寒武纪显卡MLU编译安装mmcv1.7.0、mmdetection2.26.0并测试

一、 虚拟环境搭建与torch_mlu安装

conda create -n pytorch1.9_mlu python=3.7 -y   #创建环境
source activate pytorch1.9_mlu                #激活环境

cd 到mlu的指定目录下进行pytorch和pytorch_mlu安装

首先安装torch,然后torch_mlu，最后torchvision
pip installl torch-1.9.0-cp37-cp37m-linux_x86_64.whl
pip install torch_mlu-1.10.0_torch1.9-cp37-cp37m-linux_x86_64.whl
pip install torchvision-0.10.0a0+300a8a4-cp37-cp37m-linux x86 64.whl

vim ~/.bashrc #将寒武纪驱动写入

然后验证：

如果报错OSError: libmpi.so.40: cannot open shared object file: No such file or directory

解决

conda install openmpi

二、 install mmcv-full on Cambricon MLU device

问题 ImportError libtorch_mlu.so: cahnot open shared object file: No such file or directory

export LD_LIBRARY_PATH=/home4/sre/hyliu41/anaconda3/envs/pytorch1.9_mlu/lib/python3.7/site-packages/torch_mlu/csrc/lib:$LD_LIBRARY_PATH

安装mmdetection2.26

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -v -e .

验证安装

修改：

mmcv1.7.0不支持dcn在mlu后端运行

参考，目前mmcv2.1.0版本支持
将这部分函数修改加入的mmcv1.7.0中