DCNv2 Compilation is successful but train.py fails

[niranjanravilevelup]

Hi,
Thank you for the repo. I am working with https://github.com/lbin/DCNv2 to build DCNv2.

I used this and did a git check out for version pytorch_1.6. When I did the ./make.sh, I was able to build successfully and no errors. This was done in pytorch version 1.7 and ./make.sh was modified with --user since I do not have root privilages.

Now, when I run the centerFuse train.py file, I face two problems.

In dcn_v2.py, import _ext as _backend fails and it displays "import DCN failed" and "NoneType" object is not callable. But _ext.so file exsists in the same directory but the code is unable to import it.

If i comment out import _ext as _backend, then import DCN becomes succesful and the error now is Torch not compiled with CUDA enabled.

I tried various work around and no fix.

If I compile ./make.sh with pytorch 1.7 (CUDA compiled) , does my train.py should alse be run in same pytroch 1.7 (CUDA compiled) ? Is this why I am facing the issue?

Or what other things I could be considering?

I a stuck in a loop here. Any help would be very much helpful.

Thank you.