A quick & dirty fix.
mmcv-full 1.7.2, mmlvis 10.5.3 and mmpycocotools 12.0.3 have to be built from source, and pycocotools should be uninstalled manually.
When building mmpycocotools, there is an error in linking, caused by two maskApi.o. The quick and dirty fix is running:
cc -pthread -B /home/akiko/miniconda3/envs/bonai/compiler_compat -shared -Wl,-rpath,/home/akiko/miniconda3/envs/bonai/lib -Wl,-rpath-link,/home/akiko/miniconda3/envs/bonai/lib -L/home/akiko/miniconda3/envs/bonai/lib -Wl,-rpath,/home/akiko/miniconda3/envs/bonai/lib -Wl,-rpath-link,/home/akiko/miniconda3/envs/bonai/lib -L/home/akiko/miniconda3/envs/bonai/lib build/temp.linux-x86_64-cpython-310/../common/maskApi.o build/temp.linux-x86_64-cpython-310/pycocotools/_mask.o -o build/lib.linux-x86_64-cpython-310/pycocotools/_mask.cpython-310-x86_64-linux-gnu.so
then pip install -e .
When building mmcv-full, use the following commands:
MMCV_WITH_OPS=1 FORCE_CUDA=1 pip install -e . -v
You need an environment with CUDA=11.7 and pytorch=1.13.1. A working nvcc=11.7.x is required. If your system have both CUDA 12 and 11, do following:
export PATH="/usr/local/cuda-11/bin:$PATH" export LD_LIBRARY_PATH="/usr/local/cuda-11/lib64:$LD_LIBRARY_PATH"
Conda will not solve everything here.
A quick & dirty fix. mmcv-full 1.7.2, mmlvis 10.5.3 and mmpycocotools 12.0.3 have to be built from source, and pycocotools should be uninstalled manually. When building mmpycocotools, there is an error in linking, caused by two maskApi.o. The quick and dirty fix is running:
cc -pthread -B /home/akiko/miniconda3/envs/bonai/compiler_compat -shared -Wl,-rpath,/home/akiko/miniconda3/envs/bonai/lib -Wl,-rpath-link,/home/akiko/miniconda3/envs/bonai/lib -L/home/akiko/miniconda3/envs/bonai/lib -Wl,-rpath,/home/akiko/miniconda3/envs/bonai/lib -Wl,-rpath-link,/home/akiko/miniconda3/envs/bonai/lib -L/home/akiko/miniconda3/envs/bonai/lib build/temp.linux-x86_64-cpython-310/../common/maskApi.o build/temp.linux-x86_64-cpython-310/pycocotools/_mask.o -o build/lib.linux-x86_64-cpython-310/pycocotools/_mask.cpython-310-x86_64-linux-gnu.so
then pip install -e . When building mmcv-full, use the following commands:MMCV_WITH_OPS=1 FORCE_CUDA=1 pip install -e . -v
You need an environment with CUDA=11.7 and pytorch=1.13.1. A working nvcc=11.7.x is required. If your system have both CUDA 12 and 11, do following:export PATH="/usr/local/cuda-11/bin:$PATH" export LD_LIBRARY_PATH="/usr/local/cuda-11/lib64:$LD_LIBRARY_PATH"
Conda will not solve everything here.