple G++. Then I could install cuda11.0 properly. Remember to change your cuda path in ~/.bashrc to 11.0 and source it again.
• After installing cuda 11.0. If you use the RTX series, you might encounter the issues of nvcc fatal : Unsupported gpu architecture 'compute_86' when you run install setup.py develop. The solution is comment line 1441 in cpp_extension of pytorch package -> go to your python compiler directory, for example /home/user/miniconda3/envs/your_env/lib/python3.7/site-packages/torch/include/torch/extension.h, and comment this line flags.append('-gencode=arch=compute_{},code=sm_{}'.format(num, num))