NVlabs/FocalFormer3D

the compile code of bevpool and localattn

SISTMrL opened this issue · 8 comments

SISTMrL commented

Hello, in your project/mmdet3d_plugin/models/utils/ops, there are bevpool and localttn_ops, could you please provide the code to compile these two module? And could you please provide your gcc version? Thanks very much

chenyilun95 commented

Hi, these two modules will be compiled automatically. The GCC version I used is 9.3.0. You can just install it with anaconda.

SISTMrL commented

Hi, these two modules will be compiled automatically. The GCC version I used is 9.3.0. You can just install it with anaconda.

hello, it's strange to me, i can compile the local_attn.so from the source code that comes from "https://github.com/zzd1992/Image-Local-Attention", but i can't compile the "*.so" file when run the focalformer3d code the problem i was encountered is shown as below. image
when i convert the gcc version to 4.9 it shows that the gcc version is too old, need >5

the environment i used is : torch 1.9.0+cu111，the others is same with you. but i want to verify the gnu version you have used

and the last: could you please provide the local_attn.so and bev_pool.so, if i can use them.

SISTMrL commented

Sure you can download them here:

https://drive.google.com/drive/folders/1dvQEYPY-2MM4s2AgFzgmG84ggwygxnr-?usp=drive_link

hello, i put the localattn.so into "models/utils/ops/locatt_ops", and annotate all the code in "__init.__py" to avoid automatically compiling when i run the code. when i run the command of "bash tools/dist_train.sh projects/configs/focalformer3d/FocalFormer3D_LC.py 1", the error i was encountered is shown as follow:
image
and my folder is shown as follows:
image
i just modify the FocalFormer3D_LC.py as follows:
image

chenyilun95 commented

Does the bev_pool module work? I found you used cuda-8.0 in your path. What's your nvcc version?

SISTMrL commented

Does the bev_pool module work? I found you used cuda-8.0 in your path. What's your nvcc version?

i don't see the bev_pool in the wrong message

i run the nvcc -V, the cuda version is 11.4
image

chenyilun95 commented

Perhaps you can go through the following tutorial to identify the problem.

https://pytorch.org/docs/stable/cpp_extension.html
You can find the use of compiling function in locatt_ops/init.py.

SISTMrL commented

i will have a try